Electrical Engineering

Image Processing
Tensor Transform and Discrete
Tomography with MATLAB®
Image
Focusing on mathematical methods in computer tomography, Image
Processing: Tensor Transform and Discrete Tomography with
MATLAB® introduces novel approaches to help in solving the problem
of image reconstruction on the Cartesian lattice. Specifically, it
Processing
discusses methods of image processing along parallel rays to more Tensor Transform and Discrete
quickly and accurately reconstruct images from a finite number of
projections, thereby avoiding overradiation of the body during a
computed tomography (CT) scan. Tomography with MATLAB®
The book presents several new ideas, concepts, and methods, many
of which have not been published elsewhere. New concepts include
methods of transferring the geometry of rays from the plane to the
Cartesian lattice, the point map of projections, the particle and its X-ray set
field function, and the statistical model of averaging. The authors
supply numerous examples, MATLAB®-based programs, end-of-chapter
problems, and experimental results of implementation.
image f(n,m)
The main approach for image reconstruction proposed by the authors
differs from existing methods of back-projection, iterative reconstruction,
and Fourier and Radon filtering. In this book, the authors explain how
to process each projection by a system of linear equations, or linear
convolutions, to calculate the corresponding part of the 2-D tensor
or paired transform of the discrete image. They then describe how projection
to calculate the inverse transform to obtain the reconstruction. The ø=26.565º
proposed models for image reconstruction from projections are simple
and result in more accurate reconstructions.

Introducing a new theory and methods of image reconstruction, this

book provides a solid grounding for those interested in further research
and in obtaining new results. It encourages readers to develop effective detectors
applications of these methods in CT.

K14749
ISBN: 978-1-4665-0994-8
90000

9 781466 509948
Artyom M. Grigoryan n Merughan M. Grigoryan

K14749_Cover_mech.indd 1 8/7/12 10:48 AM

 Image
Processing
Tensor Transform and Discrete
Tomography with MATLAB
®

K14749_FM.indd 1 8/13/12 2:12:07 PM

This page intentionally left blank
 Image
Processing
Tensor Transform and Discrete
Tomography with MATLAB
®

Artyom M. Grigoryan n Merughan M. Grigoryan

Boca Raton London New York

CRC Press is an imprint of the

Taylor & Francis Group, an informa business

K14749_FM.indd 3 8/13/12 2:12:07 PM

MATLAB® is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks does not
warrant the accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB® soft-
ware or related products does not constitute endorsement or sponsorship by The MathWorks of a particular
pedagogical approach or particular use of the MATLAB® software.

CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2013 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20120719

International Standard Book Number-13: 978-1-4665-0995-5 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has
not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented,
including photocopying, microfilming, and recording, or in any information storage or retrieval system,
without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.
com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and
registration for a variety of users. For organizations that have been granted a photocopy license by the CCC,
a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
 v

TO STUDENTS IN ELECTRICAL ENGINEERING

COMPUTER SCIENCE
AND
RESEARCHERS WORKING IN COMPUTED TOMOGRAPHY

♣
This page intentionally left blank
Contents

Author Bios xi

Preface xiii

1 Discrete 2-D Fourier Transform 1

1.1 Separable 2-D transforms . . . . . . . . . . . . . . . . . . . . 2

1.2 Vector forms of representation . . . . . . . . . . . . . . . . . 4
1.3 Partitioning of 2-D transforms . . . . . . . . . . . . . . . . . 5
1.3.1 Tensor representation . . . . . . . . . . . . . . . . . . 8
1.3.2 Covering with cyclic groups . . . . . . . . . . . . . . . 9
1.4 Tensor representation of the 2-D DFT . . . . . . . . . . . . . 12
1.4.0.1 Code: Splitting-signal calculation . . . . . . . 13
1.4.1 Tensor algorithm of the 2-D DFT . . . . . . . . . . . . 13
1.4.2 N is prime . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.2.1 Code: 2-D DFT by tensor transform . . . . . 20
1.4.3 N is a power of two . . . . . . . . . . . . . . . . . . . 21
1.4.4 N is a power of an odd prime . . . . . . . . . . . . . . 27
1.4.5 Case N = L1 L2 (L1 6= L2 > 1) . . . . . . . . . . . . . 29
1.4.6 General case . . . . . . . . . . . . . . . . . . . . . . . 29
1.4.7 Other orders N1 × N2 . . . . . . . . . . . . . . . . . . 30
1.5 Discrete Fourier transform and its geometry . . . . . . . . . 32
1.5.1 Inverse DFT . . . . . . . . . . . . . . . . . . . . . . . 35
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2 Direction Images 41

2.1 2-D direction images on the lattice . . . . . . . . . . . . . . . 41

2.1.1 Superposition of directions . . . . . . . . . . . . . . . 44
2.2 The inverse tensor transform: Case N is prime . . . . . . . . 51
2.2.1 Inverse tensor transform . . . . . . . . . . . . . . . . . 51
2.2.2 Formula of the inverse tensor transform . . . . . . . . 57
2.2.2.1 Code for inverse tensor transform . . . . . . 58
2.3 3-D paired representation . . . . . . . . . . . . . . . . . . . . 60
2.3.1 2D-to-3D paired transform . . . . . . . . . . . . . . . 62
2.3.2 The splitting of the 2-D DFT . . . . . . . . . . . . . . 66

vii
viii IMAGE PROCESSING

2.4 Complete system of 2-D paired functions . . . . . . . . . . . 75

2.4.0.1 Code: System of basic paired functions . . . 80
2.4.1 1-D DFT and paired transform . . . . . . . . . . . . . 81
2.5 Paired transform direction images . . . . . . . . . . . . . . . 83
2.6 L-paired representation of the image . . . . . . . . . . . . . . 87
2.6.1 Principle of superposition: General case . . . . . . . . 90
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3 Image Sampling Along Directions 97

3.1 Image reconstruction: Model I . . . . . . . . . . . . . . . . . 98

3.1.1 Coordinate systems and rays . . . . . . . . . . . . . . 100
3.2 Inverse paired transform . . . . . . . . . . . . . . . . . . . . 101
3.3 Example: Image 4 × 4 . . . . . . . . . . . . . . . . . . . . . . 103
3.3.1 Horizontal and vertical projections . . . . . . . . . . . 103
3.3.2 Diagonal projections . . . . . . . . . . . . . . . . . . . 107
3.3.3 Other projections . . . . . . . . . . . . . . . . . . . . . 109
3.3.3.1 Generator (1, 3) . . . . . . . . . . . . . . . . 109
3.3.3.2 Generator (1, 2) . . . . . . . . . . . . . . . . 111
3.3.3.3 Generator (2, 1) . . . . . . . . . . . . . . . . 115
3.4 Property of the directed multiresolution . . . . . . . . . . . . 120
3.5 Example: Image 8 × 8 . . . . . . . . . . . . . . . . . . . . . . 121
3.5.1 Horizontal projection . . . . . . . . . . . . . . . . . . . 121
3.5.2 Vertical projection . . . . . . . . . . . . . . . . . . . . 124
3.5.3 Diagonal projection . . . . . . . . . . . . . . . . . . . 125
3.5.4 (2, 1)- and (1, 2)-projections . . . . . . . . . . . . . . . 129
3.5.4.1 (2, 1)-projection . . . . . . . . . . . . . . . . 129
3.5.4.2 (1, 2)-projection . . . . . . . . . . . . . . . . 137
3.5.5 (1, 3)-projection . . . . . . . . . . . . . . . . . . . . . . 143
3.5.6 (1, 4)- and (4, 1)-projections . . . . . . . . . . . . . . . 158
3.5.7 (1, 5)-projection . . . . . . . . . . . . . . . . . . . . . . 172
3.5.8 (1, 6)-projection . . . . . . . . . . . . . . . . . . . . . . 189
3.5.9 (6, 1)-projection . . . . . . . . . . . . . . . . . . . . . . 196
3.5.10 (1, 7)-projection . . . . . . . . . . . . . . . . . . . . . . 202
3.6 Summary of results . . . . . . . . . . . . . . . . . . . . . . . 208
3.6.1 Equations of rays . . . . . . . . . . . . . . . . . . . . . 210
3.6.2 Equations for line-integrals . . . . . . . . . . . . . . . 213
3.7 Equations in the coordinate system (X, 1 − Y ) . . . . . . . . 214
3.7.1 Convolution equations . . . . . . . . . . . . . . . . . . 219
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

4 Main Program of Image Reconstruction 227

4.1 The main diagram of the reconstruction . . . . . . . . . . . . 227

4.2 Part 1: Image model . . . . . . . . . . . . . . . . . . . . . . . 229
TABLE OF CONTENTS ix

4.3 The coordinate system and rays . . . . . . . . . . . . . . . . 231

4.4 Part 2: Projection data . . . . . . . . . . . . . . . . . . . . . 232
4.5 Part 3: Transformation of geometry . . . . . . . . . . . . . . 237
4.6 Part 4: Linear transformation of projections . . . . . . . . . 241
4.7 Part 5: Calculation the 2-D paired transform . . . . . . . . . 245
4.7.1 Method of incomplete 1-D DPT . . . . . . . . . . . . 246
4.7.2 Fast 1-D paired transform . . . . . . . . . . . . . . . . 247
4.7.3 Inverse 2-D DPT . . . . . . . . . . . . . . . . . . . . . 250
4.7.4 Preliminary results . . . . . . . . . . . . . . . . . . . . 252
4.8 Fast projection integrals by squares . . . . . . . . . . . . . . 254
4.9 Selection of projections . . . . . . . . . . . . . . . . . . . . . 265
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

5 RECONSTRUCTION FOR PRIME SIZE IMAGE 271

5.1 Image reconstruction: Model II . . . . . . . . . . . . . . . . . 271

5.2 Example with image 7 × 7 . . . . . . . . . . . . . . . . . . . 272
5.2.1 Horizontal projection . . . . . . . . . . . . . . . . . . . 273
5.2.2 Vertical projection . . . . . . . . . . . . . . . . . . . . 274
5.2.3 Diagonal projection . . . . . . . . . . . . . . . . . . . 275
5.2.4 (1, 2)-Projection . . . . . . . . . . . . . . . . . . . . . 279
5.2.5 (1, 3)-projection . . . . . . . . . . . . . . . . . . . . . . 285
5.2.6 (1, 4)-projection . . . . . . . . . . . . . . . . . . . . . . 291
5.2.7 (1, 5)-projection . . . . . . . . . . . . . . . . . . . . . . 299
5.2.8 (1, 6)-projection . . . . . . . . . . . . . . . . . . . . . . 306
5.2.9 Reconstructed image 7 × 7 . . . . . . . . . . . . . . . 311
5.3 General algorithm of image reconstruction . . . . . . . . . . 313
5.4 Program description and image model . . . . . . . . . . . . . 315
5.5 System of equations . . . . . . . . . . . . . . . . . . . . . . . 318
5.6 Solutions of convolution equations . . . . . . . . . . . . . . . 319
5.6.1 Splitting-signal composition . . . . . . . . . . . . . . . 321
5.6.2 Inverse 2-D tensor transform . . . . . . . . . . . . . . 322
5.7 MATLAB R
-based code (N prime) . . . . . . . . . . . . . . 324
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

6 Method of Particles 329

6.1 Point-map of projections . . . . . . . . . . . . . . . . . . . . 329

6.1.1 A-particle and the field . . . . . . . . . . . . . . . . . 332
6.1.2 Representation by field functions . . . . . . . . . . . . 337
6.2 Method of G-rays . . . . . . . . . . . . . . . . . . . . . . . . 343
6.2.1 G-rays for the first set of generators . . . . . . . . . . 343
6.2.2 G-rays for the second set of generators . . . . . . . . . 348
6.2.3 G-rays for the third set of generators . . . . . . . . . . 351
6.2.4 G-rays for the fourth set of generators . . . . . . . . . 354
x IMAGE PROCESSING

6.2.5Map of projections for one square . . . . . . . . . . . . 355

6.2.5.1 Codes for particles . . . . . . . . . . . . . . . 360
6.3 Reconstruction by field transform . . . . . . . . . . . . . . . 365
6.4 Method of circular convolution . . . . . . . . . . . . . . . . . 374
6.4.1 Uniform frames . . . . . . . . . . . . . . . . . . . . . . 379
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380

7 Methods of Averaging Projections 383

7.1 Filtered backprojection . . . . . . . . . . . . . . . . . . . . . 384

7.2 BP and method of splitting-signals . . . . . . . . . . . . . . . 386
7.2.1 Tensor method of summation of projections . . . . . . 390
7.3 Method of summation of line-integrals . . . . . . . . . . . . . 397
7.4 Models with averaging . . . . . . . . . . . . . . . . . . . . . 398
7.4.1 Method of proportion . . . . . . . . . . . . . . . . . . 399
7.4.2 Method with probability model . . . . . . . . . . . . . 402
7.4.3 Reconstruction of the shifted image . . . . . . . . . . 404
7.4.4 Method of minimization of error . . . . . . . . . . . . 407
7.4.5 Corpuscular approach . . . . . . . . . . . . . . . . . . 409
7.5 General case: Probability model . . . . . . . . . . . . . . . . 411
7.5.0.1 Code of the reconstruction . . . . . . . . . . 414
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418

Bibliography 423

Appendix A 427

Appendix B 433

Index 441
Author Bios

Artyom M. Grigoryan received MS degrees in mathematics from Yerevan

State University (YSU), Armenia, USSR, in 1978, in imaging science from
Moscow Institute of Physics and Technology, USSR, in 1980, and in elec-
trical engineering from Texas A&M University, USA, in 1999, and a Ph.D.
degree in mathematics and physics from YUS, in 1990. In 1990-1996, he
was a senior researcher with the Department of Signal and Image Processing
at the Institute for Problems of Informatics and Automation, and Yerevan
State University, National Academy Science of Armenia. In 1996-2000 he was
a Research Engineer with the Department of Electrical Engineering, Texas
A&M University. In December 2000, he joined the Department of Electri-
cal Engineering, University of Texas at San Antonio, where he is currently
an Associate Professor. He is author of the book entitled: Multidimensional
Discrete Unitary Transforms: Representation, Partitioning and Algorithms,
Marcel Dekker, Inc., 2003, and the textbook Brief Notes in Advanced DSP:
Fourier Analysis with MATLAB R
, CRC Press Taylor and Francis Group,
2009. He is the author of two book chapters and many journal papers and
specializes in the theory and application of fast one- and multi-dimensional
Fourier transforms, elliptic Fourier transforms, tensor and paired transforms,
integer unitary heap transforms, design of robust linear and nonlinear filters,
image cryptography, computerized 2-D and 3-D tomography, and processing
of biomedical images.
Merughan M. Grigoryan received the MS degree in physics from Yere-
van State University, Armenia, USSR, in 1979, and worked as a Postdoctoral
Research Associate from 1979-1981 on the dispersion of ultrashort impulses
in the Department of Radio-Physics and Electronics, YSU. From 1982-1995
he worked as a Senior Research Engineer in different science institutes as All-
Union Scientific Associations “Astro,” “Neitron,” Scientific Research Institute
of Non-Ferrous Metals (USSR), on topics which include electronics, signal and
image processing, and acoustic emission. He is currently conducting a private
research on the following topics: theory and application of quantum mechan-
ics in signal processing, differential equations, Fourier analysis, elliptic Fourier
transforms, fast integer unitary transformations, theory and methods of the
fast unitary transforms generated by signals, and methods of encoding in
cryptography. He is the coauthor of the book Brief Notes in Advanced DSP:
Fourier Analysis with MATLAB, CRC Press Taylor and Francis Group, 2009.

xi
This page intentionally left blank
Preface

This book is devoted to one of the most interesting applications of mathema-

tical methods in digital image processing. It is computed tomography (CT)
or computerized X-tomography, wherein the projection data of the recon-
structing image are obtained by means of the roentgen radiation interac-
tion with tissue. The result of the CT is a two-dimensional (2-D) or three-
dimensional (3-D) image, which represents, with some degree of accuracy, the
image through which the rays pass. Our task is to find the solution of the
problem of image reconstruction by a finite number of projections. The fun-
damental principles and main methods of image reconstruction, which include
the Fourier slice theorem, methods of summation and filtered backprojection,
and the method of finite series expansion are well known and can be found in
many books written on CT. The problem of reconstruction of the 2-D func-
tion from its ray-sums, or line-integrals, which was solved by Radon in 1917,
faced its main obstacle in CT. The number of possible projections is finite.
All known reconstructions in CT result in approximations of the image. The
images in CT are not bounded, and Kotelnikov’s theorem of sampling is not
valid for them. We still do not know if the exact solution of the problem ex-
ists, and if it exists, then for what kind of images and models? Therefore, we
will not follow the beaten path, but instead open slightly the door of the box
where the solution of the problem can be found.
This book is a brief collection of some notes and results of our research in digi-
tal image reconstruction. We present the case of the 2-D image and the parallel
projections. The model describing the process of projection data collection is
simple. On both sides of the analyzed object, or image which is assumed to be
immovable, the X-ray set and detectors are disposed and revolved around the
object at different angles. Then, the measurement data of radiation and detec-
tion of X-rays are collected along other directions. This set of measurements,
or projections, is used by specific mathematical methods to reconstruct the
image of the observed object or tissue. Methods of image reconstruction must
be fast and accurate because of the desired high-quality images for diagnostic
purposes. More importantly, we need methods that reconstruct the image on
the discrete Cartesian lattice with a minimal number of projections, in order
to not overradiate the body in the CT.
Many concepts, ideas, and methods described in this book have not been pre-
sented or published anywhere else. This is written as a textbook with many
examples described in detail and programs that are given in short form, to

xiii
xiv IMAGE PROCESSING

demonstrate the presented concepts, their properties, and methods of image

reconstruction. New concepts include the methods of transferring the geo-
metry of rays from the plane to the Cartesian lattice, the point map of pro-
jections, the particle and its field function, the statistical model of averaging,
and others. Our goal is also to give graduate students and other readers solid
material for the presented theory of image reconstruction, to benefit those
interested in continuing this research and obtaining new results in image re-
construction.
The following describes the organization of the book. In Chapter 1 the concept
of the 2-D tensor transform and splitting-signals of the image are described.
The splitting of the 2-D discrete Fourier transform (DFT) by the tensor is
analyzed. Chapter 2 is devoted to the paired representation of the image.
This is the 2-D frequency and 1-D time transform of the image. The transform
is not separable but unitary and leads to the effective calculation of the 2-D
DFT. Both tensor and paired transforms are described by basic functions with
binary coefficients located along the parallel rays; therefore, the image can be
decomposed by direction images and used in the reconstruction. In Chapter 3,
the method of transferring the geometry for calculating the paired transform
of the image is presented. The examples for reconstructing images of sizes 4×4
and 8 × 8 are described in detail. The general algorithm for reconstructing the
image on the Cartesian lattice of size 2r ×2r and MATLAB R
-based programs
are given in Chapter 4. The case of images on a lattice of size N ×N, when N is
prime, is described separately in Chapter 5. The tensor transform is calculated
from the line-integrals and the reconstruction is obtained through the inverse
tensor transform. In Chapter 6, the concept of particles and field functions
and their application in reconstructing the image are described. The method
of backprojection and new methods of summation are given in Chapter 7.
The simple model with statistics is presented and applied for reconstructing
the image on a Cartesian lattice of arbitrary size. Each chapter contains a
list of problems that we suggest readers work on and solve. Difficult problems
are marked with an asterisk; they require computations by hand and with a
computer. In the Appendix, we include two translated papers published in the
1980s in Russian, wherein the concepts of the tensor and paired representation
for images were presented.
We appreciate all who assisted in the preparation of the book. We are grate-
ful to the reviewers for their suggestions and recommendations. Many thanks
to Taylor & Francis/CRC Press, especially Ms. Nora Konopka, for giving us
this great opportunity to prepare the main results of our research for pub-
lication. Finally, we express our gratitude and love of our families for their
understanding, support, and extraordinary patience during the preparation of
this book.

Artyom and Merughan M. Grigoryan

February 22, 2012.
TABLE OF CONTENTS xv

Symbol Description

f(x, y) Image on square [0, 1] ×[0, 1] rp,s,t Wide ray

fn,m Discrete image ˜l = l̃p,sLine xp + my = t
Fp,s 2-D DFT of the image fn,m wl line-integral of the image
Fp 1-D DFT of the signal fn along the l-ray
rect Rectangle function on vl line-sum of the discrete
[−0.5, 0.5] image along the l-ray
XN,N Cartesian lattice N × N
bt Solution of the linear equa-
σ = σJ Irreducible covering of the
tion b = A−1 w
lattice
A Toeplitz matrix
σ 0 = σJ 0 Partition of the lattice
JN,N Set of generators A-ray Arithmetical rays on the lat-
0
JN,N Set of generators tice
Vp,s,t Set of pixels (n, m) of the G-ray Geometrical rays on the
lattice XN,N , which sat- square [0, 1] × [0, 1]
isfy the Diaphanous equa- CP (p, s) Control points of the (p, s)-
tion np + ms = t mod N projection
Tp,s Cyclic group of frequencies U H(K) UH-square of the number K
generated by (p, s) dp,s
n1,n2 Direction image generated
0 by (p, s)
Tp,s Subset (orbit) of frequencies
generated by (p, s) UN,N Set of triplets (p, s, t) of the
fp,s,t Component of the splitting- 2-D paired transform
signal in tensor representa- U (p, s) Set of triplets 2k (p, s, t)
tion A(n, m) Map of projections for point
0
fp,s,t Component of the splitting- (n, m) of the lattice
signal in paired representa- C(n, m) Map of projections for G-
tion particle [n, m] on the square
χN,N Tensor transformation ψn,m Field function of the A-
χ0N,N 2-D paired transformation particle (n, m)
χp,s,t Basis function of the tensor
φn,m Field function of the G-
transform
particle [n, m]
χ0p,s,t Basis paired function of the
Ψ Base matrix of particles
2-D paired transform
χ0p,s,t;L L-paired function Ψp,s 2-D DFT of the base matrix
χ0p,t Basis paired function of the R Matrix of the field functions
(k)
1-D paired transform pp,s,t Probability density function
χ0N 1-D paired transformation ta(n, m) Number t(n, m; n0 , m0 ) of
AN 1-D N -point DHdT A-rays passing trough the
FN 1-D N -point DFT points (n, m) and (n0 , m0 )
FN,N 2-D N × N -point DFT at the same time
l = lp,s Line np + ms = t c(n, m) Coefficient of intersection of
∆lp,s Length of intersection of the rays with image element
ray with the (n, m)-IE (n, m)
This page intentionally left blank
1
Discrete 2-D Fourier Transform

The theory of the continuous-time and discrete-time Fourier transformations

has been well developed. The discrete Fourier transform has become a powerful
technique in signal processing, and in particular in image processing. Effec-
tive methods, or fast algorithms, [1]-[6], of the discrete Fourier transforms
(DFT) are used for solving many problems in image processing in the fre-
quency domain, such as image filtration, restoration, enhancement, compres-
sion, and image reconstruction by projections [7]-[10]. Other unitary transfor-
mations are also used in signal and image processing. Considerable interest
is given to many applications of the discrete Hartley transformation (DHT),
since it relates closely to the DFT and has been created as an alternative
form of the complex DFT to eliminate the necessity of using complex ope-
rations [11]-[15]. The discrete cosine transformation (DCT) is used in speech
and image processing, especially in image compression and transform coding
in telecommunication [16]-[22]. Another unitary transformation is the discrete
Hadamard transformation (DHdT), whose basic functions take value ±1 at
each point [23]-[48]. The Hadamard transform has found useful applications
in signal and image processing, communication systems, image coding, image
enhancement, pattern recognition, and general two-dimensional filtering [2, 7].
The application of the two-dimensional transform involves the calculation
of the transform, manipulation with transform coefficients, then calculation
of the inverse transform. For images of large sizes, this process requires a
great number of operations of multiplication and addition when performing
the transforms. Different methods of effective calculation of two-dimensional
unitary transforms have been developed to reduce the number of operations
needed. In most cases, the calculation of the 2-D transform is reduced by parti-
tioning the entire image by 1-D or 2-D blocks and calculating the transforms of
these blocks. In the traditional “row-column” algorithm, 1-D transforms over
all rows and then columns are calculated. For images of large sizes, the trans-
position slows down the process of calculating the transform. Other methods
of fast calculation of the transform have been developed to avoid the trans-
position. We mention here the idea of generalization of the 2-D “butterfly”
operation from the 1-D Cooley-Tukey algorithm [1] to the 4-D operation, when
dividing the transforms by four parts of size (N/2 ×N/2) each. This algorithm
reduces the number of multiplications by almost 25%, when compared with
the row-column method. The number of multiplications can be reduced ap-

1
2 CHAPTER 1: 2-D DFT

proximately twice in the method of the polynomial transformations developed

by Nussbaumer [38].
Our subject of study is the partitioning of two-dimensional discrete Fourier
transform based on the concepts of the tensor and paired representations
of images. In these representations, the images are described by sets of 1-D
signals which carry the spectral information of images in different subsets of
frequency-points. The processing of two-dimensional images is thus reduced
to processing 1-D signals, called image-signals or splitting-signals, since they
represent the images and split the transforms of these images. The splitting-
signals are described here for the two-dimensional discrete Fourier transform,
but they can be defined to other transforms, such as the Hartley, Hadamard,
and cosine transforms [46, 48, 51, 55].

1.1 Separable 2-D transforms

Many of the two-dimensional transformations are separable, meaning that
these transforms over 2-D images can be performed by calculating one-dimen-
sional transforms consequently along both dimensions of the signal. For in-
stance, for a separable two-dimensional (2-D) transformation T, the transform
of a 2-D image f = {fn,m } of size (N × N ), N > 1, can be obtained by first
calculating the 1-D transforms over all columns of the image, and then calcu-
lating the 1-D transforms by the rows of the obtained 2-D data, as shown in
Figure 1.1.

1−D DTs 1−D DTs = 2−D DT

FIGURE 1.1
Block-diagram of calculation of the 2-D discrete transform (DT) (separable).

In matrix form, the transform of the image f can be written as

[2-D T ][f] = [1-D T ][f][1-D T ]t

where t denotes the matrix operation of transposition, and square brackets [·]
are used to denote the matrices of the transformations T and image f.
As an example, we consider the 2-D DFT of the image fn,m , which is
defined by

X N−1
N−1 X
Fp,s = (FN,N ◦ f)p,s = fn,m W np+ms , p, s = 0 : (N − 1), (1.1)
n=0 m=0
Image Processing 3

where W = WN = exp(−2πj/N ) is the kernel of the transformation, and

j 2 = −1. The designation p = 0 : (N − 1) denotes p as an integer that
runs from 0 to (N − 1). (p, s) is the frequency-point. The kernel is separable,
W np+ms = W npW ms , and the transform can thus be written as
N−1
" N−1
#
X X
ms
Fp,s = Fn (s) = fn,m W W np , p, s = 0 : (N − 1), (1.2)
n=0 m=0

where Fn (s) is the value of the 1-D DFT of row number n at point s. To
calculate the 2-D DFT, 2N 1-D DFTs are used in the row-column algorithm.
This algorithm is simple, but requires many operations of multiplication and
addition. All twiddle coefficients, W t , t = 0 : (N − 1), lie on N equidistant
points of the unit circle, and many of them are irrational numbers.
We now consider the transformation whose kernel lies only on two points
±1 on the unit circle. The 2-D separable discrete Hadamard transform of order
N × N, where N = 2r , r > 1, is defined as

X N−1
N−1 X
Ap,s = (AN,N ◦ f)p,s = fn,m a(p; n)a(s; m)
n=0 m=0
N−1
" # (1.3)
X N−1 X
= fn,m a(s; m) a(p; n).
n=0 m=0

The kernel of the transformation is defined by the binary function

a(p; n) = (−1)n0 p0 +n1 p1 +...+nr pr (1.4)

where (n0 , n1 , ..., nr) and (p0 , p1 , ..., pr) are the binary representation of num-
bers n and p, respectively.
As an example, Figure 1.2 shows an image (2048 × 2048) in part a, along
with the 2-D discrete Fourier and Hadamard transforms of the image in b and
c, respectively. The realization of the Hadamard transformation requires only
operations of addition (and subtraction). From the computational point of
view, the 1-D Hadamard transform is faster than the complex Fourier trans-
form. These two different transforms can share the same fast algorithm. For
instance, the fast Fourier transform by the paired transforms can also be used
for the fast Hadamard transform, when considering all twiddle coefficients W t
equal 1 [45].
We now present the tensor approach and its improvement, for dividing
the calculation of the 2-D DFT into the minimal number of short 1-D trans-
forms. The approach is universal because it can be implemented to calculate
other discrete unitary transforms, such as the Hadamard, cosine, and Hartley
transforms [39, 45, 48], and transforms of high dimensions.
4 CHAPTER 1: 2-D DFT

Girl with bear 2−D DFT 2−D DHT

(a) (b) (c)

FIGURE 1.2
(a) Original image of size 2048 × 2048, (b) 2-D DFT (in absolute mode and
shifted to the center), and (c) 2-D DHdT of the image.

1.2 Vector forms of representation

When processing the two-dimensional image fn,m , in the frequency domain by
a specific unitary transformation, let’s say the 2-D discrete Fourier transfor-
mation, the image can be represented in a form that splits the structure of the
transformation in a way that yields an effective method of calculation of the
transform with the following image processing. Such forms are not necessarily
of the matrix form, but other figures. The presented work here does not rely
on traditional methods of processing 2-D transforms and signals, but more
effective methods which are based on the discovery that can be formulated
briefly as follows. Two-dimensional spectra are split by appointed trajectories
(such as orbits) and the movement of a spectral point along each such trajec-
tory is of great interest in the process of formation of the spectra, as well as
in processing the spectra. For many orders of the transform, trajectories do
not intersect, and it is possible to extract the spectral information from such
trajectories or change and put desired information into trajectories.
We discuss the theory of fast 2-D discrete Fourier transformation, which
is based on the concept of partitions that reveal the two-dimensional and
multi-dimensional transformations. The use of new forms of image and trans-
form representation simplifies the calculation of 2-D transforms and leads to
effective solutions of different problems in image processing, such as image
enhancement, computerized tomography, image filtration, compression, and
encryption. These new forms are the tensor and paired forms of representa-
tion. Their main task is to represent uniquely the image in the form of a set
of 1-D signals which can be processed separately and then transferred back to
the image, as shown in the diagram of Figure 1.3 (with or without block 2).
The calculation of the 2-D DFT is reduced to calculation of 1-D DFTs and
processing of the 2-D image to processing all or a few 1-D signals. The pro-
Image Processing 5
−1
fn,m - signal 1 - FN,N gn,m
1D DSP
χ -
Image - -
- -
- - Image
- 2-D IDFT
- signal L - 2
original 1D DSP 1 Fp,s processed

FIGURE 1.3
Block diagram of image processing by 1-D signals. (DSP = digital signal
processing)

cessed image can be calculated by inverting the 2-D DFT, or by calculating

directly from the processed 1-D signals. The mathematical structure of the
2-D DFT and other unitary transforms possess such representations [46, 48].
As an example, we consider the image enhancement by one such signal,
instead of calculating the 2-D DFT of the image and manipulating all of its
coefficients. Figure 1.4 shows the original image of size 256 × 256 in part a,
along with a 1-D signal derived from the image in b, the 1-D DFT of the
signal (in absolute scale and shifted to the center) in c, the coefficients to
be multiplied pointwise by the 1-D DFT in d, a new 1-D signal in e, 256
frequency-points at which the spectral information of the new signal will be
renewed in the 2-D DFT of the image in f, and the inverse 2-D DFT, or the
enhanced image in g. The enhanced image can be obtained by the inverse
2-D DFT, as well as directly from the new signal (in e) [50, 49, 55]. Thus
the problem of the 2-D image enhancement is reduced to processing one 1-D
signal (or a few such signals), by passing the calculation of the 2-D DFT of the
original image, as well as the inverse 2-D DFT (2-D IDFT) for the enhanced
image. We now describe methods of deriving such 1-D signals, which lead to
effective calculating and processing of the 2-D DFT.

1.3 Partitioning of 2-D transforms

Let a sequence g = {g0 , g1 , . . . , gN−1 } of length N > 1 be linearly and uniquely
expressed by a sequence f = {f0 , f1 , . . . , fN−1 },
N−1
X
gp = ϕp (n)fn , p = 0 : (N − 1). (1.5)
n=0

The transformation of f into g, by using this equation is called a linear trans-

formation, which we denote by A. The set of coefficients ap,n = ϕp (n) forms
the square (N × N ) matrix A =|| ap,n || of the transformation. The linear
6 CHAPTER 1: 2-D DFT

176 200

175 180 FT(0)=4388.5

174 160

140
173
120
172
100
171
80
170
60
169 40
168 20

167 0
0 50 100 150 200 250 0 50 100 150 200 250

(a) original (b) 1-D signal (c) 1-D DFT

1 176
original signal
175
0.95
174
0.9
173

0.85 172

0.8 171
new signal
170
0.75
169
0.7
168

0.65 167
0 50 100 150 200 250 0 50 100 150 200 250

(d) factors (e) new signal

250

200

150

100

50

0
0 50 100 150 200 250

(f) frequency-points (g) enhanced

FIGURE 1.4 (See color insert)

Fast transform-based method of image enhancement. (a) The original image,
(b) the image-signal, (c) the magnitude spectrum of the signal, (d) factors, (e)
the processed signal, (f) marked locations of 256 frequency-points, at which
the 2-D DFT was changed, and (g) the image enhanced by one signal.

transformation can be written in matrix representation [g] = A[f], where [g]

and [f] denote the vector-columns of sequences g and f, respectively. Every
linear 1-D transformation uniquely determines a 2-D matrix A, and vice versa,
every 2-D matrix A determines the linear 1-D transformation. Similarly, each
four-dimensional matrix determines a certain linear transformation A of two-
Image Processing 7

dimensional sequences, or images

A : f = {fn1,n2 } → g = {gp1 ,p2 } (1.6)

where nk , pk = 0 : (Nk −1) and Nk > 1, k = 1 : 2. The numbers N1 and N2 are

called orders of the two-dimensional transformation A = AN1 ,N2 . g = A ◦ f is
called the two-dimensional transform of f.
The transform of the image f is described by the following relation:
1 −1 N
NX X2 −1

gp1 ,p2 = ϕp1 ,p2 (n1 , n2 )fn1 ,n2 (1.7)

n1 =0 n2=0

where ap1 ,p2 ,n1 ,n2 = ϕp1 ,p2 (n1 , n2 ) are coefficients of the matrix of the trans-
formation A. Elements (p1 , p2 ) ∈ X are referred to as frequency-points. We
assume that the image f and the transform A ◦ f are defined on the same
two-dimensional rectangular integer lattice of size N1 × N2 ,

X = XN1 ,N2 = {(n1 , n2 ); nk = 0 : (Nk − 1), k = 1 : 2}

(1.8)
= {(p1 , p2 ); pk = 0 : (Nk − 1), k = 1 : 2}.

This set X is called the fundamental period of the transformation A.

The matrix of the N × N -point discrete Fourier transformation FN,N is
the four-dimensional matrix
 
FN,N =|| ap1 ,p2 ,n1 ,n2 ||=|| W n1p1 +n2 p2 || .

The Fourier transformation is unitary. A discrete transformation A is called

unitary if the matrix of the transformation is unitary, i.e., AA∗ = I, where I
is the identity matrix, and A∗ =|| ān1 ,n2,p1 ,p2 || is the complex conjugate to
A, where the sign − denotes the transition to the complex conjugate value.
For a fixed frequency-point (p1 , p2 ), the function ϕp1 ,p2 (n1 , n2 ) is the basis
function of the two-dimensional transformation A, and the collection of such
functions {ϕ} = {ϕp1 ,p2 (n1 , n2 )} is the basis, or kernel of A. The unitary
property of the transformation can be expressed as
X
ϕp1 ,p2 (n1 , n2 )ϕ̄s1 ,s2 (n1 , n2 ) = δ(p1 , s1 )δ(p2 , s2 )
(n1 ,n2 )∈X (1.9)
(p1 , p2 ), (s1 , s2 ) ∈ X,

where δ denotes the delta symbol, δ(p, s) = 1, if p = s, and δ(p, s) = 0,

otherwise. If the collection of functions {ϕ} satisfies this condition, then {ϕ}
is said to be a complete and orthonormal set of functions, and it is orthogonal
if there is a factor different from 1 at δ(p1 , s1 )δ(p2 , s2 ) in (1.9). Thus, for
the unitary transformation A, this collection of functions is a complete and
orthonormal set of basis functions.
8 CHAPTER 1: 2-D DFT

Example 1.1 (1-D Fourier transformation) Let f be a 1-D sequence

f = {f0 , f1 , ..., fN−1}. The N -point Fourier transform of the sequence f is
defined by
N−1
1 X np
Fp = (FN ◦ f)p = √ W fn , p = 0 : (N − 1). (1.10)
N n=0
√ √
The basis functions ϕp (n) = W np are the pairs (1/ N cos(ωp n), 1/ N sin(ωp n))
of discrete-time cosine and sine waves with frequencies ωp = (2π/N )p. The
waves are orthonormal, since
N−1 N−1
X 1 X n(p−s)
ϕp (n)ϕ̄s (n)= W
n=0
N n=0
N−1   
1 X 2π(p − s) 2π(p − s) 1,p = s
= cos n − j sin n =
N n=0 N N 0,p 6= s

and the transformation is unitary. The matrices of the transformation and its
conjugate are symmetric,
  1 j2π 1 j2π
FN = √ || e− N np ||n,p=0:(N−1), ∗
[FN ] = √ || e N np ||n,p=0:(N−1),
N N
∗
and [FN ][FN ] = I. The conjugate matrix is thus the matrix of the inverse 1-D
DFT.

The orthogonality of the basis functions of the 2-D DFT follows directly
from the orthogonality of the basis functions of the 1-D DFT, because of
separability ϕp,s (n, m) = ϕp (n)ϕs (m). Indeed, the following calculations hold:

X N−1
N−1 X X N−1
N−1 X
ϕp1 ,s1 (n, m)ϕ̄p2 ,s2 (n, m) = ϕp1 (n)ϕs1 (m)ϕ̄p2 (n)ϕ̄s2 (m)
n=0 m=0 n=0 m=0
N−1
X N−1
X
= ϕp1 (n)ϕ̄p2 (n) ϕs1 (m)ϕ̄s2 (m) = δ(p1 , p2 )δ(s1 , s2 ).
n=0 m=0

1.3.1 Tensor representation

In this section, we describe a concept of covering that reveals the mathematical
structure of many multidimensional transforms [39]-[48]. This concept will be
applied for the 2-D discrete Fourier transformation. The covering is considered
to be composed of cyclic groups of frequency-points. Such covering leads to
the tensor, or vector representation of images.
Suppose σ = (T ) is an irreducible covering of an 2-D lattice X = XN1 ,N2 .
It means the set-theoretic union of all subsets T coincides with X and any
smaller family of subsets of T from σ does not cover X. We use card to denote
Image Processing 9

the cardinality of a set. If a discrete 2-D unitary transformation with the

fundamental period X can be split into a set of card σ one-dimensional unitary
transformations A, then we say that the considered 2-D transformation is
revealed by the covering σ, or, the covering σ reveals the transformation. Let
f be an N1 × N2 sequence, or image.
Definition 1.1 2-D N1 × N2 -point discrete transformation P is said to be
revealed by a covering σ of X if, for each set T ∈ σ, there exists a 1-D ortho-
gonal transformation A = A(T ) and a sequence fT such that the restriction
of the transform P ◦ f on the set of frequency-points T equals the transform
A ◦ fT , i.e.,

P ◦ f | = A ◦ fT . (1.11)
T

Each 1-D transformation A is determined by the corresponding subset T,

not f. The set of the 1-D transforms {A(T ); T ∈ σ} is called the splitting of
the 2-D transformation P by the covering σ and denoted by R(P; σ). The set
of 1-D sequences {fT ; T ∈ σ} is the σ-representation of f with respect to the
transformation P.
It should be noted that the covering σ not only splits the 2-D transfor-
mation by 1-D transformations, but also determines the corresponding repre-
sentation of the 2-D image f as the set of 1-D sequences fT . In other words,
two representations are defined, one for the given image and another for its
transform,
f → {fT ; T ∈ σ}, and P ◦ f → {A ◦ fT ; T ∈ σ}. (1.12)

1.3.2 Covering with cyclic groups

We consider an irreducible covering σ composed only from additive cyclic
groups  
σ = σJ = Tp1 ,p2 (1.13)
(p1 ,p2 )∈J

with generators (p1 , p2 ) from a certain subset J ⊂ X = XN1 ,N2 . The cyclic
group T with a generator (p1 , p2) is defined as a set of frequency-points which
are integer multiples to the generator,
T = Tp1 ,p2 = {(kp1 , kp2 ); k = 0 : (card T − 1)} (1.14)
We use the short notation kpi = (kpi ) mod Ni for i = 1 : 2.
Example 1.2 ((Lattice 3 × 3)) The covering of the lattice 3 × 3 can be de-
fined by σ = (T1,0 , T1,1 , T1,2 , T0,1) as shown
         
• • • • ◦ ◦ • ◦ ◦ • ◦ ◦ • • •
 • • •  =  • ◦ ◦ ,  ◦ • ◦ ,  ◦ ◦ • ,  ◦ ◦ ◦  .
• • • • ◦ ◦ ◦ ◦ • ◦ • ◦ ◦ ◦ ◦
X3,3 T1,0 T1,1 T1,2 T0,1
10 CHAPTER 1: 2-D DFT

When the generator is (1, 1), all elements of the group T1,1 are located along
the diagonal of the lattice. All elements of the group T1,0 are located on the
first row of the lattice, and the elements of group T0,1 are located on the first
column of the lattice.
All other groups Tp,s coincide with the groups of the covering σ. For in-
stance, T0,2 = T0,3 = T0,1 , T2,2 = T3,3 = T1,1 , and T2,1 = T1,2 .

Example 1.3 (Lattice 4 × 4) The following covering of the lattice 4 × 4

holds:
2 3 2 3 2 3 2 3 2 3 2 3
• ◦ ◦ ◦ • ◦ ◦ ◦ • ◦ ◦ ◦ • ◦ ◦ ◦ • • • • • ◦ • ◦
6• ◦ ◦ ◦7 6◦ • ◦ ◦7 6◦ ◦ • ◦7 6◦ ◦ ◦ •7 6◦ ◦ ◦ ◦7 6◦ ◦ ◦ ◦7
6 7, 6 7, 6 7, 6 7, 6 7, 6 7.
4• ◦ ◦ ◦5 4◦ ◦ • ◦5 4• ◦ ◦ ◦5 4◦ ◦ • ◦5 4◦ ◦ ◦ ◦5 4◦ • ◦ •5
• ◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦ • ◦ ◦ • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
T1,0 T1,1 T1,2 T1,3 T0,1 T2,1

Example 1.4 (Lattice 5 × 5) Let X be the lattice 5×5, that corresponds to

the N1 = N2 = 5 case. The group Tp1 ,p2 with a generator (p1 , p2 ) ∈ X = X5,5
is
Tp1 ,p2 = {(0, 0), (p1, p2 ), (2p1 , 2p2 ), (3p1 , 3p2 ), (4p1 , 4p2)}.
There are six groups T that compose an irreducible covering σ = σJ of X,
namely
T0,1 = {(0, 0), (0, 1), (0, 2), (0, 3), (0, 4)}
T1,1 = {(0, 0), (1, 1), (2, 2), (3, 3), (4, 4)}
T2,1 = {(0, 0), (2, 1), (4, 2), (1, 3), (3, 4)}
(1.15)
T3,1 = {(0, 0), (3, 1), (1, 2), (4, 3), (2, 4)}
T4,1 = {(0, 0), (4, 1), (3, 2), (2, 3), (1, 4)}
T1,0 = {(0, 0), (1, 0), (2, 0), (3, 0), (4, 0)}
and the set of generators

J = J5,5 = {(0, 1), (1, 1), (2, 1), (3, 1), (4, 1), (1, 0)}. (1.16)

Figure 1.5 shows the location of all frequency-points of these six groups. These
groups intersect only at point (0, 0). If the covering composed of these groups
reveals the 5 × 5-point 2-D transformation, then this transformation can be
split by six five-point 1-D transformations. This transformation is the Fourier
transformation. To calculate, for instance, the 5 × 5-point 2-D DFT, only
six 1-D five-point transforms are required, instead of ten transforms in the
“row-column” method.

In the general N × N case, the elements of the group Tp1 ,p2 lie on parallel
lines at angle θ = tan−1 (p2 /p1 ) to the horizontal axis. The number l of such
lines is determined as follows. If p1 = 0 or p2 = 0, then l = 1. For other
cases, let k1 and k2 be the smallest integers satisfying the relations k1 p1 =
k2 p2 = N − 1. Then, l = k1 /k2 with k1 ≥ k2 , and l = k2 /k1 with k1 < k2 .
Image Processing 11

T0,1 T1,1 T2,1

4 4 4
3 3 3
2 2 2
1 1 1
0 0 0

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4

T3,1 T4,1 T1,0

4 4 4
3 3 3
2 2 2
1 1 1
0 0 0

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4

FIGURE 1.5
Arrangement of frequency-points of groups Tp1 ,p2 covering the lattice 5 × 5.

For instance, when N = 5 and (p1 , p2 ) = (2, 1), we obtain 2p1 = 4p2 = 4 and
l = 4/2 = 2. The frequency-points of the group T2,1 lie on two parallel lines at
angle θ = tan−1 (1/2) = 26.5651◦ to the horizontal axis (see Figure 1.5). The
points of this group can also be considered as lying on three parallel lines at
angle θ1 = θ − 90◦ = −63.4349◦ to the horizontal axis.
It is important to note, that if we splice the opposite sides of the lattice
bounds, then the lattice will be represented as a net traced on the surface
of a three-dimensional torus and the mentioned l lines will compose a closed
spiral on the torus, which will pass through those points on the net, which
correspond to the points (0, 0) and (p1 , p2 ). All elements of the cyclic group
will be the points of intersection of the spiral with the net. As an example,

(0,0)

T1,2

T1,1

FIGURE 1.6 (See color insert)

Torus of the lattice 20 × 20 with two spirals corresponding to the groups T1,1
and T1,2 .
12 CHAPTER 1: 2-D DFT

Figure 1.6 shows the points of the lattice X20,20 on the torus and two spirals
with frequency-points of the groups T1,1 and T1,2 which intersect at the knot
(0, 0).
The irreducible covering σ of the domain X composed from groups (1.14)
is unique. To illustrate this property, we consider the square lattice X5,5 . The
irreducible covering σ of X5,5 is the family of six groups given in (1.15),
σ = (T0,1 , T1,1 , T2,1 , T3,1, T4,1 , T1,0 ).
The cyclic group Tp1 ,p2 with any other generatrix (p1 , p2 ) 6= (0, 0), different
from generators (0, 1), (1, 1), (2, 1), (3, 1), (4, 1) and (1, 0), coincides with one
of the groups of the covering σ. For instance,
T1,2 = {(0, 0), (1, 2), (2, 4), (3, 1), (4, 3)} = T3,1 ,
T3,2 = {(0, 0), (3, 2), (1, 4), (4, 1), (2, 3)} = T4,1 ,
T2,2 = {(0, 0), (2, 2), (4, 4), (1, 1), (3, 3)} = T1,1 .

1.4 Tensor representation of the 2-D DFT

In this section, we discuss in detail the construction and properties of the
tensor representation of the image with respect to the two-dimensional discrete
Fourier transformation.
Let f = {fn1 ,n2 } be the image of size N1 × N2 , and let N0 =
g.c.d.(N1 , N2 ) > 1, that is, N1 = N0 N10 , N2 = N0 N20 . Let σ be the irre-
ducible covering of the rectangular lattice X = XN1 ,N2 , as defined in (1.13).
We denote by F = FN1 ,N2 the N1 ×N2 -point 2-D discrete Fourier transforma-
tion. The 2-D DFT of the image f, accurate to the normalizing factor 1/N1 N2 ,
is defined by the following relation:
1 −1 N
NX X2 −1

Fp1 ,p2 = (F ◦ f)p1 ,p2 = fn1 ,n2 WNn11p1 WNn22 p2 , (p1 , p2) ∈ X, (1.17)
n1 =0 n2 =0

where WNk = exp(−2πj/Nk ), k = 1, 2.

For an arbitrary frequency-point (p1 , p2 ), we determine in the period X
the sets of points
Vp1 ,p2 ,t = {(n1 , n2 ); N20 n1 p1 + N10 n2 p2 = t mod N } , t = 0 : (N −1), (1.18)
where N = N1 N2 /N0 . On these sets of points, we consider the sums of the
sequence f, i.e., the following N quantities:
X
fp1 ,p2 ,t = {fn1 ,n2 ; (n2 , n2 ) ∈ Vp1 ,p2 ,t }, t = 0 : (N − 1). (1.19)
For the spectral component Fp1 ,p2 , the following calculations hold:
1 −1 N
NX X2 −1 N−1
X
N 0 n1 p1 +N10 n2 p2
Fp1 ,p2 = fn1 ,n2 WN 2 = fp1 ,p2 ,t W t , (1.20)
n1 =0 n2 =0 t=0
Image Processing 13
2π
where W = WN = e−j N . The general equation is also valid,
N−1
X
Fkp1 ,kp2 = fp1 ,p2 ,t W kt , k = 0 : (N − 1). (1.21)
t=0

In other words, N components Fkp1,kp2 of the 2-D DFT of the image f can
be represented by the 1-D sequence of length N,

fTp1 ,p2 = {fp1 ,p2 ,0 , fp1 ,p2 ,1 , . . . , fp1 ,p2 ,N−1 } . (1.22)

The sequence fTp1 ,p2 determines the spectral information of the image at
frequency-points of the set Tp1 ,p2 . Such a sequence is called the splitting-signal,
or the image-signal. The components of the splitting-signal are numbered by
the set of three, or triplets (p1 , p2 , t), where two components represent the
frequency (p1 , p2 ) and t is referred to as the time. Thus the splitting-signal
is the (2-D frequency)-(1-D time) representation of the 2-D image f, which
defines the full 2-D DFT of f at the frequency-points of the set Tp1 ,p2 .
The set of splitting-signals

{fTp1 ,p2 ; Tp1 ,p2 ∈ σ} (1.23)

is called the tensor representation, or tensor transform of the image.

1.4.0.1 Code: Splitting-signal calculation

Below is the function with script ft pst.m that accomplishes the calculation
of the splitting-signal generated by (p, s). Here A stands for the image fn,m of
size N × N, and B for the splitting-signal.

% call: ft_pst / A.M. Grigoryan, 1996

function B=ft_pst(A,p,s)
[M N]=size(A);
B=zeros(1,N);
ks=0;
for m=1:M
t=ks;
for n=1:N
if t>=N t=t-N; end
t1=t+1;
B(t1)=B(t1)+A(m,n); t=t+p;
end
ks=ks+s; if ks>=M ks=ks-M; end
end

1.4.1 Tensor algorithm of the 2-D DFT

The algorithm of calculation of the 2-D discrete Fourier transform FN1 ,N2 ◦ f
of the image f = fn1 ,n2 is performed by the following two steps.
14 CHAPTER 1: 2-D DFT

Step 1. Calculate the 1-D splitting-signals fT of the tensor representation

of the image, i.e., calculate the tensor transform of the image,

χσ : f → {fT ; T ∈ σ}. (1.24)

Step 2. Calculate the 1-D DFTs of the obtained splitting-signals, FN (T )◦

fT , T ∈ σ.
Step 3. Allocate the 1-D DFTs in the 2-D data by cyclic groups T ∈ σ,

Fkp1,kp2 = (FN ◦ fT )k , k = 0 : (N − 1),

where the number N = N1 N2 /g.c.d.(N1, N2 ).

The number of 1-D transforms required for calculating the 2-D DFT, co-
incides with the cardinality, card σ, of the covering σ, or the cardinality of the
set J of generators of these groups,
 
σ = σJ = Tp1 ,p2 . (1.25)
(p1 ,p2 )∈J

We here separately consider the set of generators for the most interesting
cases, when N1 = N2 = N, and N is a general prime, the product of two
prime numbers, and then we describe the case when N is a power of prime.

1.4.2 N is prime
Let N > 1 be the prime. The irreducible covering σJ of the set XN,N has the
cardinality N + 1, i.e.,
card σJ = N + 1. (1.26)
Thus, the least number of cyclic groups Tp,s that together cover the period
X equals (N + 1). Indeed, the irreducible covering σJ is determined by the
following set of generators:

J = JN,N = {(0, 1), (1, 1), (2, 1), ..., (N − 1, 1), (1, 0)}. (1.27)

Other sets of (N + 1) generators can also be taken, for instance,

J = {(1, 0), (1, 1), (1, 2), ..., (1, N − 1), (0, 1)}.

Therefore, to calculate the N × N -point 2-D DFT, it is sufficient to fulfill

(N + 1) one-dimensional N -point DFTs.

Example 1.5 Consider the 3 ×3-point DFT of the following two-dimensional

sequence, or image:  
1 2 1
f = {fn1 ,n2 } =  2 4 2 .
1 2 1
Image Processing 15

The square lattice X3,3 = {(n1 , n2 ); n1 , n2 = 0, 1, 2} is covered by the total-

ity of sets σ = (T0,1 , T1,1, T2,1 , T1,0 ). The tensor representation of the image
consists of four splitting-signals,

χσ : {fn1 ,n2 } → {{f0,1,t }, {f1,1,t}, {f2,1,t}, {f1,0,t}} , t = 0, 1, 2. (1.28)

Step 1: We denote by f the vector-column composed from rows of the se-

quence fn1 ,n2 , i.e., f = (1, 2, 1, 2, 4, 2, 1, 2, 1)0. The first splitting-signal {f0,1,t}
is calculated by
   
f0,1,0 1 0 0 1 0 0 1 00
 f0,1,1  =  0 1 0 0 1 0 0 10  f = [4, 8, 4], (1.29)
f0,1,2 0 0 1 0 0 1 0 01
and the next three splitting-signals are calculated as follows:
   
f1,1,0 1 0 0 0 0 1 0 10
 f1,1,1  =  0 1 0 1 0 0 0 01  f = [5, 5, 6], (1.30)
f1,1,2 0 0 1 0 1 0 1 00
   
f2,1,0 1 0 0 0 1 0 0 01
 f2,1,1  =  0 1 0 0 0 1 1 00  f = [6, 5, 5], (1.31)
f2,1,2 0 0 1 1 0 0 0 10
   
f1,0,0 1 1 1 0 0 0 0 00
 f1,0,1  =  0 0 0 1 1 1 0 00  f = [4, 8, 4]. (1.32)
f1,0,2 0 0 0 0 0 0 1 11
Step 2: The three-point DFT of the first signal {f0,1,0 , f0,1,1, f0,1,2} =
[4, 8, 4] is calculated by
      
A0 1 1 1 4 16
A1  = 1 W W 2  8 = −2.0 − j3.4641 ,
A2 1W 2 W 4 −2.0 + j3.4641

where W = exp(−j2π/3). The sum of each splitting-signal equals 16, and

there is no need to calculate the three-point DFTs of other splitting-signals at
zero. Therefore, we can perform three incomplete three-point DFTs of these
splitting-signals,
 
    5  
B1 1 W W2   −0.5 + j0.8660
= 5 = ,
B2 1W 2 W −0.5 − j0.8660
6
 
   2
 6  
C1 1W W   1
= 5 = ,
C2 1W 2 W 1
5
 
    4  
D1 1 W W2   −2.0 − j3.4641
= 8 = .
D2 1W 2 W −2.0 + j3.4641
4
16 CHAPTER 1: 2-D DFT
10 10 10 10
f0,1,t f1,1,t f2,1,t f1,0,t

5 5 5 5

0 0 0 0
0 1 2 0 1 2 0 1 2 0 1 2

20 20 20 20

15
|Ak| 15
|Bk| 15
|Ck| 15
|Dk|

10 10 10 10

5 5 5 5

0 0 0 0
0 1 2 0 1 2 0 1 2 0 1 2

T0,1 T1,1 T2,1 T1,0

3 3 3 3

2 2 2 2

1 1 1 1

0 0 0 0

−1 −1 −1 −1
−1 0 1 2 3 −1 0 1 2 3 −1 0 1 2 3 −1 0 1 2 3

FIGURE 1.7
Four splitting-signals (the first row), the 3-point DFTs (in absolute scale)
of the splitting-signals (the second rows), and the location of the frequency-
points of the cyclic groups T ∈ σ (the third row).

The splitting-signals and the magnitude spectrums of these signals are

shown in the first and second rows of Figure 1.7, respectively.
Step 3: The location of frequency-points of four cyclic groups T ∈ σ on
the lattice 3 × 3, where the 1-D DFTs of the splitting-signals are placed, is
shown in the last row of the figure. As a result, we obtain the following matrix
expression for the 3 × 3-point DFT of the given sequence f:
     
F0,0 F0,1 F0,2 A0 A1 A2 16 −2.0 − j3.4641 −2 + j3.4641
 F1,0 F1,1 F1,2  =  D1 B1 C2  =  −2 − j3.4641−0.5 + j0.8660 1 .
F2,0 F2,1 F2,2 D2 C1 B2 −2 + j3.4641 1 −0.5 − j0.8660

If we unite four binary matrices 9 × 3 in (1.29)-(1.32), we obtain the follo-

wing matrix 12 × 9 of the tensor transformation in (1.28):
Image Processing 17

 
1 0 0 1 0 0 1 0 0
0 1 0 0 1 0 0 1 0
 
0 0 1 0 0 1 0 0 1
 
1 0 0 0 0 1 0 1 0
 
0 1 0 1 0 0 0 0 1
 
0 0 1 0 1 0 1 0 0
[χσ ] = 
1
. (1.33)
 0 0 0 1 0 0 0 1
0 1 0 0 0 1 1 0 0
 
0 0 1 1 0 0 0 1 0
 
1 1 1 0 0 0 0 0 0
 
0 0 0 1 1 1 0 0 0
0 0 0 0 0 0 1 1 1
In matrix form, the described tensor algorithm of the 3 × 3-point DFT can
be written as
   
F0,0 11 1
 F0,1   1 W W 2 
   
 F0,2   1 W 2 W 
   2

 F1,1   1W W 
   2

 F2,2  =  1W W [χσ ]f .
   2

 F2,1   1W W 
   2

 F1,2   1W W 
   2

 F1,0   1W W 
F2,0 1W 2 W

The tensor algorithm of the 3 × 3-point DFT uses the following number of
arithmetical operations: 4 real multiplications and 38 real additions, because
the sequence f is real. Indeed, since W 2 = −1 − W, we have the following:
 
    x  
F1 1 W W2   x + W y − (z + W z)
= y =
F2 1W 2 W x − (y + W y) + W z
z   (1.34)
x − z + W (y − z)
= .
x − y − W (y − z)

The incomplete 3-point Fourier transform

√ can thus be calculated by one com-
plex multiplication by W 1 = (− 3 − j)/2. The three-point DFT of real data
uses one operation of real multiplication, five additions, and one shifting.
Three additions are required to calculate the incomplete three-point DFT,
since F̄2 = F1 . Further, the direct calculation of the matrix [χ] of order 12 × 9
is fulfilled in the given example via 24 operations of real addition and subtrac-
tion. Therefore, the 3×3-point 2-D DFT requires (5+3×3)+24 = 38 additions
for the real sequence f. In the row-column algorithm, six 3-point DFTs are
used, namely, three transforms with real inputs and three transforms with
complex inputs. Therefore, the algorithm uses respectively 3 + 2(3) = 9 real
18 CHAPTER 1: 2-D DFT

multiplications and 3 × 5 + 3 × 16 = 63 additions, when data are real. Thus,

in the tensor and polynomial algorithms, we get the advantage of the number
of multiplications by 3 times, and 1.6 times for additions.

In the general case, for a prime number N > 3, in the traditional row-column
algorithm, 2N one-dimensional N -point DFTs are used. Therefore, the tensor
algorithm decreases the number of multiplications by 2N/(N + 1) times, i.e.,
almost by 2 times, for large N. The tensor transform χσ requires N 3 − N
additions.

Example 1.6 (5 × 5-point DFT) Let N = 5 and f be the following image

of size 5 × 5 :  
 1 2 1 3 1
 
 2 0 1 1 2
 
f =  1 3 2 2 1 .
 4 1 0 1 3
 
 2 4 1 2 1
The underlined unit shows the location of the zero point. We consider the
frequency-point (p1 , p2 ) = (2, 1). All values of t in the equations n1 p1 +n2 p2 =
t mod 5 can be written in the following matrix:
 
 0 1 2 3 4
 
 2 3 4 0 1
 
||t = (n1 · 2 + n2 · 1) mod 5||n2,n1 =0:4 =  4 0 1 2 3 .
 1 2 3 4 0
 
 3 4 0 1 2

Therefore, the components of the splitting-signal fT2,1 of f are defined as

follows:


 f2,1,0 = f0,0 + f1,3 + f2,1 + f3,4 + f4,2 = 1 + 1 + 3 + 3 + 1 = 9

 f2,1,1 = f0,1 + f1,4 + f2,2 + f3,0 + f4,3 = 2 + 2 + 2 + 4 + 2 = 12

fT2,1 = f2,1,2 = f0,2 + f1,0 + f2,3 + f3,1 + f4,4 = 1 + 2 + 2 + 1 + 1 = 7



 f2,1,3 = f0,3 + f1,1 + f2,4 + f3,2 + f4,0 = 3 + 0 + 1 + 0 + 2 = 6

f2,1,4 = f0,4 + f1,2 + f2,0 + f3,3 + f4,1 = 1 + 1 + 1 + 1 + 4 = 8

and fT2,1 = {9, 12, 7, 6, 8}. The sum of the this signal equals the sum of the
image f, i.e.,
4
X 4
X 4
X
f2,1,t = 9 + 12 + 7 + 6 + 8 = 42 = fn1 ,n2 .
t=0 n1 =0 n2=0

The five-point DFT of the splitting-signal equals

{42, 4.6631 − 4.3920j, −3.1631 − 1.4001j, −3.1631 + 1.4001j, 4.6631 + 4.3920j}.

Image Processing 19

This transform coincides with the 2-D DFT of the image f at frequency-points
of the group
T2,1 = {(0, 0), (2, 1), (4, 2), (1, 3), (3, 4)},
as shown in the following table:

2 3
42 0 0 0 0
60
6 0 0 −3.1631 + 1.4001j 0 7
7
6 0 4.6631 − 4.3920j 0 0 0 7.
6 7
40 0 0 0 4.6631 + 4.3920j 5
0 0 −3.1631 − 1.4001j 0 0

We can fill the remaining values of the 2-D DFT, by calculating the 1-D
DFTs of other splitting-signals. For instance, for the signal corresponding to
the generator (p1 , p2 ) = (3, 1), we have the following table of time-points:
 
 0 1 2 3 4
 
 3 4 0 1 2
 
||t = (n1 · 3 + n1 · 1) mod 5||n2,n1 =0:4 =  1 2 3 4 0 .
 4 0 1 2 3
 
 2 3 4 0 1

The components of the splitting-signal fT2,1 are thus calculated as


 f3,1,0 = f0,0 + f1,2 + f2,4 + f3,1 + f4,3 = 1 + 1 + 1 + 1 + 2 = 6


 f3,1,1 = f0,1 + f1,3 + f2,0 + f3,2 + f4,4 = 2 + 1 + 1 + 0 + 1 = 5

fT3,1 = f3,1,2 = f0,2 + f1,4 + f2,1 + f3,3 + f4,0 = 1 + 2 + 3 + 1 + 2 = 9



 f3,1,3 = f0,3 + f1,0 + f2,2 + f3,4 + f4,1 = 3 + 2 + 2 + 3 + 4 = 14

f3,1,4 = f0,4 + f1,1 + f2,3 + f3,0 + f4,2 = 1 + 0 + 2 + 4 + 1 = 8

and fT3,1 = {6, 5, 9, 14, 8}. The five-point DFT of this splitting-signal equals

{42, −8.5902 + 5.7921j, 2.5902 − 2.9919j, 2.5902 + 2.9919j, −8.5902 − 5.7921j}.

This transform defines the 2-D DFT of the image at frequency-points of the
group T3,1 = {(0, 0), (3, 1), (1, 2), (4, 3), (2, 4)}. In this stage, the 2-D DFT is
filled as shown in the following table:
2 3
42 0 0 0 0
60
6 0 2.5902 − 2.9919j −3.1631 + 1.4001j 0 7
7
6 0 4.6631 − 4.3920j 0 0 −8.5902 − 5.7921j 7
6 7.
4 0 −8.5902 + 5.7921j 0 0 4.6631 + 4.3920j 5
0 0 −3.1631 − 1.4001j 2.5902 + 2.9919j 0

In a similar way, the 1-D DFTs of the splitting-signals fT0,1 , fT1,1 , fT4,1 ,
and fT1,0 fill the rest of the table of the 2-D DFT of the image f. The tensor
20 CHAPTER 1: 2-D DFT

representation of the image is the following six splitting-signals:



 fT0,1 ={10, 10, 5, 9, 8}


 fT ={9, 7, 7, 12, 7}
 1,1

fT2,1 ={9, 12, 7, 6, 8}
f→
 fT ={6, 5, 9, 14, 8}
 3,1


 fT ={5, 10, 11, 8, 8}
 4,1

fT1,0 ={8, 6, 9, 9, 10}

1.4.2.1 Code: 2-D DFT by tensor transform

Below is the program with script TT2D Nisprime.m, to accomplish the calcula-
tion of the N × N -point 2-D DFT by the tensor transform. N is a prime.

% ===============================================================
% Call: TT2D_Nisprime.m / A.M. Grigoryan, 11/17/2001
% Demo code for calculating the NxN-point DFT by 2-D DTT.
% (N+1) N-point DFT are used, when N is a prime number.
% For each generator (p,s), two functions are called:
% ft_pst(image,p,s) - calculates the splitting-signal
% cyclic_group(p,s,N) - calculates the cyclic group T(p,s)
% The "tree" image can be taken from the library of
% The Signal and Image Processing Institute, University of
% Southern California, by address http://sipi.usc.edu/database/
fid=fopen(’tree.img’,’rb’);
f=fread(fid,[256,256]); fclose(fid); clear fid; f=f’;
% extend this image to the size 257x257 as
f(:,257)=f(:,256); f(257,:)=f(256,:);
% Set of generators J(257,257)
N=257; N1=N+1;
Jps=ones(2,N1); Jps(1,1:N)=0:N-1; Jps(1:2,N1)=[1,0];
% 2-D DFT by the tensor transform
FT=zeros(N);
for k=1:N1
s=Jps(1,k); p=Jps(2,k);
Image_signal=zeros(1,N); Image_signal=ft_pst(f,p,s);
TDFFT_signal=fft(Image_signal); T_ps=cyclic_group(p,s,N)+1;
for k=1:N
p1=T_ps(k,1); s1=T_ps(k,2); FT(s1,p1)=TDFFT_signal(k);
end
end
% Verify if the 2-D DFT has been calculated correctly
Inv_image=real(ifft2(FT));
Inv_image=Inv_image.*(Inv_image>=0); Inv_image=round(Inv_image);
% Display the original image and inverse 2-D DFT
figure; colormap(gray)
subplot(1,3,1); imagesc(f);
axis(’image’); axis(’off’); title(’Original image’);
Image Processing 21

subplot(1,3,2); image(abs(FT)/N);
axis(’image’); axis(’off’); title(’2-D DFT by TT’);
subplot(1,3,3); imagesc(Inv_image);
axis(’image’); axis(’off’); title(’Inverse 2-D DFT’);
% ============================================================
function T_ps=cyclic_group(p1,p2,N)
T_ps=zeros(N,2); p=0; s=0;
for k=1:N
if p>=N p=p-N; end
if s>=N s=s-N; end
T_ps(k,:)=[p,s]; p=p+p1; s=s+p2;
end

This program also prints the original image, the 2-D DFT and its inverse,
to verify the calculations.

1.4.3 N is a power of two

When N = 2r , r > 1, the irreducible covering σ of the lattice X = X2r ,2r can
be taken as the following family of 3N/2 cyclic groups:
    
σJ = Tp1 ,1 r
, T1,2p2 r−1
. (1.35)
p1 =0:(2 −1) p2 =0:(2 −1)

Thus, to calculate the 2r × 2r -point 2-D DFT, 3 · 2r−1 one-dimensional DFTs

are used in the tensor algorithm, which is described by
r r
2X −1 2X −1
Fkp,ks =(F2r ,2r ◦ f)kp,ks = fn1 ,n2 W n1 kp+n2ks
n1 =0 n2=0
r
2X −1
(1.36)
=(F2r ◦ fT )k = fp,s,t W kt, k = 0 : (2r − 1).
t=0

where the generators (p, s) are taken from the set

J = {(0, 1), (1, 1), (2, 1), ..., (2r − 1, 1)} ∪ {(1, 0), (1, 2), (1, 4), ..., (1, 2r − 2)}.
(1.37)
The components of splitting-signals fT are calculated by the characteristic
functions of sets Vp,s,t ,
r r
2X −1 2X −1 X
fp,s,t = χp,s,t ◦ f = χp,s,t (n1 , n2 )fn1 ,n2 = fn1 ,n2 .
n1 =0 n2 =0 (n1 ,n2 )∈Vp,s,t
(1.38)
These binary functions determine the tensor transform χσ and are defined as

1, if n1 p + n2 s = t mod 2r ,
χp,s,t (n1 , n2) = (n1 , n2 ) ∈ X. (1.39)
0, otherwise,
22 CHAPTER 1: 2-D DFT

All ones in the masks of the functions lie on parallel lines passing the knots
of the corresponding sets Vp,s,t .

Example 1.7 (4 × 4-point DFT) Consider the following image 4 × 4 and

its discrete Fourier transform:
   
1 2 1 3
 
 27
 j 1 −j 
 → {Fp,s } = −2 + j −3j 1 + 6j 
2 1 1 2  1
{fn,m } =  .
1 3 2 1  −3 −2 + j −5 −2 − j 
 
1 1 3 2 −2 − j 1 − 6j 3j 1 

For (p, s) = (1, 1), values of t in the equations np+ms = t mod N, t = 0 : 3,

can be written in the form of the following matrix:
 
0 1 2 3
 
1 2 3 0
||t = (n · 1 + m · 1) mod 4||n,m=0:3 =  .
2 3 0 1

3 0 1 2

Coefficients of this matrix show the points of sets V1,1,t , t = 0 : 3. The image-
signal fT1,1 is defined as


 f1,1,0 = f0,0 + f3,1 + f2,2 + f1,3 = 1 + 2 + 2 + 1 = 6

f1,1,1 = f1,0 + f0,1 + f3,2 + f2,3 = 2 + 2 + 1 + 3 = 8
fT1,1 = (1.40)

 f1,1,2 = f2,0 + f1,1 + f0,2 + f3,3 = 1 + 1 + 1 + 2 = 5

f1,1,3 = f3,0 + f2,1 + f1,2 + f0,3 = 1 + 3 + 1 + 3 = 8.

One can note that

3
X 3
X 3 X
X 3
fp,s,t = f1,1,t = fn,m = 27.
t=0 t=0 n=0 m=0

Therefore the last component f1,1,3 can be calculated as 27 − (6 + 8 + 5) = 8.

The 4-point DFT of the image-signal fT1,1 defines the 4 × 4-point DFT
along the main diagonal as
 
 27 ◦ ◦ ◦ 
 
 ◦ 1 ◦ ◦
fT1,1 = {6, 8, 5, 8} → Fk = {27, 1, −5, 1} → 
 .
 ◦ ◦ −5 ◦ 

 ◦ ◦ ◦ 1

For the generator (p, s) = (1, 2), the corresponding matrix of equations
np + ms = t mod N, t = 0 : 3, is calculated as follows:
 
0 2 0 2 

1 3 1 3 
||t = (n · 1 + m · 2) mod 4||n,m=0:3 =  .
2 0 2 0 
3 1 3 1
Image Processing 23

Therefore, the image-signal fT1,2 of fn,m is defined as



 f1,2,0 = f0,0 + f2,1 + f0,2 + f2,3 = 1 + 1 + 1 + 3 = 6

f1,2,1 = f1,0 + f3,1 + f1,2 + f3,3 = 2 + 2 + 3 + 2 = 9
fT1,2 = (1.41)

 f1,2,2 = f2,0 + f0,1 + f2,2 + f0,3 = 1 + 2 + 2 + 1 = 6

f1,2,3 = f3,0 + f1,1 + f3,2 + f1,3 = 3 + 1 + 1 + 1 = 6.
The 4-point DFT of the image-signal fT1,2 defines the 4 × 4-point DFT for
additional three samples of the set T1,2 = {(0, 0), (1, 2), (2, 0), (3, 2)} as follows:
 
 27 ◦ ◦ ◦ 

 ◦ • −3j ◦ 
fT1,2 = {6, 9, 6, 6} → Fk = {27, −3j, −3, 3j} →  .
 −3 ◦ • ◦ 
 ◦ ◦ 3j • 

To complete the 4 × 4-point DFT at frequency-points of groups T1,0 , T1,3 , T0,1,

and T2,1 , the four-point DFTs of the corresponding splitting-signals are used.
We also consider another example with different set of generators.
Example 1.8 (4 × 4-point DFT) Consider the following 2-D sequence f of
size 4 × 4 :    
 f0,0 f0,1 f0,2 f0,3   1 2 1 3 
   
f f f f  2 0 1 1
f =  1,0 1,1 1,2 1,3  =  .
 f2,0 f2,1 f2,2 f2,3   1 3 2 2 

 f3,0 f3,1 f3,2 f3,3   2 4 1 2 
The set of generators of the cyclic groups Tp,s of the covering σJ equals
J = {(0, 1), (1, 1), (2, 1), (3, 1)} ∪ {(1, 0), (1, 2))}. (1.42)
We first describe the splitting-signal corresponding to the frequency-point
(p, s) = (0, 1). For that, we write all values of t in the equations np + ms =
t mod 4 in the form of the following matrix:
 
0 1 2 3
 
0 1 2 3
||t = (n · 0 + m · 1) mod 4||n,m=0:3 = 
 .
0 1 2 3

0 1 2 3

The components of this splitting-signal are calculated as follows:

   
1 0 0 0 1 2 1 3
 1 0 0 0  2 0 1 1
f0,1,0 = χ0,1,0 ◦ f = 
 1 0 0 0  ◦ 1 3 2 2 = 1 + 2 + 1 + 2 = 6
   (1.43)
1 0 0 0 2 4 1 2
   
0 1 0 0 1 2 1 3
 0 1 0 0  2 0 1 1
f0,1,1 = χ0,1,1 ◦ f = 
 0 1 0 0  ◦ 1 3 2 2 = 2 + 0 + 3 + 4 = 9
   (1.44)
0 1 0 0 2 4 1 2
24 CHAPTER 1: 2-D DFT
   
0 0 1 0 1 2 1 3
0 0 1 0 2 0 1 1
f0,1,2 = χ0,1,2 ◦ f = 
0
◦  =1+1+2+1= 5 (1.45)
0 1 0  1 3 2 2
0 0 1 0 2 4 1 2
   
0 0 0 1 1 2 1 3
0 0 0 1◦2 0 1 1

f0,1,3 = χ0,1,3 ◦ f = 
  = 3 + 1 + 2 + 2 = 8. (1.46)
0 0 0 1 1 3 2 2
0 0 0 1 2 4 1 2
The splitting-signal fT0,1 = {6, 9, 5, 8}. The four-point DFT of this signal
equals (F0 , F1, F2 , F3 ) = (28, 1 − j, −6, 1 + j), which can be written in the
table of the 2-D DFT of f at the frequency-points (0, 0), (0, 1), (0, 2), and
(0, 3) as follows:
   
F0 F1 F2 F3 28 1 − j −6 1 + j
 0 0 0 0  0 0 0 0 
 0 0 0 0= 0 0
   .
0 0 
0 0 0 0 0 0 0 0
We also consider the splitting-signal corresponding to the next generator
(p, s) = (1, 1). For this generator, equations np + ms = t mod 4 result in the
following matrix:
 
0 1 2 3
 
1 2 3 0
||t = (n · 1 + m · 1) mod 4||n,m=0:3 = 
 .
2 3 0 1

3 0 1 2

Therefore, the first component of this splitting-signal is calculated by

   
1 0 0 0 1 2 1 3
 0 0 0 1  2 0 1 1
f1,1,0 = χ1,1,0 ◦ f = 
 0 0 1 0  ◦  1 3 2 2 = 1 + 1 + 2 + 4 = 8 (1.47)
  

0 1 0 0 2 4 1 2
and similarly we obtain the next three components
f1,1,1 = χ1,1,1 ◦ f = 2 + 2 + 2 + 1 = 7
f1,1,2 = χ1,1,2 ◦ f = 1 + 0 + 1 + 2 = 4 (1.48)
f1,1,3 = χ1,1,3 ◦ f = 3 + 1 + 3 + 2 = 9.
Thus, the splitting-signal fT1,1 = {8, 7, 4, 9}. The four-point DFT of this signal
equals (A0 , A1 , A2 , A3 ) = (28, 4 + 2j, −4, 4 − 2j), which defines the 2-D DFT
of f at the frequency-points (0, 0), (1, 1), (2, 2), and (3, 3). At this step, we can
write the other three values of the 2-D DFT as follows:
   
F0 F1 F2 F3 28 1 − j −6 1 + j
 0 A1 0 0   0 4 + 2j 0 0
 0 0 A2 0  =  0
   .
0 −4 0
0 0 0 A3 0 0 0 4 − 2j
Image Processing 25

The first component, A0 , equals F0 = 28, and could be omitted from the
calculations to avoid the redundancy. The redundancy of calculation takes
place for other splitting-signals not only at frequency-point (0, 0), but at
frequency-points with even coordinates, i.e., in the quarter of all frequency-
points, and that can be seen in Figure 1.8.

T0,1 T1,1 T2,1 T3,1 T1,0 T1,2

3 3 3 3 3 3
2 2 2 2 2 2
1 1 1 1 1 1
0 0 0 0 0 0
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3

FIGURE 1.8
The disposition of frequency-points of six groups T of the covering σ of X4,4 .

For instance, the four-point DFT of the splitting-signal fT2,1 = {4, 8, 7, 9}

equals (B0 , B1 , B2 , B3 ) = (28, −3 + j, −6, −3 − j). These values define the 2-D
DFT at the frequency-points (0, 0), (2, 1), (0, 2), and (2, 3). At this step, we
can write the other two values of the 2-D DFT as follows:
   
F0 F1 F2 F3 28 1 − j −6 1 + j
 0 A1 0 0   0 4 + 2j 0 0
 0 B1 A2 B3  =  0 −3 + j −4 −3 − j  .
   

0 0 0 A3 0 0 0 4 − 2j
It is clear, that the incomplete four-point DFT is required, to avoid calcula-
tions for components B0 and B2 , which have already been calculated.
We continue the calculation of the 2-D DFT. The four-point DFT of
the splitting-signal fT3,1 = {5, 7, 7, 9} equals (C0 , C1 , C2 , C3) = (28, −2 +
2j, −4, −2 − 2j). It defines the 2-D DFT of f at the frequency-points
(0, 0), (3, 1), (2, 2), and (1, 3). At this step, we can record two new values of
the 2-D DFT as follows:
   
F0 F1 F2 F3 28 1 − j −6 1+j
 0 A1 0 C3   0 4 + 2j 0 −2 − 2j 
 0 B1 A2 B3  =  0 −3 + j −4 −3 − j  .
   

0 C1 0 A3 0 −2 + 2j 0 4 − 2j
For this signal, the incomplete four-point DFT is required to avoid calculations
for components C0 and C2 , which have already been calculated in the second
step of our calculations (when (p, s) was (1, 1)).
The remaining five values of the 2-D DFT will be calculated by
the splitting-signals corresponding to the generators (1, 0) and (1, 2). The
splitting-signal fT1,0 = {7, 4, 8, 8}. The four-point DFT of this signal equals
(D0 , D1 , D2 , D3 ) = (28, −1 + 5j, 2, −1 − 5j) and defines the 2-D DFT of f at
26 CHAPTER 1: 2-D DFT

the frequency-points (0, 0), (1, 0), (2, 0), and (3, 0). At this step, we can record
three new values of the 2-D DFT as follows:
   
F0 F1 F2 F3 28 1 − j −6 1+j
 D1 A1 0 C3   −1 + 5j 4 + 2j 0 −2 − 2j 
 D2 B1 A2 B3  = 
   .
2 −3 + j −4 −3 − j 
D3 C1 0 A3 −1 − 5j −2 + 2j 0 4 − 2j

The redundancy of calculation on this step is only at (0, 0). At the last step,
the four-point DFT of the splitting-signal fT1,2 = {7, 9, 8, 4} is calculated. It
equals (E0 , E1, E2 , E3 ) = (28, −1 − 5j, 2, −1 + 5j) and defines the 2-D DFT of
f at the frequency-points (0, 0), (1, 2), (2, 0), and (3, 2). We fill the 2-D DFT
with the values of E1 and E2 ,

2 3 2 3 2 3
F0 F1 F2 F3 28 1−j −6 1+j F0,0 F0,1 F0,2 F0,3
7 = 6 −1 + 5j 4 + 2j −1 − 5j −2 − 2j 7 = 6 F1,0
6 D1 A1 E1 C3 7 6 7 6 F1,1 F1,2 F1,3 7
6 7.
4 D2 B1 A2 B3 5 4 2 −3 + j −4 −3 − j 5 4 F2,0 F2,1 F2,2 F2,3 5
D3 C1 E3 A3 −1 − 5j −2 + 2j −1 + 5j 4 − 2j F3,0 F3,1 F3,2 F3,3

For this signal, it is sufficient to perform an incomplete four-point DFT, to

avoid the redundancy of calculations in two points. Thus, the 4 × 4-point 2-D
DFT is calculated in the tensor algorithm by six four-point DFTs. Namely,
one full four-point DFT and five incomplete four-point DFTs are required in
this algorithm.

Example 1.9 (Image 256 × 256) We consider the representation of the

image in the group of frequencies, which is generated by (p1 , p2 ) = (16, 1).
Figure 1.9 shows (a) the clock-and-moon image 256×256 in part a, along with
the image-signal fT16,1 of length 256 in b, the 1-D DFT of the image-signal
(in absolute scale) in c, and 256 samples of this 1-D DFT at frequency-points
of the subset T16,1 of X256,256 at which the 2-D DFT of the image is filled by
the 1-D DFT in d. The value of the 1-D DFT at point 0, which is the sum
43885 of the image, has been truncated in parts c and d.

Example 1.10 (Image 256 × 256) Figure 1.10 illustrates the image 256 ×
256 in part a, along with the 1-D DFT over the image-signal fT1,3 of length
256 in c, and the magnitude spectrum of the image in d. Three bright parallel
lines on the spectrum show the samples at points of the group T1,3 at which
the 2-DFT of the image is the 1-DFT of the image-signal.
The 2-D DFT at samples at this group has been amplified, in order to see
the location of the group and directions of the projection, along which the
components of the tensor are calculated as sums of line-integrals. The image
after amplifying the 2-DFT at samples of the group T1,3 is shown in part b.
The effect of amplifying those spectral components illustrates the meaning of
Equation (1.21). The projections are calculated at angle ψ = 18.4349◦ and the
1-D DFT fills the 2-D DFT along three lines at angle θ = 90 − ψ = 71.5651◦.
Image Processing 27
176

174
fT (t)
16,1

172

170

t
168
0 50 100 150 200 250
(a) (b)
43885
150
F0=43885
200
100
150

100
50
50

0 200
200 100 p1
0 p2 100
0 50 100 150 200 250 0 0

(c) (d)

FIGURE 1.9
(a) The image 256 × 256, (b) image-signal corresponding to the generator
(p1 , p2 ) = (16, 1), (c) absolute spectrum of the image-signal (with the trun-
cated zero component), and (d) the arrangement of values of the 1-D DFT in
the 2-D DFT of the image (in the 3-D view).

Example 1.11 (Image 512 × 512) Figure 1.11 illustrates the image of size
512 × 512 in part a, along with the splitting-signal fT3,1 of length 512 in b, the
1-D DFT over this splitting-signal in c, and points of the subset T3,1 at which
the 2-D DFT of the image is filled by the 1-D DFT in d. The image is the
max-image of stacked fluorescent in situ hybridization (FISH) images arising
from normal glands. In general, the image f is real and components fp,s,t can
be considered to be coefficients of decomposition of the 2-D DFT in the group
Tp,s by the cos(ωx)- and sin(ωx)-waves with the frequency ω = 2πt/N. In
the 3-D representation fn,m → fp,s,t , two dimensions correspond to frequency
(p, s) and one dimension refers the time coordinate t of the signal fTp,s .

1.4.4 N is a power of an odd prime

We now consider a splitting of the N × N -point DFT, when N = Lr , L > 1
is an arbitrary odd prime number, and r > 1. The irreducible covering σJ =
(Tp,s ) of the lattice XLr ,Lr consists of Lr−1 (L + 1) cyclic groups and can be
28 CHAPTER 1: 2-D DFT

90°
ψ θ
(a) (b)
50

1−D DFT of fT
40 1,3
amplitude

30

20

10

0
0 50 100 150 200 250
ω

(c) (d)

FIGURE 1.10
(a) Image 256 × 256. (b) Image after amplifying the 2-DFT at samples of
the group T1,3 . (c) Absolute value of the 1-D DFT of the image-signal fT1,3
(zero component is shifted to the center and truncated). (d) 2-D DFT of the
image with amplified samples at the group T1,3 . Angles ψ = 18.4349◦ and
θ = 71.5651◦.

defined by the following set of generators:

r r−1
L[ −1 [ L [−1
J = JLr ,Lr = (p1 , 1) (1, Lp2 ). (1.49)
p1 =0 p2 =0

Thus, to calculate the Lr × Lr -point DFT, it is sufficient to fulfill Lr−1 (L + 1)

Lr -point DFTs of the splitting-signals. The column-row algorithm uses 2(Lr )
Lr -point DFTs, i.e., 2L/(L + 1) times more. For instance, the 9 × 9-point 2-D
DFT is calculated in the tensor algorithm by 12 one-dimensional nine-point
DFTs. In the column-row algorithm, this 2-D transform requires 18 nine-point
DFTs. The tensor algorithm can be improved, when removing the redundancy
of calculations in the intersections of the cyclic groups Tp,s of the covering.
The redundancy of the algorithm is in calculation of the spectral components
at all frequency-points (p1 , s1 ) with coordinates that are integer multiple to
L, i.e., when (p1 , s1 ) = (Lp mod N, Ls mod N ) and (p, s) ∈ J.
Image Processing 29
50
splitting−signal fT
48
3,1

46

44

42

40

38
0 100 200 300 400 500
(a) (b)
150
DFT(0)=2233

100
T3,1

50

0
0 100 200 300 400 500

(c) (d)

FIGURE 1.11
(a) FISH image, (b) the splitting-signal fT3,1 , (c) 1-D DFT of the splitting-
signal (in absolute scale and shifted), and (d) arrangement of values of the
1-D DFT in the 2-D DFT of the image at frequency-points of the set T3,1 .

1.4.5 Case N = L1 L2 (L1 6= L2 > 1)

We consider the N = L1 L2 case, where L1 and L2 are arbitrary coprime
numbers > 1. The irreducible covering σJ of the lattice XL1 L2 ,L1 L2 consists
of (L1 + 1)(L2 + 1) cyclic groups Tp,s . Such covering σJ can be determined by
the following set of generators:
L2 −1
L1 [  [ 
J= (p1 , 1) ∪{(1, 0)}∪ (1, p2 ) ∪{(L1 , L2 )}∪{(L2, L1 )}.
p1 =0 g.c.d.(p2 ,L1 L2 )>1
(1.50)
To calculate the 2-D (L1 L2 )×(L1 L2 )-point DFT, it is sufficient to perform
(L1 + 1)(L2 + 1) L1 L2 -point 1-D DFTs of splitting-signals fTp,s . For instance,
the calculation of the 20×20-point 2-D DFT is reduced to calculation of thirty
20-point DFTs, instead of forty 20-point DFTs in the column-row algorithm.

1.4.6 General case

In the general N × N case, the construction of the irreducible covering
 
σJ = Tp,s (1.51)
p,s∈J
30 CHAPTER 1: 2-D DFT

of the square lattice XN,N can be implemented in the following way [48]. We
first define the set BN = {n ∈ XN ; g.c.d.(n, N ) > 1} and function β(p), which
equals the number of elements s ∈ BN being coprime with p and such that
ps < N. Then the set of generators can be defined as follows:
N−1
[ [ [  [ [ 
J= (p, 1) 1, s) (p, s) . (1.52)
p=0 s∈BN p,s∈BN , g.c.d.(p,s)=1, p,s≤N

The number of generators in this set, or the number of 1-D DFTs required to
calculate the 2-D N × N -point DFT equals
X
card σJ = 2N − φ(N ) + β(p) (1.53)
p∈BN

where we denote the Euler’s function by φ(N ), that is, the number of positive
integers P
that are smaller than N and coprime with N. It is easy to verify, that
φ(N ) ≥ {β(p); p ∈ BN }, so that card σ ≤ 2N.

1.4.7 Other orders N1 × N2

The tensor algorithm, as well as the improved tensor algorithm, can be con-
structed for other orders N1 × N2 of the 2-D DFT, when N1 6= N2 > 1. The
tensor algorithm is defined by the irreducible covering σJ of the lattice XN1 ,N2
by the cyclic groups Tp,s . It is not difficult to compose such covering for each
case N1 × N2 under consideration.
Example 1.12 (The 3 × 6-point DFT) We consider the 3 × 6-point disc-
rete Fourier transformation, F3,6 , i.e., the case when N1 = 3 and N2 = 6. The
irreducible covering σJ of the lattice X3,6 can be defined by the set of four
generators; J = {(0, 1), (1, 1), (2, 1), (1, 3)}. All cyclic groups of this covering
and their intersections are shown below:
T0,1 = {(0, 0), (0, 1), (0, 2), (0, 3), (0, 4), (0, 5)}
T1,1 = {(0, 0), (1, 1), (2, 2), (0, 3), (1, 4), (2, 5)}
T2,1 = {(0, 0), (2, 1), (1, 2), (0, 3), (2, 4), (1, 5)}
 
(0, 0), (0, 3) 
T1,3 = (1, 0), (1, 3) .
 
(2, 0), (2, 3)
Figure 1.12 illustrates the locations of all frequency-points of these four groups.
The groups are intersected only at two frequency-points (0, 0) and (0, 3), i.e.,
when the coordinates of the points are integer multiple to 3. Therefore, the
transformation F3,6 can be split by one six-point DFT and three incomplete
six-point DFTs. We note for comparison, that in the column-row algorithm,
three six-point DFTs with real inputs and six three-point DFTs with complex
inputs are used. This algorithm uses therefore more operations of multiplica-
tion than the tensor algorithm.
Image Processing 31

T0,1 T1,1 T2,1 T1,3

5 5 5 5
4 4 4 4
3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 1 2 0 1 2 0 1 2 0 1 2

FIGURE 1.12
Arrangement of frequency-points of groups Tp1 ,p2 covering the lattice X3,6 .

Example 1.13 (The 6 × 8-point DFT) We consider the 6 × 8-point Fou-

rier transformation, F6,8 . The irreducible covering σJ of the lattice X6,8 can
be defined by the set of generators J = {(1, 1), (2, 1), (1, 2), (1, 4), (1, 0)}. In-
tersections of cyclic groups of this covering are shown below:
 
 (0, 0), (1, 1), (2, 2), (3, 3), (4, 4), (5, 5), (0, 6), (1, 7)
T1,1 = (2, 0), (3, 1), (4, 2), (5, 3), (0, 4), (1, 5), (2, 6), (3, 7)
 
(4, 0), (5, 1), (0, 2), (1, 3), (2, 4), (3, 5), (4, 6), (5, 7)
 
 (0, 0), (2, 1), (4, 2), (0, 3), (2, 4), (4, 5), (0, 6), (2, 7)
T2,1 = (4, 0), (0, 1), (2, 2), (4, 3), (0, 4), (2, 5), (4, 6), (0, 7)
 
(2, 0), (4, 1), (0, 2), (2, 3), (4, 4), (0, 5), (2, 6), (4, 7)
 
 (0, 0), (1, 2), (2, 4), (3, 6)
T1,2 = (4, 0), (5, 2), (0, 4), (1, 6)
 
(2, 0), (3, 2), (4, 4), (5, 6)
  (1.54)
 (0, 0), (1, 4)
T1,4 = (2, 0), (3, 4)
 
(4, 0), (5, 4)
 

 (0, 0) 

 


 (1, 0)

 
(2, 0)
T1,0 = .

 (3, 0)
 
 (4, 0) 


 
 
(5, 0)

The locations of frequency-points of the five cyclic groups Tp,s of the covering
are shown in Figure 1.13.
The groups are intersected only at 12 frequency-points with even coordi-
nates. Therefore, the transformation F6,8 can be split by one 24-point DFT,
two incomplete 24- and 12-point DFTs, and two incomplete six-point DFTs.
These incomplete DFTs are not calculated for the components with even
points, and, therefore, they can be reduced to calculation of the DFTs of
twice smaller orders. Thus the 6 × 8-point DFT can be split by the 24-, 12-,
32 CHAPTER 1: 2-D DFT

T1,1 T2,1 T1,2 T1,4 T1,0

6 6 6 6 6

4 4 4 4 4

2 2 2 2 2

0 0 0 0 0
0 5 0 5 0 5 0 5 0 5

FIGURE 1.13
Arrangement of frequency-points of groups Tp,s covering the lattice X6,8 .

and 6-point DFTs, and two 3-point DFTs, and the redundancy of calculations
in the tensor algorithm will be removed.

1.5 Discrete Fourier transform and its geometry

For a better understanding the tensor representation of the 2-D DFT, we first
consider the geometry of the 1-D DFT, as a beautiful system rotating the data
on the plane around circles. For that, the block structure of the matrix of the
N -point discrete Fourier transform (DFT) in the real space R2N is described.
All blocks 2 × 2 of this matrix correspond to the Givens transformations, or
elementary rotations describing the multiplications by twiddle coefficients.
The N -point discrete Fourier transform of the signal is defined as a de-
composition of the signal by N roots of the unit
2π 2π 2π
W k = WNk = e−i N k = ck − isk = cos( k) − i sin( k), k = 0 : (N − 1).
N N
The roots are located on the unit circle, (W k )N = 1. The DFT of the vector-
signal f = (f0 , f1 , f2 , ..., fN−1)0 , which is calculated by
N−1
X
Fp = Rp + iIp = W npfn , p = 0 : (N − 1),
n=0

has the following matrix in the complex space C N :

 
1 1 1 1 1 1
 1 W 1 W 2 W 3 · · · W N−1 
 
 1 W 2 W 4 W 6 · · · W N−2 
[FN ] = 
1 · · ·
.
 ··· ··· ··· ···  
1 
N−1 N−2 N−3 1
1W W W W
Image Processing 33

We now describe in matrix form the multiplication of the complex number

x = x1 + ix2 , which is considered as the column-vector (x1 , x2 )0 or row-vector
(x1 , x2 ), by the twiddle coefficients W k ,
   
x1 c x + sk x2
x= → W kx = k 1 , k = 0 : (N − 1),
x2 ck x2 − sk x1

where W k = (ck , −sk ) = cos ϕk − i sin ϕk , and the angles ϕk = 2πk/N. In

matrix form, this multiplication can be written as
  
k cos ϕk sin ϕk x1
T x= .
− sin ϕk cos ϕk x2

The matrix of rotation by the angle ϕk = kϕ1 is denoted by T k . In the k = 0

case, ϕ0 = 0, T = I, and there is no rotation of the signal-data.
Thus, the complex plane is transferred into the 2-D real space, C → R2 ,
and each operation of multiplication by the twiddle coefficient is considered
as the elementary rotation, or the Givens transformation, W k → T k , k = 0 :
(N − 1). We next consider the inclusion of the complex space C N into the real
space R2N by

f = (f0 , f1 , ..., fN−1)0 → f̄ = (r0 , i0 , r1 , i1 , ..., rN−1, iN−1 )0 ,

where we denote rk = Refk and ik = Imfk for k = 0 : (N − 1). The vector f̄

is composed from the original vector, or signal f , and its vector-component is
denoted by f̄k = (f¯2k , f¯2k+1)0 = (rk , ik )0 . The N -point DFT of f is represented
in the real space R2N as the 2N -point transform
  N−1   N−1
X  cos(kϕp ) sin(kϕp )   rk 
Rp X
kp rk
F̄p = = T = , (1.55)
Ip ik − sin(kϕp ) cos(kϕp ) ik
k=0 k=0

where p = 0 : (N − 1).
In matrix form, the DFT in the space R2N is described by the following
matrix 2N × 2N :
 
I I I I I I
 I T 1 T 2 T 3 · · · T N−1 
 2

X= I T T 4 T 6 · · · T N−2 . (1.56)
I · · · · · · · · · · · · · · · 
I T N−1 T N−2 T N−3 T1

The (n, p)-th blocks 2 × 2 of this matrix is T np, where n, p = 0 : (N − 1). The
matrix I = I2 is the identity matrix 2 × 2. The rotation matrices T k compose
a one-parametric group with period N. In other words

T k1 +k2 = T k1 T k2 , (T 0 = T N = I),

for any k1 , k2 = 0 : (N − 1).

34 CHAPTER 1: 2-D DFT

According to (1.55), when calculating the component of the DFT at the

frequency-point p, the vector-components, or points (rk , ik ) are rotated by
the corresponding angles kϕp . Thus the first point (r0 , i0 ) stays on p its place,
and the second point (r1 , i1 ) is rotated around the circle of radius rp 2 2
1 + i1 by
angle ϕp . The next point (r2 , i2 ) is rotated around the circle of radius r2 + i22
2

by the twice larger angle 2ϕp , or twice faster, than the point (r1 , i1 ), and so on.
Then the coordinates of all rotated points are added and one point is defined,
which represents the DFT at frequency-point p. For p = 0, there is no rotation;
the sums of coordinates of the original points define the component F0 .

Example 1.14 Consider the vector-signal of length 10 with the following real
and imaginary parts:

{rk ; k = 0 : 9} = { 5, −3, 2,−5, −1, 3,−4, −2, 7, −6}

{ik ; k = 0 : 9} = {−2, 1,4,−3, 7, 2,−5, 5, −1, 3}

This signal is plotted in the form of ten points on the plane in Figure 1.14. The
points are numbered and ten circles on which they lie are shown. The figure

10−point DFT composition: component F(0).

8
4

6
7

2
imaginary part y(n) of the signal z(n)=x(n)+jy(n)

4
9

5
2
1 F(0)

0
8

0
−2
3

−4
6

−6

−8
−8 −6 −4 −2 0 2 4 6 8

FIGURE 1.14
Calculation of the first component of the DFT, F0 .

also shows projections of all points of the signal on the X-axis and Y -axis
(shown by dash lines) and their sums (by x), and the point F0 = (−0.4, 1.1)
(by the bullet with label F0 ). The coordinates of F0 are normalized by the
factor of N = 10, i.e., F0 /10.
Image Processing 35
10−point DFT composition: component F(1).
8

6
9
6

imaginary part y(n) of the signal z(n)=x(n)+jy(n)

4
2

2
F(0)
7

0
1
F(1)
5 0
−2

3
−4

4
−6
8

−8
−8 −6 −4 −2 0 2 4 6 8

FIGURE 1.15 (See color insert)

Calculation of the second component of the DFT, F1 .

In the next stage of the DFT calculation, points of the signals are rotated as
shown in Figure 1.15, and the sums (normalized by 10) of projections of the ro-
tated points on the X-axis and Y -axis define the point F1 = (0.0684, −0.6155).
Similarly, these ten points are rotated eight more times, and during these
rotations the sums of coordinates of ten points overlay a new set of ten points
with the following coordinates:

{Rp} = {−0.4, 0.0684, 0.7729,−2.0982, 4.0537, 2.2,

−0.0648, 0.1566,−0.3618,0.6733}
{Ip } = { 1.1, −0.6155,−0.3367, 0.4982,−0.1407, −0.5,
−0.5211,−1.5836,−0.1015,0.2009}

where p = 0 : 9. These ten points Fp = (Rp , Ip ) are show in Figure 1.16.

1.5.1 Inverse DFT

The inverse discrete Fourier transform (IDFT) is defined as
N−1
1 X −np
fn = W Fp , n = 0 : (N − 1). (1.57)
N p=0
36 CHAPTER 1: 2-D DFT
10−point DFT composition: F(p), p=0:9.
8
4

6
7
imaginary part y(n) of the signal z(n)=x(n)+jy(n) 2
4
9

5
2
1 F(0)

F(8) F(9) F(4)

0 F(3)
F(2) F(5) 8
F(6)
F(1)
0
F(7)
−2
3

−4
6

−6

−8
−8 −6 −4 −2 0 2 4 6 8

FIGURE 1.16 (See color insert)

Result of ten rotations of the signal-data.

In the real space R2N , the N -point IDFT of {Fp; p = 0 : (N −1)} is represented
as
  N−1   N−1   
rn 1 X −np Rp 1 X cos(nϕp ) − sin(nϕp ) Rp
f̄n = = T = ,
in N p=0 Ip N p=0 sin(nϕp ) cos(nϕp ) Ip

where n = 0 : (N − 1).
The illustration of the IDFT is similar to the DFT, but is a clock-wise
rotation, i.e., in the direction opposite to the rotation in the DFT. If the
projections of all rotated points have been not normalized for the DFT, then
after rotating the points, the sums of the coordinates of rotated points are
normalized by the factor of N. Thus, if we start with the ten points {Fp ; p =
0 : 9} in the above considered example, as shown in Figure 1.16, the sequential
clockwise rotations of these points around the circles, on which they lie, define
the original signal, i.e., ten points with coordinates

{rk ; k = 0 : 9} = { 5, −3, 2, −5,−1, 3, −4,−2, 7, −6},

{ik ; k = 0 : 9} = {−2, 1,4, −3, 7,2, −5, 5,−1, 3}.

Example 1.15 We consider the following symmetric complex signal z(n) of

length N = 201,

z(n) = x(n) + iy(n) = n + i(n/5), z(−n) = z(n), n = 0 : 100.

Image Processing 37

Figures 1.17-1.20 illustrate the geometry of a few components of the 201-point

DFT. These figures were generated by code horoscope 1byDFTfiles.m, which
can be found at http://www.fasttransforms.com in the folder Lectures/DSP.
The star in these figures denotes the value of F0 .

DFT component, F(p), at frequency−point p=1 DFT component, F(p), at frequency−point p=2
100 100

80 80

60 60

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 −100
−100 −80 −60 −40 −20 0 20 40 60 80 100 −100 −80 −60 −40 −20 0 20 40 60 80 100

FIGURE 1.17
DFT components F1 and F2 .

DFT component, F(p), at frequency−point p=3 DFT component, F(p), at frequency−point p=4
100 100

80 80

60 60

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 −100
−100 −80 −60 −40 −20 0 20 40 60 80 100 −100 −80 −60 −40 −20 0 20 40 60 80 100

FIGURE 1.18
DFT components F3 and F4 .

When fn,m is a real image and of size N × N, all its N 2 first points are
38 CHAPTER 1: 2-D DFT

DFT component, F(p), at frequency−point p=66 DFT component, F(p), at frequency−point p=67
100 100

80 80

60 60

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 −100
−100 −80 −60 −40 −20 0 20 40 60 80 100 −100 −80 −60 −40 −20 0 20 40 60 80 100

FIGURE 1.19
DFT components F66 and F67.

DFT component, F(p), at frequency−point p=115 DFT component, F(p), at frequency−point p=120
100 100

80 80

60 60

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 −100
−100 −80 −60 −40 −20 0 20 40 60 80 100 −100 −80 −60 −40 −20 0 20 40 60 80 100

FIGURE 1.20
DFT components F115 and F120 .

located on the real line and then are rotated, step-by-step, N 2 times in the
plane R2 . The process of rotation of all points is more complicated when
comparing with the N -point DFT. For instance, the point f1,1 is rotated as
W p+s f1,1 , i.e.,
f1,1 → W 1 f1,1 → W 2 f1,1 → W 3 f1,1 → ... → W N−1 f1,1
for components F1,0 , F2,0, F3,0, ..., FN−1,0, as well as for components F0,1 , F0.2,
Image Processing 39

F0,3 , ..., F0,N−1 , and also rotated as f1,1 → W 2 f1,1 → W 3 f1,1 → W 4 f1,1 →
... → f1,1 for components F1,1 and F2,1, F3,1 , ..., FN−1,1, and so on. This pro-
cess of rotation of point f1,1 is fulfilled when combined with other points fn,m ,
that are rotated as W np+ms fn,m , where (p, s) ∈ X \ (1, 1).
In tensor representation, one can describe the entire picture of rotation, or
the geometry of all components of the 2-D DFT. Indeed, the image N × N is
the set of splitting-signals of length N each, and the process of rotation of the
splitting-signal generated by frequency-point (p, s) describes the 2-D DFT at
frequency-points of the corresponding group Tp,s .

Problems
Problem 1.1 Consider the following image: {fn,m = (n + m) mod 6, n = 0 :
7, m = 0 : 3.} Determine the number of splitting-signals that describe the
image in tensor representation. Calculate the tensor transform and plot all
splitting-signals of the image.

Problem 1.2 Consider the tree image of size 257 × 257 in tensor represen-
tation. Calculate and plot two splitting-signals generated by frequency-points
(1, 2) and (17, 34). Express one signal in terms of another signal.

Problem 1.3 For the tree image of size 256 × 256, calculate and plot the
splitting-signal generated by frequency-point (1, 7). Compute the 1-D DFT
of the splitting-signal. In the lattice 256 × 256, show the frequency-points at
which the 2-D DFT of the image is defined by this splitting-signal.

Problem 1.4 Consider the tree image of size 256 × 256 and its 2-D DFT in
tensor representation.
A. Calculate and sketch the splitting-signal of the (1, 13)-projection.
B. Assume that the (1, 12)-projection is not calculated, or is not available.
(i) How would the tree image be changed if the splitting-signal of
the (1, 12)-projection was substituted by the splitting-signal of the (1, 13)-
projection?
(ii) How would the tree image be changed if the splitting-signal of the
(1, 12)-projection was substituted by the zeros?

Problem 1.5 Consider the tree image of size 256 × 256 in tensor representa-
tion.
A. Remove the horizontal, vertical, and diagonal projections from the set
of projections. Calculate and show the image reconstructed by the remaining
255 projections.
B. Calculate and sketch the image reconstructed only by three projections,
which are the horizontal, vertical, and diagonal projections.
40 CHAPTER 1: 2-D DFT

C. Calculate and sketch the image reconstructed only by two projections,

the horizontal and vertical projections.
Problem 1.6 Repeat Problem 1.5 for the tree image extended to the size
257 × 257.
Problem 1.7 (Circular Convolution) Consider a discrete image fn,m of size
N ×N, where N > 1 is a prime, and the small window 3×3 with the coefficients

1/9, if n, m = −1, 0, 1;
hn,m =
0, otherwise.
A. Determine if the circular 2-D convolution
XX
(f ⊗ h)n,m = fn+n1 mod N,m+m1 mod N hn1 ,m1
n1 m1

can be accomplished by N + 1 circular 1-D convolutions of splitting-signals of

the image and window.
B. Sketch the results in A for the tree image fn,m of size 257 × 257.
Problem 1.8 (Motion of particles in the ring) Consider the binary image
of the following ellipse: {f = (x, y)} = {(a sin φ, b cos φ); φ ∈ [0, 2π)},where
a = 2 and b = 1. Compose the complex 1-D signal fn = (xn , yn ) = xn + jyn of
length 64, which represents this ellipse, by selecting 64 values of φ uniformly
distributed in [0, 2π). Use the 1-D DFT-based system to rotate 64 selected
points of the ellipse in the ring b2 ≤ x2 + y2 ≤ a2 , and draw the geometry of
such rotation, for rotations number p = 1, 2, 16, 21, 31, 32, 61, 62, and 63.
Problem 1.9 (Geometry of the signals) Consider the tree image of size
256 × 256 and the splitting-signal fT2,1 , which are shown in Figure 1.21 in
parts a and b, respectively.
140

130

120

110

100
0 50 100 150 200 250
(a) (b)

FIGURE 1.21
(a) Tree image and (b) the splitting-signal generated by (p, s) = (2, 1).

Use the concept of the 1-D DFT in the real space and sketch the geometry
of components F1 , F2 , F3 , F83, F110, and F127 of the 256-point DFT of the
splitting-signal. Discuss the reconstruction of the splitting-signal from the
geometry of any of these components.
2
Direction Images

This chapter discusses the decomposition of the image by direction images,

which is based on the concept of the tensor representation and its advanced
form, the paired representation [41]-[48]. The 2-D image is considered to be of
the size N × N, where N is prime, a power of two, and a power of odd prime.
The tensor and paired representations in the frequency-and-time domain de-
fine the image as a set of 1-D splitting-signals. Each of such splitting-signals
is calculated as the sum of the image along the parallel lines, and it defines
the direction image as a component of the original image. The unique decom-
position of the image by direction images is described and formulas for the
inverse tensor and paired transforms are given. These formulas can be used for
image reconstruction from projections [53]-[56], when splitting-signals or their
direction images are calculated directly from the projection data. The number
of required projections is uniquely defined by the tensor representation of the
image.

2.1 2-D direction images on the lattice

The components fp1 ,p2 ,t , t = 0 : (N − 1), of the splitting-signals of the image
fn1 ,n2 are defined as sums of the image at points lying on the corresponding
set Vp1 ,p2 ,t defined in (1.18). To describe these sets, we first consider the case
when N1 = N2 = N. Given a sample (p1 , p2 ) ∈ X and a nonnegative integer
t < N, the set

Vp1 ,p2 ,t = {(n1 , n2 ); n1 p1 + n2 p2 = t mod N, n1 , n2 = 0 : (N − 1)},

if it is not empty, is the set of points (n1 , n2 ) along a family of parallel straight
lines at the angle of ψ = − arctan(p2 /p1 ) to the horizontal axis. The equations
for these lines are 
xp1 + yp2 =t 


xp1 + yp2 =t + N
(2.1)
... . ... 


xp1 + yp2 =t + kN
where k ≤ p1 + p2 . We denote this family by Lp1 ,p2 ,t . For different values of
t1 6= t2 < N, the families of lines Lp1 ,p2 ,t1 and Lp1 ,p2 ,t2 do not intersect. All

41
42 CHAPTER 2: Direction Images

together, the sets Vp1 ,p2 ,t , t = 0 : (N − 1), compose a partition of the period
X. It is interesting to note, that the direction of parallel lines of Lp1 ,p2 ,t is
perpendicular to the direction of frequency-points of the cyclic group Tp1 ,p2 .
Example 2.1 (Lattice 8 × 8) On the lattice X8,8 , we consider two sets of
parallel lines L2,1,1 and L2,1,2. Each family contains three parallel lines. For
the family L2,1,1, the parallel lines are

y1 : 2x + y = 1, y9 : 2x + y = 9, y17 : 2x + y = 17.

One point (0, 1) of the set V2,1,1 lies on the first line of L2,1,1, four points
(1, 7), (2, 5), (3, 3), (4, 1) on the second line, and three points (5, 7), (6, 5), (7, 3)
on the third one. Therefore, f2,1,1 = (x0,1) + (x1,7 + x2,5 + x3,3 + x4,1 ) + (x5,7 +
x6,5 + x7,3 ). The parallel lines of the family L2,1,2 are defined by

y2 : 2x + y = 2, y10 : 2x + y = 10, y18 : 2x + y = 18,

and the component f2,1,2 = (x0,2 +x1,0 )+(x2,6 +x3,4 +x4,2 +x5,0 )+(x6,6 +x7,4 ).
The disposition of the points lying on the parallel lines of these sets is given

V2,1,1 V2,1,2 T2,1

8 8 8 l2
6 6 6

4 4 4
l1
2 2 2

0 0 0

0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
y1 y9 y17 y2 y10 y18

FIGURE 2.1
The locations of points of sets V2,1,1 and V2,1,2 and frequency-points of the
group T2,1 .

in Figure 2.1. The location of the frequency-points of the group T2,1 is also
shown in this figure. Two parallel lines pass through these frequency-points,
which are defined in the frequency plane (w1 , w2 ) as l1 : 2w2 − w1 = 0 and
l2 : 2w2 − w1 = 8. The parallel lines l1 and l2 are perpendicular to the parallel
lines yn of L2,1,1 and L2,1,2 , as well as all other families L2,1,t, t = 0, 3 : 7.
The disposition of the points of all disjoint eight sets V2,1,t , when t = 0 : 7,
is given in Figure 2.2.
We can identify the opposite sides of boundaries of the square Y = [0, N ]×
[0, N ] and consider Y as a torus and the 2-D lattice X as a net traced on the
Image Processing 43

V2,1,0 V2,1,1 V2,1,2 V2,1,3

8 8 8 8
6 6 6 6
4 4 4 4
2 2 2 2
0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8

V2,1,4 V2,1,5 V2,1,6 V2,1,7

8 8 8 8
6 6 6 6
4 4 4 4
2 2 2 2
0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8

FIGURE 2.2
The disposition of eight sets of points, V2,1,t , t = 0 : 7 (shown by filled circles).

torus in the 3-D space. Then, the straight lines of Lp1 ,p2 ,t will compose a
closed spiral St = Sp,s,t on the torus. As an example for N = 32, Figure 2.3
shows the location of points of sets V1,1,3 and V1,1,7 on the square grid 32 × 32
in parts (a) and (b), respectively. Each set is located on two parallel lines.

V1,1,3 V1,1,7

30 30

25 25

20 20

15 15

10 10

5 5

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(a) (b)

FIGURE 2.3
The sets (a) V1,1,3 and (b) V1,1,7 on the lattice 32 × 32.

We now consider the same sets on the discrete torus. Figure 2.4 shows the
locus of two spirals S3 and S7 on the net, for (p, s) = (1, 1). They correspond
respectively to the parallel straight lines of families L1,1,3 and L1,1,7 described
by (2.1). The sums {fp1 ,p2 ,t ; t = 0 : (N − 1)}, calculated on N parallel spirals
St , t = 0 : (N −1), represent the image fn1 ,n2 in the group Tp1 ,p2 of frequency-
44 CHAPTER 2: Direction Images

S
1,1,7

S
1,1,3

−5
10

10
5
5
0
0
−5
−5

−10 −10

FIGURE 2.4 (See color insert)

The net with knots of the grid 32 × 32 in the 3-D space with locus of two
spirals S1,1,3 and S1,1,7 .

points. These points lie on the spiral that passes through the initial point (0, 0)
of the net and make an angle π/2 with the spirals St .
Each image-signal is the set of discrete integrals along the family Lp1 ,p2 ,t
of parallel lines. Therefore, the processing of the image-signal fT yields the
change in the Fourier transform at frequency-points of the corresponding cyclic
group Tp1 ,p2 . After performing the inverse 2-D discrete Fourier transform, the
corresponding change might be observed in the image along the parallel lines
of sets Vp1 ,p2 ,t , t = 0 : (N − 1). As an example, Figure 2.5 shows the tree
image of size 256 × 256 in part a, along with the results of amplifying only
one image-signal fT2,1 by the factor of 4 in b, and signal fT5,1 by the factor
of 6 in c. The directions of parallel lines of the corresponding families L2,1,t,
and L5,1,t, t = 0 : 255, on the image can easily be observed.

2.1.1 Superposition of directions

The images of Figure 2.5 illustrate the fact that an image f = fn1 ,n2 can
be composed by specific collection of direction images. The representation of
the image as the set of splitting-signals leads to the image decomposition by
direction images. To show that, we consider the tensor representation of the
image
{fn1 ,n2 } → {fT ; T = Tp1 ,p2 ∈ σJ },
where σ = σJ is the irreducible covering of the period X. We assume that the
size of X is N × N, and J is the set of generators.
Image Processing 45

(a) (b) (c)

FIGURE 2.5
(a) Tree image before and after processing by the image-signals (b) fT2,1 and
(c) fT5,1 .

The tensor transform can be used for calculating and processing the 2-D
discrete Fourier transform (DFT) of the image, since each of splitting-signal
fTp,s defines the 2-D DFT of the image at frequency-points of the cyclic group
Tp,s . The image can thus be calculated from the tensor transform, by using
the inverse 2-D DFT. The image can also be reconstructed directly from its
tensor transform, without calculating the inverse 2-D DFT. The splitting-
signals define unique direction components of the image.
As an example, Figure 2.6 shows the bridge image 256 × 256 extended to
the size 257 × 257 in part a, along with the image of all 258 splitting-signals
in b, the set of 1-D DFTs of these splitting-signals in c, which represent the
2-D DFT of the image shown in d. The splitting-signals and their DFTs are

(a) image (b) 2−D TT (c) 1−D DFTs (d) 2−D DFT

FIGURE 2.6
(a) Image 257 × 257, (b) image of 258 splitting-signals, (c) image of 258 257-
point DFTs of the splitting-signals, and (d) 2-D DFT of the image. (All DFTs
are shown in absolute mode.)

written row-wise in images each of size 258 × 257 in (b) and (c), respectively.
These images can be reduced to the original size 257 × 257. Indeed, the sum
46 CHAPTER 2: Direction Images

of all components of the splitting-signals is equal to the sum of the image, i.e.,
N−1
X X N−1
N−1 X
fp,s,t = f0,0,0 = fn,m .
t=0 n=0 m=0

Therefore, we can save one complete splitting-signal, let say, {f0,1,t ; t = 0 :

(N − 1)}, and only (N − 1) first components in the remaining N splitting-
signals, {f1,p,t ; t = 0 : (N − 2)}, where p = 0 : (N − 1). The total number of
the required components of all signals equals N + N (N − 1) = N 2 . The DFTs
of splitting-signals carry the same information only at the frequency-point
(0, 0), which equals f0,0,0 .
For a given generator (p, s) 6= (0, 0), we consider the following incomplete
2-D DFT with values at frequency-points of the group Tp,s :

Fkp,ks, if (p1 , p2 ) = (kp, ks), k = 0 : (N − 1),
Dp1 ,p2 = Dpp,s1 ,p2
= (2.2)
0, otherwise,

where p1 , p2 = 0 : (N − 1). Examples of such incomplete 2-D DFTs were given

in Example 1.6 for the N = 5 case. We define the direction image dn1,n2 as
the inverse 2-D DFT of the data D,
N−1 N−1
−1 1 X X
dn1,n2 = d(p,s)
n1,n2 = (FN,N ◦ D
p,s
)n1 ,n2 = Dp ,p W n1 p1 +n2 p2
N 2 p =0 p =0 1 2
1 2

n1 , n2 = 0 : (N − 1).
(2.3)
Since the splitting-signal fTp,s determines the 2-D DFT of the image at
frequency-points of the group T = Tp,s ,

(FN ◦ fTp,s )k = Fkp,ks , k = 0 : (N − 1),

the following calculations hold:

N−1 N−1 N−1
1 X X n1 p1 +n2 p2 1 X
dn1 ,n2 = Dp ,p W = Fkp,ks W n1 (kp)+n2(ks)
N 2 p =0 p =0 1 2 N2
1 2 k=0
N−1
1 1 X 1
= Fkp,ks W k(n1p+n2 s) = fp,s,(n1p+n2 s) mod N .
NN N
k=0
(2.4)
Thus, N values of the splitting-signal are placed on all N 2 points of the square
lattice XN,N . Namely, each value fp,s,t is placed at all points that are located
on the parallel lines of the corresponding family Lp,s,t .
The interesting property of the tensor transform is derived. The direction
image is composed of N values of the splitting-signal, that are placed at all
points of the image N × N along the parallel lines.
As an example, Figure 2.7 shows the tree image of size 257 × 257 in part
Image Processing 47

Original image
140

135

130

125

120

115 image−signal (1,5)

110
0 50 100 150 200 250
(a) (b)

250

200

150 1−D DFT

100

50

0
0 50 100 150 200 250
(c)
(d) DI−(1,5)

FIGURE 2.7
(a) The image 257 × 257, (b) the splitting-signal {f1,5,t ; t = 0 : 256}, and (c)
the 1-D DFT of the splitting-signal and frequency-points of T1,5 , and (d) the
corresponding direction image dn1,n2 (the image has been scaled).

(a), along with the splitting-signal generated by the frequency (p, s) = (1, 5)
in b. The original size 256 × 256 of the tree image was extended to the size
257 × 257, by adding a zero row and column. The 1-D DFT of the splitting-
signal in absolute scale and the location of all frequency-points of the cyclic
group T1,5 , at which the 2-D DFT of the image is defined by this splitting-
(1,5)
signal, are shown in c. The direction image dn1,n2 is illustrated in d.
(p,s)
The direction images dn1 ,n2 of the tree image, which are generated by the
frequencies (p, s) = (0, 1), (1, 1), (2, 1), ..., and (9, 1) are shown in Figure 2.8
in parts a-j, respectively.
When N is a prime, (N +1) direction signal-images are required to compose
the image fn,m . Indeed, the covering σJ = (Tp,s ; (p, s) ∈ J) consists of (N +1)
cyclic groups Tp,s . The groups T of σJ intersect only at the point (0, 0).
(p,s)
Therefore, the sum of all (N + 1) incomplete 2-D DFTs Dp1 ,p2 , (p, s) ∈ JN,N ,
equals the 2-D DFT of the image plus N times single values of the F0,0 at
48 CHAPTER 2: Direction Images

(a) DI−(0,1) (b) DI−(1,1) (c) DI−(2,1) (d) DI−(3,1) (e) DI−(4,1)

(f) DI−(5,1) (g) DI−(6,1) (h) DI−(7,1) (i) DI−(8,1) (j) DI−(9,1)

FIGURE 2.8
(a)-(j) The first ten direction images of the tree image 257 × 257. (All images
were scaled.)

point (0, 0). We denote by Op1 p2 the complex matrix with all zero coefficients,
except the first coefficient that equals O0,0 = F0,0 .
The inverse 2-D DFT of the sum of all (N + 1) incomplete 2-D DFTs can
be calculated as follows:
 
X X h i X
−1 −1
d(p,s)
n1,n2 = FN,N ◦ Dp,s = FN,N ◦ Dp,s 
n1 ,n2
(p,s)∈JN,N (p,s)∈JN,N (p,s)∈JN,N
n1 ,n2
h
i
−1
= FN,N ◦ Fp1 ,p2 + N Op1 p2
n1 ,n2
   
−1 −1
= FN,N ◦ Fp1 ,p2 + N FN,N ◦ Op1 p2
n1 ,n2 n1 ,n2
1
= fn1 ,n2 + F0,0 = fn1 ,n2 + N E[f].
N
(2.5)
Here E[f] denotes the mean of the image,
N−1 N−1
1 X X 1 1
E[f] = fn,m = 2 F0,0 = 2 f0,0,0 .
N 2 n=0 m=0 N N

We obtain the following formula of the inverse tensor transformation:

X 1 X
fn1 ,n2 = d(p,s)
n1,n2 − N E[f] = fp,s,(n1 p+n2 s) mod N − N E[f].
N
(p,s)∈JN,N (p,s)∈JN,N
(2.6)
Considering the set JN,N as

JN,N = {(1, 0), (1, 1), (1, 2), (1, 3), ..., (1, N − 1)} ∪ {(0, 1)},
Image Processing 49

we obtain the following statement.

Statement 1: (Superposition by direction images) The image fn1 ,n2 of size
N × N, where N > 2 is a prime, can be composed of (N + 1) direction images
as follows:
"N−1 #
X
(1,s) (0,1)
fn1 ,n2 = dn1,n2 + dn1 ,n2 − N E[f]
s=0
"N−1 #
1 X (2.7)
= f1,s,(n1+sn2 ) mod N + f0,1,n2 − N E[f],
N s=0
n1 , n2 = 0 : (N − 1).

To simplify the above equation, we could assume that the image is centered
fn1 ,n2 → fn1 ,n2 − E[f]. Then, the sum of direction images equals the image,
X 1 X
fn1 ,n2 = d(p,s)
n1,n2 = fp,s,(n1p+n2 s) mod N . (2.8)
N
(p,s)∈J (p,s)∈J

If we define the set of generators to be

JN,N = {(0, 1), (1, 1), (2, 1), (3, 1), ..., (N − 1, 1)} ∪ {(1, 0)},

Statement 1 would be written as follows:

"N−1 #
X
(p,1) (1,0)
fn1 ,n2 = dn1 ,n2 + dn1,n2 − N E[f]
p=0
"N−1 # (2.9)
1 X
= fp,1,(n1p+n2 ) mod N + f1,0,n1 − N E[f].
N p=0

The image is the sum of components of (N + 1) splitting-signals, or the sum

of (N + 1) direction images of the tensor representation of the image. The
above inverse formulas for the discrete image reconstruction from its tensor
transform requires only two operations of multiplication and division by N.
The division by N can be considered in the definition of the tensor transform
components. Then, the reconstruction requires only N 2 such operations.
Each directional d(p,s) completes the original image with details, or paral-
lel straight lines of different brightness in gray scale. The direction of these
lines is defined by the generator (p, s). Equation (2.7) describes the princi-
ple of superposition of the direction components in image formation. Each
direction image is determined by the corresponding splitting-signal, which is
calculated by the discrete linear integrals (sums) of the image fn1 ,n2 along the
parallel lines of Lp,s,t , t = 0 : (N − 1). These integrals can be calculated from
the projection data along the angle defined by the generator (p, s). Thus we
obtain the simple formula of reconstruction of the discrete image by its projec-
tion data on the Cartesian lattice, by using the splitting-signals of the tensor
50 CHAPTER 2: Direction Images

representation of the image. The number of projections equals the number of

generators, i.e., (N + 1), when N is prime. This is the required number of
projections for the exact reconstruction of the image N × N.
As an example, Figure 2.9 shows the tree-image after removing the pro-
jection data fp,s,(n1 p+n2 s) mod N that correspond to the direction defined by
the generator (p, s) = (1, 0) in part a, (p, s) = (1, 1) in b, and (p, s) = (0, 1) in
c.

(a) (b) (c)

FIGURE 2.9
Images 257 × 257 reconstructed by 257 projections by splitting-signals, when
the projection data of one generator (p, s) have been removed, for (p, s) equals
(a) (1, 0), (b) (1, 1), and (c) (0, 1).

The angles of the required projections for reconstructing the image of size
257 × 257 by the splitting-signals (or direction images) compose the following
set
Φ257,257 = {arctan(p); p = 0 : 257} ∪ {π/2}.
Figure 2.10 illustrates all central angles of this set on the unit circle. One

0.8 90°
0.6
°
45
0.4 26.57° °
18.43
0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

FIGURE 2.10
Central angles of 258 projections for reconstructing the image 257 × 257.

can see that the points on the circle are not uniformly distributed, and the
increment of the angles is not a constant. The main part of the projections are
Image Processing 51

taken along the small angles. For instance, there are no projections for angles
in the intervals (45◦ , 90◦) and (26.57◦, 45◦), as well as (18.43◦, 26.57◦).

2.2 The inverse tensor transform: Case N is prime

In many practical applications when an image is processed through the fre-
quency domain, the size of the image is considered to be a power of two, such
as 256 × 256, in the case N × N, because of the fast algorithm for the 1-D
discrete Fourier transform. Many transformations possess complete systems of
basis functions, which are orthogonal, and the 2-D paired transformation is one
such transformations. The tensor transform is not orthogonal, but invertible.
Therefore, the availability of effective algorithms of the inverse tensor trans-
form would be very useful for many applications, such as image reconstruction
from projections, image filtration, and enhancement. The basic functions of
the tensor transform define the line-integrals of the discrete image and add
them in a specific way, to compose the transform. These basic functions are
binary, with values of 0 and 1 only, and therefore not orthogonal. However,
such system is complete in the sense that the system defines the inverse ten-
sor transform, as well. We here describe the system of basic functions of the
inverse tensor transform, and then show how to calculate these functions by
the basic functions of the tensor transform.
Consider the tensor representation of the image f = {fn,m } of size N × N,
χN,N : {fn,m } → {{fp,s,t ; t = 0 : (N − 1)}; (p, s) ∈ JN,N }, (2.10)
in the case when N is prime. The set of generators is
JN,N = {(1, 0), (1, 1), (1, 2), ..., (1, N − 1)} ∪ {(0, 1)}.
Given (p, s) ∈ JN,N , the components of the splitting-signal {fp,s,0 , fp,s,1 , ...,
fp,s,N−1 } are defined as the sums of the image along the parallel lines,
X
fp,s,t = fn,m , t = 0 : (N − 1), (2.11)
(n,m)∈Vp,s,t

where the sets Vp,s,t = {(n, m); np + ms = t mod N }.

2.2.1 Inverse tensor transform

The formula of image reconstruction from its tensor transform, which is given
in the principle of superposition, can be described by using the complete
system of basic functions of the tensor transform. These functions

1; if np + ms = t mod N,
χp,s,t (n, m) = t = 0 : (N − 1),
0;otherwise,
52 CHAPTER 2: Direction Images

define the components fp,s,t = χp,s,t ◦ f.

The number of basic functions, or components in the tensor transform
equals N (N + 1) = N 2 + N, which exceeds the size of the image. The re-
dundancy can be removed, if we consider that the sum of all components of
each splitting-signal {fp,s,0 , fp,s,1 , ..., fp,s,N−1} equals N 2 E(f), and the last
component of the signal can be calculated by

fp,s,N−1 = N 2 E(f) − (fp,s,0 + fp,s,1 + · · · + fp,s,N−2 ).

Therefore, we define the tensor transform as the following set of N 2 com-

ponents:

{f1,s,0, f1,s,1 , ..., fp,s,N−2}, s = 0 : (N − 1),
χ[f] = (2.12)
{f0,1,0, f0,1,1, ..., fp,s,N−2, f0,1,N−1}.

The system of basic functions of the tensor transform is thus considered to be


{χ1,s,0, χ1,s,1, ..., χp,s,N−2}, s = 0 : (N − 1),
χ= (2.13)
{χ0,1,0, χ0,1,1, ..., χp,s,N−2, χ0,1,N−1}.

All basic functions are numbered in the order shown in the above equation,
i.e., the first function is χ1,0,0 , the second is χ1,0,1 , ..., the N th is χ1,1,0,
..., the N (N − 1)th is χ1,N−1,N−2 , the N (N − 1) + 1 is χ0,1,0, ..., and the
last function is χ0,1,N−1 . Thus, the triplet-number (p, s, t) corresponds to the
following number k = k(p, s, t) from the interval [0, N 2 − 1] :

(N − 1)s + t, if p = 1, and s = 0 : (N − 1), t = 0 : (N − 2),
k(p, s, t) =
N (N − 1) + t, if p = 0, and s = 1, t = 0 : (N − 1).

The set of ordered triplet-numbers (p, s, t) is denoted by U = UN,N .

The inverse tensor transform is described by N 2 functions that are similar
to the functions χp,s,t . The basic functions in (2.13) can be written row-wise
in the matrix N × N and its inverse matrix can be calculated. Columns of
the inverse matrix correspond to the apures of the basis functions of the
inverse tensor transform. We denote these two matrices by [χN,N ] and [χ−1N,N ],
respectively.
As an example, consider the N = 5 case. The matrix 25 × 25, composed
of the apures of 25 basic functions of the tensor transform, and the transpose
inverse matrix equal, respectively,
Image Processing 53

T
[χ5,5 ] = [χ−1
5,5 ] =
1111100000000000000000000 11111000000000000000 − 1 − 1 − 1 − 1 − 1
2 3 2 3
6 0000011111000000000000000 7 6 00000111110000000000 − 1 − 1 − 1 − 1 − 1 7
6 0000000000111110000000000 7 6 00000000001111100000 − 1 − 1 − 1 − 1 − 1 7
6 7 6 7
6 0000000000000001111100000 7 6 00000000000000011111 − 1 − 1 − 1 − 1 − 1 7
6 7 6 7
6 1000000001000100010001000 7 6 1000 − 1000 − 1100 − 1100 − 1100 − 11000 7
6 7 6 7
6 0100010000000010001000100 7 6 0100 − 1100 − 1000 − 1010 − 1010 − 10100 7
6 7 6 7
6 7 6 7
6 0010001000100000000100010 7 6 0010 − 1010 − 1010 − 1000 − 1001 − 10010 7
6 7 6 7
6 0001000100010001000000001 7 6 0001 − 1001 − 1001 − 1001 − 1000 − 10001 7
6 7 6 7
6 1000000100000010100000010 7 6 10 − 1000010 − 10 − 1001010 − 10 − 10010 7
6 7 6 7
6 0001010000001000000101000 7 6 00 − 1101000 − 10 − 1100000 − 11 − 11000 7
6 7 6 7
6 0100000010100000010000001 7 6 01 − 1000001 − 11 − 1000001 − 10 − 10001 7
6 7 6 7
6 0000101000000101000000100 7 6 00 − 1010100 − 10 − 1010100 − 10 − 10100 7
6 7 6
1 6
7
6 1000000010010000000100100 7,
6 7 5 6
100 − 100 − 10100100 − 100 − 101 − 10100 7
7.
6 0010010000000100100000001 7 6 001 − 101 − 10000001 − 101 − 100 − 10001 7
6 7 6 7
6 0000100100100000001001000 7 6 000 − 110 − 11001000 − 100 − 110 − 11000 7
6 7 6 7
6 0100000001001001000000010 7 6 010 − 100 − 10010010 − 110 − 100 − 10010 7
6 7 6 7
6 1000001000001000001000001 7 6 1 − 100001 − 100001 − 100001 − 1 − 10001 7
6 7 6 7
6 0000110000010000010000010 7 6 0 − 100110 − 100010 − 100010 − 1 − 10010 7
6 7 6 7
6 0001000001100000100000100 7 6 0 − 101000 − 101100 − 100100 − 1 − 10100 7
6 7 6 7
6 0010000010000011000001000 7 6 0 − 110000 − 110000 − 111000 − 1 − 11000 7
6 7 6 7
6 1000010000100001000010000 7 6 00000 00000 00000 00000 50000 7
6 7 6 7
6 0100001000010000100001000 7 6 −11000 − 11000 − 11000 − 11000 41000 7
6 7 6 7
6 0010000100001000010000100 7 6 −10100 − 10100 − 10100 − 10100 40100 7
6 7 6 7
4 0001000010000100001000010 5 4 −10010 − 10010 − 10010 − 10010 40010 5
0000100001000010000100001 −10001 − 10001 − 10001 − 10001 40001

The tensor transform Z = (Z0 , Z1 , Z2 , ..., ZN 2−1 )T of the image {fn,m }

can be considered as the product of the matrix [χN,N ] on the image-vector
f = (fn+mN = fn,m ; n, m = 0 : N − 1)T ,
Z = [χN,N ]f ,
and fp,s,t = Zk(p,s,t), (p, s, t) ∈ U. If we pack the tensor transform into the
matrix (N × N ) by Zk → Zk1 ,k2 , where k1 + k2 N = k, k = 0 : (N 2 − 1), then
the following formula can be considered:
(
Z t−s ,if p = 1, s = 0 : (N − 1), t = 0 : (N − 2),
fp,s,t = (s(N−1)+t) mod N,bs+ N c
ZN−1,t , if p = 0, s = 1, t = 0 : (N − 1),

where b.c denotes the floor function.

One can notice that all basic functions of the inverse tensor transform have
values of ±1 and 0, except the last five functions. Each of these functions
considered as the 2-D matrix has one coefficient that equals 5 or 4 on the last
five rows (in column number 21). Such peculiarity holds for matrices N × N
for other cases of prime N > 2; these matrices have coefficients N and N − 1
in the last N rows.
54 CHAPTER 2: Direction Images

For example, the transpose inverse matrix [χ−1 T

7,7 ] of the 7 × 7-point tensor
transform presents the following table (up to the factor of 1/7):
1 1 1 1 1 1 1-0-0 0 0 0 0 0 0 0-0 0-0-0 0-0-0-0-0 0-0-0-0-0 0 0-0-0-0-0 0 0 0-0 0-0-1-1-1-1-1-1-1
0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0-1-1-1-1-1-1-1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0-1-1-1-1-1-1-1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0-1-1-1-1-1-1-1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0-1-1-1-1-1-1-1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1-1-1-1-1-1-1-1
1 0 0 0 0 0-1 0 0 0 0 0-1 1 0 0 0 0-1 1 0 0 0 0-1 1 0 0 0 0-1 1 0 0 0 0-1 1 0 0 0 0-1 1 0 0 0 0 0
0 1 0 0 0 0-1 1 0 0 0 0-1 0 0 0 0 0-1 0 1 0 0 0-1 0 1 0 0 0-1 0 1 0 0 0-1 0 1 0 0 0-1 0 1 0 0 0 0
0 0 1 0 0 0-1 0 1 0 0 0-1 0 1 0 0 0-1 0 0 0 0 0-1 0 0 1 0 0-1 0 0 1 0 0-1 0 0 1 0 0-1 0 0 1 0 0 0
0 0 0 1 0 0-1 0 0 1 0 0-1 0 0 1 0 0-1 0 0 1 0 0-1 0 0 0 0 0-1 0 0 0 1 0-1 0 0 0 1 0-1 0 0 0 1 0 0
0 0 0 0 1 0-1 0 0 0 1 0-1 0 0 0 1 0-1 0 0 0 1 0-1 0 0 0 1 0-1 0 0 0 0 0-1 0 0 0 0 1-1 0 0 0 0 1 0
0 0 0 0 0 1-1 0 0 0 0 1-1 0 0 0 0 1-1 0 0 0 0 1-1 0 0 0 0 1-1 0 0 0 0 1-1 0 0 0 0 0-1 0 0 0 0 0 1
1 0 0-1 0 0 0 0 0 0 1 0 0-1 0 0-1 0 0 0 1 0 0 1 0 0-1 0 0-1 0 0 0 1 0 0 1 0 0-1 0 0-1 0 0 0 1 0 0
0 0 0-1 1 0 0 1 0 0 0 0 0-1 0 0-1 1 0 0 0 0 0 0 0 0-1 1 0-1 1 0 0 0 0 0 0 0 0-1 1 0-1 1 0 0 0 0 0
0 1 0-1 0 0 0 0 0 0 0 1 0-1 1 0-1 0 0 0 0 0 0 0 1 0-1 0 0-1 0 0 0 0 1 0 0 1 0-1 0 0-1 0 0 0 0 1 0
0 0 0-1 0 1 0 0 1 0 0 0 0-1 0 0-1 0 1 0 0 1 0 0 0 0-1 0 0-1 0 1 0 0 0 0 0 0 0-1 0 1-1 0 1 0 0 0 0
0 0 1-1 0 0 0 0 0 0 0 0 1-1 0 1-1 0 0 0 0 0 0 0 0 1-1 0 1-1 0 0 0 0 0 0 0 0 1-1 0 0-1 0 0 0 0 0 1
0 0 0-1 0 0 1 0 0 1 0 0 0-1 0 0-1 0 0 1 0 0 1 0 0 0-1 0 0-1 0 0 1 0 0 1 0 0 0-1 0 0-1 0 0 1 0 0 0
1 0-1 0 0 0 0 0 0 1 0-1 0 0 0 0 0 0 1 0-1 0-1 0 0 0 0 1 0 1 0-1 0 0 0 0 0 0 1 0-1 0-1 0 0 0 0 1 0
0 0-1 0 0 1 0 1 0 0 0-1 0 0 0 0 1 0 0 0-1 0-1 0 0 1 0 0 0 0 0-1 0 0 1 0 1 0 0 0-1 0-1 0 0 1 0 0 0
0 0-1 1 0 0 0 0 0 0 0-1 1 0 1 0 0 0 0 0-1 0-1 1 0 0 0 0 0 0 0-1 1 0 0 0 0 0 0 0-1 1-1 1 0 0 0 0 0
0 1-1 0 0 0 0 0 0 0 1-1 0 0 0 0 0 0 0 1-1 1-1 0 0 0 0 0 0 0 1-1 0 0 0 0 0 0 0 1-1 0-1 0 0 0 0 0 1
0 0-1 0 0 0 1 0 1 0 0-1 0 0 0 0 0 1 0 0-1 0-1 0 0 0 1 0 1 0 0-1 0 0 0 0 0 1 0 0-1 0-1 0 0 0 1 0 0
0 0-1 0 1 0 0 0 0 0 0-1 0 1 0 1 0 0 0 0-1 0-1 0 1 0 0 0 0 0 0-1 0 1 0 1 0 0 0 0-1 0-1 0 1 0 0 0 0
1 0 0 0 0-1 0 0 0 0-1 0 1 0 0-1 0 1 0 0 0 0 1 0 0 0 0-1 0 0 0 0-1 0 1 0 0-1 0 1 0 0-1 0 1 0 0 0 0
0 0 1 0 0-1 0 1 0 0-1 0 0 0 0-1 0 0 0 1 0 0 0 0 1 0 0-1 0 1 0 0-1 0 0 0 0-1 0 0 0 1-1 0 0 0 1 0 0
0 0 0 0 1-1 0 0 0 1-1 0 0 0 1-1 0 0 0 0 0 0 0 0 0 0 1-1 0 0 0 1-1 0 0 0 1-1 0 0 0 0-1 0 0 0 0 0 1
0 0 0 0 0-1 1 0 0 0-1 1 0 0 0-1 1 0 0 0 0 1 0 0 0 0 0-1 0 0 0 0-1 1 0 0 0-1 1 0 0 0-1 1 0 0 0 0 0
0 1 0 0 0-1 0 0 0 0-1 0 0 1 0-1 0 0 1 0 0 0 0 1 0 0 0-1 1 0 0 0-1 0 0 0 0-1 0 0 1 0-1 0 0 1 0 0 0
0 0 0 1 0-1 0 0 1 0-1 0 0 0 0-1 0 0 0 0 1 0 0 0 0 1 0-1 0 0 1 0-1 0 0 1 0-1 0 0 0 0-1 0 0 0 0 1 0
1 0 0 0-1 0 0 0-1 0 0 1 0 0 0 1 0 0 0-1 0 0 0-1 0 0 1 0 0 0 1 0 0 0-1 0 0 0-1 0 0 1-1 0 0 1 0 0 0
0 0 0 1-1 0 0 1-1 0 0 0 0 0 0 0 0 0 1-1 0 0 1-1 0 0 0 0 0 0 0 0 0 1-1 0 0 1-1 0 0 0-1 0 0 0 0 0 1
0 0 0 0-1 0 1 0-1 0 1 0 0 0 1 0 0 0 0-1 0 0 0-1 0 1 0 0 0 1 0 0 0 0-1 0 0 0-1 0 1 0-1 0 1 0 0 0 0
0 0 1 0-1 0 0 0-1 0 0 0 0 1 0 0 0 1 0-1 0 1 0-1 0 0 0 0 0 0 0 0 1 0-1 0 1 0-1 0 0 0-1 0 0 0 0 1 0
0 0 0 0-1 1 0 0-1 1 0 0 0 0 0 0 0 0 0-1 1 0 0-1 1 0 0 0 1 0 0 0 0 0-1 0 0 0-1 1 0 0-1 1 0 0 0 0 0
0 1 0 0-1 0 0 0-1 0 0 0 1 0 0 0 1 0 0-1 0 0 0-1 0 0 0 1 0 0 0 1 0 0-1 1 0 0-1 0 0 0-1 0 0 0 1 0 0
1-1 0 0 0 0 0 0 1-1 0 0 0 0 0 0 1-1 0 0 0 0 0 0 1-1 0 0 0 0 0 0 1-1 0 0 0 0 0 0 1-1-1 0 0 0 0 0 1
0-1 0 0 0 0 1 1 0-1 0 0 0 0 0 1 0-1 0 0 0 0 0 1 0-1 0 0 0 0 0 1 0-1 0 0 0 0 0 1 0-1-1 0 0 0 0 1 0
0-1 0 0 0 1 0 0 0-1 0 0 0 1 1 0 0-1 0 0 0 0 1 0 0-1 0 0 0 0 1 0 0-1 0 0 0 0 1 0 0-1-1 0 0 0 1 0 0
0-1 0 0 1 0 0 0 0-1 0 0 1 0 0 0 0-1 0 0 1 1 0 0 0-1 0 0 0 1 0 0 0-1 0 0 0 1 0 0 0-1-1 0 0 1 0 0 0
0-1 0 1 0 0 0 0 0-1 0 1 0 0 0 0 0-1 0 1 0 0 0 0 0-1 0 1 1 0 0 0 0-1 0 0 1 0 0 0 0-1-1 0 1 0 0 0 0
0-1 1 0 0 0 0 0 0-1 1 0 0 0 0 0 0-1 1 0 0 0 0 0 0-1 1 0 0 0 0 0 0-1 1 1 0 0 0 0 0-1-1 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 7 0 0 0 0 0 0
-1 1 0 0 0 0 0-1 1 0 0 0 0 0-1 1 0 0 0 0 0-1 1 0 0 0 0 0-1 1 0 0 0 0 0-1 1 0 0 0 0 0 6 1 0 0 0 0 0
-1 0 1 0 0 0 0-1 0 1 0 0 0 0-1 0 1 0 0 0 0-1 0 1 0 0 0 0-1 0 1 0 0 0 0-1 0 1 0 0 0 0 6 0 1 0 0 0 0
-1 0 0 1 0 0 0-1 0 0 1 0 0 0-1 0 0 1 0 0 0-1 0 0 1 0 0 0-1 0 0 1 0 0 0-1 0 0 1 0 0 0 6 0 0 1 0 0 0
-1 0 0 0 1 0 0-1 0 0 0 1 0 0-1 0 0 0 1 0 0-1 0 0 0 1 0 0-1 0 0 0 1 0 0-1 0 0 0 1 0 0 6 0 0 0 1 0 0
-1 0 0 0 0 1 0-1 0 0 0 0 1 0-1 0 0 0 0 1 0-1 0 0 0 0 1 0-1 0 0 0 0 1 0-1 0 0 0 0 1 0 6 0 0 0 0 1 0
-1 0 0 0 0 0 1-1 0 0 0 0 0 1-1 0 0 0 0 0 1-1 0 0 0 0 0 1-1 0 0 0 0 0 1-1 0 0 0 0 0 1 6 0 0 0 0 0 1

A simple analytical formula can be derived for basic functions of the inverse
tensor transform (ITT). To show that, we number the basic functions of the
ITT by the same triplets (p, s, t) ∈ U and preserve their orders as in the
forward TT. Then the following holds:


 χp,s,t (n, m) − χp,s,N−1 (n, m)
1  if p = 1, s = 0 : (N − 1), t = 0 : (N − 2),
χ−1
p,s,t (n, m) = (2.14)
N  χ (n, m) − χp,s,0 (n, m) + N δN−1,0 (n, m)
 p,s,t

if p = 0, s = 1, t = 0 : (N − 1),

where n, m = 0 : (N − 1). As an example, we first consider the N = 7 case

and the triplet-number (p, s, t) = (1, 5, 3). This number corresponds to the
row with number k = k(1, 5, 3) = 34 in the matrix [χ7,7].
Therefore we consider the 34-th row in the matrix [χ−1 T
7,7] . We obtain the
Image Processing 55

following basic functions:

   
0 0 0 1 0 0 0 0 0 0 1 0 0 −1
0 0 0 0 0 1 0  0 −1 0 0 0 1 0 
   
1 0 0 0 0 0 0  1 0 0 −1 0 0 0 
 −1 1  
0 0 1 0 0
[χ1,5,3] =  0 0 , [χ1,5,3] = 7  0 0 1 0 0−1 0  ,
 
 −1 0 0 0 1 0 0 
0 0 0 0 1 0 0  

0 0 0 0 0 0 1   0 0 −1 0 0 0 1 
0 1 0 0 0 0 0 0 1 0 0 −1 0 0

and  
0 0 0 0 0 0 1
0 1 0 0 0 0 0 
 
0 0 0 1 0 0 0 
 
[χ1,5,3] − 7[χ−1
1,5,3] = [χ1,5,6] =  0 0 0 0 0 1 0  ,
 
1 0 0 0 0 0 0 
 
0 0 1 0 0 0 0 
0 0 0 0 1 0 0
−1 1


or [χ1,5,3] = 7 [χ1,5,3] − [χ1,5,6] .
One also can see that for the 35-th basic function, we have the following:
   
0 0 0 0 1 0 0 0 0 0 0 1 0 −1
 0 0 0 0 0 0 1   0 −1 0 0 0 0 1 
   
 0 1 0 0 0 0 0   0 1 0 −1 0 0 0 
   
[χ1,5,4] − 7[χ−1
1,5,4] =  0 0 0 1 0 0 0  −  0 0 0 1 0−1 0  = [χ1,5,6],
   
 0 0 0 0 0 1 0   −1 0 0 0 0 1 0 
   
 1 0 0 0 0 0 0   1 0 −1 0 0 0 0 
0 0 1 0 0 0 0 0 0 1 0 −1 0 0


or [χ−1 1
1,5,4] = 7 [χ1,5,4] − [χ1,5,6] .
Now we consider the triplet-number (p, s, t) = (0, 1, 2). This number cor-
responds to the row number k = k(0, 1, 2) = 45. The basic functions of the
direct and inverse tensor transform with number (0, 1, 2) equal, respectively,
   
0 0 0 0 0 0 0 −1 −1 −1 −1 −1 −1 6
0 0 0 0 0 0 0   0 0 0 0 0 0 0
   
1 1 1 1 1 1 1   1 1 1 1 1 1 1
  −1 1  
 0 0 0 0 0 0 0  , [χ0,1,2] = 7  0 0 0 0 0 0 0  ,
[χ0,1,2] =    
0 0 0 0 0 0 0   0 0 0 0 0 0 0
   
0 0 0 0 0 0 0   0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0
56 CHAPTER 2: Direction Images


and [χ−1
0,1,2] =
1
7 [χ0,1,2] − [χ0,1,0] + [δ6,0 ]. Indeed
 
1 1 1 1 1 1 1−7
0 0 0 0 0 0 0 
 
0 0 0 0 0 0 0 
 
[χ0,1,2] − 7[χ−1
0,1,2 ] = [χ 0,1,0 ] − 7[δ6,0 ] = 0
 0 0 0 0 0 0  .
0 0 0 0 0 0 0 
 
0 0 0 0 0 0 0 
0 0 0 0 0 0 0


For the 46-th function, we have [χ−1
0,1,3] =
1
7
[χ0,1,3] − [χ0,1,0] + [δ6,0 ]. Indeed
   
0 0 0 0 0 0 0 −1 −1 −1 −1 −1 −1 6
0 0 0 0 0 0 0  0 0 0 0 0 0 0
   
0 0 0 0 0 0 0  0 0 0 0 0 0 0

−1
 
[χ0,1,3] − 7[χ0,1,3] = 
1 1 1 1 1 1 1 − 

 1 1 1 1 1 1 1
0 0 0 0 0 0 0  0 0 0 0 0 0 0

 
0 0 0 0 0 0 0  0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0
=[χ0,1,0] − 7[δ6,0].

matrix of TT inverse matrix

10 10

20 20

30 30

40 40

10 20 30 40 10 20 30 40
(a) (b)

original image tensor transform (TT) inverse TT

2 2 2

4 4 4

6 6 6

2 4 6 2 4 6 2 4 6
(c) (d) (e)

FIGURE 2.11
(a) The matrix 49 × 49 of the DTT and (b) transpose inverse matrix, (c) the
image of size 7 × 7, (d) DTT of the image, and (e) inverse DTT.
Image Processing 57

Figure 2.11 shows the binary image of the matrix [χ7,7 ] of the tensor trans-
formation in part a, along with the gray-scale image of the transpose inverse
matrix in b. The image fn,m = (n + m − 1) mod 10, n, m = 0 : 6, is shown in
c, the discrete tensor transform (DTT) of the image in d, and the inverse ten-
sor transform in e. In the image of the inverse matrix in b, the values −1, 0, 1,
and 6, 7 are shown in grays from black to white.

2.2.2 Formula of the inverse tensor transform

As follows directly from Equation (2.14), the following calculations hold for
the inverse discrete tensor transform:
N−1
!
X X N−2
X
N fn,m = fp,s,t χp,s,t (n, m) − f1,s,t χ1,s,N−1 (n, m)
(p,s,t)∈U s=0 t=0
N−1
!
X
− f0,1,t [χ0,1,0(n, m) − N δN−1,0 (n, m)]
t=0
X N−1
X

= fp,s,t χp,s,t (n, m) − N 2 E(f) − f1,s,N−1 χ1,s,N−1 (n, m)
(p,s,t)∈U s=0
2
−N E(f) [χ0,1,0 (n, m) − N δN−1,0 (n, m)]
X N−1
X
= fp,s,t χp,s,t (n, m) + f1,s,N−1 χ1,s,N−1 (n, m)
(p,s,t)∈U s=0
N−1
!
X
2
−N E(f) χ1,s,N−1 (n, m) + χ0,1,0(n, m) − N δN−1,0 (n, m) .
s=0

It is not difficult to see that the image in the round braces is constant,
N−1
X
c(n, m) = χ1,s,N−1 (n, m) + χ0,1,0(n, m) − N δN−1,0 (n, m) ≡ 1.
s=0

Thus, the inversion formula states that the image is composed as the dif-
ference
fn,m = bU (n, m) − e(f), n, m = 0 : (N − 1), (2.15)
of the backprojection
 
N−1
1  X X
bU (n, m) = fp,s,t χp,s,t (n, m) + f1,s,N−1 χ1,s,N−1 (n, m)
N s=0
(p,s,t)∈U
(2.16)
and the constant-image, which is the mean of the splitting-signal
N−1 N−1 N−1
1 X X 1 X
e(f) = N E(f) = fn,m = f1,0,t . (2.17)
N n=0 m=0 N t=0
58 CHAPTER 2: Direction Images

Thus, although we have considered the tensor transform in the form of

(2.12) with N 2 components, the entire tensor (N + 1) × N is actually used
in the image reconstruction, when backprojecting by spirals the transform into
the points (n, m). Therefore, we denote the set of all (N + 1)N triplets in the
tensor transform by
0
U 0 = UN,N = UN,N ∪ {(1, s, N − 1); s = 0 : (N − 1)}.

The 2nd Inverse formula for DTT (Spiral Backprojection): The image of
size N × N, when N is prime, can be calculated from the tensor transform by
1 X
fn,m = fp,s,t χp,s,t (n, m) − e(f), n, m = 0 : (N − 1). (2.18)
N
(p,s,t)∈U 0

We just mentioned that this formula describes the image reconstruction

by projecting back the transform data, fp,s,t , in each point (n, m) along the
spirals, not the line as in the traditional method of backprojection which
was originated by considering the polar system of coordinates instead of the
Cartesian. Indeed, as shown above, one can identify the opposite sides of
boundaries of the square [0, N ] × [0, N ] and consider it in the 3-D space as
a torus and the 2-D Cartesian lattice XN,N = {(n, m); n, m = 0 : (N −
1)} as a net traced on the torus. Then, for a given triplet-number (p, s, t),
the straight lines in the set Vp,s,t compose one discrete spiral Sp,s,t on the
torus, because of the equality (xp + ys) mod N = (t + kN ) mod N = t for all
integers k. Each component fp,s,t of the tensor transform is the sum of line-
integrals, the integral along one spiral on the torus. Equation (2.18) describes
the backprojection by spirals, or spiral backprojection.
Thus, in the case when N is prime, the image reconstruction from its tensor
transform, or the inverse tensor transform, is determined by the same system
of basic functions {χp,s,t }. The tensor transformation is a good example of a
non-separable and non-orthogonal 2-D transformation with the simple system
of non-negative basic functions.

2.2.2.1 Code for inverse tensor transform

Below is the simple code with script demo 7x7inverseTT.m, which is written to
calculate the tensor transform of the image, all basic functions of the trans-
form, and the inverse tensor transform in the 7 × 7 case.
%%%%%%%%%%%%%%%%%%%%%%%% MAIN CODE %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% call: demo_7x7inverseTT.m
% Demo: The inverse formula for the tensor transform 7x7
% Art-Meruzhan Grigoryan, November 12, 2010
N=7; N49=N*N;
fprintf(’ Tensor transform (%g,%g) \n’,N,N);
TT_matrix=[]; k=1; p=1;
for s=0:N-1
Image Processing 59

for t=0:N-2
A=matrix_pst(N,p,s,t); TT_matrix(k,:)=reshape(A,1,N49);
fprintf(’ %g tensor matrix for (%g,%g,%g) \n’,k,p,s,t);
k=k+1;
end
end
s=1; p=0;
for t=0:N-1
A=matrix_pst(N,p,s,t); TT_matrix(k,:)=reshape(A,1,N49);
fprintf(’ %g tensor matrix for (%g,%g,%g) \n’,k,p,s,t);
k=k+1;
end
TT_imatrix=inv(TT_matrix)*N; % (*N) can be removed *
% Calculate TT:
% Example with the following original image:
% 1 2 3 4 5 6 7
% 2 3 4 5 6 7 8
% 3 4 5 6 7 8 9
% 4 5 6 7 8 9 0
% 5 6 7 8 9 0 1
% 6 7 8 9 0 1 2
% 7 8 9 0 1 2 3
for n=1:N
for m=1:N
image_1(n,m)=mod(n+m-1,10);
end
end
signal_1=reshape(image_1,N49,1);
signal_xx=TT_matrix*signal_1;
XX=reshape(signal_xx,N,N);
% XX is the TT of the image_1
% 28 49 29 29 29 29 28
% 35 9 39 39 39 29 35
% 42 19 39 39 29 29 42
% 39 29 39 29 29 39 39
% 36 39 29 29 39 39 36
% 33 29 29 39 39 39 33
% 49 29 39 39 39 39 30 * the last s-signal
% sum(sum(XX(1:7,7)))=243=sum(sum(image_1))
% --------------------------------------------------------------
% Check if the inverse TT is correct:
signal_2=round(TT_imatrix*signal_xx/N); % (/N) can be removed *
image_2=reshape(signal_2,N,N); % it equals image_1
% --------------------------------------------------------------
% The image reconstruction by the derived formulas (2.15)-(2.18)
fprintf(’ Inverse Tensor transform (%g,%g) \n’,N,N);
% TT is the vector-signal ’signal_xx’ of length 49.
B=zeros(N);
p=1; k=1;
60 CHAPTER 2: Direction Images

for s=0:N-1
for t=0:N-2
A=matrix_pst(N,p,s,t); B=B+signal_xx(k)*A; k=k+1;
end
end
s=1; p=0;
for t=0:N-1
A=matrix_pst(N,p,s,t); B=B+signal_xx(k)*A; k=k+1;
end
% Consider additional N matrices (and coefficients)
for s=0:N-1
A=matrix_pst(N,1,s,N-1);
TT=sum(sum(A.*image_1)); % i.e. f_{1,s,N-1}
B=B+TT*A;
end
B=B/N;
C=B-sum(sum(image_1))/N; % image reconstruction
%%%%%%%%%%%%%%%%%%%%%%% End of the code %%%%%%%%%%%%%%%%%%%%%%%%%
% Call: matrix_pst.m / Artyom Grigoryan, 1997.
function A=matrix_pst(N,p,s,t)
if t>N t=mod(t,N); end
A=zeros(N,N);
ms=0;
for m=1:N
np=0;
for n=1:N
t1=np+ms; t1=mod(t1,N); if t1==t A(m,n)=1; end
np=np+p;
end
ms=ms+s;
end

2.3 3-D paired representation

We introduce a more advanced form of representation of images, than the ten-
sor representation. This form associates with special partitions of the lattice
in the frequency domain, which reveal the two-dimensional discrete Fourier
transform. The tensor representation of the image is associated with an irre-
ducible covering σ not being partitions of the lattice. This is why there is a
redundancy in calculations; the same spectral information is contained in dif-
ferent groups of frequency-points and the splitting-signals together contains
more points than the size of the represented image. As an illustration, Fi-
gure 2.12 shows the image 256 × 256 in part a, along with the image of 384
1-D splitting-signals in b, which lie on rows of this image. These 1-D signals
Image Processing 61

describe uniquely the original image and, at the same time, they split the ma-
thematical structure of the 2-D discrete Fourier transform (shown in c) into
a set of separate 1-D transforms (shown all together as an image in d). The

50

100

150

200

250

300

350
(a)
50 100 150 200 250

(b)

50

100
50

150
100

200
150

250
200

300
250
50 100 150 200 250
350

(c)
50 100 150 200 250

(d)

FIGURE 2.12
Conventional and tensor representation of the image and its spectrum. (a)
Image 256 × 256, (b) the image of all 384 splitting-signals, (c) the 256 × 256-
point DFT of the image, and (d) the image of the 256-point DFTs of splitting-
signals. (DFTs are in absolute scale and shifted to the centers.)

2-D DFT is thus considered as a unique set of separate 1-D DFTs. The image
in b is the tensor representation of the image, or the tensor transform of the
image,
χ = χσ : f → {fTp,s ; Tp,s ∈ σJ }. (2.19)
62 CHAPTER 2: Direction Images

One can note that a few splitting-signals are well expressed, which can be
seen also from the energy plot of all splitting-signals, which is given in Fi-
gure 2.13. For this image, the high energy is concentrated in the splitting-

1500

1000

500

0
0 50 100 150 200 250 300 350

FIGURE 2.13
Energy of all 384 splitting-signals of the tree image.

signals of numbers 1, 129, 257, 172, 52, 258, and 2, which correspond to the
generators (p, s) = (0, 1), (128, 1), (1, 0), (171, 1), (51, 1), (1, 2), and (1, 1), res-
pectively. The processing of only these signals may lead to good results, for
instance in image enhancement [49, 50, 52]. One such example is shown in Fi-
gure 1.4, when the image is enhanced by one splitting-signal of the generator
(p, s) = (7, 1). However, all together the splitting-signals contain 384 × 256
points, which is much greater than the original size 256 × 256.
In order to remove such redundancy, we consider the concept of the paired
representation of images. Unlike the tensor representation, the paired rep-
resentation allows for distributing the spectral information of the image by
disjoint sets of frequency-points. This property makes the paired representa-
tion the most attractive in image processing, as well as in developing effective
algorithms for calculating the 2-D DFT transform.

2.3.1 2D-to-3D paired transform

A complete system of functions can be derived from the two-dimensional disc-
rete Fourier transform as a system that splits the transform [40]-[42],[46, 55].
In other words, there exists a system of functions that reveals completely the
mathematical structure of the 2-D DFT and represents it as a minimal compo-
sition of short 1-D DFTs. Such a system is three-dimensional, i.e., the system
of functions are numbered by three parameters, namely, two parameters for
the spatial frequencies and one parameter for the time. The change in time
determines a series of functions, and the total number of triples numbering
the system of functions, since the system is complete, equals the size of the
2-D DFT, let us say N ×N. The complete systems of such functions which are
called the paired functions exist also in the one- and multi-dimensional cases.
Let N be a power of two, N = 2r , r > 1. We consider an irreducible
Image Processing 63

covering σJ of the lattice XN,N , and for each generator (p, s) ∈ J, we determine
the characteristic functions of the sets Vp,s,t = {(n1 , n2 ); n1 p + n2 s = t mod
N },

1; (n1 , n2 ) ∈ Vp,s,t ,
χp,s,t (n1 , n2 ) = t = 0 : (N − 1). (2.20)
0; otherwise,
These functions describe the tensor transformation of the 2-D image f,
X N−1
N−1 X
fp,s,t = χp,s,t ◦ f = χp,s,t (n1 , n2 )fn1 ,n2 , t = 0 : (N − 1). (2.21)
n1=0 n2 =0

Therefore, the tensor transformation χσ is described by the following sys-

tem of binary functions:
χσ = {χp,s,t; Tp,s ∈ σJ , t = 0 : (N − 1)}. (2.22)
The tensor transformation is not orthogonal, but, by means of this transfor-
mation, one can synthesize unitary transformations.
Definition 2.1 For a given frequency-point (p, s) ∈ XN,N and integer t ∈
{0, 1, 2, ..., N/2 − 1}, the function
χ0p,s,t (n1 , n2 ) =χp,s,t (n1 , n2 ) − χp,s,t+N/2 (n1 , n2 )

 1, if (n1 , n2 ) ∈ Vp,s,t
= −1, if (n1 , n2 ) ∈ Vp,s,t+N/2 (2.23)

0, otherwise
(n1 , n2 ) ∈ XN,N ,
is called the paired function.
Example 2.2 We consider the N = 8 case. The paired functions are calcu-
lated by

 1; if n1 p + n2 s = t mod 8
χ0p,s,t (n1 , n2 ) = −1; if n1 p + n2 s = (t + 4) mod 8 (2.24)

0; otherwise
where t = 0 : 3. For instance, the values of the function χ03,1,2(n1 , n2 ) can be
written in the form of the following mask:
 
0 0 1 0 0 0 −1 0
 0 0 0 −1 0 0 0 1 
 
 −1 0 0 0 1 0 0 0 
 
0
 0 1 0 0 0 −1 0 0 
[χ3,1,2] = 
 .
 0 0 −1 0 0 0 1 0 

 0 0 0 1 0 0 0 −1 
 
 1 0 0 0 −1 0 0 0 
0 −1 0 0 0 1 0 0
64 CHAPTER 2: Direction Images
0
The elements fp,s,t of the paired representation of the image fn,m are
defined as
X N−1
N−1 X
0
fp,s,t = fp,s,t − fp,s,t+N/2 = χ0p,s,t ◦ fn,m = χ0p,s,t (n, m)fn,m (2.25)
n=0 m=0

where the time parameter t runs integers from the interval [0, N/2 − 1]. Thus,
the pair of elements of the tensor representation, fp,s,t and fp,s,t+N/2 , defines
0
fp,s,t .
As an example, Figure 2.14 shows the image in part a, along with the
tensor splitting-signal {f1,4,0 , f1,4,1, f1,4,2, f1,4,3, . . . f1,4,255} of length 256 in
0 0 0 0
b, and paired splitting-signal {f1,4,0 , f1,4,1 , f1,4,2 , . . . f1,4,127} of length 128 in
c. The signals are generated by the frequency-point (p, s) = (1, 4).

1.5
6.5
1

6 0.5

5.5 0

−0.5
5
−1
(a) 0 100 200 0 50 100
(b) (c)

FIGURE 2.14
(a) Image 256 × 256, (b) splitting-signal fT1,4 , and (c) splitting-signal fT1,4
0 .

The 2-D paired transform is defined as a set of N 2 such components

0
{fp,s,t ; (p, s, t) ∈ U }, (2.26)
where the set U of triplet-numbers (p, s, t) can be taken as follows:
r−1
[ n o 
U= (p, s, 2nt); (p, s) ∈ 2n JN/2n ,N/2n , t = 0 : (N/2n+1 − 1) ∪ (0, 0, 0) .
n=0
(2.27)

The selection of this set U allows us to consider the 2-D paired transform
as the representation of the 2-D image as the set of complete splitting-signals.
It means the paired transform is written as
r−1
[ [ n o
0 0 n+1
{fp,s,t ; (p, s, t) ∈ U } = fp,s,2 n t ; t = 0 : (N/2 − 1)
n=0 (p,s)∈2n JN/2n ,N/2n

∪ {f0,0,0},
(2.28)
Image Processing 65

and the paired splitting-signals generated by frequencies (p, s) are

0 0 0 0
fTp,s
0 = {fp,s,0 , fp,s,2 n , fp,s,2·2n , . . . , fp,s,N/2−2n }. (2.29)

The higher the frequency (p, s), i.e., 2n = g.c.d.(p, s), the shorter the splitting-
signal, and the time 2n t runs the interval [0, N/2 − 1] faster; with the step
2n . In the collection of the splitting-signals, there are 3N/2 signals of length
N/2 each, 3N/4 signals of length N/4 each, and so on. The total number of
splitting-signals, or generators (p, s) equals 3N − 2.
It follows directly from the definition of the set U, that the set of frequency-
points (p, s) of splitting-signals of the paired transform is defined as
( r−1 )
[
0 n
JN,N = 2 JN/2n ,N/2n ∪ {(0, 0)},
n=0

and its subsets are defined as follows:


JN/2n ,N/2n = {2n (p, 1); p = 0 : (N/2n − 1)}∪ 2n (1, 2s); s = 0 : (N/2n+1 − 1) .
0
The set JN,N is the projection of the set of triplets U on the lattice XN,N .
0
For instance, when N = 8, the set J8,8 equals


 (0, 1), (1, 1), (2, 1), (3, 1), (4, 1), (5, 1), (6, 1), (7, 1), (1, 0), (1, 2), (1, 4), (1, 6)

(0, 2), (2, 2), (4, 2), (6, 2), (2, 0), (2, 4)

 (0, 4), (4, 4), (4, 0)

(0, 0)

Figure 2.15 shows the location (marked by filled circles) of frequency-points

0
of this set, as well as the set J16,16 on the grid X16,16.

7 15

6
5 10
4
3
2 5

1
0 0
0 2 4 6 0 5 10 15

(a) N = 8 (b) N = 16

FIGURE 2.15
Sets of (3N − 2) generators (in red) of paired splitting-signals.

Figure 2.16 shows the tomo image 256 × 256 in part a, along with the
66 CHAPTER 2: Direction Images

350

300 2−D DPT

s
i
g 250
n
a
l 200

n
u 150
m
b
e 100
r

50

128 64 32 16 8 4 ..

signal length

(a) {fn,m } 0
(b) {fp,s,t }

FIGURE 2.16
(a) Image and (b) splitting-signals of lengths 128, 64, 32, 16, 8, 4, 2, 1, 1.

totality of 766 paired splitting-signals of the image in b. The first 384 splitting-
0 ; (p, s) ∈ J256,256} of length 128 each are shown in the form of
signals {fTp,s
the image 384 × 128. The next 192 splitting-signals {fTp,s 0 ; (p, s) ∈ 2J128,128}

of length 64 each are shown in the form of the image 192 × 64, and so on.
The splitting-signals of the same length are united and separated from others
according to the order of the paired transform components, which is given in
the definition of the set U. The whole picture represents the paired transform of
the tomo image. The set of all splitting-signals composing the paired transform
of the image can be packed into the original form, the square matrix 256×256,
as shown in Figure 2.17. All splitting-signals of length 128 are written in three
squares 128 × 128 of the matrix. All splitting-signals of length 64 are written
in three squares 64 × 64 of the remaining square 128 × 128, and so on.

2.3.2 The splitting of the 2-D DFT

The paired splitting-signal defines the 2-D DFT of the image at specific
frequency-points. To show that, we first consider the case when g.c.d.(p, s) = 1,
for instance (1, s), or (p, 1). The splitting-signal is of length N/2,
0 0 0 0
fTp,s
0 = {fp,s,0 , fp,s,1 , fp,s,2 , . . . , fp,s,N/2−1 }.
Image Processing 67

FIGURE 2.17
766 splitting-signals of lengths 128, 64, 32, 16, 8, 4, 2, 1, 1.

If we apply the shifted N/2 -point DFT over this signal,

  N/2−1
X (m+ 21 )t
s s 0
Fm = FN/2 ◦f Tp,s
0 = fp,s,t WN/2 , m = 0 : N/2 − 1,
m
t=0

we obtain the following [40, 41, 46]:

N/2−1
X
mt
s 0
Fm = fp,s,t W t WN/2 = F(2m+1)p,(2m+1)s (2.30)
t=0

for m = 0 : (N/2 − 1). The sum in this equation represents the N/2-point
DFT of the modified splitting-signal
0 0 0 0
gTp,s
0 = {fp,s,0 , fp,s,1 W 1 , fp,s,2 W 2 , . . . , fp,s,N/2−1 W N/2−1 }.

Thus the 2-D DFT of the image fn,m at frequency-points of the subset
n o
0
Tp,s = (p, s), (3p, 3s), , (5p, 5s) . . . , ((N − 1)p, (N − 1)s) (2.31)

of the cyclic group Tp,s is defined by the splitting-signal of length N/2.

As an example, Figure 2.18 shows the tree image of size 256 × 256 in part
a, along with the splitting-signal fT3,1 0 of length 128 together with the real part
of the modified signal fT3,1 0 in b, the 128-point DFT of the modified signal in
c, and the samples of the 2-D DFT of the image at frequency-points of the
0
subset T3,1 on the lattice X256,256, at which the 2-D DFT can be filled by the
1-D DFT of the modified splitting-signal.
If g.c.d.(p, s) = 2n , where n ≥ 0, then sets Vp,s,t = ∅ and components
0
fp,s,t = 0 if t is not divisible by 2n . Therefore, the splitting-signal fTp,s
0 can be
considered as the following signal of length N/2n+1 :
0 0 0 0
fTp,s
0 = {fp,s,0 , fp,s,2 n , fp,s,2·2n , . . . , fp,s,N/2−2n }.
68 CHAPTER 2: Direction Images

20 Re g’T
3,1
10

−10
f’ T
−20 3,1

0 50 100

(a) (b)
700 200
600

500 100

400
0
300

200 200

100 100
0 200 250
0 50 100 0 50 100 150
0
(c) (d)

FIGURE 2.18 (See color insert)

(a) Tree image 256 ×256, (b) the paired splitting-signal fT3,1
0 and the real part
of the modified signal, (c) the 1-D DFT of the modified signal, and (d) the
0
arrangement of values of the 1-D DFT at frequency-points of the subset T3,1 .
(The 1-D DFT is shown in the absolute scale.)

The N/2-point DFT in the right side of (2.30) represents the N/2n+1 -point
DFT of the splitting-signal fTp,s
0 modified by the vector of rotated coefficients
t n+1
{WN/2 n ; t = 0 : (N/2 − 1)} as follows:

N/2n+1 −1
X
mt
0 t
F(2m+1)p,(2m+1)s = fp,s,2 n t WN/2n WN/2n+1 (2.32)
t=0

when m = 0 : (N/2n+1 − 1).

As an example, Figure 2.19 shows the 256 × 256-point DFT of the tomo
image in the form of the 1-D DFTs of the modified paired splitting-signals
in part a. This totality of 1-D DFTs is shown in a form that is similar to
the figure of splitting-signals in Figure 2.16(b). These 1-D DFTs represent
the splitting of the 2-D DFT. The set of 1-D DFTs in the form of the square
matrix 256 × 256 is shown in part b.
Image Processing 69

350

300

D 250
F
T
200
n
u
m 150
b
e
r 100

50

128 64 32 16 8 4..

1−D DFT order

(a) {F(2k+1)p,(2k+1)s } (b) {F(2k+1)p,(2k+1)s }

FIGURE 2.19
(a), (b) 2-D DFT of the image composed of the 1-D DFT of modified splitting-
signals of lengths 128, 64, 32, 16, 8, 4, 2, 1, 1.

As was mentioned above, the projection of frequency-time points of the

3-D set U on the 2-D frequency domain XN,N can be taken as the required
0
set of generators (p, s) of subsets Tp,s covering XN,N . Indeed, the family of
r
2 3 − 2 subsets
  
0 0


σ = T2n p,2n s (p,s)∈J r−n r−n , {(0, 0)} (2.33)
2 ,2 n=0:(r−1)

is the partition of X, which reveals the transformation F2r ,2r . The 2r × 2r -

point DFT is thus split by this partition σ 0 into 2r 3 − 2 short DFTs, namely,
3 · 2r−1 2r−1 -point DFTs, 3 · 2r−2 2r−2 -point DFTs, . . . , and six 2-point DFTs,


 3N/2 N/2-point DFTs



 3N/4 N/4-point DFTs

 3N/8 N/8-point DFTs

N × N -point 2-D DFT → ... ... ... ...



 6 2-point DFTs



 3 1-point DFTs

1 1-point DFTs

The totality of (3N − 2) splitting-signals {fT 0 , T 0 ∈ σ 0 } consists of 2r−1 3

70 CHAPTER 2: Direction Images

signals of length 2r−1 each, 2r−2 3 signals of length 2r−2 each, and so on. The
summary length of all splitting-signals equals N 2 . This set of splitting-signals
is referred to as the σ 0 -representation (or the paired representation) of the
image fn,m .

Example 2.3 (Case 8 × 8) To compose a partition of the lattice X8,8 , we

0
first note that each subset Tp,s is the orbit of the point (p, s) with re-
spect to the movement group G = {1, 3, 5, 7}. For instance, the subset
0
T1,1 = {(1, 1), (3, 3), (5, 5), (7, 7)} is the orbit of the point (1, 1). The point
(1, 1) “is moving” starting (at time t = 0) from itself and “returning” back in
four units of time (at time t = 4) as (1, 1) → (3, 3) → (5, 5) → (7, 7) → (1, 1).
0
In the orbit T3,1 , the point “is moving” as follows: (3, 1) → (1, 3) → (7, 5) →
(5, 7) → (3, 1). The set of all orbits dividing the lattice X8,8 is unique and
given in Table 2.1.

TABLE 2.1
Set of orbits in the grid 8 × 8
0 0 0 0 0 0 0 0 0 0 0 0
T0,1 T1,1 T2,1 T3,1 T4,1 T5,1 T6,1 T7,1 T1,0 T1,2 T1,4 T1,6 4
0 0 0 0 0 0
T0,2 T2,2 T4,2 T6,2 T2,0 T2,4 2
0 0 0
T0,4 T4,4 T4,0 1
0
T0,0 1

The lattice is divided by 22 orbits and the orbits that are shown on the
same row have equal cardinalities, which are shown in the right column of
this table. For instance, the point (0, 1) runs around its orbit for time t = 4,
and the orbit of the point (0, 2) is twice shorter, (0, 2) → (0, 6) → (0, 2).
The set of generators of all orbits is shown in Figure 2.20 in part a, by the
0
filled circles. The locations of frequency-points of orbits Tp,s for generators
(p, s) = (1, 1), (3, 1), and (6, 1) are shown in parts b, c, and d, respectively.

6 6 6 6

4 4 4 4

2 2 2 2

0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
(a) (b) (c) (d)

FIGURE 2.20
(a) 22 generators on the lattice 8 × 8, and the orbits of points (b) (1, 1), (c)
(3, 1), and (d) (6, 1). (The generators and elements of the orbits are shown by
the filled circles.)

The 22 splitting-signals that are defined for the generators of subsets, or or-
bits T 0 of the partition σ 0 determine values of the 2-D DFT at frequency-points
of these subsets. For instance, the signal fT1,10 carries the spectral information
Image Processing 71

of the image at frequency-points (1, 1), (3, 3), (5, 5), and (7, 7). According to
Table 2.1, the 8 × 8-point DFT is split by twelve 4-point DFTs, six 2-point
DFTs, and four 1-point DFTs (which are the identity transformations).
This table also shows how to compose the complete set of basis paired
functions. The 64 triplets (p, s, t), or numbers of the complete system of paired
functions χ0p,s,t are composed from the generators (p, s) and time variable t,
which runs numbers 0, 1, 2, 3, for the orbits in the first row in the table. For
short orbits in the second row, t runs only values of 0 and 2. Indeed, both
coordinates of points on these orbits are even; therefore conditions (n1 p+n2 s =
t mod 8) and (n1 p + n2 s = t + 4 mod 8) do not hold for odd t = 1 and 3.
In other words, for these triples (p, s, t), we have χ0p,s,t ≡ 0, and we do not
consider such functions. There is no movement (or t = 0) in single-point orbits
for (p, s) = (0, 4), (4, 4), (4, 0), and (0, 0). The set of 64 triplet-numbers for the
complete system of paired functions is given in Table 2.2.

TABLE 2.2
Set of triplets (p, s, t)
(p, s) t
(0, 1), (1, 1), (2, 1), (3, 1), (4, 1), (5, 1), (6, 1), (7, 1), (1, 0), (1, 2), (1, 4), (1, 6) 0, 1, 2, 3
(0, 2), (2, 2), (4, 2), (6, 2), (2, 0), (2, 4) 0, 2
(0, 4), (4, 4), (4, 0) 0
(0, 0) 0

Figure 2.21 shows the first 36 basis paired functions χ0p,s,t , which are placed
in the order given in the above table, starting from the top on the left to
the bottom on the right. For instance, the first four functions have numbers
(0, 1, t), where t = 0 : 3. The filled circle is used for value 1 and the open
circle for −1. It is not difficult to see that all these functions are orthogonal.
The complete set of paired functions defines the paired representation of the
image,
0 0
f → {fTp,s
0 ; T
p,s ∈ σ }. (2.34)
The components of the paired splitting-signals fTp,s0 are calculated from the
components of the splitting-signals fTp,s of the tensor representation of f by
0
fp,s,t = fp,s,t − fp,s,t+4 , t ∈ {0, 1, 2, 3}.
The transformation
0
χ0 : f → {fp,s,2 0 0
n t ; Tp,s ∈ σ , 2
n 0
= g.c.d.(p, s), t = 0 : (card(Tp,s ) − 1)} (2.35)

is called the paired transformation. Thus the paired transform is the repre-
sentation of the image in the form of 1-D splitting-signals; components of all
splitting-signals together compose the 2-D discrete paired transform.
72 CHAPTER 2: Direction Images

FIGURE 2.21
The first 36 basic functions of the complete system of the 8 × 8-paired trans-
formation. (Values of 1 are shown by the filled circles, and −1 by open circles.)

Let (p, s) = (3, 1) and let f be the following image of size 8 × 8 :

 
12131213
2 0 1 2 2 4 2 1 
 
1 3 2 2 1 1 1 2 
 
4 1 0 1 3 1 3 1 
f =  .
2 4 1 2 1 3 2 2 

2 4 1 2 1 2 2 1 
 
2 4 1 2 1 5 2 1 
24121132

We will calculate the 2-D DFT of this image at frequency-points of the

orbit of the point (3, 1). For that, we first write all values of t in the equations
Image Processing 73

3n1 + n2 = t mod 8 in the form of the table

 
0 1 2 3 4 5 6 7
3 4 5 6 7 0 1 2
 
6 7 0 1 2 3 4 5
 
1 2 3 4 5 6 7 0
||t = (3n1 + n2 ) mod 8||n2,n1 =0:7 =
4
.
 5 6 7 0 1 2 3
7 0 1 2 3 4 5 6
 
2 3 4 5 6 7 0 1
5 6 7 0 1 2 3 4

Then, the components of the splitting-signal fT3,1

0 are defined as follows:
0

f3,1,0 =(f0,0 + f1,5 + f2,2 + f3,7 + f4,4 + f5,1 + f6,6 + f7,3 )−



 (f0,4 + f1,1 + f2,6 + f3,3 + f4,0 + f5,5 + f6,2 + f7,7 )



 =(1 + 4 + 2 + 1 + 1 + 4 + 2 + 2)−



 −(1 + 0 + 1 + 1 + 2 + 2 + 1 + 2) = 7,
 0


f3,1,1 =(f0,1 + f1,6 + f2,3 + f3,0 + f4,5 + f5,2 + f6,7 + f7,4 )−



 (f0,5 + f1,2 + f2,7 + f3,4 + f4,1 + f5,6 + f6,3 + f7,0 )



 =(2 + 2 + 2 + 4 + 3 + 1 + 1 + 1)−

−(2 + 1 + 2 + 3 + 4 + 2 + 2 + 2) = −2,
fT3,1
0 = 0

f3,1,2 =(f0,2 + f1,7 + f2,4 + f3,1 + f4,6 + f5,3 + f6,0 + f7,5 )−



 (f0,6 + f1,3 + f2,0 + f3,5 + f4,2 + f5,7 + f6,4 + f7,1 )



 =(1 + 1 + 1 + 1 + 2 + 2 + 2 + 1)−



 −(1 + 2 + 1 + 1 + 1 + 1 + 1 + 4) = −1,
 0


f3,1,3 =(f0,3 + f1,0 + f2,5 + f3,2 + f4,7 + f5,4 + f6,1 + f7,6 )−



 (f0,7 + f1,4 + f2,1 + f3,6 + f4,3 + f5,0 + f6,5 + f7,2 )



 =(3 + 2 + 1 + 0 + 2 + 1 + 4 + 3)−

−(3 + 2 + 3 + 3 + 2 + 2 + 5 + 1) = −5.

0 W, where
This splitting-signal is modified as fT3,1

W = diag{1, e−jπ/4 , e−j2π/4 , e−j3π/4} = diag{1, a(1 − j), −j, −a(1 + j)}
√
and a = 2/2 = 0.7071. The four-point DFT of the modified splitting-signal

gT3,1
0 0 W = {7, −1.4142 + j1.4142, j, 3.5355 + j3.5355}
= fT3,1

is calculated as follows:
     
1 1 1 1 1 0 0 0 7 9.1213 + j5.9497
 1 −j −1 j  0 a(1 − j) 0 0  −2   4.8787 + j3.9497
  = 
 −1   4.8787 − j3.9497 .
  
 1 −1 1 −1  0 0 −j 0
1 j −1 −j 0 0 0 −a(1 + j) −5 9.1213 − j5.9497

The obtained four-point DFT coincides with the 8×8-point DFT of the image
74 CHAPTER 2: Direction Images
0
f at four frequency-points of the orbit T3,1 , i.e.
   
F3,1 9.1213 + j5.9497
 F1,3   4.8787 + j3.9497
 F7,5  =  4.8787 − j3.9497 .
   

F5,7 9.1213 − j5.9497

One can notice from this example, that the pairs of complex conjugate com-
ponents F3,1 and F5,7 , F1,3 and F7,5 , are obtained from the same splitting-
0
signal. In general, if (p1 , p2 ) is on the orbit Tp,s , then the frequency-point
(N − p1 , N − p2 ) is also on this orbit.
Twelve subsets T 0 of the partition σ 0 of the lattice X8,8 consist of four
points each, six subsets of two points each, and the remaining four subsets are
one-point sets. The splitting of the 8 × 8-point DFT thus equals
n o
R(F8,8 ; σ 0 ) = F4 , . . . , F4 , F2 , . . . , F2 , F1 , . . . , F1 . (2.36)
| {z } | {z } | {z }
12 times 6 times 4 times

The first twelve four-point DFTs are calculated over the modified splitting-
signals
0 0 0 0
gTp,s
0 ={fp,s,0 , fp,s,1 W, fp,s,2 W 2 , fp,s,3 W 3}
0 0 0 0
(2.37)
={fp,s,0 , a(1 − j)fp,s,1 , −jfp,s,2 , −a(1 + j)fp,s,3 }
√
where a = 2/2 = 0.7071 and W = exp(−j2π/8) = a(1 − j). The generators
(p, s) of these signals lie on the first row in Table 2.2. The calculation in (2.37)
uses two operations of multiplication by the factor of a, for each splitting-
signal. The next six two-point DFTs are calculated over the short modified
splitting-signals
0 0 0 0
gTp,s
0 = {fp,s,0 , fp,s,2 W 2 } = {fp,s,0 , −jfp,s,2 }.

The generators (p, s) of these signals are given in the second row in Table 2.2.
Thus, the total number of operations of multiplication for calculating the
8 × 8-point DFT through the splitting (2.36) equals m08,8 = 12 · 2 = 24.

Example 2.4 (Case 16 × 16) The splitting of the 16 × 16-point DFT by the
partition σ 0 = (T 0 ) equals
n o
R(F16,16; σ 0 ) = F8 , . . . , F8 , F4 , . . . , F4, F2 , . . . , F2 , F1 , . . . , F1 .
| {z } | {z } | {z } | {z }
24 times 12 times 6 times 4 times

0
The cardinalities of the largest subsets Tp,s of the partition σ 0 of the lattice
X16,16 equal 8. The generators for these subsets are taken from the set J16,16,
and the corresponding modified splitting-signals have the following form:
0 0 0 2 0 3 0 7
0 W = {f
gTp,s
0 = fTp,s p,s,0 , fp,s,1 W, fp,s,2 W , fp,s,3 W , ..., fp,s,7W } (2.38)
Image Processing 75

where W = exp(−j2π/16) = 0.9239 − j0.3827. This product uses six non-

trivial operations of multiplication by the twiddle factors W t , t = 1, 2, 3, 5, 6, 7.
The eight-point DFT requires two operations of multiplication. Therefore
the calculation of 24 eight-point DFTs in the splitting of the 16 × 16-point
DFT requires 24 · 8 operations of multiplication. The calculation of the four-
point DFTs over the modified splitting-signals of length four was described
in Example 2.3. Each such transform uses two operations of multiplication.
The total number of operations of multiplication required to calculate the
16 × 16-point DFT by the paired transform is calculated by

m016,16 = 24(m8 + 8 − 2) + 12(m4 + 4 − 2) = 24(2 + 6) + 12(2) = 192 + 24 = 216.

We note for comparison, that the tensor algorithm of the 16 × 16-point DFT
uses 24m16 = 24 · 10 = 240 operations of multiplication.

In the general N = 2r case, the number of operations of multiplication

required for calculating the 2r × 2r -point DFT by the paired transforms can
be estimated as
r−2
X

m 2r ,2r = 2r−n 3 m2r−n + 2r−n − 2 ≤ 2 · 4r−1 (r − 7/3) + 8/3, (r > 2).
n=1

When estimating this number, we use the fact that the 2n -point fast paired
DFT requires m2n = 2n−1(n − 3) + 2 multiplications, when n ≥ 3 [41, 45].

2.4 Complete system of 2-D paired functions

The paired functions compose a basis in the linear space of discrete images of
size N × N, and the system of functions

χ0 = {χ0p,s,t ; (p, s, t) ∈ U }

is complete [42]. We can construct other such sets of triples, U. For instance,
we can substitute all or a part of triplets (p, s, t) by (s, p, t).
The 2-D discrete paired transform does not require multiplications. The
complete system of paired functions is defined as

 1,if np + ms = 2k t mod N,
0
χp,s,2k t (n, m) = −1, if np + ms = (2k t + N/2) mod N, (2.39)

0,otherwise,
which is the difference of two characteristic functions χp,s,2k t and χp,s,2k t+N/2 .
When (p, s) = (0, 0), the paired function is defined as χ00,0,0 (n, m) ≡ 1. The
paired function χ0p,s,2k t itself represents a 2-D plane wave, and the decompo-
sition of the image by the paired functions is the decomposition of the image
76 CHAPTER 2: Direction Images

by plane waves. All basic paired functions are orthogonal. In the masks of
paired functions, all “1”s and all “−1”s lie on the different but parallel lines,
the directions of which are defined by (p, s). Inside each series of functions
generated by frequency (p, s),

{χ0p,s,2k t ; t = 0, 1, 2, ..., N/2k+1 − 1, (2k = g.c.d.(p, s))},

the paired functions have the same frequency, direction, and they are parallel
shifted functions with respect to each other. As an example, Figure 2.22 shows,
for the N = 32 case, the gray-scale images of basic paired functions χ01,4,1 and
χ01,4,6 in parts a and b, and functions χ02,2,6 and χ02,2,12 in c and d, respectively.

BPF1,4,1(n,m) BPF1,4,6(n,m) BPF2,2,6(n,m) BPF2,2,12(n,m)

30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
(a) (b) (c) (d)

FIGURE 2.22
Images 32 × 32 of the paired functions (a) χ01,4,1, (b) χ01,4,6 , (c) χ02,2,6 , and (d)
χ02,2,12. (Values 1, −1, and 0 correspond to the white, black, and gray levels,
respectively.)

The paired function can be considered as the quantized 2-D cosine wave
  
0
  2π(np + ms − 2k t)
χp,s,2k t (n, m) = M cp,s,2k t (n, m) = M cos ,
2r
(2.40)
where the quantization operation, M, is defined as M [1] = 1, M [−1] = −1,
and M [x] = 0, if x 6= ±1.
For the N = 256 case, Figure 2.23 shows the gray-scale images of the 2-D
cosine wave with number (4, 8, 8) and its quantized function (or the paired
function) in parts a and b, respectively.
Similar to paired functions, inside each series of functions generated by
frequency (p, s) the cosine waves are shifted in parallel and defined by the
same direction. Figure 2.24 shows the images of the 2-D cosine waves with
numbers (1, 4, 1), (1, 8, 1), and (12, 4, 4) in parts a, b, and c, respectively. Four
periods of the wave c1,4,1 cover the discrete lattice X256,256; the wave c1,8,1
covers it with eight periods, and the wave c12,4,4 with twelve periods. Thus,
the length of each wave depends on the frequency (p, s) that generates this
Image Processing 77

cosine wave (4,8,8) paired basic wave (4,8,8)

50 50

100 100

150 150

200 200

250 250
50 100 150 200 250 50 100 150 200 250

(a) (b)

FIGURE 2.23
The images of (a) the 2-D cosine wave c4,8,8(n, m) and (b) the paired function
χ04,8,8(n, m).

cosine wave (1,4,1) cosine wave (1,8,1) cosine wave (12,4,4)

(a) (b) (c)

FIGURE 2.24
The images 256 × 256 of the 2-D cosine waves (a) c1,4,1, (b) c1,8,1, and (c)
c12,4,4.

wave, and the wave is oriented in the 3-D space. It covers the period 256 × 256
in the direction that is also defined by the frequency.
In addition we observe the following property of the multiresolution of
the image, which is defined along directions. We call this property the directed
multiresolution of the image by the paired transform. The paired functions ge-
nerated by the frequencies (p, s) and (2p, 2s), as well as (4p, 4s), (8p, 8s), . . .,
are defined by the same direction. Indeed, the angles along which the co-
efficients “1” and “−1” are located in the masks equal to the angle ϑ =
arctg(p/s) = arctg(2p/2s) = arctg(4p/4s). The paired function χ02p,2s,2t can
be calculated by

χ02p,2s,2t = χ2p,2s,2t−χ2p,2s,2t+N/2 = χp,s,t −χp,s,t+N/4 +χp,s,t+N/2 −χp,s,t+3N/4 ,

78 CHAPTER 2: Direction Images

and the components of the paired transform can be calculated through the
tensor transform as follows:
0 0
f2p,2s,2t = f2p,2s,2t = fp,s,t − fp,s,t+N/4 + fp,s,t+N/2 − fp,s,t+3N/4 .

Indeed, the following relation holds for tensor characteristic functions:

χ2p,2s,2t = χ2p,2s,2t = χp,s,t + χp,s,t+N/2 , t = 0 : (N/2 − 1),

where subscripts are considered modulo N. Similarly, the components of the

paired splitting-signals generated by frequencies (4p, 4s), (8p, 8s), ... can be
calculated from the same splitting-signal {fp,s,t ; t = 0 : (N − 1)}. Although
the complete family of paired functions are generated by (3N − 2) frequencies
(p, s), all functions can be calculated from 3N/2 characteristic functions of the
tensor transformation. In other words, all N 2 paired functions are determined
by directions at 3N/2 angles of the set

ΨN = {arctan(p); p = 0 : (N − 1)} ∪ {arctan(1/(2s)); s = 0 : (N/2 − 1)}.

As an example, Figure 2.25 shows the images of the 2-D cosine waves with
0
numbers (1, 2, 1), (2, 4, 2), (4, 8, 4), and (8, 16, 8) from J256,256 in parts a-d,
respectively. These waves have different frequencies, but they are oriented at

CW (1,2,1) CW (2,4,2) CW (4,8,4) CW (8,16,8)

(a) (b) (c) (d)

FIGURE 2.25
The images 256 × 256 of the 2-D cosine waves (a) c1,2,1 , (b) c2,4,2 , (c) c4,8,4,
and (d) c8,16,8.

the same angle. There are three other series of cosine waves with frequency-
numbers (16, 32), (32, 64), and (64, 128), which are defined by the direction at
this angle. Thus, seven series of the basic paired functions have the same orien-
tation, but define different resolutions. The triplet-numbers of these functions
and the number of parallel shifts of functions inside each series are given in
Table 2.3. The total number of these paired functions equals 254. In all triplet-
numbers of the above functions, the first two components are integer multiple
to (p0 , s0 ) = (1, 2) in modulo N. Therefore, we call this set of seven series of
cosine functions (and paired functions) the (1, 2)-generated set of functions.
We now consider the (1, 5)-generated set of functions. Figure 2.26 shows the
Image Processing 79
TABLE 2.3
Numbers of basic paired functions defined along the direction at angle arctg(1/2)
(p, s, t) (1, 2, t) (2, 4, 2t) (4, 8, 4t) (8, 16, 8t) (16, 32, 16t) (32, 64, 32t) (64, 128, 64t)
t 0 : 127 0 : 63 0 : 31 0 : 15 0:7 0:3 0, 1
# 128 64 32 16 8 4 2
shift by 1 2 4 8 16 32 64

CW (1,5,1) CW (2,10,2) CW (4,20,4) CW (8,40,8)

(a) (b) (c) (d)

FIGURE 2.26
The images 256 × 256 of the 2-D cosine waves (a) c1,5,1, (b) c2,10,2, (c) c4,20,4,
and (d) c8,40,8.

images for the 2-D cosine waves with numbers (1, 5, 1), (2, 10, 2), (4, 20, 4),
and (8, 40, 8) in parts a-d, respectively. The other two series of cosine waves
with frequency-numbers (16, 80) and (32, 160) are also defined by the same
direction. The triplet-numbers of these functions and the number of parallel
shifts of functions inside each series of the (1, 5)-generated set are given in
Table 2.4. In the general case of (p0 , s0 ) ∈ JN,N , the (p0 , s0 )-generated set of

TABLE 2.4
Numbers of basic paired functions defined along the direction at angle arctg(1/5)
(p, s, t) (1, 5, t) (2, 10, 2t) (4, 20, 4t) (8, 40, 8t) (16, 80, 16t) (32, 160, 32t) (64, 64, 64t)
t 0 : 127 0 : 63 0 : 31 0 : 15 0:7 0:3 0, 1
# 128 64 32 16 8 4 2
shift by 1 2 4 8 16 32 64

functions is defined similarly.

It should be noted that the series of functions with high frequencies 2k (p, s)
have very dense structures with small periods N/2k ×N/2k . Such functions de-
scribe high-resolution components of the image and they can be determined by
different directions. As an example, we consider the cosine wave with triplet-
number (8, 40, 8) = 8 · (1, 5, 1) from Table 2.4. This wave has the period
32 × 32 and is directionally defined at angle arctan(1/5). The same triplet
can be defined as (8, 40, 8) = 8 · (1, 5 + 32n, 1) in modulo 256, for any in-
teger n. We consider, for instance, the cases when n = 3, 4, and 5, i.e., the
triplets (1, 101, 1), (1, 133, 1), and (1, 165, 1). The corresponding frequency-
points (1, 101), (1, 133), and (1, 165) are from the set of generators J256,256.
80 CHAPTER 2: Direction Images

The cosine wave with number (8, 40, 8) can thus be referred to the (1, 101)-
oriented set of functions, as well as to the (1, 133)- and (1, 165)-oriented sets.
Directions of these three sets are determined by the angles arctan(1/101),
arctan(1/133), and arctan(1/165), respectively.

2.4.0.1 Code: System of basic paired functions

The complete set of basic paired functions can be calculated by the program
with script matrices 2Dpaired.m. It prints all N 2 matrices of paired functions.
% -------------------------------------------------
% call: matrices_2Dpaired.m
N=input(’ Set the value of N (power of 2) = ’);
r=log2(N); L2=N/2; p=1;
for nn=1:r
for s=0:p:N-1
for t=0:p:L2-1
fprintf(’ #(%g,%g,%g) ’,p,s,t);
A=pmatrix_pst(N,p,s,t)
end
end
s=p; step=2*s;
for p=0:step:N-1
for t=0:s:L2-1
fprintf(’ #(%g,%g,%g) ’,p,s,t);
A=pmatrix_pst(N,p,s,t)
end
end
p=2*s;
end
Alast=pmatrix_pst(N,0,0,0)
% -------------------------------------------------
% Call: pmatrix_pst.m
function A=pmatrix_pst(N,p,s,t)
if (t>N) | (t<0) t=mod(t,N); end
A=zeros(N,N);
ms=0; t2=t+N/2;
for m=1:N
np=0;
for n=1:N
t1=mod(np+ms,N);
if (t1==t) A(m,n)=1;
elseif (t1==t2) A(m,n)=-1;
end
np=np+p;
end
ms=ms+s;
end

The function pmatrix pst.m calculates one matrix with triplet (p, s, t).
Image Processing 81

2.4.1 1-D DFT and paired transform

The 2-D discrete paired transformation is not separable, but its basic functions
can be calculated by means of the convolution from the basic functions of the
1-D paired transform [46, 48]. The 1-D paired transform reveals the structure
of the 1-D DFT in a way similar to the 2-D case, and this transform can also
be defined from the cosine transform.
Given N = 2r , r > 1, the 1-D N -point discrete paired transform χ0N is
determined by the following complete system of the paired functions [41, 46]:
  
2π(n − t)
χ02k ,2k t (n) = M cos , t = 0 : (2r−k−1 − 1), k = 0 : (r − 1),
2r−k
χ00,0 (n) ≡ 1, n = 0 : (N − 1),
(2.41)
where M is the real function which differs from zero only on the bounds of
the interval [−1, 1] and takes values M (−1) = −1, M (1) = 1. The double
numbering of the paired functions refers to the frequency (p = 2k ) and time
(t). The paired transform is a transform of the discrete-time signal fn to the
set of frequency-time signals,
n o
fn → {f20 k ,0 , f20 k ,2k , f20 k ,2k 2 , ..., f20 k,N/2k+1 −1 }, k = 0 : (r − 1) ,

that splits the 2r -point DFT by the 2r−k−1 -point DFTs, k = 0 : (r − 1). The
components of these splitting-signals are calculated by
r
2X −1
f20 k ,2k t = χ02k ,2k t ◦ fn = χ02k ,2k t (n)fn , t = 0 : (2r−k−1 − 1).
n=0

Example 2.5 (Case N = 8) The matrix of the eight-point discrete paired

transform (DPT) is defined as
 0   
[χ1,0 ] 1 0 0 0 −1 0 0 0
[χ01,1] 
0 1 0 0 0 −1 0 0

 0 
[χ1,2 ] 0 0 1 0 0 0 −1 0
 0   
0
[χ1,3 ] 0 0 0 1 0 0 0−1
[χ8 ] =  0  = 
   .
[χ2,0 ] 1 0−1 0 1 0 −1 0 
[χ0 
0 1 0 −1 0 1 0−1
]  
 2,2
[χ0  ] 1 −1 1 −1 1 −1 1−1
4,0
[χ00,0 ] 1 1 1 1 1 1 1 1

The first four basis paired functions correspond to the frequency p = 1, the
next two functions correspond to frequency p = 2, and the last two functions
correspond to the frequencies p = 4 and 0, respectively. The process of com-
position of these functions from the corresponding cosine waves defined in the
interval [0, 7] is illustrated in Figure 2.27.
82 CHAPTER 2: Direction Images

1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
1 1
0 0
−1 −1
0 2 4 6 0 2 4 6
(a) (b)

FIGURE 2.27
(a) Cosine waves and (b) discrete paired functions of the eight-point DPT.

Let fn be the signal {1, 2, 2, 4, 5, 3, 1, 3}. The paired transform of this sig-
nals results in four splitting-signals as follows:
 0 0 0 0
{f , f1,1 , f1,2 , f1,3 } = {−4, −1, 1, 1},
 1,0

 0 0
{f2,0 , f2,2 } = {3, −2},
{fn } →
{f 0 } = {−3},
 4,0


0
{f0,0 } = {21}.

The splitting of the 2r -point DFT into (r + 1) short DFTs is described by

2r−k−1
X −1
0


F (2m+1)2k = fp,t W2tr−k W2mt
r−k−1 , m = 0 : (2r−k−1 − 1). (2.42)
t=0

The set of 2r frequency-points X2r = {0, 1, 2, ..., 2r − 1} is divided by (r + 1)

subsets, or orbits

Tp0 = {(2m + 1)p mod 2r ; m = 0 : (2r−1 /p − 1)},

where p = 2k , k = 0 : (r −1), and T00 = {0}. These subsets compose a partition

Image Processing 83

σ 0 of X2r , and the 2r -point DFT is split as follows:

R(F2r ; σ 0 ) = {F2r−1 , F2r−2 , F2r−3 , ..., F2, 1, 1} . (2.43)

Using the similar splitting for each short transform F2r−k−1 of this splitting, we
obtain the full decomposition of the 2r -point DFT, by the paired transforms.

Example 2.6 (8-point DFT) Let N = 8 and let {fn } be the signal of
Example 2.5. According to (2.42), the calculation of the eight-point DFT
of fn can be written in matrix form as

2 3 8 9 2 3 2 3
F7 2 8 9 3 1 > 1 −5.4142 + j
6 F3 7
2 3 1 >
> >
6 2 7 6 −2.5858 + j 7
6 7 6 [F2] W>
>
> > >
< > > >
5diag −j
=h 0i 7 >
> >
> 6 7 6 7
6 F5 7 64 1 χ −j 6 2 7 6 −2.5858 − j 7
> >
4
7 > >
1
> >
3
6 7 6 7 > > 6 7 6 7
6 F1 7 6 1 6 −5.4142 − j 7
7diag W
0 64 7
>
> >
> 7 < =h i
6 7=6 1
: ;
χ 8 6 7 6
6 7 =6 7
6 5 7 6 3 − 2j 7
6 F6 7 6 7 1 > 7
[F ]
>
2
6 7 6 7 > >
6 F2 7 6 −j > 6 3 7 6 3 + 2j 7
7 >
> >
> 6 7 6 7
1 5
6 7 4 >
> >
4 F4 5 1 41 5 4 −3
>
> >
> 5
1
>
> >
: >
F0 1 3 21
;

where the twiddle factors W = exp(−2πj/8) = 0.7071(1 − j) and W 3 =

−0.7071(1 + j).
In the general N = 2r case, where r ≥ 2, the following matrix decomposi-
tion holds for the 1-D DFT [45, 48, 51]:
r−1
hM   i
[F2r ] = F2r−k−1 ⊕ 1 D2r [χ2r ], (2.44)
k=0

where ⊕ denotes the operation of the Kronecker sum of matrices and the
diagonal matrix
r r
D2r = diag{1, W, W 2, W 3 , ..., W 2 −1 , 1, W 2 , W 4 , ..., W 2 −2
,
r
1, W 4 , W 8 , ..., W 2 −4 , 1, ..., 1, 1}.

2.5 Paired transform direction images

In this section, we describe the principle of superposition, which states that
the image {fn,m } can be composed with the direction images, which are deter-
mined by splitting-signals in paired representation. Let (p, s) be the generator
0
from the set JN,N and let 2k = g.c.d.(p, s), where k ≥ 0. We denote by Dp1 ,s1
the incomplete 2-D DFT of the image, which is defined only at frequency-
0
points of the subset Tp,s , i.e.,
 0
Fp1 ,s1 ;if (p1 , s1 ) ∈ Tp,s ,
Dp1 ,s1 = Dp(p,s)
1 ,s1
= (2.45)
0; otherwise.
84 CHAPTER 2: Direction Images

We start with the first series of generators, when g.c.d.(p, s) = 1. There are
3N/2 such generators that compose the subset JN,N . The inverse transform
of the above incomplete 2-D DFT can be calculated as follows:
N−1 N−1
1 X X
dn,m = Dp ,s W −(np1+ms1 )
N 2 p =0 s =0 1 1
1 1

1 X
= 2 Fp1 ,s1 W −(np1 +ms1 )
N
(p1 ,s1 )∈Tp,s
0

N/2−1
1 X
= F(2k+1)p,(2k+1)sW −(n(2k+1)p+m(2k+1)s)
N2
k=0
N/2−1
1 X
= F(2k+1)p,(2k+1)sW −(2k+1)(np+ms) (2.46)
N2
k=0
 
N/2−1
1 2 X −kt 
= F(2k+1)p,(2k+1)sWN/2 W −t
2N N
k=0
1 0

1 0
= fp,s,t W t W −t = f ,
2N 2N p,s,(np+ms) mod N
where we denote t = (np + ms) mod N.
Thus we obtain the following direction image of the 1st series:
1 0
dn,m = d(p,s)
n,m = f , n, m = 0 : (N − 1).
2N p,s,(np+ms) mod N
The direction image N × N is composed of N/2 values of the splitting-signal
0 , which are placed on the image along the set of parallel lines np + ms =
fTp,s
t mod N, t = 0 : (N − 1). As an example, Figure 2.28 shows for the bridge
image 256 × 256 the first nine direction images defined by frequencies (p, s) =
(0, 1), (1, 1), . . . , (8, 1) in parts a-i, respectively.
According to the definition of the paired representation, the following pro-
0 0
perty holds for its components: fp,s,t+N/2 = −fp,s,t when t = 0 : (N/2 − 1).
0
When (np + ms) mod N ≥ N/2, the values of fp,s,t are placed on the image
with the minus sign. The composition of the direction image from the splitting-
signal is simple. All values of this image can be calculated, for instance, by
the following method of permutation of values of the splitting-signal. Let us
assume that a value of the direction image at point (n, m) has been calculated.
Then the next three neighbor values of the image can be calculated as follows:
 0
fp,s,t+p ; if t + p < N/2,
dn+1,m =
−f 0 ; otherwise,
 p,s,t+p
0
0 f p,s,t+s if t + s < N/2,
;
dn,m = fp,s,t → dn,m+1 = 0 (2.47)
−fp,s,t+s ; otherwise,
 0
fp,s,t+p+s ; if t + p + s < N/2,
dn+1,m+1 = 0
−fp,s,t+p+s ; otherwise,
Image Processing 85

(a) DI−(0,1) (b) DI−(1,1) (c) DI−(2,1)

(d) DI−(3,1) (e) DI−(4,1) (f) DI−(5,1)

(g) DI−(6,1) (h) DI−(7,1) (i) DI−(8,1)

FIGURE 2.28
Nine first direction image components of the bridge image. (All images have
been scaled.)

where we omitted the factor 1/(2N ). Thus, the direction image can be con-
structed right after the corresponding projection data at the angle arctg(p/s)
have been obtained.
We now consider the (k + 1)th series of generators, i.e., such that
g.c.d.(p, s) = 2k , when k ∈ {1, 2, ..., r − 1}. The number of such generators in
0
the set JN,N equals 3N/2k+1 . The calculation of the inverse transform of the
86 CHAPTER 2: Direction Images

incomplete 2-D DFT, Dp1 ,s1 , results in the following direction image:
N−1 N−1
1 X X
dn,m = d(p,s)
n,m = Dp ,s W −(np1 +ms1 )
N 2 p =0 s =0 1 1
1 1

N/2k+1 −1
1 X
= F(2l+1)p,(2l+1)s W −(2l+1)(np+ms)
N2
l=0
 k+1

k+1 N/2X −1
1 2 −lt  W −t k
= k+1  F(2l+1)p,(2l+1)s WN/2 k+1 N/2
2 N N
l=0
1  0 t

−t
= k+1 fp,s,2k t WN/2 k WN/2 k
2 N
1 0
= k+1 fp,s,(np+ms) mod N , (2.48)
2 N
where we denoted 2k t = (np + ms) mod N.
For the bridge image, Figure 2.29 shows the first nine direction images of
the 2nd series, for the generators (p, s) = (0, 2), (2, 2), . . . , (16, 2) in parts a-i,
respectively.
The last series of generators contains only the frequency (0, 0) and the set
0
T0,0 = {(0, 0)}. The transform Dp1 ,s1 is zero at all frequency-points except
the point (0, 0), where the transform equals F0,0 . In this case, we have the
following constant-image:
1 1 0
dn,m = d(0,0)
n,m = 2
F0,0 = 2 f0,0,0 .
N N
0 0
All (3N − 2) subsets Tp,s , with generators (p, s) ∈ JN,N , compose a par-
tition of the lattice N × N. It means the sum of corresponding (3N − 2)
incomplete 2-D DFTs equals the 2-D DFT of the image. It also means the
(p,s)
sum of (3N − 2) direction images dn,m equals the image fn,m .
Statement 2: (Superposition by paired direction images) The disc-
rete image of size N × N, where N = 2r , (r > 1), can be composed from
(3N − 2) direction images by
X
fn,m = d(p,s)
n,m
(p,s)∈JN,N
0

r−1
1 X 1 X
0 1 0 (2.49)
= fp,s,(np+ms) mod N + f ,
2N 2k N 2 0,0,0
k=0 (p,s)∈J2r−k ,2r−k

n, m = 0 : (N − 1).

This is the formula of composition of the image by direction images and the
formula of reconstruction of the image from its 2-D discrete paired transform.
Image Processing 87

(a) DI−(0,2) (b) DI−(2,2) (c) DI−(4,2)

(d) DI−(6,2) (e) DI−(8,2) (f) DI−(10,2)

(g) DI−(12,2) (h) DI−(14,2) (i) DI−(16,2)

FIGURE 2.29
Nine direction image components of the bridge image. (All images have been
scaled.)

2.6 L-paired representation of the image

Consider the case when N = Lr and L > 1 is odd prime. The paired functions
are defined similarly to the L = 2 case. The components of the tensor trans-
form are united with the twiddle factors, which divide the unit circle into L
parts. Below is the definition of the paired transform in the general case of
N × N.
Definition 2.2 Let L be a non-trivial divisor of the number N and let
WL = exp(−2πi/L). Given frequency-point (p, s) ∈ XN,N and integer
t ∈ {0, 1, ..., N − 1}, the function
L−1
X
χ0p,s,t (n, m) = χ0p,s,t;L (n, m) = WLk χp,s,t+kN/L (n, m) (2.50)
k=0
88 CHAPTER 2: Direction Images

is called the L-paired function.

The L-paired function, or simply the paired-function is defined thus as


 1, if np + ms = t mod N



 W L , if np + ms = (t + N/L) mod N
 2
W L , if np + ms = (t + 2N/L) mod N
χ0p,s,t (n, m) = (2.51)

 · · · , ···
 L−1

 W , if np + ms = (t + (L − 1)N/L) mod N
 L

0, otherwise.
0
Given triplet (p, s, t), we denote by fp,s,t the coefficient of the representa-
tion of the image f = {fn,m } by the L-paired function χ0p,s,t
X N−1
N−1 X
0
fp,s,t = χ0p,s,t ◦ f = χ0p,s,t (n, m)fn,m
n=0 m=0
L−1
(2.52)
X
= WLk fp,s,t+kN/L .
k=0

We now consider the N = Lr case, when L > 2 is prime and r > 1. As

in the N = 2r case, we construct a partition σ 0 = (T 0 ) of the lattice XN,N ,
which is composed of the following subsets:
n  o
0
Tp,s;L = (mL + 1)p, (mL + 1)s ; m = 0 : (N/L − 1) . (2.53)

To construct such a partition σ 0 , we first consider the subset of generators

(p, s),
r r−1
L[ −1 [ L [−1
J = JLr ,Lr = (1, s) (Lp, 1), (2.54)
s=0 p=0
which is used for the tensor representation of the image.
Theorem 2.1 For an arbitrary prime number L > 1 and r > 1, the following
is valid:
1) The totality of (L + 1)(Lr − 1) + 1 subsets
    
σ0 = 0
TjL n p,jLn s;L , {(0, 0)}
(p,s)∈JLr−n ,Lr−n j=1:(L−1) n=0:(r−1)

is the partition of the lattice XLr ,Lr .

2) The system of the L-paired functions
χ0 = {χ0p,s,t ; (p, s, t) ∈ U }
is the complete set of functions, where the set U is defined as
[ L−1
r−1 [ n o
U= (Ln (jp, js, t); (p, s) ∈ JLr−n ,Lr−n , t = 0 : (Lr−n−1 − 1)
n=0 j=1 [
{(0, 0, 0)} 
Image Processing 89

This theorem can be proved from the following statement [43, 46]:
Theorem 2.2 Let L > 1 be a prime number and let r > 1. Then, the follo-
wing is valid:
1) For each sample (p, s) of X and m = 0 : (Lr−n−1 − 1),
Lr−n−1
X −1
0 t

mt
F(mL+1)p,(mL+1)s = fp,s,L n t WLr−n WLr−n−1 . (2.55)
t=0

0
The 2-D DFT at frequency-points of the subset Tp,s is defined by the N/Ln+1 -
point DFT of the splitting-signal modified by twiddle coefficients. Here Ln is
defined as g.c.d.(p, s) = jLn , where j ∈ {1, ..., L − 1}.
2) The Lr × Lr -point DFT is split into (L2 − 1)Lr−1 Lr−1 -point DFTs,
(L − 1)Lr−2 Lr−2 -point DFTs, . . . , and (L2 − 1)L L-point DFTs.
2

It should be noted that the sum on the right side of Equation (2.55) de-
scribes the shifted 1-D DFT of the splitting-signal. For instance, when n = 0,
we have
Lr−1
X−1 Lr−1
X−1
0 (m+ 1 )t 0


Fm+ L1 = fp,s,t WLr−1 L = fp,s,t WLt r WLmt
r−1

t=0 t=0
m = 0 : (Lr−1 − 1),

which is the shifted Lr−1 -point DFT in the frequency domain, (m → m+1/L).
The 2-D paired transform, or the L-paired transform is the representation
of the 2-D image in the 2-D frequency and 1-D time domain in the form of a
family of paired splitting-signals,
[ L−1
r−1 [ [ n o
0 n+1
{fn,m } → fp,s,L n t ; t = 0 : (N/L − 1)
n=0 j=1 (p,s)∈jLn JN/Ln ,N/Ln
[
{f0,0,0},

and the L-paired splitting-signals generated by the frequencies (p, s) are

0 0 0 0
fTp,s
0 = {fp,s,0 , fp,s,L n , fp,s,2·Ln , . . . , fp,s,N/L−Ln }. (2.56)
Figure 2.30 shows the bridge image of size 243 × 243 in part a, and the
3-paired transform of the image in b. The first two sets of 2(324) = 648
0 ; (p, s) ∈ jJ243,243}, j = 1, 2, of length 81 each are
splitting-signals {fTp,s
shown in the form of two images of size 324 × 81 each in part a. The next
0 ; (p, s) ∈ j3J81,81}, j = 1, 2, of
two sets of 2(108) = 216 splitting-signals {fTp,s
length 27 each are shown in the form of two images of size 108 × 27 each, and
so on. The whole picture of these 968 splitting-signals represents the paired
transform of the tree image.
The 2-D DFT of the tree image as the set of 968 1-D DFTs of all 3-paired
splitting-signals is shown in Figure 2.31 in a similar way.
90 CHAPTER 2: Direction Images

j=1 j=2

300 2−D DPT

s
i
g 250
n
a
l 200

n
u 150
m
b j=1 j=2
e 100
r

50

81 27 9 31
signal length

(a) {fn,m } 0
(b) {fp,s,t }

FIGURE 2.30
(a) Bridge image 243 × 243 and (b) the splitting-signals of lengths
81, 27, 9, 3, 1, 1.

j=1 j=2

300 2−D DFT

D 250
F
T
200
n
u
m 150
b
e j=1 j=2
r 100

50

81 27 9 31
1−D DFT order

FIGURE 2.31
1-D DFTs of the modified splitting-signals. (All transforms are in absolute
scale and shifted to the center).

2.6.1 Principle of superposition: General case

In L-paired representation, the image fn,m of size N × N, when N = Lr ,
r > 1, is the set of (L + 1)(Lr − 1) + 1 splitting-signals,
0 0 0 0
fTp,s
0 = {fp,s,0 , fp,s,L n , fp,s,2·Ln , . . . , fp,s,N/L−Ln }, (2.57)
Image Processing 91
0 0
which define the 2-D DFT at the corresponding subsets Tp,s = Tp,s;L . The set
of all generators (p, s) of the paired transform components is defined as

[ L−1
r−1 [
0
JN,N = {jLk JLr−k ,Lr−k } ∪ {(0, 0)},
k=0 j=1

and the subsets JLk ,Lk , for k = 1 : r, are

k k−1
L[ −1 [ L [−1
JLk ,Lk = (1, s) (Lp, 1). (2.58)
s=0 p=0

According to (2.55), the 1-D DFT of the splitting-signals weighted by the

twiddle factors carry the spectral information of the image at frequency-points
0
of the subsets Tp,s ,

Lr−n−1
X −1
F(mL+1)p,(mL+1)s = Gm = gt0 WLmt
r−n−1 , (2.59)
t=0

where we define
0
gt0 = gp,s,t 0
= fp,s,L t
n t WLr−n , t = 0 : Lr−n−1 − 1.

Note that these components are periodic,

0
gt0 = gt+L 0
r−n−1 = gt mod Lr−n−1 . (2.60)

For instance, when g.c.d.(p, s) = j, or n = 0, the following calculations are

valid:
0 0 0
gt+N/L =fp,s,t+N/L W t+N/L = (fp,s,t+N/L WL )W t
L−1
!
X
= fp,s,(t+N/L)+kN/L WLk WL W t
k=0
L−1
!
X
= fp,s,t+(k+1)N/L WLk+1 Wt
k=0
L−1
!
X
= fp,s,t+kN/L WLk + fp,s,t+N WLL Wt
k=1
L−1
!
X
0
= fp,s,t+kN/L WLk W t = fp,s,t W t = gt0 ,
k=0

since the tensor transform is periodic, i.e., fp,s,t+N = fp,s,t .

The direction images are constructed similar to the N = 2r case. The
92 CHAPTER 2: Direction Images

incomplete 2-D DFT is composed only from the components of the 2-D DFT
0
with frequency-points of the subset Tp,s ,
 0
Fp1 ,s1 ; if (p1 , s1 ) ∈ Tp,s;L
Dp1 ,s1 = (2.61)
0; otherwise.
The image of this incomplete 2-D transform is composed only of values of the
splitting-signal generated by (p, s).
We first consider this image for a generator (p, s) from the first series of
0
generators in JN,N , i.e., when g.c.d.(p, s) ∈ {1, 2, ..., L − 1}. The following
calculations hold for the inverse 2-D DFT:
N−1 N−1
1 X X
d(p,s)
n,m = Dp ,s W −(np1+ms1 )
N 2 p =0 s =0 1 1
1 1

1 X
= 2 Fp1 ,s1 W −(np1+ms1 )
N
(p1 ,s1 )∈Tp,s
0

N/L−1
1 X
= F(kL+1)p,(kL+1)s W −(n(kL+1)p+m(kL+1)s)
N2
k=0
N/L−1
1 X
= F(Lk+1)p,(Lk+1)s W −(Lk+1)(np+ms)
N2
k=0
 
N/L−1
1 L X −kt 
= F(Lk+1)p,(kL+1)sWN/L W −t
LN N
k=0

where we denote t = (np + ms) mod N. According to (2.59), the sum in

the brackets is the N/L-point inverse DFT, which represents the data of g0 .
Therefore,
1 0
d(p,s)
n,m = g W −t
N L p,s,t mod (N/L)
1 0
= f W t mod (N/L) W −t
N L p,s,t mod (N/L)
1 0
= f W −t+t mod (N/L) . (2.62)
N L p,s,t mod (N/L)
Since the image is real, we can write this direction image as
L−1
!
(p,s) 1 X
dn,m = f W n W −(t−t̃)
N L n=0 p,s,t̃+nN/L L
L−1
1 X n− t−t̃
= fp,s,t̃+nN/L WL N/L (2.63)
N L n=0
L−1  2π 
1 X 1 
= cos n − r−1 (t − t̃) fp,s,t̃+nN/L
N L n=0 L L
Image Processing 93

where t̃ = t mod (N/L).

0
For (p, s) from the second series of generators in JN,N , i.e., when
g.c.d.(p, s) = jL and j ∈ {1, 2, ..., L − 1}, the following calculations are valid:

N/L2 −1
1 X −(kL+1)(np/L+ms/L)
d(p,s)
n,m = F(kL+1)p,(kL+1)s WN/L
N2
k=0
 
N/L2 −1
1 L2 X −k(np/L+ms/L) −(np/L+ms/L)
= F(kL+1)p,(kL+1)s WN/L2 WN/L
N L2 N
k=0

( we denote t = np/L + ms/L mod N/L )

1 0 −t
= g 2 · W
N L2 p,s,t mod N/L N/L
1 0 t mod N/L2 −t
= 2
fp,s,L(t mod N/L2 ) WN/L · WN/L
NL
1 0 −t+t mod N/L2
= 2
fp,s,L(t mod N/L2 ) WN/L . (2.64)
NL
0
Similarly, for (p, s) from the remaining series of generators in JN,N , that
k
is, when g.c.d.(p, s) = jL , where j ∈ {1, 2, ..., L − 1} and k = 2 : (r − 1), the
direction images are calculated by:
1 0 −t+t mod N/Lk+1
d(p,s)
n,m = k+1
fp,s,L k (t mod N/Lk+1 ) W
N/L k (2.65)
NL
where t = (np/Lk + ms/Lk ) mod N/Lk .
We can also write this image as

(p,s) 1
dn,m = f0 k W
−t+t̃
k
N Lk+1 p,s,L t̃ N/L
L−1
1 X
−t+t̃
= k+1
fp,s,Lk t̃+nN/L WLn WN/L k
NL n=0
L−1 (2.66)
1 X t−t̃
n− r−k−1
= fp,s,L k t̃+nN/L WL
L

N Lk+1 n=0
L−1  2π 
1 X 1 
= cos n − (t − t̃) fp,s,Lk t̃+nN/L
N Lk+1 n=0 L Lr−k−1

where t̃ = t mod (N/Lk+1 ).

Statement 3: Paired Transform Slice Theorem II: The discrete
image {fn,m } of size N × N, where N = Lr , L is prime, and r > 1, can
94 CHAPTER 2: Direction Images

be decomposed by (L + 1)(N − 1) + 1 direction images as

X
fn,m = d(p,s)
n,m
(p,s)∈JN,N
0

r−1 L−1 (2.67)

X X X
= d(p,s)
n,m + d(0,0)
n,m .
k=0 j=1 (p,s)∈jLk JN/Lk ,N/Lk

Each component of the paired transform can be calculated from the pro-
jection data. To reconstruct the discrete image on the Cartesian lattice, only
(L + 1)N/L projections are needed at the angles of the set

Φ={ϕ(p, s) = arctan(p/s); (p, s) ∈ JN,N }

={arctan(p); p = 0 : (N − 1)} ∪ {arctan(1/(Ls)); s = 0 : (N/L − 1)}.

Codes for discrete image decomposition by direction images, or image re-

construction from its projections by the tensor and paired transforms can be
found in Image Reconstruction at http://www.fasttransforms.com.

Problems
Problem 2.1 On the lattice 16 × 16, sketch the parallel lines, or rays that
pass through the points of the sets {V1,4,t ; t = 0 : 3}. Determine and sketch
parallel lines that pass through the points of the group T1,4 .

Problem 2.2 Consider the lattice 7 × 7 and the frequency-point (1, 3). Show
that the set of parallel rays that pass all points of the sets V1,3,t , t = 0 : 6, is
not unique. Determine at least two different sets of such parallel rays and for
each of these sets, determine the total number of rays.

Problem 2.3 Calculate all basic functions of the 2-D tensor transformation
for the N = 11 case and pack them into the matrix 121 × 121. Calculate also
the matrix of the inverse tensor transformation. Use MATLAB R
function
imagesc to sketch these matrices as images.

Problem 2.4 For the tree image 256 × 256, calculate the direction images
(1,3) (1,16)
dn,m and dn,m from the corresponding splitting-signals generated by fre-
quencies (p, s) = (1, 3) and (1, 16), respectively.
Image Processing 95

Problem 2.5 Consider the following discrete image of size 7 × 7 :

1 1 0 1 0 1 1
0 1 1 0 1 1 0
0 1 2 1 2 1 0
[fn,m ; n, m = 0 : 6] = 1 0 1 6 1 0 1 .
0 1 2 1 2 1 0
0 1 1 0 1 1 0
1 1 0 1 0 1 1

Calculate and sketch eight direction components of this image in tensor repre-
sentation. Calculate the original image from the sum of these eight direction
images.

Problem 2.6 Accomplish the subsampling of the tree image down to the size
32 × 32 and consider this image in paired representation.
A. Splitting-signals of the image have different energies. Determine and
sketch the five paired splitting-signals of the subsampled image that have the
highest energies.
B. Sketch the direction images defined by the splitting-signals in A. The
values of the direction image across all parallel rays are periodically repeated.
Determine these periods for each of the above five direction images.
C. The number of splitting-signals for the subsampled image 32 × 32 is 94,
and they gave different lengths. It is possible to pack all splitting-signals into
a matrix 32 × 32. Develop an algorithm to pack all splitting-signals in such
matrix.

Problem 2.7 The 2-D paired transformation χ0N,N is not separable. Prove
that the 2-D basic paired functions can be obtained from the basic functions
of the 1-D paired transformation.

Problem 2.8 Suppose that you are given an image of size 256 × 256. De-
velop an algorithm for calculating the 2-D separable paired transform of this
image. Compare your result with the 2-D non-separable paired transform of
the image.

Problem 2.9 Assume that the image you are given has the size 243×243. The
number 243 is the power of three, 243 = 35 . Develop a fast algorithm of the
3-paired transform of this image. You may calculate first the tensor transform
of this image, and then use it to calculate all 3-paired splitting-signals. Sketch
the image and the magnitude of the 3-paired transform.

Problem 2.10 The given image 512 × 512 can be extended to the size 513 ×
513. Discuss the effect of using the image 513 × 513 in the frequency domain,
instead of the original image.
This page intentionally left blank
3
Image Sampling Along Directions

The concept of the CT or computed X-ray tomography is illustrated in Fi-

gure 3.1, which shows the process of projection data collection. On both sides
of the analyzed object or image, which is assumed to be immovable, the X-ray
set and detectors are disposed and revolved around the object by a small angle
in certain time intervals, and the measurement data of radiation and detection
of X-rays are collected along other directions. This set of measurements, or

FIGURE 3.1
Method of computed X-ray tomography.

projections, taken at different positions of X-ray sources and detectors, is used

by specific mathematical methods to reconstruct the two-dimensional (slice),
or in general, the three-dimensional image of the observed object or tissue.
The methods of image reconstruction must be fast and accurate because of the
desired high quality of images for diagnostic purposes. And more important,
we need methods that reconstruct the image directly on the discrete Cartesian

97
98 CHAPTER 3: Image Reconstruction

lattice and with a minimal number of projections, in order to not over-radiate

the body in CT.
In this chapter, we analyze solutions of the problem of reconstruction of
the discrete image on the Cartesian lattice from projections of the image,
which are based on the concepts of the tensor and paired transformations. In
the framework of the constructed model, we show a way of using the ray-sums,
or line-integrals of the image, or real projection data, for exact reconstructing
the discrete image. Our goal is to show the way to transfer the geometry from
the image plane to the Cartesian lattice. The sampling theorem of Kotelnikov
is not used [35] to obtain the discrete image, but rather the sampling along
the projections. We need to reconstruct the image from a finite number of its
projections. The approach we propose for image reconstruction differs from
the well-known methods of backprojection, iterative reconstruction, Fourier
filtering, Radon filtering, and convolution filtering [25]-[32]. Each projection
is processed by a system of linear equations, or linear convolutions, to calculate
the corresponding part of the 2-D paired transform of the image, and then
the inverse paired transform is calculated to obtain the discrete image. The
model described for reconstruction is simple and the reconstruction is exact.

3.1 Image reconstruction: Model I

The image f(x, y) is considered in the form of a digital image fd (x, y) which
is represented by a finite number of elements, or image elements (IE) with
constant values and of size (∆x) × (∆y) each,
Z
1
fd (x, y) = f(x, y)dxdy, if (x, y) ∈ IE. (3.1)
(∆x)2 IE

We assume that the original image f(x, y) occupies the square region [0, 1] ×
[0, 1], which is divided into N 2 image elements by the Cartesian lattice N ×N,
where N > 1; therefore ∆x = ∆y = 1/N. All image elements are numbered
by (n, m), where n, m = 0 : (N − 1), and the image is considered in matrix
form fn,m with values
Z
fn,m = f(x, y)dxdy = (∆x)2 fd (x0 , y0 ), ∀(x0 , y0 ) ∈ (n,m)-th IE,
(n,m)-th IE
(3.2)
which are placed in the center of the corresponding (n, m)-th IE (see Fi-
gure 3.2). Thus, the point (x0 , y0 ) is considered to be the center of the (n, m)-th
IE. Thus the discrete image is considered on the square grid, or the Cartesian
lattice XN,N = {(n, m); n, m = 0 : (N − 1)} inside the region [0, 1] × [0, 1]. In
such a model, the ray-sum, or line-integral along the ray l in the (n, m)-th IE
Image Processing 99

1
f(x,y)

l−ray
0.67
∆l
f(n,m)
y m
0.33 IE
n

0
0 0.33 0.67 1
x

FIGURE 3.2
Model of the discrete image.

is defined as follows:
Z
(n,m) 1
wl = f(x, y)dl = (∆l)fd (x0 , y0 ) = (∆l) fn,m = (∆l)N 2 fn,m
(n,m)-th IE (∆x)2

where ∆l = ∆ln,m is the length of the ray l in the (n, m)-th IE. Therefore, if
∆l 6= 0, i.e., the ray passes through the IE, then
1 (n,m)
fn,m = wl , n, m = 0 : (N − 1). (3.3)
N 2 ∆l n,m

The line-integral of the image f(x, y) along the ray l is denoted by wl ,

Z X N−1
N−1 X X N−1
N−1 X
(n,m)
wl = f(x, y)dl = wl = N2 ∆ln,m fn,m . (3.4)
l m=0 n=0 m=0 n=0

We also can write that

X (n,m)
wl = wl
(n,m)-th IE ∩l

where the summation is performed by those IE which lie in the direction of

the ray l. Here, the idealized rays are considered, i.e., rays of width 0 each.
The sum of the discrete image along the ray l is
X 1 X 1 (n,m)
vl = fn,m = w . (3.5)
N2 ∆ln,m l
(n,m)∈l (n,m)∈l

We will call vl the line-sum.

100 CHAPTER 3: Image Reconstruction

If the number of projections is L, and {l} is the set of rays passing through
the (n, m)-th IE, then we can write the following:
 
1 1 X 1 (n,m) 
fn,m = 2 w , n, m = 0 : (N − 1). (3.6)
N L ∆l n,m l
l∈{l}

(n,m)
Here, we can probably approximate each normalized integral wl /∆ln,m
by the full integral of the image along the ray that passes through the point
(n, m), after normalizing this integral by the length of the ray. As a result, we
obtain an approximation
1 1 X wl
fn,m ≈ , n, m = 0 : (N − 1), (3.7)
N2 L ∆l
l∈{l}

where ∆l are the lengths of the rays inside the square region [0, 1] × [0, 1].
This approximation, which can be referred to as backprojection [29, 30], will
be considered in detail in Chapter 7. Now we discuss a method of exact re-
construction of the image.

3.1.1 Coordinate systems and rays

We consider two co-ordinate systems on the plane. The first system X-Y
of coordinates (x, y) is for the image f(x, y) on the square [0, 1] × [0, 1]. The
second coordinate system (n, m), where n and m are integers, is for the lattice
XN,N placed in the center of the square. This system is used for the discrete
image fn,m . Parameters x and n run from left to right, and parameters y and
m run from the top down, as shown in Figure 3.3, for the N = 4 case. The first
point of the image, f0,0 , is in the point with coordinates (x, y) = (1/8, 1/8).
Given the frequency-point (p, s) ∈ XN,N , such that g.c.d.(p, s) = 1, we
consider the lines

l(t) = lp,s (t) = {(n, m); pn + sm = t}, t = 0 : (p + s)(N − 1),

on the square grid XN,N . These lines are referred to as the arithmetical rays.
The equation of these lines on the square [0, 1] × [0, 1] is
 
t p+s
l(t) = lp,s (t) = (x, y); px + sy = + , t = 0 : (p + s)(N − 1).
N 2N

These lines are referred to as the geometrical rays, to distinguish the discrete
and continuous cases. These two types of rays are denoted by l(t), and the
same set of t is considered for the rays, t = 0 : (p + s)(N − 1). This set of t
for the geometrical rays may be changed for effective calculation of sums vl
for different (p, s), as will be shown later.
Image Processing 101

l1,1(2)
0
0
0.25
1
y and m 0.5
2
0.75
3
1
0 1 2 3
0 0.25 0.5 0.75 1
x and n

FIGURE 3.3
Two coordinate systems for images f(x, y) and f(n, m).

The generator (p, s) defines the slope, − tan−1 (p/s), of these rays. The
sums of the image along the arithmetical rays are denoted by
X
vp,s (t) = vl(t) = fn,m .
lp,s,t

Line-integrals wl are considered to be the physical measurements of the

projection, or measurements of geometrical rays, and the sums vl are mea-
surements of arithmetical rays. The main problem that we are going to solve
is formulated simply: Determine the sums vl from the line-integrals wl . Since
the discrete tensor and paired transforms of the image are completely defined
by the direction binary 2-D functions, these transforms can be used to solve
this problem and reconstruct the image from its projections. We first describe
in detail the case in which the discrete image is considered of size N × N and
N is a power of two. In this case, the tensor representation {fp,s,t } is modified
0
into the unitary paired transform {fp,s,t }, which does not require operations
of multiplication. The case where N is prime is considered in Chapter 5.

3.2 Inverse paired transform

Our task is to obtain the discrete image fn,m of size N × N, by using the 2-D
discrete paired transform (DPT), which can be calculated from the integrals
of the image f(x, y) along the rays, which correspond to the projections to
be defined. If we find a way of calculating the 2-D DPT from the physical
measurements, the problem of image reconstruction in discrete form will then
102 CHAPTER 3: Image Reconstruction

be solved by using the inverse 2-D DPT. We here recall briefly the main
results of the paired representation of the image, which defines a unique image
superposition by direction images.
The discrete image {fn,m } of size N × N, where N = 2r , r > 1, can be
composed of (3N − 2) splitting-signals as
r−1
1 X 1 X
0 1 0
fn,m = fp,s,(np+ms) mod N + f . (3.8)
2N 2k N 2 0,0,0
k=0 (p,s)∈2k J2r−k ,2r−k

The components of the paired transform

0
fp,s,2 k t = fp,s,2k t − fp,s,2k t+N/2 , t = 0 : (N/2k+1 − 1),

where 2k = g.c.d.(p, s), are calculated by the system of orthogonal 2-D paired
functions χ0p,s,2k t (n, m) defined in (2.39). The set of N 2 triplet-numbers of
paired functions is defined as
r−1
[n o
UN,N = (p, s, 2k t); (p, s) ∈ 2k JN/2k ,N/2k , t = 0 : (N/2k+1 −1) ∪{(0, 0, 0)},
k=0
(3.9)
where the subsets of generators (p, s) are calculated by
 
JN/2k ,N/2k = (1, s); s = 0 : (N/2k − 1) ∪ (2p, 1); p = 0 : (N/2k+1 − 1) .

If we normalize the components of the 2-D DPT as

0 1 0 1
fp,s,t → fp,s,t = f0 , 0
(f0,0,0 0
→ f0,0,0 /N ),
2 k+1 2 · g.c.d(p, s) p,s,t

then the decomposition of the image in (3.8) can be written as

 
r−1
1 X X
0 0
fn,m = fp,s,(np+ms) mod N + f0,0,0 .
 (3.10)
N
k=0 (p,s)∈2 J2r−k ,2r−k
k

Due to this formula, the value of the discrete image at each point (n, m)
is defined by 3N − 2 values of the 2-D DPT, which are calculated along
arithmetical rays of 3N/2 projections that pass through points (n, m) on the
Cartesian grid on the image. The number of operations of addition which are
required to reconstruct the image by this equation equals

a(N ) = N 2 [(3N − 2) − 1] = 3N 2 (N − 1).

Additional operations are required for calculating the splitting-signals from

the physical measurements, or the line-integrals wl .
Image Processing 103

3.3 Example: Image 4 × 4

We consider the masks of all basic paired functions χ0p,s,t in detail, for the
N = 4 case. Sixteen triplets (p, s, t) of the 4 × 4-point paired transform are
defined by the set


 (1, 0, 0), (1, 0, 1), (1, 1, 0), (1, 1, 1), (1, 2, 0), (1, 2, 1), (1, 3, 0), (1, 3, 1),

(0, 1, 0), (0, 1, 1), (2, 1, 0), (2, 1, 1),
U4,4 =

 (2, 0, 0), (2, 2, 0), (0, 2, 0),

(0, 0, 0).

3.3.1 Horizontal and vertical projections

First, we separate seven paired functions which are generated by (p, s) =
(1, 0), (0, 1), (0, 2), (2, 0), and (0, 0) and defined by the horizontal and vertical
projections. For the generator (p, s) = (1, 0), the masks of two paired basic
functions are
   
1 0 −1 0 0 1 0 −1
 1 0 −1 0   0 1 0 −1 
[χ01,0,0] =  1 0 −1 0 ,
 [χ01,0,1] = 
 0 1 0 −1 .
 (3.11)
1 0 −1 0 0 1 0 −1

The summation of the image is performed along the geometrical rays

l(t) = l1,0 (t) = {(x, y); 1x + 0y = t/4 + 1/8}, t = 0, 1, 2, 3,

as shown in Figure 3.4. Note that along the horizontal axis, the integer value

l(0) l(2) l(1) l(3)

0 0
0 0
0.25 0.25
1 1
y and m

y and m

0.5 0.5
2 2
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.4
Schemes of calculating four line-integrals for (p, s) = (1, 0).

n runs from 0 to 3, and x runs the interval [0, 1]. Along the vertical axis, the
104 CHAPTER 3: Image Reconstruction

integer value m runs from 0 to 3, and y runs the interval [0, 1]. The same
system of coordinates is used for all masks of basic functions of the paired
transform.
We denote the sums of the discrete image along these lines by v1,0 (t),

X 3
X
v1,0 (t) = vl(t) = fn,m = ft,m , t = 0, 1, 2, 3.
l1,0 (t) m=0

They are four measurements of the projection by the angle ϕ = ϕ1,0 = 0 to the
horizontal axis. This horizontal projection defines the first two components of
the 2-D paired transform of the image,
0 0
f1,0,0 = v1,0 (0) − v1,0 (2), f1,0,1 = v1,0 (1) − v1,0 (3).

All sums v1,0 (t) can be calculated from the horizontal projection of the image
f(x, y), Z
w1,0 (t) = wl(t) = f(x, y)dl, t = 0, 1, 2, 3.
l(t)

The length of intersection of the geometrical ray l(t) with the (t, m)-th IE
equals ∆l = ∆lt,m = 1/N, and
N−1
1 X (t,m) 1 1 1
v1,0 (t) = w = 2 1 wl(t) = wl(t) = w1,0(t). (3.12)
N 2 ∆l m=0 l N N N N

Therefore
0 1
f1,0,t [w1,0 (t) − w1,0(t + 2)], t = 0, 1.
=
N
The next two paired functions generated by (p, s) = (0, 1) have the follo-
wing masks:
   
1 1 1 1 0 0 0 0
 0 0 0 0  1 1 1 1
[χ00,1,0] = 
 −1 −1 −1 −1 ,
 [χ00,1,1] = 
 0 0 0 0 .
 (3.13)
0 0 0 0 −1 −1 −1 −1

The summation of the image is performed along the geometrical rays

l(t) = l0,1 (t) = {(x, y); 0x + 1y = t/4 + 1/8}, t = 0, 1, 2, 3,

as shown in Figure 3.5.

Here, four measurements of the projection by the angle ϕ = ϕ0,1 = π/2 to
the horizontal axis are used. We denote the sums of the discrete image along
these lines by v0,1 (t),

X 3
X
v0,1 (t) = fn,m = fn,t , t = 0 : 3.
l0,1 (t) n=0
Image Processing 105

0 0
0 l(0) 0
0.25 0.25
1 1 l(1)
y and m

y and m
0.5 0.5
2 l(2) 2
0.75 0.75
3 3 l(3)
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.5
Schemes of calculating four line-integrals for (p, s) = (0, 1).

This vertical projection defines two components of the 2-D paired transform,
0 0
f0,1,0 = v0,1 (0) − v0,1 (2), f0,1,1 = v0,1 (1) − v0,1 (3).

All sums v1,0 (t) can be calculated from the vertical projection of the image
f(x, y), Z
w0,1 (t) = wl(t) = f(x, y)dl, t = 01, 2, 3,
l(t)

in a way similar to the horizontal projection,

N−1
1 X n,t 1 1 1
v0,1 (t) = w = 2 1 wl(t) = wl(t) = w0,1 (t) (3.14)
N 2 ∆l n=0 l N N N N

and, therefore,

0 1
f1,0,t = [w1,0 (t) − w1,0(t + 2)], t = 0, 1.
N
The next three basic functions have the following masks:
     
1 1 1 1 1 −1 1−1 1 1 1 1
 −1 −1 −1 −1   1 −1 1−1 
1
 1 1 1
[χ00,2,0] =   0   0
 1 1 1 1 , [χ2,0,0] =  1 −1 1−1 , [χ0,0,0] =  1
,
1 1 1
−1 −1 −1 −1 1 −1 1−1 1 1 1 1

and they are defined by the vertical and horizontal projections. The corres-
106 CHAPTER 3: Image Reconstruction

ponding components of the 2-D paired transform can be calculated by

0
f0,2,0 =v0,1 (0) − v0,1 (1) + v0,1 (2) − v0,1 (3)
= N1 [w0,1 (0) − w0,1 (1) + w0,1 (2) − w0,1 (3)] ,
0
f2,0,0 =v1,0 (0) − v1,0 (1) + v1,0 (2) − v1,0 (3)
= N1 [w1,0 (0) − w1,0 (1) + w1,0 (2) − w1,0 (3)] ,
0
f0,0,0 =v0,1 (0) + v0,1 (1) + v0,1 (2) + v0,1 (3)
= N1 [w0,1 (0) + w0,1 (1) + w0,1 (2) + w0,1 (3)] ,
=v1,0 (0) + v1,0 (1) + v1,0 (2) + v1,0 (3)
= N1 [w1,0 (0) + w1,0 (1) + w1,0 (2) + w1,0 (3)] .
0
To calculate the component f0,0,0 , both horizontal and vertical projections
can be used.
Summarizing the above calculations for the vertical and horizontal projec-
tions, we can write the above equations in matrix form
 0    
f0,1,0 1 0 −1 0 w0,1 (0)
0
 0  = 1  0 1 0 −1  w0,1 (1)  ,
 f0,1,1    
 f0,2,0  4  1 −1 1 −1  w0,1 (2) 
0 1 1 1 1 w0,1 (3)
f0,0,0

and     
0
f1,0,0 1 0 −1 0 w1,0 (0)
0
 0  = 1  0 1 0 −1  w1,0 (1)  ,
 f1,0,1    
 f2,0,0  4  1 −1 1 −1  w1,0 (2) 
0 1 1 1 1 w1,0 (3)
f0,0,0
where  
1 0 −1 0
 0 1 0 −1 
[χ04,4] =  
 1 −1 1 −1 
1 1 1 1
is the matrix of the 4-point 1-D paired transformation, χ04 . The calculation of
0
the component f0,0,0 is repeated and, therefore, can be omitted from one of
these equations, or considered as the mean of calculations by both projections,

0 1h1
f0,0,0 = [w0,1(0) + w0,1 (1) + w0,1(2) + w0,1 (3)]+
2 4 i
1
+ [w1,0(0) + w1,0(1) + w1,0 (2) + w1,0 (3)] .
4
Image Processing 107

3.3.2 Diagonal projections

Now we consider the second set of five paired functions, which are generated
by (p, s) = (1, 1), (1, 3), and (2, 2). The masks of the first two functions are
   
1 0 −1 0 0 1 0 −1
 0 −1 0 1   1 0 −1 0 
[χ01,1,0] =   0
 −1 0 1 0  , [χ1,1,1] =  0 −1 0 1  .
 

0 1 0 −1 −1 0 1 0

Figure 3.6 shows two masks of the paired functions for the frequency-
point (p, s) = (1, 1) and seven parallel rays along which the coefficients ±1
are located. These rays define the diagonal projection of the image when the

l(0) l(2) l(1) l(3)

0 0
0 l(4) 0
0.25 0.25
1 1 l(5)
y and m

y and m

0.5 0.5
2 l(6) 2
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.6
Schemes of calculating seven line-integrals for (p, s) = (1, 1).

angle of projection equals π/4. The sums of the image along the arithmetical
rays
l(t) = l1,1,t = {(x, y); 1x + 1y = t/4 + 1/4}, t = 0, 1, ..., 6,
define the following components of the 2-D paired transform:
0
f1,1,0 = v1,1 (0) − v1,1 (2) + v1,1 (4) − v1,1 (6),
0
f1,1,1 = v1,1 (1) − v1,1 (3) + v1,1 (5).
√
In this case, ∆l = ∆l1,1 = 2∆x and the sums v1,1 (t) are defined by the
line-integrals w1,1 (t) as

v1,1 (t) √ v1,1 (t) √

w1,1 (t) = ∆l = 2 = 2N v1,1 (t), (3.15)
(∆x)2 ∆x
√
or v1,1 (t) = w1,1(t)/( 2N ), t = 0, 1, 2, 3. Therefore these two components of
108 CHAPTER 3: Image Reconstruction

the 2-D paired transform can be calculated by

0 1
f1,1,0 = √ [w1,1(0) − w1,1 (2) + w1,1(4) − w1,1 (6)],
4 2
0 1
f1,1,1 = √ [w1,1(1) − w1,1 (3) + w1,1(5)].
4 2
The mask of the paired function χ02,2,0 contains only coefficients ±1,
 
1 −1 1 −1
 −1 1 −1 1 
[χ02,2,0] = 
 1 −1 1 −1  ,


−1 1 −1 1
which can be seen on the seven parallel lines 1x + 1y = t/4 + 1/4, t = 0 : 6,

n+sm=t, (s=1) n+(s−4)m=t, (s=3)

l(0) l(1) l(2) l(3) l(0) l(1) l(2) l(3)
0 0
0 l(4) 0
l(−1)
0.25 0.25
1 l(5) 1
y and m

y and m

l(−2)
0.5 0.5
2 l(6) 2
l(−3)
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.7
Two schemes of calculating seven line-integrals for (p, s) = (2, 2).
0
which are shown in Figure 3.7 (on the left). Therefore the component f2,2,0
of the paired transform can be calculated by
6 6
0
X 1 X
f2,2,0 = (−1)n v1,1 (t) = √ (−1)n w1,1 (t).
t=0
4 2 t=0

The above calculations can be written in the following matrix form:

 
w1,1(0)
 0
   w1,1(1) 
f1,1,0 1 0 −1 0 1 0 −1  
 w1,1(2) 
0 1
 f 
1,1,1 
 0 1 0 −1 0 1 0  
 = 4√2  1 −1 1 −1 1 −1 1  w1,1(3)  ,
 0
  
 f2,2,0
√  w1,1(4) 
0 1 1 1 1 1 1 1 
f0,0,0 / 2 
 w1,1(5) 
w1,1(6)
Image Processing 109

or     
0
f1,1,0 1 0 −1 0 w1,1 (0) + w1,1 (4)
0
 f1,1,1  1 
= √  0 1 0 −1  w1,1 (1) + w1,1 (5) 
 
 4 2  1 −1 1 −1  w1,1 (2) + w1,1 (6)  .
 0
 f2,2,0
√
0 1 1 1 1 w1,1 (3)
f0,0,0 / 2
The matrix 4 × 4 in this equation is the matrix of the four-point DPT.

3.3.3 Other projections

The remaining three projections that are generated by the frequency-points
(p, s) = (1, 3), (1, 2), and (2, 1) are described similarly to the projections ge-
nerated by (p, s) = (1, 0), (0, 1), and (1, 1). We will start with the (1, 3) case,
which corresponds to the diagonal projection, and then consider two symmet-
ric cases (1, 2) and (2, 1).

3.3.3.1 Generator (1, 3)

Consider two paired functions that are generated by the frequency-point (1, 3)
and have the following masks:
   
1 0 −1 0 0 1 0 −1
 0 1 0 −1   −1 0 1 0 
[χ01,3,0] =   0
 −1 0 1 0  , [χ1,3,1] =  0 −1 0 1  .
 

0 −1 0 1 1 0 −1 0

Coefficients ±1 on these masks are situated on the parallel rays

l(t) = l1,3 (t) = {(x, y); 1 · x + 3 · y = t/4 + 1/2},

where t runs the integer numbers from 0 to 12. The rays compose the angle
ϕ1,3 = π −tan−1 (1/3) to the horizontal axis, as shown in the following matrix:
 
0 1 2 3
3 4 5 6
[1 · n + 3 · m = t] = 
6 7 8 9 .
 (3.16)
9 10 11 12

The line connecting two 6’s shows the ray l(6) and the direction of the parallel
rays in the projection. This large number of rays can be replaced with another
set of parallel rays with number less than 13. To show that, we first consider
the location of four coefficients −1 in the mask of the function χ01,3,0(n, m). The
locations of these units are determined by the above matrix, where coefficients
t = 2, 6, and 10 equal 2 mod 4. These four −1 are located on three parallel
rays by the angle ϕ1,3 to the horizontal. One ray passes through two points
with number 6, and one ray passes through the points with numbers 2 and
10 each. Similarly, one can notice that four coefficients 1 in the mask [χ01,3,0],
which are shown by coefficients t = 0, 4, 8, and 12 which equal 0 mod 4, are
110 CHAPTER 3: Image Reconstruction

on the different four parallel rays by the angle ϕ1,3 to the horizontal. Thus,
all units of this mask are situated on seven rays.
Consider the similar matrix [t] which is determined by the point (1, 3−4) =
(1, −1),  
0 1 2 3
 −1 0 1 2
[1 · n − 1 · m = t] = 
 −2 −1 0 1 .


−3 −2 −1 0
One can see that coefficients ±2 are located on two parallel rays, and all
coefficients 0 are on one diagonal ray. These rays compose the angle ϕ1,−1 =
tan−1 (1) = π/4 with the horizontal axis,

l(t) = l1,−1 (t) = {(x, y); 1 · x − 1 · y = t/4},

and the total number of them equals (1+|−1|)3+1 = 7, as shown in Figure 3.7
(on the right).
The sums of the image along these arithmetical rays l(t) define the follo-
wing components of the 2-D paired transform:
0
f1,3,0 = −v1,−1 (−2) + v1,−1 (0) − v1,−1 (2),
0
f1,3,1 = v1,−1 (−3) − v1,−1 (−1) + v1,−1 (1) − v1,−1 (3),

as shown in Figure 3.8.

l(0) l(2) l(1) l(3)

0 0
0 0
l(−1)
0.25 0.25
1 1
y and m

y and m

l(−2)
0.5 0.5
2 2
l(−3)
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.8
Schemes of calculating seven line-integrals for (p, s) = (1, −1).

Similar to the (p, s) = (1, 1) case, the sums v1,−1 (t) of the image are defined
by the line-integrals w1,−1 (t) as follows:

v1,−1 (t) √ v1,−1 (t) √

w1,−1 (t) = ∆l 2
= 2 = N 2v1,−1 (t), (3.17)
(∆x) ∆x
Image Processing 111
√
or v1,−1 (t) = w1,−1(t)/( 2N ), t = −3 : 3. Therefore we obtain

0 1
f1,3,0 = √ [−w1,−1(−2) + w1,−1 (0) − w1,−1 (2)],
4 2
0 1
f1,3,1 = √ [w1,−1(−3) − w1,−1 (−1) + w1,−1 (1) − w1,−1 (3)].
4 2
As follows from Figure 3.7 (on the right), one can also calculate the component
0
f2,2,0 , by using these line-integrals,

3 3
0
X 1 X
f2,2,0 = (−1)n v1,−1 (t) = √ (−1)n w1,−1 (t).
t=−3
4 2 t=−3

The calculations of these three components of the paired transform to-

0
gether with f0,0,0 can be written in matrix form
 
w1,−1(−3)
 0
   w1,−1(−2) 
f1,3,0 0 −1 0 1 0 −1 0  
 w1,−1(−1) 
0
 f 
1,3,1  1  1 0 −1
 0 1 0 −1  
 = 4√2  −1 1 −1
 0
 w1,−1(0)  ,
 f2,2,0
√ 1−1 1 −1 
 w1,−1(1) 

0 1 1 1 1 1 1 1 
f0,0,0 / 2  w1,−1(2) 


w1,−1(3)

which can also be written as

 0
   
f1,3,0 1 0−1 0 w1,−1 (0)
0 1 

 f1,3,1 
= √  0 1 0 −1  w1,−1 (1) + w1,−1(−3)  .
 
0
 f2,2,0
√
 4 2  1 −1 1 −1  w1,−1 (2) + w1,−1(−2) 
0 1 1 1 1 w1,−1 (3) + w1,−1(−1)
f0,0,0 / 2

3.3.3.2 Generator (1, 2)

We now consider the masks of the paired functions for (p, s) = (1, 2),
   
1 0 −1 0 0 1 0 −1
 −1 0 1 0  , [χ0 ] =  0 −1 0 1  .
 
[χ01,2,0] =  1,2,1
 1 0 −1 0   0 1 0 −1 
−1 0 1 0 0 −1 0 1

The coefficients ±1 on these masks are situated on the parallel rays at the
angle ϕ1,2 = π − tan−1 (1/2) with the horizontal axis (see Figure 3.9),

l(t) = l1,2 (t) = {(x, y); 1 · x + 2 · y = t/4 + 3/8},

where t runs the integer numbers from 0 to 9. The sums of the image along
112 CHAPTER 3: Image Reconstruction

l(0) l(2) l(1)

0 0 l(3)
0 l(4) 0
0.25 0.25 l(5)
1 l(6) 1
y and m

y and m
0.5 0.5 l(7)
2 l(8) 2
0.75 0.75 l(9)
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.9
Schemes of calculating seven line-integrals for (p, s) = (1, 2).

the corresponding arithmetical rays define the following components of the

paired transform:
0
f1,2,0 = v1,2 (0) − v1,2 (2) + v1,2 (4) − v1,2 (6) + v1,2 (8),
0
f1,2,1 = v1,2 (1) − v1,2 (3) + v1,2 (5) − v1,2 (7) + v1,2 (9).

One can see from Figure 3.9 and Figure 3.10 (on the left), that the geo-
metrical rays l(t) pass through the additional squares, or image elements on
which the arithmetical rays are not defined. For example, the arithmetical ray
l(6) passes through only two image elements with numbers (0, 3) and (2, 2),
and the geometrical ray l(6) passes through six image elements with numbers
(0, 3), (1, 3), (1, 2), (2, 2), (3, 2), and (3, 1). Each sum v1,2 (t) cannot be calcu-
lated from only the integral w1,2 (t), which is defined along the geometrical
rays l(t), as was the case for the (p, s) = (1, 1) and (1, 3) cases. To derive a
simple expression between the measurements along the arithmetical and geo-
metrical rays for this projection, we first consider the arithmetical rays l(6)
and l(7) and the geometrical ray l = {(x, y); x +2y = 2} between them, which
can be considered as the ray ˜l(7) = l(7 − 0.5) shifted in the vertical direction.
This
√ geometrical √ ray intersects four image elements with intersection of length
5/2∆x = 5/8 in each. Therefore, the line-integral along this ray can be
expressed by two sums of the image along the neighbor rays, but arithmetical,
as follows:
√
5 1
w = w1,2 (7) = w(˜ l(7)) = 2
[f0,3 + f1,3 + f2,2 + f3,2 ]
√ (∆x)
8
5 1
= 2
[(f0,3 + f2,2 ) + (f1,3 + f3,2 )]
√ (∆x)
8  
5 v1,2 (6) + v1,2 (7) √
= 2
= 2 5 [v1,2 (6) + v1,2 (7)] .
8 (∆x)
Image Processing 113

l(0) l(1) l(2) l(3)

0 0 l(4)
0 0 l(5)
0.25 0.25 l(6)
l(6)
1 1
y and m

y and m
l(7)
0.5 0.5 l(8)
l(7)
2 2 l(9)
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.10
Schemes of calculating seven line-integrals for (p, s) = (1, 2).

Case 1: Consider the set of 10 shifted geometrical rays

˜
l(t) = l1,2 (t − 0.5) = {(x, y); 1 · x + 2 · y = t/4 + 1/8},

which are shown in Figure 3.10 (on the right) and numbered from left to right
and top to bottom (and without the fraction 1/2 in the numbers). In the
figure, rays l̃(t) are denoted by l(t). For all line-integrals of the image along
these rays, we can write the following:
√
w1,2(t) = 2 5 [v1,2 (t − 1) + v1,2 (t)] , t = 0 : 9, (3.18)

where it is assumed that v1,2 (−1) = 0. These equations can be written in

matrix form
    
w1,2(0) 1 0 0 0 0 0 0 0 0 0 v1,2 (0)
 w1,2(1)   1 1 0 0 0 0 0 0 0 0  v1,2 (1) 
    
 w1,2(2)   0 1 1 0 0 0 0 0 0 0  v1,2 (2) 
    
 w1,2(3)   0 0 1 1 0 0 0 0 0 0  v1,2 (3) 
    
 w1,2(4)  √ 
 = 2 5  0 0 0 1 1 0 0 0 0 0  v1,2 (4)  .
 
w=  w1,2(5)   0 0 0 0 1 1 0 0 0 0  v1,2 (5)  (3.19)
    
 w1,2(6)   0 0 0 0 0 1 1 0 0 0  v1,2 (6) 
    
 w1,2(7)   0 0 0 0 0 0 1 1 0 0  v1,2 (7) 
    
 w1,2(8)   0 0 0 0 0 0 0 1 1 0  v1,2 (8) 
w1,2(9) 0 0 0 0 0 0 0 0 1 1 v1,2 (9)
114 CHAPTER 3: Image Reconstruction

The above Toeplitz matrix 10 × 10 with a left bandwidth of 1 has the triangle
inverse matrix, and all sums along the arithmetical rays are calculated by
    
v1,2 (0) 1 0 0 0 0 0 0 0 0 0 w1,2 (0)
 v1,2 (1)   −1 1 0 0 0 0 0 0 0 0  w1,2 (1) 
    
 v1,2 (2)   1 −1 1 0 0 0 0 0 0 0  w1,2 (2) 
    
 v1,2 (3)   −1 1 −1 1 0 0 0 0 0 0  w1,2 (3) 
    
 v1,2 (4)  1 
= √  1 −1 1 −1 1 0 0 0 0 0  w1,2 (4) .
 

 v1,2 (5)  2 5  −1 1 −1 1 −1 1 0 0 0 0  w1,2 (5)  (3.20)
    
 v1,2 (6)   1 −1 1 0 1 −1 1 0 0 0  w1,2 (6) 
    
 v1,2 (7)   −1 1 −1 1 −1 1 −1 1 0 0  w1,2 (7) 
    
 v1,2 (8)   1 −1 1 −1 1 −1 1−1 1 0  w1,2 (8) 
v1,2 (9) −1 1 −1 1 −1 1 −1 1 −1 1 w1,2 (9)
This equation can also be written in the following simple recurrent form:
v1,2 (t)= (−1)t [w1,2(0) − w1,2 (1) + w1,2 (2) − ... + (−1)t w1,2 (t)]
= w1,2 (t) − v1,2 (t − 1), (v1,2 (−1) = 0) (3.21)
t = 0, 1, 2, . . ., 9,
√
where the factor of 1/2 5 is omitted for simplicity.
Case 2: We now consider another set of geometrical rays,
˜
l(t) = l1,2 (t + 0.5) = {(x, y); 1 · x + 2 · y = t/4 + 1/2},
which are defined by shifting the arithmetical rays to the left, as shown in
Figure 3.11. The geometrical ray l = {(x, y); x + 2y = 2} between the arith-

l(0) l(1) l(2)

0 0 l(3)
0 0 l(4)
0.25 0.25 l(5)
l(6)
1 1
y and m

y and m

l(6)
0.5 0.5 l(7)
l(7)
2 2 l(8)
0.75 0.75 l(9)
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.11
The 2nd set of rays for calculating seven line-integrals for (p, s) = (1, 2).

metical rays l(6) and l(7) is considered as the shifted ray ˜l(6) = l(6 + 0.5).
The line-integral along this ray can be expressed by two sums of the image
along these arithmetical rays, as follows:
√
w = w1,2(6) = w(˜l(6)) = 2 5 [v1,2 (6) + v1,2 (7)] .
Image Processing 115

For all line-integrals of the image along the rays ˜l(t), we can write the following:
√
w1,2(t) = 2 5 [v1,2 (t) + v1,2 (t + 1)] , t = 0 : 9, (3.22)

where it is assumed that v1,2 (10) = 0. These equations can be written in

matrix form
    
w1,2(0) 1 1 0 0 0 0 0 0 0 0 v1,2 (0)
 w1,2(1)   0 1 1 0 0 0 0 0 0 0  v1,2 (1) 
    
 w1,2(2)   0 0 1 1 0 0 0 0 0 0  v1,2 (2) 
    
 w1,2(3)   0 0 0 1 1 0 0 0 0 0  v1,2 (3) 
    
 w1,2(4)  √  0 0 0 0 1 1 0 0 0 0  v1,2 (4) 
w= 
 = 2 5  0 0 0 0 0 1 1 0 0 0  v1,2 (5)  .
    (3.23)
 w1,2(5)    
 w1,2(6)   0 0 0 0 0 0 1 1 0 0  v1,2 (6) 
    
 w1,2(7)   0 0 0 0 0 0 0 1 1 0  v1,2 (7) 
    
 w1,2(8)   0 0 0 0 0 0 0 0 1 1  v1,2 (8) 
w1,2(9) 0 0 0 0 0 0 0 0 0 1 v1,2 (9)

The Toeplitz matrix 10×10 with a right bandwidth of 1 has the upper triangle
inverse matrix, and all sums along the arithmetical rays are calculated by
    
v1,2 (0) 1 −1 1−1 1 −1 1 −1 1 −1 w1,2 (0)
 v1,2 (1)   0 1 −1 1 −1 1 −1 1 −1 1  w1,2 (1) 
    
 v1,2 (2)   0 0 1−1 1 −1 1 −1 1 −1  w1,2 (2) 
    
 v1,2 (3)   0 0 0 1 −1 1 −1 1 −1 1  w1,2 (3) 
    
 v1,2 (4)  1 
= √  0 0 0 0 1 −1 1 −1 1 −1  w1,2 (4)  .
 

 v1,2 (5)  2 5  0 0 0 0 0 1 −1 1 −1 1  w1,2 (5)  (3.24)
    
 v1,2 (6)   0 0 0 0 0 0 1 −1 1 −1  w1,2 (6) 
    
 v1,2 (7)   0 0 0 0 0 0 0 1 −1 1  w1,2 (7) 
    
 v1,2 (8)   0 0 0 0 0 0 0 0 1 −1  w1,2 (8) 
v1,2 (9) 0 0 0 0 0 0 0 0 0 1 w1,2 (9)

This equation can also be written in the following recurrent form:

v1,2 (9)= w1,2 (9)

(3.25)
v1,2 (t)= w1,2 (t) − v1,2 (t + 1), t = 8, 7, 6, . . . , 1, 0,
√
where the factor of 1/(2 5) is omitted for simplicity.

3.3.3.3 Generator (2, 1)

The last two paired functions are defined by the generator (p, s) = (2, 1), and
their masks are
   
1 −1 1 −1 0 0 0 0
 0 0 0 0  1 −1 1 −1 
[χ02,1,0] =   0
 −1 1 −1 1  , [χ2,1,1] =  0 0 0 0  .
 

0 0 0 0 −1 1 −1 1
116 CHAPTER 3: Image Reconstruction

These masks are the transposition of the corresponding matrices considered

above for the generator (1, 2). Ten parallel rays
l(t) = l2,1,t = {(x, y); 2 · x + 1 · y = t/4 + 3/8}, t = 0 : 9,
are shown separately in Figure 3.12. The sums of the image along such arith-

l(0) l(2) l(4) l(6) l(1) l(3) l(5) l(7)

0 0
0 0
l(8)
0.25 0.25
1 1 l(9)
y and m

y and m
0.5 0.5
2 2
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.12
Schemes of calculating 11 line-integrals for (p, s) = (2, 1).

metical rays define the following components of the paired transform:

0
f2,1,0 = v2,1 (0) − v2,1 (2) + v2,1 (4) − v2,1 (6) + v2,1 (8),
0
f2,1,1 = v2,1 (1) − v2,1 (3) + v2,1 (5) − v2,1 (7) + v2,1 (9).

Case 1: As in the (p, s) = (1, 2) case, the geometrical rays l(t) pass
through some image elements which are not intersected by arithmetical rays
of this projection. Therefore we consider a new set of shifted geometrical rays
˜
l(t) = l2,1 (t − 0.5) = {(x, y); 2 · x + 1 · y = t/4 + 1/4}.
The example with arithmetical rays l(4) and l(5) given in Figure 3.13 (on
the left) shows that the geometrical√ray between them intersects four image
elements with intersection of length 5/8 in each. Therefore, the line-integral
along this ray can be expressed by two neighbor rays, but arithmetical, as
follows:
√
5 1
w = w(5)= [(f1,3 + f2,1 ) + (f1,2 + f2,0 )]
8 (∆x)2
√  
5 v2,1 (5) + v1,2 (4) √
= 2
= 2 5 [v2,1 (5) + v2,1 (4)] .
8 (∆x)
Ten such shifted geometrical rays are shown in Figure 3.13 (on the right). For
all line-integrals of the image along these rays, we can write the following:
√
w2,1(t) = 2 5 [v2,1 (t) + v2,1 (t − 1)] , t = 0 : 9, (3.26)
Image Processing 117

l(4) l(5) l(0) l(1) l(3) l(5) l(7)

0 0
0 0 l(8)
0.25 0.25
1 1 l(9)
y and m

y and m
0.5 0.5
2 2
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.13
Schemes of calculating 11 line-integrals for (p, s) = (2, 1).

where v2,1 (−1) = 0. The sums of the image along the arithmetical rays can
be calculated from ten line-integrals along the geometrical rays, by using the
following equation:

v2,1 (t) =(−1)t [w2,1(0) − w2,1 (1) + w2,1(2) − . . . + (−1)t w2,1 (t)]
=w2,1 (t) − v2,1 (t − 1), (v2,1 (−1) = 0) (3.27)
t = 0, 1, 2, . . ., 9,
√
where the factor of 1/(2 5) is omitted.
We also can define two new binary orthogonal matrices
   
1 −1 1 −1 1 −1 1 −1
 1 −1 1 −1   −1 1 −1 1 
[χ02,1,0+1] =   0
 −1 1 −1 1  , [χ2,1,0−1] =  −1 1 −1 1 
 

−1 1 −1 1 1 −1 1 −1

and consider new components of superposition of these functions with the

image,
3 X
X 3
0
f2,1,0+1 = χ02,1,0+1(n, m)fn,m
m=0 n=0

and
3 X
X 3
0
f2,1,0−1 = χ02,1,0−1(n, m)fn,m .
m=0 n=0

These components can be calculated from the line-integrals along the geomet-
118 CHAPTER 3: Image Reconstruction

rical rays as
0
f2,1,0+1 = (v2,1,0 − v2,1,2 + v2,1,4 − v2,1,6 + v2,1,8 )
+(v2,1,1 − v2,1,3 + v2,1,5 − v2,1,7 + v2,1,9 )
= (v2,1,0 + v2,1,1 ) − (v2,1,2 + v2,1,3 ) + (v2,1,4 + v2,1,5)
−(v2,1,6 + v2,1,7 ) + (v2,1,8 + v2,1,9 )
1
= 2√ 5
(w2,1,1 − w2,1,3 + w2,1,5 − w2,1,7 + w2,1,9),

where v2,1,t = v2,1 (t), w2,1,t = w2,1 (t), and

0
f2,1,0−1 =(v2,1,0 − v2,1,2 + v2,1,4 − v2,1,6 + v2,1,8)
−(v2,1,1 − v2,1,3 + v2,1,5 − v2,1,7 + v2,1,9)
=v2,1,0 − (v2,1,2 + v2,1,1) + (v2,1,4 + v2,1,3)
−(v2,1,6 + v2,1,5) + (v2,1,8 + v2,1,7) − v2,1,9
1
= 2√ 5
(w2,1,0 − w2,1,2 + w2,1,4 − w2,1,6 + w2,1,8 − w2,1,10),

when considering v2,1,−1 = v2,1,10 = 0.

Case 2: Let us consider another set of geometrical rays, by shifting the
arithmetical rays to the left,
˜
l(t) = l2,1 (t + 0.5) = {(x, y); 2 · x + 1 · y = t/4 + 1/2},

as shown in Figure 3.14. The geometrical ray l = {(x, y); x + 2y = 2} between

l(4) l(5) l(0) l(2) l(4) l(6)

0 0
0 0 l(7)
0.25 0.25
1 1 l(8)
y and m

y and m

0.5 0.5
2 2 l(9)
0.75 0.75
3 3
1 1
0 1 2 3 0 1 2 3
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x and n x and n

FIGURE 3.14
The 2nd set of rays for calculating seven line-integrals for (p, s) = (2, 1).

the arithmetical rays l(4) and l(5) is considered as the shifted ray ˜l(4) =
l(4 + 0.5). The line-integral along this ray can be expressed by two sums of
the image along these arithmetical rays, as follows:
√
w = w2,1(4) = w(˜l(4)) = 2 5 [v1,2 (4) + v1,2 (5)] .
Image Processing 119

Ten geometrical rays are defined as ˜l(t) = l(t+0.5), t = 0 : 9. All line-integrals

of the image along these rays can be written as
√
w2,1(t) = 2 5 [v2,1 (t) + v2,1 (t + 1)] , t = 0 : 9, (3.28)

where it is assumed that v2,1 (10) = 0. The solution of this system of equations
can be calculated by using the following recurrent procedure:
b(9)= w2,1 (9),
(3.29)
b(t)= w2,1 (t) − b(t + 1), t = 8, 7, 6, . . ., 1, 0,
√
and v2,1 (t) = b(t)/(2 5), for t = 0 : 8.
Thus for reconstruction of the image 4 × 4, six projections are used with
total 4 + 4 + 7 + 7 + 10 + 10 = 42 measurements along the geometrical rays,
as shown in Table 3.1.
TABLE 3.1
Data of the 4 × 4 discrete image reconstruction
Projection Number (p, s, t) of basic paired functions # measurements
0◦ (1, 0, 0), (1, 0, 1), (2, 0, 0), (0, 0, 0) 4
90◦ (0, 1, 0), (0, 1, 1), (0, 2, 0) 4
45◦ (1, 1, 0), (1, 1, 1), (2, 2, 0) 7
−45 (135◦ )
◦
(1, 3, 0), (1, 3, 1) 7
63.43◦ (1, 2, 0), (1, 2, 1) 10
26.57◦ (2, 1, 0), (2, 1, 1) 10

All equations of the A-rays l(t), G-rays ˜l(t) = l(t − t0 ) with shifts t0 ,
and the sums vp,s (t) considered above for the example N = 4, are given in
Table 3.2. In the last column, the normalized factors are given for the sums
vp,s (t), so that vp,s (t) = vp,s (t)/K. For generators (p, s) = (1, 2) and (2, 1),
both cases of G-rays l̃(t) are given, when the corresponding set of A-rays are
shifted to the left and right.

TABLE 3.2
Equations of G-rays for the image 4 × 4
(p, s) A-rays t0 G-rays l(t − t0 ) t vp,s = A−1 wp,s K
(1, 0) l1,0 (t) 0 x = 4t + 81 0:3 w1,0 (t) 4
√
(1, 1) l1,1 (t) 0 x + y = 4t + 41 0:6 w1,1 (t) 4√2
(1, 2) l1,2 (t) 1/2 x + 2y = 4t + 81 0:9 w1,2 (t) − v1,2 (t − 1) 2√5
l1,2 (t) −1/2 x + 2y = 4t + 21 9:0 w1,2 (t) − v1,2 (t + 1) 2√5
(1, 3) l1,−1 (t) 0 x − y = 4t 3 : −3 w1,−1 (t) 4 2
(0, 1) l0,1 (t) 0 y = 4t + 81 0:7 w0,1 (t) 4
√
(2, 1) l2,1 (t) 1/2 2x + y = 4t + 41 0:9 w2,1 (t) − v2,1 (t − 1) 2√5
l2,1 (t) −1/2 2x + y = 4t + 2t 9:0 w2,1 (t) − v2,1 (t + 1) 2 5

Model is original: equations are np + ms = t and xp + ys = t1 .

120 CHAPTER 3: Image Reconstruction

3.4 Property of the directed multiresolution

The paired functions generated by frequency-points (p, s) and (2p, 2s), as well
as (4p, 4s), (8p, 8s), . . . , are defined by the same direction. We denote these
set of points by U (p, s). This property holds because of the relation

χ2p,2s,2t = χ2p,2s,2t = χp,s,t + χp,s,t+N/2 , t = 0 : (N/2 − 1),

where subscripts are considered modulo N. The paired function χ02p,2s,2t can
thus be calculated by

χ02p,2s,2t = χ2p,2s,2t−χ2p,2s,2t+N/2 = χp,s,t −χp,s,t+N/4 +χp,s,t+N/2 −χp,s,t+3N/4 .

0
Components of the 2-D paired transform are defined as fp,s,t = fp,s,t −
0
fp,s,t+N/2 , and the functions f2p,2s,2t can be calculated by
0 0
f2p,2s,2t = f2p,2s,2t = fp,s,t − fp,s,t+N/4 + fp,s,t+N/2 − fp,s,t+3N/4 .

Similarly, for other triplet-numbers of the set Up,s , the components of the
paired splitting-signals generated by frequencies (4p, 4s), (8p, 8s), ... can be
calculated from the same splitting-signal {fp,s,t ; t = 0 : (N − 1)}. This means
the following: For each generator (p, s) ∈ JN,N , the set U (p, s) consists of
N triplet-numbers, and the 2-D paired transform of the image on this set
is defined by the 1-D paired transform of the tensor splitting-signal fTp,s =
{fp,s,t ; t = 0 : (N − 1)}. Thus
 0

 {fp,s,t ; t = 0 : (N/2 − 1)},

 0


 {f2p,2s,2t ; t = 0 : (N/4 − 1)},

 {f 0
4p,4s,4t; t = 0 : (N/8 − 1)},
χ0N [fTp,s ] =

 ... ... ... ,

 0


 {fN/2,N/2,0 },

 0
{f0,0,0}.

This property will be used in the next N = 8 example, as well as in the general
case when N = 2r , r > 1.
Image Processing 121

3.5 Example: Image 8 × 8

The 64 triplets (p, s, t) of the 8 × 8-point paired transform are defined by the
set

 (1, 0, 0), (1, 0, 1), (1, 0, 2), (1, 0, 3), (1, 1, 0), (1, 1, 1), (1, 1, 2), (1, 1, 3),




 (1, 2, 0), (1, 2, 1), (1, 2, 2), (1, 2, 3), (1, 3, 0), (1, 3, 1), (1, 3, 2), (1, 3, 3),



 (1, 4, 0), (1, 4, 1), (1, 4, 2), (1, 4, 3), (1, 5, 0), (1, 5, 1), (1, 5, 2), (1, 5, 3),



 (1, 6, 0), (1, 6, 1), (1, 6, 2), (1, 6, 3), (1, 7, 0), (1, 7, 1), (1, 7, 2), (1, 7, 3),

(0, 1, 0), (0, 1, 1), (0, 1, 2), (0, 1, 3), (2, 1, 0), (2, 1, 1), (2, 1, 2), (2, 1, 3),
U8,8 =

 (4, 1, 0), (4, 1, 1), (4, 1, 2), (4, 1, 3), (6, 1, 0), (6, 1, 1), (6, 1, 2), (6, 1, 3),



 (2, 0, 0), (2, 0, 2), (2, 2, 0), (2, 2, 2), (2, 4, 0), (2, 4, 2), (2, 6, 0), (2, 6, 2),

 (0, 2, 0), (0, 2, 2), (4, 2, 0), (4, 2, 2),



 (4, 0, 0), (4, 4, 0), (0, 4, 0),



(0, 0, 0).

We separate these 64 triplets by subsets wherein the triplets correspond to

the components of the 2-D paired transform, which are defined by the same
directions, or projections. The number of these directions is 12 and they are
calculated by ϑ = arctg(p/s), where (p, s) ∈ J8,8 , and

J8,8 = {(1, s); s = 0 : 7} ∪ {(2p, 1); p = 0 : 3}.

Given generator (p, s) ∈ JN,N , the corresponding subset of triplets 2k (p, s, t),
where t = 0 : N/2k+1 − 1, when k = 0 : (r − 1), is called U (p, s). The triplet
(0, 0, 0) is also an element of this set.

3.5.1 Horizontal projection

Consider the first subset of triplets (p, s, t),

 (1, 0, 0), (1, 0, 1), (1, 0, 2), (1, 0, 3),


(2, 0, 0), (2, 0, 2),
U (1, 0) =

 (4, 0, 0),

(0, 0, 0).

Masks of the corresponding eight 2-D paired functions equal

   
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
 1 0 0 0 −1 0 0 0  0 1 0 0 0 −1 0 0
   
 1 0 0 0 −1 0 0 0  0 1 0 0 0 −1 0 0
   
 1 0 0 0 −1 0 0 0   0 1 0 0 0 −1 0 0
[χ01,0,0] = 
 1 0 0 0 −1 0 0 0  , [χ 0
1,0,1 ] =  ,
 
0 1 0 0
 0 −1 0 0
 1 0 0 0 −1 0 0 0  0 1 0 0 0 −1 0 0
   
 1 0 0 0 −1 0 0 0  0 1 0 0 0 −1 0 0
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
122 CHAPTER 3: Image Reconstruction
   
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1
0 0 1
 0 0 0 −1 0 

 0 0 0 1 0 0 0 −1 
 
0 0 1
 0 0 0 −1 0 

 0 0 0 1 0 0 0 −1 
 
0
0 0 1
 , [χ01,0,3] =  0 0 0 1 0 0 0 −1  ,
0 0 0 −1 0   
[χ1,0,2] = 0 0 1
 0 0 0 −1 0 

 0 0 0 1 0 0 0 −1 
 
0 0 1
 0 0 0 −1 0 

 0 0 0 1 0 0 0 −1 
 
0 0 1 0 0 0 −1 0   0 0 0 1 0 0 0 −1 
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1
   
1 0 −1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
 1 0 −1 0 1 0 −1 0   0 1 0 −1 0 1 0 −1 
   
 1 0 −1 0 1 0 −1 0 
  0 1 0 −1 0 1 0 −1 
  
 1 0 −1 0 1 0 −1 0 
 , [χ0 ] =  0 1 0 −1 0 1 0 −1  ,
0
 
[χ2,0,0] =  1 0 −1 2,0,2
 0 1 0 −1 0 

 0 1 0 −1 0 1 0 −1 
 
 1 0 −1 0 1 0 −1 0   0 1 0 −1 0 1 0 −1 
  

 1 0 −1 0 1 0 −1 0   0 1 0 −1 0 1 0 −1 
1 0 −1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
   
1 −1 1 −1 1 −1 1 −1 1 1 1 1 1 1 1 1
 1 −1 1 −1 1 −1 1 −1  1 1 1 1 1 1 1 1
   
 1 −1 1 −1 1 −1 1 −1  1 1 1 1 1 1 1 1
  
 1 −1 1 −1 1 −1 1 −1  , [χ0 ] =  1 1 1 1 1 1 1 1  .
 
[χ04,0,0] = 
 1 −1 1 0,0,0
 −1 1 −1 1 −1 
1 1 1 1 1 1 1 1
 
 1 −1 1 −1 1 −1 1 −1  1 1 1 1 1 1 1 1
  

 1 −1 1 −1 1 −1 1 −1   1 1 1 1 1 1 1 1
1 −1 1 −1 1 −1 1 −1 1 1 1 1 1 1 1 1
Eight rays of the horizontal projection are used to define all components
of the 2-D paired transform with triplet-numbers of U (1, 0). We define the
control points for these rays by the bullets, from which the rays pass. These
control points are numbered as shown below:
 
. . . . . . . .
 .. .. .. .. .. .. .. .. 
 . . . . . . . . (angle of rays is 90◦).
0 1 2 3 4 5 6 7
• • • • • • • •
We denote the sums of the image along these vertical rays by vt = v1,0 (t),
where t = 0 : 7. It is not difficult to notice that the components of the 2-D
paired transform with these triplet-numbers can be calculated as follows:
0
f1,0,0 = v0 − v4
0
f1,0,1 = v1 − v5
0
f1,0,2 = v2 − v6
0
f1,0,3 = v3 − v7
0 (3.30)
f2,0,0 = (v0 + v4 ) − (v2 + v6 )
0
f2,0,2 = (v1 + v5 ) − (v3 + v7 )
0
f4,0,0 = (v0 + v4 ) − (v1 + v5 ) + (v2 + v6 ) − (v3 + v7 )
0
f0,0,0 = (v0 + v4 ) + (v1 + v5 ) + (v2 + v6 ) + (v3 + v7 ).
Image Processing 123

These calculations correspond to the 8-point discrete paired transform (DPT)

of the vector-measurement v = (v0 , v1 , v2 , ..., v7)0 , which requires only 2N −
2 = 14 operations of addition and subtraction [48]. Thus, we obtain P(1, 0) =
[χ08 ]v. Here P(1, 0) denotes the 2-D discrete paired transform data on the left
0 0
side of the linear system of equations in (3.30), i.e., the vector (f1,0,0 , f1,0,1 , . . .,
0 0 0 0
f2,0,2 , f4,0,0, f0,0,0) . The length of intersection of the geometrical ray l(t) with
the image element is ∆x = 1/8 (see Figure 3.15). Therefore, the line-integrals

l(0) l(4) l(1) l(5)

0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

l(2) l(6) l(3) l(7)

0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m

3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.15
Schemes of calculating the eight line-integrals of the (1, 0)-projection.

equal
1
w(t) = w1,0 (t) =
v(t) = 8v(t), t = 0 : 7,
∆x
and the 2-D paired transform data in the subset U (1, 0) is calculated by
1 0
P(1, 0) = [χ ]w,
8 8
where w denotes the vector-projection (w(0), w(1), w(2), ..., w(7))0.
124 CHAPTER 3: Image Reconstruction

3.5.2 Vertical projection

The vertical projection is considered similarly. The subset of triplet-numbers
(p, s, t) for this projection is


 (0, 1, 0), (0, 1, 1), (0, 1, 2), (0, 1, 3),

(0, 2, 0), (0, 2, 2),
U (0, 1) =

 (0, 4, 0),

(0, 0, 0).

We first consider seven masks of the paired functions with triplet-numbers of

U (0, 1),
2 3 2 3
1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 1 1 1 1 1 1 17
6 7 6 7
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07
7, [χ00,1,1 ] = 6 0 0 0 0 0 0 0 0 7
6 7
[χ00,1,0 ] = 6
6 −1 −1 −1 −1 −1 −1 −1 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 −1 −1 −1 −1 −1 −1 −1 −1 7
6 7 6 7
4 0 0 0 0 0 0 0 05 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 3 2 3
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
6 1 1 1 1 1 1 1 17 6 0 0 0 0 0 0 0 07
6 7 6 7
0
6 0 0 0 0 0 0 0 07 0
6 1 1 1 1 1 1 1 17
[χ0,1,2 ] = 66
7, [χ0,1,3 ] = 6 0 0 0 0 0 0 0 0 7
7 6 7
6 0 0 0 0 0 0 0 07 6 7
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
4 −1 −1 −1 −1 −1 −1 −1 −1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 −1 −1 −1 −1 −1 −1 −1
2 3 2 3
1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 1 1 1 1 1 1 17
6 7 6 7
6 −1 −1 −1 −1 −1 −1 −1 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7, [χ00,2,2 ] = 6 −1 −1 −1 −1 −1 −1 −1 −1 7
6 0 0 0 0 0 0 0 07 6 7
[χ00,2,0 ] = 6
6 1 1 1 1 1 1 1 17 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 1 1 1 1 1 1 1 17
6 7 6 7
4 −1 −1 −1 −1 −1 −1 −1 −1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 −1 −1 −1 −1 −1 −1 −1

2 3
1 1 1 1 1 1 1 1
6 −1 −1 −1 −1 −1 −1 −1 −1 7
6 7
6 1 1 1 1 1 1 1 17
6 7
6 −1 −1 −1 −1 −1 −1 −1 −1 7
[χ00,4,0 ] = 6
6 1 1 1 1 1 1 1 17 .
7
6 7
6 −1 −1 −1 −1 −1 −1 −1 −1 7
6 7
4 1 1 1 1 1 1 1 15
−1 −1 −1 −1 −1 −1 −1 −1
Image Processing 125

The numbered control points of eight rays for this projection are denoted
by the bullets as shown below:
0 
• . . . . . . .
 1• . . . . . . . 
2 
 • . . . . . . .
(angle of rays is 0◦).
 
 .. .. .. .. .. .. .. .. 
 . . . . . . . .
 
 6• . . . . . . . 
7
• . . . . . . .

The 2-D paired transform in the subset of triplets (p, s, t) ∈ U (0, 1) can be
calculated by the 8-point DPT of the vector w, which is composed from the
projection data w(t) = w1,0 (t), t = 0 : 7,

1 0
P(0, 1) = [χ ]w.
8 8

3.5.3 Diagonal projection

Now we consider the third subset of triplet-numbers


 (1, 1, 0), (1, 1, 1), (1, 1, 2), (1, 1, 3),

(2, 2, 0), (2, 2, 2),
U (1, 1) =

 (4, 4, 0),

(0, 0, 0),

and the masks of the corresponding first seven 2-D paired functions
2 3 2 3
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
6 0 0 0 −1 0 0 0 1 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 −1 0 0 0 1 0 0 7
7, [χ01,1,1 ] = 6 0 0 −1 0 0 0 1 0 7
6 7
[χ01,1,0 ] = 6
6 −1 0 0 0 1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 0 0 1 0 0 0 −1 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
4 0 0 1 0 0 0−1 0 5 4 0 0 0 1 0 0 0−1 5
0 1 0 0 0 −1 0 0 0 0 1 0 0 0−1 0
2 3 2 3
0 0 1 0 0 0−1 0 0 0 0 1 0 0 0−1
6 0 1 0 0 0 −1 0 0 7 6 0 0 1 0 0 0−1 0 7
6 7 6 7
6 1 0 0 0 −1 0 0 0 7 6 0 1 0 0 0 −1 0 0 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7
7, [χ01,1,3 ] = 6 1 0 0 0 −1 0 0 0 7
6 7
[χ01,1,2 ] = 6
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 −1 0 0 0 1 0 0 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
4 −1 0 0 0 1 0 0 0 5 4 0 −1 0 0 0 1 0 0 5
0 0 0 1 0 0 0 −1 −1 0 0 0 1 0 0 0
126 CHAPTER 3: Image Reconstruction

2 3 2 3
1 0 −1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
6 0 −1 0 1 0 −1 0 1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
6 −1 0 1 0 −1 0 1 0 7 6 0 −1 0 1 0 −1 0 1 7
6 7 6 7
6 0 1 0 −1 0 1 0 −1 7
7, [χ02,2,2 ] = 6 −1 0 1 0 −1 0 1 0 7
6 7
[χ02,2,0 ] = 6
6 1 0 −1 0 1 0 −1 0 7 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
6 0 −1 0 1 0 −1 0 1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 −1 0 1 0 −1 0 1 5
0 1 0 −1 0 1 0 −1 −1 0 1 0 −1 0 1 0

2 3
1 −1 1 −1 1 −1 1 −1
6 −1 1 −1 1 −1 1 −1 1 7
6 7
6 1 −1 1 −1 1 −1 1 −1 7
6 7
0
6 −1 1 −1 1 −1 1 −1 1 7
[χ4,4,0 ] = 6
6 7.
6 1 −1 1 −1 1 −1 1 −1 7
7
6 −1 1 −1 1 −1 1 −1 1 7
6 7
4 1 −1 1 −1 1 −1 1 −1 5
−1 1 −1 1 −1 1 −1 1
Fifteen parallel rays of the diagonal projection are used to define components
of the 2-D paired transforms with triplet-numbers of U (1, 1). The number of
rays is calculated by (p + s)(N − 1) + 1 = 15, where (p, s) = (1, 1). We define
the control points of these rays by the bullets which are numbered as
0 
• . . . . . . .
 1• . . . . . . . 
2 
 • . . . . . . .
(angle of rays is − 45◦).
 
 .. .. .. .. .. .. .. .. 
 . . . . . . . .
 
 6• . . . . . . . 
7 8 9 10 11 12 13 14
• • • • • • • •

These masks with the rays, on which the coefficients 1 and −1 are situated,
are shown in Figures 3.16-3.18.
The sums of the discrete image along these rays are denoted by vt = v1,1 (t),
where t = 0 : 14. It is not difficult to notice that the components of the 8 × 8-
point 2-D paired transform with the triplet-numbers (p, s, t) ∈ U (1, 1) can be
calculated by
0
f1,1,0 = v0 − v4 + v8 − v12
0
f1,1,1 = v1 − v5 + v9 − v13
0
f1,1,2 = v2 − v6 + v10 − v14
0
f1,1,3 = v3 − v7 + v11 − 0
0 (3.31)
f2,2,0 = (v0 + v4 + v8 + v12 ) − (v2 + v6 + v10 + v14 )
0
f2,2,0 = (v1 + v5 + v9 + v13 ) − (v3 + v7 + v11 + 0)
0
f4,4,0 = v0 − v1 + v2 − v3 + · · · − v11 + v12 − v13 + v14
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · + v11 + v12 + v13 + v14 .
Image Processing 127

l(0) l(4) l(1) l(5)

0 l(8) 0
0 0
0.25 0.25 l(9)
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m
3 3
0.5 l(12) 0.5
4 4
0.625 0.625 l(13)
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

l(2) l(6) l(3) l(7)

0 0
0 0
0.25 0.25
1 1
0.25 l(10) 0.25
2 2
0.375 0.375 l(11)
y and m

y and m
3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 l(14) 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.16
Rays along which the line-integrals are calculated for (p, s) = (1, 1).

Introducing the new variables

X X
an = vn+8m = v1,1,n+8m, n = 0 : 7,
m m

the above system of linear equations can be written as follows:

0
f1,1,0 =a0 − a4
0
f1,1,1 =a1 − a5
0
f1,1,2 =a2 − a6
0
f1,1,3 =a3 − a7
0 (3.32)
f2,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).

These calculations correspond to the 8-point discrete paired transform of the

128 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(1) l(3) l(5) l(7)

0 0
0
l(8) 0
0.25 0.25
1 1
l(9)
0.25 0.25
2
l(10) 2
0.375 0.375
l(11)
y and m

y and m
3 3
0.5 0.5
4
l(12) 4
0.625 0.625
5 5
l(13)
0.75 0.75
6
l(14) 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 3.17
Rays along which the line-integrals are calculated for (p, s) = (2, 2).

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0 l(8)
0
0.25 l(9)
1
0.25 l(10)
2
0.375 l(11)
y and m

3
0.5 l(12)
4
0.625 l(13)
5
0.75 l(14)
6
0.875
7
1
0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1
x and n

FIGURE 3.18
Rays along which the line-integrals are calculated for (p, s) = (4, 4).

vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(1, 1) = [χ08 ]a,

where P(1, 1) denotes the 2-D paired transform data on the left side of sys-
0
tem of equations in (3.32). The calculation of the component f0,0,0 is repeated,
since it is obtained from the subsets U (0, 1) and U (1, 0) as well. Thus, the pro-
cedure of calculating the above components of the 2-D paired transform of the
discrete image fn,m from 15 arithmetical rays is accomplished by the 8-point
paired transform. The sums of the image v(t) = v1,1 (t) can be calculated from
the line-integrals of the image f(x, y) along the corresponding geometrical
Image Processing 129

rays  
t 1
l(t) = l1,1 (t) = (x, y); x + y = + , t = 0 : 14,
8 8
where (x, y) ∈ [0, 1] × [0, 1]. Indeed, as shown in Figure 3.16, the length
√ of in-
tersection
√ of the geometrical ray with the image element equals ∆l = 2∆x =
2/N. Each integral along the geometrical ray is proportional to the sum of
the image along the corresponding arithmetical ray,
v(t) √
w(t) = w1,1 (t) = ∆l 2
= 2 · 8v(t), t = 0 : 14.
(∆x)
√
Therefore, v(t) = w(t)/(8 2), t = 0 : 14.

3.5.4 (2, 1)- and (1, 2)-projections

The next subset of triplet-numbers


 (2, 1, 0), (2, 1, 1), (2, 1, 2), (2, 1, 3),

(4, 2, 0), (4, 2, 2),
U (2, 1) =

 (0, 4, 0),

(0, 0, 0)

can be considered together with the subset



 (1, 2, 0), (1, 2, 1), (1, 2, 2), (1, 2, 3),

(2, 4, 0), (2, 4, 2),
U (1, 2) =

 (4, 0, 0),

(0, 0, 0).

These two symmetric cases with generators (2, 1) and (1, 2) correspond to the
projections by angles π −tan−1 (1/2) and π −tan−1 (2/1), respectively, and are
described by similar equations. We consider in detail the case (p, s) = (2, 1).

3.5.4.1 (2, 1)-projection

Twenty-two rays of the projection are used to define all components of the
2-D paired transforms with triplet-numbers of U (2, 1). The number of rays is
calculated as (p + s)(N − 1) + 1 = 3 · 7 + 1 = 22. We define the control points
of these rays by the bullets and we number them as shown below:
0 
• . . . . . . .
1• . . . . . . .
2 
 • . . . . . . .
 
 .. .. .. .. .. .. .. .. 
 . . . . . . . . (angle of rays is − tan−1 2 = −63.44◦).
 
5• . . . . . . .
 
 6 • 8 • 10 • 12 • 14 • 16 • 18 • 20 • 
7 9 11 13 15 17 19 21
• • • • • • • •
130 CHAPTER 3: Image Reconstruction

The equations of the rays l2,1 (t) on the square [0, 1] × [0, 1] are
 
t 3
l(t) = l2,1 (t) = (x, y); 2x + y = + , t = 0 : 21.
8 16

These rays are shown in Figure 3.19.

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7) l(8) l(9) l(10)l(11)l(12)l(13)l(14)
l(15)
0

0
l(16)
0.25

1
l(17)
0.25

2
l(18)
0.375

3
y and m

l(19)
0.5

4
l(20)
0.625

5
l(21)
0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.19
Rays along which the sums of the image are calculated for (p, s) = (2, 1).

The paired functions with numbers (0, 4, 0) and (0, 0, 0) were considered
for the horizontal and vertical projections. The masks of the paired functions
with the first six triplet-numbers of U (2, 1) are

2 3 2 3
1 0 −1 0 1 0 −1 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
6 0 −1 0 1 0 −1 0 1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7 , [χ02,1,1 ] = 6 0 −1 0 1 0 −1 0 1 7
6 0 0 0 0 0 0 0 07 6 7
[χ02,1,0 ] = 6
6 −1 0 1 0 −1 0 1 0 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 −1 0 1 0 −1 0 1 0 7
6 7 6 7
4 0 1 0 −1 0 1 0 −1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 1 0 −1 0 1 0 −1
Image Processing 131

2 3 2 3
0 1 0 −1 0 1 0 −1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
6 1 0 −1 0 1 0 −1 0 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7 , [χ02,1,3 ] = 6 1 0 −1 0 1 0 −1 0 7
6 0 0 0 0 0 0 0 07 6 7
[χ02,1,2 ] = 6
6 0 −1 0 1 0 −1 0 1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 0 −1 0 1 0 −1 0 1 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 0 1 0 −1 0 1 0
2 3 2 3
1 −1 1 −1 1 −1 1 −1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 −1 1 −1 1 −1 1 −1 7
6 7 6 7
6 −1 1 −1 1 −1 1 −1 1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7 , [χ04,2,2 ] = 6 −1 1 −1 1 −1 1 −1 1 7 .
6 0 0 0 0 0 0 0 07 6 7
[χ04,2,0 ] = 6
6 1 −1 1 −1 1 −1 1 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 1 −1 1 −1 1 −1 1 −1 7
6 7 6 7
4 −1 1 −1 1 −1 1 −1 1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 1 −1 1 −1 1 −1 1

These masks with the rays, on which the coefficients 1 and −1 are situated,
are shown in Figures 3.20 and 3.21.
The sums of the image along these rays are denoted by vt = v2,1 (t), where
t = 0 : 21. It is not difficult to notice that the components of the 2-D paired
transform of the image with the triplets of U (2, 1) can be calculated by
0
f2,1,0 = v0 − v4 + v8 − v12 + v16 − v20
0
f2,1,1 = v1 − v5 + v9 − v13 + v17 − v21
0
f2,1,2 = v3 − v7 + v11 − v15 + v19 − 0
0
f2,1,3 = v2 − v6 + v10 − v14 + v18 − 0
0 (3.33)
f4,2,0 = (v0 + v4 + v8 + v12 + v16 + v20 ) − (v2 + v6 + v10 + v14 + v18 )
0
f4,2,2 = (v1 + v5 + v9 + v13 + v17 + v21 ) − (v3 + v7 + v11 + v15 + v19 )
0
f0,4,0 = v0 − v1 + v2 − v3 + · · · + v18 − v19 + v20 − v21
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · + v18 + v19 + v20 + v21 .

Introducing the new variables

X X
an = vn+8m = v2,1,n+8m, n = 0 : 7,
m m

the above system of equations can be written as follows:

0
f2,1,0 =a0 − a4
0
f2,1,1 =a1 − a5
0
f2,1,2 =a2 − a6
0
f2,1,3 =a3 − a7
0 (3.34)
f4,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f4,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f0,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
132 CHAPTER 3: Image Reconstruction

l(0) l(4) l(8) l(12) l(1) l(5) l(9) l(13)

0 0
0 0
0.25 l(16) 0.25
1 1
0.25 0.25 l(17)
2 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 l(20) 0.625
5 5
0.75 0.75 l(21)
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

l(2) l(6) l(10) l(14) l(3) l(7) l(11)

0 0 l(15)
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 l(18) 0.375
y and m

y and m
3 3
0.5 0.5 l(19)
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.20
Rays along which the sums of the image are calculated for (p, s) = (2, 1).

Again we can see that the calculation of the 2-D paired transform data
P(2, 1) corresponds to the 8-point discrete paired transform of the vector
a = (a0 , a1 , a2 , ..., a7)0 , i.e., P(2, 1) = [χ08 ]a.
Now we express the sums v(t) of the discrete image fn,m through the
line-integrals w(t), assuming first that these integrals are calculated along the
geometrical rays that coincide with the arithmetical rays l(t), i.e., w(t) =
w2,1 (t), t = 0 : 21. As an example, we consider the ray number seven, l(7).
As can be seen from Figure 3.22, this ray crosses the IE of number (0, 7) and
then two IE of numbers (0, 6) and (1, 6), and such intersections are repeated
on its way to the IE of number (4, 0). The length of the intersection ∆l0,7 of
the ray with the IE of number (0, 7) is twice the length of intersection with
each IE of number (0, 6) and (1, 6).
Image Processing 133

l(0) l(2) l(4) l(6) l(8) l(10) l(12) l(14) l(1) l(3) l(5) l(7) l(9) l(11) l(13)
0 0 l(15)
0 0
0.25 l(16) 0.25
1 1
0.25 0.25 l(17)
2 2
0.375 l(18) 0.375
y and m

y and m
3 3
0.5 0.5 l(19)
4 4
0.625 l(20) 0.625
5 5
0.75 0.75 l(21)
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 3.21
Rays along which the sums of the image are calculated for (p, s) = (4, 2).

l(0) l(4) l(7) l(8) l(9) l(12)

0
l(16)
0.25

0.25

0.375

3
y and m

0.5

4
l(20)
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.22 (See color insert)

Mask of the paired function χ02,1,0 (n, m) and rays.

√
Since ∆l0,7 =
5/2∆x, we can write the following:
  √  
∆l0,7 1 1 5 1 1
w(7) = w2,1 (7)= v(7) + v(6) + v(8) = 8 v(6) + v(7) + v(8) .
(∆x)2 2 2 2 2 2
134 CHAPTER 3: Image Reconstruction

For the line-integral along the ray l(8), we also can write
   
√ 1 1 √ 1 1
w(8) = w2,1 (8) = 4 5 v(7) + v(9) + v(8) = 4 5 v(7) + v(8) + v(9) .
2 2 2 2
Similar equations can be used for the integrals along all these geometrical
rays,
 
√ 1 1
w(t) = w2,1 (t) = 4 5 v(t − 1) + v(t) + v(t + 1) , t = 0 : 21.
2 2
We consider two rays which are between rays l(7), l(8) and l(8), l(9) and
call them l̃(8) and ˜ l(9), respectively. One can notice that new ray number 8
intersects equally the IE of number (0, 7), the IE of number (1, 6) and other
the IE, i.e., the length
√ of intersection of this ray with all these IEs is the same
and equals ∆l0,7 = 5/2∆x. New ray number 9 also equally intersects image
elements. The line-integrals along these two rays can be defined as
√ √
w(8) = w(˜ l(8)) = 4 5 [v(7) + v(8)] , w(9) = w(˜l(9)) = 4 5 [v(8) + v(9)] .
Thus we can simplify the calculations, by considering the shifting, t → t−1/2,
and the following new set of geometrical rays:
 
˜ t 1
l(t) = l2,1 (t − 1/2) = (x, y); 2x + y = + , t = 0 : 21,
8 8
which are shown in Figure 3.23. For this set of geometrical rays, we can define
the line-integrals along the rays as follows:
√
˜
w(t) = w(l(t)) = 4 5 [v(t − 1) + v(t)] , t = 0 : 21,
where v(−1) = 0. This system of equations can be written in matrix form
 
1 0 0 0 ... 0 0 0
 1 1 0 0 ... 0 0 0 
 
√  0 1 1 0 ... 0 0 0 
 
√
w = 4 5Av = 4 5  0 0 1 1 ... 0 0 0 

 v, (3.35)
 ........................ 
 
 0 0 0 0 ... 1 1 0 
0 0 0 0 ... 0 1 1
where w = (w(0), w(1), ..., w(21))0 and v = (v0 , v1 , ..., v21)0 . The above
Toeplitz matrix 22 × 22 has the triangle inverse matrix, which is similar to
the matrix described in (3.19) for the N = 4 example. The inverse matrix is
triangular and consists of unit diagonals whose signs alternate,
 
1 0 0 0 ... 0 0 0
 −1 1 0 0 ... 0 0 0 
 
 1−1 1 0 ... 0 0 0 
 
A−1 =   −1 1 −1 1 ... 0 0 0  .
 (3.36)
 ... ... ... ... ... ... ......
 
 1−1 1 −1 ...−1 1 0 
−1 1 −1 1 ... 1 −1 1
Image Processing 135
l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7) l(8) l(9) l(10)l(11)l(12)l(13)l(14) l(15)

0 l(16)

0.25

1 l(17)

0.25

2 l(18)

0.375

3 l(19)
y and m

0.5

4 l(20)

0.625

5 l(21)

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.23
New set of geometrical rays for calculating line-integrals for (p, s) = (2, 1).

The equation b = A−1 w can thus be written in the following recurrent

form:
b0 =w(0),
(3.37)
bt =w(t) − bt−1 , t = 1, 2, . . . , 21,
√
and, therefore, v = b/(4 5), t = 0 : 21. In other words, the solution v is
calculated by the method of forward substitution.

Remark 3.1 We also can consider another set of geometrical rays, by using
the shifting, t → t + 1/2, of the original arithmetical rays,
 
˜ t 1
l(t) = l2,1 (t + 1/2) = (x, y); 2x + y = + , t = 0 : 21,
8 4

which is shown in Figure 3.24. One can notice that geometrical ray number 7
intersects equally another pair of IEs of number (0, 7) and (1, √
6). Therefore, we
can define the integral along this ray as w(7) = w(˜l(7)) = 4 5 [v(7) + v(8)] .
In general, the integrals along the geometrical rays are defined as follows:
√
w(t) = w(˜l(t)) = 4 5 [v(t) + v(t + 1)] , t = 0 : 21,

considering v(22) = 0. With the exception of the IE of number (0, 0), each
other IE is intersected by two parallel geometrical rays. This system of linear
136 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7) l(8) l(9) l(10)l(11)l(12)l(13)l(14)

0 l(15)

0.25

1 l(16)

0.25

2 l(17)

0.375

3 l(18)
y and m

0.5

4 l(19)

0.625

5 l(20)

0.75

6 l(21)

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.24
New set of geometrical rays for calculating line-integrals for (p, s) = (2, 1).

equations can be written in matrix form

 
1 1 0 0 ... 0 0 0
 0 1 1 0 ... 0 0 0 
 
√  0 0 1 1 ... 0 0 0 
 
√
w = 4 5Av = 4 5  0 0 0 1 ... 0 0 0 

 v. (3.38)
 ........................ 
 
 0 0 0 0 ... 0 1 1 
0 0 0 0 ... 0 0 1

The above Toeplitz matrix 22 × 22 has an upper triangle inverse matrix that
is the transposition of the inverse matrix (3.36), which was used for the first
set of geometrical rays. The matrix is triangular and consists of unit diagonals
whose signs alternate,
 
1 −1 1−1 ...−1 ... 1 −1
 0 1 −1 1 −1 ......−1 1 
 
 0 0 1−1 1 −1 ... 1 −1 
 
−1
 0 0 0 1 −1 1 ...−1 1 
A =  . (3.39)
 0 0 0 0 1 −1 ... ...−1 

 ... ... ... ... ... ...... ... ... 
 
 0 0 0 0 0 0 ... 1 −1 
0 0 0 0 0 0 ... 0 1
Image Processing 137

The coefficients of this matrix are

n
X
Am,n = (−1)m−n δm−n,t .
k=0

The equation b = A−1 w can be written in the following recurrent form:

b21 =w(21),
(3.40)
bt =w(t) − bt+1 , t = 20, 19, . . ., 0,
√
and, therefore, v = b/(4 5), t = 0 : 21. Thus, the solution v is calculated by
the method of back substitution.

3.5.4.2 (1, 2)-projection

We now consider the rays of the projection that corresponds to the generator
(p, s) = (1, 2). Twenty-two arithmetical rays of this projection which are used
to define all components of the 2-D paired transforms with triplet-numbers of
U (1, 2) are shown in Figure 3.25. The set of 22 geometrical rays that coincide

l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0
l(7)

0 l(8)

0.25
l(9)

1
l(10)

0.25
l(11)

2
l(12)

0.375
l(13)

3 l(14)
y and m

0.5
l(15)

4 l(16)

0.625
l(17)

5 l(18)

0.75
l(19)

6 l(20)

0.875
l(21)

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.25
Set of arithmetical rays for calculating line-sums for (p, s) = (1, 2).

with these arithmetical rays are defined as

 
t 3
l(t) = l1,2 (t) = (x, y); x + 2y = + , t = 0 : 21,
8 16
138 CHAPTER 3: Image Reconstruction

where (x, y) ∈ [0, 1] × [0, 1].

The masks of the paired functions with the first six triplet-numbers of
U (1, 2) are
   
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
 0 0−1 0 0 0 1 0   0 0 0 −1 0 0 0 1 
   
 −1 0 0 0 1 0 0 0   0 −1 0 0 0 1 0 0 
   
0
 0 0 1 0 0 0 −1 0  0
 0 0 0 1 0 0 0 −1 
[χ1,2,0] = 

 , [χ1,2,1] =  0 1 0 0 0 −1 0 0  ,
  
 1 0 0 0 −1 0 0 0   
 0 0−1 0 0 0 1 0   0 0 0 −1 0 0 0 1 
   
 −1 0 0 0 1 0 0 0   0 −1 0 0 0 1 0 0 
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1
   
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1
 1
 0 0 0 −1 0 0 0
0 1
 0 0 0 −1 0 0 
 0
 0−1 0 0 0 1 0
0 0
 0 −1 0 0 0 1 
0
 −1 0 0 0 1 0 0 0 0
 0 −1 0 0 0 1 0 0
[χ1,2,2] =  , [χ1,2,3] =  ,
 0
 0 1 0 0 0 −1 0
0 0
 0 1 0 0 0 −1 

 1
 0 0 0 −1 0 0 0
0 1
 0 0 0 −1 0 0 
 0 0−1 0 0 0 1 0 0 0 0 −1 0 0 0 1
−1 0 0 0 1 0 0 0 0 −1 0 0 0 1 0 0
   
1 0−1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
 −1 0 1 0 −1 0 1 0  0 −1 0 1 0 −1 0 1
   
 1
 0−1 0 1 0 −1 0
0 1
 0 −1 0 1 0 −1 

 −1 0 1 0 −1 0 1 0  0 −1 0 1 0 −1 0 1
[χ02,4,0] =  , [χ02,4,2] =  .
 1
 0−1 0 1 0 −1 0
0 1
 0 −1 0 1 0 −1 

 −1 0 1 0 −1 0 1 0  0 −1 0 1 0 −1 0 1
   
 1 0−1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1 
−1 0 1 0 −1 0 1 0 0 −1 0 1 0 −1 0 1

These masks with the rays on which the coefficients 1 and −1 are situated,
are shown in Figures 3.26 and 3.27. We define and number the control points
for these rays by the bullets as shown below:
0 1 2 3 4 5 6 7 
• • • • • • • •
 . . . . . . 8• 9• 
 
 . . . . . . 10 • 11 • 
 
 . . . . . . 12 • 13 •  −1 ◦
 . . . . . . 14 • 15 •  (angle of rays is − tan 1/2 = −26.565 ).
 
 
 . . . . . . 16 • 17 • 
 
 . . . . . . 18 • 19 • 
. . . . . . 20 • 21 •

The components of the paired transform with the triplet-numbers of the

Image Processing 139

l(0) l(4) l(1) l(5)

0 0
0 l(8) 0
0.25 0.25 l(9)
1 1
0.25 0.25
2 l(12) 2
0.375 0.375 l(13)
y and m

y and m
3 3
0.5 0.5
4 l(16) 4
0.625 0.625 l(17)
5 5
0.75 0.75
6 l(20) 6
0.875 0.875 l(21)
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

l(2) l(6) l(3)

0 0 l(7)
0 0
0.25 0.25
1 l(10) 1
0.25 0.25 l(11)
2 2
0.375 0.375
y and m

y and m
3 l(14) 3
0.5 0.5 l(15)
4 4
0.625 0.625
5 l(18) 5
0.75 0.75 l(19)
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

FIGURE 3.26
Arithmetical rays along which the line-sums are calculated for (p, s) = (1, 2).

subset U (1, 2) are calculated by

0
f1,2,0 = v0 − v4 + v8 − v12 + v16 − v20
0
f1,2,1 = v1 − v5 + v9 − v13 + v17 − v21
0
f1,2,2 = v2 − v6 + v10 − v14 + v18
0
f1,2,3 = v3 − v7 + v11 − v15 + v19
0 (3.41)
f2,4,0 = (v0 + v4 + v8 + v12 + v16 + v20 ) − (v2 + v6 + v10 + v14 + v18 )
0
f2,4,2 = (v1 + v5 + v9 + v13 + v17 + v21 ) − (v3 + v7 + v11 + v15 + v19 )
0
f4,0,0 = v0 − v1 + v2 − v3 + · · · + v18 − v19 + v20 − v21
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · + v18 + v19 + v20 + v21 ,

where variables vt = v1,2 (t) denote the sums of the discrete image along the
140 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(1) l(3) l(5)

0 0 l(7)
0 l(8) 0
0.25 0.25 l(9)
1 l(10) 1
0.25 0.25 l(11)
2 l(12) 2
0.375 0.375 l(13)
l(14)
y and m

y and m
3 3
0.5 0.5 l(15)
4 l(16) 4
0.625 0.625 l(17)
5 l(18) 5
0.75 0.75 l(19)
6 l(20) 6
0.875 0.875 l(21)
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.27
Arithmetical rays along which the line-sums are calculated for (p, s) = (2, 4).

rays l(t), t = 0 : 21. The system of equations can also be written as

0
f1,2,0 =a0 − a4
0
f1,2,1 =a1 − a5
0
f1,2,2 =a2 − a6
0
f1,2,3 =a3 − a7
0 (3.42)
f2,4,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,4,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,0,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ),

where an are calculated from the sums vt = v1,2 (t) of the discrete image by
X X
an = vn+8m = v1,2,n+8m, n = 0 : 7.
m m

The 8-point discrete paired transform of the vector a = (a0 , a1 , a2 , ..., a7)0
results in the 2-D paired transform data at points of U (1, 2),

P(1, 2) = [χ08 ]a.

As in the (p, s) = (2, 1) case, to calculate the vector a from the projection-
data, we will simplify the equation describing the linear relation between the
line-integrals w(t) and sums v(t), t = 0 : 21, by considering another set of
22 geometric rays. For that, we first consider the mask of the paired func-
tion χ01,2,0(n, m) with coefficients ±1 of the rays l(0), l(4), l(8), l(12), l(16),
and l(20) which are shown in Figure 3.28. One can see that the geometrical
ray l(8) intersects three image elements when x runs two intervals, 2∆x, along
Image Processing 141

l(0) l(4)
0
l(7)

0 l(8)

0.25
l(9)
1

0.25

2
l(12)

0.375

3
y and m

0.5

4 l(16)

0.625

0.75

6 l(20)

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.28
The mask of number (1, 2, 0) and rays.

the horizontal. Therefore, the integral w(8) = w1,2(8) can be calculated by us-
ing three sums, v(7), v(8), and v(9). On other hand, the integral of the image
along the geometrical ray between the rays l(8) and l(9) can be calculated
from the sums of the image along two arithmetical rays by
√
w = 5/2 · 8 [v(8) + v(9)] .
This geometric ray is considered as ray number 8 in the new set of parallel
geometrical rays, which are defined as
 
t 1
l̃(t) = l1,2 (t + 0.5) = (x, y); x + 2y = + , t = 0 : 21,
8 4
and illustrated in Figure 3.29.
The relation between the measurements along the geometrical rays ˜l(t) and
arithmetical rays l(t) is described by the following system of linear equations:
√
w(t) = w(˜ l(t)) = 4 5 [v(t) + v(t + 1)] , t = 0 : 21,
where v(22) = 0. This system of equations was considered in matrix form in
(3.38) and (3.39), for the (p, s) = (2, 1) case. The
√ required sums v(t) of the
discrete image fn,m are calculated by v(t) = bt /4 5, where the coefficients bt
can be calculated in the following recurrent form:
b21 = w(21),
(3.43)
bt = w(t) − bt+1 , t = 20, 19, . . . , 1, 0.
142 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0
l(7)
0
l(8)
0.25
l(9)
1
l(10)
0.25
l(11)
2
l(12)
0.375
l(13)
3
l(14)
y and m

0.5
l(15)
4
l(16)
0.625
l(17)
5
l(18)
0.75
l(19)
6
l(20)
0.875
l(21)
7

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.29
New set of geometrical rays for calculating line-integrals for (p, s) = (1, 2).

Remark 3.2 Another set of shifted parallel geometrical rays can also be con-
sidered. As an example, we consider the rays defined as
 
t 1
l̃(t) = l1,2 (t − 0.5) = (x, y); x + 2y = + , t = 0 : 21,
8 8

which are illustrated in Figure 3.30.

The relation between the measurements along the geometrical rays ˜l(t) and
arithmetical rays l(t) is described by the following system of linear equations:
√
˜ 5
w(t) = w(l(t)) = · 8 [v(t) + v(t − 1)] , t = 0 : 21,
2
where v(−1) = 0. This system of equations was considered in matrix form
in (3.35) and (3.36), for the (p, s) = (2, 1) case.√Therefore, the required sums
v(t) of the image are calculated by v(t) = bt /4 5, and the coefficients bt can
be calculated in the following recurrent form:

b0 =w(0),
(3.44)
bt =w(t) − bt−1 , t = 1, 2, . . . , 20, 21.
Image Processing 143

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0
l(8)
0
l(9)
0.25
l(10)
1
l(11)
0.25
l(12)
2
l(13)
0.375
l(14)
3
l(15)
y and m

0.5
l(16)
4
l(17)
0.625
l(18)
5
l(19)
0.75
l(20)
6
l(21)
0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.30
New set of geometrical rays for calculating line-integrals for (p, s) = (1, 2).

3.5.5 (1, 3)-projection

We now consider the subset of triplet-numbers U (1, 3),


 (1, 3, 0), (1, 3, 1), (1, 3, 2), (1, 3, 3),

(2, 6, 0), (2, 6, 2),
U (1, 3) =

 (4, 4, 0),

(0, 0, 0).

Since the masks of the paired functions χ04,4,0 and χ00,0,0 were given above, we
consider functions with the first six triplet-numbers of the subset U (1, 3),
2 3 2 3
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
6 0 −1 0 0 0 1 0 0 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
6 0 0 1 0 0 0−1 0 7 6 0 0 0 1 0 0 0−1 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7
7 , [χ01,3,1 ] = 6 1 0 0 0 −1 0 0 0 7
6 7
[χ01,3,0 ] = 6
6 −1 0 0 0 1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7 6 0 0 1 0 0 0 −1 0 7
6 7 6 7
4 0 0 −1 0 0 0 1 0 5 4 0 0 0 −1 0 0 0 1 5
0 0 0 1 0 0 0−1 −1 0 0 0 1 0 0 0
144 CHAPTER 3: Image Reconstruction

2 3 2 3
0 0 1 0 0 0−1 0 0 0 0 1 0 0 0−1
6 0 0 0 −1 0 0 0 1 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
6 −1 0 0 0 1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7
7 , [χ01,3,3 ] = 6 0 0 1 0 0 0 −1 0 7
6 7
[χ01,3,2 ] = 6
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 0 0 1 0 0 0−1 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
4 1 0 0 0 −1 0 0 0 5 4 0 1 0 0 0 −1 0 0 5
0 −1 0 0 0 1 0 0 0 0 −1 0 0 0 1 0
2 3 2 3
1 0 −1 0 1 0 −1 0 0 −1 0 1 0 −1 0 1
6 0 1 0 −1 0 1 0 −1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
6 −1 0 1 0 −1 0 1 0 7 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
6 0 −1 0 1 0 −1 0 1 7
7 , [χ02,6,2 ] = 6 −1 0 1 0 −1 0 1 0 7
6 7
[χ02,6,0 ] = 6
6 1 0 −1 0 1 0 −1 0 7 6 0 −1 0 1 0 −1 0 1 7
6 7 6 7
6 0 1 0 −1 0 1 0 −1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 1 0 −1 0 1 0 −1 5
0 −1 0 1 0 −1 0 1 −1 0 1 0 −1 0 1 0

The masks of these paired functions are shown in Figures 3.31 and 3.32,
together with the rays the pass through the coefficients ±1 of the masks.
The set of 29 geometrical rays of this projection,
 
t 1
l(t) = l1,3 (t) = (x, y); x + 3y = + , t = 0 : 28,
8 4

where (x, y) ∈ [0, 1] × [0, 1], is shown in Figures 3.33.

We define the control points for these rays as shown below:
0 1 2 3 4 5 6 7 
• • • • • • • •
 . . . . . 8 • 9 • 10 • 
 
 . . . . . 11 • 12 • 13 • 
 
 . . . . . 14 • 15 • 16 •  −1 ◦
 . . . . . 17 • 18 • 19 •  (angle of rays is − tan (1/3) = −18.435 ).
 
 
 . . . . . 20 • 21 • 22 • 
 
 . . . . . 23 • 24 • 25 • 
. . . . . 26 • 27 • 28 •

The components of the paired transform with the triplet-numbers of the

subset U (1, 3) are calculated as follows:
0
f1,3,0 = v0 − v4 + v8 − v12 + v16 − v20 + v24 − v28 ,
0
f1,3,1 = v1 − v5 + v9 − v13 + v17 − v21 + v25 ,
0
f1,3,2 = v2 − v6 + v10 − v14 + v18 − v22 + v26 ,
0
f1,3,3 = v3 − v7 + v11 − v15 + v19 + v23 + v27 ,
Image Processing 145

l(0) l(4) l(1) l(5)

0 0
0 l(8) 0 l(9)
0.25 0.25
1 l(12) 1
0.25 0.25 l(13)
2 2
0.375 l(16) 0.375
y and m

y and m
3 3 l(17)
0.5 0.5
4 l(20) 4 l(21)
0.625 0.625
5 l(24) 5
0.75 0.75 l(25)
6 6
0.875 l(28) 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

l(2) l(6) l(3)

0 0 l(7)
0 0
0.25 l(10) 0.25
1 1 l(11)
0.25 0.25
2 l(14) 2 l(15)
0.375 0.375
y and m

y and m
3 l(18) 3
0.5 0.5 l(19)
4 4
0.625 l(22) 0.625
5 5 l(23)
0.75 0.75
6 l(26) 6 l(27)
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

FIGURE 3.31
Masks of the paired functions with triplet-numbers (1, 3, t), t = 0, 1, 2, 3.

0
f2,6,0 = (v0 + v4 + v8 + v12 + v16 + v20 + v24 + v28 )
−(v2 + v6 + v10 + v14 + v18 + v22 + v26 ),
0
f2,6,2 = (v1 + v5 + v9 + v13 + v17 + v21 + v25 )
−(v3 + v7 + v11 + v15 + v19 + v23 + v27 ),
0
f4,4,0 = v0 − v1 + v2 − v3 + · · · + v18 − v19 + v20 − v21 ,
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · + v18 + v19 + v20 + v21 ,
where variables vt = v1,3 (t) denote the sums of the discrete image along the
146 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(1) l(3) l(5)

0 0 l(7)
0 l(8) 0 l(9)
0.25 l(10) 0.25
1 1 l(11)
l(12)
0.25 0.25 l(13)
2 l(14) 2 l(15)
0.375 l(16) 0.375
l(17)
y and m

y and m
3 l(18) 3
0.5 0.5 l(19)
4 l(20) 4 l(21)
0.625 l(22) 0.625
5 5 l(23)
l(24)
0.75 0.75 l(25)
6 l(26) 6 l(27)
0.875 l(28) 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.32
Masks of the paired functions with triplet-numbers (2, 6, t), t = 0, 2.

l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0
l(7)
0 l(8)
0.25
l(9)
l(10)
1 l(11)
0.25
l(12)
l(13)
2 l(14)
0.375
l(15)
l(16)
3 l(17)
y and m

0.5
l(18)
l(19)
4 l(20)
0.625
l(21)
l(22)
5 l(23)
0.75
l(24)
l(25)
6 l(26)
0.875
l(27)
l(28)
7

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.33
The set of arithmetical rays for the (1, 3)-projection.

rays l(t), t = 0 : 28. The system of equations can also be written as

0
f1,3,0 =a0 − a4
0
f1,3,1 =a1 − a5
0
f1,3,2 =a2 − a6
0
f1,3,3 =a3 − a7
0 (3.45)
f2,6,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,6,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 )
Image Processing 147

where an are calculated from the sums vt = v1,3 (t) of the discrete image by
X X
an = vn+8m = v1,3,n+8m, n = 0 : 7.
m m

The 8-point discrete paired transform of the vector a = (a0 , a1 , a2 , ..., a7)0
results in the 2-D paired transform data at points of U (1, 3), i.e., P(1, 3) =
[χ08 ]a. To calculate the vector a from the projection data, we will define a set
of geometrical rays that allow calculation of the integrals w(t) by means of
the sums v(t), t = 0 : 28. For that, we first consider a few arithmetical rays as
the geometrical rays in the part of the square [0, 1] × [0, 1], which are shown
in Figure 3.34. One can notice that each of five rays equally intersects image

0.5
l(19)
4 l(20)
l(21)
0.625
l(22)
5 l(23)

0.75
y and m

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.34
Five rays of the projection corresponding (1, 3).

elements,√namely the length of intersection of the ray with each IE equals

∆l0,7 = 10/3∆x. The line-integrals along the geometrical rays 20, 21, and
22 can be calculated by the sums of the discrete image as follows:
√
w(20) = w1,3(20)= 8√10/3 [v(21) + v(19) + v(20)]
= 8√10/3 [v(19) + v(20) + v(21)]
w(21) = w1,3(21)= 8√10/3 [v(21) + v(22) + v(20)]
= 8√10/3 [v(20) + v(21) + v(22)]
w(22) = w1,3(22)= 8 10/3 [v(21) + v(22) + v(23)] .
148 CHAPTER 3: Image Reconstruction

Similar equations hold for all integrals w(t) of the considered projection
√
w(t) = w1,3 (t) = 8 10/3 [v(t − 1) + v(t) + v(t + 1)] , t = 0 : 28.

In matrix form, this system of linear equations is determined by the

Toeplitz matrix 29 × 29,
 
1 1 0 0 0 ... 0 0 0
 1 1 1 0 0 ... 0 0 0 
 
 0 1 1 1 0 ... 0 0 0 
 
 0 0 1 1 1 ... 0 0 0  ,
A=  (3.46)
 ........................... 
 
 0 0 0 0 0 ... 1 1 1 
0 0 0 0 0 ... 0 1 1

which is not triangular and has determinant 0. In other words, the given set
of geometrical rays cannot be used to define all required sums v(t), t = 0 : 21,
for the considered projection. Therefore, as for the projections defined by
frequency-points (p, s) = (2, 1) and (1, 2), we consider other sets of rays. The
simple way to select such a set is to remove the first ray from the set and add
one parallel new ray at the end, or remove the last ray from the set and add
one new parallel ray at the beginning. We consider the first case.
Case 1: The set of 29 geometrical rays is defined as (see Figure 3.35)
 
t 3
l̃(t) = l1,3 (t + 1) = (x, y); x + 3y = + , t = 0 : 28,
8 8

and the integrals along these rays are denoted by w(t) = w(˜l(t)).
One can see, for instance, that ˜
√ w(l(1)) = w(l(2)) = v(1) + v(2) + v(3),
up to the normalized coefficient 8 10/3. The system of equations for all new
integrals w(t) of the considered projection can be defined as follows:
√
w(t) = w(˜ l(t)) = 8 10/3 [v(t) + v(t + 1) + v(t + 2)] , t = 0 : 28,

where it is assumed that v(29) = v(30) = 0. This system of equations can be

written in matrix form as
 
1 1 1 0 ... 0 0 0
 0 1 1 1 ... 0 0 0 
 
√ √  0 0 1 1 ... 0 0 0 
8 10 8 10  
w= Av = 0 0 0 1 ... 0 0 0  v. (3.47)
3 3  
 ........................ 

 0 0 0 0 ... 0 1 1 
0 0 0 0 ... 0 0 1
Image Processing 149

l(0) l(1) l(2) l(3) l(4) l(5)

0
l(6)
0 l(7)
0.25
l(8)
l(9)
1 l(10)
0.25
l(11)
l(12)
2 l(13)
0.375
l(14)
l(15)
3 l(16)
y and m

0.5
l(17)
l(18)
4 l(19)
0.625
l(20)
l(21)
5 l(22)
0.75
l(23)
l(24)
6 l(25)
0.875
l(26)
l(27)
7 l(28)
1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.35
The set of geometrical rays for the projection defined by (1, 3).

The above Toeplitz matrix 29 × 29 has the upper triangle inverse matrix
 
1 −1 0 1 −1 0 1 −1 ... 0 1 −1
 0 1 −1 0 1 −1 0 1 ...−1 0 1 
 
 0 0 1 −1 0 1 −1 0 ... 1 −1 0 
 
 0 0 0 1 −1 0 1 −1 ... 0 1 −1 
 
−1
 0 0 0 0 1 −1 0 1 ...−1 0 1 
A =  . (3.48)
 . . . . . . . . ... . . . 

 0 0 0 0 0 0 0 1 ...−1 0 1 
 
 0 0 0 0 0 0 0 0 ... 1 −1 0 
 
 0 0 0 0 0 0 0 0 ... 0 1 −1 
0 0 0 0 0 0 0 0 ... 0 0 1

The inverse transform b = A−1 w can be calculated by the following recurrent

form:
b28 =w(28),
b27 =w(27) − b28 ,
b26 =w(26) − (b27 + b28 ), (3.49)
b25 =w(25) − (b26 + b27 ),
bt =w(t) − (bt+1 + bt+2 ), t = 24, 23, . . ., 1, 0.
150 CHAPTER 3: Image Reconstruction

The √ required sums v(t) of the discrete image are calculated by v(t) =
bt /(8 10/3), where t = 0 : 28.
Case 2: The set of 29 geometrical rays can also be defined as (see Fi-
gure 3.36)
 
t 1
l̃(t) = l1,3 (t − 1) = (x, y); x + 3y = + , t = 0 : 28.
8 8

The integrals along these rays are denoted by w(t) = w(˜l(t)).

l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0
l(7)
l(8)
0 l(9)
0.25
l(10)
l(11)
1 l(12)
0.25
l(13)
l(14)
2 l(15)
0.375
l(16)
l(17)
3 l(18)
y and m

0.5
l(19)
l(20)
4 l(21)
0.625
l(22)
l(23)
5 l(24)
0.75
l(25)
l(26)
6 l(27)
0.875
l(28)

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.36
Second set of geometrical rays for the (1, 3)-projection.

In this case, the line-integrals w(t) are defined by the sums v(t) as
√
w(t) = w(˜l(t)) = 8 10/3 [v(t) + v(t − 1) + v(t − 2)] , t = 0 : 28. (3.50)

In matrix form, this system of linear equations is described by the triangular

Image Processing 151

Toeplitz matrix 29 × 29 with left band width 2,

 
1 0 0 0 0 ... 0 0 0
 1 1 0 0 0 ... 0 0 0 
 
 1 1 1 0 0 ... 0 0 0 
 
A=  0 1 1 1 0 ... 0 0 0  , (det(A) = 1). (3.51)

 ........................... 
 
 0 0 0 0 0 ... 1 1 0 
0 0 0 0 0 ... 1 1 1

The inverse matrix equals

 
1 0 0 0 0 0 ... 0 0 0 0 0
 −1 1 0 0 0 0 ... 0 0 0 0 0
 
 0 −1 1 0 0 0 ... 0 0 0 0 0
 
 1 0−1 1 0 0 ... 0 0 0 0 0
 
A−1  −1 1 0 −1 1 0 ... 0
= 0 0 0 0. (3.52)
 0 −1 1 0 −1 1 ... 0 0 0 0 0
 
 . . . . . .... . . . . .
 
 1 0−1 1 0 −1 ... 1 0 −1 1 0
−1 1 0 −1 1 0 ...−1 1 0 −1 1

The inverse transform b = A−1 w can be calculated in the following recurrent

form:
b0 =w(0),
b1 =w(1) − b0 ,
b2 =w(2) − (b1 + b0 ), (3.53)
b3 =w(3) − (b2 + b1 ),
bt =w(t) − (bt−1 + bt−2 ), t = 4, 5, . . . , 27, 28.
The required
√ sums v(t) of the discrete image are calculated by v(t) =
3bt /(8 10), where t = 0 : 28.

Remark 3.3 Instead of the subset U (1, 3), we can consider the subset U (3, 1),


 (3, 1, 0), (3, 1, 1), (3, 1, 2), (3, 1, 3),

(6, 2, 0), (6, 2, 2),
U (3, 1) =

 (4, 4, 0),

(0, 0, 0).

Indeed, it is not difficult to see that the paired functions generated by the fre-
quency (p, s) = (3, 1) are χ03,1,t = ±χ01,3,t mod 4 , where the sign is determined
by sign(4 − t mod 8), for t = 0, 1, 2, 3. These two points (1, 3) and (3, 1) gener-
ate the same group T1,3 = T3,1 in the frequency domain. It also means that all
1’s and all −1’s are located on the parallel rays at the angles π − tan−1 (1/3)
and π − tan−1 (3/1) to the horizontal. As an example, Figure 3.37 shows the
mask of the function χ03,1,0 = χ01,3,0 and two sets of parallel rays, along which
the coefficients ±1 of the function are located. For the (3, 1)-projection, the
152 CHAPTER 3: Image Reconstruction

l(0) l(4) l(8) l(12) l(16) l(20) l(0) l(4)

0 0
0 0 l(8)
0.25 l(24) 0.25
1 1 l(12)
0.25 0.25
2 2
0.375 0.375 l(16)
y and m

y and m
3 3
0.5 0.5
4 4 l(20)
0.625 l(28) 0.625
5 5 l(24)
0.75 0.75
6 6
0.875 0.875 l(28)
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.37
Two sets of rays on the mask of the paired function with number (3, 1, 0).

control points of the second set of rays are defined as shown below:
0 
• . . . . . . .
1 • . . . . . . .
2 
 • . . . . . . .
3 
 • . . . . . . . −1 ◦
 • . . . . . . .  (angle of rays is π − tan (3) = 108.435 ).
4 
 5 8 11 14 17 20 23 26 
 • • • • • • • •
 6 9 12 15 18 21 24 27 
 • • • • • • • •
7 10 13 16 19 22 25 28
• • • • • • • •

The set of the 29 geometrical rays of this projection,

 
t 1
l(t) = l3,1 (t) = (x, y); 3x + y = + , t = 0 : 28,
8 4

where (x, y) ∈ [0, 1] × [0, 1], is shown in Figures 3.38.

Below are six masks for the first six triplet-numbers of the subset U (3, 1),
2 3 2 3
1 0 0 0 −1 0 0 0 0 0 0 1 0 0 0−1
6 0 −1 0 0 0 1 0 0 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
6 0 0 1 0 0 0−1 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7
7, [χ03,1,1 ] =6 0 0 1 0 0 0 −1 0 7
6 7
[χ03,1,0 ] =6
6 −1 0 0 0 1 0 0 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
4 0 0 −1 0 0 0 1 0 5 4 0 1 0 0 0 −1 0 0 5
0 0 0 1 0 0 0 −1 0 0 −1 0 0 0 1 0
Image Processing 153

l(0) l(2) l(4) l(6) l(8) l(10) l(12) l(14) l(16) l(18) l(20) l(22)

0 l(23)
0

0.25 l(24)
1

0.25 l(25)
2

0.375 l(26)
3
y and m

0.5 l(27)
4

0.625 l(28)
5

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.38
The set of arithmetical rays for the (3, 1)-projection.

2 3 2 3
0 0 −1 0 0 0 1 0 0 1 0 0 0 −1 0 0
6 0 0 0 1 0 0 0 −1 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
6 1 0 0 0 −1 0 0 0 7 6 0 0 0 1 0 0 0−1 7
6 7 6 7
6 0 −1 0 0 0 1 0 0 7
7, [χ03,1,3 ] =6 1 0 0 0 −1 0 0 0 7
6 7
[χ03,1,2 ] =6
6 0 0 1 0 0 0−1 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7 6 0 0 1 0 0 0 −1 0 7
6 7 6 7
4 −1 0 0 0 1 0 0 0 5 4 0 0 0 −1 0 0 0 1 5
0 1 0 0 0 −1 0 0 −1 0 0 0 1 0 0 0
2 3 2 3
1 0 −1 0 1 0 −1 0 0 −1 0 1 0 −1 0 1
6 0 1 0 −1 0 1 0 −1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
6 −1 0 1 0 −1 0 1 0 7 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
0
6 0 −1 0 1 0 −1 0 1 7 0
6 −1 0 1 0 −1 0 1 0 7
[χ6,2,0 ] =66
7, [χ6,2,2 ] =6 0 −1 0 1 0 −1 0 1 7
7 6 7
6 1 0 −1 0 1 0 −1 0 7 6 7
6 0 1 0 −1 0 1 0 −1 7 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 1 0 −1 0 1 0 −1 5
0 −1 0 1 0 −1 0 1 −1 0 1 0 −1 0 1 0

The masks of these paired functions are shown in Figures 3.39 and 3.40,
together with the rays passing through the coefficients ±1 of the masks.
154 CHAPTER 3: Image Reconstruction

l(0) l(4) l(8) l(12) l(16) l(20) l(1) l(5) l(9) l(13) l(17) l(21)
0 0
0 0
0.25 l(24) 0.25
1 1
0.25 0.25 l(25)
2 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 l(28) 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

l(2) l(6) l(10) l(14) l(18) l(22) l(3) l(7) l(11) l(15) l(19)
0 0 l(23)
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 l(26) 0.375
y and m

y and m
3 3
0.5 0.5 l(27)
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.39
Masks of the paired functions with triplet-numbers (3, 1, t), t = 0, 1, 2, 3.

The components of the 2-D paired transform with the triplet-numbers of

the subset U (1, 3) are calculated by
0
f3,1,0 = v0 − v4 + v8 − v12 + v16 − v20 + v24 − v28 ,
0
f3,1,1 = v1 − v5 + v9 − v13 + v17 − v21 + v25 ,
0
f3,1,2 = v2 − v6 + v10 − v14 + v18 − v22 + v26 ,
0
f3,1,3 = v3 − v7 + v11 − v15 + v19 + v23 + v27 ,
0
f6,2,0 = (v0 + v4 + v8 + v12 + v16 + v20 + v24 + v28 )
(3.54)
−(v2 + v6 + v10 + v14 + v18 + v22 + v26 ),
0
f6,2,2 = (v1 + v5 + v9 + v13 + v17 + v21 + v25 )
−(v3 + v7 + v11 + v15 + v19 + v23 + v27 ),
0
f4,4,0 = v0 − v1 + v2 − v3 + · · · + v18 − v19 + v20 − v21 ,
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · + v18 + v19 + v20 + v21 ,

where variables vt = v3,1 (t) denote the sums of the discrete image along the
Image Processing 155

l(0)l(2)l(4)l(6)l(8)l(10)
l(12)
l(14)
l(16)
l(18)
l(20)
l(22) l(1)l(3)l(5)l(7)l(9)l(11)
l(13)
l(15)
l(17)
l(19)
l(21)
0 0 l(23)
0 0
0.25 l(24) 0.25
1 1
0.25 0.25 l(25)
2 2
0.375 l(26) 0.375
y and m

y and m
3 3
0.5 0.5 l(27)
4 4
0.625 l(28) 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 3.40
Masks of the paired functions with triplet-numbers (6, 2, t), t = 0, 2.

rays l(t), t = 0 : 28. The system of equations can be written as

0
f3,1,0 =a0 − a4
0
f3,1,1 =a1 − a5
0
f3,1,2 =a2 − a6
0
f3,1,3 =a3 − a7
0 (3.55)
f6,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f6,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 )
where an are calculated from the sums vt = v3,1 (t) of the discrete image by
X X
an = vn+8m = v3,1,n+8m, n = 0 : 7.
m m

The 8-point discrete paired transform of the vector a = (a0 , a1 , a2 , ..., a7)0
results in the 2-D paired transform data at points of U (3, 1),
P(3, 1) = [χ08,8 ]a.
To calculate the vector a from the projection data, we will define a set of
geometrical rays that allows calculation of the integrals w(t) by means of the
sums v(t), t = 0 : 28. For that, we first consider a few arithmetical rays as the
geometrical rays in the square [0, 1] × [0, 1], which are shown in Figure 3.41.
We define two geometrical rays l̃(8) = l(7) and ˜l(9) = l(8). One can see that
the line-integrals along these rays can be calculated as follows:
√
w(8) = w(˜l(8))= 8√10/3 [v(6) + v(7) + v(8)] ,
(3.56)
w(9) = w(˜l(9))= 8 10/3 [v(7) + v(8) + v(9)] .
156 CHAPTER 3: Image Reconstruction

l(0) l(4) l(6) l(7) l(8) l(9)

0.25

0.25

0.375

3
y and m

0.5

0.625

0.75

0.875

1
0 1 2 3

0 0.125 0.25 0.375 0.5

x and n

FIGURE 3.41
Eight rays of the (3, 1)-projection.

To work with similar equations for all line-integrals w(t) of the considered
projection,
√
w(t) = w(˜
l(t)) = 8 10/3 [v(t) + v(t − 1) + v(t − 2)] , t = 0 : 28, (3.57)

we define the following set of 29 geometrical rays:

 
t 1
l̃(t) = l3,1 (t − 1) = (x, y); 3x + y = + , t = 0 : 28.
8 8

A similar system of equations was described in (3.50)-(3.53) for the (1, 3)-
projection. This system has the following solution:
√
v(t) = 3bt /(8 10), t = 0 : 28,

where coefficients b(t) are calculated recursively as

b0 =w(0),
b1 =w(1) − b0 ,
b2 =w(2) − (b1 + b0 ), (3.58)
b3 =w(3) − (b2 + b1 ),
bt =w(t) − (bt−1 + bt−2 ), t = 4, 5, . . . , 27, 28.
Image Processing 157

l(0) l(2) l(4) l(6) l(8) l(10) l(12) l(14) l(16) l(18) l(20) l(22)

0 l(24)
0

0.25 l(25)
1

0.25 l(26)
2

0.375 l(27)
3
y and m

0.5 l(28)
4

0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.42
The set of geometrical rays for the (3, 1)-projection.

Figure 3.42 shows the new set of geometrical rays.

The following should be mentioned. If we denote the lines l(7) and l(8) in
Figure 3.41 by ˜l(6) and ˜
l(7), respectively, the line-integrals along these rays
will be defined as
√
w(6) = w(˜l(6))= 8√10/3 [v(6) + v(7) + v(8)] ,
w(7) = w(˜l(7))= 8 10/3 [v(7) + v(8) + v(9)] .

We obtain the system of equations

√
w(t) = w(˜l(t)) = 8 10/3 [v(t) + v(t + 1) + v(t + 2)] , t = 0 : 28, (3.59)

for the set of 29 geometrical rays which are defined as

n t 3o
l̃(t) = l3,1 (t + 1) = (x, y); 3x + y = + , t = 0 : 28.
8 8
This set of geometrical rays is shown in Figure 3.43. The system of equations
in (3.59) is described by the same Toeplitz matrix as for the subset U (1, 3)
in (3.47). Therefore,
√ the sums v(t) of the discrete image can be calculated by
v(t) = bt /(8 10/3), where t = 0 : 28, and components bt are calculated in
158 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(8) l(10) l(12) l(14) l(16) l(18) l(20)

0 l(22)
0

0.25 l(23)
1

0.25 l(24)
2

0.375 l(25)
3
y and m

0.5 l(26)
4

0.625 l(27)
5

0.75 l(28)
6

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.43
The second set of geometrical rays for the (3, 1)-projection.

recurrent form,
b28 =w(28),
b27 =w(27) − b28 ,
(3.60)
b26 =w(26) − (b27 + b28 ),
bt =w(t) − (bt+1 + bt+2 ), t = 25, 24, . . ., 1, 0.

3.5.6 (1, 4)- and (4, 1)-projections

Consider two subsets of triplet-numbers

 (1, 4, 0), (1, 4, 1), (1, 4, 2), (1, 4, 3),


(2, 0, 0), (2, 0, 2),
U (1, 4) =

 (4, 0, 0),

(0, 0, 0),
and 

 (4, 1, 0), (4, 1, 1), (4, 1, 2), (4, 1, 3),

(0, 2, 0), (0, 2, 2),
U (4, 1) =

 (0, 4, 0),

(0, 0, 0).
Image Processing 159

The paired functions with numbers (2, 0, 0), (0, 2, 2), (0, 2, 0), (0, 2, 2),
(4, 0, 0), (0, 4, 0), and (0, 0, 0) were described above. We thus consider the
paired functions with other triplet-numbers. Four masks of the paired func-
tions with the triplet-numbers (1, 4, t), t = 0 : 3, are
   
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
 −1 0 0 0 1 0 0 0   0 −1 0 0 0 1 0 0 
   
 1 0 0 0 −1 0 0 0   0 1 0 0 0 −1 0 0 
   
0
 −1 0 0 0 1 0 0 0  0
 0 −1 0 0 0 1 0 0 
[χ1,4,0] = 
 , [χ1,4,1] =  0 1 0 0 0 −1 0 0  ,
  
 1 0 0 0 −1 0 0 0   
 −1 0 0 0 1 0 0 0   0 −1 0 0 0 1 0 0 
   
 1 0 0 0 −1 0 0 0   0 1 0 0 0 −1 0 0 
−1 0 0 0 1 0 0 0 0 −1 0 0 0 1 0 0
   
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1
0
 0 −1 0 0 0 10  
0
 0 0 −1 0 0 0 1 
0
 0 1 0 0 0 −1 0 

0
 0 0 1 0 0 0 −1 

0 0 −1 0 0 0 10  0 0 0 −1 0 0 0 1
[χ01,4,2] =  , [χ01,4,3] =  .
0
 0 1 0 0 0 −1 0 

0
 0 0 1 0 0 0 −1 

0
 0 −1 0 0 0 10  
0
 0 0 −1 0 0 0 1 
0 0 1 0 0 0 −1 0  0 0 0 1 0 0 0 −1 
0 0 −1 0 0 0 10 0 0 0 −1 0 0 0 1

(1 + 4)7 + 1 = 36 parallel rays are considered for the (1, 4)-projection. The
control points of the set of rays are defined as shown below:
0 1 2 3 4 5 6 7 
• • • • • • • •
 . . . . 8 • 9 • 10 • 11 • 
 
 . . . . 12 • 13 • 14 • 15 • 
 
 . . . . 16 • 17 • 18 • 19 •  −1 ◦
 . . . . 20 • 21 • 22 • 23 •  (angle of rays is − tan (1/4) = −14.036 ).
 
 
 . . . . 24 • 25 • 26 • 27 • 
 
 . . . . 28 • 29 • 30 • 31 • 
. . . . 32 • 33 • 34 • 35 •

The set of the 36 parallel rays of this projection,

 
t 5
l(t) = l1,4 (t) = (x, y); x + 4y = + , t = 0 : 35,
8 16

where (x, y) ∈ [0, 1] × [0, 1], is shown in Figure 3.44. These rays are also shown
separately on the masks of the paired functions χ01,4,t, t = 0 : 3, in Figure 3.45.

We denote by vt = v1,4 (t), where t = 0 : 35, the sums of the discrete

image fn,m along these rays. The components of the paired transform with
160 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5)

0 l(6)
l(7)
0 l(8)
l(9)
0.25 l(10)
l(11)
1 l(12)
l(13)
0.25 l(14)
l(15)
2 l(16)
l(17)
0.375 l(18)
l(19)
3 l(20)
y and m

l(21)
0.5 l(22)
l(23)
4 l(24)
l(25)
0.625 l(26)
l(27)
5 l(28)
l(29)
0.75 l(30)
l(31)
6 l(32)
l(33)
0.875 l(34)
l(35)
7

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.44
The sets of rays for the (1, 4)-projection.

the triplet-numbers (1, 4, t) can be calculated as follows:

0
f1,4,0 = v0 − v4 + v8 − v12 + v16 − v20 + v24 − v28 + v32 ,
0
f1,4,1 = v1 − v5 + v9 − v13 + v17 − v21 + v25 − v29 + v33 ,
0 (3.61)
f1,4,2 = v2 − v6 + v10 − v14 + v18 − v22 + v26 − v30 + v34 ,
0
f1,4,3 = v3 − v7 + v11 − v15 + v19 − v23 + v27 − v31 + v35 .

Two masks for the paired functions χ02,0,t, t = 0, 2, are shown in Fi-
gure 3.46. The corresponding components of the 2-D paired transform are
calculated by
0
f2,0,0 =(v0 + v4 + v8 + v12 + v16 + v20 + v24 + v28 + v32 )
−(v2 + v6 + v10 + v14 + v18 + v22 + v26 + v30 + v34 ),
0 (3.62)
f2,0,2 =(v1 + v5 + v9 + v13 + v17 + v21 + v25 + v29 + v33 )
−(v3 + v7 + v11 + v15 + v19 + v23 + v27 + v31 + v35 ).
0
Similarly, for the remaining two components, we have the following: f4,0,0 =
0
v0 − v1 + v2 − v3 + · · · + v34 − v35 and f0,0,0 = v0 + v1 + v2 + v3 + · · · + v34 + v35 .
Introducing the new variables
X X
an = vn+8m = v1,4,n+8m, n = 0 : 7,
m m
Image Processing 161

l(0) l(4) l(1) l(5)

0 0
0 l(8) 0 l(9)
0.25 0.25
1 l(12) 1 l(13)
0.25 0.25
2 l(16) 2 l(17)
0.375 0.375
l(20)
y and m

y and m
3 3 l(21)
0.5 0.5
4 l(24) 4 l(25)
0.625 0.625
5 l(28) 5 l(29)
0.75 0.75
6 l(32) 6 l(33)
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

l(2) l(6) l(3)

0 0 l(7)
0 0
0.25 l(10) 0.25 l(11)
1 1
0.25 l(14) 0.25 l(15)
2 2
0.375 l(18) 0.375 l(19)
y and m

3
y and m 3
0.5 l(22) 0.5 l(23)
4 4
0.625 l(26) 0.625 l(27)
5 5
0.75 l(30) 0.75 l(31)
6 6
0.875 l(34) 0.875 l(35)
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

FIGURE 3.45
Four masks with the set of arithmetical rays for the paired functions with
triplet-numbers (1, 4, t), t = 0 : 3.

the complete system of linear equations for the set P(1, 4) with eight compo-
nents fp0 1 ,s1 ,t , (p1 , s1 , t) ∈ U (1, 4), can be written as follows:
0
f1,4,0 =a0 − a4
0
f1,4,1 =a1 − a5
0
f1,4,2 =a2 − a6
0
f1,4,3 =a3 − a7
0 (3.63)
f2,0,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,0,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,0,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
These calculations correspond to the 8-point discrete paired transform of the
vector a = (a0 , a1 , a2 , ..., a7)0 , i.e., P(1, 4) = [χ08 ]a.
To express the sums v(t) of the discrete image fn,m through the line-
integrals w(t) = w1,4 (t), we consider the part of the mask of the paired func-
162 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(1) l(3) l(5)

0 0 l(7)
0 l(8) 0 l(9)
0.25 l(10) 0.25 l(11)
1 l(12) 1 l(13)
0.25 l(14) 0.25 l(15)
2 l(16) 2 l(17)
0.375
l(18) 0.375 l(19)
l(20)
y and m

y and m
3 3 l(21)
0.5
l(22) 0.5 l(23)
4 l(24) 4 l(25)
0.625
l(26) 0.625 l(27)
5 l(28) 5 l(29)
0.75
l(30) 0.75 l(31)
6 l(32) 6 l(33)
0.875 l(34) 0.875 l(35)
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.46
Two masks with the set of arithmetical rays for the paired functions with
triplet-numbers (2, 0, t), t = 0, 1.

tion with number (1, 4, 0), which is shown in Figure 3.47. The geometrical ray
l(29) between two points of the discrete image, (0, 7) and (5, 6) passes through
six image elements of numbers (0, 7), (1, 7), (2, 7), (3, 7), (3, 6), and (4, 6). The
length of the intersection ∆l0,7 of the ray with the IE of number (0, 7) is twice
the length of intersection with each IE of number (3, 7) and (3, 6), but equal
to the length
√ of intersection with the IE of numbers (1, 7) and (2, 7). Since
∆l0,7 = 17/4∆x, we can write the following:
 
∆l0,7 1 1
w(29) = w1,4(29) = v(28) + v(29) + v(30) + v(31) + v(27) .
(∆x)2 2 2
Similar equations can be used for calculating other integrals w(t), t = 0 : 35,
and then the sums v(t) can be found.
Now we consider the geometrical rays l1 and l2 between rays 28, 29, and
29, 30, respectively. The length of the intersection of ray l1 with √the image
elements of numbers (0, 7), (1, 7), (2, 7), and (3, 6) is the same, 17/4∆x.
The length of the intersection of ray l2 √ with each image element of num-
ber (0, 7), (1, 7), (2, 7), and (3, 7) is also 17/4∆x. Therefore, the integrals
w1 and w2 of the image along the shifted rays l1 and l2 , respectively, can be
written as follows:
√
17/4∆x
w1 = [v(28) + v(29) + v(30) + v(27)]
(∆x)2
√
=2 17 [v(27) + v(28) + v(29) + v(30)] ,
√
w2 =2 17 [v(28) + v(29) + v(30) + v(31)] .
These equations can be generalized and we consider two ways to do that.
Image Processing 163

0.625

l(27)

5 l(28)

l(29)

0.75 l(30)

l(31)

6 l(32)
y and m

0.875

l1
1
l2

0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.47
Set of geometrical rays for calculating line-integrals when (p, s) = (1, 4).

Case 1: The integral w1 can be denoted by w(30), and w2 by w(31).

Case 2: The integral w1 can be denoted by w(27), and w2 by w(28).
In both cases, a simple solution for sums v(t) can be obtained. Indeed, in
the first case, we consider a new set of parallel rays defined by the shifting,
t → t − 3/2,
 
˜ t 1
l(t) = l1,4 (t − 3/2) = (x, y); x + 4y = + , t = 0 : 35,
8 8
which is shown in Figure 3.48. In the second case, we consider the set of
parallel rays defined by the shifting, t → t + 3/2,
 
˜ t 1
l(t) = l1,4 (t + 3/2) = (x, y); x + 4y = + , t = 0 : 35,
8 2
which is shown in Figure 3.49. The line-integrals along the new rays are defined
as w(t) = w(˜ l(t)), where t = 0 : 35.
In the first case, for the integrals along the set of geometrical rays, we can
write the following equations:
√
w(t) = 2 17 [v(t) + v(t − 1) + v(t − 2) + v(t − 3)] , t = 0 : 35,
164 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0 l(7)

0 l(9)

0.25 l(11)

1 l(13)

0.25 l(15)

2 l(17)

0.375 l(19)

3 l(21)
y and m

0.5 l(23)

4 l(25)

0.625 l(27)

5 l(29)

0.75 l(31)

6 l(33)

0.875 l(35)

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.48
The first set of geometrical rays for the (1, 4)-projection.

where v(−1) = v(−2) = v(−3) = 0. This system of linear equations can be

written in matrix form as
 
1 0 0 0 0 0 ... 0 0 0 0 0
 1 1 0 0 0 0 ... 0 0 0 0 0 
 
 1 1 1 0 0 0 ... 0 0 0 0 0 
 
 1 1 1 1 0 0 ... 0 0 0 0 0 
 
√ √  0 1 1 1 1 0 ... 0 0 0 0 0 
w = 2 17Av = 2 17 
 0 0 1 1 1 1 ... 0 0 0 0 0 v,
 (3.64)
 
 . . . . . . ... . . . . . 
 
 . . . . . . ... . . . 0 0 
 
 0 0 0 0 0 0 ... 1 1 1 1 0 
0 0 0 0 0 0 ... 0 1 1 1 1

where w = (w0 , w1 , ..., w35)0 and v = (v0 , v1 , ..., v35)0 . The above Toeplitz
Image Processing 165

l(0) l(1) l(2) l(3) l(4)

0
l(5)
0
l(7)
0.25
l(9)
1
l(11)
0.25
l(13)
2
l(15)
0.375
l(17)
3
l(19)
y and m

0.5
l(21)
4
l(23)
0.625
l(25)
5
l(27)
0.75
l(29)
6
l(31)
0.875
l(33)
7
l(35)
1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.49
The second set of geometrical rays for the (1, 4)-projection.

matrix 36 × 36 has the following inverse lower triangle Toeplitz matrix:

 
1 0 0 0 0 0 0 ... 0 0 0 0
 −1 1 0 0 0 0 0 ... 0 0 0 0 
 
 0 −1 1 0 0 0 0 ... 0 0 0 0 
 
 0 0 −1 1 0 0 0 ... 0 0 0 0 
 
 1 0 0 −1 1 0 0 ... 0 0 0 0 
 
 −1 1 0 0 −1 1 0 ... 0 0 0 0 
 
A−1 =  0 −1 1 0 0 −1 1 ... 0 0 0 0  .
 (3.65)
 0 0 −1 1 0 0 −1 ... 0 0 0 0 
 
 ... ... ... ... ... ... ...... ... ... ...... 
 
 1 0 0 −1 1 0 0 ... 1 0 0 0 
 
 −1 1 0 0 −1 1 0 ...−1 1 0 0 
 
 0 −1 1 0 0 −1 1 ... 0 −1 1 0 
0 0 −1 1 0 0 −1 ... 0 0 −1 1
−1
√ the solution of the equation b = A w, we obtain the sums v =
From
b/(2 17). The vector b can be calculated by using the following recurrent
166 CHAPTER 3: Image Reconstruction

procedure:
b0 =w(0),
b1 =w(1) − b0 ,
b2 =w(2) − (b1 + b0 ),
(3.66)
b3 =w(3) − (b2 + b1 + b0 ),
b4 =w(4) − (b3 + b2 + b1 ),
bt =w(t) − (bt−1 + bt−2 + bt−3 ), t = 5, 6, . . . , 35.
In the second case, the integrals along the set of geometrical rays can be
defined by equations
√
w(t) = 2 17 [v(t) + v(t + 1) + v(t + 2) + v(t + 3)] , t = 0 : 35,
where v(36) = v(37) = v(38) = 0. This system of linear equations can be
written in matrix form as
 
111100 ... 0 0 0
0 1 1 1 1 0 ... 0 0 0
 
0 0 1 1 1 0 ... 0 0 0
√ √  
w = 2 17Av = 2 17  0 0 0 1 1 1 ... 0 0 0

 v, (3.67)
 .. .. .. .. .. .. . .. 
. . . . . .
 ... .. . 

0 0 0 0 0 0 ... 0 1 1
000000 ... 0 01
where w = (w0 , w1 , ..., w35)0 and v = (v0 , v1 , ..., v35)0 . The above Toeplitz
matrix 36 × 36 has the following triangle inverse matrix:
 
1 −1 0 0 1 −1 0 0 1 −1 0 ... 1 −1 0 0
 0 1−1 0 0 1 −1 0 0 1 −1 ... 0 1 −1 0 
 
 0 0 1 −1 0 0 1−1 0 0 1 ... 0 0 1 −1 
 
 0 0 0 1 −1 0 0 1 −1 0 0 ...−1 0 0 1 
 
 0 0 0 0 1 −1 0 0 1 −1 0 ... 1 −1 0 0 
 
 0 0 0 0 0 1 −1 0 0 1 −1 ... 0 1 −1 0 
 
A−1 =   . . . . . . . . . . . ... . . . .  .
 (3.68)
 0 0 0 0 0 0 0 0 0 0 1 ... 0 0 1 −1 
 
 0 0 0 0 0 0 0 0 0 0 0 ...−1 0 0 1 
 
 0 0 0 0 0 0 0 0 0 0 0 ... 1 −1 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 ... 0 1 −1 0 
 
 0 0 0 0 0 0 0 0 0 0 0 ... 0 0 1 −1 
0 0 0 0 0 0 0 0 0 0 0 ... 0 0 0 1

The equation b = A−1 w can be written in the following recurrent form:

b35 =w(35),
b34 =w(34) − b35 ,
b33 =w(33) − (b34 + b35 ),
(3.69)
b32 =w(32) − (b33 + b34 + b35 ),
b31 =w(31) − (b32 + b33 + b34 ),
bt =w(t) − (bt+1 + bt+2 + bt+3 ), t = 30, 29, . . ., 1, 0.
Image Processing 167
√
Therefore, the sums are calculated by v = b/(2 17).
Now we consider the first four matrices of the subset U (4, 1),
2 3 2 3
1 −1 1 −1 1 −1 1 −1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 −1 1 −1 1 −1 1 −1 7
6 7 6 7
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07
7, [χ04,1,1 ] =6 0 0 0 0 0 0 0 0 7
6 7
[χ04,1,0 ] =6
6 −1 1 −1 1 −1 1 −1 1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 −1 1 −1 1 −1 1 −1 1 7
6 7 6 7
4 0 0 0 0 0 0 0 05 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 3 2 3
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
6 1 −1 1 −1 1 −1 1 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7, [χ04,1,3 ] =6 1 −1 1 −1 1 −1 1 −1 7
6 0 0 0 0 0 0 0 07 6 7
[χ04,1,2 ] =6
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 0 0 0 0 0 0 0 07
6 7 6 7
4 −1 1 −1 1 −1 1 −1 1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 1 −1 1 −1 1 −1 1

Thirty-six rays of this projection are used to define all components of the
2-D paired transforms with triplet-numbers of U (4, 1). We define the control
points of the set of rays as shown below:
0 
• . . . . . . .
1 • . . . . . . . 
2 
 • . . . . . . .
3 
 • . . . . . . .
 5 8 12 16 20 24 28 32  (angle of rays is − tan−1 (4) = −75.964◦).
 • • • • • • • •
 5 9 13 17 21 25 29 33 
 • • • • • • • •
 6 10 14 18 22 26 30 34 
 • • • • • • • •
7 11 15 19 23 27 31 35
• • • • • • • •

The set of the 36 rays of this projection,

 
t 5
l(t) = l4,1 (t) = (x, y); 4x + y = + , t = 0 : 35,
8 16

where (x, y) ∈ [0, 1] × [0, 1], is shown in Figure 3.50 on the masks of the paired
functions χ04,1,t, t = 0 : 3.
The components of the 2-D paired transform with the triplet-numbers
(4, 1, t) are calculated by
0
f4,1,0 =v0 − v4 + v8 − v12 + v16 − v20 + v24 − v28 + v32
0
f4,1,1 =v1 − v5 + v9 − v13 + v17 − v21 + v25 − v29 + v33
0 (3.70)
f4,1,2 =v2 − v6 + v10 − v14 + v18 − v22 + v26 − v30 + v34
0
f4,1,3 =v3 − v7 + v11 − v15 + v19 − v23 + v27 − v31 + v35
168 CHAPTER 3: Image Reconstruction

l(0) l(4) l(8) l(12)l(16)l(20)l(24)l(28) l(1) l(5) l(9) l(13)l(17)l(21)l(25)l(29)

0 0
0 0
0.25 0.25
1 l(32) 1
0.25 0.25
2 2 l(33)
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

l(2) l(6) l(10)l(14)l(18)l(22)l(26)l(30) l(3) l(7) l(11)l(15)l(19)l(23)l(27)

0 0
0 0 l(31)
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m
3 l(34) 3
0.5 0.5
4 4 l(35)
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.1250.250.3750.50.6250.750.875 1 0 0.1250.250.3750.50.6250.750.875 1
x and n x and n

FIGURE 3.50
Masks and rays for calculating line-integrals for the (4, 1)-projection.

where the sums of the discrete image along the rays are denoted by vt = v4,1 (t),
t = 0 : 35. With the new variables
X X
an = vn+8m = v4,1,n+8m, n = 0 : 7,
m m

the complete system of linear equations, which defines the subset P(4, 1) of
components of the 2-D paired transform with the triplet-numbers of the subset
Image Processing 169

U (4, 1), can be written as follows:

0
f4,1,0 =a0 − a4
0
f4,1,1 =a1 − a5
0
f4,1,2 =a2 − a6
0
f4,1,3 =a3 − a7
0 (3.71)
f0,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f0,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f0,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).

These calculations correspond to the 8-point discrete paired transform of the

vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(4, 1) = [χ08 ]a.

To express the sums v(t) of the discrete image fn,m through the line-
integrals w(t), we consider in detail the part of the mask of the paired func-
tion with number (4, 1, 0), which is shown in Figure 3.51. The geometrical
ray l1 between rays l(6) and l(7) passes through six image elements of num-
bers (0, 7), (0, 6), (0, 5), (1, 4), (1, 3), (1, 2), (1, 1), and (2, 0). The length of the
intersection ∆l0,7 of the ray with the IE of number (0, 7) equals the length
of the intersection
√ of the ray with each of the remaining seven IEs. Since
∆l0,7 = 17/4∆x, we can write the following:

∆l0,7
w1 = w(l1 ) = [v(7) + v(6) + v(5) + v(8)] .
(∆x)2

For the geometrical ray l2 , which is between rays l(7) and l(8), we obtain

∆l0,7
w2 = w(l2 ) = [v(7) + v(6) + v(9) + v(8)] .
(∆x)2

These equations can be used for calculating other integrals w(t), t = 0 : 35,
and then the sums v(t) can be found. Similar to the (p, s) = (1, 4) case, these
equations can be generalized, and we consider two ways to do that.
Case 1: The line-integral w1 can be denoted by w(8), and w2 by w(9).
Case 2: The line-integral w1 can be denoted by w(5), and w2 by w(6).
In the first case, we consider the set of geometrical parallel rays, which are
defined by the shifting, t → t − 3/2,
 
t 1
l̃(t) = l4,1 (t − 3/2) = (x, y); 4x + 1y = + , t = 0 : 35.
8 8

These rays are shown in Figure 3.52, and the original set of arithmetical rays
is shown in Figure 3.53.
170 CHAPTER 3: Image Reconstruction

l(5) l(6) l(7) l(8) l(9)

0

0.25

0.25

0.375

3
y and m

0.5

0.625

0.75

0.875

1
l l 0 1 2
1 2
0 0.125 0.25 0.375
x and n

FIGURE 3.51
Set of geometrical rays for calculating line-integrals when (p, s) = (4, 1).

The line-integrals along the set of geometrical rays can be described by

the following equations:
√
w(t) = w(˜l(t)) = 2 17 [v(t) + v(t − 1) + v(t − 2) + v(t − 3)] , t = 0 : 35,

where v(−1) = v(−2) = v(−3) = 0. As shown in (3.66),√ the solution of this

system of linear equations is calculated by v = b/(2 17), where the vector b
can be calculated by means of the following recursion:

b0 =w(0),
b1 =w(1) − b0 ,
b2 =w(2) − (b1 + b0 ),
(3.72)
b3 =w(3) − (b2 + b1 + b0 ),
b4 =w(4) − (b3 + b2 + b1 ),
bt =w(t) − (bt−1 + bt−2 + bt−3 ), t = 5, 6, . . . , 35.

For the second case, we consider the set of geometrical parallel rays defined
Image Processing 171

l(0) l(4) l(8) l(12) l(16) l(20) l(24) l(28)

0 l(32)

0.25 l(33)

0.25 l(34)

0.375 l(35)

3
y and m

0.5

0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.52
The set of geometrical rays for the (4, 1)-projection.

by the shifting, t → t + 3/2,

 
t 1
l̃(t) = l4,1 (t + 3/2) = (x, y); 4x + 1y = + , t = 0 : 35.
8 2
The integrals along the set of geometrical rays can be defined by the following
equations:
√
w(t) = w(˜
l(t)) = 2 17 [v(t) + v(t + 1) + v(t + 2) + v(t + 3)] , t = 0 : 35,
where√ v(36) = v(37) = v(38) = 0. The sums v(t) are defined by v(t) =
bt /(2 17), where the components bt are calculated by the following recurrent
procedure:
b35 =w(35),
b34 =w(34) − b35 ,
b33 =w(33) − (b34 + b35 ),
(3.73)
b32 =w(32) − (b33 + b34 + b35 ),
b31 =w(31) − (b32 + b33 + b34 ),
bt =w(t) − (bt+1 + bt+2 + bt+3 ), t = 30, 29, . . . , 0.
172 CHAPTER 3: Image Reconstruction

l(0) l(4) l(8) l(12) l(16) l(20) l(24) l(28) l(30)

0 l(31)

0.25

1 l(32)

0.25

2 l(33)

0.375

3 l(34)
y and m

0.5

4 l(35)

0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.53
The set of arithmetical rays for the (4, 1)-projection.

The set of 36 geometrical rays ˜l(t) is shown in Figure 3.54.

3.5.7 (1, 5)-projection

The set of triplet-numbers generated by the frequency-point (p, s) = (1, 5) is


 (1, 5, 0), (1, 5, 1), (1, 5, 2), (1, 5, 3),

(2, 2, 0), (2, 2, 2),
U (1, 5) =

 (4, 4, 0),

(0, 0, 0).

The masks of the paired functions χ04,4,0 and χ00,0,0 , as well as χ02,2,0 and
χ02,2,2
were given above. Therefore, we consider only masks with the first four
Image Processing 173

l(0) l(4) l(8) l(12) l(16) l(20) l(24) l(28)

0.25 l(30)

0.25 l(31)

0.375 l(32)

3
y and m

0.5 l(33)

0.625 l(34)

0.75 l(35)

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.54
The 2nd set of geometrical rays for the (4, 1)-projection.

triplet-numbers of the subset U (1, 5),

2 3 2 3
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
6 0 0 0 1 0 0 0 −1 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
0
6 0 1 0 0 0 −1 0 0 7 0
6 0 0 1 0 0 0 −1 0 7
[χ1,5,0 ] =6
6
7, [χ1,5,1 ] =6 0 −1 0 0 0 1 0 0 7
7 6 7
6 −1 0 0 0 1 0 0 0 7 6 7
6 0 0 0 −1 0 0 0 1 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
4 0 0 1 0 0 0−1 0 5 4 0 0 0 1 0 0 0−1 5
0 −1 0 0 0 1 0 0 0 0 −1 0 0 0 1 0
2 3 2 3
0 0 1 0 0 0−1 0 0 0 0 1 0 0 0−1
6 0 −1 0 0 0 1 0 0 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
6 1 0 0 0 −1 0 0 0 7 6 0 1 0 0 0 −1 0 0 7
6 7 6 7
7 , [χ01,5,3 ] =6 −1 0 0 0 1 0 0 0 7
6 0 0 0 1 0 0 0−1 7 6 7
[χ01,5,2 ] =6
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7 6 0 0 1 0 0 0−1 0 7
6 7 6 7
4 −1 0 0 0 1 0 0 0 5 4 0 −1 0 0 0 1 0 0 5
0 0 0 −1 0 0 0 1 1 0 0 0 −1 0 0 0
174 CHAPTER 3: Image Reconstruction

It should be noted that the same set of masks can be referred to as the
masks for the subset U (5, 1),


 (5, 1, 0), (5, 1, 1), (5, 1, 2), (5, 1, 3),

(2, 2, 0), (2, 2, 2),
U (5, 1) =

 (4, 4, 0),

(0, 0, 0).

Indeed, the following is valid: χ05,1,2t = χ01,5,2t and χ05,1,2t+1 = −χ01,5,2t+1, for
t = 0, 1.
The set of all (p + s)(N − 1) + 1 = 6(7) + 1 = 43 arithmetical rays of the
projection by the angle π − tan−1 (5) is shown in Figure 3.55.

l(0) l(1) l(2) l(3) l(4) l(5)

0 l(6)
0 l(8)
0.25 l(10)
l(12)
1
l(14)
0.25 l(16)
2 l(18)
0.375 l(20)
l(22)
3
l(24)
y and m

0.5 l(26)
4 l(28)
0.625 l(30)
l(32)
5
l(34)
0.75 l(36)
6 l(38)
0.875 l(40)
l(42)
7

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.55
The set of arithmetical rays for the (5, 1)-projection.

To reduce this large number of rays, we consider another direction, or

projection which defines the same set of masks of the paired functions. The
similar substitution of the projections was described in the N = 4 example, for
the (p, s) = (1, 3) case. In the (p, s) = (1, 5) case, this alternative projection
is determined by the frequency-point (p, s − N ) = (1, −3).
As an example, Figures 3.56 and 3.57 show the mask of the paired function
Image Processing 175

l(0) l(4)
0

0 l(8)
0.25
l(12)
1

0.25 l(16)
2

0.375 l(20)

3
l(24)
y and m

0.5

4 l(28)
0.625
l(32)
5

0.75 l(36)
6
l(40)
0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.56
The mask [χ01,5,0] with the set of arithmetical rays of the (1, 5)-projection.

χ01,5,0 with the parallel arithmetical rays of two different projections. Eleven
rays of the (1, 5)-projection are shown in the first figure, and seven rays of the
(1, −3)-projection are shown in the second figure.
The number of parallel rays for the (1, −3)-projection equals (1 +3)7 +1 =
29, which is less than 43. The set of all rays for this projection is shown in
Figure 3.58. The control points of this set of rays are given below
 0 
• 1 • 2 • 3 • 4• 5 • 6• 7 •
 −3 • −2 • −1 • . . . . . 
 −6 −5 −4 
 −9 • −8 • −7 • . . . . . 
 
 −12 • −11 • −10 • . . . . .  (angle of rays is tan−1 (1/3) = 18.435◦).
 
 −15 • −14 • −13 • . . . . . 
 
 −18 • −17 • −16 • . . . . . 
 
 • • • . . . . .
−21 −20 −19
• • • . . . . .
The set of the 29 geometrical rays of this projection is described as
 
t 1
l(t) = l1,−3 (t) = (x, y); x − 3y = − , t = 7 : −1 : −21, (3.74)
8 8
176 CHAPTER 3: Image Reconstruction

l(4)
0

0.25

1
l(4)
0.25

0.375 l(0)
3
y and m

0.5

4
l(−4)

0.625

5
l(−8)
0.75

0.875 l(−12)
7

1
l(−20) l(−16)
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.57
The mask [χ01,5,0] with the set of arithmetical rays of the (1, −3)-projection.

where (x, y) ∈ [0, 1] × [0, 1]. The first six masks for the paired functions are
shown in Figure 3.59 and 3.60.
The components of the 2-D paired transform with the triplet-numbers
(1, 5, t) are calculated as follows:
0
f1,5,0 =−v4 + v0 − v−4 + v−8 − v−12 + v−16 − v−20
=(v0 + v−8 + v−16 ) − (v4 + v−4 + v−12 + v−20 ),
0
f1,5,1 =−v5 + v1 − v−3 + v−7 − v−11 + v−15 − v−19
=(v1 + v−7 + v−15 ) − (v5 + v−3 + v−11 + v−19 ),
0 (3.75)
f1,5,2 =−v6 + v2 − v−2 + v−6 − v−10 + v−14 − v−18
=(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 + v−18 ),
0
f1,5,3 =−v7 + v3 − v−1 + v−5 − v−9 + v−13 − v−17 + v−21
=(v3 + v−5 + v−13 + v−21 ) − (v7 + v−1 + v−9 + v−17 ),

where the sums of the discrete image along the rays are denoted by vt =
v1,−3 (t), t = 7 : −1 : −21. For the components of the 2-D paired transform
Image Processing 177

l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0

0
l(7)
0.25 l(6)
1
l(5)
l(4)
0.25 l(3)
2
l(2)
l(1)
0.375 l(0)
3
l(−1)
l(−2)
y and m

0.5 l(−3)
4
l(−4)
l(−5)
0.625 l(−6)
5
l(−7)
l(−8)
0.75 l(−9)
6
l(−10)
l(−11)
0.875 l(−12)
7
l(−13)
l(−14)
1 l(−15)
l(−21) l(−20) l(−19) l(−18) l(−17) l(−16)
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.58
The set of arithmetical rays for the (1, −3)-projection.

with the remaining numbers of the subset U (1, 5), we have the following:
0
f2,2,0 =−v6 + v4 − v2 + v0 − v−2 + v−4 − v−6 + v−8 −
−v−10 + v−12 − v−14 + v−16 − v−18 + v−20
=(v0 + v−8 + v−16 ) + (v4 + v−4 + v−12 + v−20 )−
−(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 + v−18 ),
0
f2,2,2 =−v7 + v5 − v3 + v1 − v−1 + v−3 − v−5 + v−7
−v−9 + v−11 − v−13 + v−15 − v−17 + v−19 − v−21
=(v1 + v−7 + v−15 ) + (v5 + v−3 + v−11 + v−19 )−
−(v3 + v−5 + v−13 + v−21 ) − (v7 + v−1 + v−9 + v−17 ),
0
f4,4,0 =−v7 + v6 − v5 + v4 − v3 + v2 − v1 + v0 −
−v−1 + v−2 − v−3 + v−4 − v−5 + ... + v−20 − v21
=(v0 + v−8 + v−16 ) + (v4 + v−4 + v−12 + v−20 )−
−(v1 + v−7 + v−15 ) − (v5 + v−3 + v−11 + v−19 )+
+(v2 + v−6 + v−14 ) + (v6 + v−2 + v−10 + v−18 )−
−(v3 + v−5 + v−13 − v−21 ) − (v7 + v−1 + v−9 + v−17 ),
0
f0,0,0 =v0 + v1 + v2 + v3 + · · · + v7 + v−1 + v−2 + v−3 +
+v−4 + v−5 + ... + v−20 + v21 .
178 CHAPTER 3: Image Reconstruction

0
l(4) 0
l(5)
0 0
0.25 0.25
1 1 l(5)
0.25
l(4) 0.25
2 2 l(1)
0.375 l(0) 0.375
y and m

y and m
3 3
0.5 0.5 l(−3)
4 l(−4) 4
0.625 0.625
5 5 l(−7)
0.75
l(−8) 0.75
6 6 l(−11)
0.875 l(−12) 0.875
7 7
1 l(−20) 5 6 l(−16) 1 l(−19) 6 7 l(−15)
0 1 2 0 1 2 3
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

0
l(6) 0
l(7)
0 0 l(7)
0.25 l(6) 0.25
1 1
0.25 0.25 l(3)
2 l(2) 2
0.375 0.375
y and m

3 y and m 3 l(−1)
0.5
l(−2) 0.5
4 4 l(−5)
0.625 l(−6) 0.625
5 5
0.75 0.75 l(−9)
6 l(−10) 6
0.875 0.875
7 7 l(−13)
1
l(−14) 1
0 1 2 3 4 l(−18) 7 0 1 l(−21) 4 5 l(−17)
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.59
Four masks [χ01,5,k ], k = 0 : 3, with the arithmetical rays l1,−3 (t).

l(2) l(4) l(6) l(1) l(3) l(5) l(7)

0 0
0 0
l(7)
0.25 l(6) 0.25
1 1 l(5)
l(4)
0.25 0.25 l(3)
2 l(2) 2
l(1)
0.375 l(0) 0.375
l(−1)
y and m

y and m

3 3
l(−2)
0.5 0.5 l(−3)
4 l(−4) 4
l(−5)
0.625 l(−6) 0.625
5 5 l(−7)
l(−8)
0.75 0.75 l(−9)
6 l(−10) 6
l(−11)
0.875 l(−12) 0.875
7 7 l(−13)
l(−14)
1 1 l(−15)
0 1 2 l(−20) l(−18) l(−16) 0 1 l(−21) l(−19) l(−17)
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.60
Two masks [χ02,2,2k ], k = 0, 1, with the arithmetical rays l1,−3 (t).

With the new variables

X X
an = vn−8m = v1,−3,n−8m, n = 0 : 7,
m m
Image Processing 179

the complete system of linear equations, for calculating the components of

the 2-D paired transform with triplet-numbers of the subset U (1, 5), can be
written as follows:
0
f1,5,0 =a0 − a4
0
f1,5,1 =a1 − a5
0
f1,5,2 =a2 − a6
0
f1,5,3 =a3 − a7
0 (3.76)
f2,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
These calculations correspond to the 8-point discrete paired transform of the
vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(1, 5) = P(1, −3) = [χ08 ]a.

As in the (1, 3) case, we consider the shifted set of the arithmetical rays,
l(t) → l(t + 1), as the set of 29 geometrical rays:
 
t
l̃(t) = l1,−3 (t + 1) = (x, y); x − 3y = , t = 7 : −1 : −21,
8
which are shown in Figure 3.61. We obtain the same system of equations for
the line-integrals, as in the (1, 3) case considered above (see (3.51)-(3.53)),
√
w(t) = w(˜ l(t)) = 8 10/3 [v(t) + v(t + 1) + v(t + 2)] , t = 7 : −1 : −21,
(3.77)
where v(8) = v(9) = 0. The required sums √ v(t) of the discrete image fn,m
are therefore defined as v(t) = 3bt /(8 10), where the components bt are
calculated recursively by
b7 = w(7),
b6 = w(6) − b7 ,
(3.78)
b5 = w(5) − (b6 + b7 ),
bt = w(t) − (bt+1 + bt+2 ), t = 4, 3, 2, 1, 0, −1, . . . , −20, −21.
Remark 3.4 To remove negative numbers in control points, we can consider
the change in numbers by t → 7 − t,
 7 6 5 4 3 2 1 0 
• • • • • • • •
 10 • 9 • 8 • . . . . . 
 13 12 11 
 • • • . . . . .
 16 15 14 
 • • • . . . . . −1 ◦
 • • • . . . . .  (angle of rays is tan (1/3) = 18.435 ).
 19 18 17 
 22 21 20 
 • • • . . . . .
 25 24 23 
 • • • . . . . .
28 27 26
• • • . . . . .
180 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0

0
l(7)
l(6)
0.25 l(5)
1
l(4)
l(3)
0.25 l(2)
2
l(1)
l(0)
0.375 l(−1)
3
l(−2)
l(−3)
y and m

0.5 l(−4)
4
l(−5)
l(−6)
0.625 l(−7)
5
l(−8)
l(−9)
0.75 l(−10)
6
l(−11)
l(−12)
0.875 l(−13)
7
l(−14)
l(−15)
1
l(−21) l(−20) l(−19) l(−18) l(−17) l(−16)
0 1 2

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.61
The set of geometrical rays for the (1, −3)-projection.

The set of 29 geometrical rays of this projection in (3.74) can be written as

 
t 3
l(t) = l1,−3 (7 − t) = (x, y); x − 3y = − + , t = 0 : 28.
8 4
Let v̄t = v7−t and w̄t = w7−t . Then, the system of equations in (3.77) can be
described as
√
w̄(t) = w(˜l(t)) = 8 10/3 [v̄(t) + v̄(t − 1) + v̄(t − 2)] , t = 0 : 28, (3.79)
√
and the equations in (3.78) for new coefficients b̄t = v̄(t)8 10/3 are described
as
b̄0 = w̄(0),
b̄1 = w̄(1) − b̄0 ,
(3.80)
b̄2 = w̄(2) − (b̄1 + b̄0 ),
b̄t = w̄(t) − (b̄t−1 + b̄t−2 ), t = 3, 4, . . . , 27, 28.

Remark 3.5 We can consider the shifted set of the arithmetical rays, l(t) →
l(t − 1), as the set of 29 geometrical rays:
 
˜ t 1
l(t) = l1,−3 (t − 1) = (x, y); x − 3y = − , t = 7 : −1 : −21.
8 4
Image Processing 181

These parallel rays are shown in Figure 3.62.

l(2) l(3) l(4) l(5) l(6) l(7)

0

0.25 l(7)
1
l(6)
l(5)
0.25 l(4)
2
l(3)
l(2)
0.375 l(1)
3
l(0)
l(−1)
y and m

0.5 l(−2)
4
l(−3)
l(−4)
0.625 l(−5)
5
l(−6)
l(−7)
0.75 l(−8)
6
l(−9)
l(−10)
0.875 l(−11)
7
l(−12)
l(−13)
1 l(−14)
l(−21) l(−20) l(−19) l(−18) l(−17) l(−16)
0 1 7 l(−15)
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n

FIGURE 3.62
Another set of geometrical rays for the (1, −3)-projection.

The system of equations for the line-integrals is described as

√
w(t) = w(˜
l(t)) = 8 10/3 [v(t) + v(t − 1) + v(t − 2)] ,
(3.81)
t = 7 : −1 : −21,

where v(−22) = v(−23) = 0. A similar system of equations was described

in (3.57) for the (3, 1)-projection with integrals w(t) numbered from 0 to 28.
The required
√ sums v(t) of the discrete image are thus calculated by v(t) =
3bt /(8 10), where coefficients bt are calculated recursively

b−21 =w(−21),
b−20 =w(−20) − b−21 ,
(3.82)
b−19 =w(−19) − (b−20 + b−21),
bt =w(t) − (bt−1 + bt−2 ), t = −18, −17, . . ., 6, 7.

When using the numbering of the control points and rays from 0 to
182 CHAPTER 3: Image Reconstruction

28, as suggested in Remark 5.1, the solution of the system in (3.57) can
be written exactly as in the (p, s) = (3, 1) case (see (3.60)). Then, the
system
√ of equations in (3.77) can be described as w̄(t) = w(˜l(t)) =
8 10/3 [v̄(t) + v̄(t + 1) + v̄(t + 2)] , t = 0 : 28, when the coefficients b̄t are
calculated by

b̄28 = w̄(28),
b̄27 = w̄(27) − b̄28 ,
b̄26 = w̄(26) − (b̄27 + b̄28 ),
b̄t = w̄(t) − (b̄t+1 + b̄t+2 ), t = 25, 24, . . ., 1, 0.

Remark 3.6 Since χ01,5,t = (−1)t χ05,1,t, for t = 0 : 3, we can consider the
(−3, 1)-projection for calculating the components of the 2-D paired transform
of the image, whose triplet-numbers are from the subset U (1, 5), or U (5, 1).
Four masks with the first four triplet-numbers (5, 1, t) are
2 3 2 3
1 0 0 0 −1 0 0 0 0 −1 0 0 0 1 0 0
6 0 0 0 1 0 0 0 −1 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 1 0 0 0−1 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7
7, [χ05,1,1 ] =6 0 0 −1 0 0 0 1 0 7
6 7
[χ05,1,0 ] =6
6 −1 0 0 0 1 0 0 0 7 6 0 1 0 0 0 −1 0 0 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
4 0 0 1 0 0 0 −1 0 5 4 0 0 0 −1 0 0 0 1 5
0 −1 0 0 0 1 0 0 0 0 1 0 0 0 −1 0
2 3 2 3
0 0 1 0 0 0 −1 0 0 0 0 −1 0 0 0 1
6 0 −1 0 0 0 1 0 0 7 6 0 0 1 0 0 0 −1 0 7
6 7 6 7
6 1 0 0 0 −1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 0 0 1 0 0 0 −1 7
7, [χ05,1,3 ] =6 1 0 0 0 −1 0 0 0 7
6 7
[χ05,1,2 ] =6
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 1 0 0 0 −1 7
6 7 6 7
6 0 1 0 0 0 −1 0 0 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
4 −1 0 0 0 1 0 0 0 5 4 0 1 0 0 0 −1 0 0 5
0 0 0 −1 0 0 0 1 −1 0 0 0 1 0 0 0
In each of these masks, three 1’s are underlined, to show the direction of the
rays l−3,1 (t). The set of 29 geometrical rays for this projection is described as
 
t 1
l(t) = l−3,1 (t) = (x, y); −3x + y = − , t = 7 : −1 : −21.
8 8

For this set of rays, we consider the following set of control points:
20
• −3 • −6 • −9 • −12 • −15 • −18 • −21 •
3
6 1 −2 −5 −8 −11 −14 −17 −20 7
6 2 • −1 • −4 • −7 • −10 • −13 • −16 • −19 • 7
6 • • • • • • • •7
63 7
6 • . . . . . . .7
64
6 • . . .
7 (angle of rays is tan−1 (3) = 71.565◦ ).
65 . . . .77
6 • . . . . . . .7
66 7
4 • . . . . . . .5
7
• . . . . . . .
Image Processing 183

It is clear that the diagram of these control points is similar to the dia-
gram of the control points considered above for the (1, −3)-projection. Namely,
these two diagrams are equivalent up to operation of the matrix transposition.
Therefore, our future calculations are similar to the (1, −3) case.
The complete system of linear equations for calculating the components of
the 2-D paired transform with triplet-numbers of the subset U (5, 1), or U (1, 5)
can be written as follows:
0 0
f5,1,0 = f1,5,0 =a0 − a4
0 0
f5,1,1 = −f1,5,1 =a1 − a5
0 0
f5,1,2 = f1,5,2 =a2 − a6
0 0
f5,1,3 = −f1,5,3 =a3 − a7
0 (3.83)
f2,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 )

where variables an are defined as

X X
an = vn−8m = v−3,1,n−8m, n = 0 : 7.
m m

These calculations correspond to the 8-point discrete paired transform of

the vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(5, 1) = P(−3, 1) = [χ08 ]a,

which can also be written as P(1, 5) = diag{1, −1, 1, −1, 1, 1, 1, 1, }[χ08]a.

The set of all rays for the (−3, 1)-projection is shown in Figure 3.63.
0
The components f5,1,t , t = 0 : 3, are calculated by the same system of
equations as in (3.75), where the sums vt are defined as v−3,1 (t), where t =
7 : −1 : −21. For instance, for the mask [χ05,1,0] shown in Figure 3.64, we have
the following:
0
f5,1,0 =−v4 + v0 − v−4 + v−8 − v−12 + v−16 − v−20
=(v0 + v−8 + v−16 ) − (v4 + v−4 + v−12 + v−20 ).

As in the (1, 3)-projection case, the set of 29 geometrical rays can be defined
as the shifted set of the arithmetical rays, l(t) → l(t + 1),
 
˜ t
l(t) = l−3,1 (t + 1) = (x, y); −3x + y = , t = 7 : −1 : −21,
8
which are shown in Figure 3.65. We obtain the following system of equations
for the line-integrals:
√
w(t) = w(˜l(t)) = 8 10/3 [v(t) + v(t + 1) + v(t + 2)] ,
(3.84)
t = 7 : −1 : −21,
184 CHAPTER 3: Image Reconstruction

l(1) l(0) l(−1) l(−3) l(−5) l(−7) l(−9) l(−11)l(−13)l(−15)l(−17)l(−19)l(−21)

0

l(2)
0.25

l(3)
0.25

l(4)
0.375

l(5)
y and m

0.5

l(6)
0.625

l(7)
0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.63
The set of arithmetical rays for the (−3, 1)-projection.

where v(8) = v(9) = 0. The required sums

√ v(t) of the discrete image fn,m
are therefore defined as v(t) = 3bt /(8 10), where the components bt are
calculated recursively by

b7 =w(7),
b6 =w(6) − b7 ,
(3.85)
b5 =w(5) − (b6 + b7 ),
bt =w(t) − (bt+1 + bt+2 ), t = 4, 3, . . ., −20, −21.

We now consider another set of control points for the (−3, 1)-projection,
 
. . . . . . . 0•
 . . . . . . . 1•
 
 . . . . . . . 2•
 
 . . . . . . . 3• −1 1 ◦
 . . . . . . . 4 •  (angle of rays is tan ( 3 = 18.435 ).
 
 26 23 20 17 14 11 8 5 
 • • • • • • • •
 27 24 21 18 15 12 9 6 
 • • • • • • • •
28 25 22 19 16 13 10 7
• • • • • • • •
Image Processing 185
l(0) l(−4) l(−8) l(−12) l(−16) l(−20)
0

0.25

0.25

l(4)
0.375

3
y and m

0.5

0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.64
The mask [χ05,1,0] with the rays of the (−3, 1)-projection.

The set of 29 parallel rays with these control points is shown in Figure 3.66,
after shifting them to the right, t → t − 1, to compose the geometrical rays.
The components of the paired transform with the triplet-numbers (5, 1, t)
are calculated as follows:
0
f5,1,3 = v0 − v4 + v8 − v12 + v16 − v20 + v24 − v28
= (v0 + v8 + v16 + v24 ) − (v4 + v12 + v20 + v28 ),
0
f5,1,0 = −v1 + v5 − v9 + v13 − v17 + v21 − v25
= −(v1 + v9 + v17 + v25 ) + (v5 + v13 + v21 ),
0 (3.86)
f5,1,1 = −v2 + v6 − v10 + v14 − v18 + v22 − v26
= −(v2 + v10 + v18 + v26 ) + (v6 + v14 + v22 ),
0
f5,1,2 = −v3 + v7 − v11 + v15 − v19 + v23 − v27
= −(v3 + v11 + v19 + v27 ) + (v7 + v15 + v23 ),

where the sums of the discrete image along the rays are denoted by vt , t = 0 :
28. Introducing new variables
X X
an = vn+8m = v−3,1,n+8m, n = 0 : 7,
m m

0
we can write the complete system of linear equations (3.86) as f5,1,0 = −a1 +
0 0 0
a5 , f5,1,1 = −a2 + a6 , f5,1,2 = −a3 + a7 , and f5,1,3 = a0 − a4 .
186 CHAPTER 3: Image Reconstruction

l(0) l(−1) l(−3) l(−5) l(−7) l(−9) l(−11)l(−13)l(−15)l(−17)l(−19)l(−21)

0

l(1)
0.25

l(2)
0.25

l(3)
0.375

l(4)
y and m

0.5

l(5)
0.625

l(6)
0.75

l(7)
0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.65
The set of geometrical rays for the (−3, 1)-projection.

The next four components of the 2-D paired transform are

0
f2,2,0 = v1 + v5 + v9 + v13 + v17 + v21 + v25
−v3 − v7 − v11 − v15 − v19 − v23 − v27
= (v1 + v9 + v17 + v25 ) + (v5 + v13 + v21 )
−(v3 + v11 + v19 + v27 ) − (v7 + v15 + v23 )
= (a1 + a5 ) − (a3 + a7 ),
0
f2,2,2 = −v0 − v4 − v8 − v12 − v16 − v20 − v24 − v28
+v2 + v6 + v10 + v14 + v18 + v22 + v26
(3.87)
= −(v0 + v8 + v16 + v24 ) − (v4 + v12 + v20 + v28 )
+(v2 + v10 + v18 + v26 ) + (v6 + v14 + v22 )
= −(a0 + a4 ) + (a2 + a6 ),
0
f4,4,0 = −v0 + v1 − v2 + v3 − v4 + v5 − ... − v26 + v27 − v28
= −(a0 + a4 ) + (a1 + a5 ) − (a2 + a6 ) + (a3 + a7 ),
0
f0,0,0 = v0 + v1 + v2 + v3 + v4 + v5 + ... + v26 + v27 + v28
= (a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
Image Processing 187

l(23) l(22) l(20) l(18) l(16) l(14) l(12) l(10) l(8) l(6) l(4) l(2) l(0)
0

l(24)
0.25

l(25)
0.25

l(26)
0.375

l(27)
y and m

0.5

l(28)
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.66
The set of geometrical rays for the (−3, 1)-projection.

Thus, we obtain the following algorithm:

0
−f5,1,0 = a1 − a5
0
−f5,1,1 = a2 − a6
0
−f5,1,2 = a3 − a7
0
f5,1,3 = a0 − a4
0 (3.88)
f2,2,0 = (a1 + a5 ) − (a3 + a7 )
0
−f2,2,2 = (a0 + a4 ) − (a2 + a6 )
0
−f4,4,0 = (a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f4,0,0 = (a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).

The calculations correspond to the 8-point discrete paired transform of the

vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(5, 1) = diag{−1, −1, −1, 1, 1, −1, −1, 1}[χ08]T a,

where T is the matrix of the following permutation of the vector:

T : {{a0 , a1 , a2 , a3 }, {a4 , a5 , a6 , a7 }} → {{a1 , a2 , a3 , a0 }, {a5, a6 , a7 , a4 }}.

This set of geometrical rays corresponds to the case considered above (see
188 CHAPTER 3: Image Reconstruction

Figure 3.63), when 29 arithmetical rays l(t), t = 7 : −1 : −21, are shifted by

t → t − 1, to compose the set of geometrical rays,
 
t 1
l̃(t) = l−3,1 (t − 1) = (x, y); −3x + y = − , t = 7 : −1 : −21,
8 4
which are shown in Figure 3.67. For this set of geometrical rays, the system

l(2) l(1) l(−1) l(−3) l(−5) l(−7) l(−9) l(−11)l(−13)l(−15)l(−17)l(−19)l(−21)

0

l(3)
0.25

l(4)
0.25

l(5)
0.375

l(6)
y and m

0.5

l(7)
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.67
The set of geometrical rays for the (−3, 1)-projection.

of equations for the line-integrals is

√
w(t) = w(˜ l(t)) = 8 10/3 [v(t) + v(t − 1) + v(t − 2)] ,
(3.89)
t = 7 : −1 : −21,
where v(−22) = v(−23) = 0. The √ required sums v(t) of the discrete image
are calculated by v(t) = 3bt /(8 10), where coefficients bt are calculated re-
cursively
b−21 =w(−21),
b−20 =w(−20) − b−21 ,
(3.90)
b−19 =w(−19) − (b−20 + b−21 ),
bt =w(t) − (bt−1 + bt−2 ), t = −18, −17, . . . , 6, 7.
Image Processing 189

3.5.8 (1, 6)-projection

The set of triplet-numbers generated by the frequency-point (p, s) = (1, 6) is


 (1, 6, 0), (1, 6, 1), (1, 6, 2), (1, 6, 3),

(2, 4, 0), (2, 4, 2),
U (1, 6) =

 (4, 0, 0),

(0, 0, 0).

The masks of the 2-D paired functions with the first six triplet-numbers
of the subset U (1, 6) are
 
1 0 0 0 −1 0 0 0  
 0 0 1 0 0 0 −1 0  0 1 0 00 −1 0 0
   0 0 0 10 0 0 −1 
 −1 0 0 0 1 0 0 0   
   0 −1 0 00 1 0 0 
 −1 0 0 0 1 0 0 0   
 , [χ0 ] =  0 0 0 −1 0 0 0 1  ,
   
[χ01,6,0] = 
 0 0 −1 0 0 0 1 0  1,6,1  0 1 0 00 −1 0 0 
 1 0 0 0 −1 0 0 0   
   0 0 0 10 0 0 −1 
 0 0 1 0 0 0 −1 0   
   0 −1 0 00 1 0 0 
 −1 0 0 0 1 0 0 0 
0 0 0 −1 0 0 0 1
0 0 −1 0 0 0 1 0
   
0 0 1 0 0 0 −1 0 0 0 0 10 0 0 −1
 −1 0 0 0 1 0 0 0   0 −1 0 00 1 0 0 
   
 0 0 −1 0 0 0 1 0   0 0 0 −1 0 0 0 1 
   
0
 1 0 0 0 −1 0 0 0  0
 0 1 0 00 −1 0 0 
[χ1,6,2] =  , [χ1,6,3] =  ,
 0 0 1 0 0 0 −1 0   0 0 0 10 0 0 −1 
 
 −1 0 0 0 1 0 0 0   0 −1 0 00 1 0 0 
   
 0 0 −1 0 0 0 1 0   0 0 0 −1 0 0 0 1 
1 0 0 0 −1 0 0 0 0 1 0 00 −1 0 0
   
1 0 −1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
 −1 0 1 0 −1 0 1 0   0 −1 0 10 −1 0 1 
   
 1 0 −1 0 1 0 −1 0   0 1 0 −1 0 1 0 −1 
   
 −1 0 1 0 −1 0 1 0   0 −1 0 10 −1 0 1 
[χ02,4,0] = 
 1 0 −1 0 1 0 −1 0  ,
 [χ02,4,2] = 
 0 1 0 −1 0 1 0 −1  .

   
 −1 0 1 0 −1 0 1 0   0 −1 0 10 −1 0 1 
   
 1 0 −1 0 1 0 −1 0   0 1 0 −1 0 1 0 −1 
−1 0 1 0 −1 0 1 0 0 −1 0 10 −1 0 1

Instead of the parallel rays corresponding to the (1, 6)-projection, we con-

sider the (1, −2)-projection which defines the same set of masks of the paired
functions χ0p1 ,s1 ,t1 , (p1 , s1 , t1 ) ∈ U (1, 6). The change in direction of the rays
allows for reducing the total number of rays from (1 + 6)7 + 1 = 50 to
(1 + 2)7 + 1 = 22. As an example, Figures 3.68 and 3.69 show the mask
of the paired function χ01,6,0 with the parallel arithmetical rays of two diffe-
rent projections. Thirteen rays of the (1, 6)-projection are shown in the first
190 CHAPTER 3: Image Reconstruction

l(0) l(4)
0

0 l(8)
0.25
l(12)
1
l(16)
0.25

2 l(20)
0.375
l(24)
3
y and m

l(28)
0.5

4 l(32)
0.625
l(36)
5
l(40)
0.75

6 l(44)
0.875 l(48)

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.68
The mask [χ01,6,0] with the set of 13 parallel arithmetical rays.

figure, and 5 rays of the (1, −2)-projection are shown in the second figure.
The set of all 22 rays for this projection is shown in Figure 3.70, and it is
described by the following system of equations:
 
t 1
l(t) = l1,−2 (t) = (x, y); x − 2y = − , t = 7 : −1 : −14, (3.91)
8 16
where (x, y) ∈ [0, 1] × [0, 1],
The control points of this set of parallel rays are given below
 0 
• 1• 2 • 3• 4 • 5• 6 • 7•
 −2 • −1 • . . . . . . 
 −4 −3 
 −6 • −5 • . . . . . . 
 
• • . . . . . .  (angle of rays is tan−1 ( 1 ) = 26.5651◦).
 
 −8 −7
 •
 −10 −9 • . . . . . . 
 2
 •
 −12 −11 • . . . . . . 

 • • . . . . . .
−14 −13
• • . . . . . .

The first six masks for the paired functions which are defined by the (1, 6)-
projection are shown in Figures 3.71 and 3.72.
Image Processing 191

l(4)
0

0.25

0.25
l(4)
2

0.375

3
y and m

0.5
l(0)
4

0.625

0.75
l(−4)
6

0.875

1
l(−12) l(−8)
0 1 2 3 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.69
The mask [χ01,6,0] with the set of 5 parallel arithmetical rays.

The components of the 2-D paired transform with the triplet-numbers

(1, 6, t) and (2, 4, 2t) are calculated by
0
f1,6,0 =(v0 + v−8 ) − (v4 + v−4 + v−12 ),
0
f1,6,1 =(v1 + v−7 ) − (v5 + v−3 + v−11 ),
0
f1,6,2 =(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 ),
0
f1,6,3 =(v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),
(3.92)
0
f2,4,0 =(v0 + v−8 ) + (v4 + v−4 + v−12 )−
−(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 ),
0
f2,4,2 =(v1 + v−7 ) + (v5 Chapters/chapter3/ + v−3 + v−11 )−
−(v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),

where the sums of the discrete image along the rays are denoted by vt =
v1,−2 (t), t = 7 : −1 : −14. For the remaining two components with the triplet-
192 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0

0
l(7)
0.25
l(6)
1
l(5)
0.25
l(4)
2
l(3)
0.375
l(2)
3
y and m

l(1)
0.5
l(0)
4
l(−1)
0.625
l(−2)
5
l(−3)
0.75
l(−4)
6
l(−5)
0.875
l(−6)
7
l(−7)
1
l(−14) l(−13) l(−12) l(−11) l(−10) l(−9) l(−8)
0 1

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.70
The set of arithmetical rays for the (1, −2)-projection.

numbers (4, 0, 0) and (0, 0, 0) we have the following:

0
f4,0,0 = (v0 + v−8 ) + (v4 + v−4 + v−12 ) − (v1 + v−7 ) − (v5 + v−3 + v−11 )
+(v2 + v−6 + v−14 ) + (v6 + v−2 + v−10 )−
−(v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),
0
f0,0,0 = v0 + v1 + v2 + v3 + · · · +
+v7 + v−1 + v−2 + v−3 + v−4 + v−5 + ... + v−13 + v−14 .

With the new variables

X X
an = vn−8m = v1,−2,n−8m, n = 0 : 7,
m m

the complete system of linear equations, for calculating the eight components
of the 2-D paired transform with triplet-numbers of the subset U (1, 6), can be
Image Processing 193

0
l(0) l(4) 0
l(1) l(5)
0 0
0.25 0.25
1 1
0.25 0.25 l(5)
2 l(4) 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5 l(1)
4 l(0) 4
0.625 0.625
5 5 l(−3)
0.75 l(−4) 0.75
6 6
0.875 0.875
7 7 l(−7)
1 1
0 1 2 l(−12) 6 7 l(−8) 0 1 2 3 l(−11) 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

0
l(2) l(6) 0
l(3) l(7)
0 0
0.25 0.25 l(7)
1 l(6) 1
0.25 0.25
2 2
0.375 0.375 l(3)
y and m

3 l(2) y and m 3
0.5 0.5
4 4 l(−1)
0.625 l(−2) 0.625
5 5
0.75 0.75
6 6 l(−5)
0.875 l(−6) 0.875
7 7
1 1
0 1 l(−14) 4 l(−10) 0 1 2 l(−13) 5 l(−9)
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.71
Masks [χ01,6,k ], k = 0 : 3, with the arithmetical rays l1,−2 (t), t = 7 : −1 : 14.

written as follows:
0
f1,6,0 =a0 − a4
0
f1,6,1 =a1 − a5
0
f1,6,2 =a2 − a6
0
f1,6,3 =a3 − a7
0 (3.93)
f2,4,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,4,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,0,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).

These calculations correspond to the 8-point discrete paired transform of the

vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(1, 6) = P(1, −2) = [χ08 ]a.

As in the case of the (1, 2)-projection, two sets of geometrical rays can be
194 CHAPTER 3: Image Reconstruction

l(2) l(4) l(6) l(1) l(3) l(5) l(7)

0 0
0 0
l(7)
0.25 l(6) 0.25
1 1 l(5)
l(4)
0.25 0.25 l(3)
2 l(2) 2
l(1)
0.375 l(0) 0.375
l(−1)
y and m

y and m
3 3
l(−2)
0.5 0.5 l(−3)
4 l(−4) 4
l(−5)
0.625 l(−6) 0.625
5 5 l(−7)
l(−8)
0.75 0.75 l(−9)
6 l(−10) 6
l(−11)
0.875 l(−12) 0.875
7 7 l(−13)
l(−14)
1 1 l(−15)
0 1 2 l(−20) l(−18) l(−16) 0 1 l(−21) l(−19) l(−17)
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
x and n x and n

FIGURE 3.72
Two masks [χ02,4,2k ], k = 0, 1, with the arithmetical rays l1,−2 (t).

considered to derive a solvable system of linear equations between geometri-

cal and arithmetical rays. The first set is defined by shifting the set of the
arithmetical rays by l(t) → l(t + 0.5). This set of 22 rays is
 
˜ t
l(t) = l1,−2 (t + 0.5) = (x, y); x − 2y = , t = 7 : −1 : −14,
8

and is shown in Figure 3.73. For this set of geometrical rays, we obtain the
following system of equations for the line-integrals:
√
w(t) = w(˜
l(t)) = 4 5 [v(t) + v(t + 1)] , t = 7 : −1 : −14, (3.94)

where v(8) = 0. The required

√ sums v(t) of the discrete image fn,m are therefore
defined as v(t) = bt /(4 5), where values of bt are calculated recursively by

b7 =w(7),
b6 =w(6) − b7 , (3.95)
bt =w(t) − bt+1 , t = 5, 4, . . . , 0, 1, −1, −2, ..., −14.

The second set is defined by shifting the set of the arithmetical rays by
l(t) → l(t − 0.5),
 
t 1
l̃(t) = l1,−2 (t − 0.5) = (x, y); x − 2y = − , t = 7 : −1 : −14,
8 8

which is given in Figure 3.74. The system of equations for the line-integrals is
defined as
√
w(t) = w(˜l(t)) = 4 5 [v(t) + v(t − 1)] , t = 7 : −1 : −14, (3.96)
Image Processing 195

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0

0 l(7)
0.25 l(6)
1 l(5)
0.25 l(4)
2 l(3)
0.375 l(2)
3 l(1)
y and m

0.5 l(0)
4 l(−1)
0.625 l(−2)
5 l(−3)
0.75 l(−4)
6 l(−5)
0.875 l(−6)
7 l(−7)
1 l(−8)
l(−14) l(−13) l(−12) l(−11) l(−10) l(−9)
0 1

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.73
The set of 22 geometrical rays for the (1, −2)-projection.

where v(−15) = 0. The required √ sums v(t) of the discrete image fn,m are
therefore defined as v(t) = bt /(4 5), where the components bt are calculated
recursively by
b−14 =w(−14),
b−13 =w(−13) − b−14 , (3.97)
bt =w(t) − bt−1 , t = −12, −11, . . . , 0, 1, ..., 7.
It should also be noted that the numbering of the control points can be
changed, as well as the numbering of all line-integrals w(t) and sums vt = v(t)
by the transformation t → 7 − t, where t = 7 : −1 : −21,
 7 6 5 4 3 2 1 0 
• • • • • • • •
 9• 8• . . . . . .
 11 10 
 • • . . . . . .
 13 12 
 • • . . . . . .
 • • . . . . . . .
 15 14 
 17 16 
 • • . . . . . .
 19 18 
 • • . . . . . .
21 20
• • . . . . . .
196 CHAPTER 3: Image Reconstruction

l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0
l(0)
0

0.25 l(7)
1 l(6)
0.25 l(5)
2 l(4)
0.375 l(3)
3 l(2)
y and m

0.5 l(1)
4 l(0)
0.625 l(−1)
5 l(−2)
0.75 l(−3)
6 l(−4)
0.875 l(−5)
7 l(−6)
1 l(−7)
l(−14) l(−13) l(−12) l(−11) l(−10) l(−9) l(−8)
0

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.74
The second set of 22 geometrical rays for the (1, −2)-projection.

3.5.9 (6, 1)-projection

The set of triplet-numbers generated by the frequency-point (p, s) = (6, 1) is


 (6, 1, 0), (6, 1, 1), (6, 1, 2), (6, 1, 3),

(4, 2, 0), (4, 2, 2),
U (6, 1) =

 (0, 4, 0),

(0, 0, 0).

The masks of the 2-D paired functions with first six triplet-numbers of the
subset U (6, 1) are
2 3 2 3
1 0 −1 0 1 0 −1 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 0 −1 0 1 0 −1 0 7
6 7 6 7
6 0 1 0 −1 0 1 0 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7, [χ06,1,1 ] =6 0 1 0 −1 0 1 0 −1 7
6 0 0 0 0 0 0 0 07 6 7
[χ06,1,0 ] =6
6 −1 0 1 0 −1 0 1 0 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 −1 0 1 0 −1 0 1 0 7
6 7 6 7
4 0 −1 0 1 0 −1 0 1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 −1 0 1 0 −1 0 1
Image Processing 197

2 3 2 3
0 −1 0 1 0 −1 0 1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 0 −1 0 1 0 −1 0 1 7
6 7 6 7
6 1 0 −1 0 1 0 −1 0 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7, [χ06,1,3 ] =6 1 0 −1 0 1 0 −1 0 7
6 0 0 0 0 0 0 0 07 6 7
[χ06,1,2 ] =6
6 0 1 0 −1 0 1 0 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 0 1 0 −1 0 1 0
2 3 2 3
1 −1 1 −1 1 −1 1 −1 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 07 6 1 −1 1 −1 1 −1 1 −1 7
6 7 6 7
6 −1 1 −1 1 −1 1 −1 1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
7, [χ04,2,2 ] =6 −1 1 −1 1 −1 1 −1 1 7.
6 0 0 0 0 0 0 0 07 6 7
[χ04,2,0 ] =6
6 1 −1 1 −1 1 −1 1 −1 7 6 0 0 0 0 0 0 0 07
6 7 6 7
6 0 0 0 0 0 0 0 07 6 1 −1 1 −1 1 −1 1 −1 7
6 7 6 7
4 −1 1 −1 1 −1 1 −1 1 5 4 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 −1 1 −1 1 −1 1 −1 1

Instead of parallel rays of the (6, 1)-projection, we consider the (−2, 1)-
projection which defines the same set of masks of the paired functions. Indeed,
the equation 6n + m = t mod 8 can be written as (8 − 2)n + m = −2n + m =
t mod 8. The change of the projection allows for reducing the number of rays
from 50 to 22. The set of all parallel rays for this projection is shown in
Figure 3.75.
The control points of this set of rays are given below
 0 −2 −4 −6 −8 −10 −12 −14 
• • • • • • • •
 1 • −1 • −3 • −5 • −7 • −9 • −11 • −13 • 
2 
 • . . . . . . .
3 
 • . . . . . . .
4  (angle of rays is tan−1 (2) = 63.435◦)
 • . . . . . . . 
5 
 • . . . . . . .
6 
 • . . . . . . .
7
• . . . . . . .

The set of the 29 geometrical rays of this projection is described as

 
t 1
l(t) = l−2,1 (t) = (x, y); −2x + y = − , t = 7 : −1 : −14, (3.98)
8 16

where (x, y) ∈ [0, 1]×[0, 1]. The masks with parallel rays for the first six paired
functions are shown in Figures 3.76 and 3.77.
The components of the 2-D paired transform with the triplet-numbers of
198 CHAPTER 3: Image Reconstruction

l(0) l(−2) l(−4) l(−6) l(−8) l(−10) l(−12) l(−14)

0 l(1)

0.25 l(2)

0.25 l(3)

0.375 l(4)

3
y and m

0.5 l(5)

0.625 l(6)

0.75 l(7)

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.75
The set of arithmetical rays for the (−2, 1)-projection.

U (6, 1) are calculated as follows:

0
f6,1,0 = (v0 + v−8 ) − (v4 + v−4 + v−12 ),
0
f6,1,1 = (v1 + v−7 ) − (v5 + v−3 + v−11 ),
0
f6,1,2 = (v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 ),
0
f6,1,3 = (v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),
(3.99)
0
f4,2,0 = (v0 + v−8 ) + (v4 + v−4 + v−12 )−
−(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 ),
0
f4,2,2 = (v1 + v−7 ) + (v5 + v−3 + v−11 )−
−(v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),
where the sums of the discrete image along the rays are denoted by vt =
v−2,1 (t), t = 7 : −1 : −14. For the components with two remaining numbers
(0, 4, 0) and (0, 0, 0), we have the following:
0
f0,4,0 =(v0 + v−8 ) + (v4 + v−4 + v−12 ) − (v1 + v−7 )−
−(v5 + v−3 + v−11 ) + (v2 + v−6 + v−14 ) + (v6 + v−2 + v−10 )
−(v3 + v−5 + v−13 ) − (v7 + v−1 + v−9 ),
0
f0,0,0 =v0 + v1 + v2 + v3 + · · · + v7 +
+v−1 + v−2 + v−3 + v−4 + v−5 + ... + v−13 + v−14 .
Image Processing 199

l(−4) l(−8) l(−12) l(−3) l(−7) l(−11)

0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 1 l(4) 3 l(0) 5 6 7 0 l(5) l(1) 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

l(−2) l(−6) l(−10) l(−14) l(−1) l(−5) l(−9) l(−13)

0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m
3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 l(6) 2 l(2) 4 5 6 7 0 l(7) l(3) 4 5 6 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.76
Four masks [χ06,1,k ], k = 0 : 3, with the arithmetical rays l−2,1 (t).

l(−2) l(−4) l(−6) l(−8) l(−10)l(−12) l(−14) l(−1) l(−3) l(−5) l(−7) l(−9) l(−11) l(−13)
0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m

3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 l(6) l(4) l(2) l(0) 5 6 7 0 l(7) l(5) l(3) l(1) 5 6 7
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 3.77
Two masks [χ04,2,2k ], k = 0, 1, with the arithmetical rays l−2,1 (t).
200 CHAPTER 3: Image Reconstruction

With the new variables

X X
an = vn−8m = v−2,1,n−8m, n = 0 : 7,
m m

the complete system of linear equations, for calculating the eight components
of the 2-D paired transform with triplets of the subset U (6, 1), can be written
as follows:
0
f6,1,0 =a0 − a4
0
f6,1,1 =a1 − a5
0
f6,1,2 =a2 − a6
0
f6,1,3 =a3 − a7
0 (3.100)
f4,2,0 =(a0 + a4 ) − (a2 + a6 )
0
f4,2,2 =(a1 + a5 ) − (a3 + a7 )
0
f0,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).

These calculations correspond to the 8-point discrete paired transform of the

vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(6, 1) = P(−2, 1) = [χ08 ]a.

Two sets of geometrical rays can be used to determine the sums of the
image along the arithmetical rays. The first set is defined by shifting the set
of arithmetical rays by l(t) → l(t + 0.5). This set of 22 rays is described by
 
˜ t
l(t) = l−2,1 (t + 0.5) = (x, y); −2x + y = , t = 7 : −1 : −14,
8
and is shown in Figure 3.78. The line-integrals along this set of geometrical
rays can be written as
√
w(t) = w(˜l(t)) = 4 5 [v(t) + v(t + 1)] , t = 7 : −1 : −14, (3.101)

where v(8) = 0.√The required sums v(t) of the image fn,m are therefore defined
as v(t) = bt /(4 5), where the components bt are calculated recursively by

b7 = w(7),
b6 = w(6) − b7 , (3.102)
bt = w(t) − bt+1 , t = 5, 4, . . . , 0, 1, −1, −2, ..., −14.

The second set of geometrical rays is defined by shifting the set of the
arithmetical rays by l(t) → l(t − 0.5),
 
t 1
l̃(t) = l−2,1 (t − 0.5) = (x, y); −2x + y = − , t = 7 : −1 : −14,
8 8
which is given in Figure 3.79. In this case, the system of equations for the
Image Processing 201

l(0) l(−2) l(−4) l(−6) l(−8) l(−10) l(−12) l(−14)

0

l(1)
0.25

l(2)
0.25

l(3)
0.375

l(4)
y and m

0.5

l(5)
0.625

l(6)
0.75

l(7)
0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.78
The set of geometrical rays for the (−2, 1)-projection.

line-integrals is described by
√
w(t) = w(˜ l(t)) = 4 5 [v(t) + v(t − 1)] , t = 7 : −1 : −14, (3.103)

where v(−15) = 0. The required √ sums v(t) of the discrete image fn,m are
therefore defined as v(t) = bt /(4 5), where the components bt are calculated
recursively by

b−14 =w(−14),
b−13 =w(−13) − b−14 , (3.104)
bt =w(t) − bt−1 , t = −12, −11, . . . , 0, 1, ..., 7.

It should be noted that the numbering of the control points can be changed,
as well as the numbering of all line-integrals w(t) and sums vt = v(t) by the
202 CHAPTER 3: Image Reconstruction

l(1) l(0) l(−2) l(−4) l(−6) l(−8) l(−10) l(−12) l(−14)

0

l(2)
0.25

l(3)
0.25

l(4)
0.375

l(5)
y and m

0.5

l(6)
0.625

l(7)
0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.79
The second set of geometrical rays for the (−2, 1)-projection.

transform t → 7 − t, where t = 7 : −1 : −21,

 7 9 11 13 15 17 19 21 
• • • • • • • •
 6 • 8 • 10 • 12 • 14 • 16 • 18 • 20 • 
5 
 • . . . . . . .
4 
 • . . . . . . . −1 ◦
 • . . . . . . .  (angle of rays is tan (2) = 63.435 ).
3 
2 
 • . . . . . . .
1 
 • . . . . . . .
0
• . . . . . . .

3.5.10 (1, 7)-projection

This is the last set of triplet-numbers in the 8×8 example. The set is generated
by the frequency-point (p, s) = (1, 7),


 (1, 7, 0), (1, 7, 1), (1, 7, 2), (1, 7, 3),

(2, 6, 0), (2, 6, 2),
U (1, 7) =

 (4, 4, 0),

(0, 0, 0).
Image Processing 203

We consider masks of the 2-D paired functions with the first six triplet-
numbers of the subset U (1, 7),
2 3 2 3
1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0 0
6 0 1 0 0 0 −1 0 0 7 6 0 0 1 0 0 0 −1 0 7
6 7 6 7
6 0 0 1 0 0 0 −1 0 7 6 0 0 0 1 0 0 0−1 7
6 7 6 7
7, [χ01,7,1 ] =6 −1 0 0 0 1 0 0 0 7
6 0 0 0 1 0 0 0−1 7 6 7
[χ01,7,0 ] =6
6 −1 0 0 0 1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 −1 0 0 0 1 0 0 7 6 0 0 −1 0 0 0 1 0 7
6 7 6 7
4 0 0 −1 0 0 0 1 0 5 4 0 0 0 −1 0 0 0 1 5
0 0 0 −1 0 0 0 1 1 0 0 0 −1 0 0 0
2 3 2 3
0 0 1 0 0 0 −1 0 0 0 0 1 0 0 0−1
6 0 0 0 1 0 0 0−1 7 6 −1 0 0 0 1 0 0 0 7
6 7 6 7
6 −1 0 0 0 1 0 0 0 7 6 0 −1 0 0 0 1 0 0 7
6 7 6 7
6 0 −1 0 0 0 1 0 0 7
7, [χ01,7,3 ] =6 0 0 −1 0 0 0 1 0 7
6 7
[χ01,7,2 ] =6
6 0 0 −1 0 0 0 1 0 7 6 0 0 0 −1 0 0 0 1 7
6 7 6 7
6 0 0 0 −1 0 0 0 1 7 6 1 0 0 0 −1 0 0 0 7
6 7 6 7
4 1 0 0 0 −1 0 0 0 5 4 0 1 0 0 0 −1 0 0 5
0 1 0 0 0 −1 0 0 0 0 1 0 0 0 −1 0
2 3 2 3
1 0 −1 0 1 0 −1 0 0 1 0 −1 0 1 0 −1
6 0 1 0 −1 0 1 0 −1 7 6 −1 0 1 0 −1 0 1 0 7
6 7 6 7
6 −1 0 1 0 −1 0 1 0 7 6 0 −1 0 1 0 −1 0 1 7
6 7 6 7
6 0 −1 0 1 0 −1 0 1 7
7, [χ02,6,2 ] =6 1 0 −1 0 1 0 −1 0 7
6 7
[χ02,6,0 ] =6
6 1 0 −1 0 1 0 −1 0 7 6 0 1 0 −1 0 1 0 −1 7
6 7 6 7
6 0 1 0 −1 0 1 0 −1 7 6 −1 0 1 0 −1 0 1 0 7
6 7 6 7
4 −1 0 1 0 −1 0 1 0 5 4 0 −1 0 1 0 −1 0 1 5
0 −1 0 1 0 −1 0 1 1 0 −1 0 1 0 −1 0

Instead of the parallel rays of the (1, 7)-projection, we consider the (1, −1)-
projection which defines the same set of masks of paired functions. This change
allows for reducing the number of rays from (1+7)7+1 = 57 to (1+1)7+1 = 15.
As an example, Figures 3.80 and 3.81 show the same mask of the paired
function χ01,7,0 with the parallel arithmetical rays of two different projections.
Fifteen rays of the (1, 7)-projection are shown in the first figure, and three
rays of the (1, −1)-projection are shown in the second figure.
The set of the 15 geometrical rays of this projection is described by
 
t
l(t) = l1,−1 (t) = (x, y); x − y = , t = 7 : −1 : −7, (3.105)
8

where (x, y) ∈ [0, 1] × [0, 1]. This set is shown in Figure 3.82. The control
204 CHAPTER 3: Image Reconstruction

l(0) l(4)
0
l(8)
0
l(12)
0.25

1 l(16)

0.25 l(20)
2 l(24)
0.375
l(28)
3
l(32)
y and m

0.5
l(36)
4
l(40)
0.625

5 l(44)

0.75 l(48)
6 l(52)
0.875
l(56)
7

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.80
The mask [χ01,7,0] with 15 arithmetical rays of the (1, 7)-projection.

points of this set of rays are shown below:

 0 1 2 3 4 5 6 7 
• • • • • • • •
 −1 • . . . . . . . 
 −2 
 • . . . . . . .
 −3 
 • . . . . . . . −1 ◦
 • . . . . . . .  (angle of rays is tan (1) = 45 ).
 −4 
 −5 
 • . . . . . . .
 −6 
 • . . . . . . .
−7
• . . . . . . .

The first six masks for the paired functions with the triplet-numbers of
U (1, 7) are shown in Figures 3.83 and 3.84.
The components of the 2-D paired transform with the triplet-numbers
Image Processing 205

l(4)
0

0.25

0.25

0.375

3
y and m

0.5
l(4)
4

0.625

0.75

0.875

1 l(0)
l(−4)
0 1 2 3 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.81
The mask [χ01,7,0] with 3 arithmetical rays of the (1, −1)-projection.

(1, 7, t) and (2, 6, 2t) are calculated by

0
f1,7,0 =(v0 ) − (v4 + v−4 ),
0
f1,7,1 =(v1 + v−7 ) − (v5 + v−3 ),
0
f1,7,2 =(v2 + v−6 ) − (v6 + v−2 ),
0 (3.106)
f1,7,3 =(v3 + v−5 ) − (v7 + v−1 ),
0
f2,6,0 =(v0 ) + (v4 + v−4 ) − (v2 + v−6 ) − (v6 + v−2 ),
0
f2,6,2 =(v1 + v−7 ) + (v5 + v−3 ) − (v3 + v−5 ) − (v7 + v−1 ),

where the sums of the discrete image along the rays are denoted by vt =
v1,−1 (t), t = 7 : −1 : −7. For the remaining two components of the 2-D paired
transform, which have the triplet-numbers (4, 4, 0) and (0, 0, 0), we have the
following:
0
f4,4,0 =(v0 ) + (v4 + v−4 ) − (v1 + v−7 ) − (v5 + v−3 )+
+(v2 + v−6 ) + (v6 + v−2 ) − (v3 + v−5 ) − (v7 + v−1 ),
0
f0,0,0 =v0 + v1 + v2 + v3 + · · · + v7 + v−1 + v−2 + ... + v−6 + v−7 .
206 CHAPTER 3: Image Reconstruction

l(0) l(1) l(2) l(3) l(4) l(5) l(6) l(7)

0

0.25
l(7)
1

0.25
l(6)
2

0.375
l(5)
3
y and m

0.5
l(4)
4

0.625
l(3)
5

0.75
l(2)
6

0.875
l(1)
7

1
l(−7) l(−6) l(−5) l(−4) l(−3) l(−2) l(−1) l(0)
0

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.82
The set of arithmetical rays for the (1, −1)-projection.

Introducing the new variables

X X
an = vn−8m = v1,−1,n−8m, n = 0 : 7,
m m

the complete system of linear equations for calculating the eight components
of the 2-D paired transform with triplet-numbers of the subset U (1, 7) can be
written as follows:
0
f1,7,0 =a0 − a4
0
f1,7,1 =a1 − a5
0
f1,7,2 =a2 − a6
0
f1,7,3 =a3 − a7
0 (3.107)
f2,6,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,6,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,4,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
These calculations correspond to the 8-point discrete paired transform of the
vector a = (a0 , a1 , a2 , ..., a7)0 ,

P(1, 7) = P(1, −1) = [χ08 ]a.

Image Processing 207

0
l(0) l(4) 0
l(1) l(5)
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375 l(5)
y and m

y and m
3 3
0.5 l(4) 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875 l(1)
7 7
1 1
0 1 2 3 l(−4) 6 7 l(0) 0 l(−7)2 3 4 l(−3) 7
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

0
l(2) l(6) 0
l(3) l(7)
0 0
0.25 0.25 l(7)
1 1
0.25 l(6) 0.25
2 2
0.375 0.375
y and m

3
y and m 3
0.5 0.5
4 4
0.625 0.625 l(3)
5 5
0.75 l(2) 0.75
6 6
0.875 0.875
7 7
1 1
0 1 l(−6) 4 5 l(−2) 0 1 2 l(−5) 5 6 l(−1)
0 0.120.250.37 0.5 0.620.750.87 1 0 0.120.250.37 0.5 0.620.750.87 1
x and n x and n

FIGURE 3.83
Four masks [χ01,7,k ], k = 0 : 3, with the arithmetical rays l1,−1 (t).

As in the (p, s) = (1, 1) case of the diagonal projection, the sums v(t) of
the discrete image fn,m can be calculated directly from the line-integrals w(t).
The set of geometrical rays coincides with the set of arithmetical
√ rays.√Each
ray intersects the image element√along the path of length ∆l
√ = 2∆x = 2/8.
Therefore, w(t) = w1,−1 (t) = 8 2v(t), or v(t)
√ = w(t)/(8 2), t = 7 : −1 : −7.
We obtain the following: P(1, 7) = 1/(8 2)[χ08 ]w, where the vector w =
(w7 , w6 , ..., w0, w−1 , ..., w−7)0 .
The numbering of the control points can be changed, as well as the numbe-
ring of all line-integrals w(t) and sums vt = v(t), by using the transformation
208 CHAPTER 3: Image Reconstruction

l(0) l(2) l(4) l(6) l(1) l(3) l(5) l(7)

0 0
0 0
0.25 0.25 l(7)
1 1
0.25 l(6) 0.25
2 2
0.375 0.375
l(5)
y and m

y and m
3 3
0.5 l(4) 0.5
4 4
0.625 0.625 l(3)
5 5
0.75 l(2) 0.75
6 6
0.875 0.875 l(1)
7 7
1 1
0 1 l(−6) 3 l(−4) 5 l(−2) 7 l(0) 0 l(−7) 2 l(−5) 4 l(−3) 6 l(−1)
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 3.84
Two masks [χ02,6,2k ], k = 0, 1, with the arithmetical rays l1,−1 (t).

t → 7 − t, where t = 7 : −1 : −7,
 7 6 5 4 3 2 1 0 
• • • • • • • •
 8• . . . . . . . 
 9 
 • . . . . . . .
 10 
 • . . . . . . .
 11
 • . . . . . . .
 (angle of rays is tan−1 (1) = 45◦).
 12 
 • . . . . . . .
 13 
 • . . . . . . .
14
• . . . . . . .

3.6 Summary of results

For the N = 8 case, all 3N/2 = 12 projections that are required for recon-
structing the image N × N were considered in detail, and geometrical rays for
each of these projections were described. The number of parallel geometrical
Image Processing 209

rays used in the projections is defined as follows:

n(1, s)=(1 + s)(N − 1) + 1 = N + s(N − 1), s = 0, 1, ..., N/2,

n(1, N − s) =N 2 − s(N − 1), s = N/2 + 1, N/2 + 2, ..., N − 1,
n(N − p, 1)=N 2 − p(N − 1), p = N/2 + 2, N/2 + 4, ..., N − 2,
n(p, 1)=n(1, p) = N + p(N − 1), p = 0, 2, 4, ..., N/2.
(3.108)
Values of these numbers n = n(p, s), for all 12 generators (p, s), are given
in Table 3.3. The total number of geometrical rays is 190. The subsets U (p, s)
and their intersections (which are underlined) are also given in the table. It
shows that the total number of intersections between these subsets equals
96 − 64 = 32.

TABLE 3.3
Number of geometrical rays and subsets U (p, s) for the 8 × 8 case
(p, s) n triplet-numbers in U (p, s)
(1, 0) 8 (1, 0, 0), (1, 0, 1), (1, 0, 2), (1, 0, 3), (2, 0, 0), (2, 0, 2), (4, 0, 0), (0, 0, 0) 8
(1, 1) 15 (1, 1, 0), (1, 1, 1), (1, 1, 2), (1, 1, 3), (2, 2, 0), (2, 2, 2), (4, 4, 0), (0, 0, 0) 7
(1, 2) 22 (1, 2, 0), (1, 2, 1), (1, 2, 2), (1, 2, 3), (2, 4, 0), (2, 4, 2), (4, 0, 0), (0, 0, 0) 6
(1, 3) 29 (1, 3, 0), (1, 3, 1), (1, 3, 2), (1, 3, 3), (2, 6, 0), (2, 6, 2), (4, 4, 0), (0, 0, 0) 6
(1, 4) 36 (1, 4, 0), (1, 4, 1), (1, 4, 2), (1, 4, 3), (2, 0, 0), (2, 0, 2), (4, 0, 0), (0, 0, 0) 4
(1, 5) 29 (1, 5, 0), (1, 5, 1), (1, 5, 2), (1, 5, 3), (2, 2, 0), (2, 2, 2), (4, 4, 0), (0, 0, 0) 4
(1, 6) 22 (1, 6, 0), (1, 6, 1), (1, 6, 2), (1, 6, 3), (2, 4, 0), (2, 4, 2), (4, 0, 0), (0, 0, 0) 4
(1, 7) 15 (1, 7, 0), (1, 7, 1), (1, 7, 2), (1, 7, 3), (2, 6, 0), (2, 6, 2), (4, 4, 0), (0, 0, 0) 4
(0, 1) 8 (0, 1, 0), (0, 1, 1), (0, 1, 2), (0, 1, 3), (0, 2, 0), (0, 2, 2), (0, 4, 0), (0, 0, 0) 7
(2, 1) 22 (2, 1, 0), (2, 1, 1), (2, 1, 2), (2, 1, 3), (4, 2, 0), (4, 2, 2), (0, 4, 0), (0, 0, 0) 6
(4, 1) 36 (4, 1, 0), (4, 1, 1), (4, 1, 2), (4, 1, 3), (0, 2, 0), (0, 2, 2), (0, 4, 0), (0, 0, 0) 4
(6, 1) 22 (6, 1, 0), (6, 1, 1), (6, 1, 2), (6, 1, 3), (4, 2, 0), (4, 2, 2), (0, 4, 0), (0, 0, 0) 4
190 64

Each triplet-number (p, s, t) of masks of the paired functions in this table

has the form 2k (p1 , s1 , t1 ), where 2k =g.c.d.(p, s), and t1 = 0 : N/2k+1 −
1, and its multiplicity (i.e., the number of subsets U covering this triplet-
number) equals 2k . Therefore, there are two ways to use this property in
image reconstruction.
1. One can use the complete 1-D paired transforms over all vectors a,
whose components are defined as
X X
an = vn±Nm = vp,s,n±Nm , n = 0 : N − 1.
m m

0
Then, the components fp,s,t of the 2-D paired transform, which have been
calculated by 1-D paired transform P(p, s) = [χ0N ]a, are normalized by the
factor of 2k = g.c.d.(p, s), when (p, s) 6= (0, 0), and N, when (p, s) = (0, 0).
2. One can use the incomplete complete 1-D paired transforms over vec-
tors a, to avoid the repetition of calculation of components of the 2-D paired
210 CHAPTER 3: Image Reconstruction

transform of the image. The last column in the table shows the number of
outputs to be calculated when performing the incomplete 1-D paired trans-
forms. For instance, the complete 1-D paired transform can be used for the
first (1, 0)-projection, but only six outputs for the transform when calculating
the transform P(1, 2) = [χ08 ]a, for the (1, 2)-projection, and half of calculations
when calculating the transform P(1, 4) = [χ08 ]a, for the (1, 4)-projection, and
so on. The use of incomplete 1-D paired transforms reduces the time necessary
to calculate the reconstructed image, when compared with the first method.

3.6.1 Equations of rays

The systems of equations for both arithmetical and geometrical rays (or sim-
ply, A-rays and G-rays) are given in Table 3.4, for the 8 × 8 case.

TABLE 3.4
Equations of G-rays on the grid 8 × 8. (The coordinate system is original.)
p+s
(p, s) A-rays t ± t0 G-rays l(t ∓ t0 ) t0 = 2 −1 t
(1, 0) l1,0 (t) t x = t/8 + 1/16 0 0:7
(1, 1) l1,1 (t) t x + y = t/8 + 1/8 0 0 : 14
(1, 2) l1,2 (t) t − 1/2 x + 2y = t/8 + 1/8 1/2 0 : 21
l1,2 (t) t + 1/2 x + 2y = t/8 + 1/4 1/2 21 : −1 : 0
(1, 3) l1,3 (t) t−1 x + 3y = t/8 + 1/8 1 0 : 28
l1,3 (t) t+1 x + 3y = t/8 + 3/8 1 28 : −1 : 0
(1, 4) l1,4 (t) t − 3/2 x + 4y = t/8 + 1/8 3/2 0 : 35
l1,4 (t) t + 3/2 x + 4y = t/8 + 1/2 3/2 35 : −1 : 0
p+N −s
(p, s) A-rays t ± t0 G-rays l(t ∓ t0 ) t0 = 2 −1 t
(1, 5) l1,−3 (t) t−1 x − 3y = t/8 − 1/4 1 7 : −1 : −21
l1,−3 (t) t+1 x − 3y = t/8 1 −21 : 1 : 7
(1, 6) l1,−2 (t) t − 1/2 x − 2y = t/8 − 1/8 1/2 7 : −1 : −14
l1,−2 (t) t + 1/2 x − 2y = t/8 1/2 −14 : 1 : 7
(1, 7) l1,−1 (t) t x − y = t/8 0 7 : −1 : −7
p+s
(p, s) A-rays t ± t0 G-rays l(t ∓ t0 ) t0 = 2 −1 t
(0, 1) l0,1 (t) t y = t/8 + 1/16 0 0:7
(2, 1) l2,1 (t) t − 1/2 2x + y = t/8 + 1/8 1/2 0 : 21
l2,1 (t) t + 1/2 2x + y = t/8 + 1/4 1/2 21 : 0
(4, 1) l4,1 (t) t − 3/2 4x + y = t/8 + 1/8 3/2 0 : 35
l4,1 (t) t + 3/2 4x + y = t/8 + 1/2 3/2 35 : 0
N −p+s
(p, s) A-rays t ± t0 G-rays l(t ∓ t0 ) t0 = 2
−1 t
(6, 1) l−2,1 (t) t − 1/2 −2x + y = t/8 − 1/8 1/2 7 : −1 : −14
l−2,1 (t) t + 1/2 −2x + y = t/8 1/2 −14 : 1 : 7

It should be noted that in this case, as well as in the general N = 2r case

where r ≥ 2, the system of equations for the G-rays is defined in the following
way. We first consider the A-rays l(t) = lp,s (t) which are described by the
equation
t p+s
Lp,s,t (x, y) = xp + ys = + .
N 2N
Image Processing 211

The set of all generators of JN,N is divided into four subsets, and we consider
each of them separately. The first and second subsets contain all generators
(1, s), where s = 0 : N/2 and s = N/2 + 1 : N − 1, respectively. The third and
fourth subsets contain all generators (p, 1), where p = 2p1 and p1 = 0 : N/4
and p1 = N/4 +1 : N/2 −1, respectively. For each of these subsets we consider
case A which is referred to as the case when the arithmetical rays are shifted
to the right, to define the corresponding geometrical rays. Case B is referred
to as the case when the arithmetical rays are shifted to the left, to define the
corresponding geometrical rays.
Subsets I and III: p = 1 and s ≤ N/2, and s = 1 and p ≤ N/2.
G-rays l̃(t) are defined by the shift
p+s
t → t ∓ t0 , t0 = − 1,
2
of A-rays, i.e.
t ∓ t0 p+s
L̃p,s,t (x, y) = Lp,s,t∓t0 (x, y) = + .
N 2N
Case A:
t − t0 p+s
L̃p,s,t (x, y) =Lp,s,t−t0 (x, y) = +
 N 2N
t p+s 1 p+s t 1
= − − + = + .
N 2N N 2N N N

Case B:
t + t0 p+s
L̃p,s,t (x, y)= Lp,s,t+t0 (x, y) = +
 N 2N
t p+s 1 p+s t p+s−1
= + − + = + .
N 2N N 2N N N

Thus, if p = 1,
t s
L̃1,s,t (x, y) = L1,s,t+t0 (x, y) = + ,
N N
and if s = 1,
t p
L̃p,1,t (x, y) = Lp,1,t+t0 (x, y) =
+ .
N N
The equation for case A is simple, when compared with case B.
Subset II: p = 1 and s > N/2.
A-rays are considered to be l(t) = l1,s−N (t), which are defined by

t 1 + (s − N )
L1,s−N,t (x, y) = x + y(s − N ) = + .
N 2N
212 CHAPTER 3: Image Reconstruction

The G-rays ˜
l(t) are defined by the shift of A-rays:

1 + (N − s) N −s−1
t → t ∓ t0 , t0 = −1= ,
2 2
t ∓ t0 1 + (s − N )
L̃1,s−N,t (x, y) = L1,s−N,t∓t0 (x, y) = + .
N 2N
Case A:
t − t0 1+s−N
L̃1,s−N,t (x, y)= L1,s−N,t−t0 (x, y) = +
  N 2N  
t N −s−1 s−N +1 t s+1
= − + = − 1− .
N 2N 2N N N

Case B:
t + t0 1+s−N
L̃1,s−N,t (x, y) =L1,s−N,t+t0 (x, y) = +
  N 2N
t N −s−1 s−N +1 t
= + + = .
N 2N 2N N

Subset IV: p > N/2 and s = 1.

A-rays are considered to be l(t) = lp−N,1 (t), which are defined by

t (p − N ) + 1
Lp−N,1,t (x, y) = x(p − N ) + y = + .
N 2N
The G-rays ˜
l(t) are defined by the shift

(N − p) + 1 N −p−1
t → t ∓ t0 , t0 = −1=
2 2
of A-rays, i.e.
t ∓ t0 (p − N ) + 1
L̃p−N,1,t (x, y) = Lp−N,1,t∓t0 (x, y) = + .
N 2N
Case A:
t − t0 p−N +1
L̃p−N,1,t (x, y) =Lp−N,1,t−t0 (x, y) = +
  N 2N  
t N −p−1 p−N +1 t p+1
= − + = − 1− .
N 2N 2N N N
Case B:
t + t0 p−N +1
L̃p−N,1,t (x, y)=Lp−N,1,t+t0 (x, y) = +
  N 2N
t N −p−1 p−N +1 t
= + + = .
N 2N 2N N
Image Processing 213

Thus, for subsets II and IV, equations for case B have a simple form and can
therefore be selected for image reconstruction. For subsets I and III, equations
for case A are simple. In both cases A and B, geometrical rays are defined
from the arithmetical rays by shifting ˜l(t) = l(t − t0 ) and ˜l(t) = l(t + t0 ),
respectively. Here the shift t0 is calculated by
 p+s

 − 1, when 0 < p, s ≤ N/2,


 2
p + (N − s)
t0 = − 1, when p = 1, s > N/2, (3.109)

 2
 (N − p) + s − 1, when p > N/2, s = 1.


2
It should be noted that for each of the subsets of generators (p, s), the use
of only case A, or case B leads to similar systems of equations that describe
the procedure of calculating the sums v(t) from integrals w(t). Indeed, these
equations are described respectively by the upper or lower triangular Toeplitz
matrices. One can use only, for instance, the upper triangular Toeplitz ma-
trices, if only case A will be used for all subsets of generators. It is also not
difficult to notice that the solution of the system described by the low upper
triangular Toeplitz matrix, b = A−1 w, is reduced to the solution of the sys-
tem with the upper triangular Toeplitz matrix, b̃ = Ã−1 w̃, after permutation
t → M − t of rows of the vectors b, w, and the matrix A−1 . Here M × M is
the size of the Toeplitz matrix.

3.6.2 Equations for line-integrals

For each (p, s)-projection, except the horizontal, vertical, and diagonal pro-
jections, the set of arithmetical rays is used for image reconstruction, after
shifting the rays either to the left, t → t − t0 , or to the right t → t + t0 ,
where the positive parameter t0 = t0 (p, s) is calculated by (3.109). For the
described N = 8 example, the systems of equations which describe the so-
lutions b = A−1 w of the convolution equations w = KAv, for all (p, s)-
projections, are given together in Table 3.5 for case A, i.e., when all shifts are
by t → t − t0 . Constants K are defined by the lengths√ √of intersection
√ of the
ray with the image element. For instance, K = 8 2, 8 5/2, and 8 10/3 for
the (1, 1), (1, 2), and (1, 3)-projections, respectively.
Another set of systems of equations, which describe the solutions of the
convolution equations w = KAv, is given together in Table 3.6, for case B,
i.e., when all shifting of arithmetical rays are performed by t → t + t0 .
214 CHAPTER 3: Image Reconstruction

TABLE 3.5
Model I (case A): Convolution equations for the 8 × 8 case
G-rays w = KAv, b = A−1 w v = Bw (v = Bb)
l1,0 (t) w(t) = 8v(t)
√ t = 0 : 7, (K = 8) √ v(t) = w(t)/8 √
l1,1 (t) w(t) = 8√2v(t) t = 0 : 14, (K = 8 2) v(t) = w(t)/(8
√ 2)
l1,2 (t − 0.5) w(t) = 4 5[v(t − 1) + v(t)] v(t) = bt /(4 5)
√ bt = w(t) − bt−1
b0 = w(0), t = 0 : 21 √
l1,3 (t − 1) w(t) = 8 10/3[v(t) + v(t − 1) + v(t − 2)] v(t) = 3bt /(8 10)
√ bt = w(t) − (bt−1 + bt−2 )
b0 = w(0), t = 0 : 28 √
l1,4 (t − 1.5) w(t) = 2 17[v(t) + v(t − 1) + v(t − 2) + v(t − 3)] v(t) = bt /(2 17)
b0 = w(0), bt = w(t) − (bt−1 + bt−2 + bt−3 ) t = 0 : 35
√ √
l1,−3 (t − 1) w(t) = 8 10/3[v(t) + v(t − 1) + v(t − 2)] v(t) = 3bt /(8 10)
b−21 = w(−21),
√ bt = w(t) − (bt−1 + bt−2 ) t = −21 : 7 √
l1,−2 (t − 0.5) w(t) = 4 5[v(t) + v(t − 1))] v(t) = bt /(4 5)
b−14 = w(−14),
√ bt = w(t) − bt−1 t = −14 : 7 √
l1,−1 (t) w(t) = 8 2v(t) v(t) = w(t)/(8 2)
bt = w(t) t = −7 : 7
l0,1 (t) w(t) = 8v(t),
√ t=0:7 v(t) = w(t)/8√
l2,1 (t − 0.5) w(t) = 4 5[v(t) + v(t − 1)] v(t) = bt /(4 5)
√ bt = w(t) − bt−1
b0 = w(0), t = 0 : 21 √
l4,1 (t − 1.5) w(t) = 2 17[v(t) + v(t − 1) + v(t − 2) + v(t − 3)] v(t) = bt /(2 17)
b0 = w(0), bt = w(t) − (bt−1 + bt−2 + bt−3 ) t = 0 : 35
√ √
l−2,1 (t − 0.5) w(t) = 4 5[v(t) + v(t − 1))] v(t) = bt /(4 5)
b−14 = w(−14), bt = w(t) − bt−1 t = −14 : 7

3.7 Equations in the coordinate system (X, 1 − Y )

It is common to consider the coordinate system as shown in Figure 3.2. In
the first realization of the proposed method of image reconstruction in the
MATLAB R
-based program, this system was used for calculating all rays,
arithmetical and geometrical. Thus, the realization was implemented in the
system of coordinates (x, 1 − y) instead of the original system (x, y).
In the new coordinate system, A-rays l(t) = lp,s (t) within the unit square
[0, 1] × [0, 1] can be written as
t p+s
Lp,s,t (x, 1 − y) = xp + (1 − y)s = + .
N 2N
Therefore, we obtain another set of equations, as shown in Table 3.7, for the
N = 8 case. We now consider separately each subset of generators and write
the corresponding equations.
Subsets I and III: p = 1 and s ≤ N/2, and s = 1 and p ≤ N/2.
The G-rays ˜l(t) are defined by the shift
p+s
t → t ∓ t0 , t0 = − 1,
2
Image Processing 215

TABLE 3.6
Model I (case B): Convolution equations for the 8 × 8 case
G-rays w = KAv, b = A−1 w v = Bw (v = Bb)
l1,0 (t) w(t) = 8v(t),
√ t = 0 : 7, (K = 8) √ v(t) = w(t)/8 √
l1,1 (t) w(t) = 8√2v(t), t = 0 : 14 (K = 8 2) v(t) = w(t)/(8
√ 2)
l1,2 (t + 0.5) w(t) = 4 5[v(t) + v(t + 1)] v(t) = bt /(4 5)
√ bt = w(t) − bt+1
b21 = w(21), t = 21 : −1 : 0√
l1,3 (t + 1) w(t) = 8 10/3[v(t) + v(t + 1) + v(t + 2)] v(t) = bt /(8/3 10)
√ bt = w(t) − (bt+1 + bt+2 )
b28 = w(28), t = 28 : −1 :√0
l1,4 (t + 1.5) w(t) = 2 17[v(t) + v(t + 1) + v(t + 2) + v(t + 3)] v(t) = bt /(2 17)
b35 = w(35), bt = w(t) − (bt+1 + bt+2 + bt+3 ) t = 35 : −1 : 0
√ √
l1,−3 (t + 1) w(t) = 8 10/3[v(t) + v(t + 1) + v(t + 2)] v(t) = bt /(8/3 10)
√ bt = w(t) − (bt+1 + bt+2 )
b7 = w(7), t = 7 : −1 : −21
√
l1,−2 (t + 0.5) w(t) = 4 5[v(t) + v(t + 1))] v(t) = bt /(4 5)
√ bt = w(t) − bt+1
b7 = w(7), t = 7 : −1 : −14√
l1,−1 (t) w(t) = 2/8v(t) v(t) = w(t)/(8 2)
bt = w(t) t = 7 : −1 : −7
l0,1 (t) w(t) = 8v(t),
√ t=0:7 v(t) = w(t)/8√
l2,1 (t + 0.5) w(t) = 4 5[v(t) + v(t + 1)] v(t) = bt /(4 5)
√ bt = w(t) − bt+1
b21 = w(21), t = 21 : −1 :√0
l4,1 (t + 1.5) w(t) = 2 17[v(t) + v(t + 1) + v(t + 2) + v(t + 3)] v(t) = bt /(2 17)
b35 = w(35), bt = w(t) − (bt+1 + bt+2 + bt+3 ) t = 35 : −1 : 0
√ √
l−2,1 (t + 0.5) w(t) = 4 5[v(t) + v(t + 1))] v(t) = bt /(4 5)
b7 = w(7), bt = w(t) − bt+1 t = 7 : −1 : −14

of A-rays,
t ∓ t0 p+s
L̃p,s,t (x, 1 − y) = Lp,s,t∓t0 (x, 1 − y) = + .
N 2N
Case A:
t − t0 p+s
L̃p,s,t (x, 1 − y) = Lp,s,t−t0 (x, 1 − y) = +
  N 2N
t p+s 1 p+s t 1
= − − + = + .
N 2N N 2N N N

Thus we have the equation:

t 1
xp + (1 − y)s = + , t = 0 : (p + s)(N − 1). (3.110)
N N
If p = 1,
t 1
x + (1 − y)s = + , t = 0 : (s + 1)(N − 1).
N N
If s = 1,
t 1
xp + (1 − y) = + , t = 0 : (p + 1)(N − 1).
N N
216 CHAPTER 3: Image Reconstruction
TABLE 3.7
Model II: Equations of G-rays on the grid 8 × 8
(p, s) A-rays G-rays l(t) ˜ t
(1, 0) l1,0 (t) x = t/8 + 1/16 0:7
(1, 1) l1,1 (t) x + (1 − y) = t/8 + 1/8 0 : 14
(1, 2) l1,2 (t) x + 2(1 − y) = t/8 + 1/8 0 : 21
l1,2 (t) x + 2(1 − y) = t/8 + 1/4 21 : −1 : 0
(1, 3) l1,3 (t) x + 3(1 − y) = t/8 + 1/8 0 : 28
l1,3 (t) x + 3(1 − y) = t/8 + 3/8 28 : −1 : 0
(1, 4) l1,4 (t) x + 4(1 − y) = t/8 + 1/8 0 : 35
l1,4 (t) x + 4(1 − y) = t/8 + 1/2 35 : −1 : 0
(p, s) A-rays G-rays ˜l(t) t
(1, 5) l1,−3 (t) x − 3(1 − y) = t/8 − 1/4 7 : −1 : −21
l1,−3 (t) x − 3(1 − y) = t/8 −21 : 1 : 7
(1, 6) l1,−2 (t) x − 2(1 − y) = t/8 − 1/8 7 : −1 : −14
l1,−2 (t) x − 2(1 − y) = t/8 −14 : 1 : 7
(1, 7) l1,−1 (t) x − (1 − y) = t/8 7 : −1 : −7
(p, s) A-rays G-rays ˜l(t) t
(0, 1) l0,1 (t) (1 − y) = t/8 + 1/16 0:7
(2, 1) l2,1 (t) 2x + (1 − y) = t/8 + 1/8 0 : 21
l2,1 (t) 2x + (1 − y) = t/8 + 1/4 21 : −1 : 0
(4, 1) l4,1 (t) 4x + (1 − y) = t/8 + 1/8 0 : 35
l4,1 (t) 4x + (1 − y) = t/8 + 1/2 35 : −1 : 0
(p, s) A-rays G-rays ˜l(t) t
(6, 1) l−2,1 (t) −2x + (1 − y) = t/8 − 1/8 7 : −1 : −14
l−2,1 (t) −2x + (1 − y) = t/8 −14 : 1 : 7

Case B:
t + t0 p+s
L̃p,s,t (x, 1 − y) =Lp,s,t+t0 (x, 1 − y) = +
  N 2N
t p+s 1 p+s t p+s−1
= + − + = + ,
N 2N N 2N N N
where t = (p + s)(N − 1) : −1 : 0.
If p = 1,
t s
L̃1,s,t (x, 1 − y) = L1,s,t+t0 (x, 1 − y) = + ,
N N
and, if s = 1,
t p
L̃p,1,t (x, 1 − y) = Lp,1,t+t0 (x, 1 − y) = + .
N N
Thus, if p = 1,
t s
x + (1 − y)s = + , t = (s + 1)(N − 1) : −1 : 0,
N N
Image Processing 217

if s = 1, the G-rays are described by

t p
xp + (1 − y) = + , t = (p + 1)(N − 1) : −1 : 0.
N N
Subset II: p = 1 and s > N/2.
A-rays are considered to be l(t) = l1,s−N (t), which are defined by
t 1 + (s − N )
L1,s−N,t (x, 1 − y) = x + (1 − y)(s − N ) = +
N 2N
The G-rays ˜
l(t) are defined by the shift of A-rays:
1 + (N − s) N −s−1
t → t ∓ t0 , t0 = −1= ,
2 2
t ∓ t0 1 + (s − N )
L̃1,s−N,t (x, 1 − y) = L1,s−N,t∓t0 (x, 1 − y) = + .
N 2N
Case A:
t − t0 1+s−N
L̃1,s−N,t (x, 1 − y) =L1,s−N,t−t0 (x, 1 − y) = +
  N 2N  
t N −s−1 s−N +1 t s+1
= − + = − 1− .
N 2N 2N N N
Thus
 
t s+1
x + (1 − y)(s − N ) = − 1− , t = (N − 1) : −1 : −(N − s)(N − 1).
N N
Case B:
t + t0 1+s−N
L̃1,s−N,t (x, 1 − y) =L1,s−N,t+t0 (x, 1 − y) = +
  N 2N
t N −s−1 s−N +1 t
= + + = .
N 2N 2N N
Thus
t
x + (1 − y)(s − N ) = , t = −(N − s)(N − 1) : 1 : (N − 1).
N
Subset IV: p > N/2 and s = 1.
A-rays are considered to be l(t) = lp−N,1 (t), which are defined by
t (p − N ) + 1
Lp−N,1,t (x, 1 − y) = x(p − N ) + (1 − y) = + .
N 2N
The G-rays ˜
l(t) are defined by the shift
(N − p) + 1 N −p−1
t → t ∓ t0 , t0 = −1=
2 2
218 CHAPTER 3: Image Reconstruction

of A-rays
t ∓ t0 (p − N ) + 1
L̃p−N,1,t (x, 1 − y) = Lp−N,1,t∓t0 (x, 1 − y) = + .
N 2N
Case A:
t − t0 p−N +1
L̃p−N,1,t (x, 1 − y) =Lp−N,1,t−t0 (x, 1 − y) = +
  N 2N  
t N −p−1 p−N +1 t p+1
= − + = − 1− .
N 2N 2N N N
Thus
 
t p+1
x(p − N ) + (1 − y) = − 1− , t = (N − 1) : −1 : −(N − p)(N − 1).
N N
Case B:
t + t0 p−N +1
L̃p−N,1,t (x, y)=Lp−N,1,t+t0 (x, y) = +
  N 2N
t N −p−1 p−N +1 t
= + + = .
N 2N 2N N
Thus
t
x(p − N ) + (1 − y) = , t = −(N − p)(N − 1) : 1 : (N − 1).
N
For each subset of generators, there are two cases, A or B, and one of
them can be selected to accomplish the method of image reconstruction. For
two scenarios, when only case A or B is considered for all four subsets of
generators of the set JN,N , the equations for G-rays are given in Tables 3.8
and 3.9, respectively. In the first column, the number I, II, III, or IV is
assigned to the subset containing generators (p, s).
Here we recall that for the generators (1, s), and (p, 1), such that p, s >
N/2, all equations for rays are written and the components of the 2-D paired
transform, or splitting-signals fTp,s 0 that are generated by these frequency-
points, are calculated by considering the change in direction of projections.
The (1, s)-projections are changed by the corresponding (1, s−N )-projections,
and the (p, 1)-projections by the (p − N, 1)-projections.
For the N = 8 case, Figure 3.85 shows two sets of angles of 12 projections
before and after changing the directions, Φ8,8 = {arctan(s/p); (p, s) ∈ J8,8 }
and Φ?8,8 = arctan(s/p); (p, s) ∈ J8,8 ?
. The sets of generators are defined as

J8,8 = {(p, 1); p = 0 : 7}, {(1, 2s); s = 0 : 3} ,
?

J8,8 = {(p, 1); p = 0 : 4}, {(p̄, 1); p̄ = −3 : −1}, {(1, 2s); s = 0 : 2}, (1, −2) .
Herewith, all these projections together with the remaining ones, are used
to calculate the 2-D discrete paired transform of the image, with genera-
tors (p, s) from the original set JN,N . That allows for using the existing
program for the inverse 2-D discrete paired transform (which is available at
http://www.fasttransforms.com).
Image Processing 219
TABLE 3.8
Model II (case A): Equations of G-rays for image N × N
(p, s) G-rays
t 1
I (1, 0) x− − =0 s=0
N 2N
t = 0 : (N − 1)
t 1
I (1, s) x − ys + s − − =0 s ≤ N/2
N N
t = 0 : (s + 1)(N − 1)
 
t s+1
II (1, s) x + y(N − s) + (s − N ) − + 1− =0 s > N/2
N N
t = (N − 1) : −1 : −(N − s)(N − 1)
t 1
III (0, 1) −y + 1 − − = 0, p=0
N 2N
t = 0 : (N − 1)
t 1
III (p, 1) xp − y + 1 − − =0 p ≤ N/2
N N
t = 0 : (p + 1)(N − 1)
 
t p+1
IV (p, 1) x(p − N ) − y + 1 − + 1− =0 p > N/2
N N
t = (N − 1) : −1 : −(N − p)(N − 1)

TABLE 3.9
Model II (case B): Equations of G-rays for image N × N
(p, s) G-rays
t 1
I (1, 0) x− − =0 s=0
N 2N
t = 0 : (N − 1)
t s
I (1, s) x − ys + s − − =0 s ≤ N/2
N N
t = (s + 1)(N − 1) : −1 : 0
t
II (1, s) x − y(s − N ) + (s − N ) − =0 s > N/2
N
t = −(N − s)(N − 1) : 1 : (N − 1)
t 1
III (0, 1) −y + 1 − − = 0, p=0
N 2N
t = 0 : (N − 1)
t p
III (p, 1) xp − y + 1 − − =0 p ≤ N/2
N N
t = (p + 1)(N − 1) : −1 : 0
t
IV (p, 1) x(p − N ) − y + 1 − =0 p > N/2
N
t = −(N − p)(N − 1) : 1 : (N − 1)

3.7.1 Convolution equations

For all generators (p, s) ∈ JN,N , the relations between the sums v(t) = vp,s (t)
of the discrete image fn,m and line-integrals w(t) = wp,s (t) of the original
220 CHAPTER 3: Image Reconstruction

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 0.5 1 0 0.5 1
(a) (b)

FIGURE 3.85
Central angles of (a) original and (b) new 12 projections.

unknown image f(x, y) are described in matrix form by the Toeplitz matrices,
which are similar to the described N = 8 and N = 4 examples. We consider
separately the convolution equations of line-integrals for cases A and B.
1. Equations in case A
Subset I: p = 1 and s ≤ N/2.
When s 6= 0, the convolution equation w = Av is described as
√
1 + s2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s + 1)],
s (3.111)
t = 0 : (1 + s)(N − 1),

where it is assumed that v(−1) = v(−2) = ... = v(−s + 1) = 0. The inverse

transform b = A−1 w is calculated by the following recurrent form:

b0 = w(0),
b1 = w(1) − b0 ,
b2 = w(2) − (b1 + b0 ),
b3 = w(3) − (b2 + b1 + b0 ),
(3.112)
... ... ... ... ... ...
bs−1 = w(s − 1) − (bs−2 + bs−3 + ... + b1 + b0 ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−s+2 + bt−s+1)),
t = s, s + 1, . . . , (1 + s)(N − 1).

The required sums v(t) of the discrete image are calculated by

s
v(t) = bt √ , t = 0 : (1 + s)(N − 1). (3.113)
N 1 + s2
In the s = 0 case, we have the following simple solution: v(t) = w(t)/N,
t = 0 : (N − 1).
Subset II: p = 1 and s > N/2.
Image Processing 221

The (1, s)-projection is described as the (1, s − N )-projection, i.e., s is

considered as s − N. Therefore, the following convolution equation w = Av
is used:
√
1 + s̄2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s̄ + 1)], (3.114)
s̄
t = (N − 1) : −1 : −s̄(N − 1),

where s̄ = N − s, and v(−k) = 0, when k > s̄(N − 1)).

We denote by M the negative number −s̄(N − 1). The inverse transform
b = A−1 w can be calculated by the following recurrent form:
bM = w(M ),
bM +1 = w(M + 1) − bM ,
bM +2 = w(M + 2) − (bM +1 + bM ),
bM +3 = w(M + 3) − (bM +2 + bM +1 + bM ),
(3.115)
... ... ... ... ... ... ... ...
bM +s̄−1 = w(M + s̄ − 1) − (bM +s̄−2 + bM +s̄−3 + ... + bM +1 + bM ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−s̄+2 + bt−s̄+1) ),
t = M + s̄, M + s̄ + 1, . . . , (N − 1).

The required sums v(t) of the discrete image are calculated by

s̄
v(t) = bt √ , t = (N − 1) : −1 : −s̄(N − 1). (3.116)
N 1 + s̄2
Subset III: p ≤ N/2 and s = 1.
This subset is described similarly to subset I. When p 6= 0, the convolution
equation w = Av is described as
p
1 + p2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − p + 1)],
p (3.117)
t = 0 : (p + 1)(N − 1).

The inverse transform b = A−1 w can be calculated in the recurrent form

b0 = w(0),
b1 = w(1) − b0 ,
b2 = w(2) − (b1 + b0 ),
b3 = w(3) − (b2 + b1 + b0 ),
(3.118)
... ... ... ... ... ...
bp−1 = w(p − 1) − (bp−2 + bp−3 + ... + b1 + b0 ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−p+2 + bt−p+1) ),
t = p, p + 1, . . . , (p + 1)(N − 1).

The required sums v(t) of the discrete image are calculated by

p
v(t) = bt p , t = 0 : M. (3.119)
N 1 + p2
222 CHAPTER 3: Image Reconstruction

In the p = 0 case, we have the following simple solution: v(t) = w(t)/N,

t = 0 : (N − 1).
Subset IV: p > N/2 and s = 1.
The projection (p, 1) is considered as the (p − N, 1)-projection, i.e., p is
considered as p − N. The convolution equation w = Av is described as
p
1 + p̄2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − p̄ + 1)],
p̄ (3.120)
t = (N − 1) : −1 : −p̄(N − 1),
where p̄ = N − p, and v(k) = 0, when k < −p̄(N − 1)). The inverse transform
b = A−1 w can be calculated by the recurrent form
bM = w(M ),
bM +1 = w(M + 1) − bM ,
bM +2 = w(M + 2) − (bM +1 + bM ),
bM +3 = w(M + 3) − (bM +2 + bM +1 + bM ),
(3.121)
... ... ... ... ... ... ... ...
bM +p̄−1 = w(M + p̄ − 1) − (bM +p̄−2 + bM +p̄−3 + ... + bM +1 + bM ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−p̄+2 + bt−p̄+1) ),
t = M + p̄, M + p̄ + 1, . . . , (N − 1),
where M = −p̄(N − 1). The required sums v(t) of the discrete image are
calculated by
p̄
v(t) = bt p , t = (N − 1) : −1 : −p̄(N − 1). (3.122)
N 1 + p̄2
We now write all convolution equations for case B.
2. Equations in case B
Subset I: p = 1 and s ≤ N/2.
When s 6= 0, the convolution equation w = Av is described as
√
1 + s2
w(t) = N [v(t) + v(t + 1) + v(t + 2) + ... + v(t + s − 1)],
s (3.123)
t = (1 + s)(N − 1) : −1 : 0,
where it is assumed that v(k) = 0, for k > (1+s)(N −1). The inverse transform
b = A−1 w is calculated by the recurrent form
bM = w(M ),
bM −1 = w(M − 1) − bM ,
bM −2 = w(M − 2) − (bM −1 + bM ),
bM −3 = w(M − 3) − (bM −2 + bM −1 + bM ),
(3.124)
... ... ... ... ... ...
bM −s+1 = w(M − s + 1) − (bM −s+2 + bM −s+3 + ... + bM −1 + bM ),
bt = w(t) − (bt+1 + bt+2 + ... + bt+s−2 + bt+s−1)),
t = M − s, M − s − 1, . . . , 0,
Image Processing 223

where (M = (1 + s)(N − 1). The required sums v(t) of the discrete image are
calculated by
s
v(t) = bt √ , t = 0 : (1 + s)(N − 1). (3.125)
N 1 + s2
In the s = 0 case, we have the following: v(t) = w(t)/N, t = 0 : (N − 1).
Subset II: p = 1 and s > N/2.
The projection (1, s) is described as the (1, s−N )-projection. The following
convolution equation w = Av is considered:
√
1 + s̄2
w(t) = N [v(t) + v(t + 1) + v(t + 2) + ... + v(t + s̄ − 1)],
s̄ (3.126)
t = (N − 1) : −1 : −s̄(N − 1),
where s̄ = N − s, and v(N ) = v(N + 1) = ... = v(N + s̄ − 2) = 0. The inverse
transform b = A−1 w is calculated by the recurrent form
bN−1 = w(N − 1),
bN−2 = w(N − 2) − bN−1 ,
bN−3 = w(N − 3) − (bN−2 + bN−1 ),
bN−4 = w(N − 4) − (bN−3 + bN−2 + bN−1 ),
... ... ... ... ... ... ... ... (3.127)
bN−s̄ = w(N − s̄) − (bN−s̄+1 + bN−s̄+2 + ... + bN−2 + bN−1 ),
bt = w(t) − (bt+1 + bt+2 + ... + bt+s̄+2 + bt+s̄−1)),
t = N − s̄ − 1, N − s̄ − 2, . . . , −s̄(N − 1).
where M = −s̄(N − 1). The required sums v(t) of the discrete image are
calculated by
s̄
v(t) = bt √ , t = (N − 1) : −1 : −s̄(N − 1). (3.128)
N 1 + s̄2
Subset III: p ≤ N/2 and s = 1.
When p 6= 0, the convolution equation w = Av is described as
p
1 + p2
w(t) = N [v(t) + v(t + 1) + v(t + 2) + ... + v(t + p − 1)],
p (3.129)
t = (p + 1)(N − 1) : −1 : 0,
where it is assumed that v(k) = 0, for k > (p+1)(N −1). The inverse transform
b = A−1 w is calculated by the following recurrent form: (M = (p+1)(N −1)):
bM = w(M ),
bM −1 = w(M − 1) − bM ,
bM −2 = w(M − 2) − (bM −1 + bM ),
bM −3 = w(M − 3) − (bM −2 + bM −1 + bM ),
(3.130)
... ... ... ... ... ...
bM −p+1 = w(M − p + 1) − (bM −p+2 + bM −p+3 + ... + bM −1 + bM ),
bt = w(t) − (bt+1 + bt+2 + ... + bt+p−2 + bt+p−1) ),
t = M − p, M − p − 1, . . . , 0.
224 CHAPTER 3: Image Reconstruction

The required sums v(t) of the discrete image are calculated by

p
v(t) = bt p , t = 0 : (p + 1)(N − 1). (3.131)
N 1 + p2

In the p = 0 case, v(t) = w(t)/N, t = 0 : (N − 1).

Subset IV: p > N/2 and s = 1.
The projection (p, 1) is described as the (p−N, 1)-projection. The following
convolution equation w = Av is considered:
p
1 + p̄2
w(t) = N [v(t) + v(t + 1) + v(t + 2) + ... + v(t + p̄ − 1)],
p̄ (3.132)
t = (N − 1) : −1 : −p̄(N − 1),

where p̄ = N − p, and v(N ) = v(N + 1) = ... = v(N + p̄ − 1) = 0. The inverse

transform b = A−1 w is calculated by the recurrent form

bN−1 = w(N − 1),

bN−2 = w(N − 2) − bN−1 ,
bN−3 = w(N − 3) − (bN−2 + bN−1 ),
bN−4 = w(N − 4) − (bN−3 + bN−2 + bN−1 ),
... ... ... ... ... ... ... ... (3.133)
bN−p̄ = w(N − p̄) − (bN−p̄+1 + bN−p̄+2 + ... + bN−2 + bN−1 ),
bt = w(t) − (bt+1 + bt+2 + ... + bt+p̄+2 + bt+p̄−1)),
t = N − p̄ − 1, N − p̄ − 2, . . . , −p̄(N − 1).

where M = −p̄(N − 1). The required sums v(t) of the discrete image are
calculated by
p̄
v(t) = bt p , t = (N − 1) : −1 : −p̄(N − 1). (3.134)
N 1 + p̄2

Problems
Problem 3.1 Different sets of generators (p, s) can be used to cover the Cartesian
lattice 8 × 8 by sets Tp,s , but each such set will have a cardinality greater than
or equal to 12. For each (p, s)-projection, there are different ways to select the set
of G-rays which will be used to calculate the line-integrals wp,s (t). The number of
these integrals depends on (p, s). How would you choose the sets of generators and
G-rays, to obtain the minimum total number of G-rays for reconstructing the image
composed of rectangles on the lattice?

Problem 3.2 Consider the following image on the square [0, 1] × [0, 1] to be recon-
Image Processing 225

structed as the discrete image 8 × 8 :

0 0 4 4 0 0 2 2
0 0 4 4 0 0 2 2
3 3 0 0 1 1 0 0
3 3 0 0 1 1 0 0
[f (x, y); 0 ≤ x, y ≤ 1 ] = .
0 0 2 2 0 0 4 4
0 0 2 2 0 0 4 4
1 1 0 0 3 3 0 0
1 1 0 0 3 3 0 0

Each square is referred to as an image element of size 1/8 × 1/8. Sketch the graphs
of line-integrals wp,s (t) of the (1, 2)- and (1, 5)-projections, i.e., when (p, s) = (1, 2)
and (1, 5). Use these integrals to calculate the line-sums vp,s (t) and splitting-signals
0
{fp,s,t ; t = 0 : 3}. Verify if these splitting-signals can be used to calculate the 2-D
0 0
DFT of the image in the sets of frequency-points T1,2 and T1,5 .

Problem 3.3 Suppose that you are given an image in the square [0, 1]×[0, 1], which
is composed of image elements of size 1/16 × 1/16 each with a constant intensity.
In this case, we say that the image is on the Cartesian lattice 16 × 16. Sketch line-
integrals along G-rays of the (1, 4)-projection and prove that these integrals can be
0
used to calculate the paired splitting-signal {f1,4,t ; t = 0 : 7} of the reconstructed
image 16×16. This means you can calculate the 2-D DFT of this image at frequency-
0
points of the set T1,4 .

Problem 3.4 Suppose that you are given an image on the square [0, 1] × [0, 1] and
the Cartesian lattice 10 × 10.
A. Determine the number of projections required to reconstruct the image.
B. Show that for each (p, s)-projection in A, a set of G-rays exists such that
the line-integrals along these rays can be used to calculate the splitting-signals
{fp,s,t ; t = 0 : 9}.
C. Develop the algorithm for full reconstruction of the given image from its pro-
jections and show your results. As an example, you can use the image in Problem 3.2
after extending the image to the size 10 × 10, by adding a zero column or row to
each side of the image.

Problem 3.5 ∗ In the described model 8 × 8, the G-rays are considered as lines.
Assume that the projections were obtained from the rays of width 6= 0. It means the
integrals wp,s (t) represent the integrals of the image along the wide rays. Assume
that the width of G-rays is a small number, for instance d < 1/64. Is it possible to
generalize the method of reconstruction described in this chapter to the case with
wide rays? Propose the method of transferring the geometry of wide G-rays to line
A-rays, to solve the problem of image reconstruction from its ray-projections.

Problem 3.6 ∗ For N = 2r , where r > 1, the basic functions of the 2-D discrete
paired transformation are described by the masks, or matrices with coefficients −1, 1,
and 0. It was shown in the N = 8 example, that all coefficients 1 and −1 are located
on parallel lines. Propose the concept of 2-D paired functions with the width 6= 0.
In other words, propose such 2-D binary functions on the square [0, 1] × [0, 1], whose
226 CHAPTER 3: Image Reconstruction

masks contain coefficients 1 and −1, which are located on different sets of parallel
wide rays.
A. Give examples of such wide-ray paired functions.
B. Discuss the property of orthogonality for the functions in A.
For instance, we studied the following functions. Given (p, s) 6= (0, 0), we consider
the set of parallel rays px+sy = t/N at the angle ϑ = − tan−1 (p/s) to the horizontal
line, where t runs a set of integers. We consider rays with width ∆ = ∆(p, s) and
the central line px + sy = t/N. Let a be the maximum width of the rays, which
is assumed to be the width of the disjoint next rays for the vertical or horizontal
projections, when (p, s) = (0, 1) and (1, 0), respectively. For wide rays with central
lines y = t/N or x = t/N, the maximum width is a = 1/N. This is the distance
between two neighbor points of the lattice N × N places in the square [0, 1] × [0, 1].
Definition 1: On the square [0, 1] × [0, 1] with the lattice XN,N on it, the wide
ray is defined by
1, if − a2 < px + sy − Nt ≤ a2 ,

rp,s,t (x, y; a) = (3.135)
0, otherwise,
where x, y ∈ [0, 1].
pThe width ∆(p, s) of the ray rp,s,t is calculated by ∆(p, s) = a/s cos(ϑ) =
a/ p2 + s2 , (s 6= 0). In the (1, s) = (1, 0) case, we consider ∆ = a.
Definition 2: Given triplet (p, s, t), the paired function of wide rays, or the wide
ray paired function on the square [0, 1] × [0, 1] is defined as
∞
X
χ0p,s,t (x, y; a) = (−1)k rp,s,t+k2r−n−1 (x, y; a), x, y ∈ [0, 1], (3.136)
k=−∞

where t ∈ [0, N/2 − 1]. For (p, s) = (0, 0), the paired function χ00,0,0 (x, y; a) ≡ 1.
As an example for the N = 8 case, Figure 3.86 shows two gray-scale images of
the paired wide ray functions for (p, s, t) = (1, 2, 1) and (4, 1, 3) in parts a and b,
respectively. White, black, and gray colors on the image correspond to the values of
1, −1, and 0, respectively.

(p,s,t)=(1,2,1), [N=8] (p,s,t)=(4,1,3), [N=8]

0 0

0.25 0.25

0.5 0.5

0.75 0.75

1 1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
(a) (b)

FIGURE 3.86
Images of functions (a) χ01,2,1 (x, y; a) and (b) χ04,1,3 (x, y; a), when a = 1/8.
4
Main Program of Image Reconstruction

The method of transformation of the geometry of projection data, i.e., line-

integrals wp,s (t) along the parallel rays lp,s (t), into the sums vp,s (t) of the
discrete image to be reconstructed can be used for writing an uncomplicated
code for exact reconstructing the images on the Cartesian lattice N × N,
with N being a power of two. Different schemes for changing the geometry of
the arithmetic rays with following solutions of the convolution equations were
considered in Chapter 3, and the examples for the reconstruction on lattices
4×4 and 8×8 were described in detail. In this chapter, we give the description
of one of our codes, which was first written in MATLAB R
and then in C+
and used for reconstruction of an image up to the size 2048 × 2048. The image
f(x, y) is considered on the unit square with the Cartesian lattice, i.e. the
image is the composition of image elements (squares) of size 1/N × 1/N each.
Each image element has a constant intensity at all its points.

4.1 The main diagram of the reconstruction

The image is generated from the random rectangles in our program. The
calculation of the reconstructed image from its projections by the proposed
method of image reconstruction can be separated into the following parts,
which are shown in the block-diagram in Figure 4.1.
1. The image N × N is composed by random rectangles on the square
[0, 1] × [0, 1].
2. For the given frequency-point (p, s), the (p, s)-projection is calculated,
i.e., all line-integrals w(t) along the set of geometrical rays l̃p,s (t).
3. The value of (p, s) is taken from the set J = JN,N of 3N/2 generators,
which is calculated for the given N.
4. The geometry of the G-rays of the (p, s)-projection is transformed to A-
rays, i.e., the set of sums v(t) of the discrete image fn,m is calculated from the
line-integrals w(t), by solving the system of linear equations with a Toeplitz
matrix (M × M ), where M = (p̄ + s̄ − 1)(N − 1) + 1. Here, p̄ = N − p and
s̄ = N − s when p and s are greater than N/2, and p̄ = p and s̄ = s when
p, s ≤ N/2. The Toeplitz matrix is triangular, therefore the transformation of
geometry of G-rays to A-rays is fast.

227
228 CHAPTER 4: Main Program

image f (x, y)
1
Image composition: Random rectangles

?
Calculate (p, s)-projection 2 3

Generator set
?
4 J = {(p, s)}
Transform G-rays into A-rays, w → v

?
5
Composition of the vector a

?
Permutation of the vector a 6

? ?
7 8
N -point paired transform of a stop-points

?
Composition of 2-D paired transform 9

?
Inverse 2-D paired transform 10

?
reconstructed image fn,m

FIGURE 4.1
Block-diagram of image reconstruction (case N = 2r , r > 1), when the image
is composed of rectangles on the grid N × N placed in the square [0, 1] × [0, 1].

5. The N -dimensional vector a is calculated from the sums v(t).

6. The correction in indexing of this vector is performed, when the gene-
rator (p, s) has a coordinate which greater than N/2.
7. The 1-D N -point fast discrete paired transform (DPT) is calculated over
the vector a. For many projections, the incomplete paired transform can be
used to reduce the calculations.
8. Stop-points are calculated to accomplish the incomplete 1-D DPT, when
the case with incomplete 1-D DPT is selected in the program.
9. The 2-D DPT of the discrete image is filled by the 1-D paired transform.
Image Processing 229

10. After processing all 3N/2 projections, the inverse 2-D DPT is calcu-
lated to obtain the reconstructed discrete image fn,m .
Now we describe each part in this block-diagram in more detail.

4.2 Part 1: Image model

The image f(x, y) to be reconstructed is considered on the unit square [0, 1] ×
[0, 1] and in the form of N 2 image elements (IE) with intensities fn,m ,
N−1 N−1  
1 X X 1 1
f(x, y) = fn,m rect 2 N x − n − , N y − m − , (4.1)
(∆x)2 m=0 n=0 2 2

where the function rect2 (x, y) = rect(x)rect(y), and rect(t) = 1, if |t| < 1/2,
and 0, otherwise. The image f(x, y) is generated from a few rectangles rk (x, y)
defined by positions (xk , yk , ∆xk , ∆yk ), which are randomly placed in the
square on the grid N × N,
K  
1 X x − xk 1 y − yk 1
f(x, y) = Rk rect − , − . (4.2)
(∆x)2 ∆xk 2 ∆yk 2
k=1

Here, Rk is the intensity of the k-th rectangle, and integer K ≥ 1. It is

assumed that the sizes of the IEs are equal, ∆x = ∆y = 1/N, for all k. As an
example, for the N = 64 case, Figure 4.2 shows the gray-scale image of size
64 × 64, which is composed of seven rectangles (on the left), and its discrete
representation (on the right), after the image reconstruction. The image is the
sum of these rectangles.
The coordinates of all rectangles and their intensities are given in Table 4.1.

TABLE 4.1
Data of rectangles
k xk yk ∆xk ∆yk Rk
1 0.4219 0.4531 0.5000 0.2500 5
2 0.4688 0.2969 0.2188 0.4688 11
3 0.2188 0.4219 0.2656 0.2813 12
4 0.1875 0.1250 0.2969 0.4375 8
5 0.4219 0.2031 0.1875 0.1875 3
6 0.0938 0.0781 0.1406 0.0938 3
7 0.2969 0.4531 0.4688 0.1094 9

All data of random rectangles are generated first in the integer form as
shown in Table 4.2, and then they are used for the discrete grid N × N by
xk = x0k /N, yk = yk0 /N, ∆xk = ∆x0k /N, ∆yk = ∆yk0 /N, k = 1 : 7.
230 CHAPTER 4: Main Program

image f(x,y) with 7 rectangles discrete image f(n,m)

1

10
0.8

20
0.6 1
3 30
72

0.4 40
4
5
50
0.2
6
60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
intensities: 5 11 12 8 3 3 9 grid 64x64

FIGURE 4.2 (See color insert)

Seven color rectangles and the discrete gray-scale image composed of these
rectangles.

TABLE 4.2
Integer data of rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 27 29 32 16 5
2 30 19 14 30 11
3 14 27 17 18 12
4 12 8 19 28 8
5 27 13 12 12 3
6 6 5 9 6 3
7 19 29 30 7 9

Below is the beginning of the script image reconstionA3A4B1A2.m, where

the random rectangles are generated and saved as data data allrect in integer
format [x,y,dx,dy,d intensity].
% ----------------------------------------------------------------
% call: image_reconstionA3A4B1A2.m
% Art-Meruzhan Grigoryan, January 04, 2011
. . .
fprintf(’\n Start the code \n’);
N=input(’ input the value of N: ’);
n=input(’ input the number of rectangles (1,2,...,16): ’);
many_rectangles=n;
data_allrect=zeros(many_rectangles,5);
for k=1:many_rectangles
kk=1;
while kk==1
data_1rect=round(rand(1,4)/2*N);
Image Processing 231

if (min(data_1rect)>1)
kk=0;
end
end
% data_1rect is in integer format [x,y,dx,dy].
% intensity (<=16) of the image is added:
data_1rect(5)=round(rand*16)+1;
data_allrect(k,:)=data_1rect;
end
draw_allfigures; % draw all rectangles on the square [0,1]x[0,1].
. . .

4.3 The coordinate system and rays

When writing the first version of our program of image reconstruction, the
discrete image fn,m and all masks of the paired functions χ0p,s,t (n, m) were
considered in the original coordinate system (n, m), as shown in Figure 4.2
in part b. The parallel arithmetical rays lp,s (t) along only the units +1 and
−1 in each such mask were described with respect to the coordinate system
(x, 1−y), instead of the original system with coordinates (x, y). In this system,
all rectangles and lines are drawn using MATLAB commands. This means the
top left corner of the square [0, 1] × [0, 1] has coordinates (0, 1), not (0, 0) as
in the original coordinate system (x, y), as shown in Figure 4.2 in part a. In
these two systems of coordinates, the arithmetical rays and geometrical rays
are described by np + ms = t on discrete grid XN,N and xp + (1 − y)s = t1
on the square [0, 1] × [0, 1], respectively.
In addition, all diaphanous equations for the A-rays and G-rays are consi-
dered in the form of ns + mp = t and xs + (1 − y)p = t, instead of np + ms = t
and xp + (1 − y)s = t, respectively. In other words, the parameter s is referred
to as p, and p as s. The generators (p, s) of the set JN,N composed of sub-
sets I, II, III, and IV are taken in the order III, IV, I, and II. We denote by
M1, M2, M3, and M4 these four cases with corresponding families of equa-
tions for the G-rays and convolutions. Equations for M1, M2, M3, and M4
are selected of types A, A, B, and A, respectively. In other words, different
types of equations for geometrical rays are considered. Therefore the program
of image reconstruction is called A3-A4-B1-A2 code.
It should be said that in the program all control points of the rays are
numbered from 0 to (p + s)(N − 1), for the (p, s)-projections, in order to avoid
negative indexing. Therefore, all equations which were given above for the
A-rays and G-rays have been changed, as described below.
232 CHAPTER 4: Main Program

Case M1 (A3): Generators are (1, s), 0 ≤ s ≤ N/2, and equations are

 xs − y + 1 − t − 1 = 0, when s 6= 0

N N
t 1
 −y + 1 −
 − = 0, when s = 0,
N 2N
where t = 0 : (s + 1)(N − 1). This is the case of A3 (see Table 3.8), when
changing the parameter p to s, and (p, 1) to (1, s).
Case M2 (A4): Generators are (1, s), s > N/2, and equations are
 
t 1
x(s − N ) − y − + (N − s) + 1 − = 0,
N N
where t = 0 : (N − s + 1)(N − 1). These equations correspond to case A4
 
t p+1
x(p − N ) + (1 − y) − + 1− = 0,
N N
t = (N − 1) : −1 : −(N − p)(N − 1),

after permutation t → t − (N − p)(N − 1) and the change p → s.

Case M3 (B1): Generators are (p, 1), 0 < p ≤ N/2, and equations are
t p
x − yp + p − − = 0.
N N
This is the case of B1 after changing the parameter p to s.
Case M4 (A2): Generators are (p, 1), p > N/2, and equations are
t 1
x + y(N − p) − − = 0, t = 0 : (N − p + 1)(N − 1).
N N
These equations correspond to equations in case A2,
 
t s+1
x + y(N − s) + (s − N ) − + 1− =0
N N
t = (N − 1) : −1 : −(N − s)(N − 1),

after permutation t → t − (N − s)(N − 1) and replacement of s by p.

All equations above for G-rays are given in Tables 4.3.

4.4 Part 2: Projection data

The set of projections is defined by the following set of generators (p, s):

JN,N = {{(1, s); s = 0 : (N − 1)}, {(2p, 1); p = 0 : (N/2 − 1)}} .

Image Processing 233
TABLE 4.3
Model II (case M): Equations of G-rays for image N × N
(p, s) G-rays
t 1
I (1, 0) −y + 1 − − =0 s=0
N 2N
t = 0 : (N − 1)
t 1
I (1, s) xs − y + 1 − − = 0, s ≤ N/2
N N
t = 0 : (s + 1)(N − 1)
t 1
II (1, s) x(s − N ) − y − + (N − s) + (1 − ) = 0 s > N/2
N N
t = 0 : (N − s + 1)(N − 1)
t 1
III (0, 1) x− − =0 p=0
N 2N
t = 0 : (N − 1)
t p
III (p, 1) x − yp + p − − =0 p ≤ N/2
N N
t = 0 : (p + 1)(N − 1)
t 1
IV (p, 1) x + y(N − p) − − =0 p > N/2
N N
t = 0 : (N − p + 1)(N − 1)

For each (p, s) of this set, the corresponding (p, s)-projection is calculated.
The measurements of the projections, i.e., the line-integrals wl (t) of the image
f(x, y) along the rays l(t) = lp,s (t),
Z
wl (t) = fd (x, y)dl, (4.3)
l(t)

are calculated as the sum of measurements wl;k (t) of each rectangle rk (x, y).
In other words,
K
X K
X Z
wl (t) = wl;k (t) = Rk rk (x, y)dl. (4.4)
k=1 k=1 l(t)

Each such integral wl;k (t) is calculated by finding two points of intersection
of the ray l(t) with the k-th rectangle. We denote these points by A(t) =
p1 , b1 ) and B(t) = (a2 , b2 ), and the distance between them as ∆lk (t) =
(a
(a2 − a1 )2 + (b2 − b1 )2 . Then, the value of the line-integral wl;k (t) equals

wl;k (t) = Rk ∆lk (t).

To find the points of intersection of the line lp,s (t) with the rectangle
rk (x, y), one can perform a few verifications, namely, the maximum four fol-
lowing verifications:

t − s(yk + ∆yk )
1. x= ∈ [xk , xk + ∆xk ],
p
234 CHAPTER 4: Main Program
1 1

0.8 0.8

0.6 0.6
2 2

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) (b)

FIGURE 4.3
The rectangle and rays passing through it.

if the line px + sy = t intersects the horizontal line y = yk + ∆yk ,

t − p(xk + ∆xk )
2. y= ∈ [yk , yk + ∆yk ],
s
if the line px + sy = t intersects the vertical line x = xk + ∆xk ,
t − syk
3. x= ∈ [xk , xk + ∆xk ],
p
if the line px + sy = t intersects the horizontal line y = yk ,
t − pxk
4. y= ∈ [yk , yk + ∆yk ],
s
if the line px + sy = t intersects the vertical line x = xk . Verification #1
includes the (1, 0) case, when s = 0, and verification #4 includes the (0, 1)
case, when p = 0.
The positive answer in verification #1 results in the intersection point
(x, yk + ∆yk ). The positive answers in verifications #2, 3, and 4 result in
the intersection points (xk + ∆xk , y), (x, yk ), and (xk , y), respectively. These
different examples can be seen in Figure 4.3 for the rectangle r2 (x, y).
Each of these verifications requires a maximum of one multiplication or
division, because all generators (p, s) ∈ JN,N have the form (1, s) or (p, 1). Two
multiplications and one square root are required for computing ∆lk (t). Thus, a
total of six multiplications and one square root are used to find the intersection
of the line with the rectangle and calculate the length of the intersection. One
additional multiplication is used in the product wl;k (t) = Rk ∆lk (t).
It should be noted that three positive answers can be obtained in the case
when the line lp,s (t) passes along a side of the rectangle. These cases are
(p, s) = (1, 0) and (0, 1), because a rotation of rectangles in the image f(x, y)
is not considered. Four positive answers can also be obtained in the case where
Image Processing 235

the line lp,s (t) passes along a diagonal of the rectangle. In both such cases,
only two different points of intersection are considered. The case where the
line lp,s (t) intersects the rectangle in one point, i.e., the intersection is in a
corner of the rectangle, can be omitted.
The calculation of points of intersection of the line px + sy = t with one
of the sides of the rectangle can also be written on the unique form. Indeed,
consider the general form of the vertical and horizontal lines, p̂x + ŝy = t̂,
where p̂, ŝ = 0 or 1, and t̂ is one value from {xk , yk , xk + ∆xk , yk + ∆yk }. The
the solution of the system 
px + sy = t
p̂x + ŝy = t̂
is    −1     
x ps t 1 ŝ −s t
= = .
y p̂ ŝ t̂ pŝ − sp̂ −p̂ p t̂
The determinant does not require a multiplication. It is probably the simplest
method of calculating the points of intersection of the line with the rectangle,
which uses two multiplications.
Below is the script of function intersection withrectangleA, which is used
in the main program for computing the intersection of the line with one rect-
angle.
% ------------------------------------------------------------
% call: intersection_withrectangleA.m / A.-M. Grigoryan, 2011
% The calculation of coordinates (x,y) of intersection
% of line "p_line=[p,s,-t]" which defines px+sy-t=0
% with the rectangle "p_rect=[x0,y0,x1,y1]"
% This code can be optimized, because p or s equals +1,-1.
% ------------------------------------------------------------
function p=intersection_withrectangleA(p_rect,p_line)
pp=[];
x0=p_rect(1,1); y0=p_rect(1,2);
x1=p_rect(2,1); y1=p_rect(2,2);
p=p_line(1); s=p_line(2); t=-p_line(3);
k=0; % number of intersections
if s==0 % (1,0) case, i.e. p=1 -> line is x=t
if (t>=x0) & (t<=x1)
r1=[t,y0]; r2=[t,y1]; pp=[pp;r1;r2]; k=k+2;
end
elseif p==0 % (0,1) case, i.e. s=1 -> y=t
if (t>=y0) & (t<=y1)
r1=[x0,t]; r2=[x1,t]; pp=[pp;r1;r2]; k=k+2;
end
else % case when p,s>0
x=(t-s*y1)/p; % 1. for line y=y1
if (x>=x0) & (x<=x1)
r1=[x,y1]; pp=[pp;r1]; k=k+1;
end
236 CHAPTER 4: Main Program

y=(t-p*x1)/s;
if (y>=y0) & (y<=y1) % 2. for line x=x1
r1=[x1,y]; pp=[pp;r1]; k=k+1;
end
x=(t-s*y0)/p; % 3. for line y=y0
if (x>=x0) & (x<=x1)
r1=[x,y0]; pp=[pp;r1]; k=k+1;
end
y=(t-p*x0)/s;
if (y>=y0) & (y<=y1) % 4. for line x=x0
r1=[x0,y]; pp=[pp;r1]; k=k+1;
end
end
% Calculate the points of intersection (all cases):
p=[];
if k==0 return end
if k==1 p=[pp;pp]; return end
if k==2
p=pp; return
elseif k>2
if pp(1,:)~=pp(2,:)
p(1:2,:)=pp(1:2,:); return
else
p(1,:)=pp(1,:); p(2,:)=pp(3,:);
end
end
% ---------------------------------------------------------

Another Method: We now describe another method of calculating the

points of intersection of the line with a rectangle. Consider the four parallel
rays passing the corners of the rectangle rk (x, y), as shown in Figure 4.4. We

0.8 (x1,y2)
(x2,y2)
l(t1)
0.6
2
0.4
l(t2)
(x1,y1) (x2,y1)
0.2
l(t3) l(t4)
0
0 0.5 1

FIGURE 4.4
The rectangle and four parallel rays passing its corners.

denote the coordinates of the corners by (x1 , y2 ), (x2 , y2 ), (x1 , y1 ), and (x2 , y1 ),
Image Processing 237

which are on the parallel lines l(t1 ), l(t2 ), l(t3 ), and l(t4 ), respectively. Given
numbers p and s, we consider equations of these parallel lines px + sy = tk ,
k = 1 : 4, where

t1 = px1 + sy2 , t2 = px2 + sy2 , t3 = px1 + sy1 , t4 = px2 + sy1 ,

by assuming t1 > t2 > t3 > t4 .

Let l(t) be the line px + sy = t. The intersection of this line with the
rectangle can be defined by looking the location of the point t in the partition
of the real line (−∞, t4 , t3 , t2 , t1 , ∞). Indeed, it is not difficult to verify that


 0, if t ≥ t1 ,

 ∆1 (t), if t ∈ [t2 , t1 ),

∆l(t) = ∆2 (t), if t ∈ [t3 , t2 ), (4.5)



 ∆ 3 (t), if t ∈ (t 4 , t 3 ),

0, if t ≤ t4 .

The length ∆2 is calculated by

p q
2
∆2 (t) = (x2 − x1 )2 + (y0 − y1 )2 = (x2 − x1 )2 + ((x1 − x2 )p/s)
p
=|x2 − x1 | 1 + (p/s)2 ,

where y0 is the coordinate of the point of intersection of the line with the side
x = x2 , i.e., y0 = y1 + (x1 − x2 )p/s = x1 + (t1 − t2 )/s. The values of ∆1 (t)
and ∆3 (t) are calculated by
t − t1 t − t4
∆1 (t) = ∆2 (t), ∆3 (t) = ∆2 (t).
t2 − t1 t3 − t4
To calculate each value of tk , one operation of multiplication is required,
p one multiplication for ∆2 (t), assuming that the coefficient k(p, s) =
and
1 + (s/p)2 can be calculated in advance. Calculation of ∆1 (t) or ∆3 (t) re-
quired two multiplications. Thus, seven operations of multiplication and one
square root are used in this method.

4.5 Part 3: Transformation of geometry

Given generator (p, s), the set w of line-integrals wl;k (t) is used for calculating
the line-sums v(t), by multiplying the inverse Toeplitz matrix by the integrals,
as b = A−1 w, with the following calculations
1
v= b,
∆l(p, s)
238 CHAPTER 4: Main Program

where ∆l(p, s) denotes the length of the geometrical ray in the image element.
The calculation of the vector b from the set of integrals w is performed as
follows. We consider separately each of four subsets of generators (p, s).
Subset M1: (p = 1 and s ≤ N/2).
The number of line-integrals w(t) calculated along the geometrical rays
¯
l1,s (t) equals (1 + s)(N − 1) + 1. In the s = 0 case, b = w, and
v(t) = b(t)/N, t = 0 : (N − 1).
When s 6= 0, the convolution equation w = Av is described by
√
1 + s2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s + 1)],
s (4.6)
t = 0 : (s + 1)(N − 1).
Therefore, the inverse transform b is calculated in the recurrent form,
b0 = w(0),
b1 = w(1) − b0 ,
b2 = w(2) − (b1 + b0 ),
b3 = w(3) − (b2 + b1 + b0 ),
(4.7)
... ... ... ... ... ...
bs−1 = w(s − 1) − (bs−2 + bs−3 + ... + b1 + b0 ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−s+2 + bt−s+1)),
t = s, s + 1, . . . , (s + 1)(N − 1).
The required sums v(t) of the discrete image are calculated by
s
v(t) = bt √ , t = 0 : (1 + s)(N − 1). (4.8)
N 1 + s2
Below is the script (solution of linearsystemA.m) of the function that is
used for calculating the vector b.
% ---------------------------------------------------------
% call: solution_of_linearsystemA.m
% solution of the system with the upper triangular T-matrix
% w(t)= v(t)+v(t-1)+...+v(t-p+1)
% w=Av
% b=inv(A)w
% such systems of equations correspond to the case A.
function b=solution_of_linearsystemA(p,w)
N=length(w);
z=0;
for i1=1:p
b(i1)=w(i1)-z; z=z+b(i1);
end
for i1=p+1:N
b(i1)=w(i1)-z; z=z-b(i1-p)+b(i1);
end
% ---------------------------------------------------------
Image Processing 239

Subset M2: (p = 1 and s > N/2).

The number of line-integrals w(t) calculated along the geometrical rays
¯
l1,s−N (t) equals (1 + N − s)(N − 1) + 1. The convolution equation w = Av is
described as in (4.6), and the inverse transform b is calculated in the recurrent
form as in (4.7), where s is substituted by N − s. The required sums v(t) of
the discrete image are calculated by
N −s
v(t) = bt p , t = 0 : (1 + N − s)(N − 1).
N 1 + (N − s)2

The numbering is not negative, as it was in A4 (t = (N − 1) : −1 : (N −

s)(N − 1)), because of the permutation of rays by t → t − (N − s)(N − 1).
Subset M3: (p ≤ N/2 and s = 1).
The number of line-integrals w(t) calculated along the geometrical rays
¯
lp,1 (t) equals (p + 1)(N − 1) + 1. In the p = 0 case, b = w and

v(t) = b(t)/N, t = 0 : (N − 1).

When p 6= 0, the convolution equation w = Av is described by

p
1 + p2
w(t) = N [v(t) + v(t + 1) + v(t + 2) + ... + v(t + p − 1)],
p (4.9)
t = (1 + p)(N − 1) : −1 : 0,

where it is assumed that v(k) = 0, for k > (1+p)(N −1). The inverse transform
b = A−1 w is calculated in the following recurrent form (M = (p + 1)(N − 1)):

bM = w(M ),
bM −1 = w(M − 1) − bM ,
bM −2 = w(M − 2) − (bM −1 + bM ),
bM −3 = w(M − 3) − (bM −2 + bM −1 + bM ),
... ... ... ... ... ... (4.10)
bM −p+1 = w(M − p + 1) − (bM −p+2 + bM −p+3 + ... + bM −1 + bM ),
bt = w(t) − (bt+1 + bt+2 + ... + bt+p−2 + bt+p−1) ),
t = M − p, M − p − 1, . . . , 0.

The required sums v(t) of the discrete image are calculated by

p
v(t) = bt p , t = 0 : (1 + p)(N − 1).
N 1 + p2

By changing the parameter t as (p + 1)(N − 1) − t, Equation (4.9) becomes

p
1 + p2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − p + 1)],
p (4.11)
t = 0 : (p + 1)(N − 1),
240 CHAPTER 4: Main Program

where it is assumed that v(k) = 0, for k < 0. Therefore, the solution in (4.10)
is calculated by

b00 = w(0),
b01 = w(1) − b00 ,
b02 = w(2) − (b01 + b00 ),
b03 = w(3) − (b02 + b01 + b00 ),
... ... ... ... ... ... (4.12)
b0p−1 = w(p − 1) − (b0p−2 + b0p−3 + ... + b01 + b0 ),
b0t = w(t) − (b0t−1 + b0t−2 + ... + b0t−p+2 + bt−p+1) ),
t = p, p + 1, . . . , (p + 1)(N − 1),

where b0t = bM −t (or bt = b0M −t ), for t = 0 : M.

This is the same system of equations that was given above in (4.7) for
cases M1 and M2. Therefore, the function solution of linearsystemA can be
used with the following commands:

w=fliplr(w);
b=solution_of_linearsystemA(s,w);
b=fliplr(b);

Subset M4: (p > N/2 and s = 1).

The (p, 1) case is described by the (p−N, 1)-projection, i.e., p is considered
as p − N. The number of line-integrals w(t) calculated along the geometrical
rays ¯
lp−N,1 (t) equals (N −p+1)(N −1)+1. The convolution equation w = Av
is described as
p
1 + (N − p)2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − N + p + 1)],
N −p
t = −(N − p)(N − 1) : (N − 1),
(4.13)
where v(−k) = 0, when k > (N − p)(N − 1). After permutation t → t −
(N − p)(N − 1), the numbering in this convolution equation runs the numbers
t = 0 : (N − p + 1)(N − 1).
The inverse transform b = A−1 w is calculated in the recurrent form as
b0 = w(0),
b1 = w(1) − b0 ,
b2 = w(2) − (b1 + b0 ),
b3 = w(3) − (b2 + b1 + b0 ),
... ... ... ... ... ... (4.14)
bN−p−1 = w(N − p − 1) − (bN−p−2 + bN−p−3 + ... + b1 + b0 ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−N+p+2 + bt−N+p+1) ),
t = N − p, N − p + 1, . . . , (N − p + 1)(N − 1).

Therefore the function solution of linearsystemA.m is used to calculate the

Image Processing 241

vector b, and the required sums v(t) of the discrete image are calculated by

N −p
v(t) = bt p , t = 0 : (N − p + 1)(N − 1). (4.15)
N 1 + (N − p)2

For all generators (p, s) ∈ JN,N , in equations v(t) = bt /∆l(p, s), the corres-
ponding normalized coefficients ∆l(p, s) equal the lengths of the geometrical
rays in the image element. These coefficients are calculated in advance (as
delta r in the program) by
p
1 p2 + s2
∆l(p, s) = ,
N p+s−1

when p, s ≤ N/2. Otherwise, if p > N/2,

p
1 (N − p)2 + 1
∆l(p, 1) = ,
N N −p

and, if s > N/2, p

1 (1 + (N − s)2
∆l(1, s) = .
N N −s

4.6 Part 4: Linear transformation of projections

For the given (p, s)-projection, all line-sums v(t) are used to compose the
0
components fp,s,t of the 2-D discrete paired transform of the reconstructed
image fn,m . After solving the convolution equation w = Av, the vector v
is transferred to the corresponding vector a = (a0 , a1 , ..., aN−1)0 with the
following transformation into the 2-D discrete paired transform.
This transformation is described in detail in Chapter 3 for the N = 8
case. For instance, the first four components of the paired transform with the
triplet-numbers of the subset U (1, 2) are calculated as follows (see (3.54)):
0
f1,2,0 = v0 − v4 + v8 − v12 + v16 − v20 ,
0
f1,2,1 = v1 − v5 + v9 − v13 + v17 − v21 ,
0
f1,2,2 = v2 − v6 + v10 − v14 + v18 ,
0
f1,2,3 = v3 − v7 + v11 − v15 + v19 .

From the set of 22 components vt = v1,2 (t), t = 0 : 21, the components of the
vector a = (a0 , a1 , ..., a7)0 are composed as an = vn + vn+8 + vn+16 , n = 0 : 7.
It is assumed that vn+16 = 0, if n + 16 > 21. By means of the vector a, we
242 CHAPTER 4: Main Program

obtain the following system of equations:

0
f1,2,0 =a0 − a4
0
f1,2,1 =a1 − a5
0
f1,2,2 =a2 − a6
0
f1,2,3 =a3 − a7
0
f2,4,0 =(a0 + a4 ) − (a2 + a6 )
0
f2,4,2 =(a1 + a5 ) − (a3 + a7 )
0
f4,0,0 =(a0 + a4 ) − (a1 + a5 ) + (a2 + a6 ) − (a3 + a7 )
0
f0,0,0 =(a0 + a4 ) + (a1 + a5 ) + (a2 + a6 ) + (a3 + a7 ).
The 2-D paired transform data at triplet-points of U (1, 2) is thus calculated
by the 8-point 1-D discrete paired transform, P (1, 2) = [χ08 ]a.
Similar composition of the vector a holds in the general case, when N = 2r ,
r > 1,
an = vn + vn+N + vn+2N + · · · + vn+(p+s−1)N , n = 0 : (N − 1).
Here, vt = vp,s (t), and t = 0 : (p + s − 1)(N − 1), for the first and third subsets
of generators, when (p, s) = (1, s), s ≤ N/2, and (p, s) = (p, 1), p ≤ N/2. It is
assumed that vt = 0, if t > (p + s − 1)(N − 1).
Below is the script of the function called Avector composition, to accom-
plish the composition of the N -dimensional vector a from the sums v(t).
% -------------------------------------
% call: Avector_composition.m
function A=Avector_composition(N,E)
M=length(E);
A=zeros(1,N);
for i1=1:N
A(i1)=sum(E(i1:N:M));
end
% -------------------------------------

When calculating vector a from the sums vp,s (t) for the second and fourths
subsets of generators (i.e., for subsets M2 and M4), we need to take into ac-
count the fact that the control points of geometrical rays ¯lp,s (t) have negative
values as well. For the N = 8 case, Table 4.4 shows the original numbering of
the rays by t, for these subsets of generators, when (p, s) = (1, s), s > N/2,
and (p, s) = (p, 1), p > N/2, respectively.

Because in the program negative numbering is not used, the numbering

of control points for rays of (p, s)-projections, for such generators, should be
changed. To compose new sets of control points, we first consider the (1, 5)-,
or (1, −3)-projection in the N = 8 case. This projection is described by the
set of 29 geometrical rays which are described by
 
t 1
l(t) = l1,−3 (t) = (x, y); x − 3y = − , t = 7 : −1 : −21,
8 8
Image Processing 243
TABLE 4.4
Case M: Composition of the vector a for the image 8 × 8
(p, s) (a0 , ..., an, ..., a7) v = Bw
(1, 0) an = vn v(t) = b(t)/8, √t = 0 : 7
P
(1, 1) an = m vn+8m v(t) = b(t)/(8
√ 2), t = 0 : 14
P
(1, 2) an = m vn+8m v(t) = bt /(4 5),
√ t = 0 : 21
P
(1, 3) an = m vn+8m v(t) = bt /(8/3
√ 10), t = 0 : 28
P
(1, 4) an = m vn+8m v(t) = bt /(2 17), t = 0 : 35
P √
(1, 5) an = m vn−8m v(t) = bt /(2√17), t = 7 : −1 : −21
P
(1, 6) an = m vn−8m v(t) = bt /(4 √5), t = 7 : −1 : −14
P
(1, 7) an = m vn−8m v(t) = b(t)/(8 2), t = 7 : +1 : −7
(0, 1) an = vn v(t) = b(t)/8√ t = 0 : 7
P
(2, 1) an = m vn+8m v(t) = bt /(4√5), t = 0 : 21
P
(4, 1) an = m vn+8m v(t) = bt /(2 17), t = 0 : 35
P √
(6, 1) an = m vn−8m v(t) = bt /(4 5), t = 7 : −1 : −14

where (x, y) ∈ [0, 1] × [0, 1].

The control points of this set of rays are given below:
0
• 1 • 2• 3• 4 • 5• 6• 7 •
2 3
6 −3 • −2 • −1 • . . . . . 7
6 −6 −5 −4 7
6 −9 • −8 • −7 • . . . . . 7
6 7
6 • •
6 −12 −11 −10 • . . . . . 7
7.
6 • •
6 −15 −14 −13 • . . . . . 7
7
6 −18 • −17 • −16 • . . . . . 7
6 7
4 • • • . . . . .5
−21 −20 −19
• • • . . . . .

As shown in (3.75), the four components of the paired transform with the
triplet-numbers (1, 5, 0), (1, 5, 1), (1, 5, 2), and (1, 5, 3) are calculated as follows:
0
f1,5,0 =(v0 + v−8 + v−16 ) − (v4 + v−4 + v−12 + v−20 ),
0
f1,5,1 =(v1 + v−7 + v−15 ) − (v5 + v−3 + v−11 + v−19 ),
0 (4.16)
f1,5,2 =(v2 + v−6 + v−14 ) − (v6 + v−2 + v−10 + v−18 ),
0
f1,5,3 =(v3 + v−5 + v−13 + v−21 ) − (v7 + v−1 + v−9 + v−17 ).

These components are written in (3.83) by components of the vector a as

0 0
f1,5,0 =a0 − a4 , f1,5,1 = a1 − a5 ,
0 0 (4.17)
f1,5,2 =a2 − a6 , f1,5,3 = a3 − a7 .

The components an , are calculated by

X
an = vn−8m = vn + vn−8 + vn−16 + vn−24 , n = 0 : 7,
m
244 CHAPTER 4: Main Program

where v−22 = v−23 = v−24 = 0. Because subset M2 corresponds to subset A4

with the transformation t → t − (N − p)(N − 1), where p = 5, we consider the
change t → t − 21 in numbering of the control points,
2 21
• 22 • 23 • 24 • 25 • 26 • 27 • 28 •
3
18 19 20
6 • • • . . . . .7
6 15 16 17 7
6 • • • . . . . .7
6 12 13 14 7
6 • • • . . . . .7
6 • • • . . . . . 7.
6 9 10 11 7
6 6 7 8 7
6 • • • . . . . .7
6 3 4 5 7
4 • • • . . . . .5
0
• 1• 2 • . . . . .

Denoting the sums vk0 = vk−21, equations in (4.16) and (4.17) can be
written respectively as
0 0 0
f1,5,0 = (v21 + v13 + v50 ) − (v25
0
+ v170
+ v90 + v10 ),
0 0 0 0 0 0 0
f1,5,1 = (v22 + v14 + v6 ) − (v26 + v18 + v10 + v20 ),
(4.18)
f1,5,2 = (v23 + v15 + v7 ) − (v27 + v19 + v11 + v30 ),
0 0 0 0 0 0 0
0 0 0
f1,5,3 = (v24 + v16 + v80 + v00 ) − (v28
0 0
+ v20 0
+ v12 + v40 ),
and
0
f1,5,0 =−a01 + a05 , 0
f1,5,1 = −a02 + a06 ,
0 (4.19)
f1,5,2 =−a03 + a07 , 0
f1,5,3 = −a04 + a00 .
The new components a0n are calculated by
X
a0n = 0
vn+8m = vn0 + vn+8 0
+ vn+16 0
+ vn+24 , n = 0 : 7,
m
0 0 0
and v29 = v30 = v31 = 0. Therefore, to obtain the order of components
0
an as in (4.17), the sign of all components can be changed with the following
permutation: {1, 2, 3, 4, 5, 6, 7, 0} → {0, 1, 2, 3, 4, 5, 6, 7}, or n → (n−1) mod 8,
n = 0 : 7.
It is not difficult to see that for the (1, 6)-projection, the similar renum-
bering, t → t − 14, of the control points results in the following representation
of the components with triplet-numbers (1, 6, 0), (1, 6, 1), (1, 6, 2) and (1, 6, 3):
0
f1,6,0 =−a02 + a06 , 0
f1,6,1 = −a03 + a07 ,
f1,6,2 =−a04 + a00 ,
0
f1,6,3 = −a05 + a01 .
0

Therefore, the permutation n → (n−2) mod 8 of components a0n can be used,

with following change of the sign, in order to write the above equations as
0
f1,6,0 =a00 − a04 , 0
f1,6,1 = a01 − a05 ,
f1,6,2 =a02 − a06 ,
0
f1,6,3 = a03 − a07 .
0

In the general case of N, for s > N/2, the permutation of components a0n
0 0 0
for f1,s,0 , f1,s,1 , ..., f1,s,N/2−1) is described by

n → (n − s + N/2) mod N, n = 0 : (N − 1). (4.20)

Image Processing 245

Therefore, for such generators (1, s), the function called vectorB indexcorection
is used after Avector composition, to accomplish the correct composition of
the N -dimensional vector a from the sums v(t).
% -------------------------------------------
% call: vectorB_indexcorrection.m
function A=vectorB_indexcorection(p,s,N,E)
N2=N/2;
if p==1
sN2=s-N2;
for t=0:N-1
t1=mod(t-sN2,N)+1;
a=E(t+1); A(t1)=-a;
end
end
if s==1
pN2=p-N2;
for t=0:N-1
t1=mod(t-pN2,N)+1;
a=E(t+1); A(t1)=-a;
end
end
% -------------------------------------------

For the subset M4 (or A2), when the generators are (p, 1) and p is even
and greater that N/2, the change in the numbering of the control points is
calculated in the same way and the function vectorB indexcorection is used af-
ter Avector composition, to accomplish the composition of the N -dimensional
vector a from the sums v(t) = vp,1 (t).

4.7 Part 5: Calculation the 2-D paired transform

After calculating the vector a = (a0 , a1 , a2 , ..., aN−1)0 for the (p, s)-projection,
the N -point 1-D discrete paired transform (DPT) of this vector results in the
part of the 2-D DPT of the reconstructed image, which is denoted by P(p, s),

P(p, s) = [χ0N ]a.

This part contains N components that belong to r + 1, or log2 (N ) + 1 paired

splitting-signals, which are generated by the frequency-points (p, s), (2p, 2s),
246 CHAPTER 4: Main Program

(4p, 4s), (8p, 8s), ..., (N/2p, N/2s), and (0, 0). In other words,

0 0 0 0 0 0

 {fp,s,0 , fp,s,1 , fp,s,2 , fp,s,3 , fp,s,4 , ..., fp,s,N/2−1},

 0 0 0 0


 {f2p,2s,0 , f2p,2s,2 , f2p,2s,4 , ..., f2p,2s,N/2−2 },


 {f 0 , f0 0
, ..., f4p,4s,N/2−4 },
P(p, s) = 4p,4s,0 4p,4s,4 (4.21)


 ... ... ... ...
 0

 {f },

 N/2p,N/2s,0

 {f 0 }.
0,0,0

The number of (p, s)-projections is 3N/2, and the number of splitting-

signals in the 2-D DPT equals 3N − 2, from which, 3N/2 signals are of length
N/2 each, 3N/4 signals are of length N/4 each, ..., 6 signals are of length 2
0
each, 3 signals are of length 1 each, and the last one is {f0,0,0 }.
The part P(p, s) of the 2-D discrete paired transform contains r splitting-
0
signals plus one-point signal {f0,0,0 } which is common for all parts. For diffe-
rent (p, s), parts P(p, s) may contain the same splitting-signals of length N/4,
or N/8, ..., 1. Indeed, each splitting-signal of length N/4 can be seen in two
different parts, each splitting-signal of length N/8 can be seen in four diffe-
rent parts, and so on. Therefore, when performing the complete calculation of
the N -point DPT over the vectors a calculated for all 3N/2 projections, the
0
components fp,s,t of the 2-D DPT of the discrete image fn,m are normalized
as
0 1 0
fp,s,t → k fp,s,t , 2k = g.c.d.(p, s), (p, s) ∈ JN,N \ (0, 0),
2
0 0
and f0,0,0 → f0,0,0 /N. In our program, this method is implemented separately
in function reconstruction ofAllrectanglesOpt2fast.

4.7.1 Method of incomplete 1-D DPT

Another method, which is based on calculation of incomplete 1-D DPT over
vectors a, is implemented in the function reconstruction ofAllrectanglesFast.
For this method, the set of 3N/2 “stop-points” is calculated first and then used
for calculating the incomplete 1-D DFT. Each such stop-point indicates how
many splitting-signals are calculated.
For instance, in the N = 8 case, J8,8 = {(1, 0), (1, 1), (1, 2), ..., (1, 7), (0, 1),
(2, 1), (4, 1), (6, 1)}, and each part P(p, s) contains four splitting-signals of
lengths 4, 2, 1, and 1. The set of stop-points equals

SP (8) = {4, 3, 2, 2, 1, 1, 1, 1, 3, 2, 1, 1}.

This means the following (see Table 3.3): the 8-point DPT for the (1, 0)-
projection is complete, the 8-point DPT for the (1, 1)-projection is used to
calculate only the first 3 splitting-signals, the 8-point DPT for the (1, 2)-
projection is used to calculate only the first 2 splitting-signals, ..., and the 8-
point DPT for the (6, 1)-projection is used to calculate only the first splitting-
Image Processing 247

signal. The sum of stop-points is 22 and equals the number of splitting-signals

in the 8 × 8-point 2-D DPT.
Below is the script of the function called find stoppoints, to calculate the
set of stop-points for all 3N/2 projections in image reconstruction.

% ---------------------------------------------------------
% call: find_stoppoints.m (M. Grigoryan, 2011)
function Stop_points=find_stoppoints(N)
r=log2(N);
Stop_points=zeros(1,N-1);
o=zeros(1,N-1);
N2=N/2; n1=1; n2=N2; M2=N2;
for k=1:r
o(n1:n2)=M2;
N2=N2/2; M2=M2+N2; n1=n2+1; n2=n2+N2;
end
o=N-o(N-1:-1:1);
Stop_points=r-log2(o(1:N-1));
Stop_points=[Stop_points(1)+1 Stop_points];
% add the second part for the generators of parts 3,4.
s1=Stop_points(1)-1;
Stop_points=[Stop_points s1 Stop_points(3:2:N)];
% sum(Stop_points) = 3N-2
% ---------------------------------------------------------

We now consider briefly the concepts of the 1-D DPT and incomplete 1-D
DPT, and then we give the scripts of codes that perform these transforms.

4.7.2 Fast 1-D paired transform

The 1-D N -point discrete paired transformation χ0N , which was described in
§2.4.1, can be defined similarly to the paired transformation in the 2-D case.
The complete system of the paired functions for the N -point DPT is composed
of the paired functions [46, 51]:

 1, if 2k n = (2k t) mod N
0
χ2k ,2k t (n) = −1, if 2k n = (2k t + N/2) mod N

0, otherwise (4.22)
t = 0 : (2r−k−1 − 1), k = 0 : (r − 1),

and χ00,0 (n) ≡ 1, n = 0 : (N − 1).

The paired transform is the unitary frequency-time representation of the
discrete-time signal fn in the form of the set of (r + 1) splitting-signals,
n o
0
χ0N : fn → {f20 k ,0 , f20 k ,2k , f20 k ,2k 2 , ..., f20 k,N/2k+1 −1 }, k = 0 : (r − 1) , {f0,0 },

which are generated by frequencies p = 1, 2, 4, 8, ..., N/2, and 0.

248 CHAPTER 4: Main Program

The 1-D DPT is fast and requires 2N − 2 operations of addition and

subtraction. Figure 4.5 shows the signal-flow graph of the 8-point transform
with 7 operations of addition and 7 subtractions.
f0 c s f0
Q 3 1,0

f1 c Q 0
 3s f1,1
Q Q  
f2 c Q Q   s f0
Q QQ  3 1,2

c QQQ Q s 0
f3 QQQ 3 f1,3
 

Q
Q Q
c  Q Q s 0
f4  Q QsH * f2,0
QQ
 Q Q H 
c  Q Qs H  s 0
f5 
  QQQ H H
H * f2,2

H
 HX X
c  0
f6 Q Qs H s :s f4,0
 Q  HH s X  X
f7 c Qs H XXs f 0
0,0

FIGURE 4.5
Signal-flow graph of the fast eight-point DPT.

Below is the script fst 1d.m of the function which is used for calculating
the N -point DPT, when N > 1 is a power of two.
% ---------------------------------------
% call: fst_1d.m (A.M. Grigoryan, 1996)
function A=fst_1d(A,N,M)
% N=length(A); M=log2(N);
LK=0; NK=N; I=1; II=1;
while I <= M
NK=NK/2; LK=LK+NK; J=II;
while J <= LK
J1=J+NK;
T=A(J1); T1=A(J);
A(J1)=T1+T; A(J)=T1-T; J=J+1;
end
II=LK+1; I=I+1;
end
% ---------------------------------------

One can also use another function with script fastpaired 1d.m for cal-
culating the 1-D DPT, which is written in a simple form, but it does not
accomplish the calculation faster than fst 1d.m.
% ----------------------------------------------
% call: fastpaired_1d.m (M.M. Grigoryan, 2007)
function y=fastpaired_1d(x)
N=length(x);
if N==1
y=x;
else
N2=N/2;
Image Processing 249

x1=x(1:N2); x2=x(N2+1:N);
y1=x1+x2; y2=x1-x2;
y=[y2 fastpaired_1d(y1)];
end
% ----------------------------------------------

To accomplish the calculation of the incomplete N -point DPT with the

stop-point Ls ∈ SP (N ), the function incomplete dpt can be used. For simplic-
ity of reading, this function is also written with the iteration, as the function
fastpaired 1d.

% ----------------------------------------------
% call: incomplete_dpt.m (M.M. Grigoryan, 2011)
function y=incomplete_dpt(x,Ls)
N=length(x);
if (length(x)==1) | (Ls==0)
y=x;
else
N2=N/2;
y(1,1:N2)=x(1,1:N2)-x(1,N2+1:end);
Ls=Ls-1;
if Ls~=0
z(1,1:N2)=x(1,1:N2)+x(1,N2+1:end);
y=[y incomplete_dpt(z,Ls)];
end
end
% ----------------------------------------------

For fast implementation of the inverse 1-D DPT, the function

incopletefst 1d is used in the main program, which works as the function
fst 1d, when the stop-point Ls equals r + 1.

% ------------------------------------------------
% call: incopletefst_1d.m (A.M. Grigoryan, 2011)
function B=incompletefst_1d(A,Ls)
N=length(A); M=log2(N);
LK=0; NK=N;
if Ls>M Ls=M+1; end
I=1; II=1;
while I <= Ls
NK=NK/2; LK=LK+NK; J=II;
while J <= LK
J1=J+NK;
T=A(J1); T1=A(J);
A(J1)=T1+T; A(J)=T1-T; J=J+1;
end
II=LK+1; I=I+1;
end
if Ls<M B=A(1:LK);
250 CHAPTER 4: Main Program

elseif Ls==M B=A(1:N-1);

else B=A;
end
% ------------------------------------------------

4.7.3 Inverse 2-D DPT

The 1-D N -point discrete paired transforms of vectors a, which are calculated
for all 3N/2 projections, are used to compose the 2-D paired transform of
the reconstructed image, which we consider in the form of the square matrix
0
N × N. For example, the components fp,s,t of the 4 × 4-point 2-D DPT are
located in such a matrix as follows:
’1,0,0’ ’1,0,1’ ’0,1,0’ ’0,1,1’
’1,1,0’ ’1,1,1’ ’2,1,0’ ’2,1,1’
’1,2,0’ ’1,2,1’ ’2,0,0’ ’0,2,0’
’1,3,0’ ’1,3,1’ ’2,2,0’ ’0,0,0’

For the 8 × 8 case, the 2-D DPT is transformed into the square matrix as
’1,0,0’ ’1,0,1’ ’1,0,2’ ’1,0,3’ ’0,1,0’ ’0,1,1’ ’0,1,2’ ’0,1,3’
’1,1,0’ ’1,1,1’ ’1,1,2’ ’1,1,3’ ’2,1,0’ ’2,1,1’ ’2,1,2’ ’2,1,3’
’1,2,0’ ’1,2,1’ ’1,2,2’ ’1,2,3’ ’4,1,0’ ’4,1,1’ ’4,1,2’ ’4,1,3’
’1,3,0’ ’1,3,1’ ’1,3,2’ ’1,3,3’ ’6,1,0’ ’6,1,1’ ’6,1,2’ ’6,1,3’
’1,4,0’ ’1,4,1’ ’1,4,2’ ’1,4,3’ ’2,0,0’ ’2,0,2’ ’0,2,0’ ’0,2,2’
’1,5,0’ ’1,5,1’ ’1,5,2’ ’1,5,3’ ’2,2,0’ ’2,2,2’ ’4,2,0’ ’4,2,2’
’1,6,0’ ’1,6,1’ ’1,6,2’ ’1,6,3’ ’2,4,0’ ’2,4,2’ ’4,0,0’ ’0,4,0’
’1,7,0’ ’1,7,1’ ’1,7,2’ ’1,7,3’ ’2,6,0’ ’2,6,2’ ’4,4,0’ ’0,0,0’

This representation of the 2-D DPT was illustrated for the “tomo” image
in Figures 2.16 and 2.17.
Therefore, we first need to locate each part P(p, s) in this square matrix.
Since the components of the 2-D DPT are numbered by triplets (p, s, t), we
use function map pst2nm to transfer the triplets into the corresponding numbers
0
(n, m), which show the locations of components fp,s,t in the matrix.

% ----------------------------------------
% call: map_pst2nm.m
function nm=map_pst2nm(N,p,s,t)
if (p==0) & (s==0) & (t==0)
nm=[N,N];
return
end
if p==1
nm=[s+1,t+1];
elseif s==1
nm=[p/2+1,N/2+t+1];
else
N2=N/2;
nm=N2+map_pst2nm(N2,p/2,s/2,t/2);
Image Processing 251

end
% ---------------------------------------

For example, for the image 256×256, the command map pst2nm(256,p,s,t)
gives nm = (1, 2), (4, 3), and (130, 133), when (p, s, t) = (1, 0, 1), (1, 3, 2) and
(2, 2, 8), respectively. The function checking table is written to calculate the
entire matrix of such triplet-numbers in the general N = 2r case, when r ≥ 2.

% ---------------------------------------------------
% call: checking_table.m
% To calculate the table NxN of triplets (p,s,t).
N=8; L=3*N-2-1; N2=N/2-1;
Jps=setof_ps(N);
A=cell(N);
for k=1:L
p=Jps(k,1); s=Jps(k,2); g=gcd(p,s);
p_instring=num2str(p); s_instring=num2str(s);
ss=strcat(p_instring,’,’,s_instring,’,’);
for t=0:g:N2
t_instring=num2str(t); ss2=strcat(ss,t_instring);
nm=map_pst2nm(N,p,s,t); n=nm(1); m=nm(2);
A(n,m)={ss2};
end
end
A(N,N)={’0,0,0’};
% ---------------------------------------------------

The inverse 2-D DPT is calculated by the formula of reconstruction of the

image, which is given in Statement 2 in (2.49),
r−1
1 X 1 X
0 1 0
fn,m = fp,s,(np+ms) mod N + f ,
2N 2k N 2 0,0,0 (4.23)
k=0 (p,s)∈J2r−k ,2r−k

n, m = 0 : (N − 1).

The inverse 2-D DPT requires only operations of addition/subtraction and

division by powers of two. To accomplish this inverse transform, the function
inverse 2DDPT with the following scripts is used.

% --------------------------------------------------------------
% Call: inverse_2DDPT.m
% split_matrix is the image-matrix of the 2-D DPT.
% Jps=setof_ps(N) is the set of all 3N-2 generators (p,s).
% A.M. Grigoryan and Nan Du, 10/27/2009
% --------------------------------------------------------------
function image_fromPT=inverse_2DDPT(split_matrix,Jps,N)
L=length(Jps);
% The inverse formula of the 2-D paired transform:
image_fromPT=zeros(N);
252 CHAPTER 4: Main Program

NN=2*N; N1=N/2; N2=NN; Np1=N+1;

for m=0:N-1
m1=m+1;
for n=0:N-1
n1=n+1;
total=0; con=0; N_new=N; row=1; col=1; M=1;
r=1; col_add=N; col_r=0; mm=1; Ncol=N1; Nnn=N1;
for k=1:L
p=Jps(k,1); s=Jps(k,2); t=mod(p*n+m*s,N);
if t>=N1
t=t-N1; t=t/mm;
total=total-split_matrix(row,col+t);
else
t=t/mm;
total=total+ split_matrix(row,col+t);
end
N_new2=N_new/2; C=N_new+N_new2;
if k==con+C
con=k; N_new=N_new2; mm=2*mm;
end
row=row+1;
if row==Np1
row=M; M=M+Nnn; r=r+1; Nnn=Nnn/2;
end
if k==col_add
cc=Ncol; col=col+cc;
col_add=col_add+2*cc;
col_r=col_r+1; Ncol=Ncol/2;
end
end
image_fromPT(m1,n1)=total;
end
end
image_fromPT=image_fromPT/NN;
% --------------------------------------------------------------

4.7.4 Preliminary results

Main program main reconstionofRCT8fast.m, which is written in MATLAB,
accomplishes the reconstruction of the image f(x, y) of size N × N, N = 2r ,
from 3N/2 projections. The image is generated randomly as a set of rectangles
with coordinates on the Cartesian lattice XN,N located in the square [0, 1] ×
[0, 1].
As an example, Figure 4.6 shows ten rectangles on the square with different
intensities on the left. These rectangles are on the grid 128 × 128. The image
calculated from 192 projections is shown on the right. The reconstruction is
exact, and the main program required 24 seconds to calculate all projections,
i.e., line-integrals w(t), and reconstruct the image. Incomplete 1-D DPTs were
Image Processing 253

10 rectangles and grid 128x128 reconstructed image 128x128

1

20
0.8
5 4
40
0.6 9 2
7 60
8
10
0.4 6 1 80

0.2 100
3

120
0
0 0.2 0.4 0.6 0.8 1 20 40 60 80 100 120
intensities: 8 10 16 9 16 9 15 4 11 8 cpu time/rect=0.40 min

FIGURE 4.6 (See color insert)

Ten random rectangles (left) and the reconstructed image (right).

used in calculations. The use of the complete 1-D DPT for all projections
required 24.87 seconds. The CPU time is calculated as the average time for
one rectangle. The process of calculation is organized so that for each (p, s)-
projection, the integrals are calculated for each rectangle separately, and then
the results are added.
Figure 4.7 shows ten rectangles on the square [0, 1] × [0, 1] on the left.
The coordinates of these rectangles are on the grid 256 × 256. The image
calculated from 384 projections is shown on the right. The reconstruction is

10 rectangles and grid 256x256 reconstructed image 256x256

1

0.8 1 8 50

9
0.6 100
5
3
0.4 10 150
4

6 7
0.2 200
2

0 250
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 4 7 11 6 14 9 15 5 11 1 cpu time/rect=4.56 min

FIGURE 4.7 (See color insert)

Ten random rectangles (left) and the reconstructed image (right).
254 CHAPTER 4: Main Program

exact, and the main program required 4.56 minutes to calculate all projections
and reconstruct the image. Incomplete 1-D DPTs were used in calculations.
The use of complete 1-D DPTs for all projections required 4.68 minutes, which
exceeds the time with incomplete DPTs, because of many repetitions in the
process of calculating the 2-D DPT of the discrete image.
Time characteristics of calculation, when the program was implemented in
MATLAB are given in Table 4.5. The time for calculation of the 2-D paired
transform and its inverse transform are also given.
TABLE 4.5
Time for MATLAB-based program
N incomplete 1-D DPT complete 1-D DPT 2-D DPT 2-D iDPT
32 0.48s/rec 0.59s/rec 0.0312s 0.0156s
64 2.58s/rec 2.68s/rec 0.0468s 0.0780s
128 24.01s/rec 24.87s/rec 0.2028s 0.5304s
256 4.56m/rec 4.68m/rec 1.8564s 4.4460s

Time data of the program implemented in Ci+ is given in Table 4.6, when
the reconstruction is performed by incomplete 1-D DPTs. The time includes
the calculation of all line-integrals w(t), line-sums v(t), and the calculation of
the 2-D direct and inverse paired transforms. The data in these tables were
TABLE 4.6
Total time for the Ci-based program (14 random rectangles)
N ×N Image scanning & reconstruction with incomplete 1-D DPTs
32 × 32 00 : 00 : 00.03
64 × 64 00 : 00 : 00.11
128 × 128 00 : 00 : 00.79
256 × 256 00 : 00 : 06.17
512 × 512 00 : 00 : 49.76
1024 × 1024 00 : 06 : 59.09

obtained from running programs on a personal computer with Intel Dual CPU
Processor at 3.20-GHz speed. This is the first realization of the method on
Ci+, which can be improved, in order to achieve a fast reconstruction of images
on the large size Cartesian lattices.

4.8 Fast projection integrals by squares

When calculating the line-integrals w(t) for each (p, s)-projection by the pro-
gram with the block-diagram of Figure 4.1, the intersections of G-rays ˆlp,s (t)
with all rectangles are calculated, which slows the program. To make this
process more effective, we consider the following method of G-ray and A-ray
calculation, which can be used for any image f(x, y) on the grid N × N.
Image Processing 255

All image elements (IE) have the same size ∆x × ∆y, which is conside-
red equal to the size of the small square 1/N × 1/N in the program. There
is no need to calculate the coordinates of intersections of G-rays with each
rectangle and then calculate the lengths of intersections, because the lengths
of intersection of rays with each square are equal. It is sufficient, therefore, to
determine the set of squares, along which the rays pass. In other words, one
can represent each ray by the set of such squares

lp,s (t) = {(nk , mk ); k = 1 : K}.

We number each square by the coordinates (nk , mk ) of its bottom left corner.
The number K of such squares depends on p, s, and t, i.e., K = K(p, s, t).
Then, the line-integrals of the image can be calculated by
X
w(t) = ∆l(p, s) f(nk − 0.5, mk − 0.5), t = 0 : (p + s)(N − 1).
(nk ,mk )∈lp,s (t)

Now we describe this method in detail, which we call the method of fast
projection integrals by squares (FPIS).
For that we first consider the simple script (control points rays.m) of the
function that calculates all control points of rays lp,s (t) for the given (p, s)-
projection. The control points show how the rays are numbered. If we con-
sider the process of projection data collection for the parallel X-ray set, as in
computerized tomography [29], the control points determine the location and
numbering of radiation sources on the transmission side and the detectors on
the receiving side.

% ================================================================
function nm=control_points_rays(p,s,N) / M.M. Grigoryan, May 2011
%---------------------------------- 1st part of (p,s)
if (p==1) & (s<=N/2)
b=1;
if s==0
b=0;
end
for i1=1:N-s+b
nm(i1,1:2)=[i1,1];
end
for i2=1:N
for s1=s:-1:1
nm(i1,1:2)=[N-s1+1 i2]; i1=i1+1;
end
end
return
end
%---------------------------------- 2nd part of (p,s)
if (p==1) & (s>N/2)
s=N-s;
256 CHAPTER 4: Main Program

for i1=1:N-s+1
nm(i1,1:2)=[i1,N];
end
for i2=N:-1:1
for s1=s:-1:1
nm(i1,1:2)=[N-s1+1 i2]; i1=i1+1;
end
end
return
end
%---------------------------------- 3rd part of (p,s)
if (s==1) & (p~=1) & (p<=N/2)
b=0;
if p~=0 b=1; end
for i1=1:N-p+b
nm(i1,1:2)=[1,i1];
end
for i2=1:N
for p1=p:-1:1
nm(i1,1:2)=[i2 N-p1+1]; i1=i1+1;
end
end
return
end
%---------------------------------- 4th part of (p,s)
if (s==1) & (p>N/2)
p=N-p;
i1=1;
for i2=N:-1:1
for p1=1:p
nm(i1,1:2)=[i2 p1]; i1=i1+1;
end
end
for i2=p+1:N
nm(i1,1:2)=[1 i2]; i1=i1+1;
end
return
end
% ==============================================================

The number of rays lp,s (t) in the (p, s)-projection equals (p+s)(N −1)+1,
if p, s ≤ N/2, and (N − p + s)(N − 1) + 1, and (p + N − s)(N − 1) + 1, if
p > N/2 and s > N/2, respectively.
For example, when N = 8, the set of 22 control points on the lattice 8 × 8
Image Processing 257

for the (2, 1)-projection equals



 (1, 1), (1, 2), (1, 3), (1, 4), ..., (1, 8),

 (2, 7), (2, 8),

CP (2, 1) = (3, 7), (3, 8),



 ...

(8, 7), (8, 8).

For the (1, 5)-projection, the set of 29 control points equals



 (1, 8), (2, 8), (3, 8), (4, 8), ..., (8, 8),

 (6, 7), (7, 7), (8, 7),

CP (1, 5) = (6, 6), (7, 6), (8, 6),



 ...

(6, 1), (7, 1), (8, 1).

In the program, the numbering of coordinates is performed from 1, not 0.

These two sets of control-points are shown in Figure 4.8.

control points for (2,1) control points for (1,5)

0 0
0 7 6 5 4 3 2 1 0
0 0
0.25 0.25
1 10 9 8
1 1
0.25 0.25
2 13 12 11
2 2
0.375 0.375
3 16 15 14
y and m

y and m

3 3
0.5 0.5
4 19 18 17
4 4
0.625 0.625
5 22 21 20
5 5
0.75 0.75
6 8 10 12 14 16 18 20 25 24 23
6 6
0.875 0.875
7 9 11 13 15 17 19 21 28 27 26
7 7
1 1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1 0 0.12 0.25 0.37 0.5 0.62 0.75 0.87 1
x and n x and n

FIGURE 4.8
Control points for the (2, 1)- and (1, 5)-projections in the 8 × 8 case.

Now we can determine and separate all centers of image elements, or

squares through which the t-th rays lp,s (t) pass. These squares and their
centers are numbered by the same integer coordinates (nk , mk ) of the lat-
tice N × N. For example, when N = 8 and (p, s) = (1, 1), the A-rays are
258 CHAPTER 4: Main Program

represented by the following pairs of coordinates in the lattice 8 × 8:

  

 l(0)  
 (1, 1)
  


 l(1) 




 (2, 1) (1, 2)
  


 l(2) 




 (3, 1) (2, 2) (1, 3)
  


 l(3) 




 (4, 1) (3, 2) (2, 3) (1, 4)
  


 l(4) 




 (5, 1) (4, 2) (3, 3) (2, 4) (1, 5)
  


 l(5) 




 (6, 1) (5, 2) (4, 3) (3, 4) (2, 5) (1, 6)
  
 l(6) 
   (7, 1) (6, 2) (5, 3) (4, 4) (3, 5) (2, 6) (1, 7)

{l1,1 (t)} = l(7) → (8, 1) (7, 2) (6, 3) (5, 4) (4, 5) (3, 6) (2, 7) (1, 8)
  


 l(8) 




 (8, 2) (7, 3) (6, 4) (5, 5) (4, 6) (3, 7) (2, 8)
  


 l(9) 




 (8, 3) (7, 4) (6, 5) (5, 6) (4, 7) (3, 8)
  


 l(10) 




 (8, 4) (7, 5) (6, 6) (5, 7) (4, 8)
  


 l(11) 




 (8, 5) (7, 6) (6, 7) (5, 8)
  


 l(12) 




 (8, 6) (7, 7) (6, 8)
  


 l(13) 




 (8, 7) (7, 8)
  
l(14) (8, 8)

This projection is defined by the corresponding splitting of the grid 8 × 8 by

15 subsets of points.
For the (1, 2)-projection, 22 A-rays are represented by the coordinates in
following way:
  

 l(0)  
 (1, 1)
  


 l(1) 



 (2, 1)
  


 l(2) 

  (3, 1) (1, 2)


  


 l(3) 

  (4, 1) (2, 2)


  


 l(4) 

  (5, 1) (3, 2) (1, 3)


  


 l(5) 

  (6, 1) (4, 2) (2, 3)


  


 l(6) 




 (7, 1) (5, 2) (3, 3) (1, 4)
  


 l(7) 




 (8, 1) (6, 2) (4, 3) (2, 4)
  


 l(8) 




 (7, 2) (5, 3) (3, 4) (1, 5)
  


 l(9) 




 (8, 2) (6, 3) (4, 4) (2, 5)
  
l(10) (7, 3) (5, 4) (3, 5) (1, 6)
{l1,2 (t)} = →

 l(11) 
 
 (8, 3) (6, 4) (4, 5) (2, 6)
  


 l(12) 




 (7, 4) (5, 5) (3, 6) (1, 7)
  


 l(13) 

  (8, 4) (6, 5) (4, 6) (2, 7)


  


 l(14) 




 (7, 5) (5, 6) (3, 7) (1, 8)
  


 l(15) 




 (8, 5) (6, 6) (4, 7) (2, 8)
  


 l(16) 




 (7, 6) (5, 7) (3, 8)
  


 l(17) 




 (8, 6) (6, 7) (4, 8)
  


 l(18) 




 (7, 7) (5, 8)
  


 l(19) 




 (8, 7) (6, 8)
  


 l(20) 




 (7, 8)
  
l(21) (8, 8)
Image Processing 259

In general, for the given (p, s)-projection, the corresponding set of points
can be calculated by the program ps Arays points.m with the following script.
% ==========================================================
function S=ps_Arays_points(p,s,N) / M.M. Grigoryan, May 2011
% ----------------------------------------------------------
S=1;
nm=control_points_rays(p,s,N); sz=size(nm);
%-------- 1th set of generators ---------
if (p==1) & (s<=N/2)
for i1=1:sz(1)
iy=nm(i1,1); ix=nm(i1,2);
six=s*ix; t=iy+six;
while(iy>0) & (ix<=N)
fprintf(’ (%d,%d) ’ ,iy,ix); % (*)
ix=ix+1; six=six+s; iy=t-six;
end
fprintf(’\n’);
end
return
end
%-------- 2nd set of generators ---------
if (p==1) & (s>N/2)
s=N-s;
for i1=1:sz(1)
iy=nm(i1,1); ix=nm(i1,2);
six=s*ix; t=iy-six;
while(ix>0) & (iy>0)
fprintf(’ (%d,%d) ’ ,iy,ix); % (*)
ix=ix-1; six=six+s; iy=t+six;
end
fprintf(’\n’);
end
return
end
%-------- 3rd set of generators ---------
if (s==1) & (p<=N/2)
for i1=1:sz(1)
iy=nm(i1,1); ix=nm(i1,2);
piy=p*ix; t=piy+ix;
while(ix>0) & (iy<=N)
fprintf(’ (%d,%d)’ ,iy,ix); % (*)
iy=iy+1; piy=piy+p; ix=t-piy;
end
fprintf(’\n’);
end
return
end
%-------- 4th set of generators ---------
if (s==1) & (p>N/2)
260 CHAPTER 4: Main Program

p=N-p;
for i1=1:sz(1)
iy=nm(i1,1); ix=nm(i1,2);
piy=p*iy; t=ix-piy;
while(iy<=N & ix<=N)
fprintf(’ (%d,%d)’ ,iy,ix); % (*)
iy=iy+1; piy=piy+p; ix=t+piy;
end
fprintf(’\n’);
end
return
end
% ====================================================
The above two sets of coordinates for rays of the (1, 1)- and (1, 2)-
projections are calculated by using the following commands:
N=8;
p=1; s=1; S=ps_Arays_points(p,s,N);
p=1; s=2; S=ps_Arays_points(p,s,N);
To obtain values of the sums v(t) along each A-ray of the (p, s)-projection,
the sums of the image f(x, y) in the corresponding set of points is calcu-
lated. To accomplish this calculation, we can make a few changes in pro-
gram ps Arays points.m. Namely, we substitute the first command S=1 by
S=zeros(1,size(sz)), the printing commands which are marked as (*) by
S(1,i1)=S(1,i1)+A(iy,ix), and remove commands fprintf(’\n’). Here “A”
stands for the image f(x, y). We add this new function with the syntax defi-
nition function S=ps Arays(A,p,s,N).
Example 4.1 We consider the following image on the lattice 4 × 4 :
 
1 5 9 13
 2 6 10 14 
 3 7 11 15  .
[fd (x, y)] =  

4 8 12 16
The call of the program
N=4; A=1:N*N; A=reshape(A,4,4);
p=1;
for s=0:3
Vps=ps_Arays(A,p,s,N)
end
results in the following values of the (1, 0)-, (1, 1)-, (1, 2)-, and (1, 3)-
projections.
{v1,0 (t); t = 0 : 3} = {28, 32, 36, 40},
{v1,1 (t); t = 0 : 6} = {1, 7, 18, 34, 33, 27, 16},
{v1,2 (t); t = 0 : 9} = {1, 2, 8, 10, 16, 18, 24, 26, 15, 16},
{v1,−1 (t); t = 0 : 6} = {13, 23, 30, 34, 21, 11, 4}.
Image Processing 261

It is not difficult to modify the script of the function ps Arays points.m

and save all coordinates (ix,iy) of points which represent the A-rays in one
array S of size L × N × 2. The first dimension L = (p + s)(N − 1) + 1 is
for the number of A-rays in the (p, s)-projection, the second one is for the
maximal number of coordinates of points, and the last dimension is for the
coordinates. This function in our library is called Arays2grid.m with syntax
definition S=Array2grid(p,s,N). The part of the script of this function with
commands for the first set of generators (1, s), when s ≤ N/2, is given below.

% ======================================================
function S=Arays2grid(p,s,N) / M.M. Grigoryan, June 2011
% ------------------------------------------------------
nm=control_points_rays(p,s,N); sz=size(nm);
S=zeros(sz(1),N,2);
%-------- 1th set of generators ---------
if (p==1) & (s<=N/2)
for i1=1:sz(1)
iy=nm(i1,1); ix=nm(i1,2);
six=s*ix; t=iy+six;
n_pst=0;
while(iy>0) & (ix<=N)
n_pst=n_pst+1; S(i1,n_pst,:)=[iy,ix];
ix=ix+1; six=six+s; iy=t-six;
end
end
return
end
... ... ...

For fast calculation of line-integrals w(t) = wp,s (t) of the (p, s)-projection,
we can use the system of linear equations that is described in Section 4.5. For
instance, for the generator (p, s) = (1, s), when 0 < s ≤ N/2, we consider the
convolution equation

w(t) = ∆l1,s [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s + 1)],

(4.24)
t = 0 : (s + 1)(N − 1),
√
where the length ∆l1,s = N 1 + s2 /s. This system of equations also describes
two other cases, when generators (p, s) are from the subsets M 2 and M 4.
Below is the script of the function Aray2Gray case1.m which accomplishes
the summation along the G-rays from A-rays. The multiplication of the inte-
grals w(t) by the factor of ∆l1,s can be omitted and not used when reconstruc-
ting the sums v(t) from the line-integrals w(t), to simplify the calculations.

% ================================================
% call: Aray2Gray_case1.m / A.-M. Grigoryan 2011
% Vps is the set of sums v(t) of (p,s)-projections:
% Vps={Vps(t),t=1:L}, L=(1+s)(N-1)+1.
262 CHAPTER 4: Main Program

% The generators (p,s)=(1,s), 0<s<=N/2.

% ------------------------------------------------
function Wps=Aray2Gray_case1(Vps,s,L)
Wps=zeros(1,L);
s1=s-1;
for t=1:s1
Wps(t)=sum(Vps(1:t));
end
for t=s:L
Wps(t)=sum(Vps(t-s1:t));
end
dl=N*sqrt(1+s^2)/s; Wps=Wps*dl;
% ================================================

For generators of the subset M 3, when (p, s) = (p, 1) and 0 < p ≤ N/2,
we have another system of equations

w(t) = ∆lp,1 [v(t) + v(t + 1) + v(t + 2) + ... + v(t + p − 1)],

(4.25)
t = (1 + p)(N − 1) : −1 : 0,
p
where the length ∆lp,1 = N p2 + 1/p. The function Aray2Gray case2.m ac-
complishes the calculation of G-rays from A-rays in this case.

% =================================================
% call: Aray2Gray_case2.m / A.-M. Grigoryan 2011
% Vps is the set of sums v(t) of (p,s)-projections:
% Vps={Vps(t),t=1:L}, L=(p+1)(N-1)+1.
% The generators (p,s)=(p,1), 0<p<=N/2.
% ------------------------------------------------
function Wps=Aray2Gray_case2(Vps,p,L)
Wps=zeros(1,L);
p1=p-1;
for t=1:L-p1
Wps(t)=sum(Vps(t:t+p1));
end
for t=L-p1+1:L
Wps(t)=sum(Vps(t:L));
end
dl=N*sqrt(1+p^2)/p; Wps=Wps*dl;
% =================================================

As described in Section 4.5, the program solution of linearsystemA.m cal-

culates the inverse transformation, i.e., the sums v(t) from the integrals w(t).

Example 4.2 To illustrate the line-integrals w(t) and sums v(t), we consider
the image composed of seven rectangles shown in Figure 4.9 in part a. The
data of these rectangles are given in integer form in Table 4.7, where

x0k = 64xk , yk0 = 64yk , ∆x0k = 64∆xk , ∆yk0 = 64∆yk , k = 1 : 7.

Image Processing 263
TABLE 4.7
Integer data of seven
rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 18 19 6 15 9
2 29 29 14 25 6
3 39 39 18 18 13
4 41 14 17 17 3
5 14 38 2 10 10
6 36 11 10 8 15
7 24 9 7 25 4

rectangles on grid 64x64 model: image on grid 64x64

1

10
0.8
3
5 20
2
0.6
30

0.4 1 40
7 4
6 50
0.2
60
0
0 0.5 1 20 40 60
(a) (b)

190 sums v(t) of (1,2)−projection 4 190 integrals w(t) of (1,2)−projection

x 10
3
200
2.5

150 2

1.5
100
1
50
0.5

0 0
0 50 100 150 0 50 100 150
(c) (d)

FIGURE 4.9
(a) Seven rectangles on the square [0, 1] × [0, 1], (b) image composed from
these rectangles, (c) sums v(t), and (d) integrals w(t) of the (1, 2)-projection.

The gray-scale image f(x, y) composed of these seven rectangles on the

grid 64 × 64 is shown in part b. The geometries of A-rays and G-rays, i.e.,
the sums v(t) and integrals w(t) of the (1, 2)-projection, are shown in parts c
and d, respectively. The number of rays in the (1, 2)-projection equals 190 =
(1 + 2)63 + 1. This projection is at angle 90◦ − arctan(1/2) = 63.4349◦ to the
horizontal axis.
264 CHAPTER 4: Main Program

Below is the script of the program for calculating the image f(x, y) from
the random rectangles. Here, imageSQ model stands for the discrete image
f(x, y) and data allrect is used for the data of seven rectangles, as in code
image reconstructionA3A4B1A2.m, which is given in Section 4.2.

% ----------------------------------------------------------------
% composition of the gray-scale image fd(x,y) from rectangles
N=64;
many_rectangles=7;
imageSQ_model=zeros(N,N);
for k=1:many_rectangles
data_1rect=data_allrect(k,:);
x=data_1rect(1); dx=data_1rect(3);
y=data_1rect(2); dy=data_1rect(4); x1=x+dx; y1=y+dy;
d_intensity=data_1rect(5);
imageSQ_model(x:x1,y:y1)=imageSQ_model(x:x1,y:y1)+d_intensity;
end
imageSQ_model=imageSQ_model(:,N:-1:1);
% to display the image use: imagesc(imageSQ_model’);
% ----------------------------------------------------------------

For the same image, Figure 4.10 shows sums v(t) and integrals w(t)
of the (1, 3)- and (2, 1)-projections in parts (a), (b), and (c), (d), respec-
tively. These projections are at angles 90◦ − arctan(1/3) = 71.5651◦ and
90◦ − arctan(2) = 26.5651◦ to the horizontal axis. From these graphs, one can
notice the difference between A-ray and G-ray integrals, namely, the straigh-
tening out of oscillations in the line-integrals w(t).

Example 4.3 We consider the gray-scale discrete image fn,m , which is the
modified Shepp-Logan phantom from the MATLAB library. We can assume
that this image represents the image f(x, y) on the grid 256 × 256. Each value
of the discrete image corresponds to the value of f(x, y) on the (n, m)-th
square of size 1/256 × 1/256. To call this image and put it in the gray-scale
[0, 255], we use the following commands:

N=256; X=phantom(’Modified Shepp-Logan’,N);

imageSQ_model=round(double(X)*255);

Figure 4.10 shows the phantom image in part a, along with the sums v(t)
and integrals w(t) of the (1, 2)-projection in parts c and d, respectively. The
number of rays in the (1, 2)-projection equals 766 = (1 + 2)255 + 1.

The FIPS method is fast. For comparison we consider the time data given
in Table 4.5 for the main program A3-A4-B1-A2. The total time of processing
images of sizes 64 × 64, 128 × 128, and 256 × 256 equals 0.9048 s, 8.4865 s,
and 3.48 min, respectively.
Image Processing 265

253 sums v(t) of (1,3)−projection 4 253 integrals w(t) of (1,3)−projection

x 10
150
2.5

2
100
1.5

50 1

0.5

0 0
0 50 100 150 200 250 0 50 100 150 200 250
(a) (b)

190 sums v(t) of (2,1)−projection 4 190 integrals w(t) of (2,1)−projection

x 10
4

200
3
150
2
100

1
50

0 0
0 50 100 150 0 50 100 150
(c) (d)

FIGURE 4.10
(a,b) Sums v(t) and integrals w(t) of the (1, 3)-projection, and (c,d) sums v(t)
and integrals w(t) of the (2, 1)-projection.

4.9 Selection of projections

In the described case N = 2r , r > 1, we consider the (1, s − N )-projections
instead of (1, s)-projections when s > N/2, and the (p − N, 1)-projections
instead of (p, 1)-projections when p > N/2. This change allows for reducing the
number of projections for these projections, as well as moving the half of the
angles to the second part of the interval of [0, π]. As an example, Figure 4.12
shows these angles before and after the changing projections in the N = 8
case.
Figure 4.13 shows new angles of projections in the interval [0, π] in the
N = 8 case, as well as for the N = 16 and 32 cases. One can observe that a
few angles are located very close to each other, which may lead to difficulties
in obtaining projections for such angles in practical applications. We therefore
consider the problem of changing this set of angles by another more uniform
distributed set of angles in the interval [0, π], which can also be used in our
proposed method of image reconstruction for the set of generators JN,N .
266 CHAPTER 4: Main Program

phantom on the grid 256x256

(a)

766 sums v(t) of (1,2)−projection 766 integrals w(t) of (1,2)−projection

8000
15000

6000
10000
4000

5000
2000

0 0
0 200 400 600 0 200 400 600
(b) (c)

FIGURE 4.11
(a) Phantom image, (b) sums v(t), and (c) integrals w(t) of the (1, 2)-
projection.
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
(a) (b)

FIGURE 4.12
Sets of angles for 12 projections for model 8 × 8.

For that, we first consider the N = 16 case. The set of angles Ψ(16) is


 1.5708 (π/2),

 0.7854, 0.4636, 0.3218, 0.2450,
 0.1974, 0.1651, 0.1419, 0.1244,
Ψ = 2.9997, 2.9764, 2.9442, 2.8966, 2.8198, 2.6779, 2.3562,



 0,

1.1071, 1.3258, 1.4056, 1.4464, 1.7359, 1.8158, 2.0344.
Image Processing 267

N=8

N=16

N=32

0 π/4 π/2 3π/4 π

FIGURE 4.13
Sets of angles for 3N/2 projections for the cases when N = 8, 16, and 32.

Below is the script of the commands to calculate these angles.

N=16; L=N+N/2;
Jps=ps_generators(N);
A1=zeros(1,L); A1(1)=pi/2;
for k=2:L
p=Jps(k,1); if p>N/2 p=p-N; end
s=Jps(k,2); if s>N/2 s=s-N; end
A1(k)=-atan(s/p); % in radians
end
A1=pi/2+A1; A1(1)=pi/2+0;

Let us try to substitute one of these angles, let it be ϑ(1, 7) = 0.1419,

by another angle. The projection number 8 in the set Ψ(16) is the (1, 7)-
projection, which is used to calculate a few splitting-signals, including fT1,7 0 =
{f1,7,0, f1,7,1 , f1,7,2, . . . , f1,7,7}. This signals determines the 2-D DFT of the
0
image at frequency-points of T1,7 = {(1, 7), (3, 5), (5, 3), (7, 1), (9, 15), . . . .}.

n=8; a=A1(n); % 0.1419

p=Jps(n,1); s=Jps(n,2); % (p,s)=(1,7)
Tps=zeros(N/2,2);
k=1;
for m=1:2:N/2
Tps(k,1)=mod(m*p,N); Tps(k,2)=mod(m*s,N); k=k+1;
end; Tps’
% 1 3 5 7
% 7 5 3 1

The generators (1, 7), (3, 5), (5, 3), and (7, 1) are equivalent in the sense
0 0 0 0
that T1,7 = T3,5 = T5,3 = T7,1 , and the number of parallel rays for the (1, 7)-,
(3, 5)-, (5, 3)-, and (7, 1)-projections is the same, (1+7)15+1 = (3+5)15+1 =
. . . = 121. The angles of three new projections are 0.5404, 1.0304, and 1.4289,
respectively (in radians).
268 CHAPTER 4: Main Program

An=zeros(1,N/2);
for k=1:N/2
p1=Tps(k,1); if p1>N/2 p1=p1-N; end
s1=Tps(k,2); if s1>N/2 s1=s1-N; end
An(k)=-atan(s1/p1);
end
An=pi/2+An
% 0.1419 0.5404 1.0304 1.4289 1.4289 1.0304 0.5404 0.1419

The first three angles 0.1419, 0.5404, and 1.0304 are shown beneath the first
0 , as well as two other splitting-
line in Figure 4.14. The splitting-signal fT1,7

N=16

0 π/4 π/2 3π/4 π

FIGURE 4.14
Two sets of angles for 24 projections for the N = 16 case.

signals fT3,5
0 and fT5,3
0 determine the 2-D DFT of the image at the same set
0
of frequency-points, T1,7 . Thus, one of these projections can be used instead
of the (1, 7)-projection. For example, we consider the (3, 5)-projection. That
means the angle 0.1419 of the projection is changed by 0.5404, as shown in
the figure in the second line of angles. In the obtained new set of generators
J16,16, we can also substitute other angles and make all angles more uniformly
distributed in the interval [0, 1]. The question is, can one write an optimal code
to obtain such distribution of the angles in the interval [0, π], for the general
N case?

Problems
Problem 4.1 Develop the program for image reconstruction in the square
[0, 1] × [0, 1], by using the method of fast projection integrals by squares,
which is given in Section 4.8. Discuss the performance of your program.
How fast is your program in comparison with the main program with script
image reconstionA3A4B1A2.m? Show by example that this method is effective.
Assume that the reconstruction is on the lattice 256 × 256 and the image
f(x, y) is composed of a few random rectangles.

Problem 4.2 In the 2nd and 3rd columns of Table 4.5, the time data de-
scribe the total time of program performance. It includes the time required
Image Processing 269

for calculating the line-integrals w(t) for all projections, the time for calcu-
lating the sums v(t) for these integrals, and the time for calculating the 2-D
paired and its inverse transformations. Assume that the measurement data,
or line-integrals are given. Determine how much of the time is used for recon-
structing the image directly from the line-integrals. Give your answer in the
form of a table for images of sizes 64 × 64, 128 × 128, and 256 × 256.

Problem 4.3 Different sets JN,N of generators (p, s) can be used when im-
plementing the proposed method of image reconstruction by transferring the
geometry of G-rays to A-rays. The angles can be spread approximately uni-
formly in the interval [0, π), or they can be moved into a small subinterval of
[0, π). How you would select the angles for projections for image reconstruc-
tion? Discuss the N = 32 example.

Problem 4.4 As mentioned in Section 4.3, the main program of image recon-
struction is referred to as A3-A4-B1-A2 code. Different types of equations for
geometrical rays and convolution are considered. Propose the A3-A4-A1-A2
code, when the same type, A, equations are selected for all four subsets M1,
M2, M3, and M4 of generators (p, s) of the set JN,N .

Problem 4.5 Discuss how would you improve the main program for image
reconstruction, which is described in this chapter. Propose your program for
image reconstruction and give an example of the reconstruction.

Problem 4.6∗ Assume that you are given the image with a rectangle whose
one or a few vertices are not on the Cartesian lattice. For instance, let N = 8
and let the image be the following rectangle of length 1.5(1/8) :

1; if x ∈ [1/4, 3/8 + 1/16], y ∈ [1/4, 3/8],
f(x, y) =
0; otherwise.

The image consists of one square (which we call particle) and one half-square
(which we call semi-particle) and it is shown in Figure 4.15 in part a. In the
ideal case, the image we would like to obtain is
 
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0
f1.5 =
0
.
 0 0 0 0 0 0 0
0 0 1 0.5 0 0 0 0
 
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

Discuss the result of reconstruction of this image. Would you receive a good re-
construction after thresholding the reconstruction? If yes, propose an example
of such thresholding and show your result.
270 CHAPTER 4: Main Program

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) (b)

FIGURE 4.15
Image with (a) one rectangle and (b) shifted particle on the lattice 8 × 8.

Problem 4.7∗ Consider an image that is composed of the following particle

shifted in the lattice 8 × 8 :

1; if x, y ∈ [1/4 − 1/8 : 1/4 + 1/8],
fs (x, y) =
0; otherwise.

This image is shown in Figure 4.15 in part b. The image we would like to
obtain is probably the following:
 
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0
fs = 
0
.
 0 0 0 0 0 0 0
0 1 1 0 0 0 0 0
 
0 1 1 0 0 0 0 0
0 0 0 0 0 0 0 0

Discuss the result of reconstruction of this image. Would you receive a

good reconstruction after thresholding the reconstruction? If yes, propose an
example of such thresholding and show your result.
5
RECONSTRUCTION FOR PRIME SIZE
IMAGE

In this chapter, we describe the method of transferring the geometry of line-

integrals of the image along the rays from the image plane to the Cartesian
lattice N × N, for the case when N is prime. The tensor representation of the
image is considered and the calculation of the discrete image is accomplished
by the inverse tensor transform. The proposed method of image reconstruction
is described in detail for the N = 7 case. A general algorithm and programs are
also given. The tensor representation is very effective for discrete images of size
N × N. The tensor transformation is not orthogonal, but not redundant [48,
55]. The transform and its inverse transform are fast.

5.1 Image reconstruction: Model II

In tensor representation, the discrete image fn,m of size N × N is defined as
a set of splitting-signals of length N each [39],
n o
χ : {fn,m } → fTp,s = {fp,s,t ; t = 0 : (N − 1)} . (5.1)
(p,s)∈J

For the N prime case, the set of generators of splitting-signals is defined as

J = JN,N = {(1, 0), (1, 1), (1, 2), . . . , (1, N − 1), (0, 1)} .
The components of the splitting-signals are calculated as the sum of the disc-
rete image along the parallel lines
X
fp,s,t = {fn,m ; np + ms = t mod N }.
n,m

According to the principle of superposition by direction images, the image

fn,m of size N × N can be composed of (N + 1) splitting-signals as follows:
"N−1 #
1 X
fn,m = f1,s,(n+sm) mod N + f0,1,m − N E[f],
N s=0 (5.2)
n, m = 0 : (N − 1).

271
272 CHAPTER 5: TENSOR TRANSFORM METHOD

For image reconstruction, we will use this simple formula of the inverse tensor
transformation, which is the transformation from the 2-D frequency and 1-D
time space to the 2-D image space.

5.2 Example with image 7 × 7

The N = 7 case is described. N (N + 1) = 56 triplet-numbers (p, s, t) of
components of the 7 × 7-point tensor transform can be defined by the set


 (1, 0, 0),(1, 0, 1),(1, 0, 2),(1, 0, 3),(1, 0, 4),(1, 0, 5),(1, 0, 6)



 (1, 1, 0),(1, 1, 1),(1, 1, 2),(1, 1, 3),(1, 1, 4),(1, 1, 5),(1, 1, 6)



 (1, 2, 0),(1, 2, 1),(1, 2, 2),(1, 2, 3),(1, 2, 4),(1, 2, 5),(1, 2, 6)

(1, 3, 0),(1, 3, 1),(1, 3, 2),(1, 3, 3),(1, 3, 4),(1, 3, 5),(1, 3, 6)
U7,7 =

 (1, 4, 0),(1, 4, 1),(1, 4, 2),(1, 4, 3),(1, 4, 4),(1, 4, 5),(1, 4, 6)



 (1, 5, 0),(1, 5, 1),(1, 5, 2),(1, 5, 3),(1, 5, 4),(1, 5, 5),(1, 5, 6)



 (1, 6, 0),(1, 6, 1),(1, 6, 2),(1, 6, 3),(1, 6, 4),(1, 6, 5),(1, 6, 6)

(0, 1, 0),(0, 1, 1),(0, 1, 2),(0, 1, 3),(0, 1, 4),(0, 1, 5),(0, 1, 6)

In this set, the number of directions is N + 1 = 8. These directions are cal-

culated by the angles ϑ = π − arctg(p/s) to the horizontal, where (p, s) ∈
J7,7 . Given the generator (p, s), the corresponding subset of triplet-numbers
{(p, s, t); t = 0 : 6} is called U (p, s).
We consider the image f(x, y) on the square [0, 1]×[0, 1] with the Cartesian
lattice 7 × 7. The image elements (IE) are considered to have the following
values in the square [0, 1] × [0, 1] :
 
1 1 3 1 2 2 5 1131225
1 1 1 2 1 1 4 1 1 1 2 1 1 4 
 
0 2 2 1 1 1 2 0 2 2 1 1 1 2 
 
[f(x, y)] = 2 3 4 2 1 1 2 3 4 2 1 1 3  .
3 × 49 ⇒ {fn,m } =  
0 3 4 8 9 2 0 0 3 4 8 9 2 0 
 
2 3 1 8 7 6 6  2 3 1 8 7 6 6
1 9 2 4 3 1 2 1924312

Here the image f(x, y) is given together with the discrete image fn,m to be
reconstructed from eight projections of f(x, y). For simplicity of calculations,
all values of the image f(x, y) are given with the factor of 49, which is N 2 .
Each IE has the size ∆x∆y = (1/7)· (1/7) = 1/49. Therefore, for instance, the
value of f2,0 is calculated as the integral of the image f(x, y) in the (2, 0)-th
IE, which equals (3 × 49)∆x∆y = 3.
Image Processing 273

5.2.1 Horizontal projection

We consider the first subset of triplet-numbers (1, 0, t),

U (1, 0) = (1, 0, 0), (1, 0, 1), (1, 0, 2), (1, 0, 3), (1, 0, 4), (1, 0, 5), (1, 0, 6) .
The masks of the corresponding seven tensor functions χ1,0,t, t = 0 : 6, equal
2 3 2 3 2 3
1000000 0100000 0000001
6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07 60 0 0 0 0 0 17
6 7 6 7 6 7
6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07 60 0 0 0 0 0 17
6 7 6 7 6 7
6 1 0 0 0 0 0 0 7, [χ1,0,1 ] = 6 0 1 0 0 0 0 0 7, ..., [χ1,0,6 ] = 6 0 0 0 0 0 0 1 7.
[χ1,0,0 ] = 6 7 6 7 6 7
6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07 60 0 0 0 0 0 17
6 7 6 7 6 7
4 1 0 0 0 0 0 05 4 0 1 0 0 0 0 05 40 0 0 0 0 0 15
1000000 0100000 0000001

Figure 5.1 shows the masks of the first two basic functions. Seven rays
l(0) l(1)
0 0
0 0
0.14 0.14
1 1
0.28 0.28
2 2
0.42 0.42
y and m

y and m

3 3
0.57 0.57
4 4
0.71 0.71
5 5
0.85 0.85
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6

0 0.14 0.28 0.42 0.57 0.71 0.85 1 0 0.14 0.28 0.42 0.57 0.71 0.85 1
x and n x and n

FIGURE 5.1
Masks [χ1,0,0] and [χ1,0,1] of the tensor functions for the (1, 0)-projection.

of the horizontal projection are used to define all components of the tensor
transforms with triplet-numbers of U (1, 0). We define the control points of
these rays by the bullets through which the rays pass. These control points
are numbered as
 
. . . . . . .
 .. .. .. .. .. .. .. 
 . . . . . . . (angle of rays is 90◦ ).
0 1 2 3 4 5 6
• • • • • • •

The sums of the image along the rays l(t) = {x = t + 1/(2N )} are denoted
by vt = v1,0 (t), where t = 0 : 6. The components of the tensor transform with
the triplet-numbers (1, 0, t) are equal to these sums:
f1,0,0 = v0 , f1,0,1 = v1 , f1,0,2 = v2 , . . ., f1,0,6 = v6 . (5.3)
274 CHAPTER 5: TENSOR TRANSFORM METHOD

The length of intersection of the geometrical ray l(t) with the image element
equals ∆x = 1/7. Therefore, the line-integrals equal
1
w(t) = w1,0 (t) = v(t) = 7v(t), t = 0 : 6,
∆x
and v(t) = w1,0 (t)/7, t = 0 : 6. The values of these line-integrals for the given
image f(x, y) can be calculated by

w(0) = 7(1 + 1 + 0 + 2 + 0 + 2 + 1) = 7 × 7 = 49
w(1) = 7(1 + 1 + 2 + 3 + 3 + 3 + 9) = 7 × 22 = 154
w(2) = 7(3 + 1 + 2 + 4 + 4 + 1 + 2) = 7 × 17 = 119
w(3) = 7(1 + 2 + 1 + 2 + 8 + 8 + 4) = 7 × 26 = 182
w(4) = 7(2 + 1 + 1 + 1 + 9 + 7 + 3) = 7 × 24 = 168
w(5) = 7(2 + 1 + 1 + 1 + 2 + 6 + 1) = 7 × 14 = 98
w(6) = 7(5 + 4 + 2 + 3 + 0 + 6 + 2) = 7 × 22 = 154.

The splitting-signal {f1,0,t; t = 0 : 6} = v = w/7 = {7, 22, 17, 26, 24, 14, 22}.
Here, we denote by v and w the vectors of data v(t) and w(t), respectively.

5.2.2 Vertical projection

Consider the vertical projection, or the (0, 1)-projection. The subset of triplet-
numbers (p, s, t) for this projection is

U (0, 1) = (0, 1, 0), (0, 1, 1), (0, 1, 2), (0, 1, 3), (0, 1, 4), (0, 1, 5), (0, 1, 6) ,

and the masks of the corresponding seven tensor functions are

2 3 2 3 2 3
1111111 0000000 0000000
6 0 0 0 0 0 0 07 6 1 1 1 1 1 1 17 6 0 0 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 0 0 0 07 6 0 0 0 0 0 0 07 6 0 0 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 0 0 0 0 7, [χ0,1,1 ] = 6 0 0 0 0 0 0 0 7, ..., [χ0,1,6 ] = 6 0 0 0 0 0 0 0 7.
[χ0,1,0 ] = 6 7 6 7 6 7
6 0 0 0 0 0 0 07 6 0 0 0 0 0 0 07 6 0 0 0 0 0 0 07
6 7 6 7 6 7
4 0 0 0 0 0 0 05 4 0 0 0 0 0 0 05 4 0 0 0 0 0 0 05
0000000 0000000 1111111

The numbered control points of seven rays for this projection are denoted
by the bullets, as shown below
0 
• . .. . ..
 1 • . .. . .. 
2 
 • . .. . .. 
(the angle of rays is 0◦ ).
 
 .. .. .. .. .. .. .. 
 ....... 
 
 6 • . .. . .. 
7
• . .. . ..
Image Processing 275

From the projection data w(t) = w0,1 (t), t = 0 : 6, the vector v, or

the splitting-signal {f0,1,t ; t = 0 : 6} is calculated by f0,1,t = v0,1 (t) =
w0,1 (t)/7, t = 0 : 6. For the given image f(x, y), seven line-integrals w(t)
along the horizontal lines l(t) = {y = t + 1/(2N )} are calculated by

w(0) = 7(1 + 1 + 3 + 1 + 2 + 2 + 5) = 7 × 15 = 105

w(1) = 7(1 + 1 + 1 + 2 + 1 + 1 + 4) = 7 × 11 = 77
w(2) = 7(0 + 2 + 2 + 1 + 1 + 1 + 2) = 7 × 9 = 63
w(3) = 7(2 + 3 + 4 + 2 + 1 + 1 + 3) = 7 × 16 = 112
w(4) = 7(0 + 3 + 4 + 8 + 9 + 2 + 0) = 7 × 26 = 182
w(5) = 7(2 + 3 + 1 + 8 + 7 + 6 + 6) = 7 × 33 = 231
w(6) = 7(1 + 9 + 2 + 4 + 3 + 1 + 2) = 7 × 22 = 154.

Therefore, the splitting-signal {f0,1,t} = v = w/7 = {15, 11, 9, 16, 26, 33, 22}.
Figure 5.2 shows the plots of line-integrals w(t) and splitting-signals for the
(1, 0)- and (0, 1)-projections.

250 40 250 40
w0,1(t) f0,1,t w1,0(t) f1,0,t
200 200
30 30

150 150
20 20
100 100

10 10
50 50

0 0 0 0
0 2 4 6 t 0 2 4 6 t 0 2 4 6 t 0 2 4 6 t

FIGURE 5.2
Line-integrals w(t) and splitting-signals for the (0, 1)- and (1, 0)-projections.

5.2.3 Diagonal projection

Now we consider the subset of triplet-numbers

U (1, 1) = (1, 1, 0), (1, 1, 1), (1, 1, 2), (1, 1, 3), (1, 1, 4), (1, 1, 5), (1, 1, 6) .

The masks of the corresponding seven tensor functions equal

2 3 2 3 2 3
1000000 0100000 0010000
6 0 0 0 0 0 0 17 61 0 0 0 0 0 07 60 1 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 0 0 1 07 60 0 0 0 0 0 17 61 0 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 0 1 0 0 7, [χ1,1,1 ] = 6 0 0 0 0 0 1 0 7, [χ1,1,2 ] = 6 0 0 0 0 0 0 1 7
[χ1,1,0 ] = 6 7 6 7 6 7
6 0 0 0 1 0 0 07 60 0 0 0 1 0 07 60 0 0 0 0 1 07
6 7 6 7 6 7
4 0 0 1 0 0 0 05 40 0 0 1 0 0 05 40 0 0 0 1 0 05
0100000 0010000 0001000
276 CHAPTER 5: TENSOR TRANSFORM METHOD
2 3 2 3 2 3
0001000 0000100 0000010
6 0 0 1 0 0 0 07 6 0 0 0 1 0 0 07 60 0 0 0 1 0 07
6 7 6 7 6 7
6 0 1 0 0 0 0 07 6 0 0 1 0 0 0 07 60 0 0 1 0 0 07
6 7 6 7 6 7
6 1 0 0 0 0 0 0 7, [χ1,1,4 ] = 6 0 1 0 0 0 0 0 7, [χ1,1,5 ] = 6 0 0 1 0 0 0 0 7
[χ1,1,3 ] = 6 7 6 7 6 7
6 0 0 0 0 0 0 17 6 1 0 0 0 0 0 07 60 1 0 0 0 0 07
6 7 6 7 6 7
4 0 0 0 0 0 1 05 4 0 0 0 0 0 0 15 41 0 0 0 0 0 05
0000100 0000010 0000001
2 3 2 3
0000001 1111111
60 0 0 0 0 1 0 7 61 1 1 1 1 1 1 7
6 7 6 7
60 0 0 0 1 0 0 7 61 1 1 1 1 1 1 7 5
6 7 6 7 X
[χ1,1,6 ] = 6 0 0 0 1 0 0 0 7 = 61 1 1 1 1 1 1 7 − [χ1,1,t ]. (5.4)
6 7 6 7
6 0 0 1 0 0 0 0 7 6 1 1 1 1 1 1 1 7 t=0
6 7 6 7
40 1 0 0 0 0 0 5 41 1 1 1 1 1 1 5
1000000 1111111

The composition of the last matrix by the first six matrices shows the
redundancy of the 2-D tensor transform, when the component f1,1,6 can be
calculated as
X 5
5 X 5
X 6
X 5
X
f1,1,6 = fn,m − f1,1,t = f1,0,t − f1,1,t .
n=0 m=0 t=0 t=0 t=0

We define and number the control points of the parallel rays for this pro-
jection by the bullets as
0 
• . . . . . .
 1• . . . . . . 
2 
 • . . . . . .
(the angle of rays is − 45◦ ).
 
 .. .. .. .. .. .. .. 
 . . . . . . .
 
 5• . . . . . . 
6 7 8 9 10 11 12
• • • • • • •

Seven masks of tensor functions χ1,1,t(n, m) with the parallel rays on which
the coefficients 1 are situated are shown in Figures 5.3–5.5. The set of all 14
rays is also shown in Figure 5.5.
The sums of the discrete image along these rays are denoted by vt = v1,1 (t),
where t = 0 : 12. One can notice that the components of the tensor transform
with the triplet-numbers (p, s, t) ∈ U (1, 1) can be calculated as follows:

f1,1,0 = v0 + v7 , f1,1,1 = v1 + v8 ,
f1,1,2 = v2 + v9 , f1,1,3 = v3 + v10 , (5.5)
f1,1,4 = v4 + v11 , f1,1,5 = v5 + v12 , f1,1,6 = v6 .

Thus, the above system of linear equations can be written as

f1,1,t = vt + vt+7 = v1,1,t + v1,1,t+7 , t = 0 : 6,

Image Processing 277

l(0) l(1)
0 0
0 l(7) 0
0.14 0.14
1 1 l(8)
0.29 0.29
2 2
y and m

y and m
0.43 0.43
3 3
0.57 0.57
4 4
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

l(2) l(3)
0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 l(9) 2
y and m

y and m
0.43 0.43
3 3 l(10)
0.57 0.57
4 4
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.3
Parallel rays in the masks [χ1,1,t], t = 0 : 3, for the (1, 1)-projection.

l(4) l(5)
0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 2
y and m

y and m

0.43 0.43
3 3
0.57 0.57
4 l(11) 4
0.71 0.71
5 5 l(12)
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.4
Parallel rays in the masks [χ1,1,t], t = 4, 5 for the (1, 1)-projection.

where v13 = 0. The sums of the image v(t) = v1,1,t can be calculated from the
integrals of the image f(x, y) along the corresponding geometrical rays,
 t 1
l(t) = l1,1 (t) = (x, y); x + y = + , t = 0 : 12,
7 7
278 CHAPTER 5: TENSOR TRANSFORM METHOD

l(6) l(0) l(1) l(2) l(3) l(4) l(5) l(6)

0 0
0 0 l(7)
0.14 0.14
1 1 l(8)
0.29 0.29
2 2 l(9)
y and m

y and m
0.43 0.43
3 3 l(10)
0.57 0.57
4 4 l(11)
0.71 0.71
5 5 l(12)
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.5
Mask [χ1,1,6] and the set of arithmetical rays of the (1, 1)-projection.

where (x, y) ∈ [0, 1] × [0, 1]. Indeed, the length of

√ the intersection
√ of the geo-
metrical rays in the image element equals ∆l = 2∆x = 2/7. Each integral
along the geometrical ray is proportional to the sum of the image along the
corresponding arithmetical ray,

v(t) √
w(t) = w1,1 (t) = ∆l = 7 2v(t), t = 0 : 12.
(∆x)2
√
Therefore v(t) = w(t)/(7 2), t = 0 : 12. The number of rays is calculated by
(1 + 1)(7 − 1) + 1 = 13.
For the given image f(x, y), the line-integrals equal
√ √
w(0)= 7√2(1) = 7√2
w(1)= 7√2(1 + 1) = 7√2 · 2
w(2)= 7√2(3 + 1 + 0) = 7√2 · 4
w(3)= 7√2(1 + 1 + 2 + 2) = 7√2 · 6
w(4)= 7√2(2 + 2 + 2 + 3 + 0) = 7√2 · 9
w(5)= 7√2(2 + 1 + 1 + 4 + 3 + 2) = 7√2 · 13
w(6)= 7√2(5 + 1 + 1 + 2 + 4 + 3 + 1) = 7√2 · 17
w(7)= 7√2(4 + 1 + 1 + 8 + 1 + 9) = 7√2 · 24
w(8)= 7√2(2 + 1 + 9 + 8 + 2) = 7√2 · 22
w(9)= 7√2(3 + 2 + 7 + 4) = 7√2 · 16
w(10) = 7√2(0 + 6 + 3) = 7√2 · 9
w(11) = 7√2(6 + 1) = 7√2 · 7
w(12) = 7 2(2) =7 2 · 2

The sums of the image along the corresponding arithmetical rays can be writ-
ten as the vector v = (1, 2, 4, 6, 9, 13, 17, 24, 22, 16, 9, 7, 2)0. The components of
Image Processing 279

the splitting-signal are calculated by

f1,1,0 = v(0) + v(7) = 1 + 24 = 25

f1,1,1 = v(1) + v(8) = 2 + 22 = 24
f1,1,2 = v(2) + v(9) = 4 + 16 = 20
f1,1,3 = v(3) + v(10) = 6 + 9 = 15
f1,1,4 = v(4) + v(11) = 9 + 7 = 16
f1,1,5 = v(5) + v(12) = 13 + 2 = 15
f1,1,6 = v(6) = 17.

Figure 5.6 shows the graphs of the line-integrals w(t), sums v(t), and splitting-
signal for the (1, 1)-projection.

250 25 30
w1,1(t) v1,1(t) f1,1,t
200 20 25

20
150 15
15
100 10
10
50 5 5

0 0 0
0 5 10 t 0 5 10 t 0 2 4 6
t

FIGURE 5.6
Line-integrals w(t), sums v(t), and splitting-signal of the (1, 1)-projection.

5.2.4 (1, 2)-Projection

Consider the next subset of triplet-numbers,

U (1, 2) = (1, 2, 0), (1, 2, 1), (1, 2, 2), (1, 2, 3), (1, 2, 4), (1, 2, 5), (1, 2, 6) .

The masks of the tensor functions with seven triplet-numbers of U (1, 2) are

2 3 2 3 2 3
1000000 0100000 0010000
6 0 0 0 0 0 1 07 6 0 0 0 0 0 0 17 6 1 0 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 1 0 0 07 6 0 0 0 0 1 0 07 6 0 0 0 0 0 1 07
6 7 6 7 6 7
6 0 1 0 0 0 0 0 7, [χ1,2,1 ] = 6 0 0 1 0 0 0 0 7, [χ1,2,2 ] = 6 0 0 0 1 0 0 0 7,
[χ1,2,0 ] = 6 7 6 7 6 7
6 0 0 0 0 0 0 17 6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07
6 7 6 7 6 7
4 0 0 0 0 1 0 05 4 0 0 0 0 0 1 05 4 0 0 0 0 0 0 15
0010000 0001000 0000100
280 CHAPTER 5: TENSOR TRANSFORM METHOD
2 3 2 3 2 3
0001000 0000100 0000010
6 0 1 0 0 0 0 07 6 0 0 1 0 0 0 07 6 0 0 0 1 0 0 07
6 7 6 7 6 7
6 0 0 0 0 0 1 07 6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07
6 7 6 7 6 7
6 0 0 0 1 0 0 0 7, [χ1,2,4 ] = 6 0 0 0 0 0 1 0 7, [χ1,2,5 ] = 6 0 0 0 0 0 0 1 7,
[χ1,2,3 ] = 6 7 6 7 6 7
6 0 1 0 0 0 0 07 6 0 0 0 1 0 0 07 6 0 0 0 0 1 0 07
6 7 6 7 6 7
4 0 0 0 0 0 0 15 4 0 1 0 0 0 0 05 4 0 0 1 0 0 0 05
0000100 0000001 1000000
2 3 2 3
0000001 1111111
60 0 0 0 1 0 07 61 1 1 1 1 1 17
6 7 6 7
60 0 1 0 0 0 07 61 1 1 1 1 1 17 X 5
6 7 6 7
[χ1,2,6 ] = 66 1 0 0 0 0 0 0 7 = 61 1 1 1 1 1 17 −
7 6 7 [χ1,2,t ].
6 0 0 0 0 0 1 0 7 6 1 1 1 1 1 1 1 7 t=0
6 7 6 7
40 0 0 1 0 0 05 41 1 1 1 1 1 15
0100000 1111111

We define the control points of these rays by the bullets and number them
as shown below:
0 1
• 2• 3• 4 • 5• 6•
2 3
•
7 8 7
6 . . . . . • •7
. . . . 9 • 10 • 7
6
6 .
. . . . 11 • 12 • 7
6 7
6 . (angle of rays is − arctan 1/2 = −26.565◦ ).
. . . . 13 • 14 • 7
6 7
6
6 . 7
15 16 5
4 . . . . . • •
. . . . . 17 • 18 •

The number of parallel rays is calculated by (p + s)(N − 1) + 1 = 3· 6 + 1 =

19. The angle of this projection is arctan(2). These masks with the rays, on
which the coefficients 1 are situated, are shown in Figures 5.7 and 5.8. Nineteen
arithmetical rays of this projection
 
t 3
l(t) = l1,2 (t) = (x, y); x + 2y = + , t = 0 : 18,
7 14

are also shown in Figure 5.8. The components of the tensor transform with
the triplet-numbers of the subset U (1, 2) can be calculated by

f1,2,0 = v0 + v7 + v14 , f1,2,1 = v1 + v8 + v15 ,

f1,2,2 = v2 + v9 + v16 , f1,2,3 = v3 + v10 + v17 , (5.6)
f1,2,4 = v4 + v11 + v18 , f1,2,5 = v5 + v12 , f1,2,6 = v6 + v13 ,

where variables vt = v1,2 (t) denote the sums of the discrete image along the
rays l(t), t = 0 : 18. Therefore, the system of equations can be written as

f1,2,t = vt + vt+7 + vt+14 = v1,2,t + v1,2,t+7 + v1,2,t+14, t = 0 : 6,

where v19 = v20 = 0.

To calculate the signal {f1,2,t } from the projection data, we will simplify
the equation describing the linear relation between the line-integral w(t) and
Image Processing 281

l(0) l(1)
0 0
0 l(7) 0
0.14 0.14 l(8)
1 1
0.29 0.29
2 2
y and m

y and m
0.43 0.43
3 3
0.57 l(14) 0.57
4 4 l(15)
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

l(2) l(3)
0 0
0 0
0.14 0.14
1 l(9) 1
0.29 0.29 l(10)
2 2
y and m

y and m
0.43 0.43
3 3
0.57 0.57
4 4
0.71 l(16) 0.71
5 5 l(17)
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.7
Four masks [χ1,2,t], t = 0 : 3, for the (1, 2)-projection.

sums v(t), t = 0 : 18, by considering another set of 19 geometric rays. For

that, we first consider the part of the mask of the tensor function χ1,2,0 (n, m)
with coefficients 1 on the rays l(0) and l(7),√which is shown√ in Figure 5.9.
The first ray intersects the IE by the length 5/2∆x = 7 5/2. The ray l(7)
intersects the (1, 3)-th IE by the same length, but each of the (0, 3)- and (2, 3)-
th IEs by a length that is twice as small. The integral of the image along the
geometrical ray between the rays l(6) and l(7) can √be calculated from the sums
of the image along two arithmetical rays by w = 7 5/2[v1,2 (6) + v1,2(7)]. This
geometrical ray is considered as ray number 7 in the new set of parallel rays,
which are shown in Figure 5.10. These rays are defined by

˜ t 1
l(t) = l1,2 (t − 0.5) = {(x, y); x + 2y = + }, t = 0 : 18.
7 7

The relation between the line-integrals along the geometrical rays ˜l(t) and
the line-sums along the arithmetical rays l(t) is described by the following
system of linear equations:
√
w(t) = w(˜l(t)) = 7 5/2[v1,2 (t − 1) + v1,2 (t)], t = 0 : 18,
282 CHAPTER 5: TENSOR TRANSFORM METHOD

l(4) l(5)
0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 l(11) 2
l(12)
y and m

y and m
0.43 0.43
3 3
0.57 0.57
4 4
0.71 0.71
5 5
0.86 l(18) 0.86
6 6 l(19)
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

l(0) l(1) l(2) l(3) l(4) l(5)

0 l(6) 0 l(6)
0 0 l(7)
0.14 0.14 l(8)
1 1 l(9)
0.29 0.29 l(10)
2 2 l(11)
y and m

y and m
0.43 0.43 l(12)
3 l(13) 3 l(13)
0.57 0.57 l(14)
4 4 l(15)
0.71 0.71 l(16)
5 5 l(17)
0.86 0.86 l(18)
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.8
Masks [χ1,2,t], t = 4 : 6, and the set of 19 arithmetical rays.

l(0)
0 l(6)
0
l(7)
0.14
y and m

1
0.29
2
0.43
3
0.57
4 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n

FIGURE 5.9
Four parallel rays of the (1, 2)-projection.

where v1,2 (−1) = 0. These equations can be written in matrix form as

    
w1,2 (0) 1 0 0 0 0 · · · 00 0 v1,2 (0)
 w1,2 (1)   1 1 0 0 0 · · · 00 0  v1,2 (1) 
    
 w1,2 (2)   0 1 1 0 0 · · · 00 0  v1,2 (2) 
  √   
 0 0 1 1 0 · · · 00 0  v1,2 (3) 
 w1,2 (3)  7 5  
 
w =  w1,2 (4)  =  0 0 0 1 1 · · · 00 0  v1,2 (4) .
  2   
 ..   .. .. .. .. .. . . .. .. ..  .. 
 .   . . . . . . . . .  . 
    
 w1,2 (17)   0 0 0 0 0 · · · 11 0  v1,2 (17) 
w1,2 (18) 0 0 0 0 0 · · · 01 1 v1,2 (18)
Image Processing 283
˜l(0) ˜l(1) ˜l(2) ˜l(3) ˜l(4) ˜l(5) ˜l(6)
0 ˜l(7)
0 ˜l(8)
0.14 ˜l(9)
1 ˜l(10)
0.29 ˜l(11)
2 ˜l(12)
0.43 ˜l(13)
y and m

3 ˜l(14)
0.57 ˜l(15)
4 ˜l(16)
0.71 ˜l(17)
5 ˜l(18)
0.86

1
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.10
Geometrical rays for calculating line-integrals of the (1, 2)-projection.

The above Toeplitz matrix 19 × 19 has the triangle inverse matrix, and all
sums along the arithmetical rays are calculated by
    
v1,2 (0) 1 0 0 0 0 ··· 0 0 w1,2 (0)
 v1,2 (1)   −1 1 0 0 0 · · · 0 0  w1,2 (1) 
    
 v1,2 (2)   1 −1 1 0 0 · · · 0 0  w1,2 (2) 
    
 −1 1 −1 1 0 · · · 0 0  w1,2 (3) 
 v1,2 (3)  2   
 
 v1,2 (4)  = √  1 −1 1−1 1 · · · 0 0  w1,2 (4)  .
  7 5  
 ..   .. .. .. .. .. . . .. ..  .. 
.  . . . . . . . .  .
    
  
 v1,2 (17)   −1 1 −1 1 −1 · · · 1 0  w1,2 (17)
v1,2 (18) 1 −1 1−1 1 · · · −1 1 w1,2 (18)
The required
√ sums v(t) of the discrete image fn,m are calculated by v(t) =
bt /(7 5/2), where the coefficients bt are calculated as follows:
b0 =w(0),
b1 =w(1) − b0 , (5.7)
bt =w(t) − bt−1 , t = 2, 3, . . . , 18.
For the given image f(x, y), the line-integrals are calculated by
w(0) = K(1) =K · 1
w(1) = K(1 + 1) =K · 2
w(2) = K(3 + 1 + 1) =K · 5
w(3) = K(1 + 3 + 1 + 1) = K · 6
284 CHAPTER 5: TENSOR TRANSFORM METHOD

w(4) = K(2 + 1 + 1 + 1 + 0) =K ·5
w(5) = K(2 + 2 + 2 + 1 + 2 + 0) =K ·9
w(6) = K(5 + 2 + 1 + 2 + 2 + 2 + 2) = K · 16
w(7) = K(5 + 1 + 1 + 1 + 2 + 3 + 2) = K · 15
w(8) = K(4 + 1 + 1 + 1 + 4 + 3 + 0) = K · 14
w(9) = K(4 + 1 + 1 + 2 + 4 + 3 + 0) = K · 15
w(10) = K(2 + 1 + 1 + 2 + 4 + 3 + 2) = K · 15
w(11) = K(2 + 1 + 1 + 8 + 4 + 3 + 2) = K · 21
w(12) = K(3 + 1 + 9 + 8 + 1 + 3 + 1) = K · 26
w(13) = K(3 + 2 + 9 + 8 + 1 + 9 + 1) = K · 33
w(14) = K(0 + 2 + 7 + 8 + 2 + 9) =K · 28
w(15) = K(0 + 6 + 7 + 4 + 2) =K · 19
w(16) = K(6 + 6 + 3 + 4) =K · 19
w(17) = K(6 + 1 + 3) =K · 10
w(18) = K(2 + 1) =K ·3
√
where K = 7 5/2. Therefore, the sums v(t) = K −1 w(t) − v(t − 1) of the
discrete image fn,m can be calculated as follows:

v(0) = K −1 w(0) =1 =1
v(1) = K −1 w(1) − v(0) = 2 − 1 = 1
v(2) = K −1 w(2) − v(1) = 5 − 1 = 4
v(3) = K −1 w(3) − v(2) = 6 − 4 = 2
v(4) = K −1 w(4) − v(3) = 5 − 2 = 3
v(5) = K −1 w(5) − v(4) = 9 − 3 = 6
v(6) = K −1 w(6) − v(5) = 16 − 6 = 10
v(7) = K −1 w(7) − v(6) = 15 − 10 = 5
v(8) = K −1 w(8) − v(7) = 14 − 5 = 9
v(9) = K −1 w(9) − v(8) = 15 − 9 = 6
v(10) = K −1 w(10) − v(9) = 15 − 6 = 9
v(11) = K −1 w(11) − v(10) = 21 − 9 = 12
v(12) = K −1 w(12) − v(11) = 26 − 12 = 14
v(13) = K −1 w(13) − v(12) = 33 − 14 = 19
v(14) = K −1 w(14) − v(13) = 28 − 19 = 9
v(15) = K −1 w(15) − v(14) = 19 − 9 = 10
v(16) = K −1 w(16) − v(15) = 19 − 10 = 9
v(17) = K −1 w(17) − v(16) = 10 − 9 = 1
v(18) = K −1 w(18) − v(17) = 3 − 1 = 2.

The components of the splitting-signal {f1,2,t ; t = 0 : 6} are calculated by

f1,2,0 = v(0) + v(7) + v(14) = 1 + 5 + 9 = 15

f1,2,1 = v(1) + v(8) + v(15) = 1 + 9 + 10 = 20
f1,2,2 = v(2) + v(9) + v(16) = 4 + 6 + 9 = 19
f1,2,3 = v(3) + v(10) + v(17) = 2 + 9 + 1 = 12
f1,2,4 = v(4) + v(11) + v(18) = 3 + 12 + 2 = 17
f1,2,5 = v(5) + v(12) = 6 + 14 = 20
f1,2,6 = v(6) + v(13) = 10 + 19 = 29.
Image Processing 285

Figure 5.11 shows the graphs of the integrals w(t), sums v(t), and splitting-
signal for the (1, 2)-projection.

20 30
250
w1,2(t) v1,2(t) f1,2,t
25
200 15
20
150
10 15
100
10
5
50 5

0 0 0
0 5 10 15 t 0 5 10 15 t 0 2 4 6 t

FIGURE 5.11
Line-integrals w(t), sums v(t), and splitting-signal for the (1, 2)-projection.

5.2.5 (1, 3)-projection

We consider the subset of triplets

U (1, 3) = (1, 3, 0), (1, 3, 1), (1, 3, 2), (1, 3, 3), (1, 3, 4), (1, 3, 5), (1, 3, 6) .

Masks of the seven tensor functions with triplet-numbers of this subset are
2 3 2 3 2 3
1000000 0100000 0010000
6 0 0 0 0 1 0 07 6 0 0 0 0 0 1 07 6 0 0 0 0 0 0 17
6 7 6 7 6 7
6 0 1 0 0 0 0 07 6 0 0 1 0 0 0 07 6 0 0 0 1 0 0 07
6 7 6 7 6 7
6 0 0 0 0 0 1 0 7, [χ1,3,1 ] = 6 0 0 0 0 0 0 1 7, |χ1,3,2 ] = 6 1 0 0 0 0 0 0 7,
[χ1,3,0 ] = 6 7 6 7 6 7
6 0 0 1 0 0 0 07 6 0 0 0 1 0 0 07 6 0 0 0 0 1 0 07
6 7 6 7 6 7
4 0 0 0 0 0 0 15 4 1 0 0 0 0 0 05 4 0 1 0 0 0 0 05
0001000 0000100 0000010
2 3 2 3 2 3
0001000 0000100 0000010
6 1 0 0 0 0 0 07 6 0 1 0 0 0 0 07 6 0 0 1 0 0 0 07
6 7 6 7 6 7
6 0 0 0 0 1 0 07 6 0 0 0 0 0 1 07 6 0 0 0 0 0 0 17
6 7 6 7 6 7
|χ1,3,3 ] = 6
6 0 1 0 0 0 0 0 7, [χ1,3,4 ] = 6 0 0 1 0 0 0 0 7, [χ1,3,5 ] = 6 0 0 0 1 0 0 0 7,
7 6 7 6 7
6 0 0 0 0 0 1 07 6 0 0 0 0 0 0 17 6 1 0 0 0 0 0 07
6 7 6 7 6 7
4 0 0 1 0 0 0 05 4 0 0 0 1 0 0 05 4 0 0 0 0 1 0 05
0000001 1000000 0100000
2 3 2 3
0000001 1111111
60 0 0 1 0 0 07 61 1 1 1 1 1 17
6 7 6 7
61 0 0 0 0 0 07 61 1 1 1 1 1 17 X 5
6 7 6 7
[χ1,3,6 ] = 66 0 0 0 0 1 0 0 7 = 61 1 1 1 1 1 17 −
7 6 7 [χ1,3,t ].
6 0 1 0 0 0 0 0 7 6 1 1 1 1 1 1 1 7 t=0
6 7 6 7
40 0 0 0 0 1 05 41 1 1 1 1 1 15
0010000 1111111
286 CHAPTER 5: TENSOR TRANSFORM METHOD

The masks of these tensor functions are shown in Figures 5.12 and 5.13,
together with the rays passing through coefficients 1 of the masks.

l(0) l(1)
0 0
0 l(7) 0 l(8)
0.14 0.14
1 1
0.29 0.29
2 l(14) 2
y and m

y and m
0.43 0.43 l(15)
3 3
0.57 0.57
4 4
0.71 l(21) 0.71
5 5 l(22)
0.86 0.86
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

l(2) l(3)
0 0
0 0
0.14 l(9) 0.14
1 1 l(10)
0.29 0.29
2 2
y and m

y and m

0.43 0.43
3 l(16) 3 l(17)
0.57 0.57
4 4
0.71 0.71
5 l(23) 5
0.86 0.86 l(24)
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.12
Four masks [χ1,3,t], t = 0 : 3, and parallel rays for the (1, 3)-projection.

The number of parallel rays of this projection equals (1 + 3)6 + 1 = 25,

and they are defined as
 t 2
l(t) = l1,3 (t) = (x, y); x + 3y = + , t = 0 : 24,
7 7
where (x, y) ∈ [0, 1] × [0, 1]. Figure 5.13 shows these rays, which are numbered
in accordance with the following set of control points:
20
• 1• 2• 3 • 4• 5• 6•
3
6 . . . . 7• 8• 9• 7
. . . . 10 • 11 • 12 • 7
6 7
6
. . . . 13 • 14 • 15 • 7
6 7
6
6 7 (angle of rays is − tan−1 (1/3) = −18.435◦ ).
6
6 . . . . 16 • 17 • 18 • 7
7
19 20 21 5
4 . . . . • • •
. . . . 22 • 23 • 24 •
Image Processing 287

l(4) l(5)
0 0
0 0
0.14 0.14
1 l(11) 1
0.29 0.29 l(12)
2 2
y and m

y and m
0.43 0.43
3 3
0.57 l(18) 0.57
4 4 l(19)
0.71 0.71
5 5
0.86 0.86
6 l(25) 6 l(26)
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

0 l(6) 0
l(0) l(1) l(2) l(3) l(4) l(5)
l(6)
0 0
0.14 0.14 l(8)
1 1 l(10)
0.29 0.29 l(12)
2 l(13) 2
l(14)
y and m

y and m
0.43 0.43
3 3 l(16)
0.57 0.57 l(18)
4 l(20) 4
0.71 0.71 l(20)
5 5 l(22)
0.86 0.86 l(24)
6 6
1 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.13
Three masks [χ1,3,t], t = 4 : 6, and the set of 25 rays for the (1, 3)-projection.

The components of the tensor transform with the triplet-numbers of the

set U (1, 3) are calculated as

f1,3,0 = v0 + v7 + v14 + v21 , f1,3,1 = v1 + v8 + v15 + v22 ,

f1,3,2 = v2 + v9 + v16 + v23 , f1,3,3 = v3 + v10 + v17 + v24 ,
(5.8)
f1,3,4 = v4 + v11 + v18 , f1,3,5 = v5 + v12 + v19 ,
f1,3,6 = v6 + v13 + v20 ,

where variables vt = v1,3 (t) denote the sums of the discrete image along the
corresponding rays l(t), t = 0 : 24. The system of equations can also be written
as
X3 X3
f1,3,t = vt+7m = v1,3,t+7m, t = 0 : 6,
m=0 m=0

where vt = 0, if t > 24.

We now define a set of geometrical rays that allows for expressing the line-
integrals w(t) by means of the sums v(t), t = 0 : 24. For that, we consider a few
arithmetical rays as the geometrical rays in the part of the square [0, 1] ×[0, 1],
which are shown in Figure 5.14. Each of these rays equally intersects image
288 CHAPTER 5: TENSOR TRANSFORM METHOD

elements,√that is, the length of intersection of the ray with each IE equals
∆l0,6 = 10/3∆x. The integrals along the geometrical rays 15, 16, and 17
can be calculated by the line-sums of the discrete image as
√
7 10
w(15)= w1,3 (15)= √3
[v(15) + v(14) + v(16)],
7 10
w(16)= w1,3 (16)= √3
[v(15) + v(16) + v(17)],
7 10
w(17)= w1,3 (17)= 3
[v(18) + v(16) + v(17)].

Similar equations hold for all integrals w(t) of the considered projections:
√
w(t) = w1,3 (t) = 7 10/3[v(t − 1) + v(t) + v(t + 1)], t = 0 : 24.

In matrix form, this system of linear equations is determined by the Toeplitz

l(14)
0.43
l(15)
3 l(16)
l(17)
0.57
l(18)
4
y and m

0.71
5
0.86
6
1
0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n

FIGURE 5.14
Five parallel geometrical rays of the (1, 3)-projection.

matrix 25 × 25 with right bandwidth of 2. This matrix has determinant 0 and

the given set of geometrical rays cannot be used directly to define all required
sums v(t), t = 0 : 24. We consider other sets of geometrical rays. For instance,
this set of 25 geometrical rays can be defined as (see Figure 5.15)
 t 1
l̃(t) = l1,3 (t − 1) = (x, y); x + 3y = + , t = 0 : 24.
7 7

The integrals of the image along these rays are denoted by w(t) = w(˜l(t)).
The corresponding system of equations for integrals w(t) is described by
√
7 10
w(t) = [v(t) + v(t − 1) + v(t − 2)], t = 0 : 24,
3
where it is assumed that v(−1) and v(−2) are 0. This system of equations can
Image Processing 289

˜l(0) ˜l(1) ˜l(2) ˜l(3) ˜l(4) ˜l(5)

˜l(6)
0 ˜l(7)
0 ˜l(8)
˜l(9)
0.14 ˜l(10)
1 ˜l(11)
˜l(12)
0.29 ˜l(13)
2 ˜l(14)
˜l(15)
0.43 ˜l(16)
y and m

3 ˜l(17)
˜l(18)
0.57 ˜l(19)
4 ˜l(20)
˜l(21)
0.71 ˜l(22)
5 ˜l(23)
˜l(24)
0.86

1
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.15
A set of geometrical rays for the (1, 3)-projection.

be written in matrix form


10000· · · 000
1 1 0 0 0 · · · 0 0 0
 
√ √  1 1 1 0 0 · · · 0 0 0
7 10 7 10  0 1 1 1 0 · · · 0 0 0 


w= Av =   v. (5.9)
3 3  .. .. .. .. .. . . .. .. .. 
 . . . . . . . . .
 
0 0 0 0 0 · · · 1 1 0
00000· · · 111
290 CHAPTER 5: TENSOR TRANSFORM METHOD

The above Toeplitz matrix 25 × 25 has the lower triangle inverse matrix
 
1 0 0 0 0 0 0 0 · · · 0 00
 −1 1 0 0 0 0 0 0 · · · 0 00 
 
 0 −1 1 0 0 0 0 0 · · · 0 00 
 
 1 0 −1 1 0 0 0 0 · · · 0 00 
 
 −1 1 0 −1 1 0 0 0 · · · 0 00 
A−1 =  0 −1 1 0 −1 1 0 0 · · · 0 00  .
 
(5.10)
 
 1 0 −1 1 0 −1 1 0 · · · 0 00 
 
 .. .. .. .. .. .. .. .. . . . . .
 . . . . . . . . . .. .. .. 
 
 1 0 −1 1 0 −1 1 0 · · · −1 10 
−1 1 0 −1 1 0 −1 1 · · · 0 −1 1

The inverse transform b = A−1 w can be calculated in the recurrent form

b0 =w(0), b1 = w(1) − b0 ,
b2 =w(2) − (b1 + b0 ), (5.11)
bt =w(t) − (bt−1 + bt−2 ), t = 3, 4, 5, · · · , 24,
√
and v(t) = bt /(7 10/3), where t = 0 : 24.
For the considered image f(x, y), the integrals w(t) are calculated by
w(0) = K(1) =K
w(1) = K(1 + 1) =K ·2
w(2) = K(3 + 1 + 1) =K ·5
w(3) = K(1 + 3 + 1 + 1) =K ·6
w(4) = K(2 + 1 + 3 + 1 + 1 + 0) =K ·8
w(5) = K(2 + 2 + 1 + 1 + 1 + 1) =K ·8
w(6) = K(5 + 2 + 2 + 2 + 1 + 1 + 0) = K · 13
w(7) = K(5 + 2 + 1 + 2 + 1 + 2 + 0) = K · 13
w(8) = K(5 + 1 + 1 + 2 + 2 + 2 + 0) = K · 13
w(9) = K(4 + 1 + 1 + 1 + 2 + 2 + 2) = K · 13
w(10) = K(4 + 1 + 1 + 1 + 2 + 3 + 2) = K · 14
w(11) = K(4 + 1 + 1 + 1 + 4 + 3 + 2) = K · 16
w(12) = K(2 + 1 + 1 + 2 + 4 + 3 + 0) = K · 13
w(13) = K(2 + 1 + 1 + 2 + 4 + 3 + 0) = K · 13
w(14) = K(2 + 1 + 1 + 2 + 4 + 3 + 0) = K · 13
w(15) = K(3 + 1 + 1 + 8 + 4 + 3 + 2) = K · 22
w(16) = K(3 + 1 + 9 + 8 + 4 + 3 + 2) = K · 30
w(17) = K(3 + 2 + 9 + 8 + 1 + 3 + 2) = K · 28
w(18) = K(0 + 2 + 9 + 8 + 1 + 3 + 1) = K · 24
w(19) = K(0 + 2 + 7 + 8 + 1 + 9 + 1) = K · 28
w(20) = K(0 + 6 + 7 + 8 + 2 + 9 + 1) = K · 33
w(21) = K(6 + 6 + 7 + 4 + 2 + 9) =K · 34
w(22) = K(6 + 6 + 3 + 4 + 2) =K · 21
w(23) = K(6 + 1 + 3 + 4) =K · 14
w(24) = K(2 + 1 + 3) =K · 6,
√
where K = 7 10/3.
Image Processing 291

Therefore, the sums v(t) of the discrete image fn,m are calculated by
v(0) = K −1 w(0) =1 =1
v(1) = K −1 w(1) − v(0)
` ´ =2 − 1 =1
v(2) = K −1 w(2) − `v(1) + v(0)´ = 5 − (1 + 1) =3
v(3) = K −1 w(3) − `v(2) + v(1)´ = 6 − (3 + 1) =2
v(4) = K −1 w(4) − `v(3) + v(2)´ = 8 − (2 + 3) =3
v(5) = K −1 w(5) − `v(4) + v(3)´ = 8 − (3 + 2) =3
v(6) = K −1 w(6) − `v(5) + v(4)´ = 13 − (3 + 3) = 7
v(7) = K −1 w(7) − `v(6) + v(5)´ = 13 − (7 + 3) = 3
v(8) = K −1 w(8) − `v(7) + v(6)´ = 13 − (3 + 7) = 3
v(9) = K −1 w(9) − v(8)
` + v(7) ´ = 13 − (3 + 3) = 7
v(10) = K −1 w(10) − `v(9) + v(8) ´ = 14 − (7 + 3) = 4
v(11) = K −1 w(11) − `v(10) + v(9) ´ = 16 − (4 + 7) = 5
v(12) = K −1 w(12) − `v(11) + v(10)´ = 13 − (5 + 4) = 4
v(13) = K −1 w(13) − `v(12) + v(11)´ = 13 − (4 + 5) = 4
v(14) = K −1 w(14) − `v(13) + v(12)´ = 13 − (4 + 4) = 5
v(15) = K −1 w(15) − `v(14) + v(13)´ = 22 − (5 + 4) = 13
v(16) = K −1 w(16) − `v(15) + v(14)´ = 30 − (13 + 5) = 12
v(17) = K −1 w(17) − `v(16) + v(15)´ = 28 − (12 + 13) = 3
v(18) = K −1 w(18) − `v(17) + v(16)´ = 24 − (3 + 12) = 9
v(19) = K −1 w(15) − `v(14) + v(17)´ = 28 − (9 + 3) = 16
v(20) = K −1 w(16) − `v(15) + v(18)´ = 33 − (16 + 9) = 8
v(21) = K −1 w(17) − `v(16) + v(19)´ = 34 − (8 + 16) = 10
v(22) = K −1 w(18) − `v(17) + v(20)´ = 21 − (10 + 8) = 3
v(23) = K −1 w(17) − `v(16) + v(21)´ = 14 − (3 + 10) = 1
v(24) = K −1 w(18) − v(17) + v(22) = 6 − (1 + 3) = 2.

The components of the splitting-signal {f1,3,t} are calculated by

f1,3,0 =v(0) + v(7) + v(14) + v(21) = 1 + 3 + 5 + 10 = 19

f1,3,1 =v(1) + v(8) + v(15) + v(22) = 1 + 3 + 13 + 3 = 20
f1,3,2 =v(2) + v(9) + v(16) + v(23) = 3 + 7 + 12 + 1 = 23
f1,3,3 =v(3) + v(10) + v(17) + v(24) = 2 + 4 + 3 + 2 = 11
f1,3,4 =v(4) + v(11) + v(18) = 3 + 5 + 9 = 17
f1,3,5 =v(5) + v(12) = 3 + 4 + 16 = 23
f1,3,6 =v(6) + v(13) = 7 + 4 + 8 = 19.

Figure 5.16 shows the graphs of the line-integrals w(t), sums v(t), and the
splitting-signal for the (1, 3)-projection.

5.2.6 (1, 4)-projection

We now consider the subset of triplet-numbers

U (1, 4) = (1, 4, 0), (1, 4, 1), (1, 4, 2), (1, 4, 3), (1, 4, 4), (1, 4, 5), (1, 4, 6) .

Masks of the tensor functions which are numbered by these triplets are:
292 CHAPTER 5: TENSOR TRANSFORM METHOD

20 30
250
w1,3(t) v1,3(t) f1,3,t
25
200 15
20
150
10 15
100
10
5
50 5

0 0 0
0 10 20 t 0 10 20 t 0 2 4 6 t

FIGURE 5.16
Line-integrals w(t), sums v(t), and splitting-signal for the (1, 3)-projection.

     
1000000 0 1 0 0 0 00 0 01 0 0 0 0
0 0 0 1 0 0 0   0 0 0 0 1 00   0 00 0 0 1 0 
     
0 0 0 0 0 0 1   1 0 0 0 0 00   0 10 0 0 0 0 
     
 0 0 1 0 0 0 0 , [χ1,4,1] =  0 0 0 1 0 00 , [χ1,4,2] =  0 00 0 1 0 0 ,
[χ1,4,0] =      
0 0 0 0 0 1 0   0 0 0 0 0 01   1 00 0 0 0 0 
     
0 1 0 0 0 0 0   0 0 1 0 0 00   0 00 1 0 0 0 
0000100 0 0 0 0 0 10 0 00 0 0 0 1
     
0001000 0 0 0 0 1 00 0 00 0 0 1 0
0 0 0 0 0 0 1   1 0 0 0 0 00   0 10 0 0 0 0 
     
0 0 1 0 0 0 0   0 0 0 1 0 00   0 00 0 1 0 0 
     
 0 0 0 0 0 1 0 , [χ1,4,4] =  0 0 0 0 0 01 , [χ1,4,5] =  1 00 0 0 0 0 ,
[χ1,4,3] =      
0 1 0 0 0 0 0   0 0 1 0 0 00   0 00 1 0 0 0 
     
0 0 0 0 1 0 0   0 0 0 0 0 10   0 00 0 0 0 1 
1000000 0 1 0 0 0 00 0 01 0 0 0 0
   
0 0 0 00 0 1 1 1 1 1 1 11
 0 0 1 00 0 0   1 1 1 1 1 11 
   
 0 0 0 00 1 0   1 1 1 1 1 11  X 5
   
[χ1,4,6] =   0 1 0 00 0 0 =
  1 1 1 1 1 11 −
 [χ1,4,t].
 0 0 0 01 0 0   1 1 1 1 1 11  t=0
   
 1 0 0 00 0 0   1 1 1 1 1 11 
0 0 0 10 0 0 1 1 1 1 1 11

The set of all (p + s)(N − 1) + 1 = 31 arithmetical rays of the projection

by the angle arctg(4) are shown in Figure 5.17.
To reduce the large number of rays, we consider another direction, namely,
the (1, −3)-projection, which defines the same set of masks of tensor functions.
This projection is defined as (1, s − N ) = (1, 4 − 7). Masks of the tensor
Image Processing 293

l(0) l(1) l(2) l(3) l(4)

0 l(5)
l(6)
0 l(7)
l(8)
0.142 l(9)
l(10)
1 l(11)
l(12)
0.285 l(13)
l(14)
2 l(15)
l(16)
0.428 l(17)
y and m

l(18)
3 l(19)
l(20)
0.571 l(21)
l(22)
4 l(23)
l(24)
0.714 l(25)
l(26)
5 l(27)
l(28)
0.857 l(29)
l(30)
6
1
0 1 2 3 4 5 6

0 0.142 0.285 0.428 0.571 0.714 0.857 1

x and n

FIGURE 5.17
The set of 25 parallel rays for the (1, 4)-projection.

functions with the first three triplet-numbers of the subset U (1, −3) are
     
1 0 0 00 0 0 0100000 0010000
 0 0 0 10 0 0  0 0 0 0 1 0 0   0 0 0 0 0 1 0
     
 0 0 0 00 0 1  1 0 0 0 0 0 0   0 1 0 0 0 0 0
     
 0 0 1 00 0 0 , [χ1,−3,1] =  0 0 0 1 0 0 0 , [χ1,−3,2] =  0 0 0 0 1 0 0 .
[χ1,−3,0] =      
 0 0 0 00 1 0  0 0 0 0 0 0 1   1 0 0 0 0 0 0
     
 0 1 0 00 0 0  0 0 1 0 0 0 0   0 0 0 1 0 0 0
0 0 0 01 0 0 0000010 0000001

As an example, Figure 5.18 shows the mask of the tensor function χ1,4,0
with the parallel rays of two different projections. Five rays of the (1, 4)-
projection are shown on the left, and three rays of the (1, −3)-projection are
shown on the right.
For the remaining four triplet-numbers, we also have the equality of masks,
[χ1,−3,t] = [χ1,4,t], t = 3, 4, 5, 6.
We can see that for the subsets U (1, 4) and U (1, −3), we have the same
masks of the tensor functions. The number of parallel rays for (1, −3)-
projection equals (1 + 3)6 + 1 = 25, which is less than 31. The set of all
rays for this projection is shown in Figure 5.19.
294 CHAPTER 5: TENSOR TRANSFORM METHOD

l(0)
0 0
0 l(7) 0
0.14 0.14
1 1
0.28 0.28
2 l(14) 2 l(0)
0.42 0.42
y and m

y and m
3 3
0.57 l(21) 0.57
4 4
0.71 0.71 l(−7)
5 5
0.85 l(28) 0.85
6 6
1 1
l(−14)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.28 0.42 0.57 0.71 0.85 1 0 0.14 0.28 0.42 0.57 0.71 0.85 1
x and n x and n

FIGURE 5.18
The mask [χ1,4,0] with arithmetical rays of the (1, 4)- and (1, −3)-projections.

0
0
l(6)
0.142 l(5)
1 l(4)
l(3)
0.285 l(2)
l(1)
2
l(0)
0.428 l(−1)
y and m

l(−2)
3
l(−3)
0.571 l(−4)
l(−5)
4
l(−6)
0.714 l(−7)
l(−8)
5 l(−9)
0.857 l(−10)
l(−11)
6 l(−12)
1 l(−13)
l(−18) l(−17) l(−16) l(−15) l(−14)
0 1 2 3 4 5 6
0 0.142 0.285 0.428 0.571 0.714 0.857 1
x and n

FIGURE 5.19
The set of parallel rays for the (1, −3)-projection.

The control points of this set of rays are defined as

0
• 1 • 2• 3• 4 • 5• 6•
2 3
6 −3 • −2 • −1 • . . . . 7
6 −6 −5 −4 7
6 −9 • −8 • −7 • . . . . 7
6 7
6 • •
6 −12 −11 −10 • . . . . 7
7 (angle of rays is arctan(1/3) = 18.435◦ ).
6 • •
6 −15 −14 −13 • . . . . 7
7
4 • • • . . . .5
−18 −17 −16
• • • . . . .
Image Processing 295

The set of the 25 geometrical rays of this projection is described by

n t 1o
l(t) = l1,−3 (t) = (x, y); x − 3y = − , t = 6 : −1 : −18, (5.12)
7 7
where (x, y) ∈ [0, 1] × [0, 1]. Six masks of the tensor functions χ1,4,t, t = 1 : 6,
are shown in Figures 5.20 and 5.21. The components of the tensor transform

0 0
0 0
0.14 0.14
1 1
0.29 0.29 l(2)
l(1)
2 2
y and m

y and m
0.43 0.43
3 3
0.57 0.57 l(−5)
4 l(−6) 4
0.71 0.71
5 5
0.86 0.86
6 6 l(−12)
1 l(−13) 1
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

0 0
0 0
0.14 0.14 l(4)
1 l(3) 1
0.29 0.29
2 2
y and m

y and m

0.43 0.43
3 3 l(−3)
0.57 l(−4) 0.57
4 4
0.71 0.71
5 5
0.86 0.86 l(−10)
l(−11)
6 6
1 l(−18) 1 l(−17)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.20
Masks [χ1,4,t], t = 1 : 4, and the rays of the (1, −3)-projection.

with the triplet-numbers (1, 4, t) can be calculated by

f1,4,0 =v0 + v−7 + v−14 , f1,4,1 = v1 + v−6 + v−13 ,

f1,4,2 =v2 + v−5 + v−12 , f1,4,3 = v3 + v−4 + v−11 + v−18 ,
(5.13)
f1,4,4 =v4 + v−3 + v−10 + v−17 , f1,4,5 = v5 + v−2 + v−9 + v−16 ,
f1,4,6 =v6 + v−1 + v−8 + v−15 ,

where the sums of the discrete image along the rays are denoted by vt =
v(t) = v1,−3 (t), t = 6 : −1 : −18. This system of linear equations can also be
written as follows:
X X
f1,4,t = vt−7m = v1,−3,t−7m, t = 0 : 6.
m m
296 CHAPTER 5: TENSOR TRANSFORM METHOD

0 0
0 0 l(6)
0.14 l(5) 0.14
1 1
0.29 0.29
2 2
l(−1)
y and m

y and m
0.43 l(−2) 0.43
3 3
0.57 0.57
4 4
0.71 0.71 l(−8)
5 l(−9) 5
0.86 0.86
6 6
1 l(−16) 1 l(−15)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.21
Two masks [χ1,4,t], t = 5, 6, and the rays of the (1, −3)-projection.

As in the (1, 3) case, we consider the shifted set of the rays, l(t) → l(t − 1),
as the set of 25 geometrical rays
n t 2o
˜
l(t) = l1,−3 (t − 1) = (x, y); x − 3y = − , t = 6 : −1 : −18,
7 7
which are shown in Figure 5.22.
For the line-integrals, we obtain the system of equations, which is similar
to the system described for the (1, 3)-projection,
√
7 10 h i
w(t) = w(˜ l(t)) = v(t) + v(t − 1) + v(t − 2) , t = 6 : −1 : −18, (5.14)
3
where v(−19) = v(−20) = 0. The required sums √ v(t) of the discrete image
fn,m can therefore be defined as v(t) = 3bt /(7 10), where the components bt
are calculated recursively by
b−18 =w(−18), b−17 = w(−17) − b−18 ,
b−16 =w(−16) − (b−17 + b−18 ), (5.15)
bt =w(t) − (bt−1 + bt−2 ), t = −15, −14, · · · , 6.
The line-integrals w(t) for the given image f(x, y) can be calculated by
w(6) = K(5 + 2 + 2) =K ·9
w(5) = K(4 + 2 + 2 + 1) =K ·9
w(4) = K(4 + 1 + 2 + 1 + 3) =K · 11
w(3) = K(4 + 1 + 1 + 1 + 3 + 1) =K · 11
w(2) = K(2 + 1 + 1 + 2 + 3 + 1 + 1) = K · 11
w(1) = K(2 + 1 + 1 + 2 + 1 + 1 + 1) = K ·9
w(0) = K(2 + 1 + 1 + 2 + 1 + 1 + 1) = K ·9
w(−1) = K(3 + 1 + 1 + 1 + 1 + 1 + 1) = K ·9
w(−2) = K(3 + 1 + 1 + 1 + 2 + 1 + 1) = K · 10
w(−3) = K(3 + 1 + 1 + 1 + 2 + 2 + 1) = K · 11
w(−4) = K(0 + 1 + 1 + 2 + 2 + 2 + 0) = K ·8
w(−5) = K(0 + 2 + 1 + 2 + 4 + 2 + 0) = K · 11
w(−6) = K(0 + 2 + 9 + 2 + 4 + 3 + 0) = K · 20
Image Processing 297

0.14 ˜l(6)
˜l(5)
1 ˜l(4)
0.29 ˜l(3)
˜l(2)
2 ˜l(1)
0.43 ˜l(0)
˜l(−1)
y and m

3 ˜l(−2)
0.57 ˜l(−3)
˜l(−4)
4 ˜l(−5)
0.71 ˜l(−6)
˜l(−7)
5 ˜l(−8)
0.86 ˜l(−9)
˜l(−10)
6 ˜l(−11)
1 ˜l(−12)
˜l(−18) ˜l(−17) ˜l(−16) ˜l(−15) ˜l(−14) ˜l(−13)
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.22
The set of geometrical rays for the (1, −3)-projection.

w(−7) = K(6 + 2 + 9 + 8 + 4 + 3 + 2) = K · 34
w(−8) = K(6 + 6 + 9 + 8 + 4 + 3 + 2) = K · 38
w(−9) = K(6 + 6 + 7 + 8 + 4 + 3 + 2) = K · 36
w(−10) = K(2 + 6 + 7 + 8 + 4 + 3 + 0) = K · 30
w(−11) = K(2 + 1 + 7 + 8 + 1 + 3 + 0) = K · 22
w(−12) = K(2 + 1 + 3 + 8 + 1 + 3 + 0) = K · 18
w(−13) = K(1 + 3 + 4 + 1 + 3 + 2) = K · 14
w(−14) = K(3 + 4 + 2 + 3 + 2) = K · 14
w(−15) = K(4 + 2 + 9 + 2) = K · 17
w(−16) = K(2 + 9 + 1) = K · 12
w(−17) = K(9 + 1) = K · 10
w(−18) = K(1) =K · 1
√
where K = 7 10/3.
The sums v(t) of the discrete image fn,m are calculated as follows:
v(−18) = K −1w(−18) =1 =1
v(−17) = K −1w(−17) − v(−18) = 10 − 1 =9
v(−16) = K −1w(−16) − [v(−17) + v(−18)] = 12 − [9 + 1] = 2
v(−15) = K −1w(−15) − [v(−16) + v(−17)] = 17 − [2 + 9] = 6
v(−14) = K −1w(−14) − [v(−15) + v(−16)] = 14 − [6 + 2] = 6
v(−13) = K −1w(−13) − [v(−14) + v(−15)] = 14 − [6 + 6] = 2
298 CHAPTER 5: TENSOR TRANSFORM METHOD

v(−12) = K −1w(−12) − [v(−13) + v(−14)] = 18 − [2 + 6] = 10

v(−11) = K −1w(−11) − [v(−12) + v(−13)] = 22 − [10 + 2] = 10
v(−10) = K −1w(−10) − [v(−11) + v(−12)] = 30 − [10 + 10] = 10
v(−9) = K −1w(−9) − [v(−10) + v(−11)] = 36 − [10 + 10] = 16
v(−8) = K −1w(−8) − [v(−9) + v(−10)] = 38 − [16 + 10] = 12
v(−7) = K −1w(−7) − [v(−8) + v(−9)] = 34 − [12 + 16] = 6
v(−6) = K −1w(−6) − [v(−7) + v(−8)] = 20 − [6 + 12] = 2
v(−5) = K −1w(−5) − [v(−6) + v(−7)] = 11 − [2 + 6] = 3
v(−4) = K −1w(−4) − [v(−5) + v(−6)] = 8 − [3 + 2] =3
v(−3) = K −1w(−3) − [v(−4) + v(−5)] = 11 − [3 + 3] = 5
v(−2) = K −1w(−2) − [v(−3) + v(−4)] = 10 − [5 + 3] = 2
v(−1) = K −1w(−1) − [v(−2) + v(−3)] = 9 − [2 + 5] =2
v(0) = K −1w(0) − [v(−1) + v(−2)] = 9 − [2 + 2] =5
v(1) = K −1w(1) − [v(0) + v(−1)] = 9 − [5 + 2] =2
v(2) = K −1w(2) − [v(1) + v(0)] = 11 − [2 + 5] = 4
v(3) = K −1w(3) − [v(2) + v(1)] = 11 − [4 + 2] = 5
v(4) = K −1w(4) − [v(3) + v(2)] = 11 − [5 + 4] = 2
v(5) = K −1w(5) − [v(4) + v(3)] = 9 − [2 + 5] =2
v(6) = K −1w(6) − [v(5) + v(4)] = 9 − [2 + 2] = 5.

The splitting-signal {f1,4,t } is calculated by

f1,4,0 = v(0) + v(−7) + v(−14) =5 + 6 + 6 = 17

f1,4,1 = v(1) + v(−6) + v(−13) =2 + 2 + 2 = 6
f1,4,2 = v(2) + v(−5) + v(−12) =4 + 3 + 10 = 17
f1,4,3 = v(3) + v(−4) + v(−11) + v(−18) =5 + 3 + 10 + 1 = 19
f1,4,4 = v(4) + v(−3) + v(−10) + v(−17) =2 + 5 + 10 + 9 = 26
f1,4,5 = v(5) + v(−2) + v(−9) + v(−16) =2 + 2 + 16 + 2 = 22
f1,4,6 = v(6) + v(−1) + v(−8) + v(−15) =5 + 2 + 12 + 6 = 25.

Figure 5.23 shows the graphs of the integrals w(t), sums v(t), and splitting-
signal for the (1, −3)-projection.

300 20 30
w1,4(t) v1,4(t) f1,4,t
250 25
15
200 20

150 10 15

100 10
5
50 5

0 0 0
6 1 −4 −9 −14 t 6 1 −4 −9 −14 t 0 2 4 6 t

FIGURE 5.23
Line-integrals w(t), sums v(t), and splitting-signal of the (1, −3)-projection.
Image Processing 299

Remark 5.1 To work with only nonnegative numbers for control points, we
consider the transformation t → 6 − t,
 6 5 4 3 2 1 0 
• • • • • • •
 9• 8• 7• . . . . 
 12 11 10 
 • • • . . . .
 15 14 13 
 • • • . . . .  (angle of rays is tan−1 (1/3) = 18.435◦).
 18 17 16 
 • • • . . . .
 21 20 19 
 • • • . . . .
24 23 22
• • • . . . .

With such numbering of the control points, the set of 25 geometrical rays of
the (1, −3)-projection in (5.12) can be written as
 t 5
l(t) = l1,−3 (6 − t) = (x, y); x − 3y = − + .
7 7
Therefore, the system of equations in (5.14) can be written as
h i
w(t) = w(˜ l(t)) = K v(t) + v(t + 1) + v(t + 2) , t = 0 : 24, (5.16)

where v(25) = v(26) = 0. The system of equations in (5.15) for the coefficients
bt = Kv(t) is described by

b24 = w(24),
b23 = w(23) − b24 ,
(5.17)
b22 = w(22) − (b23 + b24 ),
bt = w(t) − (bt+1 + bt+2 ), t = 21, 20, · · · , 1, 0.

5.2.7 (1, 5)-projection

Consider the subset of triplets

U (1, 5) = (1, 5, 0), (1, 5, 1), (1, 5, 2), (1, 5, 3), (1, 5, 4), (1, 5, 5), (1, 5, 6) .

Seven masks of the tensor functions with triplet-numbers of U (1, 5) are

     
1000000 0100000 0 0 1 00 0 0
 0 0 1 0 0 0 0 0 0 0 1 0 0 0   0 0 0 01 0 0 
     
 0 0 0 0 1 0 0 0 0 0 0 0 1 0   0 0 0 00 0 1 
     
 0 0 0 0 0 0 1 , [χ1,5,1] =  1 0 0 0 0 0 0 , [χ1,5,2] =  0 1 0 00 0 0 ,
[χ1,5,0] =      
 0 1 0 0 0 1 0 0 0 1 0 0 0 0   0 0 0 10 0 0 
     
 0 0 0 1 0 0 0 0 0 0 0 1 0 0   0 0 0 00 1 0 
0000010 0000001 1 0 0 00 0 0
300 CHAPTER 5: TENSOR TRANSFORM METHOD
     
0001000 0000100 0 0 0 00 1 0
 0 0 0 0 0 1 0 0 0 0 0 0 0 1   1 0 0 00 0 0 
     
 1 0 0 0 0 0 0 0 1 0 0 0 0 0   0 0 1 00 0 0 
     
 0 0 1 0 0 1 0 , [χ1,5,4] =  0 0 0 1 0 0 0 , [χ1,5,5] =  0 0 0 01 0 0 ,
[χ1,5,3] =      
 0 0 0 0 1 0 0 0 0 0 0 0 1 0   0 0 0 00 0 1 
     
 0 0 0 0 0 0 1 1 0 0 0 0 0 0   0 1 0 00 0 0 
0100000 0010000 0 0 0 10 0 0
   
0000001 1111111
 0 1 0 0 0 0 0  1 1 1 1 1 1 1 
   
 0 0 0 1 0 0 0  1 1 1 1 1 1 1  5
    X
[χ1,5,6] =  0 0 0 0 0 1 0  = 1 1 1 1 1 1 1  − [χ1,5,t].
   
 1 0 0 0 0 0 0   1 1 1 1 1 1 1  t=0
   
 0 0 1 0 0 0 0  1 1 1 1 1 1 1 
0000100 1111111

Because (1, 5 − 7) = (1, −2), the functions χ1,5,t equal χ1,−2,t , for t =
0 : 6. Therefore, to reduce the number of rays from (1 + 5)6 + 1 = 37 to
(1+|−2|)6+1 = 19, we consider the direction of the rays, which is determined
by the (1, −2)-projection. As an example, Figure 5.24 shows the mask of the
tensor function χ1,5,0 with the parallel rays of two different projections. Six

l(0)
0 0
0 l(7) 0
0.14 0.14
1 1
0.28 l(14) 0.28
2 2
0.42 0.42
y and m

y and m

3 l(21) 3
0.57 0.57 l(0)
4 4
l(28)
0.71 0.71
5 5
0.85 l(35) 0.85
6 6
1 1
l(−7)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.28 0.42 0.57 0.71 0.85 1 0 0.14 0.28 0.42 0.57 0.71 0.85 1
x and n x and n

FIGURE 5.24
The mask [χ1,5,0] with two different sets of parallel rays.

rays of the (1, 5)-projection are shown on the left mask, and two rays of the
(1, −2)-projection are shown on the right mask. The set of parallel rays for
the (1, −2)-projection is shown in Figure 5.25.
The control points of parallel rays of the (1, 5)-projection are numbered
Image Processing 301

0
˜l(6)
0.14
˜l(5)
1
˜l(4)
0.29
˜l(3)
2
˜l(2)
0.43
˜l(1)
y and m

3
˜l(0)
0.57
˜l(−1)
4
˜l(−2)
0.71
˜l(−3)
5
˜l(−4)
0.86
˜l(−5)
6
˜l(−6)
1
˜l(−12) ˜l(−11) ˜l(−10) ˜l(−9) ˜l(−8) ˜l(−7)
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.25
The set of parallel rays for the (1, −2)-projection.

and located as
0
• 1 • 2• 3• 4 • 5• 6•
2 3
6 −2 • −1 • . . . . . 7
6 −4 −3 7
6 −6 • −5 • . . . . . 7
6 7
6 •
6 −8 −7 • . . . . . 7
7 (angle of rays is arctan(1/2) = 26.5651◦ ).
6 −10 • −9 • . . . . . 7
6 7
4 • • . . . . .5
−12 −11
• • . . . . .

The set of 19 geometrical rays of this projection is described by

n t 1o
l(t) = l1,−2 (t) = (x, y); x − 2y = − , t = 6 : −1 : −12, (5.18)
7 14
where (x, y) ∈ [0, 1] × [0, 1]. Six masks [χ1,5,t], t = 1 : 6, of the tensor functions
are shown in Figures 5.26 and 5.27.
The components of the tensor transform with the triplet-numbers of
U (1, 5) can be calculated by

f1,5,0 =v0 + v−7 , f1,5,1 = v1 + v−6 , f1,5,2 = v2 + v−5 + v−12 ,

f1,5,3 =v3 + v−4 + v−11 , f1,5,4 = v4 + v−3 + v−10 , (5.19)
f1,5,5 =v5 + v−2 + v−9 , f1,5,6 = v6 + v−1 + v−8 ,

where the sums of the discrete image along the rays are denoted by vt =
v(t) = v1,−2 (t), t = 6 : −1 : −12. The above system of linear equations can be
302 CHAPTER 5: TENSOR TRANSFORM METHOD

0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 2 l(2)
y and m

y and m
0.43 l(1) 0.43
3 3
0.57 0.57
4 4
0.71 0.71
5 5
0.86 0.86 l(−5)
6 l(−6) 6
1 1
0 1 2 3 4 5 6 0 1 l(−12)
2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

0 0
0 0
0.14 0.14
1 1 l(4)
0.29 l(3) 0.29
2 2
y and m

y and m
0.43 0.43
3 3
0.57 0.57
4 4
0.71 0.71 l(−3)
5 l(−4) 5
0.86 0.86
6 6
1 1
0 1 2 l(−11)
3 4 5 6 0 1 2 3 l(−10)
4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.26
Masks [χ1,5,t], t = 1 : 4, and parallel rays of the (1, −2)-projection.

0 0
0 0 l(6)
0.14 l(5) 0.14
1 1
0.29 0.29
2 2
y and m

y and m

0.43 0.43
3 3
0.57 0.57 l(−1)
4 l(−2) 4
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 3 4 l(−9)
5 6 0 1 2 3 4 5 l(−8)
6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.27
Masks [χ1,5,t], t = 5, 6, and parallel rays of the (1, −2)-projection.

written as
3
X 3
X
f1,5,t = vt−7m = v1,−2,t−7m, t = 0 : 6,
m=0 m=0

where v−13 = v−14 = 0. As in the (1, 2)-projection case, we consider the

Image Processing 303

shifted set of the rays, l(t) → l(t − 0.5), to be the set of 19 geometrical rays
 
˜ t 1
l(t) = l1,−2 (t − 0.5) = (x, y); x − 2y = − , t = 6 : −1 : −12.
7 7
These rays are shown in Figure 5.28. For this set of geometrical rays, the

0
0
0.142 l(6)
1 l(5)
0.285 l(4)
2 l(3)
0.428 l(2)
y and m

3 l(1)
0.571 ˜
l(0)
4 ˜
l(−1)
0.714 ˜
l(−2)
5 ˜
l(−3)
0.857 ˜
l(−4)
6 ˜
l(−5)
1 ˜
l(−6)
l(−12) l(−11) l(−10) l(−9) l(−8) l(−7)
0 1 2 3 4 5 6
0 0.142 0.285 0.428 0.571 0.714 0.857 1
x and n

FIGURE 5.28
The set of geometrical rays for the (1, −2)-projection.

system of equations for the line-integrals is described by

√
˜ 7 5h i
w(t) = w(l(t)) = v(t) + v(t − 1) , t = 6 : −1 : −12, (5.20)
2
where v(−13) = 0. The required√sums v(t) of the discrete image fn,m are
therefore defined as v(t) = bt /(7 5/2), where components bt are calculated
recursively by
b−12 = w(−12), b−11 = w(−11) − b−12 , b−10 = w(−10) − b−11
(5.21)
bt = w(t) − bt−1 , t = −9, −8, −7, −6, · · · , 6.
For the considered image f(x, y), the line-integrals w(t) of the (1, −2)-
projection are calculated by

w(6) = K(5 + 2) =K ·7
w(5) = K(4 + 2 + 2) =K ·8
w(4) = K(4 + 1 + 2 + 1) =K ·8
w(3) = K(2 + 1 + 1 + 1 + 3) =K ·8
w(2) = K(2 + 1 + 1 + 2 + 3 + 1) =K · 10
w(1) = K(3 + 1 + 1 + 2 + 1 + 1 + 1) = K · 10
w(0) = K(3 + 1 + 1 + 1 + 1 + 1 + 1) = K ·9
304 CHAPTER 5: TENSOR TRANSFORM METHOD

w(−1) = K(0 + 1 + 1 + 1 + 2 + 1 + 1) = K · 7
w(−2) = K(0 + 2 + 1 + 2 + 2 + 2 + 1) = K · 10
w(−3) = K(6 + 2 + 9 + 2 + 4 + 2 + 0) = K · 25
w(−4) = K(6 + 6 + 9 + 8 + 4 + 3 + 0) = K · 36
w(−5) = K(2 + 6 + 7 + 8 + 4 + 3 + 2) = K · 32
w(−6) = K(2 + 1 + 7 + 8 + 4 + 3 + 2) = K · 27
w(−7) = K(1 + 3 + 8 + 1 + 3 + 0) = K · 16
w(−8) = K(3 + 4 + 1 + 3 + 0) = K · 11
w(−9) = K(4 + 2 + 3 + 2) = K · 11
w(−10) = K(2 + 9 + 2) = K · 13
w(−11) = K(9 + 1) = K · 10
w(−12) = K(1) =K · 1
√
where K = 7 5/2. The sums v(t) of the image fn,m are calculated by
v(−12) = K −1 w(−12) =1 =1
v(−11) = K −1 w(−11) − v(−12) = 10 − 1 = 9
v(−10) = K −1 w(−10) − v(−11) = 13 − 9 = 4
v(−9) = K −1 w(−9) − v(−10) = 11 − 4 = 7
v(−8) = K −1 w(−8) − v(−9) = 11 − 7 = 4
v(−7) = K −1 w(−7) − v(−8) = 16 − 4 = 12
v(−6) = K −1 w(−6) − v(−7) = 27 − 12 = 15
v(−5) = K −1 w(−5) − v(−6) = 32 − 15 = 17
v(−4) = K −1 w(−4) − v(−5) = 36 − 17 = 19
v(−3) = K −1 w(−3) − v(−4) = 25 − 19 = 6
v(−2) = K −1 w(−2) − v(−3) = 10 − 6 = 4
v(−1) = K −1 w(−1) − v(−2) = 7 − 4 = 3
v(0) = K −1 w(0) − v(−1) =9 − 3 =6
v(1) = K −1 w(1) − v(0) = 10 − 6 = 4
v(2) = K −1 w(2) − v(1) = 10 − 4 = 6
v(3) = K −1 w(3) − v(2) =8 − 6 =2
v(4) = K −1 w(4) − v(3) =8 − 2 =6
v(5) = K −1 w(5) − v(4) =8 − 6 =2
v(6) = K −1 w(6) − v(5) = 7 − 2 = 5.
The components of the splitting-signal {f1,5,t } are calculated by

f1,5,0 =v(0) + v(−7) = 6 + 12 = 18

f1,5,1 =v(1) + v(−6) = 4 + 15 = 19
f1,5,2 =v(2) + v(−5) + v(−12) = 6 + 17 + 1 = 24
f1,5,3 =v(3) + v(−4) + v(−11) = 2 + 19 + 9 = 30
f1,5,4 =v(4) + v(−3) + v(−10) = 6 + 6 + 4 = 16
f1,5,5 =v(5) + v(−2) + v(−9) = 2 + 4 + 7 = 13
f1,5,6 =v(6) + v(−1) + v(−8) = 5 + 3 + 4 = 12.

Figure 5.29 shows the graphs of the line-integrals w(t), sums v(t), and the
splitting-signal for the (1, −2)-projection.
To avoid the negative numbering of rays, or control points, we consider
Image Processing 305

−1 20
w1,5(t) v1,5(t) 30
f1,5,t
6
15 25
−16
20
−11 10 15

−6 10
5
−1 5

6 0 0
6 −1 −6 −11 t 6 −1 −6 −11 t 0 2 4 6 t

FIGURE 5.29
Line-integrals w(t), sums v(t), and splitting-signal of the (1, 5)-projection.

the transformation t → 6 − t,
6
• 5 • 4• 3• 2 • 1• 0•
2 3
6 8• 7• . . . . . 7
6 10 9 7
6 • • . . . . .7
6 12 11 7
6 • • . . . . .7
6 14 13 7 (angle of rays is tan−1 (1/2) = 26.5651◦ ).
6 • • . . . . .7
6 16 15 7
4 • • . . . . .5
18 17
• • . . . . .

The set of 19 geometrical rays of the projection in (5.18) can be written

as n 1 11 o
l(t) = l1,−2 (6 − t) = (x, y); x − 2y = − + .
7 14
Then, the system of equations in (5.20) is written in the form of
h i
w(t) = w(˜l(t)) = K v(t) + v(t + 1) , t = 0 : 18, (5.22)

where v(19) = 0. The system of equations (5.21) for calculating the coefficients
bt = Kv(t) is described by
b18 = w(18),
b17 = w(17) − b18 ,
(5.23)
b16 = w(16) − b17 ,
bt = w(t) − bt+1 , t = 15, 14, · · · , 0,

and the equations for sums v(t) are written as

v(18) =K −1 w(18),
v(t) =K −1 w(t) − v(t + 1), t = 17 : −1 : 0.

This numbering was considered when plotting the integrals w(t) and sums
v(t) in Figure 5.29. Data {w(6), w(5), w(4), ..., w(−11), w(−12)} and {v(6),
v(5), v(4), ..., v(−11), v(−12)} were plotted versus time-points 0, 1, 2, ..., 17, 18.
306 CHAPTER 5: TENSOR TRANSFORM METHOD

5.2.8 (1, 6)-projection

We consider the subset of triplet-numbers corresponding to the generator
(1, 6),

U (1, 6) = (1, 6, 0), (1, 6, 1), (1, 6, 2), (1, 6, 3), (1, 6, 4), (1, 6, 5), (1, 6, 6) .

Masks of the tensor functions χ1,6,t with triplet-numbers of this subset are
     
1 0 0 00 0 0 0100000 0010000
 0 1 0 00 0 0  0 0 1 0 0 0 0   0 0 0 1 0 0 0
     
 0 0 1 00 0 0  0 0 0 1 0 0 0   0 0 0 0 1 0 0
     
 0 0 0 10 0 0 , [χ1,6,1] =  0 0 0 0 1 0 0 , [χ1,6,2] =  0 0 0 0 0 1 0 ,
[χ1,6,0] =      
 0 0 0 01 0 0  0 0 0 0 0 1 0   0 0 0 0 0 0 1
     
 0 0 0 00 1 0  0 0 0 0 0 0 1   1 0 0 0 0 0 0
0 0 0 00 0 1 1000000 0100000
     
0 0 0 10 0 0 0000100 0000010
 0 0 0 01 0 0  0 0 0 0 0 1 0   0 0 0 0 0 0 1
     
 0 0 0 00 1 0  0 0 0 0 0 0 1   1 0 0 0 0 0 0
     
 0 0 0 00 0 1 , [χ1,6,4] =  1 0 0 0 0 0 0 , [χ1,6,5] =  0 1 0 0 0 0 0 ,
[χ1,6,3] =      
 1 0 0 00 0 0  0 1 0 0 0 0 0   0 0 1 0 0 0 0
     
 0 1 0 00 0 0  0 0 1 0 0 0 0   0 0 0 1 0 0 0
0 0 1 00 0 0 0001000 0000100
   
0000001 1111111
 1 0 0 0 0 0 0  1 1 1 1 1 1 1 
   
 0 1 0 0 0 0 0  1 1 1 1 1 1 1  5
    X
[χ1,6,6] =   0 0 1 0 0 0 0  = 1 1 1 1 1 1 1  −
   [χ1,6,t].
 0 0 0 1 0 0 0   1 1 1 1 1 1 1  t=0
   
 0 0 0 0 1 0 0  1 1 1 1 1 1 1 
0000010 1111111
To reduce the number of parallel rays, along which the corresponding compo-
nents f1,6,t of the tensor transform are defined (see Figure 5.30), we consider
the parallel rays of the (1, −1)-projection. The masks of the tensor functions
for these two projections are the same, i.e., [χ1,6,t] = [χ1,−1,t], for t = 0 : 6.
The use of the (1, −1)-projection allows for reducing the number of rays from
(1 + 6)6 + 1 = 43 to (1 + | − 1|)6 + 1 = 13.
As an example, Figure 5.31 shows the mask of the tensor function χ1,5,0
with the parallel arithmetical rays of two different projections. Seven rays of
the (1, 6)-projection are shown in the mask on the left, and one ray of the
(1, −1)-projection is shown in the same mask on the right. The set of 13 rays
for the (1, −1)-projection is shown in Figure 5.32.
Image Processing 307

l(0) l(1) l(2) l(3)

0
l(6)
0

0.14 l(10)

1
l(14)
0.29
l(18)
2

0.43 l(22)
y and m

3
l(26)
0.57
l(30)
4

0.71 l(34)

5
l(38)
0.86
l(42)
6

1
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.30
The set of parallel rays for the (1, 6)-projection.

l(0)
0 0
0 l(7) 0
0.14 0.14
1 l(14) 1
0.29 0.29
2 2
0.43 l(21) 0.43
y and m

y and m

3 3
0.57 l(28) 0.57
4 4
0.71 l(35) 0.71
5 5
0.86 0.86
6 l(42) 6
1 1 l(0)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.31
The mask |χ01,6,0| with two different sets of arithmetical rays.

The control points of this set of rays are defined as

 0 1 2 3 4 5 6 
• • • • • • •
 −1 • . . . . . . 
 −2 
 • . . . . . .
 −3 
 • . . . . . .  (angle of rays is tan−1 (1) = 45◦).
 −4 
 • . . . . . .
 −5 
 • . . . . . .
−6
• . . . . . .
308 CHAPTER 5: TENSOR TRANSFORM METHOD

0.14 l(6)

0.29 l(5)

0.43 l(4)
y and m

0.57 l(3)

0.71 l(2)

0.86 l(1)

1 l(0)
l(−6) l(−5) l(−4) l(−3) l(−2) l(−1)
0 1 2 3 4 5 6

0 0.14 0.29 0.43 0.57 0.71 0.86 1

x and n

FIGURE 5.32
The set of parallel rays for the (1, −1)-projection.

The set of 13 parallel rays of this projection is described by

n to
l(t) = l1,−1 (t) = (x, y); x − y = , t = 6 : −1 : −6, (5.24)
7
where (x, y) ∈ [0, 1] × [0, 1]. Six masks for the tensor functions χ1,6,t, t = 1 : 6,
are shown in Figures 5.33 and 5.34.
The components of the tensor transform with the triplet-numbers (1, 6, t)
can be calculated by

f1,6,0 =v0 ,
f1,6,1 =v1 + v−6 , f1,6,2 = v2 + v−5 ,
(5.25)
f1,6,3 =v3 + v−4 , f1,6,4 = v4 + v−3 ,
f1,6,5 =v5 + v−2 , f1,6,6 = v6 + v−1 ,

where the sums of the discrete image along the rays are denoted by vt =
v(t) = v1,−1 (t), t = 6 : −1 : −6. This complete system of linear equations for
calculating the components of the 2-D tensor transform with triplet-numbers
of the subset U (1, 6) can be written as

f1,6,t = vt + vt−7 = v1,−1,t + v1,−1,t−7, t = 0 : 6,

where v−7 = 0. As in the (1, 1)-projection case, the line-sums v(t) of the
discrete image fn,m can be calculated from the line-integrals w(t). The set of
geometrical rays coincides with the set of arithmetical rays. Each ray intersects
Image Processing 309

0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 2
y and m

y and m
0.43 0.43
3 3
0.57 0.57
4 4
0.71 0.71 l(2)
5 5
0.86 l(1) 0.86
6 6
1 1
0 l(−6)
1 2 3 4 5 6 0 1 l(−5)
2 3 4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

0 0
0 0
0.14 0.14
1 1
0.29 0.29
2 2
y and m

y and m
0.43 0.43 l(4)
3 3
0.57 l(3) 0.57
4 4
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 l(−4)
3 4 5 6 0 1 2 3 l(−3)
4 5 6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.33
Masks [χ1,6,t], t = 1 : 4, and rays of the (1, −1)-projection.

0 0
0 0
0.14 0.14 l(6)
1 1
0.29 l(5) 0.29
2 2
y and m

y and m

0.43 0.43
3 3
0.57 0.57
4 4
0.71 0.71
5 5
0.86 0.86
6 6
1 1
0 1 2 3 4 l(−2)
5 6 0 1 2 3 4 5 l(−1)
6
0 0.14 0.29 0.43 0.57 0.71 0.86 1 0 0.14 0.29 0.43 0.57 0.71 0.86 1
x and n x and n

FIGURE 5.34
Masks [χ1,6,t], t = 5, 6 and rays of the (1, −1)-projection.

√ √
the image element along
√ the path of length ∆l = 2∆x = 2/7. Therefore,
w(t) = w1,−1 (t) = 7 2v(t), or

1
v(t) = √ w(t), t = 6 : −1 : −6. (5.26)
7 2
310 CHAPTER 5: TENSOR TRANSFORM METHOD

For the given image f(x, y), the line-integrals w(t) are calculated by
√ √
w(6) = 7√2(5) = 7√2 · 5
w(5) = 7√2(2 + 4) = 7√2 · 6
w(4) = 7√2(2 + 1 + 2) = 7√2 · 5
w(3) = 7√2(1 + 1 + 1 + 3) = 7√2 · 6
w(2) = 7√2(3 + 2 + 1 + 1 + 0) = 7√2 · 7
w(1) = 7√2(1 + 1 + 1 + 1 + 2 + 6) = 7√2 · 12
w(0) = 7√2(1 + 1 + 2 + 2 + 9 + 6 + 2) = 7√2 · 23
w(−1) = 7√2(1 + 2 + 4 + 8 + 7 + 1) = 7√2 · 23
w(−2) = 7√2(0 + 3 + 4 + 8 + 3) = 7√2 · 18
w(−3) = 7√2(2 + 3 + 1 + 4) = 7√2 · 10
w(−4) = 7√2(0 + 3 + 2) = 7√2 · 5
w(−5) = 7√2(2 + 9) = 7√2 · 11
w(−6) = 7 2(1) = 7 2 · 1.

The sums v(t), t = 6 : −1 : −6, of the image along the corresponding arith-
metical rays compose the vector v = (5, 6, 5, 6, 7, 12, 23, 23, 18, 10, 5, 11, 1)0,
and the components of the splitting-signal {f1,6,t; t = 0 : 6} are calculated by

f1,6,0 =v(0) = 23
f1,6,1 =v(1) + v(−6) =12 + 1 = 13
f1,6,2 =v(2) + v(−5) =7 + 11 = 18
f1,6,3 =v(3) + v(−4) =6 + 5 = 11
f1,6,4 =v(4) + v(−3) =5 + 10 = 15
f1,6,5 =v(5) + v(−2) =6 + 18 = 24
f1,6,6 =v(6) + v(−1) =5 + 23 = 28.

Figure 5.35 shows the graphs of the line-integrals w(t), sums v(t), and
splitting-signal for the (1, −1)-projection. The control points can be numbered

30
250 25
w1,6(t) v1,6(t) f1,6,t
25
200 20
20
150 15
15
100 10
10

50 5 5

0 0 0
6 3 0 −3 −6 t 6 3 0 −3 −6 t 0 2 4 6 t

FIGURE 5.35
Line-integrals w(t), sums v(t), and splitting-signal for the (1, −1)-projection.
Image Processing 311

from 0 : 12, by using the transformation t → 6 − t,

 6 5 4 3 2 1 0 
• • • • • • •
 7• . . . . . . 
 8 
 • . . . . . .
 9 
 • . . . . . .  (angle of rays is tan−1 (1) = 45◦ ).
 10 
 • . . . . . .
 1 
 • . . . . . .
12
• . . . . . .
The set of 13 geometrical rays of this projection, which is given in (5.24), can
be written as
 t 6
l(t) = l1,−1 (6 − t) = (x, y); x − y = − + ,
7 7
and the system of equations in (5.26) as
√ √
v(t) = w(t)/(7 2) = w(˜l(t))/(7 2), t = 0 : 12. (5.27)

5.2.9 Reconstructed image 7 × 7

To summarize all calculations accomplished above for obtaining eight
splitting-signals from the given image f(x, y), we pack these signals into the
matrix 8 × 7 of the 2-D discrete tensor transform (DTT) as shown below
   
f0,1,0 f0,1,1 f0,1,2 f0,1,3 f0,1,4 f0,1,5 f0,1,6 15 11 9 16 26 33 22
 f1,0,0 f1,0,1 f1,0,2 f1,0,3 f1,0,4 f1,0,5 f1,0,6   7 22 17 26 24 14 22 
   
 f1,1,0 f1,1,1 f1,1,2 f1,1,3 f1,1,4 f1,1,5 f1,1,6   25 24 20 15 16 15 17 
   
 f1,2,0 f1,2,1 f1,2,2 f1,2,3 f1,2,4 f1,2,5 f1,2,6   15 20 19 12 17 20 29 
 f1,3,0 f1,3,1 f1,3,2 f1,3,3 f1,3,4 f1,3,5 f1,3,6   19 20 23 11 17 23 19  .
 = 
   
 f1,4,0 f1,4,1 f1,4,2 f1,4,3 f1,4,4 f1,4,5 f1,4,6   17 6 17 19 26 22 25 
   
 f1,5,0 f1,5,1 f1,5,2 f1,5,3 f1,5,4 f1,5,5 f1,5,6   18 19 24 30 16 13 12 
f1,6,0 f1,6,1 f1,6,2 f1,6,3 f1,6,4 f1,6,5 f1,6,6 23 13 18 11 15 24 28
By using the inverse tensor transformation, which is given in (5.2), we
obtain the following reconstruction of the image f(x, y):
 
1131225
 1 1 1 2 1 1 4
 
 0 2 2 1 1 1 2
 
 2 3 4 2 1 1 3 .
[fn,m ] =  
 0 3 4 8 9 2 0
 
 2 3 1 8 7 6 6
1924312
Below is the script example7x7 inverseDTT.m of the program that computes
the inverse 7 ×7-point DTT. The calculations are performed without the mat-
rix of the tensor transform, as is done in the program example7x7 inverseDTT.m
given in §2.2.2.1.
312 CHAPTER 5: TENSOR TRANSFORM METHOD

% ----------------------------------------------------
% call: example7x7_inverseDTT.m
% The inverse 7x7-point 2-D DTT, for Example 5.1
% A.M. Grigoryan, November 2, 2011
% ----------------------------------------------------
% Set of generators is considered to be calculated as
% J77=ones(2,8); J77(:,1)=[0 1]; J77(2,2:8)=0:6;
DTT_ofimage=[15 11 9 16 26 33 22
7 22 17 26 24 14 22
25 24 20 15 16 15 17
15 20 19 12 17 20 29
19 20 23 11 17 23 19
17 6 17 19 26 22 25
18 19 24 30 16 13 12
23 13 18 11 15 24 28];
% Eq 5.2 of the inverse 2-D discrete tensor transform:
image_fromTT=zeros(7,7);
for n1=0:6
n=n1+1;
for m1=0:6
m=m1+1;
image_fromTT(n,m)=DTT_ofimage(1,n);
col=m1-n1;
for p=2:8
col=col+n1;
if col>=7 col=col-7; end
image_fromTT(n,m)=image_fromTT(n,m)+...
DTT_ofimage(p,col+1);
end
end
end
image_power=sum(DTT_ofimage(1,:));
image_fromTT=(image_fromTT-image_power)/7
% 1 1 3 1 2 2 5
% 1 1 1 2 1 1 4
% 0 2 2 1 1 1 2
% 2 3 4 2 1 1 3
% 0 3 4 8 9 2 0
% 2 3 1 8 7 6 6
% 1 9 2 4 3 1 2
% ----------------------------------------------------

The number of parallel geometrical rays used in each projection is defined

as follows (see Table 5.1):
n(1, 0)=n(0, 1) = N,
n(1, s)=(1 + s)(N − 1) + 1 = N − 1 + s(N − 1), s = 1 : (N − 1)/2,
n(1, N − s) =N 2 − s(N − 1), s = (N + 1)/2 : N − 1.
The total number of rays is 138. The sufficient number of rays is less than
Image Processing 313

138, if we consider the redundancy of the tensor transform. The calculation

of the transform requires one full splitting-signal, let it be {f0,1,t; t = 0 : 6},
and seven incomplete splitting-signals. The mask of the tensor function χ1,0,6
is determined by one ray. The masks of the tensor functions χ1,s,6 , when
s = 1 : 6, are defined by 1, 2, 3, 4, 3, 2, 1 rays, respectively. Therefore the
sufficient number of rays to reconstruct the image on the Cartesian grid 7 × 7,
equals 138 −(1 +2 +3 +4 +3 +2 + 1) = 122. Thus, we need 122 measurements,

TABLE 5.1
Projection data for reconstruction of the discrete image 7 × 7
(p, s) n(p, s) masks in U (p, s)
(1, 0) 7 (1, 0, 0), (1, 0, 1), (1, 0, 2), (1, 0, 3), (1, 0, 4), (1, 0, 5), (1, 0, 6)
(1, 1) 13 (1, 1, 0), (1, 1, 1), (1, 1, 2), (1, 1, 3), (1, 1, 4), (1, 1, 5), (1, 1, 6)
(1, 2) 19 (1, 2, 0), (1, 2, 1), (1, 2, 2), (1, 2, 3), (1, 2, 4), (1, 2, 5), (1, 2, 6)
(1, 3) 25 (1, 3, 0), (1, 3, 1), (1, 3, 2), (1, 3, 3), (1, 3, 4), (1, 3, 5), (1, 3, 6)
(1, 4) 25 (1, 4, 0), (1, 4, 1), (1, 4, 2), (1, 4, 3), (1, 4, 4), (1, 4, 5), (1, 4, 6)
(1, 5) 19 (1, 5, 0), (1, 5, 1), (1, 5, 2), (1, 5, 3), (1, 5, 4), (1, 5, 5), (1, 5, 6)
(1, 6) 13 (1, 6, 0), (1, 6, 1), (1, 6, 2), (1, 6, 3), (1, 6, 4), (1, 6, 5), (1, 6, 6)
(0, 1) 7 (0, 1, 0), (0, 1, 1), (0, 1, 2), (0, 1, 3), (0, 1, 4), (0, 1, 5), (0, 1, 6)
138 [122]

or line-integrals w(t) = wp,s (t) of the image f(x, y), to reconstruct 49 values
of the discrete image fn,m . Equations of the geometrical rays for the image
7 × 7 are given in Table 5.2.

TABLE 5.2
Equations of geometrical rays for the image 7 × 7 (case A)
(p, s) A-rays t → t − t0 G-rays ˜
l(t) = l(t − t0 ) t
(1, 0) l1,0 (t) t→t x = t/7 + 1/14 0:6
(1, 1) l1,1 (t) t→t x + y = t/7 + 1/7 0 : 12
(1, 2) l1,2 (t) t → t − 1/2 x + 2y = t/7 + 1/7 0 : 18
(1, 3) l1,3 (t) t→t−1 x + 3y = t/7 + 1/7 0 : 24
(1, 4) l1,−3 (t) t→t−1 x − 3y = t/7 − 2/7 6 : −1 : −18
(1, 5) l1,−2 (t) t → t − 1/2 x − 2y = t/7 − 1/7 6 : −1 : −12
(1, 6) l1,−1 (t) t→t x − y = t/7 6 : −1 : −6
(0, 1) l0,1 (t) t→t y = t/7 + 1/14 0:6

Equations for solving the convolution equations are given in Table 5.3.

5.3 General algorithm of image reconstruction

For N prime, the block-diagram of the transform-based method of reconstruc-
tion of the image on the Cartesian grid N ×N from N +1 projections is shown
in Figure 5.36.
314 CHAPTER 5: TENSOR TRANSFORM METHOD
TABLE 5.3
Solutions of equations: Case 7 × 7
(p, s) a, an w = KAv, b = A−1 w v = Bw(Bb)
(1, 0) v wt = 7vt√ (K = 7) v(t) = 17 w(t)
P n √ 1
(1, 1) m vn+7m w(t) = 7 2v(t) (K = 7 2) v(t) = 7√ 2
w(t)
√
w(t) = 7 2 5 [v(t − 1) + v(t)] v(t) = 7√b5/2
P
(1, 2) m vn+7m
t

b0 = w(0),
√
bt = w(t) − bt−1 t = 0 : 18
w(t) = 7 310 [v(t) + v(t − 1) + v(t − 2)] v(t) = 7√b10/3
P
(1, 3) m vn+7m t

b0 = w(0),
√
bt = w(t) − (bt−1 + bt−2 ) t = 0 : 24
w(t) = 7 310 [v(t) + v(t − 1) + v(t − 2)] v(t) = 7√b10/3
P
(1, 4) m vn−7m t

b−18 = w(−18),
√
bt = w(t) − (bt−1 + bt−2 ) t = −18 : 6
w(t) = 7 2 5 [v(t) + v(t − 1)] v(t) = 7√b5/2
P
(1, 5) m vn−7m t

b−12 = w(−12),
√ bt = w(t) − bt−1 t = −12 : −1 : 6
P 1
(1, 6) m vn−7m w(t) = 7 2v(t), t = −6 : 6 v(t) = 7√ 2
w(t)
1
(0, 1) vn wt = 7vt v(t) = 7 w(t)

image f (x, y)
1
Image composition: Random rectangles

? 2
3
Line-integrals of the (p, s)-projection Set of generators
JN,N = {(p, s)}
?
4
Transform of G-rays into A-rays, w → v

?
5
Calculation of the splitting-signal {fp,s,t }

?
6
Inverse 2-D discrete tensor transform

?
reconstructed image fn,m

FIGURE 5.36
Block-diagram of the reconstruction of the image f(x, y) composed of rectan-
gles on the Cartesian grid N × N placed in the square [0, 1] × [0, 1].

The following steps describe the proposed method of image reconstruction.

1. The image N × N is composed of a few random rectangles on the square
[0, 1] × [0, 1].
Image Processing 315

2. The set JN,N of generators (p, s) is calculated to define N +1 projections.

The considered set is JN,N = {(1, s); s = 0 : (N − 1)} ∪ {(0, 1)}.
3. Given frequency-point (p, s), the (p, s)-projection is calculated, i.e. all
line-integrals w(t) along the set of geometrical rays ˜lp,s (t). For that, the in-
tersections of each ray of the projection with the rectangles are calculated.
4. The geometry of G-rays of the (p, s)-projection is transformed to A-
rays, i.e. the set of sums v(t) of the discrete image fn,m is calculated from
the integrals w(t), by solving the system of linear equations with the Toeplitz
matrix (M × M ), where M = (p + s − 1)(N − 1) + 1, if s ≤ (N − 1)/2, and
M = (p + (N − s) − 1)(N − 1) + 1, if s > (N − 1)/2. The Toeplitz matrix is
triangular and the transform of geometry of G-rays to A-rays is fast.
5. The splitting-signal {fp,s,0 , fp,s,1, ..., fp,s,N−1} is calculated from v(t).
This signal is written into the corresponding row of the matrix N × (N + 1)
of the 2-D discrete tensor transform.
6. The inverse 2-D tensor transform is calculated, which is the recon-
structed discrete image fn,m .

5.4 Program description and image model

The image f(x, y) is considered in the form of N 2 cells, or image elements (IE)
with intensities fn,m for the (n, m)-th IE, where n, m = 0 : (N −1). To compose
such an image, a few rectangles rk (x, y), k = 1 : K are randomly placed on
the grid N × N in the square [0, 1] × [0, 1]. We denote by (xk , yk , ∆xk , ∆yk )
the rectangle data which include the position and size of the rectangle. The
intensity of the k-th rectangle is denoted by Rk . Thus, the image is described
as
N−1 N−1  
1 X X 1 1
fd (x, y) = fn,m rect 2 N x − n − , N y − m −
(∆x)2 m=0 n=0 2 2
K   (5.28)
1 X x − xk 1 y − yk 1
= Rk rect2 − , − .
(∆x)2 ∆xk 2 ∆yk 2
k=1

The value of fn,m is defined as the sum of intensities of rectangles that cover
the (n, m)-th IE.
As an example, Figure 5.37 shows the gray-scale image of size 257 × 257,
which is composed of ten rectangles (on the left), and its discrete representa-
tion (on the right), after the image reconstruction. The image is the sum of
these rectangles.
The program generates data of random rectangles in the integer form as
shown in Table 5.4, for the image shown in the figure.
The coordinates of these rectangles are used for the discrete grid N × N,
by normalizing them, xk = x0k /N, yk = yk0 /N, ∆xk = ∆x0k /N, ∆yk = ∆yk0 /N,
316 CHAPTER 5: TENSOR TRANSFORM METHOD

random rectangles and grid 257x257 reconstructed image 257x257

1

0.8 50
2
8 100
0.6
7

0.4 1 4 150
6
5 3

0.2 9 200
10

250
0
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 1 11 3 5 5 8 4 10 3 13 cpu time/rect=1.86 min

FIGURE 5.37 (See color insert)

Ten colored rectangles and the discrete gray-scale image of size 257 × 257,
which is composed of these rectangles.

TABLE 5.4
Integer data of 10 rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 25 84 22 40 1
2 68 151 116 65 11
3 132 28 112 117 3
4 198 101 10 1 5
5 16 72 72 32 5
6 27 53 20 87 8
7 122 95 34 86 4
8 154 105 15 114 10
9 99 45 79 25 3
10 35 10 106 54 13

k = 1 : 10. These coordinates of all rectangles and their intensities are given
in Table 5.4.
Below is the beginning of the script (image reconstionA3A4.m), where the
random rectangles are generated and saved as data data allrect in integer
format [x,y,dx,dy,d intensity].

% --------------------------------------------------------------
% call: image_reconstructionA3A4m.m
% Modified General code (N is prime)
% The rectangles are assumed on the grid NxN.
% One set of projections is used to scan all rectangles.
% The inverse 2-D tensor transform is used
% Artyom and Meruzhan Grigoryan, June 30 / July 12, 2011
Image Processing 317
TABLE 5.5
Data of 10 rectangles
k xk yk ∆xk ∆yk Rk
1 0.0973 0.3268 0.0856 0.1556 1
2 0.2646 0.5875 0.4514 0.2529 11
3 0.5136 0.1089 0.4358 0.4553 3
4 0.7704 0.3930 0.0389 0.0039 5
5 0.0623 0.2802 0.2802 0.1245 5
6 0.1051 0.2062 0.0778 0.3385 8
7 0.4747 0.3696 0.1323 0.3346 4
8 0.5992 0.4086 0.0584 0.4436 10
9 0.3852 0.1751 0.3074 0.0973 3
10 0.1362 0.0389 0.4125 0.2101 13

% Input data ---------------------------------------------------

fprintf(’\n Start the code (case N is prime) \n’);
N=input(’ input the value of N: ’);
if ~isprime(N)
P=primes(N); mprime=length(P); N=P(mprime);
fprintf(’ N to be the prime as %g \n’,N);
end
% --------------------------------------------------------------
n=input(’ input the number of rectangles (1,2,...,16): ’);
many_rectangles=n;
% --------------------------------------------------------------
data_allrect=zeros(many_rectangles,5);
NM=(N-1)/2; N10=N-10;
for k=1:many_rectangles
kk=1;
while kk==1
for kkk=1:7
x=rand*N10+4; dx=rand*NM+1; y=rand*N10+4; dy=rand*NM+1;
end
if (x+dx<N-4) & (y+dy<N-4)
data_1rect=round([x,y,dx,dy]); kk=0;
end
end
% data_1rect is in integer format [x,y,dx,dy]
% plus the intensity (<=16) of the image:
d_intensity=round(rand*16)+1; data_1rect(5)=d_intensity;
data_allrect(k,:)=data_1rect;
end
draw_allfigures; % draw all rectangles on the square [0,1]x[0,1].
. . .
318 CHAPTER 5: TENSOR TRANSFORM METHOD

5.5 System of equations

Three different systems of equations are considered for the parallel rays of
(p, s)-projections. All N + 1 projections can be divided into three parts. The
first part is defined by the (1, s)-projections, where 0 < s ≤ (N −1)/2, and the
second part contains (1, s)-projections, where (N − 1)/2 < s < N. The first
part is referred to as case A3, and the second part of projections is referred
to as case A4, similar to the N = 2r case, which is described in Chapter 4.
The (1, 0)- and (0, 1)-projections, i.e., the horizontal and vertical projections,
are considered separately.
Case (A3): For generators (1, s), 0 < s ≤ N/2, the A-rays are defined as
n t 1 + so
l(t) = l1,s (t) = x + sy = + ,
N 2N
where t = 0 : (s + 1)(N − 1). The set of G-rays are the shifted versions of the
A-rays; the shift is t0 = (1 + s)/2 − 1. Therefore, the following equations are
used for the G-rays:
n t 1o
˜
l(t) = l̃1,s (t) = l1,s (t − t0 ) = x + sy = + . (5.29)
N N

Case (A4): For generators (1, s), (N − 1)/2 < s < N, the (1, s − N )-
projections are considered. Therefore, the A-rays are described by equations
n t 1+s−No
l(t) = l1,s−N (t) = x − (N − s)y = + ,
N 2N
where t = (N − 1) : −1 : (s − N )(N − 1). The set of G-rays are defined from
the A-rays by the shift t0 = (1 + N − s)/2 − 1. The equations for the G-rays
are described as
n t 1 + so
l(t) = ˜
˜ l1,s (t) = l1,s (t − t0 ) = x − (N − s)y = −1+ . (5.30)
N N

When (p, s) = (0, 1) or (1, 0), the G-rays are defined as

n t 1 o
l(t) = lp,s (t) = px + sy = + , t = 0 : (N − 1).
N 2N
It should be noted that the above equations correspond to the considered
(X, Y ) coordinate system. In the system (X, 1 − Y ), the same equations of
parallel rays can be considered, after changing y by (1 − y). In the program,
the control points of parallel rays for (1, s)-projections, where s > (N − 1)/2,
are denoted from 0 to (p + s)(N − 1), in order to avoid negative indexing. This
new numbering of rays is accomplished by the transformation: t → (N −1)−t.
Therefore, the equations given above for the A-rays and G-rays are described
in the following way.
Image Processing 319

Case (A3): Generators are (1, s), when 0 < s ≤ N/2. The equations for
A-and G-rays are:
n t 1 + so
l(t) = l1,s (t)= x − sy = −s− ,
n N 2N
o (5.31)
˜ t 1
l(t) = ˜
l1,s (t)= x − sy = −s+ ,
N N
where t = 0 : (s + 1)(N − 1).
Case (A4): Generators are (1, s), when (N − 1)/2 < s < N. The A-rays
and G-rays are described by equations
n t N + s − 1o
l(t) = l1,s−N (t)= x − (N − s)y = − + ,
n N o2N
˜ t s (5.32)
l(t) = ˜
l1,s (t)= x − (N − s)y = − + ,
N N
t= 0 : (1 + N − s)(N − 1).

When (p, s) = (0, 1) and (1, 0), the G-rays are defined as
n t 1 o
l(t) = lp,s (t) = px − sy = + −s , t = 0 : (N − 1).
N 2N

5.6 Solutions of convolution equations

For generators (p, s) ∈ JN,N , the relations between the sums v(t) = vp,s (t) of
the discrete image fn,m and line-integrals w(t) = wp,s (t) of the original image
f(x, y) are described in matrix form by Toeplitz matrices, which are similar
to the described 7 × 7 example. The set of generators JN,N is divided by three
subsets. Therefore, we consider the corresponding convolution equations for
each subset separately.
Subset I: p = 1 and 0 < s ≤ (N − 1)/2.
The convolution equation w = Av is described as
√
1 + s2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s + 1)], (5.33)
s
t = 0 : (1 + s)(N − 1).

Here, it is assumed that v(−1) = v(−2) = ... = v(−s + 1) = 0. The inverse

320 CHAPTER 5: TENSOR TRANSFORM METHOD

transform b = A−1 w is calculated in the following recurrent form:

b0 = w(0),
b1 = w(1) − b0 ,
b2 = w(2) − (b1 + b0 ),
b3 = w(3) − (b2 + b1 + b0 ),
(5.34)
... ... ... ... ... ...
bs−1 = w(s − 1) − (bs−2 + bs−3 + ... + b1 + b0 ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−s+2 + bt−s+1)),
t = s, s + 1, . . . , (1 + s)(N − 1).

The required sums v(t) of the discrete image are calculated by

s
v(t) = bt √ , t = 0 : (1 + s)(N − 1). (5.35)
N 1 + s2
In the s = 0 case, the line-integrals and sums are equal, up to a normalized
factor, i.e., v(t) = w(t)/N, t = 0 : (N − 1).
Subset II: p = 1 and s ≥ (N + 1)/2.
For the generator (1, s), the (1, s − N )-projection is considered, i.e., s is
changed by s − N. The convolution equation w = Av is described by
√
1 + s̄2
w(t) = N [v(t) + v(t − 1) + v(t − 2) + ... + v(t − s̄ + 1)],
s̄ (5.36)
t = (N − 1) : −1 : −s̄(N − 1),

where s̄ = N − s, and v(−k) = 0, when k > s̄(N − 1)).

We denote by M the negative number −s̄(N − 1). The inverse transform
b = A−1 w can be calculated in the following recurrent form:

bM = w(M ),
bM +1 = w(M + 1) − bM ,
bM +2 = w(M + 2) − (bM +1 + bM ),
bM +3 = w(M + 3) − (bM +2 + bM +1 + bM ),
... ... ... ... ... ... ... ... (5.37)
bM +s̄−1 = w(M + s̄ − 1) − (bM +s̄−2 + bM +s̄−3 + ... + bM +1 + bM ),
bt = w(t) − (bt−1 + bt−2 + ... + bt−s̄+2 + bt−s̄+1) ),
t = M + s̄, M + s̄ + 1, . . . , (N − 1).

The required sums v(t) of the discrete image are calculated by

s̄
v(t) = bt √ , t = (N − 1) : −1 : −s̄(N − 1). (5.38)
N 1 + s̄2
Subset III: p = 0 and s = 1. For the vertical projection, the line-integrals
define the sums by v(t) = w(t)/N, t = 0 : (N − 1).
Image Processing 321

5.6.1 Splitting-signal composition

For (p, s)-projections, the corresponding splitting-signals {fp,s,t ; t = 0 : (N −
1)} are calculated from the sums vt = vp,s (t), by using only operations of
addition.
Subset I: p = 1 and 0 < s ≤ (N − 1)/2.
The components of the splitting-signal are calculated by

fp,s,t = vt + vt+N + · · · + vt+sN , t = 0 : (N − 1),

where vk = 0, for k > (1 + s)(N − 1).

Subset II: p = 1 and s ≥ (N + 1)/2.
The components of the splitting-signal are calculated by

fp,s,t = fp,−s̄,t = vt + vt−N + · · · + vt−s̄N , t = (N − 1) : −1 : −s̄(N − 1),

where s̄ = N − s, and vk = 0, for k < −s̄(N − 1).

For the horizontal and vertical projections, when (p, s) = (1, 0) and (0, 1),
respectively, the components of the splitting-signals are equal to the sums
vt = vp,s (t), i.e. fp,s,t = vt , t = 0 : (N − 1).
For the image f(x, y) shown in Figure 5.37, we consider a few line-integrals,
sums, and splitting-signals. Figure 5.38 shows these characteristics for the
(1, 0)-projection. The splitting-signal {f1,0,t; t = 0 : 256} and the vector of

10
w1,0(t) 0.035 v1,0(t) 0.035 f1,0,t
8 0.03 0.03
0.025 0.025
6
0.02 0.02
4 0.015 0.015
0.01 0.01
2
0.005 0.005
0 0 0
0 100 200 0 100 200 0 100 200

FIGURE 5.38
Line-integrals, sums, and the splitting-signal for the horizontal projection.

sums v(t) = v1,0 (t) are the same, and equal the line-integrals, up to the
constant 257, i.e., v(t) = w(t)/257. Figure 5.39 shows the data for the (1, 2)-
and (1, 254)-projections. For the (1, 2)-projection, the line-sums v(t) = v1,2 (t),
where t = 0 : 768, are calculated from the line-integrals w(t) = w1,2(t) by
√
v(t) = bt /(257 5/2), t = 0 : 768,
b0 = w(0), bt = w(t) − bt−1 , t = 1 : 768.

The components of the splitting-signal {f1,2,t ; t = 0 : 256} are calculated by

f1,2,t = vt + vt+257 + vt+514 , t = 0 : 256, where v769 = v770 = 0.
322 CHAPTER 5: TENSOR TRANSFORM METHOD

10 0.01
w1,2(t) v1,2(t) 2 f1,2,t
8 0.008
1.8
6 0.006

1.6
4 0.004

2 0.002 1.4

0 0 1.2
0 200 400 600 0 200 400 600 0 100 200

−5
x 10
8 4 0.022
w1,254(t) v1,254(t) f1,254,t
0.02
6 3
0.018

4 2 0.016

0.014
2 1
0.012

0 0 0.01
256 −256 −768 256 −256 −768 0 100 200

FIGURE 5.39
Line-integral, sums, and splitting-signal for the (1, 2)- and (1, 254)-projections.

For the (1, 254)-projection, the (1, −3)-projection is calculated instead. A

total of 1025 line-sums v(t) = v1,−3 (t), where t = 256 : −1 : −768, are
calculated from the line-integrals w(t) = w1,−3 (t) by
√
 = bt /(257 10/3), t = 256 : −1 : −768,
v(t)
 b−768 = w(−768),
b−767 = w(−767) − b−768 ,

bt = w(t) − (bt−1 + bt−2 ) t = −766 : 256.

The splitting-signal {f1,254,t; t = 0 : 256} is calculated by

f1,254,t = vt + vt−257 + vt−514 + vt−771 , t = 0 : 256,

where v−769 = v−770 = v−771 = 0.

5.6.2 Inverse 2-D tensor transform

The 2-D discrete tensor transform of the image fn,m is considered in the form
of the matrix N ×(N +1). This matrix is composed from N +1 splitting-signals
Image Processing 323

as shown  
f0,1,0 f0,1,1 . . . f0,1,N−1
f0,1,2

 f1,0,0 f1,0,1 . . . f1,0,N−1 
f1,0,2 

 f1,1,0 f1,1,1 . . . f1,1,N−1 
f1,1,2 .

 f1,2,0 f1,2,1 . . . f1,2,N−1 
f1,2,2 
 ... ... ...
... ... 
f1,N−1,0 f1,N−1,1 f1,N−1,2 . . .f1,N−1,N−1
For the image f(x, y) shown in Figure 5.37, the 2-D DTT with 258
splitting-signals on its rows is shown in the form of the gray-scale image in
Figure 5.40 in part a. The graph describing the energies of all splitting-signals
4
x 10
2
E1,0
1.95
50
1.9 E1,1

100 1.85

1.8
E0,1
150
1.75 E1,130

200 1.7

1.65
250
1.6
50 100 150 200 250 0 50 100 150 200 250
(a) (b)

FIGURE 5.40
(a) Two-dimensional 258 × 257-point DTT of the image (in scaled format),
and (b) the graph of the energy of splitting-signals.

is given in part b. Four splitting-signals have high energies

q
2
Ep,s = fp,s,0 2
+ fp,s,1 2
+ ... + fp,s,256

which are equal to 19718, 19060, 17235, and 17544, for the (1, 0), (1, 1), (0, 1),
and (1, 130)-projections, respectively.
The reconstructed discrete image fn,m is obtained by the inverse tensor
transformation, which is calculated by (5.2). The calculation of the inverse
2-D DTT by this formula is used in the script (example7x7 inverseDTT.m) in
the considered 7 × 7 example. The following script (inverse 2DDTT.m) can be
used for computing the inverse N × N -point 2-D DTT in the general case,
when N is prime.

% --------------------------------------------------
% Call: inverse_2DDTT.m
% Jps={(1,s);s=0:N-1},{(0,1)}
% A.M. Grigoryan and Nan Du, 10/27/2009
324 CHAPTER 5: TENSOR TRANSFORM METHOD

% --------------------------------------------------
function image_fromTT=inverse_2DDTT(DTT_ofimage,N)
% The inverse formula of the tensor transform:
image_fromTT=zeros(N,N);
N1=N-1;
for n1=0:N1
n=n1+1;
for m1=0:N1
m=m1+1;
image_fromTT(n,m)=DTT_ofimage(1,n);
col=m1-n1;
for p=2:N+1
col=col+n1;
if col>=N
col=col-N;
end
image_fromTT(n,m)=image_fromTT(n,m)+...
DTT_ofimage(p,col+1);
end
end
end
% Remove the redundancy of the tensor transform
image_power=sum(DTT_ofimage(1,:));
image_fromTT=(image_fromTT-image_power)/N;
% --------------------------------------------------

5.7 MATLAB
R
-based code (N prime)
In this section, we give the script image reconstructionA3A4m.m of the main
program for image reconstruction. The beginning of this program is given in
§5.4, and below is the remaining part of this program.
% -----------------------------------------------------
% call: image_reconstructionA3A4m.m
....
% =====================================================
% START THE IMAGE RECONSTRUCTION
% 1. The set J of all (N+1) generators (p,s):
for p=0:N-1
Jps(p+1,:)=[p,1];
end
Jps(p+2,:)=[1,0];
Nps=N+1;
% Reconstruction of all rectangles
% 1. Composition of the tensor from the integrals
image_pst=reconstruct_AllrectanglesA34m(data_allrect,N,Jps);
Image Processing 325

% 2. Inverse 2-D tensor transform

imageR=inverse_2DDTT(image_pst,N);
% --------------------------------------------
figure(h_fig);
subplot(1,2,2);
imagesc(imageR’);
axis image;
s_title=sprintf(’reconstructed image %gx%g’,N,N);
title(s_title);
display(’ End of the code: image_reconstructionA3A4.m ’);
% ========================================================

The script of the function reconstruct AllrectanglesA34m.m, which ac-

complishes the calculation of all line-integrals w(t) and the sums v(t) is given
below. This function uses the rectangle data (data allrect) and set of gene-
rators (Jps) calculated in the main program (image reconstructionA3A4m.m).

% ---------------------------------------------------------------------
% call: reconstruct_AllrectanglesA34m.m
% Processing: one projection for all rectangles.
% Set of generators: Jps={{(p,1);p=0:N-1},(1,0)}
% Each (p,s)-projection is processing and the 2-D tensor
% is composed in row # (p+s)+1 (and 1, when (p,s)=(0,1))
% Artyom and Meruzhan Grigoryan, June 30 / July 12, 2011
% ---------------------------------------------------------------------
function image_pst=reconstruct_AllrectanglesA34m(data_allrect,N,Jps)
%I 1. Determine the data of rectangles
many_rectangles=size(data_allrect,1);
d_xin1=1/N; % the length of 1 cell (IE)
d_yin1=d_xin1; % the high of the cell (IE)
% 2. The set of values of np=p,N-p in Jps
Nps=N+1; N_half=Nps/2-1;
np_formatrixT=zeros(1,Nps);
np_formatrixT(1:N_half+1)=0:N_half; % case 1
tnp=N-N_half-1:-1:1;
np_formatrixT(N_half+2:N)=tnp; % case 2
np_formatrixT(Nps)=0; % case 3
%II Process each (p,s)-projection
image_pst=zeros(Nps,N); % matrix for the 2-D DTT
number_ofallintegrals=0;
for kps=1:Nps
p=Jps(kps,1); s=Jps(kps,2);
if (s==0) % & (p==1) % for (1,0)
nrow=1; % the row in 2-D DTT
else nrow=p+2; % for (p,1), p=0:(N-1)
end
np=np_formatrixT(kps);
if np==0 % cases (p,s)=(1,0) and (0,1)
delta_r=d_xin1; % 1/N;
326 CHAPTER 5: TENSOR TRANSFORM METHOD

else
delta_r=sqrt(1+np^2)/d/N;
end
% delta_r = length of the path of the ray in IE
for k=1:many_rectangles
% 1. Read the rectangle data
data_1rect=data_allrect(k,:);
n0=data_1rect(1); m0=data_1rect(2);
dx=data_1rect(3); dy=data_1rect(4); k_in=data_1rect(5);
% The coordinates of the k-th rectangle:
x0=n0; y0=m0; x1=(n0+dx); y1=(m0+dy);
p_rect=[x0 y0; x1 y1];
% 2. Calculation of the (p,s)-projection, w(t):
[W,number_ofw]=projection_inrectangle2N(p,s,N,p_rect,N_half);
number_ofallintegrals=number_ofallintegrals+number_ofw;
W=W/N; % to go back to the grid [0,1]x[0,1]
W=W*k_in; % consider the rectangle intensity
% 3. Translate geometry: G-rays into A-rays
if np==0 % cases (p,s)=(1,0) and (0,1)
B=W;
else
B=solution_of_linearsystemA(np-1,W);
end
% 4. Composition of the N-dim. vector a(n) from B(t)
A=Avector_composition(N,B);
% 5. Normalization by the path
A=A/delta_r;
% 6. Correction in case 2: reordering of data in A(n)
A1=A;
if (s==1) && (p>N_half)
A1=vectorA_indexcorection(p-N,N,A);
end
% 7. Matrix composition for the Nx(N+1)-point DTT
for i1=1:N
image_pst(nrow,i1)=image_pst(nrow,i1)+A1(i1);
end
clear W B A A1;
end
end
fprintf(’\n # of all integrals w(t) is %g \n’,number_ofallintegrals);
% ---------------------------------------------------------------------

The time characteristics of calculation, when the program was imple-

mented in MATLAB are given in Table 5.6. The time for calculation of the
2-D tensor transform and its inverse transform are also given.
Table 5.7 shows the total time of image processing when the program was
implemented in Ci+. The time includes the calculation of all line-integrals
w(t), line-sums v(t), and the calculation of the 2-D direct and inverse tensor
transforms. The data in these tables were obtained from running programs on
Image Processing 327
TABLE 5.6
Time for MATLAB-based program [A3-A4]
N ×N scan & reconstruct 2-D DTT 2-D iDTT
61 × 61 0.88s/rec 0.s 0s
101 × 101 4.11s/rec 0.2340s 0.5304s
127 × 127 8.45s/rec 0.2340s 0.5304s
257 × 257 1.86m/rec 1.9188s 4.5084s
401 × 401 24.01m/rec 0.2340s 0.5304s
509 × 509 24.01m/rec 0.2340s 0.5304s

TABLE 5.7
Time for the Ci-based program (the 1st version)
N ×N projection data processing & reconstruction
37 × 37 00 : 00 : 00.03
67 × 67 00 : 00 : 00.08
101 × 101 00 : 00 : 00.27
127 × 127 00 : 00 : 00.50
257 × 257 00 : 00 : 04.21
401 × 401 00 : 00 : 16.67
509 × 509 00 : 00 : xx.xx
521 × 521 00 : 00 : 37.24
797 × 797 00 : 02 : 15.82
1031 × 1031 00 : 04 : 55.69
1777 × 1777 00 : 25 : 29.33
2053 × 2053 00 : 39 : 40.44

a Dell computer with Intel Dual CPU processor at 3.20-GHz speed.

Problems
Problem 5.1 Consider case A for all generators (1, s), s = 0 : (N − 1),
when all G-rays are described by equation (5.29). Consider also that the
(1, s)-projections are not substituted by (1, s − N )-projections, when s >
(N − 1)/2. Write the code for image reconstruction using these rays. Dis-
cuss how fast or slow your program is, when compared with the program
image reconstructionA3A4m.m. For that, you may consider images on the square
[0, 1] × [0, 1], which are composed of random rectangles, each with constant
intensity, and the Cartesian lattices N × N, for different prime N.

Problem 5.2 Suppose that you are given an image in the square [0, 1] × [0, 1]
with image elements in the lattice 257 × 257. Generate such an image from
the random rectangles.
A. For the (1, 4)-projection, calculate the line-integrals w(t), sums v(t),
and the tensor splitting-signal {f1,4,t; t = 0 : 256}. Sketch your results.
328 CHAPTER 5: TENSOR TRANSFORM METHOD

B. In most cases, the splitting-signals for the vertical and horizontal pro-
jections have high energy and represent big contributions in the reconstruction
of the image. Discuss the reconstruction without these projections. In other
words, calculate the image which is reconstructed from the (1, s)-projections,
where s = 1 : 256.
Problem 5.3 Suppose that you are given an image in the square [0, 1]×[0, 1],
which is composed of a few rectangles with constant intensity each. Each image
element has the size 1/10 × 1/10. Discuss the reconstruction of the image on
the lattice 10 × 10. Write your code and show results of reconstruction on
examples.
Problem 5.4 The method of G-rays and tensor transform for image recon-
struction, which is described in this chapter, can be applied for any size lattice
N × N, where N > 1 is an integer. Suppose that you are given an image in
the square [0, 1] × [0, 1], which is composed of a few rectangles with constant
intensity each. Each image element has the size 1/N × 1/N. Discuss the re-
construction of the image on the lattice N × N, when N = 100.
Problem 5.5 Suppose that you are given an image in the square [0, 1]×[0, 1].
As an example, generate such an image from the random rectangles on the
Cartesian lattice 128 × 128. Sample this image on the lattice 127 × 127 and
calculate the reconstruction of the image on this lattice. Discuss the result
and calculate the minimum mean-square-root error of reconstruction.
Problem 5.6∗ Our program image reconstructionA3A4m.m works 1.80 minu-
tes (with graphics), when reconstructing the image on the lattice 257 × 257.
The image is composed of ten random rectangles and it is shown in Figure 5.41.
The realization of this program in Ci+ is much faster (by about 10 s). Develop

random rectangles and grid 257x257 reconstructed image 257x257

1

9
0.8 6 50

5 1
0.6 100
28
10
4
0.4 150
7
3

0.2 200

250
0
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 10 4 15 12 5 11 16 7 9 17 cpu time/rect=1.80 min

FIGURE 5.41 (See color insert)

Image with ten rectangles and its reconstruction 257 × 257.

the Ci version of this code, or write your own code in Ci+ to reconstruct the
image from its projections.
6
Method of Particles

In this chapter, we describe a new method of summation of projection data by

G-rays, which is based on the concept of the point map of projections. For this
concept, the map of each point on the Cartesian lattice N × N is considered
in the form of a matrix which describes all A-rays passing through this point.
Each point is considered as a particle on the lattice, which is described by the
field which we call the field of the particle. This concept is then generalized
for the G-rays, and the field functions are defined for the G-particles, that
represent the image elements as particles, or the smallest squares on the lattice.
The field functions of such particles define all G-rays passing through this
particle. The consideration of field functions for both A- and G-particles leads
to a representation of the image by the field functions, and this representation
allows us to reconstruct the image from its projections. We do consider that
the image has a corpuscular structure, i.e. is composed of quantum-s, or small
image elements which are described by their fields, and study of these fields
is important for solving the problem of image reconstruction.

6.1 Point-map of projections

For the image f(x, y) in the square [0, 1] ×[0, 1], which is represented by image
elements of size ∆x∆y = 1/N 2 each, we derived the formula for the integral
along each ray lp,s (t) of the (p, s)-projection. This integral is the linear sum of
intensities of the image on the image elements, which are intersected by the
ray. The coefficients of this sum are equal to the lengths of intersection of the
ray with the image elements.
Thus, the ray is described by a set of image elements {[nk , mk ]; k = 1 : K},
where K is the discrete length of the ray lp,s (t), or the number of image
elements intersected by the ray. We denote the lengths of the intersections by
∆l(nk , mk ). It is assumed that ∆l(nk , mk ) 6= 0 and intersection in only one
point is not considered. The integral of the image f(x, y) along the ray equals
K
X
wp,s (t) = fnk ,mk ∆l(nk , mk ). (6.1)
k=1

For the scanning scheme that is used in the tensor and paired methods of

329
330 CHAPTER 6: Method of Particles

image reconstructions, the lengths ∆l(nk , mk ) are equal for all rays of the
projection, i.e.. ∆l(nk , mk ) = ∆l(p, s). Therefore, the above equation can be
written as
K
X
wp,s (t) = ∆l(p, s) fnk ,mk . (6.2)
k=1

Let rp,s (t) be the real value of the f(x, y) image integral along the ray
lp,s (t). We consider the mean square deviation of the sum of real rays from
their discrete representation
X 2
ε= (wp,s (t) − rp,s (t)) , (6.3)
(p,s)∈JN,N

where JN,N is the set of all generators (p, s). By minimizing this error by
derivatives, we consider the system of equations

∂ε X ∂wp,s (t)
= 2 (wp,s (t) − rp,s (t)) = 0, n, m = 0 : (N − 1),
∂fn,m ∂fn,m
(p,s)∈JN,N

or
X ∂wp,s (t) X ∂wp,s (t)
wp,s (t) = rp,s (t) .
∂fn,m ∂fn,m
(p,s)∈JN,N (p,s)∈JN,N

When solving this system, we obtain the following:

X X
wp,s (t)l(p, s)δn,nk δm,mk = rp,s (t)l(p, s)δn,nk δm,mk ,
(p,s)∈JN,N (p,s)∈JN,N
X X
wp,s (t)l(p, s) = rp,s (t)l(p, s).
lp,s (t)∩[n,m]6=∅ lp,s (t)∩[n,m]6=∅

One can notice that for each image element, or square [n, m], the summations
of integrals are accomplished only by the rays lp,s (t) passing through [n, m].
Therefore, we are faced with the concept of the “particle” as the sum of rays
passing through the given image element. It should also be noted that if we
consider equations (6.2) in normalized form,
K
wp,s (t) X
w̃p,s (t) = = fnk ,mk ,
∆l(p, s)
k=1

the minimization of the mean-square error results in the following condition:

X X
w̃p,s (t) = rp,s (t).
lp,s (t)∩[n,m]6=∅ lp,s (t)∩[n,m]6=∅

This means the equations and the concept of the particle, which we want to
introduce, can be described in integer form, such as integer matrices.
Image Processing 331

This concept will be described first for the A-rays. For that, we create a
picture of all rays passing through the center of a given (n0 , m0 )-th square,
where n0 , m0 ∈ {0, 1, ..., (N − 1)}, in the discrete model of the image. First
we consider the example. For the case when N = 8 and (n0 , m0 ) = (3, 3), we
introduce the following table:
 
1 0 1 1 1 0 1 0
0 1 1 1 1 1 0 1
 
0 1 1 1 1 1 0 1
 
 1 1 1 12 1 1 1 1 
A(3, 3) = A(8; 3, 3) = 0 1 1 1 1 1 0 0 .
 (6.4)
 
0 1 1 1 1 1 0 1
 
1 0 1 1 1 0 1 0
0 1 1 1 0 1 0 1

We call this table the map of projections for point (3, 3), or the point map of
projections for (3, 3). For (n, m) 6= (3, 3), the coefficient a(n, m) = a(3, 3, n, m)
of this matrix equals 1, if there is an A-ray passing through the points (3, 3)
and (n, m) at the same time. This coefficient equals 0, if there is no such A-ray.
In the center, the coefficient a(3, 3) equals 12, or the number of projections.
There are 12 rays passing through the point (3, 3), which cover 46 points
(n, m) on the Cartesian grid 8 × 8.
The descriptions of all 12 rays passing through the point (3, 3) are given
in Table 6.1. The coordinates of points (n, m), or squares [n, m] that lie on

TABLE 6.1
Coordinates of the projection map for point (3, 3)
(p, s) points [n, m] card g(p, s)
(1, 0) [4, 1] [4, 2] [4, 3] [4, 4] [4, 5] [4, 6] [4, 7] [4, 8] 8 1
(1, 1) [7, 1] [6, 2] [5, 3] [4, 4] [3, 5] [2, 6] [1, 7] 7 1
(1, 2) [8, 2] [6, 3] [4, 4] [2, 5] 4 2
(1, 3) [7, 3] [4, 4] [1, 5] 3 3
(1, 4) [8, 3] [4, 4] 2 4
(1, 5) [7, 5] [4, 4] [1, 3] 3 3
(1, 6) [8, 6] [6, 5] [4, 4] [2, 3] 4 2
(1, 7) [8, 8] [7, 7] [6, 6] [5, 5] [4, 4] [3, 3] [2, 2] [1, 1] 8 1
(0, 1) [1, 4] [2, 4] [3, 4] [4, 4] [5, 4] [6, 4] [7, 4] [8, 4] 8 1
(2, 1) [2, 8] [3, 6] [4, 4] [5, 2] 4 2
(4, 1) [3, 8] [4, 4] 2 4
(6, 1) [3, 2] [4, 4] [5, 6] [6, 8] 4 2

the rays passing through the point (3, 3) are given. On the 3-rd column of the
table, the cardinalities c(p, s), i.e., the numbers of the points (n, m) on the
rays are shown.
To calculate tables A(N ; n0 , m0 ) in the general case of N and (n0 , m0 ), we
can use the commands and scripts of the functions Arays nm.m and Aray throu
gh nm.m, which are given below. The variable Anm stands for the point map of
332 CHAPTER 6: Method of Particles

projections A(N ; n0 , m0 ). The numbering of squares is performing from 1 : N,

not 0 : (N − 1). The function Aray through nm.m calculates the point map A
only for the single (p, s)-projection. It also calculates the ray number t, that
passes through the point (n0 , m0 ), as well as, it saves coordinates (n, m) of this
ray in the 2-D array Ray squares. We also include the command of printing
these coordinates, to compose Table 6.1. In the code Arays nm.m, all single
maps are added together, to obtain the point map A(N ; m0 , n0 ).

% from: run_pointmap.m
N=8; n0=4; m0=4;
Anm=Arays_nm(N,m,n);
% ------------------------------------------------------
% call: Arays_nm.m / A.-M. Grigoryan, June 2011
function Anm=Arays_nm(N,m,n)
Anm=zeros(N);
ps=ps_generators(N); L=size(ps,1);
for n_ps=1:L
p=ps(n_ps,1); s=ps(n_ps,2);
A1=Aray_through_nm(p,s,N,m,n);
Anm=Anm+A1;
end
% -----------------------------------------------------
% call: Aray_through_nm.m / A.-M. Grigoryan, June 2011
function [A,t,Ray_squares]=Aray_through_nm(p,s,N,m0,n0)
S=Arays2grid(p,s,N); L=size(S,1);
[t,ny]=find( S(:,:,1)==m0 & S(:,:,2)==n0 );
card_of_IE=max(find(S(t,:,2),N,’first’));
A=zeros(N);
Ray_squares=zeros(card_of_IE,2);
for k=1:card_of_IE
m1=S(t,k,1); n1=S(t,k,2);
A(m1,n1)=1; fprintf(’[%g,%g] ’,m1,n1);
Ray_squares(k,1)=m1; Ray_squares(k,2)=n1;
end
fprintf(’\n’);
% -----------------------------------------------------

6.1.1 A-particle and the field

In this section, we introduce the concept of the particle and its field function.
For that, we first consider a point (n0 , m0 ) in the Cartesian lattice N × N and
the set of A-rays passing through this point. The point map of projections is a
key function for reconstructing the discrete image where each point is referred
to as a particle on the lattice.

Definition 6.1 The arithmetical particle, or A-particle of dimension N with

the center in (n0 , m0 ) is called the point (n0 , m0 ) on the Cartesian lattice
N ×N, which is described by the matrix ||a(n0 , m0 , n, m)|| defining the number
Image Processing 333

of A-rays passing through the points (n0 , m0 ) and (n, m) at the same time.
We call the two-dimensional function ψn0,m0 (n, m) = a(n0 , m0 , n, m), where
n, m = 0 : (N − 1), the field function of this particle.

Example 6.1 In matrix form, the field functions of the A-particles (0, 1) and
(3, 2) are    
1 1 1 0 0 1 0 1
 6 1 1 1 0 0 1 1 
ψ0,1 = 
1 1 0 0
 and ψ3,2 = 1 1 1 6 .
  (6.5)
1 0 1 0 0 1 1 1
Figure 6.1 illustrates two A-particles (0, 1) and (3, 2) and their field
functions. Points (0, 1), (1, 1), (2, 1), and (3, 1) are on the first A-ray of the

PMP at (0,1) PMP at (3,2)

0 0

0 0

0.25 0.25

1 1
y and m

y and m

0.5 0.5

2 2

0.75 0.75

3 3

1 1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
0 1 x and n 2 3 0 1 x and n 2 3

FIGURE 6.1
The field functions of the particles (0, 1) and (3, 2).

(0, 1)-projection. Points (0, 1), and (1, 0) are on the first A-ray of the (1, 1)-
projection. Points (2, 3), (1, 2), and (0, 1) are on the first A-ray of the (−1, 1)-
projection (or (3, 1)-projection). Points (0, 0), (0, 1), (0, 2), and (0, 3) are on
A-ray number 0 of the (1, 0)-projection, and (2, 0), (0, 1) are on the second
A-ray of the (1, 2)-projection. The only point on the first A-ray of the (2, 1)-
projection is (0, 1). This single-point ray is shown by the bullet; therefore the
number of rays seen on the figure is 5, but it is actually 6. The description
of the particle (0, 1) with its field function is also given in Table 6.2. The
description of the A-particle (3, 2) with its field function is given in Table 6.3.
A-ray number 8 of the (2, 1)-projection passes through only the point (3, 2) in
the lattice 4 × 4, and this single-point ray is shown by the bullet in the figure.
Figure 6.2 illustrates two A-particles (2, 2) and (2, 1) and their field func-
tions. One can notice six rays for both A-particles, because each of six A-rays
passes through at least two points at the same time.
We can write the field functions of all 16 A-particles on the lattice 4 × 4
in the form of a block-table [ψn,m ]n,m=0:3 , as shown in Table 6.4.
This table, or block-table 4 × 4, with all field functions can be packed into
334 CHAPTER 6: Method of Particles
TABLE 6.2
Description of the A-particle (0, 1)
(p, s) points (n, m) A-ray
(0, 1) (0, 1) , (1, 1) , (2, 1) , (3, 1) #1
(1, 1) (0, 1) , (1, 0) #1
(2, 1) (0, 1) #1
(−1, 1) (2, 3) , (1, 2) , (0, 1) #1
(1, 0) (0, 0) , (0, 1) , (0, 2) , (0, 3) #0
(1, 2) (2, 0) , (0, 1) #2

TABLE 6.3
Description of the A-particle (3, 2)
(p, s) points (n, m) A-ray
(0, 1) (0, 2) , (1, 2) , (2, 2) , (3, 2) #2
(1, 1) (2, 3) , (3, 2) #5
(2, 1) (3, 2) #8
(−1, 1) (3, 2) , (2, 1) , (1, 0) #−1
(1, 0) (3, 0) , (3, 1) , (3, 2) , (3, 3) #3
(1, 2) (3, 2) , (1, 3) #7

PMP at (2,2) PMP at (2,1)

0 0

0 0

0.25 0.25

1 1
y and m

y and m

0.5 0.5

2 2

0.75 0.75

3 3

1 1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
0 1 x and n 2 3 0 1 x and n 2 3

FIGURE 6.2
The field functions of the particles (2, 2) and (2, 1).

the following matrix of size 7 × 7, which we call the base matrix of fields of
A-particles:  
1 0 0 1 0 0 1
0 1 0 1 1 1 0
 
0 0 1 1 1 1 0
 
Ψ= 1 1 1 6 1 1 1. (6.6)
0 1 1 1 1 0 0
 
0 1 1 1 0 1 0
1 0 0 1 0 0 1
Image Processing 335
TABLE 6.4
Basis of the field
6 1 1 1 1 6 1 1 1 1 6 1 1 1 1 6
1 1 0 0 1 1 1 0 1 1 1 1 0 1 1 1
1 0 1 0 1 1 0 1 1 1 1 0 0 1 1 1
1 0 0 1 0 1 0 0 0 0 1 0 1 0 0 1
1 1 1 0 1 1 1 1 0 1 1 1 0 0 1 1
6 1 1 1 1 6 1 1 1 1 6 1 1 1 1 6
1 1 0 0 1 1 1 0 1 1 1 1 0 1 1 1
1 0 1 0 1 1 0 1 1 1 1 0 0 1 1 1
1 1 1 0 0 1 1 1 1 0 1 1 0 1 0 1
1 1 1 0 1 1 1 1 0 1 1 1 0 0 1 1
6 1 1 1 1 6 1 1 1 1 6 1 1 1 1 6
1 1 0 0 1 1 1 0 1 1 1 1 0 1 1 1
1 0 0 1 0 1 0 0 0 0 1 0 1 0 0 1
1 1 1 0 0 1 1 1 1 0 1 1 0 1 0 1
1 1 1 0 1 1 1 1 0 1 1 1 0 0 1 1
6 1 1 1 1 6 1 1 1 1 6 1 1 1 1 6

Indeed, it is not difficult to notice that the following equality holds:

ψn,m (i, j) = Ψ(3 − n + i, 3 − m + j), i, j = 0 : 3, (6.7)

for any A-particle (n, m). The base field matrix is symmetric, i.e., Ψ(i, j) =
Ψ(j, i), for i, j = 0 : 6.

We here state that for other values of N = 2r , r > 2, the similar base
matrix of fields of A-particles can also be defined. As an example, we consider
the N = 8 case, and two A-particles (3, 3) and (5, 4), whose field functions
are illustrated in Figure 6.3. There are twelve A-rays passing through the A-

PMP at (3,3) PMP at (5,4)

0 0
0 0
0.25 0.25
1 1
0.25 0.25
2 2
0.375 0.375
y and m

y and m

3 3
0.5 0.5
4 4
0.625 0.625
5 5
0.75 0.75
6 6
0.875 0.875
7 7
1 1
0 0 0.12 1 0.25 2 0.37 3 0.5 4 0.62 5 0.75 6 0.87 7 1 0 0 0.12 1 0.25 2 0.37 3 0.5 4 0.62 5 0.75 6 0.87 7 1
x and n x and n

FIGURE 6.3
The field functions of the particles (3, 3) and (5, 4).

particle (3, 3) in this example. The base matrix of fields of all 64 A-particles
336 CHAPTER 6: Method of Particles

is the following matrix 15 × 15 :

 
1 0 0 0 0 0 0 1 0 0 0 0 0 0 1
0 1 0 0 1 1 0 1 0 1 1 0 0 1 0
 
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
 
0 0 0 1 0 1 0 1 1 1 0 1 0 0 0
 
0 1 0 0 1 0 1 1 1 0 1 0 0 1 0
 
0 0 0 1 0 1 1 1 1 1 0 1 0 0 0
 
0 0 0 0 0 1 1 1 1 1 0 1 0 0 0
 
Ψ= 1 1 1 1 1 1 1 12 1 1 1 1 1 1 1. (6.8)
0 0 0 1 0 1 1 1 1 1 0 0 0 0 0
 
0 0 0 1 0 1 1 1 1 1 0 1 0 0 0
 
0 1 0 0 1 0 1 1 1 0 1 0 0 1 0
 
0 0 0 1 0 1 1 1 0 1 0 1 0 0 0
 
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
 
0 1 0 0 1 1 0 1 0 1 1 0 0 1 0
1 0 0 0 0 0 0 1 0 0 0 0 0 0 1
The field function of each A-particle (n, m), where n, m ∈ {0, 1, ..., 7}, can be
calculated by
ψn,m (i, j) = Ψ(7 − n + i, 7 − m + j), i, j = 0 : 7. (6.9)
One can notice that the matrix given in (6.4) represents the field function
ψ3,3 (i, j) of the A-particle (3, 3). In general, this equation can be written as
ψn,m (i, j) = Ψ(N − 1 − n + i, N − 1 − m + j), i, j = 0 : N − 1, (6.10)
where n, m ∈ {0, 1, ..., N −1}. The size of the base matrix is (2N −1)×(2N −1).
To calculate the base matrix Ψ in the general case of N = 2r , when r > 1,
we use the function bmf Amatrix.m with the following script.
% call: bmf_Amatrix.m, M. Grigoryan, 01/04/2012
% Calculate the base matrix of field of A-particles
function P=bmf_Amatrix(N)
ps=ps_generatorsM(N);
P=zeros(2*N-1);
L=N+N/2;
for n_ps=1:L
p=ps(n_ps,1);
s=ps(n_ps,2);
for m=-(N-1):(N-1)
sm=s*m; m1=N-m;
for n=-(N-1):(N-1)
if (p*n+sm)==0
P(N-n,m1)=R(N-n,m1)+1;
end
end
end
end
Image Processing 337

The matrix Ψ in (6.8) is calculated by calling the command P=bmf Amatrix(8).

The function ps generatorsM is the modification of the function ps generators,
where generators (1, s) and (p, 1) are substituted by (1, s − N ) and (p − N, 1),
respectively, when s and p > N/2.
% call: ps_generatorsM.m
function Jps=ps_generatorsM(N)
L=N+N/2;
Jps=ones(2,L);
Jps(1,1:N/2+1)=0:N/2; Jps(1,N/2+2:N)=-N/2+1:1:-1;
Jps(2,N+1:N+N/4+1)=0:2:N/2; Jps(2,N+N/4+2:L)=-N/2+2:2:-2;

Figure 6.4 illustrates the base matrix of fields of A-particles for the N = 4
and 8 cases in parts a and b, respectively. These matrices are shown in the

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
(a) N=4 (b) N=8

FIGURE 6.4
Base matrices of A-particles for (a) N = 4 and for (b) N = 8.

square [−1, 1] × [−1, 1] with the lattice (2N − 1) × (2N − 1). All A-rays passing
through the center (0, 0) of the base matrices are shown as well. Six A-rays
pass through the original in the N = 4 case, and twelve when N = 8. If the
A-ray passes through the centers of other squares [n, m], then the coefficients
Ψ(n, m) = 1.

6.1.2 Representation by field functions

It is clear that by introducing the field functions of A-particles, we define a
new basis in the space of discrete images fn,m , which is different from the
basis of the 2-D paired transformation. In paired representation, the image
0
{fn,m } is defined as the 3-D array {fp,s,t },

X N−1
N−1 X
0
fp,s,t = fn,m χ0p,s,t (n, m), 0
(p, s, t) ∈ JN,N .
m=0 n=0
338 CHAPTER 6: Method of Particles

The paired representation is a transformation of the image into a set of 1-D

splitting-signals generated by different frequency-points, which we call gene-
rators (p, s). Therefore the paired representation is referred to as the 2-D
frequency and 1-D time transformation. All N 2 triplet-numbers (p, s, t) of the
2-D paired functions can be written in the matrix N × N, as shown in §4.7.3.
0 0
N 2 components fp,s,t can be packed into the 2-D array {fv,w } and the paired
representation can be written as

X N−1
N−1 X
0
fv,w = fn,m χ0v,w (n, m), v, w = 0 : (N − 1).
m=0 n=0

Now we can represent the image by the field functions

X N−1
N−1 X
a(i, j) = fn,m ψn,m (i, j), i, j = 0 : (N − 1). (6.11)
m=0 n=0

The physical meaning of this representation is clear. The image is transformed

into the matrix, or field, the value of which at point (i, j), or A-particle (i, j),
is determined by the field functions of A-particles which can be seen by A-
particle (i, j) along the A-rays. In other words, a(i, j) is the sum of the image
along all rays passing through the point (i, j). By considering the base matrix
of fields of A-particles, we can write the following:

X N−1
N−1 X
a(i, j) = fn,m Ψ(N − 1 − n + i, N − 1 − m + j). (6.12)
m=0 n=0

We should notice that the obtained new basis of field functions ψn,m is not
orthogonal. That may be considered as the disadvantage of this basis. On the
contrary, the paired transform is orthogonal, because of the complex way of
combining parallel rays in each projection. This transform was originated from
the tensor transform being non-orthogonal. However, in the paired transform
the orthogonality was achieved by assigning the plus and minus signs to diffe-
rent parts of rays for each projection. In other words, each particle (p, s, t) in
the 3-D space was considered with its co-particle (p, s, t + N/2), and this pair
0
of particles was described by the function fp,s,t = (fp,s,t −fp,s,t+N/2 ), and as a
result, the principle of energy conservation was preserved. The same method
may be considered for the basis field functions, to define the corresponding
bi-orthogonal basis. For that, we first describe the N = 4 case.
All matrices of the field functions can be presented in the form of 1-D
vectors, and then written into the matrix 16 × 16. To accomplish these calcu-
lations, we can run the code as R=A rays2R(4), with the following script of the
function.
function R=A_rays2R(N)
NN=N*N; R=zeros(NN); k=1;
for n=1:N
Image Processing 339

for m=1:N
Anm=Arays_nm(N,m,n);
y=reshape(Anm,1,NN); R(k,:)=y; k=k+1;
end
end

As a result, we obtain the following matrix:

 
6 1 1 1 1 1 0 01 0 1 0 1 0 0 1
1 6 1 1 1 1 1 01 1 0 1 0 1 0 0
 
1 1 6 1 1 1 1 11 1 1 0 0 0 1 0
 
1 1 1 6 0 1 1 10 1 1 1 1 0 0 1
 
1 1 1 0 6 1 1 11 1 0 0 1 0 1 0
 
1 1 1 1 1 6 1 11 1 1 0 1 1 0 1
 
0 1 1 1 1 1 6 11 1 1 1 1 1 1 0
 
0 0 1 1 1 1 1 60 1 1 1 0 1 1 1
R= 1 1 1 0 1 1 1 06 1 1
. (6.13)
 1 1 1 0 0
0 1 1 1 1 1 1 11 6 1 1 1 1 1 0
 
1 0 1 1 0 1 1 11 1 6 1 1 1 1 1
 
0 1 0 1 0 0 1 11 1 1 6 0 1 1 1
 
1 0 0 1 1 1 1 01 1 1 0 6 1 1 1
 
0 1 0 0 0 1 1 11 1 1 1 1 6 1 1
 
0 0 1 0 1 0 1 10 1 1 1 1 1 6 1
1 0 0 1 0 1 0 10 0 1 1 1 1 1 6

This matrix is symmetric, R = R0 , and its inverse matrix exists, although the
determinant is the large number, det(R) = 4939060313245. To calculate the
inverse matrix R−1 and its 16 basis functions, we can use the code with script
show matrixIR.m, which is given below.

% call: show_matrixIR.m
N=4;
R=A_rays2R(N);
fprintf(’%16.0f \n’,det(R));
% inverse matrix 16x16
Ri=inv(R);
% represent this matrix as the block-matrix 4x4
NN=N*N;
iR=zeros(N,N,NN);
for k=1:NN
y=reshape(Ri(:,k),N,N); iR(:,:,k)=y;
end
% the 1st basis functions of transforms R and Ri
k=1;
field_function=reshape(R(:,k),N,N); % at (0,0)
cof_function=iR(:,:,k)
y=field_function.*cof_function
340 CHAPTER 6: Method of Particles

sum(sum(y))
% ========= calculated data ========================
% det(R)) = 4939060313245
% field_function
6 1 1 1
1 1 0 0
1 0 1 0
1 0 0 1
% cof_function
0.2001 -0.0262 -0.0210 -0.0209
-0.0262 -0.0120 0.0193 0.0110
-0.0210 0.0193 -0.0259 0.0086
-0.0209 0.0110 0.0086 -0.0266
% y
1.2007 -0.0262 -0.0210 -0.0209
-0.0262 -0.0120 0 0
-0.0210 0 -0.0259 0
-0.0209 0 0 -0.0266
% sum(sum(y))=1
% ==================================================

The following three matrices are printed when running this code:
   
6 1 1 1 0.2001−0.0262−0.0210−0.0209
1 1 0 0   −0.0262−0.0120 0.0193 0.0110
ψ0,0 =  1 0 1 0  , φ0,0 =  −0.0210 0.0193−0.0259 0.0086 ,
  

1 0 0 1 −0.0209 0.0110 0.0086−0.0266

and  
1.2007−0.0262−0.0210−0.0209
 −0.0262−0.0120 0 0
ψ0,0 .× φ0,0 =
 −0.0210
.
0 −0.0259 0
−0.0209 0 0 −0.0266
The first matrix describes the field function of the A-particle (0, 0); the second
matrix is the first basis function of the inverse transform R−1 . This function is
shown in matrix form. One can notice that in this matrix the maximum value
has the coefficient at the same point (0, 0). This matrix probably describes a
particle (0, 0) in another field. The third matrix is the point-wise multiplication
of these two functions. The sum of all coefficients of the last matrix equals 1.
The similar properties are valid for all other particles (n, m).
For instance, by using the same code for k = 11, we obtain the following
matrices for the A-particle (2, 2) :
   
1 0 1 1 −0.0259 0.0217−0.0186−0.0130
0 1 1 1
 , φ2,2 =  0.0217−0.0142−0.0148−0.0144 ,
 
ψ2,2 =  1 1 6 1  −0.0186−0.0148 0.2014−0.0173
1 1 1 1 −0.0130−0.0144−0.0173−0.0120
Image Processing 341
 
−0.0259 0 −0.0186−0.0130
 0−0.0142−0.0148−0.0144 
ψ2,2 .× φ2,2  −0.0186−0.0148 1.2084−0.0173 ,
= 

−0.0130−0.0144−0.0173−0.0120
and
3 X
X 3
ψ2,2 (n, m)φ2,2(n, m) = 1.
n=0 m=0

Thus, the discrete model of image reconstruction by A-rays can be solved

by means of field functions of particles. It is assumed that each A-particle
(n, m) which represents the corresponding square [n, m] on the lattice, con-
tributes equally to all rays passing through the center of this square. Here,
we also assume that the normalization of the sums by the lengths of the rays
inside the square has been considered.
We now consider the case when the particle does not contribute equally to
different projections.
Example 6.2 For the N = 4 case, we consider the A-particle (0, 1), its field
function    
1 1 1 0 1 2 3 0
6 1 1 1 • 5 5 5
 1 1 0 0 ,
ψ0,1 =   and N P =  1 6 0 0 .


1 0 1 0 1 0 6 0
Here, the second matrix shows the way we number A-rays passing through
the point (0, 1). It is clock-wise numbering, starting from the horizontal (0, 1)-
projection and considering the last projection to be the (−1, 1)-projection,
which is the (3, 1)-projection. The (2, 1)-projection is the single-point (0, 1)
and has number 4, but is not shown. We denote the linear sums along these
rays by v1 , v2 , ..., v6. Therefore, the sums of all six projections can be written
as  
v1 v2 v3 0
X 6 
 
 vk v5 v5 v5 
Y = .
 k=1 
 v1 v6 0 0 
v1 0 v6 0
This matrix can be written column-wise into the raw-vector
6
!
X
Y →y= v1 , vk , v1 , v1 , v2 , v5 , v6 , 0, v3, v5 , 0, v6 , 0, v5, 0, 0 ,
k=1

and then the matrix equation y0 = Rx0 can be solved with respect to the
original image-vector x0 = (x0,0 , x1,0, x2,0 , x3,0, x1,0, ..., x3,2, x3,3)0 .
It is expected, that by using this direct method, the reconstructed image
will be smooth. To analyze this reconstruction, we first write the pro-
gram to accomplish the reconstruction by the given equations. Below is
342 CHAPTER 6: Method of Particles

the script projection4.m of this program. Here, V stands for the vector
V = (v1 , v2 , ..., v6).

% -------------------------------------------------
% call: projection4.m
function X=projection4(V)
if (length(V)~=6) return; end
y=zeros(1,16);
y(1)=V(1); y(3)=V(1); y(4)=V(1); y(2)=sum(V);
y(5)=V(2); y(6)=V(5); y(10)=V(5); y(14)=V(5);
y(7)=V(6); y(12)=V(6); y(9)=V(3);
R=A_rays2R(4); x=inv(R)*y’; X=reshape(x’,4,4);
% -------------------------------------------------

If we consider the vector y with equal components, which corresponds

to the case when all sums along A-rays passing through the point (0, 1) are
equal (the case of the discrete model), then we obtain the exact reconstruction.
Indeed, the commands V=ones(1,6); X=projection4(V) result in the following
image:  
0 0 0 0
1 0 0 0
X =0 0 0 0 .


0 0 0 0
We now consider the modified projection data
 
1 1 1 0
 6.2 1 1 1 
Y =  1 1.2 0 0  .


1 0 1.2 0

Here we changed the (−1, 1)-projection along the ray passing through the
squares [0, 1], [1, 2], and [2, 3]. For this case, the vector V = (1, 1, 1, 1, 1, 1.2)
and, therefore, the vector y equals (1, 6.2, 1, 1, 1, 1, 1.2, 0, 1, 1, 0, 1.2, 0, 1, 0, 0).
The commands V1=[1 1 1 1 1 1.2]; X1=projection4(V1) result in the follo-
wing image:
 
0.0003 −0.0064 −0.0117 0.0047
 1.0317 −0.0021 −0.0070 −0.0117
X1 =   −0.0021 0.0330 −0.0021 −0.0064


−0.0144 −0.0021 0.0317 0.0003

   
0 0 0 0 0.0003 −0.0064 −0.0117 0.0047
 1 0 0 0   0.0317 −0.0021 −0.0070 −0.0117
=  0 0 0 0  +  −0.0021 0.0330 −0.0021 −0.0064 .
  

0 0 0 0 −0.0144 −0.0021 0.0317 0.0003

The particle (0, 1) is not reconstructed exactly, as are not the remaining parti-
cles on the lattice, although they have small values (and can be thresholded).
Image Processing 343

The square error of this change equals 0.044. To obtain the full picture of
image reconstruction, it is necessary to analyze other particles as well, and we
will leave details of such analysis to the reader.
In general, the small squares, or image elements, contribute to the line-
integrals of projections with different weights, which are defined by different
characteristics, such as the lengths of intersections of squares with rays. There-
fore we define the concept of a particle with the geometrical field function,
which generalizes the concept of the A-particle.

6.2 Method of G-rays

In this section, we describe the process of summation and distribution of
projection data between the points of the image that lie on the same rays
of the projection. This method will be applied for reconstructing the image
f(x, y) which is represented by image elements in the Cartesian lattice N × N
on the square [0, 1] × [0, 1].
For that, we consider the representation of each G-ray of the (p, s)-
projection by the corresponding set of coordinates (n, m) of image elements, or
small squares through which the ray passes. Our goal is to determine, for each
small square with coordinate (n, m), the corresponding set of G-rays lp,s (t)
that pass through this square. This set of rays will be used in the method
of averaging the projection data. Given the (p, s)-projection, we denote by
ta(p, s; n, m) the number t, which shows the A-ray lp,s (t) passing through the
center of the (n, m)-th square. To calculate the coordinates of these points
on the lattice, we will use the function with script ps Arays points.m which is
given in Chapter 4. It should be noted that in this function, the numbering
of coordinates of points is performed from (0, 0) to (N − 1, N − 1). These sets
of points will be used to determine the set of G-rays lp,s (t) passing through
the (n, m)-th image element in the square [0, 1] × [0, 1], and determine the
corresponding ray-sets, or set of image elements passed through by these rays.

6.2.1 G-rays for the first set of generators

We start with the N = 8 case and the generator (p, s) = (1, 1). A-rays of the
(1, 1)-projection are represented by the following pairs of coordinates (n, m)
344 CHAPTER 6: Method of Particles

on the lattice 8 × 8:
  

 l1,1 (0)  
 (0, 0)
 l (1) 
  
 (1, 0) (0, 1)
 1,1  
  



 l1,1 (2) 






 (2, 0) (1, 1) (0, 2)



 l1,1 (3) 






 (3, 0) (2, 1) (1, 2) (0, 3)



 l1,1 (4) 






 (4, 0) (3, 1) (2, 2) (1, 3) (0, 4)

 l (5) 
 
 (5, 0) (4, 1) (3, 2) (2, 3) (1, 4) (0, 5)

 1,1 
 

  
 l1,1 (6) 
   (6, 0) (5, 1) (4, 2) (3, 3) (2, 4) (1, 5) (0, 6)

l1,1 (7) → (7, 0) (6, 1) (5, 2) (4, 3) (3, 4) (2, 5) (1, 6) (0, 7)
  


 l1,1 (8) 




 (7, 1) (6, 2) (5, 3) (4, 4) (3, 5) (2, 6) (1, 7)
  


 l1,1 (9) 




 (7, 2) (6, 3) (5, 4) (4, 5) (3, 6) (2, 7)
  


 l1,1 (10) 




 (7, 3) (6, 4) (5, 5) (4, 6) (3, 7)
  


 l1,1 (11) 




 (7, 4) (6, 5) (5, 6) (4, 7)
  

 l (12) 
 
 (7, 5) (6, 6) (5, 7)
 1,1
 
 


 l (13) 
 
 (7, 6) (6, 7)
 1,1
 
 

l1,1 (14) (7, 7)

It is not difficult to see that ta(1, 1; n, m) = n + m, n, m = 0 : 7. The rays are

numbered from 0 to L(p, s) = (p + s)(N − 1) = 14. For instance, the point
(0, 5) lies on the A-ray with number ta(1, 1; 0, 5) = 5, and the (4, 6)-square
lies on the A-ray with number ta(1, 1; 4, 6) = 10. For this diagonal projection,
each G-ray l1,1 (t) passes through the squares whose coordinates coincide with
the coordinates of the points that determine the A-ray l1,1 (t). Thus, the G-
ray passing through the (n, m)-th square has the number t = n + m and it is
described by the set of squares with the following coordinates:

l1,1 (t) = {(n0 , m0 ), (n0 − 1, m0 + 1), ..., (n, m), ..., (m0 − 1, n0 + 1), (m0 , n0 )}

where n0 = t and m0 = 0, if t ≤ 7, and n0 = 7 and m0 = t − 7, if t > 7.

Each square is intersected by a single G-ray. For example, if (n, m) = (4, 6),
then t = 4 + 6 = 10 = 7 + 3, and the first square (from the left), which is
intersected by the G-ray l1,1 (10) is the (7, 3)-th square.
We now consider the (1, 2)-projection, i.e., the (p, s) = (1, 2) case. From
the full set of A-rays of the (1, 2)-projection, which is given in §4.8, we consider
a few A-rays and their representation by the coordinates of the points on the
lattice,   
 l1,2 (8) 
   (6, 1) (4, 2) (2, 3) (0, 4)

 l (9)  
1,2 (7, 1) (5, 2) (3, 3) (1, 4)
→
 l1,2 (10)
   2) (4, 3) (2, 4) (0, 5)
 (6,
  
l1,2 (11) (7, 2) (5, 3) (3, 4) (1, 5)
The A-ray passing through the point (n, m) on the lattice has the number
calculated by
ta(1, 2; n, m) = n + 2m, n, m = 0 : 7.
For instance, the point (3, 3) lies on the A-ray l1,2 (t) with number t =
Image Processing 345

ta(1, 2; 3, 3) = 3 + 2(3) = 9, and the point (4, 3) lies on the A-ray l1,2 (10), be-
cause ta(1, 2; 4, 3) = 4 + 2(3) = 10. Figure 6.5 shows the G-rays with numbers
9 and 10, as well as the filled (3, 3)-th square. These two rays are determined

0
l(9)
0.25
l(10)
1 1 2
0.25

2 1 2 1
0.375

3 1 2 1
y and m

0.5

4 1 2 1
0.625

5 1
0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.5
G-rays passing through the (3, 3)-th square in the (1, 2)-projection (case A).

by the squares with the following numbers (written from left to right):

l1,2 (9) ={[0, 4], [1, 4], [2, 3], [3, 3], [4, 2], [5, 2], [6, 1], [7, 1]},
l1,2 (10)={[0, 5], [1, 4], [2, 4], [3, 3], [4, 3], [5, 2], [6, 2], [7, 1]}.

Here, and later on, we use the square brackets in equations to distinguish the
difference between the numbering of points and squares; the (n, m)-th square
is denoted by [n, m]. We can write these two ray-sets as

l1,2 (9)= {[n, m]; n + 2m ∈ {8, 9}},

l1,2 (10)= {[n, m]; n + 2m ∈ {9, 10}}.

These rays pass together through the squares [1, 4], [3, 3], [5, 2], and [7, 1]. In
other words, these four squares are intersected twice by these G-rays. The
remaining squares from these ray-sets are intersected only once, as shown in
the figure. These coefficients of the intersections c(n, m) can be calculated by

c(n, m) = 2 − |(n + 2m) − 9| = 2 − |(n − 3) + 2(m − 3)|,

346 CHAPTER 6: Method of Particles

considering only the squares of rays l1,2 (9) and l1,2 (10). Here 2 stands for s,
and 9 is the number of the first G-rays passing through the square [3, 3].
In general, the G-ray l1,2 (t) can be defined by the following set of squares:

l1,2 (t) = {[n, m]; n + 2m ∈ {t − 1, t}}, t = 0 : 21.

We consider the (p, s) = (1, 3) case. The set of A-rays of the (1, 3)-
projection contains 29 rays. We consider only the following three A-rays and
their representations by points (n, m) on the lattice:
 
 l1,3 (12)  (6, 2) (3, 3) (0, 4)
l1,3 (13) → (7, 2) (4, 3) (1, 4)
 
l1,3 (14) (5, 3) (2, 4)

The A-ray passing through the point (n, m) has the number which is calculated
by ta(1, 3; n, m) = n + 3m, n, m = 0 : 7. To verify this formula, we consider
the points (3, 3) and (7, 2). The first point is on the A-ray l1,3 (t) with number

0.25

1 1 l(12)
0.25
l(13)
l(14)
2 1 2 3 2
0.375

3 1 2 3 2 1
y and m

0.5

4 3 2 1
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.6
G-rays passing through the square [3, 3] in the (1, 3)-projection (case A).

ta(1, 3; 3, 3) = 3 + 3(3) = 12. This is the number of the first G-ray passing
through the (3, 3)-th square. Figure 6.6 shows the G-rays with numbers 12, 13,
and 14, which pass through the (3, 3)-th square. The point (7, 2) is on the A-
ray l1,3 (13), and ta(1, 3; 7, 2) = 7 + 3(2) = 13. This is the number of the first
Image Processing 347

G-ray passing through the (7, 2)-th square. This square is also on the G-ray
l1,3 (14). We now consider the square-representation of these three G-rays,

l1,3 (12)= {[0, 4], [1, 3], [2, 3], [3, 3], [4, 2], [5, 2], [6, 2], [7, 1]}
l1,3 (13)= {[0, 4], [1, 4], [2, 3], [3, 3], [4, 3], [5, 2], [6, 2], [7, 2]}
l1,3 (14)= {[0, 4], [1, 4], [2, 4], [3, 3], [4, 3], [5, 3], [6, 2], [7, 2]}

and for each of these rays, we have the following description:

l1,3 (t) = {[n, m]; n + 3m ∈ {t − 2, t − 1, t}}, t = 12, 13, 14.

These three rays together pass through the squares [0, 4], [3, 3], and [6, 2].
Other squares are intersected once or twice by the G-rays, as shown in the
figure. These coefficients of the intersections c(n, m) can be calculated by

c(n, m) = 3 − |(n + 3m) − 12| = 3 − |(n − 3) + 3(m − 3)|,

when considering only the three rays l1,3 (12), l1,3 (13), and l1,3 (14). Here, 3
stands for s, and 12 is the number of the first G-rays passing through the
square [3, 3]. The square-set representation of all G-rays of this projection is
described by

l1,3 (t) = {[n, m]; n + 3m ∈ {t − 2, t − 1, t}}, t = 0 : 21.

We consider the next (p, s)-projection, when (p, s) = (1, 4). G-rays of this
projection can be described by the square-set similar to the cases (p, s) = (1, 2)
and (1, 3). The unique A-ray l1,4 (t) of this projection, which intersects the
point (n, m) on the lattice, has the number t = n + 4m. Let us consider as an
example, the point (3, 3), which is on the A-ray with number 15. This A-ray
passes through the point (7, 2) on the lattice. Figure 6.7 shows four G-rays,
starting with the ray l1,4 (15), which is the first ray passing through this point.
It is not difficult to see from this figure that these rays can be described
by the following squares:

l1,4 (15)={[0, 3], [1, 3], [2, 3], [3, 3], [4, 2], [5, 2], [6, 2], [7, 2]}
l1,4 (16)={[0, 4], [1, 3], [2, 3], [3, 3], [4, 3], [5, 2], [6, 2], [7, 2]}
l1,4 (17)={[0, 4], [1, 2], [2, 3], [3, 3], [4, 3], [5, 3], [6, 2], [7, 2]}
l1,4 (18)={[0, 4], [1, 2], [2, 4], [3, 3], [4, 3], [5, 3], [6, 3], [7, 2]}.

We also can write that

l1,4 (15) = {[n, m]; n + 4m ∈ {12, 13, 14, 15}}

and, in general,

l1,4 (t) = {[n, m]; n + 4m ∈ {t − 3, t − 2, t − 1, t}}.

These four rays together pass through the squares [3, 3] and [7, 2]. Other
348 CHAPTER 6: Method of Particles

0.25

0.25 l(15)

2 1 2 3 4 l(17)

0.375

3 1 2 3 4 3 2 1
y and m

0.5

4 3 2 1
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.7
G-rays passing through the square [3, 3] in (1, 4)-projection (case A).

squares are intersected once, twice, or three times by the G-rays, as shown in
the figure. The coefficients of the intersections c(n, m) for other squares can
be calculated by

c(n, m) = 4 − |(n + 4m) − 15| = 4 − |(n − 3) + 4(m − 3)|,

when considering only these four rays. Here, 4 stands for s, and 15 is the
number of the first G-rays passing through the square [3, 3].
Thus, for the G-rays of the (1, s)-projection, when 1 ≤ s ≤ 4, the following
holds:
l1,s (t) = {[n, m]; n + sm ∈ {t − s − 1, ..., t − 1, t}}, t = 0 : 7(1 + s),
(6.14)
c(n, m) = s − |(n − 3) + s(m − 3)|, n, m = 0 : 7,

and ta(1, s; n, m) = n + sm.

6.2.2 G-rays for the second set of generators

We now consider the (p, s) = (1, 5) case. In this case, s = 5 > N/2 = 4 and
the (1, −3)-projection is used. We consider the A-ray of the projection that
intersects the point (3, 3) of the lattice. As shown in Chapter 4, this A-ray
l1,−3 (t) has the number calculated by t = n + 3(7 − m) = 3 + 3(4) = 15, when
Image Processing 349

using the positive numbering, and the number 15 − 21 = −6 in the original

numbering of the rays. This A-ray also passes through the points (0, 2) and
(6, 4) in the lattice, and we can write l1,−3 (−6) = {(0, 2), (3, 3), (6, 4)}.
Figure 6.8 shows three G-rays which pass through the (3, 3)-th square. The
ray l1,4 (−6) is considered the first ray that passes through this square. The
other two G-rays are l1,4 (−5) and l1,4 (−4). In positive numbering, the ray
l1,4 (−5) is l1,4 (16), and the ray l1,4 (−4) is l1,4 (17).

0.25

0.25

2 3 2 1
0.375

3 1 2 3 2 1
y and m

0.5

4 1 2 3 2
l(−4)
0.625 l(−5)
l(−6)
5 1
0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.8
G-rays passing through the square [3, 3] for the (1, −3)-projection (case A).

The following square-set representation holds for these rays:

l1,−3 (−4)= {[0, 2], [1, 2], [2, 2], [3, 3], [4, 3], [5, 3], [6, 4], [7, 4]}
l1,−3 (−5)= {[0, 2], [1, 2], [2, 3], [3, 3], [4, 3], [5, 4], [6, 4], [7, 4]}
l1,−3 (−6)= {[0, 2], [1, 3], [2, 3], [3, 3], [4, 4], [5, 4], [6, 4], [7, 3]}.
These rays can also be written as
l1,−3 (−4) = {[n, m]; n + 4m ∈ {15, 16, 17}}
l1,−3 (−5) = {[n, m]; n + 4m ∈ {14, 15, 16}}
l1,−3 (−6) = {[n, m]; n + 4m ∈ {13, 14, 15}}.

For positive numbering, when t → t̄ = t+21, we can use the following equation
for these rays:

l1,−3 (t) = l1,−3 (t̄) = {[n, m]; n + 3(7 − m) ∈ {t̄ − 2, t̄ − 1, t̄}}.

350 CHAPTER 6: Method of Particles

These three rays together pass through the squares [0, 2], [3, 3], and [6, 4].
Other squares are intersected once or twice by these G-rays, as shown in the
figure. The coefficients of the intersections c(n, m) for other squares can be
calculated by
c(n, m) = 3 − |(n + 3(7 − m) − 15| = 3 − |(n − 3) + 3(7 − (m − 3)) − 21|,
when considering only these three rays. Here 3 stands for −s, and 15 is the
number of the first G-rays passing through the square [3, 3].
We consider the next projection, the (1, −2)-projection, when (p, s) =
(1, 6). Two G-rays of this projection, which pass through the square [3, 3], are
shown in Figure 6.9. These rays have numbers −2 and −3, respectively, or 12

0.25

1 1
0.25

2 1 2 1
0.375

3 1 2 1
y and m

0.5

4 1 2 1
0.625

5 1 2 l(−2)=l(12)
0.75 l(−3)=l(11)
6

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.9
G-rays pass the square [3, 3] in the (1, −2)-projection (case A).

and 11, when numbering after the transform t → t̄ = t + 14. The number 11
is calculated by 3 + 2(7 − 3), and the 11-th G-ray is considered to be the first
ray intersecting the square [3, 3]. The square set representation of these two
rays is described by
l1,−2 (−2)= {[0, 1], [1, 2], [2, 2], [3, 3], [4, 3], [5, 4], [6, 4], [7, 5]}
l1,−2 (−3)= {[0, 2], [1, 2], [2, 3], [3, 3], [4, 4], [5, 4], [6, 5], [7, 5]}.
We can also write the following:
l1,−2 (−2) ={[n, m]; n + 2(7 − m) ∈ {11, 12}}
l1,−2 (−3) ={[n, m]; n + 2(7 − m) ∈ {10, 11}}.
Image Processing 351

In general, for positive numbering of the rays, we write these two rays as

l1,−2 (t) = l1,−2 (t̄) = {[n, m]; n + 2(7 − m) ∈ {t̄ − 1, t̄}}.

The rays together pass through the squares [1, 2], [3, 3], [5, 4], and [7, 5]. Other
squares are intersected once by these G-rays, as shown in the figure. The
coefficients of the intersections c(n, m) for other squares can be calculated by

c(n, m) = 2 − |(n + 2(7 − m) − 11| = 2 − |(n − 3) + 2(7 − (m − 3)) − 14|,

when considering only these two rays. Here, 2 stands for −s, and 11 is the
number of the first G-rays passing through the square [3, 3].
For the (−1, 1)-projection, the G-rays are described by the following
squares:
  

 l(0)  
 [7, 0]

 l(1)   
 [6, 0] [7, 1]

 
 


 l(2)   
 [5, 0] [6, 1] [7, 2]

 
 

  


 l(3) 




 [4, 0] [5, 1] [6, 2] [7, 3]
  


 l(4) 




 [3, 0] [4, 1] [5, 2] [6, 3] [7, 4]
  


 l(5) 




 [2, 0] [3, 1] [4, 2] [5, 3] [6, 4] [7, 5]
  
 l(6) 
   [1, 0] [2, 1] [3, 2] [4, 3] [5, 4] [6, 5] [7, 6]

{l−1,1 (t)} = l(7) → [0, 0] [1, 1] [2, 2] [3, 3] [4, 4] [5, 5] [6, 6] [7, 7]
  


 l(8) 




 [0, 1] [1, 2] [2, 3] [3, 4] [4, 5] [5, 6] [6, 7]
  


 l(9) 




 [0, 2] [1, 3] [2, 4] [3, 5] [4, 6] [5, 7]
  


 l(10) 




 [0, 3] [1, 4] [2, 5] [3, 6] [4, 7]
  


 l(11) 




 [0, 4] [1, 5] [2, 6] [3, 7]
  


 l(12) 




 [0, 5] [1, 6] [2, 7]
  


 l(13) 




 [0, 6] [1, 7]
  
l(14) [0, 7]

The number of the G-ray passing through the square [n, m] is defined by the
same formula as for the A-ray passing through the point (n, m) on the grid,

ta(7, 1; n, m) = ta(−1, 1; n, m) = (7 − n) + m, n, m = 0 : 7.

For example, the square [3, 3] is in the same direction as the G-ray number
(7 − 3) + 3 = 7. In this diagonal projection, each square is intersected by only
one G-ray.

6.2.3 G-rays for the third set of generators

For the (2, 1)-projection, two G-rays passing through the (3, 3)-th square are
shown in Figure 6.10. The first G-ray has the number 3(2) + 3(1) = 9. These
rays are determined by the following squares:
l2,1 (9) ={[1, 7], [1, 6], [2, 5], [2, 4], [3, 3], [3, 2], [4, 1], [4, 0]}
l2,1 (10)={[1, 7], [2, 6], [2, 5], [3, 4], [3, 3], [4, 2], [4, 1], [5, 0]}.
352 CHAPTER 6: Method of Particles

l(9) l(10)

0 1 1
0.25

1 2
0.25

2 1 1
0.375

3 2
y and m

0.5

4 1 1
0.625

5 2
0.75

6 1 1
0.875

7 2
1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.10
G-rays passing through the square [3, 3] in the (2, 1)-projection (case A).

We can also write l2,1 (9) = {[n, m]; 2n + m ∈ {8, 9}}, and in general

l2,1 (t) = {[n, m]; 2n + m ∈ {t − 1, t}}.

The coefficients of the intersections c(n, m) of squares along these two G-rays
with number 8 and 9 can be calculated by

c(n, m) = 2 − |(n + 2m − 9| = 2 − |2(n − 3) + (m − 3))|.

Here 2 stands for p, and 9 is the number of the first G-rays passing through
the square [3, 3].
The above given description of the G-rays corresponds to case A, when
the A-rays are shifted by t → t − 0.5. In our code, for the (p, 1)-th projections,
when p ≤ N/2, we use case B, when the A-rays are shifted by t → t + 0.5.
Therefore, the first ray passing through the (3, 3)-th square has number 8, and
therefore, l2,1 (8) = {[n, m]; 2n + m ∈ {8, 9}}. In general for case B, the G-ray
number t can be described by

l2,1 (t) = {[n, m]; 2n + m ∈ {t, t + 1}}.

For the (4, 1)-projection, the first of four G-rays intersecting the square
[3, 3] has the number 4(3) + 1(3) = 15. In Figure 6.11, these four rays are
Image Processing 353

l(15) l(17)

0 1 3
0.25

1 2 2
0.25

2 3 1
0.375

3 4
y and m

0.5

4 1 3
0.625

5 2 2
0.75

6 3 1
0.875

7 4
1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.11
G-rays passing through (3, 3) for the (4, 1)-projection (case A).

shown. The first G-ray can be written by squares as

l4,1 (15)= {[2, 7], [2, 6], [2, 5], [2, 4], [3, 3], [3, 2], [3, 1], [3, 0]}

or l4,1 (15) = {[n, m]; 4n + m ∈ {12, 13, 14, 15}}. In general, the G-rays are
described similarly,

l4,1 (t) = {[n, m]; 4n + m ∈ {t − 3, t − 2, t − 1, t}}, t = 15 : 18.

The coefficients of the intersections c(n, m) of squares along these four G-rays
can be calculated by

c(n, m) = 4 − |(4n + m − 15| = 4 − |4(n − 3) + (m − 3))|.

Here 4 stands for p, and 15 is the number of the first G-rays passing through
the square [3, 3].
The numbering of the G-rays in the figure corresponds to case A, when A-
rays are shifted by t → t − 3/2. In case B, the shift of A-rays is by t → t + 3/2,
and ray number 15 has number 12, and ray number 17 is 14. Therefore, we
can write that

l4,1 (12) = {[n, m]; 4n + m ∈ {12, 13, 14, 15}}.

354 CHAPTER 6: Method of Particles

In general case B, the representation of the G-rays by squares is described by

l4,1 (t) = {[n, m]; 4n + m ∈ {t, t + 1, t + 2, t + 3}}, t = 0 : 35.

6.2.4 G-rays for the fourth set of generators

The last projection to be described is the (6, 1)-projection, which is conside-
red as the (−2, 1)-projection. Two G-rays intersecting the square [3, 3] have
numbers −3 and −2 and, are shown in Figure 6.12. The first G-ray passing

l(−2) l(−3)
0

0 1 1
0.25

1 2
0.25

2 1 1
0.375

3 2
y and m

0.5

4 1 1
0.625

5 2
0.75

6 1 1
0.875

7 2
1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 6.12
G-rays passing through the square [3, 3] in the (−2, 1)-projection (case A).

through the square is the ray with number −2(3) + 1(3) = −3. The square
representations of this ray and the next ray l(−2) are the following:

l−2,1 (−3)= {[2, 0], [2, 1], [3, 2], [3, 3], [4, 4], [4, 5], [5, 6], [5, 7]}
l−2,1 (−2)= {[1, 0], [2, 1], [2, 2], [3, 3], [3, 4], [4, 5], [4, 6], [5, 7]}.

We also can write that

l−2,1 (−3) = {[n, m]; 2(7 − n) + m ∈ {10, 11}}
l−2,1 (−2) = {[n, m]; 2(7 − n) + m ∈ {11, 12}}.

For positive numbering, when t → t̄ + 14, the rays with numbers −3 and −2
have numbers 11 and 12, respectively, and the above representations can be
Image Processing 355

written as
l1,−2 (t) = l1,−2 (t̄) = {[n, m]; 2(7 − n) + m ∈ {t̄ − 1, t̄}},
where t̄ = 11 and 12. In general, this representation holds for all G-rays of the
(−2, 1)-projection. It is clear, that for the horizontal and vertical projections,
i.e., for the (1, 0)- and (0, 1)-projections, each square [n, m], including [3, 3], is
intersected by only one G-ray.

6.2.5 Map of projections for one square

Considering all 12 projections for the case N = 8, and the number of G-rays
in each of these projections that pass through the square [3, 3], we obtain the
following matrix:
 
1 1 1 2 4 1 1 0
1 1 2 3 4 1 1 3
 
4 4 4 6 6 6 7 6
 
 2 5 10 26 10 5 2 1 
G(3, 3) = 
 . (6.15)
7 6 6 6 4 4 4 2

1 1 4 3 2 1 1 3
 
1 1 4 2 1 1 1 0
0 2 4 1 0 2 0 1
We call this matrix the map of projections for square [3, 3]. The coefficient
cn,m of this matrix equals the number of G-rays passing through the squares
[n, m] and [3, 3] at the same time. If there is no G-ray passing through the
squares [n, m] and [3, 3] at the same time, then cn,m = 0. There are 26 rays
passing through the square [3, 3], which all together pass through 46 squares.
The set descriptions of all 12 rays passing through the square [3, 3] are
given in Table 6.5. The coordinates of squares [n, m], which together with the
square [3, 3] pass through by G-rays, are shown in the table.
It can be seen from the matrix in (6.15), that the coefficient c(3, 3) = 26 is
in the center for the sub-matrix of size 7 × 7, and this sub-matrix is symmetric
with respect to this center. It is interesting to note the following property of
the frame-sets of this matrix. The k-th frame-set at center (3, 3) is defined as
the following set of points:
Υk = Υk (n0 , m0 ) = {(n, m); max(|n − n0 |, |m − m0 |) = k}, k > 1.
The values of the map of projections C(3, 3) on the first three frame-sets have
1 1 1 2 4 1 1 
 
1 2 34 1 1 1
" #
4 6 6 4 6 4
 7
 
10 10 , 5 5 , 2
 · 2,
6 6 4 6 4 7 3
 
1 4 32 1 1 1
1 1 4 2 11 1
356 CHAPTER 6: Method of Particles
TABLE 6.5
Coordinates of the map of projection for square [3, 3]
(p, s) squares [n, m] card g(p, s)
(1, 0) [3, 0] [3, 1] [3, 2] [3, 3] [3, 4] [3, 5] [3, 6] [3, 7] 8 1
(1, 1) [6, 0] [5, 1] [4, 2] [3, 3] [2, 4] [1, 5] [0, 6] 7 1
(1, 2) [0, 4] [1, 4] [2, 3] [3, 3] [4, 2] [5, 2] [6, 1] [7, 1] 8 1
[0, 5] [1, 4] [2, 4] [3, 3] [4, 3] [5, 2] [6, 2] [7, 1] 8 1
(1, 3) [0, 4] [1, 3] [2, 3] [3, 3] [4, 2] [5, 2] [6, 2] [7, 1] 8 1
[0, 4] [1, 4] [2, 3] [3, 3] [4, 3] [5, 2] [6, 2] [7, 2] 8 1
[0, 4] [1, 4] [2, 4] [3, 3] [4, 3] [5, 3] [6, 2] [7, 2] 8 1
(1, 4) [0, 3] [1, 3] [2, 3] [3, 3] [4, 2] [5, 2] [6, 2] [7, 2] 8 1
[0, 4] [1, 3] [2, 3] [3, 3] [4, 3] [5, 2] [6, 2] [7, 2] 8 1
[0, 4] [1, 2] [2, 3] [3, 3] [4, 3] [5, 3] [6, 2] [7, 2] 8 1
[0, 4] [1, 2] [2, 4] [3, 3] [4, 3] [5, 3] [6, 3] [7, 2] 8 1
(1, −3) [0, 2] [1, 2] [2, 2] [3, 3] [4, 3] [5, 3] [6, 4] [7, 4] 8 1
[0, 2] [1, 2] [2, 3] [3, 3] [4, 3] [5, 4] [6, 4] [7, 4] 8 1
[0, 2] [1, 3] [2, 3] [3, 3] [4, 4] [5, 4] [6, 4] [7, 3] 8 1
(1, −2) [0, 1] [1, 2] [2, 2] [3, 3] [4, 3] [5, 4] [6, 4] [7, 5] 8 1
[0, 2] [1, 2] [2, 3] [3, 3] [4, 4] [5, 4] [6, 5] [7, 5] 8 1
(1, −1) [7, 7] [6, 6] [5, 5] [4, 4] [3, 3] [2, 2] [1, 1] [0, 0] 8 1
(0, 1) [0, 3] [1, 3] [2, 3] [3, 3] [4, 3] [5, 3] [6, 3] [7, 3] 8 1
(2, 1) [1, 7] [1, 6] [2, 5] [2, 4] [3, 3] [3, 2] [4, 1] [4, 0] 8 1
(1, 7] [2, 6] [2, 5] [3, 4] [3, 3] [4, 2] [4, 1] [5, 0] 8 1
(4, 1) [2, 7] [2, 6] [2, 5] [2, 4] [3, 3] [3, 2] [3, 1] [3, 0] 8 1
[2, 7] [2, 6] [2, 5] [3, 4] [3, 3] [3, 2] [3, 1] [4, 0] 8 1
[2, 7] [2, 6] [3, 5] [3, 4] [3, 3] [3, 2] [4, 1] [4, 0] 8 1
[2, 7] [3, 6] [3, 5] [3, 4] [3, 3] [4, 2] [4, 1] [4, 0] 8 1
(−2, 1) [2, 0] [2, 1] [3, 2] [3, 3] [4, 4] [4, 5] [5, 6] [5, 7] 8 1
[1, 0] [2, 1] [2, 2] [3, 3] [3, 4] [4, 5] [4, 6] [5, 7] 8 1

the same sum of values on the frame-sets: 2 · 26 = 52.

The total number of intersections of G-rays with the (3, 3)-th square is 26,
and each of these G-rays intersects this square twice. The next frame-set has
 
1 0 2 0 1 4 2 01
0 0
 
3 3
 
2 6
 
Υ4 =  1 1

6 2
 
3 3
 
0 0
1 0 2 4 1 0 2 01

the same sum of values 2 · 26 = 52. One also can notice that all frame-sets are
the same for two maps of projections C(3, 3) and C(4, 4).
We now extend the map C(3, 3) to the square matrix 9 × 9 with the 4-th
Image Processing 357

frame,  
10 2 0 1 4 2 0 1
01 1 1 2 4 1 1 0
 
31 1 2 3 4 1 1 3
 
24 4 4 6 6 6 7 6
 
 1 2 5 10 26 10 5 2 1  .
U9,9 (26) =  (6.16)

67 6 6 6 4 4 4 2
 
31 1 4 3 2 1 1 3
 
01 1 4 2 1 1 1 0
10 2 4 1 0 2 0 1
We call this matrix the UH-square 9 × 9 of the number 26. In our study, we
also use the UH-square 15 × 15 of the number 26, which equals
 
1 0 0 1 1 0 0 1 1 3 1 1 0 0 1
0 1 0 0 2 0 0 1 2 2 2 0 0 1 0
 
0 0 1 0 1 1 0 1 3 2 1 0 1 0 0
 
1 0 0 1 0 2 0 1 4 2 0 1 0 0 1
 
2 2 1 0 1 1 1 2 4 1 1 0 1 2 2
 
2 3 3 3 1 1 2 3 4 1 1 3 4 5 5
 
0 0 1 2 4 4 4 6 6 6 7 6 4 2 1
 
 1 1 1 1 2 5 10 26 10 5 2 1 1 1 1  .
U15,15 (26) =  (6.17)

1 2 4 6 7 6 6 6 4 4 4 2 1 0 0
 
5 5 4 3 1 1 4 3 2 1 1 3 3 3 2
 
2 2 1 0 1 1 4 2 1 1 1 0 1 2 2
 
1 0 0 1 0 2 4 1 0 2 0 1 0 0 1
 
0 0 1 0 1 2 3 1 0 1 1 0 1 0 0
 
0 1 0 0 2 2 2 1 0 0 2 0 0 1 0
1 0 0 1 1 3 1 1 0 0 1 1 0 0 1
In general, the UH-square of the number M is such a square matrix, that all
frame-sets of this matrix satisfy the following property: the sum of coefficients
on the frame-set equals 2M.
Thus, for the square [3, 3] we constructed the matrix C(3, 3) with coeffi-
cients c(n, m) = c(3, 3, n, m) defining the number of G-rays passing through
the squares [3, 3] and [n, m] at the same time. The coefficient c(3, 3) is consi-
dered to be the center of this matrix. In general, for a given square [n0 , m0 ]),
we can construct a similar matrix with the center in (n0 , m0 ). For instance,
for the square [4, 4], we obtain the following map of projections, or the matrix:
 
1 0 2 0 1 4 2 0
0 1 1 1 2 4 1 1
 
3 1 1 2 3 4 1 1
 
2 4 4 4 6 6 6 7
C(4, 4) = 
 1 2 5 10 26 10 5 2 
 (6.18)
 
6 7 6 6 6 4 4 4
 
3 1 1 4 3 2 1 1
0 1 1 4 2 1 1 1
358 CHAPTER 6: Method of Particles

with the coefficient c(4, 4, 4, 4) = 26 at the center. The sum of all coefficients
in this matrix equals 207 as for the matrix C(3, 3), and it is a sub-matrix of
the UH-square 15 × 15.

Definition 6.2 The geometrical particle, or G-particle (n0 , m0 ) of dimen-

sion N is called the square [n0 , m0 ], which is described by the matrix
||c(n0 , m0 , n, m)|| defining the number of G-rays passing through the squares
[n0 , m0 ] and [n, m] at the same time. We call the two-dimensional function
φn0,m0 (n, m) = c(n0 , m0 , n, m), where n, m = 0 : (N − 1), the field function of
this particle.

Hereinafter we consider the set of projections which are defined by the tensor
representation of the discrete image. Therefore, G-particles are referred to as
the tensor G-particles. For the same reason, the A-particles defined earlier are
referred to as the tensor A-particles.

Example 6.3 In the N = 4 case, the field functions of the G-particles (1, 1)
and (2, 1) are
   
3 2 1 0 2 1 0 1
2 8 2 1 3 2 1 0
φ1,1 = 
1
 and φ2,1 = . (6.19)
2 3 2 2 8 2 1
0 1 2 1 1 2 3 2

Figure 6.13 illustrates two G-particles (1, 1) and (2, 1) and their field func-
tions. The number of G-rays passing through each G-particle equals 8. One

G−particle (1,1) G−particle (2,1)

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 0 0.5 1

FIGURE 6.13 (See color insert)

The field functions of two G-particles (1, 1) and (2, 1).

can notice that the picture of the intersection of G-rays with each square [n, m]
is the same. It is the geometric representation, or symbol, of the G-particle. To
have such a “uniform” set of rays, we add a few rays for each (p, s)-projection,
when (p, s) 6= (0, 1), (1, 0), (1, 1) and (−1, 1).
Image Processing 359
TABLE 6.6
Field functions of G-particles
8 2 1 1 2 8 2 1 1 2 8 2 1 1 2 8
2 3 2 1 1 2 3 2 0 1 2 3 0 0 1 2
1 2 1 1 0 1 2 1 1 0 1 2 0 1 0 1
1 1 1 1 0 1 1 1 0 0 1 1 1 0 0 1
2 1 0 0 3 2 1 0 2 3 2 1 1 2 3 2
8 2 1 1 2 8 2 1 1 2 8 2 1 1 2 8
2 3 3 1 1 2 3 2 0 1 2 3 0 0 1 2
1 2 1 1 0 1 2 1 1 1 1 2 0 1 0 1
1 0 1 0 2 1 0 1 1 2 1 0 1 1 2 1
2 1 0 0 3 2 1 0 2 3 2 1 1 2 3 2
8 2 1 1 2 8 2 1 1 2 8 2 1 1 2 8
2 3 2 1 1 2 3 2 0 1 2 3 0 0 1 2
1 0 0 1 1 1 0 0 1 1 1 0 1 1 1 1
1 0 1 0 2 1 0 1 1 2 1 0 1 1 2 1
2 1 0 0 3 2 1 0 2 3 2 1 1 2 3 2
8 2 1 1 2 8 2 1 1 2 8 2 1 1 2 8

We can write the field functions of all 16 G-particles in the lattice 4 × 4 in

the form of the block-table [ψn,m ]n,m=0:3 , as shown in Table 6.6.
This table, or block-table 4 × 4 with all field functions can be packed into
the following matrix of size 7 × 7, which we call the base matrix of fields of
G-particles:  
1 1 1 10 0 1
1 1 2 1 0 1 0
 
1 2 3 2 1 0 0
 
Ψ= 1 1 2 8 2 1 1 .
 (6.20)
0 0 1 2 3 2 1
 
0 1 0 1 2 1 1
1 0 0 11 1 1
Indeed, it is not difficult to notice that the following equality holds:

ψn,m (i, j) = Ψ(3 − n + i, 3 − m + j), i, j = 0 : 3, (6.21)

for any A-particle (n, m). The base matrix is symmetric, i.e. Ψ(i, j) = Ψ(j, i),
for i, j = 0 : 6. One can notice that the base matrix is the UH-square 7 × 7 of
number 8.
All 16 field functions of G-particles can be represented in the form of 1-D
vectors of length 16 each, and then written into the matrix 16 × 16. The result
360 CHAPTER 6: Method of Particles

of this operation is the following matrix:

 
8 2 1 1 2 3 2 1 1 2 1 1 1 1 1 1
2 8 2 1 1 2 3 2 0 1 3 1 0 1 1 1
 
1 2 8 2 0 1 2 3 1 0 1 2 0 0 1 1
 
1 1 2 8 0 0 1 2 0 1 0 1 1 0 0 1
 
2 1 0 0 8 2 1 1 2 3 2 1 1 2 1 1
 
3 2 1 0 2 8 2 1 1 2 3 2 0 1 2 1
 
2 3 2 1 1 2 8 2 0 1 2 3 1 0 1 2
 
1 2 3 2 1 1 2 8 0 0 1 2 0 1 0 1
R= 1 0 1 0 2 1 0 0 8
. (6.22)
 2 1 1 2 3 2 1
2 1 0 1 3 2 1 1 2 8 2 1 1 2 3 2
 
1 2 1 0 2 3 2 1 1 2 8 2 0 1 2 3
 
1 1 2 1 1 2 3 2 1 1 2 8 0 0 1 2
 
1 0 0 1 1 0 1 0 2 1 0 0 8 2 1 1
 
1 1 0 0 2 1 0 1 3 2 1 0 2 8 2 1
 
1 1 1 0 1 2 1 0 2 3 2 1 1 2 8 2
1 1 1 1 1 1 2 1 1 2 3 2 1 1 2 8

This matrix is symmetric, R = R0 , its inverse matrix exists, and the de-
terminant is det(R) = 17023383606024.

In general, the field functions of the G-particle can be calculated by means

of the base field function as

ψn,m (i, j) = Ψ(N − 1 − n + i, N − 1 − m + j), i, j = 0 : N − 1, (6.23)

where n, m ∈ {0, 1, ..., N −1}. The size of the base matrix is (2N −1)×(2N −1)
and the matrix is symmetric. The base field function for the N = 8 case is
given in (6.17), which is the UH-square 15 × 15 of number 26. This number,
as the central coefficient of the base field function, can be calculated by
3
Ψ(N − 1, N − 1) = 2 + N 2 , N = 4, 8, 16, 32, ... .
8
6.2.5.1 Codes for particles
We present a few codes that were used to calculate all matrices that were
described above for the N = 4 and 8 examples. First, we consider the script
UH square.m of the program for calculating the UH-squares, for the cases when
N = 2r , r > 1. Here, the UH-squares are defined for the set of generators
(p, s), which is used in the tensor and paired transforms. Therefore, we call
such UH-squares the tensor UH-squares.
The UH-square has size (2N − 1) × (2N − 1); therefore, we put this square
into the square [−1, 1] × [−1, 1] with the lattice (2N − 1) × (2N − 1). The
central point of the UH-square is in the central square [0, 0]. The UH-square
defines all G-rays passing through the square [0, 0].
Image Processing 361

% call: UH_square.m / M. Grigoryan, 2011

% The tensor UH-square (2N-1)x(2N-1), for N=2,4,8,16, ...
function Q=UH_square(N)
N2=2*N; N12=N/2;
delta=2/(N2-1); delta2=delta/2;
ps=ps_generators(N); L=size(ps,1);
Q=zeros(N2-1);
for i1=1:L
p=ps(i1,1); s=ps(i1,2);
if p==1
if (s>N12) s=-N+s; end
t_min=-(abs(s)+1)/N2; k=1;
if (s~=0) k=1:abs(s); end
else
if (p>N12) p=-N+p; end
t_min=-(abs(p)+1)/N2;
if (p==0) k=1; else k=1:abs(p); end
end
t=t_min+k/N;
if (s==0)
Q(:,N)=Q(:,N)+ones(N2-1,1);
elseif (p==0)
Q(N,:)=Q(N,:)+ones(1,N2-1);
else
if (s==1)
for n1=1:length(k)
for xi=-1+delta2:delta:1
n=ceil(xi/delta-0.5)+N;
m=N-ceil((t(n1)-xi)/delta/p+0.5)+1;
Q(m,n)=Q(m,n)+1;
end
end
else
for n1=1:length(k)
for yi=-1+delta2:delta:1
m=N-ceil(yi/delta+0.5)+1;
n=ceil((t(n1)-yi)/delta/s-0.5)+N;
Q(m,n)=Q(m,n)+1;
end
end
end
end
end

Figure 6.14 illustrates the UH-square 15 × 15 of number 26 and all G-rays

passing through the central component of the square, which is the square
1/8 × 1/8. This central component is shown separately on the right. The UH-
square is in the square [−1, 1] × [−1, 1]. All field functions are defined from
this UH-square by sliding the window of size 1 × 1.
362 CHAPTER 6: Method of Particles

UH−square of number 26 central square

1 0.06

0.04
0.5
0.02

0 0

−0.02
−0.5
−0.04

−1 −0.06
−1 −0.5 0 0.5 1 −0.05 0 0.05

FIGURE 6.14 (See color insert)

Map of G-rays in the UH-square of number 26 and its central component.

Figure 6.15 illustrates the UH-square 7 × 7 of number 8 and its central

component with all eight G-rays passing through this component.

UH−square of number 8 central square

1
0.1

0.5
0.05

0 0

−0.05
−0.5

−0.1
−1
−1 −0.5 0 0.5 1 −0.1 −0.05 0 0.05 0.1

FIGURE 6.15
Map of G-rays in the UH-square of number 8 and its central component.

For reconstruction of the image f(x, y) with high resolution, i.e. as the
discrete image fn,m , where n, m = 0 : N1 − 1, M1 − 1 and N1 , M1 > N, it
will be desired to reconstruct the values of the image at all or a few points of
intersections of G-rays inside the central square, as well as other squares.
To calculate the block-table [ψn,m ]n,m=0:(N−1) of the field functions of the
G-particles in the lattice 4 ×4, we use the code tableR FFofGrays.m. The script
Image Processing 363

is given below for the N = 4 example. Here, the 2-D array UH8 stands for the
base matrix.
% call: tableR_FFofGrays.m / A.-M. Grigoryan, 2011
N=4;
UH8=UH_square(N); % Base matrix, UH-square (2N-1)x(2N-1)
TR=zeros(N*N); % For the table of all field functions
R2=zeros(N);
m=1;
for i1=1:N
n=1;
for j1=1:N
R2=UH8(N-j1+1:N-j1+N,N-i1+1:N-i1+N);
TR(n:n+N-1,m:m+N-1)=R2; n=n+4;
end
m=m+4;
end

To calculate the matrix R in (6.22) from the base matrix Ψ, we use the
code matrixR FFofGrays.m with the following script.
% call: matrixR_FFofGrays.m / A.-M. Grigoryan, 2011
N=4;
UH8=UH_square(N); % Base matrix, or the UH-square in (6.20)
R=zeros(N*N); % Array for the matrix R in (6.22)
R2=zeros(N); k=1; NN=N*N;
for i1=1:N
for j1=1:N
R2=UH8(N-j1+1:N-j1+N,N-i1+1:N-i1+N);
R(k,:)=reshape(R2,1,NN); k=k+1;
end
end

To calculate the inverse matrix R−1 and print all of its 16 basic functions,
we use the code matrixR FFofGrays.m, which is given below.
% -------------------------------------------------------------
% call: matrixRin_FFofGrays / A.-M. Grigoryan, 2011
% -------------------------------------------------------------
matrixR_FFofGrays; % It calculates matrix R for the N=4 case
fprintf(’%16.0f \n’,det(R)); % 17128697013225
% The inverse matrix NxN (or 16x16)
Ri=inv(R);
% Represent this matrix as the block-matrix 4x4
NN=N*N;
iR=zeros(N,N,NN);
for k=1:NN
y=reshape(Ri(:,k),N,N); iR(:,:,k)=y;
end
% The 1st basis functions of R and Ri
364 CHAPTER 6: Method of Particles

k=1; % for G-particle (0,0)

field_function=reshape(R(:,k),N,N)
cof_function=iR(:,:,k)
sum(sum(field_function.*cof_function)) %1
% ========= calculated data ===============
% field_function #1
% 8 2 1 1
% 2 3 2 1
% 1 2 1 1
% 1 1 1 1
% cof_function
% 0.1636 -0.0192 -0.0049 -0.0107
% -0.0192 -0.0500 -0.0183 -0.0008
% -0.0049 -0.0183 0.0177 0.0056
% -0.0107 -0.0008 0.0056 -0.0056
% =========================================
k=6; % for G-particle (1,2)
field_function=reshape(R(:,k),N,N)
cof_function=iR(:,:,k)
sum(sum(field_function.*cof_function)) %1
% ========= calculated data ===============
% field_function #6
% 3 2 1 0
% 2 8 2 1
% 1 2 3 2
% 0 1 2 1
% cof_function
% -0.0500 -0.0120 -0.0008 0.0102
% -0.0120 0.1812 -0.0093 -0.0025
% -0.0008 -0.0093 -0.0459 -0.0219
% 0.0102 -0.0025 -0.0219 0.0177
% ========================================

For the first G-particle (0, 0), we obtain the following matrices of the first
basic functions of the transform, R and R−1 , respectively:
   
8 2 1 1 0.1636−0.0192−0.0049−0.0107
2 3 2 1   −0.0192−0.0500−0.0183−0.0008
ψ0,0 =  1 2 1 1  , φ0,0 =  −0.0049−0.0183 0.0177 0.0056 .
  

1 1 1 1 −0.0107−0.0008 0.0056−0.0056

The basic functions of the G-particle (1, 2) are also calculated separately in
this code:
   
3 2 1 0 −0.0500−0.0120−0.0008 0.0102
2 8 2 1  −0.0120 0.1812−0.0093−0.0025
ψ1,2 = 
 , φ1,2  −0.0008−0.0093−0.0459−0.0219 .
= 
1 2 3 2
0 1 2 1 0.0102−0.0025−0.0219 0.0177
One can notice that in matrices φ0,0 and φ1,2, the maximum values are at the
Image Processing 365

points (0, 0) and (1, 2), respectively, i.e. at the exact locations of the considered
G-particles.

6.3 Reconstruction by field transform

As shown above for A-rays and G-rays, we can define the systems of field
functions that describe the image in terms of A- and G-particles, respectively.
In both cases, the unique form of these fields allows for calculating the sums
of fields of such particles
X N−1
N−1 X
a(i, j) = ψn,m (i, j)fn,m , i, j = 0 : (N − 1). (6.24)
m=0 n=0

Here, {fn,m } stands for the image to be reconstructed, and ψn,m (i, j) are co-
efficients c(i, j, n, m) defining the number of rays passing through the particle
(i, j). The 4-D matrix ||c(i, j, n, m)|| of size (N × N × N × N ) can be rep-
resented in the form of a 2-D matrix of size (N 2 × N 2 ). We denote this 2-D
matrix by R = ||rk,l||, as shown in (6.13) and (6.22) for the example when
N = 4. We here remind the reader that the matrix R is composed of all N 2
field functions (N × N ) written row-wise in the form of 1-D vectors,
r(i−1)N+j,(n−1)N+m = c(i, j, n, m), i, j, n, m = 0 : (N − 1).
The matrices of data a(i, j) and fn,m can also be written row-wise
ai,j → a(i−1)N+j = ai,j , fn,m → f(n−1)N+m = fn,m ,
in the form of 1-D vectors. Equation 6.24 can therefore be written as
2
NX −1
ak = rk,lfl , k = 0 : (N 2 − 1). (6.25)
l=0

In matrix form, we obtain the equation a = Rf , where a = (a0 , a1 , ..., aN 2−1 )0

and f = (f0 , f1 , ..., fN 2−1 )0 . The existence of the inverse matrix, R−1 , allows
for calculating the vector f ,
f = R−1 a. (6.26)
Thus, the reconstruction of the image from its projections is achieved by using
the sums of fields of particles. The reconstruction is by the G-rays, i.e. the
matrix R and the solution in the above equation are considered with respect
the systems of field functions of G-particles. The matrix R is referred to as
the matrix of the field transformation of G-particles. Therefore, we call the
described method the method of field transformation of G-particles (FTGP).
We here consider the example of such reconstruction when the image is on
the Cartesian lattice 64 × 64.
366 CHAPTER 6: Method of Particles

Example 6.4 In the square [0, 1]×[0, 1] we consider the lattice 64×64 and the
following image composed randomly of ten rectangles. Each of the rectangles
is given by coordinates (xk,1, yk,1 ) and (xk,2 , yk,2), k = 1 : 10, as shown in
Table 6.7. The coordinates are written in integer form, i.e. they should be
normalized by the factor of N = 64. The intensities of these rectangles are
given in the last column.

TABLE 6.7
Data of 10 rectangles
k xk,1 yk,1 xk,2 yk,2 Rk
1 10 15 20 30 1
2 12 40 30 50 2
3 40 40 50 60 6
4 45 10 50 30 3
5 15 20 30 35 4
6 24 20 40 35 2
7 32 40 38 50 5
8 5 15 10 20 7
9 50 20 55 25 2
10 10 30 25 40 3

These ten rectangles are shown in Figure 6.16 in part a, along with recon-
struction of the image from 3N/2 = 96 projections in part b.

10 rectangles on 64x64 64x64: Covering by 96 projections

1

10
0.8 3
2 7 20
0.6
10 30

5 6
0.4 40
1 9
4
8
50
0.2

60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
(a) (b)

FIGURE 6.16
(a) Ten rectangles on the lattice 64 × 64 and (b) the reconstructed image.

We also describe another example, where a few rectangles are shifted and
do not lie on the lattice. For that, we modify the coordinates of five rectangles,
as shown in Table 6.8.
Figure 6.17 shows ten rectangles, five of which are not on the lattice,
in part a. The result of image reconstruction from 96 projections is shown
Image Processing 367
TABLE 6.8
Data of five shifted rectangles
k xk,1 yk,1 xk,2 yk,2 Rk
2 12 40.5 30 50 2
3 40 40.25 50 60 6
4 45.35 10 50 30 3
6 24 20 40 35.2 2
9 50 20 55.45 25 2

in part b. One can notice that the rectangles are well reconstructed with

10 rectangles on 64x64 64x64: Covering by 96 projections

1

10
0.8 3
2 7 20
0.6
10 30

5 6
0.4 40
1 9
4
8
50
0.2

60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
(a) (b)

FIGURE 6.17
(a) Ten rectangles and the lattice 64 × 64 and (b) the reconstructed image.

their exact edges, and additional work can be done to remove the effect of
shifting the rectangles. The result of reconstruction will look better when
reconstructing with high resolution. Figure 6.18 shows the same ten rectangles
and the lattice 128×128 on the square [0, 1]×[0, 1] in part a. The result of image
reconstruction from 192 projections is shown in part b. The 2nd rectangle is
now on the lattice and does not cause the error of reconstruction.

To accomplish the image reconstruction by FTGP, which is described in

Example 6.4, we use the code reconstruction byR.m with the following script.

% ----------------------------------------------------------
% call: reconstruction_byR / A.-M. Grigoryan, 2012
% ----------------------------------------------------------
N=64;
R=zeros(N,N,N,N); % 4-D matrix R
N_rect=10;
MNrects=[10 15 20 30
12 40 30 50
368 CHAPTER 6: Method of Particles

10 rectangles on 128x128 128x128: Covering by 192 projections

1

20
0.8 3
2 7 40
0.6
10 60

5 6
0.4 80
1 9
4
8
0.2 100

120
0
0 0.2 0.4 0.6 0.8 1 20 40 60 80 100 120
(a) (b)

FIGURE 6.18 (See color insert)

(a) Ten rectangles and the lattice 128 × 128 and (b) the reconstructed image.

40 40 50 60
45 10 50 30
15 20 30 35
24 20 40 35
32 40 38 50
5 15 10 20
50 20 55 25
10 30 25 40];
Matr_rect=MNrects/N;
E=[1 2 6 3 4 2 5 7 2 3]; % intensities of rectangles
% Plot the lattice NxN and rectangles -----------------
figure;
colormap(gray);
subplot(1,2,1); hold on
draw_grid; axis image;
h_x1=xlabel(’(a)’); set(h_x1,’FontSize’,12);
S_title=sprintf(’%g rectangles on %gx%g’,N_rect,N,N);
h_title=title(S_title);
% ------------------------------------------------------
ps=ps_generators(N);
L=N+N/2; % number of projections
A=zeros(N,N); % for sums a((i,j) of fields of particles
for k=1:L % loop for (p,s)-projections
p=ps(k,1); s=ps(k,2);
[p1,s1,t]=tensor_scheme_pst(p,s,N);
Len_t=length(t);
for it=1:Len_t % loop for (p,s,t)-G-rays
sum_A=fieldpst_tensor_scheme(p1,s1,t(it),N,N_rect,Matr_rect,E);
f=general_scheme_pst2(p1,s1,t(it),N);
Image Processing 369

f_sz=size(f);
for nf=1:f_sz(2)
m1=f(1,nf); n1=f(2,nf); L1=f(3,nf);
A(m1,n1)=A(m1,n1)+sum_A/L1; % Eq. 6.24
end
%========= calculate the 4-D matrix R ===========
for nf=1:f_sz(2)
for mf=1:f_sz(2)
R(f(1,nf),f(2,nf),f(1,mf),f(2,mf))=...
R(f(1,nf),f(2,nf),f(1,mf),f(2,mf))+1;
end
end
%=================================================
end % for (p,s,t)-rays
end % for (p,s)-projection
fprintf(’ 4-D Matrix R has been calculated \n’);
NN=N^2;
R2=reshape(R,NN,NN);
a=reshape(A,NN,1);
x=R2\a; % instead of x=inv(R2)*a; % Eq. 6.26
X=reshape(x,N,N); % reconstructed image
% Plot the reconstructed image
subplot(1,2,2);
imagesc(X); axis image;
S_title=sprintf(’%gx%g: Covering by %g projections ’,N,N,L);
h_title=title(S_title);
h_x2=xlabel(’(b)’); set(h_x2,’FontSize’,12);
% print -dps Figure6_16.ps
% --------------------------------------------------------------------

This main program uses the following functions:

1. draw_grid.m (to draw the rectangles and grid as in Figure 6.17(a))
2. ps_generators.m (to calculate the generators (p,s) of projections)
3. tensor_scheme_pst.m (to calculate all values of ’t’ in equations
pn+ms=t of G-rays that pass through the square [0,1]x[0,1], for a
given (p,s). It also calculates p=p-N and s=s-N, when p,s>N/2).
4. general_scheme_pst2.m (to represent the G-ray as a set of triplets
[x,y,d] of image-elements [x,y] and lengths d of intersections of
this ray with [x,y])
5. fieldpst_tensor_scheme.m (to calculate the field function along
the G-ray l(p1,s1,t1) for the sum of rectangles).
5a. raylength_inie.m (to calculate the ray-rectangle intersections)
6. sort_data.m (to reorder the coordinates of the ray-image-elements)

These programs are written not only for the tensor scheme of scanning,
when the set of 3N/2 generators (p,s) is defined as {(1,s)},{(2p,1)},
and sets of G-rays and A-rays correspond to these (p,s)-projections.
Arbitrary scheme of scanning can be used, including the fan scheme,
and not only integer values of generators (p,s) can be considered.
370 CHAPTER 6: Method of Particles

However in this section, these codes are used for the tensor scheme.
% ====================================================================
% call: tensor_scheme_pst.m
function [p,s,t]=tensor_scheme_pst(p,s,N)
N2=N/2;
delta=1/N; delta2=delta/2;
if p*s==0
t_min=delta2; t_max=1-delta2;
else
if p>N2
p=p-N; t_min=p+delta; t_max=1-delta;
elseif s>N2
s=s-N; t_min=s+delta; t_max=1-delta;
else
ps=max(p,s); t_min=delta; t_max=1+ps-delta;
end
end
t=t_min:delta:t_max;
% ====================================================================
% call: general_scheme_pst2.m
% Brief description:
% G-ray is represented as a set of triplets [x,y,d] of image-elements
% and lengths d of intersections of this ray with [x,y]. The triplets
% are written in the 3-D array F={f(1,2,3,:)=(x,y,d,:)}.
% Algorithm is simple:
% 1. Calculate all intersections with horizontal lines of the grid.
% 2. Calculate all intersections with vertical lines of the grid.
% 3. Unite these coordinates in array xy=[x y].
% 4. Sort xy by the first coordinate.
% 5. Find the coordinates of the image-elements (IE) along the G-ray.
% 6. Calculate the lengths of intersection of the ray with these IEs.
% If the intersection is zero or close to zero (as eps*100), then
% it is considered that the G-ray does not pass through the IE.
% The small value (eps*100) should be chosen more accurately.
% --------------------------------------------------------------------
function F=general_scheme_pst2(p,s,t,N)
t=-t;
delta=1/N;
eps100=eps*100;
ones1N=ones(1,N)*delta;
% 1,2
if abs(p)<eps100; % (p==0) in the tensor scheme
f=zeros(3,N);
x=-t/s;
if x>0 & x<=1
x_n=ceil(x/delta);
end
f(2,:)=ones(1,N)*x_n;
f(1,:)=1:N; f(3,:)=ones1N;
Image Processing 371

F=f; return
elseif abs(s)<eps100; % (s==0) in the tensor scheme
f=zeros(3,N);
y=-t/p;
if y>0 & y<=1
y_n=ceil(y/delta);
end
f(1,:)=1+N-ones(1,N)*y_n;
f(2,:)=1:N; f(3,:)=ones1N;
F=f; return
else % (p,s)~=(0,0) in the tensor scheme
a(1,:)=0:N; a=a*delta;
in_x=1; in_y=1; x=[];y=[];
for i1=1:length(a)
ai1=a(i1);
xn=(-t-p*ai1)/s;
if(xn >=0 & xn<=1)
x(:,in_x)=[xn;ai1]; in_x=in_x+1;
end
yn=(-t-s*ai1)/p;
if(yn >=0 & yn<=1)
y(:,in_y)=[ai1;yn]; in_y=in_y+1;
end
end
% 3.
xy=[x y];
% 4.
t=sort_data(xy); sz=size(t);
% 5.
t=t/delta;
for i1=1:sz(2)-1
i2=i1+1;
ff=(t(1,i2)+t(1,i1))/2; f(2,i1)= ceil(ff);
ff=(t(2,i2)+t(2,i1))/2; f(1,i1)= N-ceil(ff)+1;
end
xs(1,:)=t(1,:); ys(1,:)=t(2,:);
% 6.
m=1;
for i1=1:sz(2)-1
i2=i1+1;
sd=(xs(i2)-xs(i1))^2+(ys(i2)-ys(i1))^2;
f(3,i1)=sqrt(sd)*delta;
% consider only image elements with length > eps*100
if f(3,i1)>eps100
F(:,m)=f(:,i1); m=m+1;
end
end
end
372 CHAPTER 6: Method of Particles

It should be noted, that after the command "t=sort_data(xy);"

we probably should remove the equal pairs in ’xy’, in a case
if such repetition takes place. For the tensor scheme, such
situation will not happen, but not in the general case.
Therefore in the script of the above function, we could add
the following command: "t=delete_equalpairs(t);" and use the
following script for the function ’delete_equalpairs.m’.
% -----------------------------------------------------------------
function y=delete_equalpairs(x)
sz=size(x);
y(:,1)=x(:,1);
z=x(:,1); m=2;
for i1=2:sz(2)
if x(1,i1)~=z(1) || x(2,i1)~=z(2)
z=x(:,i1); y(:,m)=z; m=m+1;
end
end
% =================================================================
function sum_E=fieldpst_tensor_scheme(p,s,t,N,N_rect,Matr_rect,E)
sum_E=0;
for k=1:N_rect
[Length_inter,F]=raylength_inie(p,s,t,N,Matr_rect(k,:));
if isempty(Lengt_inter)
Length_inter=0;
end
sum_E=sum_E+E(k)*Length_inter;
end
% =================================================================
% call: raylength_inie.m
% Brief description:
% Coordinates (x1,y1) and (x2,y2) and length of intersection of the
% given G-ray l(p,s,t) with the rectangle (’rec’) are calculated.
% These coordinates are written into the 2-D array F=[x1 y1; x2 y2]
% and the length of intersection is written in ’Length_inter’.
% Algorithm is the same as in the code "general_scheme_pst2.m", the
% only difference is that instead of image elements we consider the
% one given rectangle. Calculations for the horizontal and vertical
% projections are separated from all other (p,s)-projections.
% ------------------------------------------------------------------
function [Length_inter,F]=raylength_inie(p,s,t,N,rec)
rx1=rec(1); ry1=rec(2); rx2=rec(3); ry2=rec(4);
x=[]; y=[]; eps100=eps*100;
if abs(p)<eps100
x=t/s;
if x>rx2 | x<rx1
Length_inter=[]; F=[];
else
Length_inter=ry2-ry1; F=[x,ry1;x,ry2];
end
Image Processing 373

elseif abs(s)<eps100
y=t/p;
if y>ry2 | y<ry1
Length_inter=[]; F=[];
else
Length_inter=rx2-rx1; F=[rx1,y;rx2,y];
end
else
x=[]; y=[];
for i1=1:2
i12=2*i1; ri12=rec(i12);
xn=(t-p*ri12)/s;
if(xn >=rx1 & xn<=rx2)
x=[x [xn; ri12]];
end
ri12=rec(i12-1);
yn=(t-s*ri12)/p;
if(yn >=ry1 & yn<=ry2)
y=[y [ri12;yn]];
end
end
xy=[x y];
if isempty(xy)
Length_inter=[]; F=[]; return
end
xy=sort_data(xy);
end
% calculate the length of the intersection
if (exist(’xy’)==1)
x1=xy(1,1); x2=xy(1,end);
y1=xy(2,1); y2=xy(2,end);
Length_inter=sqrt((x2-x1)^2+(y2-y1)^2);
F=[x1 y1;x2 y2];
% do not consider such intersection if length<eps*100
if Length_inter<eps100
Length_inter=[]; F=[];
end
end
% =================================================================
function f=sort_data(x)
sz=size(x);
[y,I]=sort(x(1,:));
for i1=1:sz(2)
f(:,i1)=x(:,I(i1));
end
% =================================================================

The scripts of these programs can be found at http://fasttransforms.com in

the folder Lectures/Tensor and Tomography.
374 CHAPTER 6: Method of Particles

In the examples considered above for the G- and A-particles on the lattice
4 × 4, the reconstruction of the image by FTGP has been obtained by using
the inverse matrix R−1 . The described program reconstruction byR.m also
accomplishes the calculation of this inverse matrix. The calculation of this
matrix can be separated and calculated and stored in advance. It is clear
that the calculation of such matrix (N 2 × N 2 ) is difficult for large values
of N = 2r , when r > 2. For instance, our personal computer can perform
the reconstruction of the image with maximal size 128 × 128, when using this
program. The N = 256 case faces a lack of the required memory when running
the program, and therefore new methods of calculation should be applied when
programming in MATLAB R
. In the main program reconstruction byR.m, the
command x=R2\a is used instead of calculation of the inverse matrix, when
using the command x=inv(R2)*a. We also mention the minimization techniques
of nonlinear programming [36], which may be applied for calculating the vector
x.
The direct method of FTGP may not be very effective for high values of N.
We therefore consider another, more effective solution of the problem, without
calculation of the inverse matrix R−1 .

6.4 Method of circular convolution

We now consider the transformation of the image, which is determined by the
system of field functions in (6.24)

X N−1
N−1 X
a(i, j) = fn,m ψn,m (i, j), i, j = 0 : (N − 1). (6.27)
m=0 n=0

As mentioned above, each field function can be calculated by the base field
matrix, ψn,m (i, j) = Ψ(N −1 −n+i, N −1 −m+j), i, j = 0 : N −1. Therefore,
we have the following property for the field function of the G-particle (n, m):

ψn,m (i, j) = ψ(i − n, j − m), i, j = 0 : N − 1. (6.28)

The transform of the image in (6.27) can be written as the linear convolution

X N−1
N−1 X
a(i, j) = (f ∗ ψ)(n, m) = fn,m ψ(i − n, j − m), (6.29)
m=0 n=0

where i, j = 0 : (N −1). The simple way to solve this 2-D convolution equation
is based on the 2-D discrete Fourier transform. Because the base field matrix
Ψ is of size (2N − 1) × (2N − 1), we will use the zero padding to extend the
image fn,m to this size. We denote the extended image by f˜n,m . Then, the
Image Processing 375

following is valid:

X 2N−1
2N−1 X
a(i, j) = f˜n,m ψ(i − n, j − m)
m=0 n=0
X 2N−1
2N−1 X (6.30)
= f˜n,m Ψ(N − 1 + i − n, N − 1 + j − m)
m=0 n=0
i, j = 0 : (N − 1).

Thus, the value of the 2-D linear convolution a(i, j) can be calculated from
the incomplete 2-D circular convolution of the extended image f˜n,m with the
base field matrix Ψ. To illustrate this property of zero padding, we describe
the following example.
Example 6.5 For the N = 4 case, we consider the script of the program
run ccN4.m.
% call: run_ccN4.m
% f=round(4*rand(4)); % random input 4x4
f=[ 1 3 2 1
0 1 2 2
0 4 3 2
3 0 3 3];
UH8=[ 1 1 1 1 0 0 1
1 1 2 1 0 1 0
1 2 3 2 1 0 0
1 1 2 8 2 1 1
0 0 1 2 3 2 1
0 1 0 1 2 1 1
1 0 0 1 1 1 1];
% The direct method of 2-D convolution:
R=zeros(4);
for i1=1:4
for j1=1:4
R=UH8(4-1+j1-3:4+j1-1,4-1+i1-3:4+i1-1);
A1(j1,i1)=sum(sum(R.*f));
end
end; A1
% 48 61 49 34
% 40 55 66 46
% 32 68 69 56
% 38 33 63 68
% The method of 2-D circular convolution:
F=zeros(7);
F(1:4,1:4)=f(:,:);
for i1=1:4
for j1=1:4
s=0;
for m1=1:7
376 CHAPTER 6: Method of Particles

m2=4-1+j1-3+m1-1;
if m2<1 m2=m2+7; end
if m2>7 m2=m2-7; end
for n1=1:7
n2=4-1+i1-3+n1-1;
if n2<1 n2=n2+7; end
if n2>7 n2=n2-7; end
s=s+UH8(m2,n2)*F(m1,n1);
end
end
A2(j1,i1)=s;
end
end; A2

The input image 4 × 4 as a result of command f=round(4*rand(4)) is the

matrix f = [1 3 2 1; 0 1 2 2; 0 4 3 2; 3 0 3 3]. The linear convolution calculated by
the direct method and the method of circular convolution results in the same
image A1 = A2 = [48 61 49 34; 40 55 66 46; 32 68 69 56; 38 33 63 68].

Because the linear convolution (f ∗ ψ) can be enclosed in the circular con-

volution (f˜ ⊗ ψ), one can use the discrete Fourier transform to solve equation
(6.30) and obtain the reconstruction f = {fn,m }. At the same time, this me-
thod allows for processing the reconstructed image in the frequency domain
for filtration or enhancement, if such processing is desired or required.
Let W be the twiddle coefficient W = W2N−1 = e−2πi/(2N−1) . The 2-D
DFT of the circular convolution a(n, m) can be calculated as follows:

X 2N−2
2N−2 X
Ap,s = a(i, j)W pi+sj
i=0 i=0
X 2N−2
2N−2 X X 2N−2
2N−2 X
= f˜n,m Ψ(N − 1 + i − n, N − 1 + j − m)W pi+sj
m=0 n=0 i=0 i=0
X 2N−2
2N−2 X X 2N−2
2N−2 X
= f˜n,m Ψ(i, j)W p(i−N+1+n)+s(j−N+1+m)
m=0 n=0 i=0 i=0
X 2N−1
2N−2 X X 2N−2
2N−1 X
= f˜n,m W pn+sm
Ψ(i, j)W pi+sj · W −(N−1)(p+s)
m=0 n=0 i=0 i=0
−(N−1)(p+s)
= F̃p,s Ψp,s · W

where we denote the (2N −1)×(2N −1)-point DFTs of the convolution a(n, m),
extended image f˜, and base matrix Ψ by Ap,s , F̃p,s , and Ψp,s , respectively.
Thus, we obtain the following formulas for image reconstruction:
Ap,s (N−1)(p+s)
F̃p,s = W , p, s = 0 : (2N − 1), (6.31)
Ψp,s
Image Processing 377

and
X 2N−1
2N−1 X
fn,m = F̃p,s W −(np+ms) , n, m = 0 : (N − 1). (6.32)
p=0 s=0

This formula can be used if the 2-D DFT of the base field matrix does
not have zero values. This matrix and its 2-D DFT can be calculated for any
values of N = 2r , and the simple program can be written to verify the fact
that Ψp,s 6= 0, for all (p, s).
As an example, we consider the DFT of the UH-square for the N = 4
case. The required calculations are accomplished by using the program with
the following script.

% call: checkzeros_inDFTofUHsquare.m
UH8=[ 1 1 1 1 0 0 1
1 1 2 1 0 1 0
1 2 3 2 1 0 0
1 1 2 8 2 1 1
0 0 1 2 3 2 1
0 1 0 1 2 1 1
1 0 0 1 1 1 1];
FUH8=fft2(UH8);
mn=min(min(abs(FUH8))); % 0.1491
R_ofFUH8=real(FUH8)
56.0000 -14.0036 4.6250 -1.1213 -1.1213 4.6250 -14.0036
-14.0036 0.5625 -0.3351 -0.0332 4.7409 -7.1528 24.0858
4.6250 -0.3351 -2.3107 3.6860 -4.4058 7.8358 -7.1528
-1.1213 -0.0332 3.6860 -8.7518 3.0785 -4.4058 4.7409
-1.1213 4.7409 -4.4058 3.0785 -8.7518 3.6860 -0.0332
4.6250 -7.1528 7.8358 -4.4058 3.6860 -2.3107 -0.3351
-14.0036 24.0858 -7.1528 4.7409 -0.0332 -0.3351 0.5625
I_ofFUH8=imag(FUH8)
0 -6.7438 5.7995 -4.9129 4.9129 -5.7995 6.7438
-6.7438 0.7053 -1.4683 0.1454 -5.9449 3.4446 -0.0000
5.7995 -1.4683 10.1237 -4.6221 2.1217 0.0000 -3.4446
-4.9129 0.1454 -4.6221 4.2147 0 -2.1217 5.9449
4.9129 -5.9449 2.1217 0 -4.2147 4.6221 -0.1454
-5.7995 3.4446 -0.0000 -2.1217 4.6221 -10.1237 1.4683
6.7438 0.0000 -3.4446 5.9449 -0.1454 1.4683 -0.7053

The real and imaginary parts of the DFT of the UH-square are printed
separately, and one can see that there is no zero value in this DFT. In absolute
scale, the minimum number of the DFT is 0.1491.
Below is the remaining part of the script, which calculates the DFT of the
UH-square of number 26 for the N = 8 case. The minimum number of the
DFT is 0.6347 in absolute scale.

UH26=[ 1 0 0 1 1 0 0 1 1 3 1 1 0 0 1
378 CHAPTER 6: Method of Particles

0 1 0 0 2 0 0 1 2 2 2 0 0 1 0
0 0 1 0 1 1 0 1 3 2 1 0 1 0 0
1 0 0 1 0 2 0 1 4 2 0 1 0 0 1
2 2 1 0 1 1 1 2 4 1 1 0 1 2 2
2 3 3 3 1 1 2 3 4 1 1 3 4 5 5
0 0 1 2 4 4 4 6 6 6 7 6 4 2 1
1 1 1 1 2 5 10 26 10 5 2 1 1 1 1
1 2 4 6 7 6 6 6 4 4 4 2 1 0 0
5 5 4 3 1 1 4 3 2 1 1 3 3 3 2
2 2 1 0 1 1 4 2 1 1 1 0 1 2 2
1 0 0 1 0 2 4 1 0 2 0 1 0 0 1
0 0 1 0 1 2 3 1 0 1 1 0 1 0 0
0 1 0 0 2 2 2 1 0 0 2 0 0 1 0
1 0 0 1 1 3 1 1 0 0 1 1 0 0 1 ];
FUH26=fft2(UH26);
mn=min(min(abs(FUH26))) % 0.6347

It is not difficult to verify that the 2-D DFT of the UH-squares does
not have zeros (see also Table 6.9, for other values of N = 2r , r > 3. The
number, k, of these squares UH(k) is also shown in the table. We here omit
the analytical proof of this fact and leave it to the reader.

TABLE 6.9
Tensor UH(k)-squares
N k min{|Ψp,s )|}
16 98 0.1585
32 386 0.1204
64 1538 0.0283
128 6146 0.0385
256 24578 0.0346
512 98306 0.0086
1024 393218 0.0004

We consider the base matrix Ψ as the function φ(n, m) with the center at
(0, 0), i.e.

φ(n, m) = Ψ(N + n − 1, N + m − 1), n, m = −(N − 1) : (N − 1).

We now calculate the 2-D DFT of the base function,

2N−2
X
Ψp,s = Ψ(n, m)W np+ms
n,m=0
N−1
X
= Ψ(N + n − 1, N + m − 1)W (N+n−1)p+(N+m−1)s
n,m=−(N−1)
N−1
X
= φ(n, m)W np+ms · W (N−1)(p+s) = φp,s W (N−1)(p+s) ,
n,m=−(N−1)
Image Processing 379

where W = W2N−1 = exp(−2πi/(2N − 1)). Now we consider the transform

of the function φ(n, m) shifted as (n, m) → (n + N, m + N ),
Ψp,s W N(p+s) = φp,s , p, s = 0 : (2N − 2).
This function is real and, therefore, the phase of the base matrix equals
2πN π
phase(Ψp,s ) = (p + s) = π(p + s) + (p + s).
2N − 1 2N − 1
Below is the script of the program DFTofUHsquare.m to calculate the real
function φp,s and show that it is real, namely, that the imaginary part of
this spectral function is zero. The minimum absolute value 0.038486 of this
function is also calculated. The N = 128 case is written, and it can also be
used for other values of N = 2r , when integer r > 2.
% -----------------------------------------------------------
% call: DFTofUHsquare.m / A.-M. Grigoryan 2011
% ----------------------------------------------------
N=input(’ Set the value of N=4,8,16,32, ... : ’);
N=128;
UHN=UH_square(N); % UH-square
FUH=fft2(UHN);
mn=min(min(abs(FUH))); % 0.0385
fprintf(’\n min(abs(FUH)=%8.6f \n’,mn);
M=2*N-1;
wN=exp(-1j*2*N*pi/M);
AUH=zeros(M);
for p=1:M
for s=1:M
ww=wN^(p+s-2);
AUH(p,s)=FUH(p,s)*ww;
end
end
i1=min(min(imag(AUH))); i2=max(max(imag(AUH)));
absmin=min(min(abs(AUH)));
fprintf(’ imag is in [%f,%f], min(abs(FUH))=%8.6f \n’,...
i1,i2,absmin’);
% imag is in [-0.000000,0.000000], min(abs(FUH))=0.038486
% -----------------------------------------------------------

6.4.1 Uniform frames

We consider equation (6.29) of the linear convolution, where the values of
basic functions ψ(n, m, i, j) = ψ(i − n, j − m) are written with consideration
of the lengths of the intersection of rays with the image elements, or squares
X N−1
N−1 X
a(i, j) = fn,m ψ(i − n, j − m), i, j = 0 : (N − 1). (6.33)
m=0 n=0
380 CHAPTER 6: Method of Particles

This convolution equation can be written in terms of frames, and it is easy to

do that for the following “ideal scheme of scanning.” This scanning scheme is
explained as follows.
The frames Υk of the tensor UT-square for number 26 have different coef-
ficients, the number of which equals c(k) = (2k + 1)2 − (2k − 1)2 = 8k, when
k ≥ 1, and c(0) = 1 for the center. We now assume that the numbers on
the frames are distributed uniformly. Let rk be the coefficients on frames Υk ,
respectively. Then the linear convolution in (6.33) can be written as
N−1
X X
a(i, j) = fn,m ψ(i − n, j − m)
k=0 (n,m)∈Υk
N−1
X X N−1
X X
= rk fi−n,j−m = rk fi−n,j−m.
k=0 (n,m)∈Υk k=0 (n,m)∈Υk

The projection data in point (i, j) is calculated by integrating values of the

image on the frames, or frames shifted to the center with coordinates (i, j). It
should be noted for comparison, that in the method of backprojections, the
each projection is amplified by the filter |ω1 | (see (7.3)), which in the frequency
domain (w1 cos θ, w1 sin θ) describes the circle of radius |w1 |.

Problems
Problem 6.1 (Model 1) Consider the symmetric tensor scheme of scan-
ning, when 2N − 1 generators (p, s) are used from the following set:
1
JN,N = {(1, s), s = 0 : (N − 1)} ∪ {(p, 1), p = 0, 2 : (N − 1)}.
A. Determine the base matrix Ψ of G-particles in the N = 4 case. Show
that this matrix Ψ is the UH-square.
B. Determine the field functions of the G-particles (1, 1) and (3, 2) in the
N = 4 case.
C. All N 2 field functions of G-particles can be written row-wise into the
matrix N 2 × N 2 . Determine such matrix for N = 4.
Problem 6.2 Consider the symmetric tensor scheme of scanning with the set
1
of generators JN,N . Determine the base matrix Ψ of G-particles in the N = 8
case. Show that this matrix Ψ is the UH-square.
Problem 6.3 Consider the symmetric tensor scheme of scanning with the set
1
of generators JN,N . Show that, similar to the tensor scanning scheme, the 4-
dimensional matrix R = ||c(i, j, n, m)|| provides the representation in the form
of the linear convolution. Here the coefficients c(i, j, n, m) equal the number
of rays passing through the particles (i, j) and (n, m) at the same time.
Image Processing 381

Problem 6.4 Write the code and show that reconstruction of the image N ×
N, where N is a power of two, is possible when using the scanning scheme
1
with the set JN,N .

Problem 6.5 (Model 2) Consider the following scheme of scanning, when

N − 1 generators (p, s) are used from the following set:
2
JN,N = {(1, s), s = 0 : (N/2 − 1)} ∪ {(p, 1), p = 0, 2 : (N/2 − 1)}.

For the N = 4 case, determine the following:

A. The base matrix Ψ of G-particles and show that the matrix Ψ is the
UH-square. Calculate det(Ψ).
B. The field functions of the G-particles (1, 0) and (1, 1).
C. All N 2 field functions of G-particles can be written row-wise into the
matrix N 2 × N 2 . Determine this matrix for N = 4. Determine the rank of this
matrix and determine if the inverse matrix exists.
D. Show that, similar to the tensor scanning scheme, the 4-dimensional
matrix R = ||c(i, j, n, m)|| provides the representation in the form of the linear
convolution.

Problem 6.6 Repeat Problem 6.4 for the N = 8 case.

Problem 6.7 Write the code and show that reconstruction of the image N ×
N, where N is a power of two, is possible when using the scanning scheme
2
with the set JN,N .

The solutions of the above problems show that different scanning schemes can
be used with a number of projections greater or less than 3N/2.

Problem 6.8 (Model 3) Given integer M > 0, consider the following

scheme of scanning, when 2M − 1 generators (p, s) are used from the set
3
JN,N = {(1, s), s = 0 : (M − 1)} ∪ {(p, 1), p = 0, 2 : (M − 1)}.

Write the code and find the minimum value of M = M (N ) for which recon-
struction of the image N × N is possible. Here, N is a power of two.

Problem 6.9 (Model 4) Given integers M > 0 and K > 0, consider the
following scheme of scanning, when M + K − 1 generators (p, s) are used from
the following set:
4
JN,N = {(1, s), s = 0 : (M − 1)} ∪ {(p, 1), p = 0, 2 : (K − 1)}.

Write the code and find the minimum values of M = M (N ) and K = K(N )
for which reconstruction of the image N × N is possible. Here N is a power
of two.
382 CHAPTER 6: Method of Particles

Problem 6.10 (Model X) Given integer M > 0, consider the scheme of

scanning, when M generators (p, s) are selected randomly with p and s uni-
formly distributed in the interval [0, N − 1].
x
JN,N = {(p, s), s, p ∈ [0, N − 1]}.

Write the code and find out if the exact reconstruction of the image N × N
is possible with this random scanning scheme. Here, N is a power of two.
Determine for which values of M the reconstruction can be achieved. Provide
examples and explain the results. (Note that p and s may have non-integer
values.)
7
Methods of Averaging Projections

In this chapter we describe a few methods of averaging for image reconstruc-

tion. The main result with a statistical model of the image is described in §7.4.
It allows us to achieve a high-quality image reconstruction through the pro-
cess of iterations and renewal of statistics after each iteration. This approach
does not require any transformation and its effective realization can be used
for reconstructing the image in discrete form with any size.
In Chapters 3, 4, and 5, the method of image reconstruction, which is
based on transferring the geometry from A-rays to G-rays, was described.
The reconstruction of the image is exact, when considering the image f(x, y)
as a set of small cells, or image elements inside the Cartesian lattice N ×
N. We started the description of this method with the images composed of
random rectangles, and then the proposed method was applied to other images
as well. The exact reconstruction was achieved under the assumption that
the image elements are exactly on the lattice. This is a real model of the
image for reconstruction, which is used in the known methods of finite series
expansion [29, 31, 47].
We now consider an example where the image consists of a few rectangles
rk (x, y) with intensities Rk and positions (xk , yk , ∆xk , ∆yk ),
K  
1 X x − xk 1 y − yk 1
f(x, y) = R k rect − , − . (7.1)
(∆x)2 ∆xk 2 ∆yk 2
k=1

When all data of random rectangles, xk , yk , ∆xk , and ∆yk , are integer multi-
ples of ∆x = 1/N, the image can be reconstructed exactly in the form of the
corresponding discrete analog {fn,m } of this image. Given (n, m), the value of
fn,m equals the average image intensity in the square [n, m]. We now consider
the case when data of such rectangles may not be integer multiple to 1/N, i.e.,
when one or a few rectangles of the image are not divided by the Cartesian
lattice N × N by small squares of size 1/N × 1/N each. In this case, we say
that the image f(x, y) is not on the lattice, and it is shifted.
As a example, Figure 7.1 shows the image f(x, y) with seven random rect-
angles, which is not on the Cartesian lattice 256 × 256. The reconstructed
image of the size 256 × 256 is also shown. One can observe that although the
rectangles with their straight edges have almost been reconstructed and can
be observed well, there are many “noisy” points in the image. This image
therefore needs to be processed in order to filter such a noise, as is done, for

383
384 CHAPTER 7: Methods of Averaging

7 rectangles and grid 256x256 reconstructed image 256x256

1

0.8 50
6

0.6 100
5

2
0.4 7 150

3 1
4 200
0.2

0 250
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 4 2 14 11 9 13 5 cpu time/rect=4.90 min

FIGURE 7.1 (See color insert)

Image with seven random rectangles and the reconstructed discrete image
256 × 256.

instance, in the known method of filtered backprojection (BP). This method

is also called the method of summation of projection data by rays, and it is
used widely in computed tomography [28, 29, 30].

7.1 Filtered backprojection

First we describe the method of summation of projection data by the rays for
the set of projections, which is described in the tensor representation of the
image. In other words, we consider the method of summation with the tensor
scheme of scanning the image. The concept of back projecting the measured
data obtained on the receiver side of X-ray tomography can be described
briefly as follows. Let L be the set of all G-rays l which are used for scanning
the object or image f(x, y), and let ∆l be the length of the ray l in the image.
The value wl of the projection along the ray l is divided equally among all
points (x, y) lying on this ray. Then, for every point (x, y) on the image plane,
all rays passing through this point are considered. The sum of averaging line-
integrals calculated along these rays determines the value of the reconstruction
at the point (x, y),
X X wl
fbp (x, y) = . (7.2)
∆l
l∈L l3(x,y)

We first apply this idea on the discrete images, and then on images modeled
by random rectangles in the square [0, 1]×[0, 1]. To all points in the rectangles,
 Image Processing 385

as well as outside them, some values of intensity are assigned. The resulting
image is smooth, because of averaging.
This method can be described by considering the 2-D Fourier transform
of the image f(x, y) in the polar system of coordinates in the frequency do-
main [30]. Indeed, let us consider the line-integral of the image

Z Z1Z1
w(t) = wθ (t) = f(x, y)dlθ (t) = f(x, y)δ(x cos θ + y sin θ − t)dxdy
0 0

along the line lθ (t) = {x cos θ + y sin θ = t}. The Fourier transform of w(t)
coincides with the 2-D Fourier transform of the image along the line ω2 =
ω1 tan(θ) in the frequency domain (ω1 , ω2 ). In other words,

Z1Z1
F (w1 cos θ, w1 sin θ) = f(x, y)e−j(xω1 cos θ+yω1 sin θ) dxdy =
0 0

Z∞Z1Z1 Z∞
−jω1 t
f(x, y)e δ(x cos θ + y sin θ − t)dxdydt = wθ (t)e−jω1 t dt = Wθ (ω1 ).
−∞ 0 0 −∞

This is the statement of the Fourier Slice Theorem. Therefore, by using the
inverse Fourier transform, the image can be written as
 
Zπ Z∞
1  1
f(x, y) = Wθ (ω1 )|ω1 |ejω1t dω1  dθ, (7.3)
2π 2π
0 −∞

where t = x cos θ + y sin θ. For a given angle θ, the same value of the integral
Z∞
1
bθ (t) = Wθ (ω1 )|ω1 |ejω1t dω1
2π
−∞

will be added to the image f(x, y) at all points (x, y) on the line lθ (t) =
{x cos θ + y sin θ = t}. This value is the inverse Fourier transform of the pro-
jection amplified by the filter |ω1 |. Therefore, Equation (7.3) describes the
process of filtered backprojection,
Zπ
1
f(x, y) = bθ (x cos θ + y sin θ)dθ.
2π
0

The realization by this formula requires the full, or infinite number of projec-
tions {wθ (t); θ ∈ [0, π]} and convolution with the impulse response h(t), which
is defined by the frequency characteristics H(ω1 ) = |ω1 |, where ω1 ∈ (−∞, ∞).
These two complex operations cannot be accomplished exactly in practice,
386 CHAPTER 7: Methods of Averaging

but only approximated under some assumptions. The number of projections

is reduced to a certain number by considering, for instance, only 180 angles
distributed uniformly in the interval [0, π]. The filter H(ω1 ) is not bounded,
but it will be cut by some frequency ω0 ,

|ω1 |,when ω1 ∈ (−ω0 , ω0 ),
H̃(ω1 ) = rect (ω1 /(2ω0 )) − tr (ω1 /(2ω0 )) =
0, otherwise,

when assuming that there is no aliasing for the projection data when sampling
with a period T < π/ω0 . To smooth the sharp cutoffs at ±ω0 and reduce the
ringing artifacts in the reconstruction, different smoothing window functions
can be used to modify the Ran-Lak filter H̃(ω1 ). For instance, we consider
the generalized Hamming window [29, 30]

α + (1 − α) cos(πω1 /ω0 ), when ω1 ∈ (−ω0 , ω0 ),
H(ω1 ) = |w1 | ·
0, otherwise,

where the value of α is selected from the interval (0, 1). As an example, Fi-
gure 7.2 shows the Ran-Lak filter with cutoff frequency ω0 = 1.45rad/s, the

1.5
|ω1|
smoothing window (α=0.5)
1

0.5 ω0=1.45

H(ω1)
0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 ω1 2

FIGURE 7.2
Modification of the Ran-Lak filter by Humming smoothing window.

Humming window with the parameter α = 0.5, and the function H(ω1 ).

7.2 BP and method of splitting-signals

Each splitting-signal of the discrete image fn,m can be defined as a sum of
the image along parallel lines, i.e., by projection data. At the same time,
the splitting-signal defines the direction image as being a component of the
image, as stated in (2.49), for the image of size N × N, where N is a power
of 2. This direction image has constant values along the parallel lines. The
number of such constants equals N/2 for the first series of generators (p, s),
Image Processing 387

and N/2k , for the kth series of generators, where k = 2, 3, ..., r. This property
explains the main difference between the paired transform-based method of
image reconstruction and the method of backprojection. In the method of
backprojection, direction images also are used, and for many projections, these
images are incomplete. This means they are not calculated for all points on
the Cartesian grid N × N. The number of such images equals the number of
projections, and the value of each of these images at the point (n, m) is set
to be equal to the average (and filtered) value of the projection along the ray
passing through this point. This rough assumption does not allow for accurate
reconstruction of the image, even for a large number of projections.
For example, consider the case when the full set of 2(128)−1 measurements
are used for calculating the projection by the angle arctan(2) of the image on
the lattice 128 × 128. The corresponding direction image in BP will contain
these 255 values along the parallel lines. In the paired method, from these
255 measurements, 128 values of the splitting-signal in tensor representation,
and then 64 values in paired representation are calculated and used in the
direction image. The number of projections, or direction images, for image
reconstruction by BP is considered to be 180, assuming the uniform change
of the angle of projections by 1 degree. The projections which are used in the
tensor and paired methods are calculated by the angles

Φ = Φ(256) = {π/2 − arctan(s/p); (p, s) ∈ J256,256},

which differ from the angles used in BP. It follows from the principle of super-
position, that the number of required projections equals 192, from which 382
direction images are calculated for image reconstruction by the paired method.
All direction images in BP are non-negative, and in paired representation the
direction images have both positive and negative values. The BP method re-
sults in a smooth image, and the paired method allows reconstruction of the
discrete image (noise is not considered).
As an example, Figure 7.3 shows the discrete tomo image of size 256 × 256
in part a, together with the results of image reconstruction by the filtered
backprojection in b and c, when respectively 384 and 360 projections are
used at angles equally spaced between 0 and 180 degrees. The mean-square-
root error of image reconstruction equals 0.010308 in (c) and 0.011086 in
(d). One can notice that many details on these images are not seen well. For
comparison, the reconstruction of the image by the paired transform with the
same number, 360, of projections is shown in d. The error of reconstruction
equals 0.002065 and 360 angles are defined by the frequency-points (p, s) of
the set of generators J256,256, whose splitting-signals have high energies,
255
1 X 2
Ep,s = f > 8773.85.
2562 t=0 p,s,t

The graph of the energy function Ep,s is given in Figure 7.4, where the
generators (p, s) are ordered, as written in the set J256,256, i.e., the first is
388 CHAPTER 7: Methods of Averaging

(a) Original image (b) MSR=0.010308

(c) MSR=0.011086 (d) MSR=0.002065

FIGURE 7.3
(a) The image 256 × 256. The image reconstruction by the filtered backpro-
jection algorithm by (b) 384 and (c) 360 projections (by MATLAB R
function
iradon.m), and (d) reconstruction by the 2-D DPT with 360 projections.

4
x 10
1.25
1.2
1.15
1.1
1.05
1
0.95
0.9
0.85
0 50 100 150 200 250 300 350

FIGURE 7.4
The energy curve of all 384 splitting-signals of the tomo image.

(1, 0), the 257-th is (0, 1), and the last is (254, 1). The splitting-signal genera-
ted by (p, s) = (0, 1) has the highest energy 12290.08, and the splitting-signal
with generator (1, 0) is the second with high energy signal, E(1, 0) = 9719.93.
These two signals are defined by the vertical and horizontal projections, res-
pectively. It is interesting to note that the next high-energy splitting-signal
Image Processing 389

with energy 8896.03 is the signal generated by (1, 128), not the signal defined
by the diagonal projection with E(1, 1) = 8827.80. The value of the minimum
energy of the splitting-signals equals 8771.83.
In the case when N is prime, the image fn,m can be reconstructed from
f(x, y) by direction images of number N + 1,
"N−1 #
1 X
fn,m = f1,s,(n+sm) mod N + f0,1,m − N E[f], n, m = 0 : (N − 1).
N s=0

The image is the average of direction images d(p,s)(n, m) = fp,s,(pn+sm) mod N

generated by frequencies (1, s), s = 0 : N − 1, and (0, 1). We consider, for
example, the image of size 257 × 257 and the (1, 5)-projection with L = (1 +
5)(256 − 1) + 1 = 1531 parallel rays. In the traditional backprojection, all
values of L line-integrals w(t), t = 0 : L − 1, along these rays will be put back
and distributed equally at points of these rays. The splitting-signal {f1,5,t ; t =
0 : 256} has 257 different values. Thus, in the tensor method, from these 1531
integrals, 257 values of the splitting-signal are calculated, and these 257 values
(not 1531) will be distributed back along 1531 rays. This is the main difference:
along the same rays, these two methods put back different values. BP adds
average values of the integrals. In the tensor method, these integrals w(t) are
transformed into 257 new combinations f1,5,t of integrals, and then put back
along the rays, to get the exact reconstruction. We here remind the reader
that the tensor method works on the Cartesian grid, and the BP uses the
polar grid, which does not cover the Cartesian grid.
The following should also be noted when applying the Fourier Slice Theo-
rem for image reconstruction [29, 30, 54]. The calculation of the 2-D Fourier
transform of the unknown image f(x, y) is incomplete. Such an incomplete
2-D DFT is defined on the polar grid along a finite number of radial lines, or
better said, is filled by calculating 1-D DFTs over the projections. Then, these
spectral components on the radial points are transferred to the Cartesian grid,
by using methods of interpolation. Even in the case when such transformation
of spectral data from the polar grid to the Cartesian grid is performed with
a small error, this transformation does not preserve, but breaks the mathe-
matical structure of the 2-D DFT of the image. Consider for example, any
projection which is different from the projections by 0◦, 45◦ , and 90◦ , such as
the projection calculated by the angle arctg(2). Our example is for the discrete
image 16×16. When using the Fourier Slice Theorem, the 2-D DFT of the disc-
rete image will be calculated at 8 frequency-points on the line l1 which is consi-
dered in the square grid Y16,16 = {(p, s); p, s = −8 : 7} with the original (0, 0)
in the center, as shown in Figure 7.5 part a. Since the tensor and paired rep-
resentations are described on the square grid X16,16 = {(p, s); p, s = 0 : 15},
and the 2-D DFT is periodic, we transfer the four parts of Y16,16 into X16,16
by the following simple rule:
   
AB DC
Y = → = X.
CD BA
390 CHAPTER 7: Methods of Averaging

l2
15 15
5
l1
10 10
0

5 5
l3
−5

0 0
−5 0 5 0 5 10 15 0 5 10 15
(a) (b) (c) T2,1

FIGURE 7.5
(a) The grid Y16,16 with 16 points on the lines l1 , l2 , and l3 , and the grid X16,16
with (b) 8 points on the line l1 , and (c) 16 points of the group T2,1 .

As a result, eight frequency-points on the line l1 will be located on the grid

X16,16 as shown in part b. According to the tensor representation, the pro-
jection data at the angle arct(2) define the 2-D DFT of the image fn,m at 16
frequency-points of the cyclic group T2,1 . This complete group of frequency-
points on two parallel lines with eight points each is shown in part c. The same
group on the grid Y16,16 is shown in part a, as the set of frequency-points on
the line l1 and two additional parallel lines l2 and l3 , which are missing when
the Fourier Slice Theorem is used to calculate the 2-D DFT of the discrete
image by the given projection. The 2-D DFT of the image does not have a
smooth form. The missing frequency-points on these two additional lines may
only be substituted by interpolating data from other projections (which are
assumed to be “nearly independent” [30]), or filled by zeros, which leads to
an error of reconstruction in both cases. Similar reasoning can be applied for
many other projections as well.

7.2.1 Tensor method of summation of projections

As mentioned above, we consider that the set of uniformly distributed angles
in the interval [0◦, 180◦], for collecting the projection data, does not provide
the correct calculation of the reconstruction image in the frequency and spatial
Cartesian grids. The discrete images of size N ×N, when N is a power of 2 and
prime, are defined by very different sets of angles for projections and in the
frequency domain they are defined by different sets Tp,s of frequency-points,
and this fact is not taken into consideration in the method of summation.
As examples, we consider the tomo image of size 256 × 256, as well as the
same image in 257 × 257, by adding a zero column and row to the image.
Following the tensor representation, we take 384 and 258 projections defined
by the angles of sets Φ(256) and Φ(257), respectively. We use the following
simple formula for reconstructing the discrete image fˆn,m from projection data
Image Processing 391

(without any filtration):

1 X pϑ (np + ms)
fˆn,m = ,
M card(lϑ (np + ms)) (7.4)
ϑ=ϑ(p,s)∈Φ(N)

n, m = 0 : (N − 1),

where M is the number of projections, i.e., M = 394 and M = 258, for

N = 256 and N = 257, respectively. Here, pϑ (t) = v(t) denotes the sums of

MSR=0.009257 MSR=0.009520

50 50

100 100

150 150

200 200

250 250
50 100 150 200 250 50 100 150 200 250
(a) 256x256 (b) 257x257

FIGURE 7.6
The reconstruction of the tomo image of size (a) 256×256, by 384 projections,
and (b) 257 × 257, by 258 projections.

the image fn,m at the points on the lines

lϑ (t) = {(n1 , m1 ); n1 p + m1 s = t}.

The values of v(t) can be calculated from the line-integrals w(t) of the image
f(x, y) along the G-rays of the (p, s)-projection, as described in Chapter 3.
Function card(lϑ (t)) is used for the cardinality, or number of grid points on
these lines. The weight card(lϑ (t)) of the projection pϑ (t) takes values from
the interval [1, N ], and the higher the frequency (p, s) generating the angle ϑ,
the smaller this weight. Figure 7.6 shows the result of image reconstruction
in part a, for the 256 × 256 case, and in b, for the 257 × 257 case. The mean-
square-root error of image reconstruction equals 0.009257 in (a) and 0.009520
in (b). The obtained reconstructions are the high-quality images with all fine
details of the original image. The time data from running MATLAB-based
codes on a personal computer with Intel Core Duo processor at 2.99 GHz
speed are equal to 4.2 min and 3 min, for calculating the reconstruction for
images 256 × 256 and 257 × 257, respectively.
Below is the script of the program bptt tomo256book.m, which is used to
accomplish the above described tensor method of backprojection of the tomo
392 CHAPTER 7: Methods of Averaging

image 256 × 256, which is shown in Figure 7.6 in part a. The image seen
when backprojecting only one (p, s)-projection is calculated by the function
projection cal1.m. The concept of tensor representation is considered with
direct calculation of the coordinates of points (x, y) on the grid N × N, which
lie on A-rays of the projection. Namely, for a given (p, s)-projection and each
integer t from [0, N − 1], the line-integrals, or sums of the image along the
coordinates of the following set of A-rays

lp,s (t), lp,s (t + N ), lp,s (t + 2N ), · · · , lp,s (t + (p + s − 1)N )

are calculated separately. Each such set of rays determines the corresponding
component of the tensor transform, fp,s,t .

%==================================================
% call: bptt_tomo256book.m
% Backprojection method with 3N/2 specified angles.
% Function "projection_cal1.m" is used to calculate
% the image back-projected by one (p,s)-projection.
%
% Art Grigoryan, 07/21/2010, EE UTSA
%==================================================
N=256;
fid=fopen(’tomo256.img’,’rb’);
X=fread(fid,[N,N]); fclose(fid); clear fid; X=X’;
% The set of generators (p,s)
L=N+N/2;
Jps=ones(2,L);
Jps(1,1:N)=0:N-1; Jps(2,N+1:end)=0:2:N-2;
% Reconstruction of the image
image_out=zeros(N);
for k=1:L
p=Jps(1,k); s=Jps(2,k);
image_out=image_out+projection_cal1(X,p,s,N);
end
image_out=image_out/L; % MSR=0.009257
% Display the images
figure
subplot(1,2,1); imshow(X,[]);
subplot(1,2,2); imshow(image_out,[]);
% ------------------------------------------------
function image_back=projection_cal1(X,p,s,N)
image_back=zeros(N);
TN=0:N:max(p,s)*N;
ps1=p+s;
for t=0:N-1
value=zeros(1,ps1); % for line-integrals
L=zeros(1,ps1); % for number of IE
save_xy=zeros(2,ps1,ps1); % for coordinates of IE
np=0;
Image Processing 393

for n=1:N
ms=0;
for m=1:N
t_see=np+ms;
k=floor(t_see/N); k1=k+1;
t1=t_see-TN(k1);
if t1==t
Lk1=L(k1)+1; L(k1)=Lk1;
save_xy(1,k1,Lk1)=m;
save_xy(2,k1,Lk1)=n;
value(k1)=value(k1)+X(m,n);
end
ms=ms+s;
end
np=np+p;
end
value_projected=value(1:k1)./L(1:k1);
for k2=1:k1
for n=1:L(k2)
x_onray=save_xy(1,k2,n);y_onray=save_xy(2,k2,n);
image_back(x_onray,y_onray)=value_projected(k2);
end
end
end
% ==============================================================

This code can be written in an effective way, if we determine the rays

that intersect all image elements, or squares 1/N × 1/N, on which the square
[0, 1] × [0, 1] is divided by the Cartesian grid N × N, as in the method of fast
projection integrals by squares (FPIS).
Example 7.1 Consider the N = 64 case, where the gray-scale image f(x, y)
of size 64×64 is composed of seven rectangles with the data given in Table 7.1.

TABLE 7.1
Integer data of rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 18 19 6 15 9
2 29 29 14 25 6
3 39 39 18 18 13
4 41 14 17 17 3
5 14 38 2 10 10
6 36 11 10 8 15
7 24 9 7 25 4

Figure 7.7 shows seven rectangles that compose the image f(x, y) (on the
left), together with the gray-scale discrete image fn,m (on the right). The re-
construction fˆn,m of the image f(x, y) by the described tensor method, when
394 CHAPTER 7: Methods of Averaging

7 rectangles model: image on grid 64x64

1

10
0.8 2
20
0.6 3
5
30

0.4 2
40
1
50
0.2 4
6
7
60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
intensities: 9 6 13 3 10 15 4

FIGURE 7.7
The image with seven rectangles and gray-scale image of size 64 × 64.

96 projections are used, is shown Figure 7.8 in part a. The mean-square-root

error (MSR) of reconstruction equals 0.035878. All rectangles can be seen in
their exact form with clear contours. For comparison, the image reconstructed
by the filtered BP with 100 projections is shown in b. BP is accomplished
by using the MATLAB function iradon.m with the standard Ram-Lak filter.
These projections are defined by 100 angles distributed uniformly in the in-
terval [0, 180◦]. The reconstruction is an unfocused image with the error equal
to 0.037464.

BP by Art [96] BP by Ram−Lak [100]

10

20

30

40

50

60

10 20 30 40 50 60
(a) (b)

FIGURE 7.8
The reconstruction of the image (a) by 96 projections (without any filtering),
and (b) by 100 projections in BP method with the cropped Ram-Lak filter.
Image Processing 395

Below is the main part of the program bptt rectangles.m, which accomp-
lishes the reconstruction of the image with the tensor method described by
(7.4), as well as the reconstruction by the filtered BP method with the ramp,
or Ram-Lak filter, which is shown in Figure 7.8.

%====================================================================
% call: bptt_rectangles.m
% Backprojection methods with 3N/2 specified angles.
% Art Grigoryan, 07/21/2010, EE UTSA
%==================================================
N=64; many_rectangles=7;
data_allrect=zeros(many_rectangles,5);
data_allrect=[ 18 19 6 15 9
29 29 14 25 6
39 39 18 18 13
41 14 17 17 3
14 38 2 10 10
36 11 10 8 15
24 9 7 25 4];
% composition of the gray-scale image fd(x,y) from rectangles
imageSQ_model=zeros(N,N);
for k=1:many_rectangles
data_1rect=data_allrect(k,:);
x=data_1rect(1); dx=data_1rect(3);
y=data_1rect(2); dy=data_1rect(4);
d_intensity=data_1rect(5);
x1=x+dx; y1=y+dy;
imageSQ_model(x:x1,y:y1)=imageSQ_model(x:x1,y:y1)+d_intensity;
end
imageSQ_model=imageSQ_model(:,N:-1:1);
% Art Method [96 projections] ------------------
X=imageSQ_model’;
% The set of generators (p,s)
L=N+N/2;
Jps=ones(2,L);
Jps(1,1:N)=0:N-1; Jps(2,N+1:end)=0:2:N-2;
% Reconstruction of the image
image_out=zeros(N);
for k=1:L
p=Jps(1,k); s=Jps(2,k);
image_out=image_out+projection_cal1(X,p,s,N);
end
image_out=image_out/N;
% BP method [100 projections] -------------------
f=linspace(0,180,100);
Radon_ofX = radon(X,f); % 95x100
X_iradon = iradon(Radon_ofX,f,’Ram-Lak’); % 66x66
% Display the images
figure; colormap(gray);
396 CHAPTER 7: Methods of Averaging

subplot(1,2,1); imshow(image_out,[]); title(’BP by Art [96]’);

subplot(1,2,2); imshow(X_iradon,[]); title(’BP by Ram-Lak [100]’);
%====================================================================

Example 7.2 Consider the image f(x, y) on the square [0, 1] × [0, 1], which
is composed of eight circles and two ellipses,

10 10  (x − xk )2 (y − yk )2
X X Ik ; if + ≤ rk2
f(x, y) = fk (x, y) = a2k b2k

k=1 k=1 0; otherwise
The spatial data for these circles and ellipses, as well as the intensities, are
given below.
k 1 2 3 4 5 6 7 8 9 10
xk 24 48 88 80 64 20 112 48 24 116
yk 40 52 60 24 92 104 112 16 76 24
ak 1 1 1 5 5 1 1 1 1 1
bk 1 1 1 4 8 1 1 1 1 1
rk 12 20 28 4 4 8 4 4 4 4
Ik 20 8 12 16 20 20 24 20 24 24

The geometry of these figures is in the integer form, but it will be normal-
ized by factor of 127. The number 127 is prime and the reconstruction of the
image is considered on the lattice 127 × 127. Figure 7.9 shows the sampled
discrete image in part a. The method of backprojection by the tensor scanning
scheme uses 128 projections. The reconstruction of the image is shown in b.
The image reconstructed by the method of BP with 128 projections is shown
in c.

original image BP by Art [128] BP by Ram−Lak [128]

20 20 20

40 40 40

60 60 60

80 80 80

100 100 100

120 120 120

20 40 60 80 100 120 20 40 60 80 100 120 20 40 60 80 100 120
(a) (b) (c)

FIGURE 7.9
(a) The image 127 × 127, the reconstruction with 128 projections by the (b)
tensor method and (c) BP method with the cropped Ram-Lak filter.

In the tensor method of summation, we can consider different sets of ge-

nerators JN,N . For example, the result of reconstruction in Figure 7.6(b) for
Image Processing 397

the tomo image 257 × 257 is calculated when the set of generators is J1 =
{{(1, s); s = 0 : 256}, (0, 1)}. Figure 7.10 shows the reconstruction by another
set of 258 generators in part a, which equals J2 = {{(p, 1); p = 0 : 256}, (1, 0)}.

MSR=0.011716 MSR=0.008598

50 50

100 100

150 150

200 200

250 250
50 100 150 200 250 50 100 150 200 250
(a) 258 projections (b) 513 projections

FIGURE 7.10
The reconstruction of the tomo image 257 × 257 by (a) the set of 384 projec-
tions J2 and (b) the set J3 of 513 projections.

MSR error of the reconstruction is 0.011716. One can also consider the sym-
metric scheme of scanning, when the set of generators is combined by these
two sets, namely when

J3 = J1 ∪ J2 = {(1, s); s = 0 : 256} ∪ {(p, 1); p = 0, 2 : 256}}.

The reconstruction of the image by 513 projections with this set is shown in
part b. MSR error of this reconstruction is 0.008598.

7.3 Method of summation of line-integrals

Let P be the set of projections, i.e., the set of such triplets (p, s, t) for which
integrals w = wp,s (t) of the image along the G-rays px + ys = t are calcu-
lated. The image f(x, y) is considered on the square S = [0, 1] × [0, 1], which
is divided by the set of N 2 disjoint squares, or image elements, by the Carte-
sian lattice N × N. If a G-ray passes through the square [i, j], we denote
(i,j)
the component of the integral in this square by wt . Each G-ray is deter-
mined by the corresponding set of squares along which the ray passes. By
introducing the field functions of G-particles, we calculate the 4-dimensional
matrix ||c(n, m, i, j)|| = ||ψn,m(i, j)||, coefficients of which define the number
of G-rays passing through the squares [n, m] and [i, j] at the same time.
398 CHAPTER 7: Methods of Averaging

The backprojection method of averaging the values of all integrals along

the rays passing through the squares on S can be described briefly as follows.
Let P (n, m, i, j) be the set of G-rays passing the squares [n, m] and [i, j] at
the same time. Then, we consider the following system of equations:
c(n,m,i,j)
X X (i,j)
q(n, m) = wt , n, m = 0 : (N − 1).
[i,j]∈S, P (n,m,i,j)6=∅ t=1

Here, q(n, m) is the sum of all components of the line-integrals in the inter-
section with the square [n, m].
If we propose the value of the reconstructed image at each point (n, m) in
the Cartesian lattice by collecting all rays passing through the square [n, m]
and averaging their contribution, then this value will be calculated by
c(n,m,i,j)
X X (i,j)
¯ m) =< q(n, m) >=
f(n, < wt >.
[i,j]∈S, P (n,m,i,j)6=∅ t=1

In the model of image reconstruction under consideration with the tensor

scanning scheme, the lengths of each G-ray intersecting the squares are equal,
(i,j)
for each (p, s)-projection. This means the values < wt > are equal, i.e.,
(i,j)
<wt >= w (i,j), and we can write that
c(n,m,i,j)
X X X
f¯(n, m)= w (i,j) = ψn,m (i, j)w (i,j).
[i,j]∈S, P (n,m,i,j)6=∅ t=1 [i,j]∈S

7.4 Models with averaging

In this section, we analyze models of image reconstruction and their solutions,
without considering orthogonal basis functions. The G-ray passes through a
certain number of image elements that have, in general, a non-uniform dis-
tribution of intensities. The contribution of each image element in the line-
integral is different. One can consider that each G-ray represents a sum of
random variables, ξ. With no preliminary information about such variables,
we first can consider that the contribution of each variable is uniform with the
remaining elements of this sum. In other words, if the G-ray passes through
m image elements [ik , jk ], k = 1 : m, then the line-integral along this ray
is w = ξ[i1 ,j1 ] + ξ[i2 ,j2 ] + · · · + ξ[im ,jm ] , where we assume ξ[ik ,jk ] = w/m. By
collecting the information about all G-rays passing through image elements,
we can obtain some information about the mean or other statistics of random
variables. We will describe simple models of image reconstruction and ana-
lyze approximations of the image, by considering first the mean values. The
Image Processing 399

study of such models will allows us to approach the next steps of processing, to
achieve a better approximation of the reconstruction. What is important to us
is to keep the scanning schemes proposed in the tensor and paired transform-
based methods of image reconstruction. In such scanning schemes, there is
no coefficients in the geometry w = ξ[i1 ,j1 ] + ξ[i2 ,j2 ] + · · · + ξ[im ,jm ] , because
the length of intersection of the G-rays with any image element is the same.
This is the main principle of scanning in our algorithms, which is valid for all
projections.
We now consider the example of the nonuniform image and the lattice 2×2
on it. Let f(x, y) be the following image:
1 2
a(x, y) b(x, y) 3 4
f(x, y) = = . (7.5)
c(x, y) d(x, y) 5 6
7 8
This image has two different values on each square, as shown in Figure 7.11.
For instance, in the first image element, or square [0, 0], the image has two
values of c(x, y), 5 and 7. The image has value 5 on the area which is two
times smaller than the area with intensity 7. For simplicity of calculation, this
image is considered on the square [0, 2] × [0, 2].
Our goal is to obtain a reconstruction, which is a discrete image fn,m of
size 2 × 2. If the lattice 2 × 6 or 4 × 6 is considered on the image, then the
reconstruction of the image as the discrete image of size 2 × 6 or 4 × 6 is exact.
We here stand on the case 2 × 2. What 2 × 2-point discrete representation
for this image should be considered as the discrete analog or reconstruction
of the image? Can we consider it as one of the following images
       
12 34 23 4 6
, , , and ?
56 78 67 12 14
It is difficult to select one of these or other images, without describing the
mathematical model of image reconstruction and its solution.

7.4.1 Method of proportion

Because the tensor transform-based scanning scheme is used, we consider four
projections, which are generated by (p, s) = (1, 0), (0, 1), (1, 1), and (−1, 1).
For the given image, ten integrals of four projections can be written as
w1,0 (0) = a1 + b1 w0,1 (0) = a2 + c2
w1,0 (1) = c1 + d1 w0,1 (1) = b2 + d2
w−1,1 (0) = a3 w1,1 (0) = c4
w−1,1 (1) = c3 + b3 w1,1 (1) = a4 + d4
w−1,1 (2) = d3 w1,1 (2) = b4
The numbers ak , bk , ck , and dk define the contributions of the image elements
a(x, y), b(x, y), c(x, y), and d(x, y) in the integrals.
400 CHAPTER 7: Methods of Averaging

w(0) w(1)
2 2
1 2 1 2
w(0)
3 4 3 4
1 1
5 6 5 6
w(1)
7 8 7 8
0 0

0 0.5 1 1.5 2 0 0.5 1 1.5 2

(1,0)−projection (0,1)−projection

w(0) w(1) w(1) w(2)

2 2
1 2 1 2

3 4 w(2) w(0) 3 4
1 1
5 6 5 6

7 8 7 8
0 0

0 0.5 1 1.5 2 0 0.5 1 1.5 2

(−1,1)−projection (1,1)−projection

FIGURE 7.11
The lattice 2 × 2 on the image and four projections.

For the image f(x, y) given above, these integrals are calculated as follows:
w1,0 (0) = 3 + 4 = 7 w0,1 (0) = (1 + 3) + (5 + 7) = 16
w1,0 (1) = 7 + 8 = 15 w0,1 (1) = (2 + 4) + (6 + 8) = 20
2·3+1 7 2·7+5 19
w−1,1 (0) = = w1,1 (0) = =
3 3 3 3
(2 · 7 + 5) + (2 · 4 + 2) 29 (2 · 3 + 1) + (2 · 8 + 6) 29
w−1,1 (1) = = w1,1 (1) = =
3 3 3 3
2·8+6 22 2·4+2 10
w−1,1 (2) = = w1,1 (2) = =
3 3 3 3

The mean values of the image in image elements equal

2·3+1 7 2·4+2 10 2·7+5 19 ¯ 2 · 8 + 6 22
ā = = , b̄ = = , c̄ = = ,d= = ,
3 3 3 3 3 3 3 3
and therefore, the discrete block-mean image is considered to be

7/3 10/3 2.33 3.33

f¯ = {f¯n,m ; n, m = 0, 1} = = . (7.6)
19/3 22/3 6.33 7.33

We now consider the simple method of summation, when all line-integrals

Image Processing 401

of rays passing through the certain image element are added. The values of
a, b, c, and d are estimated as follows:
w1,0 (0) + w0,1 (0) + w−1,1 (0) + w1,1 (1) b 1 + c2 + d 4
â = = ā + ,
4 4
w1,0(0) + w0,1 (1) + w−1,1(1) + w1,1 (2) a 1 + d 2 + c3
b̂ = = b̄ + ,
4 4
w1,0 (1) + w0,1 (0) + w−1,1 (1) + w1,1 (0) d 1 + a2 + b 3
ĉ = = c̄ + ,
4 4
w1,0 (1) + w0,1 (1) + w−1,1(2) + w1,1 (1) c1 + b 2 + a 4
dˆ = = d¯ + .
4 4
For the image given above, the estimated values differ from their means as
follows: â − ā = 5.83, b̂ − b̄ = 5.83, ĉ − c̄ = 3.83, and dˆ − d¯ = 3.83. We can
conduct the evaluation of this method with respect to the exact model for
image f = [a, b; c, d]. The estimations of image values can be written as
b+c+d 3 a+b+c+d 3 e
â = ā + = ā + = (ā + ) ,
4 4 4 4 3
a+d+c 3 e
b̂ = b̄ + = (b̄ + ) ,
4 4 3
d+a+b 3 e
ĉ = c̄ + = (c̄ + ) ,
4 4 3
c+b+a 3 ¯ e
dˆ = d¯ + = (d + ) ,
4 4 3
where e = a + b + c + d. The error of estimation is thus defined as e/3.
Another algorithm with the following estimations can also be considered:
1
+ 21 w0,1 (0) + w−1,1 (0) + 12 w1,1 (1)
2 w1,0 (0)
â = ,
4
1 1 1
w1,0 (0) + 2 w0,1 (1) + 2 w−1,1 (1) + w1,1 (2)
b̂ = 2 ,
4
1
w1,0 (1) + 12 w0,1 (0) + 12 w−1,1 (1) + w1,1 (0)
ĉ = 2 ,
4
1
w1,0(1) + 12 w0,1(1) + w−1,1 (2) + 12 w1,1 (1)
dˆ = 2 .
4
For the exact model, these equations can be written as
1
2
+ b) + 12 (a + c) + a + 12 (a + d)
(a 1 e 1 e
â = = a + = (a + ) ,
4 2 8 2 4
1 1 1
(a + b) + (b + d) + (c + b) + b 1 e 1 e
b̂ = 2 2 2
= b + = (b + ) ,
4 2 8 2 4
1 1 1
(c + d) + 2 (c + a) + 2 (c + b) + c 1 e 1 e
ĉ = 2 = c + = (c + ) ,
4 2 8 2 4
1
(c + d) + 12 (b + d) + d + 12 (a + d) 1 e 1 e
dˆ = 2 = d + = (d + ) .
4 2 8 2 4
402 CHAPTER 7: Methods of Averaging

The error of estimation for this method is defined as e/4, which is less than
e/3 in the first method. This example shows that the first averaging method
results in a big error, in comparison with the second method. It also gave
us the idea of using probability characteristics of the image along each ray,
and defining a recursive method of image reconstruction. For that, we need
to know only which image elements are intersected by each G-ray.
In the method described above, we use the information of the proportio-
nality of the areas of rectangles into which the square [0, 2] × [0, 2] is divided.
In other words, we use the proportional approach of distribution of random
variables in the line-integrals. We therefore can call this method the method
of proportion. In general, such information is unknown.

7.4.2 Method with probability model

We now describe a method of image reconstruction which is based on proba-
bility characteristics of the image. In the first stage, we consider the initial
states; the probabilities of random variables along the rays equal 1/2. The
values â, b̂, ĉ, and dˆ are calculated as shown above. The second stage of itera-
tion will lead to non-uniform distribution of the contributions of each image
element in line-integrals. For example, for the integral w−1,1 (1) = c3 + b3 , the
contributions of image elements c and b are considered to be ĉ/(ĉ+ b̂)·w−1,1 (1)
and b̂/(ĉ + b̂) · w1,1 (1), respectively. By continuing such iteration, we obtain
the image reconstruction if the convergence holds for such iteration.
Thus, the initial values of â, b̂, ĉ, and dˆ are considered to be 1. Then, the
marginal probabilities along the rays equal
â b̂ â
Pab (ξ = a) = , Pab (ξ = b) = , Pac (ξ = a) = ,
â + b̂ â + b̂ â + ĉ
ĉ â dˆ
Pac (ξ = c) = , Pad (ξ = a) = , Pad (ξ = d) = ,
â + b̂ â + dˆ â + dˆ (7.7)
b̂ ĉ b̂
Pbc(ξ = b) = , Pbc (ξ = c) = , Pbd (ξ = b) = ,
b̂ + ĉ b̂ + ĉ b̂ + dˆ
dˆ ĉ dˆ
Pbd (ξ = d) = , Pcd (ξ = c) = , Pcd (ξ = d) = .
b̂ + dˆ ĉ + dˆ ĉ + dˆ
In the next stage of the iteration, new values of variables â, b̂, ĉ, and dˆ are
calculated as
Pab (ξ = a)w1,0 (0) + Pac(ξ = a)w0,1 (0) + w−1,1 (0) + Pad (ξ = a)w1,1(1)
â = ,
4
Pab (ξ = b)w1,0 (0) + Pbd (ξ = b)w0,1(1) + Pbc(ξ = b)w−1,1 (1) + w1,1 (2)
b̂ = ,
4
Pcd (ξ = c)w1,0 (1) + Pac (ξ = c)w0,1 (0) + Pbc(ξ = c)w−1,1 (1) + w1,1 (0)
ĉ = ,
4
Pcd (ξ = d)w1,0 (1) + Pbd(ξ = d)w0,1(1) + w−1,1 (2) + Pad (ξ = d)w1,1 (1)
dˆ = .
4
Image Processing 403

Consider first the implementation of this algorithm on the exact model,

when the image is f = [1, 2; 3, 4]. Below is the script run iterations.m of the
program, which can be used to accomplish the reconstruction of this image.

% ==================================================
% run_iterations.m M.M. Grigoryan, 11/17/2012
x=[1 2
3 4];
N=4;
y=mit_2by2Model2(N)
% 0.8873 2.0452
% 2.9937 3.9738
e1=norm(x-y);
% N = 1 2 3 4 10
% e1= 0.9045 0.3723 0.1326 0.1215 0.2198
% --------------------------------------------------
function y=mit_2by2Model2(N)
% image is x=[1 2; 3 4];
w101=3; w102=7;
w011=4; w012=6;
w1m11=1; w1m12=5; w1m13=4;
w111=3; w112=5; w113=2;
% initial values of states:
ap=1; bp=1; cp=1; dp=1;
for i1=1:N
Pab_a=ap/(ap+bp); Pab_b=1-Pab_a;
Pac_a=ap/(ap+cp); Pac_c=1-Pac_a;
Pad_a=ap/(ap+dp); Pad_d=1-Pad_a;
Pbc_b=bp/(bp+cp); Pbc_c=1-Pbc_b;
Pbd_b=bp/(bp+dp); Pbd_d=1-Pbd_b;
Pcd_c=cp/(cp+dp); Pcd_d=1-Pcd_c;
ap=(Pab_a*w101+Pac_a*w011+w1m11+Pad_a*w111)/4/2;
bp=(Pab_b*w101+Pbd_b*w012+Pbc_b*w1m12+w113)/4/2;
cp=(Pcd_c*w102+Pac_c*w011+Pbc_c*w1m12+w111)/4/2;
dp=(Pcd_d*w102+Pbd_d*w012+w1m13+Pad_d*w112)/4/2;
y=[ap, bp; cp, dp];
end
y=2*y;
% ==================================================

With N = 1, 2, 3, and 4 steps of iteration, we obtain, respectively, the

following approximations of the image:
       
1.50002.2500 1.13982.1261 0.97182.0668 0.88732.0452
, , ,
2.75003.2500 2.88663.6895 2.95493.8884 2.99373.9738

The curve of the mean-square-error of approximation of the image on the

first 30 iterations is shown in Figure 7.12, along with the approximation of
the image after ten iterations. The convergence takes place in this algorithm,
404 CHAPTER 7: Methods of Averaging

and the minimum error (0.1215) of approximation is achieved on the 4-th

iteration.
1

0.8
f(10)=[0.7883 2.0486
0.6 3.0425 4.0388]

0.4

0.2

0
0 5 10 15 20 25 30

FIGURE 7.12
Errors of reconstruction of the image.

To apply the described method to the image f(x, y) with a non-uniform

distribution of intensities on image elements, we use the code with script
mit 2by2Model1.m, which is similar to the script mit 2by2Model2.m, where only
the values of integrals have been changed, as shown below.
% ==================================================
function y=mit_2by2Model1(N)
% integrals wps(t) for 4 projections
w101=7; w102=15;
w011=16; w012=20;
w1m11=7/3; w1m12=29/3; w1m13=24/3;
w111=19/3; w112=29/3; w113=10/3;
...
% ==================================================

With N = 2, 3, 4, and 10 steps of iteration, we obtain, respectively, the

following approximations of the image:
       
3.47864.9806 3.07794.7116 2.85634.5628 2.55104.4206
, , , .
7.11388.4610 7.40408.8971 7.58899.1112 7.86019.3220

7.4.3 Reconstruction of the shifted image

In this section, we evaluate the method of proportion for an image which is
not on the lattice 2 × 2. To simplify our calculations, we modify the image
given in (7.5) in the following way. In the square [0, 2] × [0, 2], we consider the
image
0 0
a(x, y) b(x, y) 3 4
f(x, y) = = . (7.8)
c(x, y) d(x, y) 5 6
7 8
Image Processing 405

This image and rays of four projections are shown in Figure 7.13.

w(0)=15 w(1)=18
2 2
0 0 0 0
w(0)=7
3 4 3 4
1 1
5 6 5 6
w(1)=15
7 8 7 8
0 0

0 0.5 1 1.5 2 0 0.5 1 1.5 2

(1,0)−projection (0,1)−projection

w(0)=2 w(1)=9 w(1)=28/3 w(2)=8/3

2 2
0 0 0 0

3 4 w(2)=22/3 w(0) 3 4
1 1
5 6 5 6

7 8 7 8
0 0
w(0)=19/3
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(−1,1)−projection (1,1)−projection

FIGURE 7.13
The lattice 2 × 2 on the image and four projections.

It is not difficult to calculate all ten line-integrals for this image,

w1,0 (0) = a1 + b1 = 7 w0,1 (0) = a2 + c2 = 15

w1,0 (1) = c1 + d1 = 15 w0,1 (1) = b2 + d2 = 18
w−1,1(0) = a3 = 2 w1,1 (0) = c4 = 19
3
(7.9)
w−1,1(1) = c3 + b3 = 9 w1,1 (1) = a4 + d4 = 28
3
w−1,1(2) = d3 = 22 3 w1,1 (2) = b4 = 83

The block-mean discrete image is considered to be

2 8/3 2 2.67
f¯ = = .
19/3 22/3 6.33 7.33

To accomplish the calculation of image reconstruction, we use the code with

script run iterations3.m, which is given below. The function mit 2by2Model3.m
is similar to mit 2by2Model2.m, where only the values of line-integrals are taken
from (7.9). This main program calculates the approximation of the image on
the 4-th stage of iteration and the mean-square distance. The distance is
relative to the block-mean image f¯.
406 CHAPTER 7: Methods of Averaging

% ==================================================
% run_iterations3.m A.-M. Grigoryan, 01/11/2012
N=3;
y3=mit_2by2Model3(N)
% 2.9209 4.1578
% 7.2468 8.3708
% the block-mean image
xm =[2 8/3
19/3 22/3];
d1=norm(xm-y3) % 2.2247
% ------------------------------------
function y=mit_2by2Model3(N)
% integrals wps(t) for 4 projections
w101=7; w102=15;
w011=15; w012=18;
w1m11=2; w1m12=9; w1m13=22/3;
w111=19/3; w112=28/3; w113=8/3;
...
% ==================================================

With N = 3, 4, and 5 steps of iteration, we obtain, respectively, the follo-

wing approximations fˆ(N ) of the image:
     
2.92094.1578 2.70904.0082 2.58813.9326
fˆ(3) = , fˆ(4) = , fˆ(5) = .
7.24688.3708 7.42588.5796 7.53218.6840

The curve of the mean-square distance of approximation of the image in

the first 30 iterations is shown in Figure 7.14, along with two approximations
of the image after three and ten iterations. The minimum distance (2.2247)
of approximation is achieved on the 3-rd iteration.

3.2

3
f(3)=[2.9209 4.1578 f(10)=[2.4215 3.8745
2.8 7.2468 8.3708] 7.6722 8.7858]

2.6

2.4

2.2
0 5 10 15 20 25 30

FIGURE 7.14
The distance of the approximation to the block-mean image.
Image Processing 407

7.4.4 Method of minimization of error

We now consider another approach and find the approximation of the re-
constructed image by minimizing the mean-square error. For that, we again
consider the example 2 × 2, with the known integrals, which we denote by yk
as follows:
w1,0(0) = a1 + b1 = y0 w0,1 (0) = a2 + c2 = y2
w1,0(1) = c1 + d1 = y1 w0,1 (1) = b2 + d2 = y3
w1,1(0) = c4 = y4 w−1,1 (0) = a3 = y7
w1,1(1) = a4 + d4 = y5 w−1,1 (1) = c3 + b3 = y8
w1,1(2) = b4 = y6 w−1,1 (2) = d3 = y9

To find the approximation of values for a, b, c, and d, we demand the min-

imum of the mean-square-root error

L =(a + b − y0 )2 + (c + d − y1 )2 + (a + c − y2 )2 + (b + d − y3 )2 + (c − y4 )2
+(a + d − y5 )2 + (b − y6 )2 + (a − y7 )2 + (c + b − y8 )2 + (d − y9 )2 .

By differentiating this expression by a, b, c, and d, we obtain the following

system of linear equations:

(a + b − y0 ) + (a + c − y2 ) + (a + d − y5 ) + (a − y7 ) = 0,
(a + b − y0 ) + (b + d − y3 ) + (b − y6 ) + (c + b − y8 ) = 0,
(c + d − y1 ) + (a + c − y2 ) + (c − y4 ) + (c + b − y8 ) = 0,
(c + d − y1 ) + (b + d − y3 ) + (a + d − y5 ) + (d − y9 ) = 0.

These equations can be written in a new form and new variables Ωx, which
we introduce for x = a, b, c, d, as follows:

(a + b) + (a + c) + (a + d) + a=Ωa = y0 + y2 + y5 + y7 ,
(a + b) + (b + d) + b + (c + b) =Ωb = y0 + y3 + y6 + y8 ,
(c + d) + (a + c) + c + (c + b) =Ωc = y1 + y2 + y4 + y8 ,
(c + d) + (b + d) + (a + d) + d =Ωd = y1 + y3 + y5 + y9 .

We write this system of linear equations in matrix form

    
4 1 1 1 a Ωa
1 4 1 1  b   Ωb 
   =   (7.10)
1 1 4 1   c   Ωc 
1 1 1 4 d Ωd

as well as its solution

    
a 6 −1 −1 −1 Ωa
b
 = 1  −1 6 −1 −1   Ωb 
 c  21  −1 −1 6 −1   Ωc  . (7.11)
  

d −1 −1 −1 6 Ωd
408 CHAPTER 7: Methods of Averaging

The matrix 4 × 4 can be written as

     
6 −1 −1 −1 1 0 0 0 −1 −1 −1 −1
 −1 6 −1 −1  0 1 0 0  −1 −1 −1 −1 
 −1 −1 6 −1  = 7  0
   + .
0 1 0   −1 −1 −1 −1 
−1 −1 −1 6 0 0 0 1 −1 −1 −1 −1

By denoting the sum of projections by Sw = Ωa +Ωb +Ωc +Ωd , the solution

can be written as    
a Ωa
 b  1  Ωb 
  =   − 1 Sw . (7.12)
 c  3  Ωc  21
d Ωd
We now describe this solution for the case when the image f(x, y) is de-
termined by (7.5). To accomplish the calculations, we consider the program
with the following script mse 2by2Model1.m.

% call: mse_2by2Model1.m A.-M. Grigoryan, Jan 11, 2012

% integrals wps(t) for 4 projections
w101=7; w102=15;
w011=16; w012=20;
w111=19/3; w112=29/3; w113=10/3;
w1m11=7/3; w1m12=29/3; w1m13=24/3;
% =======================================
y0=w101; y1=w102;
y2=w011; y3=w012;
y4=w111; y5=w112; y6=w113;
y7=w1m11; y8=w1m12; y9=w1m13;
Qa=y0+y2+y5+y7;
Qb=y0+y3+y6+y8;
Qc=y1+y2+y4+y8;
Qd=y1+y3+y5+y9;
Q=[Qa Qb Qc Qd];
A=Q/3-sum(Q)/21;
y=reshape(A,2,2)’
% 3.3492 5.0159
% 7.3492 9.2381
% =======================================

The result of calculations is the approximation

 
3.3492 5.0159
f¯ = .
7.3492 9.2381

We now change the values of all ten integrals, yk , for the image f = [1, 2; 3, 4],
as shown below in the script mse 2by2Model2.m.

% call: mse_2by2Model2.m A.-M. Grigoryan, Jan 11, 2012

Image Processing 409

% integrals wps(t) for 4 projections

w101=3; w102=7;
w011=4; w012=6;
w1m11=1; w1m12=5; w1m13=4;
w111=3; w112=5; w113=2;
% =======================================
... ... ...
A=Q/3-sum(Q)/21;
y=reshape(A,2,2)’
% 1.0000 2.0000
% 3.0000 4.0000
% =======================================

The calculation by this program results in the exact reconstruction of the

image.
Thus we obtain the simple solution in (7.12) of the image reconstruction,
when stating the problem in terms of minimization of the mean-square error.
The only question is whatever this solution may have negative values. We
leave the answer to this question to the reader as an exercise.

7.4.5 Corpuscular approach

In the considered tensor scanning scheme, four rays pass through each image
element in the lattice 2 × 2. The geometrical presentation of each image ele-
\/
ment as a particle has the form of . In matrix form, the following matrices
/\
correspond to four particles:
       
41 14 11 11
φ0,0 = , φ0,1 = , φ1,0 = , φ1,1 = .
11 11 41 14

Presentation of the field by the particles is

 
4a + b + c + d a + 4b + c + d
A = aφ0,0 + bφ0,1 + cφ1,0 + dφ1,1 = .
a + b + 4c + d a + b + c + 4d

Considering this 2-D field in the form of a 1-D vector

    
4a + b + c + d 4111 a
 a + 4b + c + d   1 4 1 1  b 
 a + b + 4c + d  =  1 1 4 1  c  ,
    

a + b + c + 4d 1114 d

we obtain the same equation as in (7.10). If we consider the inverse matrix

given in (7.11) as a field of particles, which we call anti-particles, these anti-
particles in matrix form will be written as
       
6 −1 −1 6 −1 −1 −1 −1
ψ0,0 = , ψ0,1 = , ψ1,0 = , ψ1,1 = .
−1 −1 −1 −1 6 −1 −1 6
410 CHAPTER 7: Methods of Averaging

Normalizing the particles as

1 1
φn,m = √ φn,m , ψn,m = √ ψn,m , n, m = 0, 1,
21 21
we obtain the orthogonality between particles and anti-particles
(φn,m , ψn1,m1 ) = δn,n1 δm,m1 , n, m, n1 , m1 = 0, 1.
It should be noted that the anti-particles ψn,m do not have representation
in the form of four directions as the particles φn,m. One may think about
simplification of the orthogonal basis for effective realization. For instance, let
us remove one projection from the scanning scheme, for instance, the (−1, 1)-
\
projection. In this case, the geometrical presentation of the particle is , and
\
four particles are defined as
       
31 13 10 11
φ0,0 = , φ0,1 = , φ1,0 = , φ1,1 = .
11 01 31 13
The field defined by these particles is described by the following matrix:
   
3111 21 −6 −6 −3
 1 3 0 1 1 
 −6 20 4 −6  .

R=  1 0 3 1 ,
 and R−1 =
48 −6 4 20 −6 

1113 −3 −6 −6 21
Then, the set of anti-particles becomes “heavy,” in the sense that the coef-
ficients of the matrices of these particles contain large numbers, when com-
pared with the original case. Indeed, the following matrices determine these
anti-particles:
       
21 −6 −6 20 −6 4 −3 −6
ψ0,0 = , ψ0,1 = , ψ1,0 = , ψ1,1 = .
−6 −3 4 −6 20 −6 −6 21
When using four projections, one can observe the movement of the particles
from one to another. The same movement holds for their anti-particles. The
movement is periodic, and four consecutive movements correspond to one full
circle. What we have just seen is the fact that the removal of one projection
breaks this property.
If in addition we remove another diagonal projection, the (1, 1)-projection,
and consider the geometry of the particle as , the particles will be defined
as        
21 12 10 01
φ0,0 = , φ0,1 = , φ1,0 = , φ1,1 = ,
10 01 21 12
and we obtain a singular field
 
2 1 1 0
1 2 0 1
R=
1
, det(R) = 0.
0 2 1
0 1 1 2
Image Processing 411

These examples show that for a large size N ×N of image in reconstruction,

an extra number of projections may significantly simplify the structure of the
anti-particles, which in turn will result in an effective reconstruction of the
image. We may assume that the tensor scanning scheme is one of the “optimal”
schemes in image reconstruction.

7.5 General case: Probability model

In this section, we describe the probability model method for reconstructing
the image of size N × N, where N > 2. Each G-rays lp,s (t) of the (p, s)-
projection is described by the following set of triplets:

lp,s (t) = {(nk , mk , ∆lk ); k = 1 : card lp,s (t)} (7.13)

where the pairs (nk , mk ) define the image elements (IEs) trough which the
G-ray passes. We denote by I(p, s) the set of these IEs, and by card its cardi-
nality. It is the discrete length of the G-ray. ∆lk are the lengths of intersections
of the ray with the corresponding k-th IEs in the set I(p, s).
On the first stage of the algorithm, the line-integral wp,s (t) along the G-ray
is distributed between image elements [nk , mk ] of the set I(p, s) in proportion
to the lengths ∆lk ,

(n ,mk )
k ∆lk
wp,s,t (nk , mk ) = wp,s,t = wp,s (t) P . (7.14)
∆li

Here, the summation of ∆li is by all image elements of the set I(p, s). The
intersection of the G-ray with each image element has the same length. There-
fore, the above equation can be written as
wp,s (t)
wp,s,t (nk , mk ) = , k = 1 : card lp,s (t). (7.15)
card lp,s (t)

The values of the discrete image are therefore defined as

1 X
an,m = wp,s,t (n, m), n, m = 0 : (N − 1). (7.16)
c(n, m)
(p,s,t)

Here c(n, m) is the number of G-rays lp,s (t) passing through the image ele-
ment, or square [n, m]. The summation of wp,s,t (n, m) is by all triplets (p, s, t)
defining such G-rays. In the considered tensor scheme of scanning the image
by G-rays, the matrix ||c(n, m)|| is known, and there is no need to calculate
these coefficients. All coefficients c(n, m) are equal in this matrix.
To demonstrate the result of approximation of the image, an,m, on this
first stage, we consider the following example when N = 64. The gray-scale
412 CHAPTER 7: Methods of Averaging
TABLE 7.2
Integer data of rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 10 15 10 15 1
2 12 40 18 10 2
3 40 40 10 20 1
4 45 10 5 20 3

Rectangles and grid 64x64 1 iteration

1

10
0.8 3
2 20
0.6
30

0.4 40
1
4
50
0.2

60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
(a) (b)

FIGURE 7.15
(a) Image with 4 rectangles on the square 64 × 64 and (b) the first iteration.

image on the square [0, 1] × [0, 1] is composed of four rectangles with the data
given in Table 7.2. The data is normalized with respect to the lattice 64 × 64.
Figure 7.15 shows these four rectangles in part a. The result of the calculation
of the image an,m after the first iteration is shown in part b. The appearance
of each rectangle is spread in two major directions, vertical and horizontal.
On the first stage of the algorithm, which we also call the zero stage, we
receive the image which is considered to be a rough estimation of the original
image; it gives us some information about the probability of the image.
We now consider the next stage of the reconstruction. The line-integrals
are considered along the new image an,m . The values of the new image on
the squares [nk , mk ], through which the ray lp,s (t) passes, will be denoted
by s1p,s,t (nk , mk ). Along this ray, the probability density function at point
(nk , mk ), which is referred to as the square [nk , mk ], is calculated by
(1)
(1) sp,s,t (nk , mk )
pp,s,t (nk , mk ) = X (1) , k, (i) = 1 : card lp,s (t). (7.17)
sp,s,t (ni , mi )
i

Therefore, the calculation in (7.15) will be changed to

(1) (1)
wp,s,t (nk , mk ) = wp,s (t)pp,s,t (nk , mk ), k = 1 : card lp,s (t). (7.18)
Image Processing 413

The image at this stage is calculated by

1 X (1)
a(1)
n,m = wp,s,t (n, m), n, m = 0 : (N − 1). (7.19)
c(n, m)
(p,s,t)

This process will be repeated a few times, and the probability density
functions will be recalculated on each new image, until the given number of
iterations is accomplished. At each new stage of the iteration, the obtained
(n)
image an,m , n ≥ 1, is used to get probability information for the next image,
by calculating the probability density function by
(n+1)
(n+1) sp,s,t (nk , mk )
pp,s,t (nk , mk ) = X (n+1) , k, (i) = 1 : card lp,s (t). (7.20)
sp,s,t (ni , mi )
i

Here, sn+1 denotes the values of the new image on the squares [nk , mk ]
p,s,t (nk , mk )
through which the ray lp,s (t) passes. The image at this stage is calculated by
1 X (n+1)
a(n+1)
n,m = wp,s,t (n, m), n, m = 0 : (N − 1), (7.21)
c(n, m)
(p,s,t)

and the distribution of the line-integral among the image elements [nk , mk ] is
defined as
(n+1) (n+1)
wp,s,t (nk , mk ) = wp,s (t)pp,s,t (nk , mk ), k = 1 : card lp,s (t). (7.22)
As an example, we consider the same image 64 × 64 with four rectangles.
Figure 7.16 illustrates the process of reconstruction of this image on diffe-
rent stages of the calculation. In parts a, b, c, and d, the results of image
reconstruction after 5, 10, 20, and 50 iterations are shown, respectively.
The described process of iterations converges and results in a good-quality
reconstruction. Figure 7.17 shows the curve of differences between the approx-
imations after k and k − 1 iterations, where k = 2 : 100,
v
uN−1 N−1
(k) (k−1) 1 u X X (k) (k−1)
εk = ε(a , a ) = 2t [an,m − an,m ]2 , (N = 64).
N n=0 n=0

The difference εk decreases exponentially to zero, and ε100 = 8.3545 · 10−6 .

The approximation process results in a high-quality image when N increases.
Figure 7.18 shows the results of image reconstruction after 100 and 200 iter-
ations in parts a and b, respectively. The following is important to mention.
Given the number of iterations, the final image should be normalized by the
factor M1 , which is approximately equal to 1/N, for large N. This factor is
used to display the image, and calculate the differences εk and the errors of
reconstruction. This factor is not used in the process of iterations described
above. Below are a few values of the factor M1 :
N 8 16 32 64 128 256
M1 0.1362667 0.06465907 0.03160831 0.0156805 0.0078208 0.0039045
414 CHAPTER 7: Methods of Averaging

5 iterations 10 iterations

10 10

20 20

30 30

40 40

50 50

60 60
20 40 60 20 40 60
(a) (b)

20 iterations 50 iterations

10 10

20 20

30 30

40 40

50 50

60 60
20 40 60 20 40 60
(c) (d)

FIGURE 7.16
Reconstruction of the image with 5, 10, 20, and 50 iterations.

−3
x 10
2

1.5

0.5

0
0 10 20 30 40 50 60 70 80 90 100

FIGURE 7.17
Difference between the first 100 iterations.

7.5.0.1 Code of the reconstruction

To implement the proposed algorithm with the probability model, and print
all figures described above for this method, we use the program with the script
mreconstruction bypm1.m, which is given below. The image is composed of four
Image Processing 415
100 iterations 200 iterations

R#3
10 10
R#1

20 20

30 30
R#2

40 40 R#4

50 50

60 60

10 20 30 40 50 60 10 20 30 40 50 60
(a) (b)

FIGURE 7.18
Reconstruction of the image with (a) 100 and (b) 200 iterations.

rectangles with data given in Table 7.2. The number of iterations is 200. This
program uses the same functions tensor scheme pst, general scheme pst2, and
fieldpst tensor scheme, which are described in Section 6.3 in the program
reconstruction byR.m.

% -------------------------------------------------------
% call: mreconstruction_bypm1.m / A.-M. Grigoryan, 2011
N=64;
M=3*N*N/8+2; M1=1/N;
delta=1/N;
N_rect=4;
MNrects=[10 15 20 30
12 40 30 50
40 40 50 60
45 10 50 30];
Matr_rect=delta*MNrects;
E=[1 2 1 3]; % intensities of 4 rectangles
% Plot the grid NxN
figure; colormap(gray);
subplot(1,2,1);
draw_grid; axis image; pause(1);
% ===================== Processing ========================
ps=ps_generators(N); L=size(ps,1);
eps100=eps*100;
N_iterations=200; % number of iteration
subplot(1,2,2);
for iter=1:N_iterations % loop for iterations
Y=zeros(N,N);
for k=1:L % loop for (p,s)-projections
p=ps(k,1); s=ps(k,2);
%----------------------------------
[p1,s1,t]=tensor_scheme_pst(p,s,N);
416 CHAPTER 7: Methods of Averaging

%----------------------------------
Length_t=length(t);
for it=1:Length_t % loop for (p,s,it) G-rays
sum_E=fieldpst_tensor_scheme(p1,s1,t(it), ...
N,N_rect,Matr_rect,E);
if (sum_E>eps)
f=general_scheme_pst2(p1,s1,t(it),N);
f_sz2=size(f,2);
L0=0;
for nf=1:f_sz2
L0=L0+f(3,nf);
end
if iter==1
for nf=1:f_sz2
m1=f(1,nf); n1=f(2,nf);
Y(m1,n1)=Y(m1,n1)+sum_E/L0;
end
else
% ========= STATISTICS =============
% Integrate for statistical rays
S_po=0;
for nf=1:f_sz2
m1=f(1,nf); n1=f(2,nf);
S_po=S_po+R(m1,n1);
end
if (S_po>eps100)
% Calculate statistical rays
for nf=1:f_sz2
m1=f(1,nf); n1=f(2,nf);
Y(m1,n1)=Y(m1,n1)+ss*R(m1,n1);
end
end
% ===================================
end
end
end % for loop: (p,s,t)-rays
end % for loop: (p,s)-projections
R=Y/M ;
% display the normalized reconstruction on this stage
subplot(1,2,2);
if iter==1
XR=R;
else
XR=R/M1;
end
imagesc(XR); axis image;
title(sprintf(’ %g Iterations’,iter)); pause(0.1);
end % for loop: iterations
% --------------------------------------------------------
Image Processing 417

In conclusion, we consider the example when image f(x, y) is composed of

rectangles that are not on the grid. Let f(x, y) be the gray-scale image on the
square [0, 1] × [0, 1], which is composed of four rectangles with the data given
in Table 7.3. The data is normalized with respect to the lattice 64 × 64.

TABLE 7.3
Integer data of rectangles
k x0k yk0 ∆x0k ∆yk0 Rk
1 10 15 10 15 + 0.25 1
2 12 + 0.1 40 18 10 2
3 40 40 + 0.2 10 20 1
4 45 10 5 + 0.5 20 3

Figure 7.19 shows four shifted rectangles on the square [0, 1] × [0, 1] in part
a, along with the results of image reconstruction after 10 iterations in part b.

Rectangles are not on the grid 10 iterations

1

10
0.8 3
2 20
0.6
30

0.4 40
1
4
50
0.2

60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
(a) (b)

FIGURE 7.19
(a) Image with 4 rectangles and (b) image approximation after 10 iterations.

Figure 7.20 shows the results of image reconstruction after 50 and 100
iterations in parts a and b, respectively.
418 CHAPTER 7: Methods of Averaging

50 iterations 100 iterations

10 10

20 20

30 30

40 40

50 50

60 60

10 20 30 40 50 60 10 20 30 40 50 60
(a) (b)

FIGURE 7.20
Reconstruction of the image after (a) 50 and (b) 100 iterations.

Problems
Problem 7.1 (Discrete Model) Consider the following two images on the
square [0, 1] × [0, 1] :
8
< 2; if x ∈ [0.6875, 0.8125], y ∈ [0.750, 0.875],
f1 (x, y) = 3; if x ∈ [0.1875, 0.3125], y ∈ [0.125, 0.250],
0; otherwise,
:

(x − 0.5)2 (y − 0.5)2
8
1
1; if + ≤ ,
<
f2 (x, y) = a2 b2 4
0; otherwise.
:

where a = 0.25 and b = 0.5. Figure 7.21 shows rectangles and the ellipse to-
gether in part a, along with the discrete representation of the image composed
of these figures in b. The discrete image is of size 61 × 61, and 61 is prime.
Therefore, only 62 projections are required for the reconstruction of the image
in b.
A. Use the tensor method of summation by (7.4) and calculate the recon-
struction of the image f1 (x, y) on the Cartesian lattice 61 × 61.
B. Repeat A for the image f2 (x, y), as well as for the sum of images
f(x, y) = f1 (x, y) + f2 (x, y).
Display the results and calculate errors of reconstruction in A, B, and C.

Problem 7.2 (Continuous Model)∗ Consider the following three-rectangle

image on the square [0, 1] × [0, 1] :

 1; if x, y ∈ [3/8 − 1/16, 3/8 + 1/16],

2; if x, y ∈ [1/2 − 1/16, 1/2 + 1/16],
f(x, y) =
 1; if x, y ∈ [5/8 − 1/16, 5/8 + 1/16],

0; otherwise.
Image Processing 419
objects and grid 61x61 image on the grid 61x61
1

10
0.8 1

20
0.6
3 30

0.4
40

0.2 2 50

0 60
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
(a) (b)

FIGURE 7.21
(a) Rectangles and ellipse on [0, 1] × [0, 1] and (b) the discrete image 61 × 61.

A. Use the tensor method of summation by (7.4) and calculate the recon-
struction of this image on the Cartesian lattices 128 × 128 and 256 × 256.
Calculate all required line-sums pϑ (np + ms) through the line-integrals, by
transferring the geometry of G-rays to A-rays. Display the results and calcu-
late the MRS error of reconstruction in both cases.
B. Find a way to improve the results of the image reconstruction by re-
moving the shadows created by these rectangles. Show your results.

Problem 7.3 Consider the image on the square [0, 1] × [0, 1], which is com-
posed of the sum of seven circles with the following coordinates (xk , yk ), radii
rk , and intensities Ik :

k 1 2 3 4 5 6 7
xk 6 12 22 20 16 5 28
yk 10 13 15 6 23 26 28
rk 3 5 7 4 7 2 1
Ik 5 2 3 4 5 5 6

The data of coordinates and radii are given in integer form, and should be
normalized by the factor of N, when reconstructing on the lattice N × N.
A. Use the tensor method of summation and calculate the reconstruction
of this image on the Cartesian lattices 128 × 128 and 256 × 256. Display the
results and calculate the MRS error of reconstruction in each case.
B. Remove the vertical and horizontal projections in the tensor scheme of
scanning and reconstruct the image on the lattices 128 × 128 and 256 × 256.
C. Compare MRS errors of reconstruction in A and B. If errors of recon-
struction in B are smaller than in A, explain this fact.
420 CHAPTER 7: Methods of Averaging

Problem 7.4 (Probability Model) Consider the following image on the

square [0, 1] × [0, 1] :
0 0 0 0 0 0 0 0
0 0 1 1 1 1 1 0
0 0 1 1 1 1 1 0
0 0 1 1 1 1 1 0
[f(x, y); 0 ≤ x, y ≤ 1 ] = .
0 0 1 1 3 1 1 0
0 0 1 1 1 1 1 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

Each square is referred to as an image element of size 1/8 × 1/8. Use the
method with statistics and calculate the 8 × 8 reconstruction of this image
after 1, 2, 20, and 30 iterations. Show the result and calculate the MSR error
of reconstruction for each of these iterations.
Problem 7.5 (Probability Model) Consider the following image on the
square [0, 1] × [0, 1] :
1 0 0 0 0 0 1
0 0 0 1 0 0 0
0 0 0 1 0 0 0
[f(x, y); 0 ≤ x, y ≤ 1 ] = 0 1 1 7 1 1 0 .
0 0 0 1 0 0 0
0 0 0 1 0 0 0
1 0 0 0 0 0 1

Each square is referred to as an image element of size 1/7 × 1/7. Use the
method with statistics and calculate the 7 × 7 reconstruction of this image
after 10, 20, and 30 iterations. Show the result and calculate the MRS error
of reconstruction for each of these iterations.
Problem 7.6 (Probability Model) Consider the following image on the
square [0, 1] × [0, 1]:
0 0 2 2 0 2 0 0
1 1 0 0 4 6 4 0
1 1 2 0 0 2 0 1
0 0 2 0 0 2 0 1
[f(x, y); 0 ≤ x, y ≤ 1 ] = .
2 0 2 0 0 2 0 0
0 4 0 1 1 0 0 0
3 3 3 0 0 2 2 2
0 0 0 2 2 0 0 0

Each square is referred to as an image element of size 1/8 × 1/8.

A. Use the method with statistics and calculate the 8 × 8 approximation
of this image after 10, 50, and 100 iterations.
B. Calculate the errors of approximation in 10, 50, and 100 iterations.
C. Repeat A and B for the 16 × 16 approximation of the image.
Image Processing 421

Problem 7.7 In the program mreconstruction bypm1.m,P which is given in

§7.5.0.1, the calculation of normalized integrals wp,s (t)/ i ∆li in (7.18) is
accomplished for each projection inside the loop for iterations.
A. Improve this program and write a new program which calculates all
these values separately, saves them, and then uses these integrals during the
next process of iterations.
B. Determine if your program performs all calculations in MATLAB faster
than the program mreconstruction bypm1.m for the same four rectangles shown
in Figure 7.15 in part a. Show your results for the N = 64 and 128 cases.

Problem 7.8 Consider the image on the square [0, 1] × [0, 1], which is com-
posed of the sum of seven rectangles with the following coordinates [xk,1, yk,1,
xk,2 , yk,2] and intensities Ik :

k 1 2 3 4 5 6 7
xk,1 10 12 40 45 29 20 30
yk,2 15 40 40 10 29 7 2
xk,1 20 30 50 50 35 40 33
yk,2 30 50 60 30 35 10 16
Ik 1 2 1 3 2 1 1

The data of coordinates are given in integer form, and should be normalized
by the factor of 64.
A. Use the method with statistics and the set of 96 generators (p, s)
J 1 = {(1, s), s = 0 : 63} ∪ {(2p, 1), p = 0 : 31}, and calculate the 64 × 64
approximation of this image after 10, 50, and 100 iterations.
B. Repeat A for the following set of 127 generators: J 3 = {(1, s), s = 0 :
63} ∪ {(p, 1), p = 0, 2 : 63}.
C. Show the results in A and B, calculate the MSR errors of the obtained
approximations, and compare the results.

Problem 7.9 Consider the image composed from the ten circles and ellipses,

10  (x − xk )2 (y − yk )2
X Ik ; if + ≤ rk2
f(x, y) = a2k b2k

k=1 0; otherwise

whose data (in integer form) and intensities Ik are the following:

k 1 2 3 4 5 6 7 8 9 10
xk 24 48 88 80 64 20 112 48 24 116
yk 40 52 60 24 92 104 112 16 76 24
ak 1 1 1 5 5 1 1 1 1 1
bk 1 1 1 4 8 1 1 1 1 1
rk 12 20 28 4 4 8 4 4 4 4
Ik 20 8 12 16 20 20 24 20 24 24

The data of coordinates are given in integer form, and should be normalized
422 CHAPTER 7: Methods of Averaging

by the factor of N = 128. Use the method with statistics and the set of 3N/2
generators (p, s) J 1 = {(1, s), s = 0 : N − 1} ∪ {(2p, 1), p = 0 : N/2 − 1}, and
calculate the N × N approximation of this image after 50 and 100 iterations.
Show the results and calculate the MRS errors of the obtained approximations.

Problem 7.10 Consider the image composed of one ellipse,


 [(x − 0.5) cos α − (y − 0.5) sin α]2

 1; if +

 0.22
f(x, y) = [(x − 0.5) sin α + (y − 0.5) cos α]2
 + ≤1


 0.12

0; otherwise

where x, y ∈ [0, 1], and the angle α = 30◦.

A. Calculate the results of reconstruction of this image on the lattice 128×
128 after performing 20 and 50 iterations. Round the results of reconstruction
and print them.
B. Sample the given image f(x, y) on the lattice 128 × 128 and calculate
the error of reconstructions in A.
Bibliography

[1] J.W. Cooley and J.W. Tukey, “An algorithm the machine computation of
complex Fourier series,” Math. Comput, vol. 9, no. 2, pp. 297-301, 1965.
[2] D.F. Elliot and K.R. Rao, Fast Transforms: Algorithms, Analyzes, and Appli-
cations, New York, Academic, 1982.
[3] L. Rabiner and R.W. Schaefer, Speech Signal Processing, Prentice Hall, Engle-
wood Cliffs, New Jersey, 1983.
[4] R.E. Blahut, Fast Algorithms for Digital Signal Processing, Addison-Wesley,
Reading, Massachusetts, 1985.
[5] P. Duhamel and M. Vetterli, “Fast Fourier transforms: A tutorial review and
state of the art,” Signal Processing, vol. 19, pp. 259-299, 1990.
[6] Handbook for Digital Signal Processing, edited by J.F. Kaiser and S.K. Mitra,
Wiley-Interscience, New York, 1993.
[7] W.K. Pratt Digital Image Processing, Wiley-Interscience, New York, 1978.
[8] A. Rosenfeld and C. Kak, Digital Image Processing, vols. I and II. Academic
Press, New York, 1982.
[9] Anil K. Jain, Fundamentals of Digital Image Processing, Prentice Hall, New
Jersey, 1989.
[10] R.C. Gonzalez and R.E. Woods, Digital Image Processing, 2nd edition, Prentice
Hall, New-Jersey, 2002.
[11] R.V. L. Hartley, A more symmetrical Fourier analysis applied to transmission
problems,” Proc. IRE, 30, pp. 144-150, March 1942.
[12] R.N. Bracewell, The Hartley Transform, Oxford University Press, 1986.
[13] R.N. Bracewell, O. Buneman, H. Hao, and J. Villasenor, “Fast two-dimensional
Hartley transform,” Proc. IEEE, 1986, 74, pp. 1282-1283.
[14] J.L. Wu and Pei Soo-Chang, “The vector split-radix algorithm for 2-D FHT,”
IEEE Transactions on Signal Processing, 1993, vol. 41, no. 2, pp. 960-965.
[15] G. Bi, “New split-radix algorithm for the discrete Hartley transform,” IEEE
Transactions on Signal Processing, February 1997, vol. 45, no. 2, pp. 297-302.
[16] N. Ahmed and R. Rao, Orthogonal Transforms for Digital Signal Processing,
Springer-Verlag, Berlin, 1975.
[17] K.R. Rao and P. Yip, Discrete Cosine Transform—Algorithms, Advantages,
Applications, Academic Press, Boston, 1990.
[18] W.B. Pennebaker and J.L. Mitchell, JPEG Still Image Compression Standard,
Van Nostrand Reinhold, New York, 1993.

423
424 Image Processing

[19] G. Mandyam, N. Ahmed, and N. Magotra, “Lossless image compression using

the discrete cosine transform,” JVCIR, vol. 8, no. 1, pp. 21-26, Mar. 1997.
[20] M. Iwahashi, N. Kambayashi, and H. Kiya, “Bit reduction of DCT basis for
transform coding,” ECJ, vol. 80, no. 8, pp. 81-91, Aug. 1997.
[21] W. Philips, “Lossless DCT for combined lossy/lossless image coding,” Proc.
IEEE Int. Conf. Image Process., vol. 3, pp. 871-875, 1998.
[22] M.S. Moellenhoff and M.W. Maier, “DCT transform coding of stereo images
for multimedia applications,” IndEle, vol. 45, no. 1, pp. 38-43, Feb. 1998.
[23] L.R. Welch, “Walsh functions and Hadamard matrices. Application of Walsh
functions,” Proc. 1970, Symp. Washington, D.C., pp. 163-165, 1970.
[24] S.S. Agaian, Hadamard Matrices and Their Applications, Lecture Notes in
Mathematics 1168. Springer Verlag, Berlin, Heitelbery, New York, Tokyo, 1985.
[25] R. Gordon, “A tutorial on ART (algebraic reconstruction techniques),” IEEE
Trans. Nuclear Science, vol. 21, pp. 78-93, 1974.
[26] R. A. Brooks and G. Di Chiro, “Principles of computer assisted tomography
(CAT) in radiographic and radioisotopic imaging,” Phys. Med. Biol., vol. 21,
no. 5, pp. 689-732, 1976.
[27] G.T. Herman and A. Lent, “Iterative reconstruction algorithms,” Comput. Biol.
Med., vol. 6, pp. 273-294, 1976.
[28] S. Helgason, The Radon Transform. Progress in Mathematics 5, Birkhäuser
Boston, 1980.
[29] G.T. Herman, Image Reconstruction From Projections. The Fundamentals of
Computerized Tomography, Academic Press, New York, 1980.
[30] A. Rosenfeld and A.C. Kak, Digital Picture Processing: Vol. 1, Academic Press,
Orlando, 1982.
[31] Y. Censor, “Finite series-expansion reconstruction methods,” Proc. of IEEE,
vol. 71, no. 3, pp. 409-419, 1983.
[32] A.H. Andersen and A.C. Kak, “Simultaneous algebraic reconstruction tech-
nique (SART): A superior implementation of the ART algorithm,” Ultrasonic
Imaging, vol. 6, pp. 81-94, 1984.
[33] R.C. Gonzalez and Wintz, Digital Image Processing, 2nd edition, Prentice Hall,
Englewood Cliffs, New Jersey, 2001.
[34] G.N. Ramachandran and A.V. Lakshminaryanan, “Three-dimensional recon-
struction from radiographs and electron micrographs,” Proc. Nat. Acad. Sci.
USA, vol. 68, pp. 2236-2240, 1971.
[35] A.N. Kotelnikov, Theory of Potential Noise Stability, Moscow, Nauka, 1956.
[36] A.V. Fiacco and G.P. McCormick, Nonlinear Programming: Sequential Un-
constrained Minimization Techniques, John Wiley and Sons, Inc., New York,
1968.
[37] Lankaster P., Theory of Matrices. Academic Press, New York, London, 1969.
[38] H.J. Nussbaumer, Fast Fourier Transform and Convolution Algorithms, 2nd
ed., Springer Verlag, Berlin, 1982.
REFERENCES 425

[39] A.M. Grigoryan, “An algorithm of the two-dimensional Fourier trans-

form,” Izvestiya VUZov SSSR, Radioelectronica, vol. 27, no. 10, pp. 52-
57, USSR, Kiev 1984 (translated at http://fasttransforms.com/Art-USSR-
papers/Art− IzvuzovSSSR1984.pdf).
[40] A.M. Grigoryan, “An optimal algorithm for computing the two-
dimensional discrete Fourier transform,” Izvestiya VUZ SSSR, Radio-
electronica, vol. 29, no. 12, pp. 20-25, USSR, Kiev 1986 (translated at
http://fasttransforms.com/Art-USSR-papers/Art− IzvuzovSSSR1986.pdf).
[41] A.M. Grigoryan, “New algorithms for calculating discrete Fou-
rier transforms,” Journal Vichislitelnoi Matematiki i Matem-
aticheskoi Fiziki, vol. 26, no. 9, pp. 1407-1412, AS USSR,
Moscow, 1986 (translated at http://fasttransforms.com/Art-USSR-
papers/Art− JVMATofASUSSR1986.pdf).
[42] A.M. Grigoryan and M.M. Grigoryan, “Two-dimensional Fourier transform in
the tensor presentation and new orthogonal functions,” Avtometria, AS USSR
Siberian section, no. 1, pp. 21-27, Novosibirsk, 1986.
[43] A.M. Grigoryan, “An algorithm for computing the two-dimensional Fou-
rier transform of equal orders,” Proceedings of the IX All-Union Con-
ference on Coding and Transmission of Information, Part 2, pp. 49-52,
USSR, Odessa, 1988 (translated at http://fasttransforms.com/Art-USSR-
papers/Art− AllUnConf1988.pdf).
[44] A.M. Grigoryan, “An algorithm for computing the discrete Fou-
rier transform with arbitrary orders,” Journal Vichislitelnoi Matem-
atiki i Matematicheskoi Fiziki, AS USSR, vol. 30, no. 10, pp. 1576-
1581, Moscow, 1991 (translated at http://fasttransforms.com/Art-USSR-
papers/Art− JVMATofASUSSR1991.pdf).
[45] A.M. Grigoryan, and S.S. Agaian, “Split manageable efficient algorithm for
Fourier and Hadamard transforms,” IEEE Trans. on Signal Processing, vol. 48,
no. 1, pp. 172-183, Jan. 2000.
[46] A.M. Grigoryan, “2-D and 1-D multi-paired transforms: Frequency-time type
wavelets,” IEEE Trans. on Signal Processing, vol. 49, no. 2, pp. 344-353, Feb.
2001.
[47] A.M. Grigoryan, “Method of paired transforms for reconstruction of images
from projections: Discrete model,” IEEE Trans. on Image Processing, vol. 12,
no. 9, pp. 985-994, Sep. 2003.
[48] A.M. Grigoryan and S.S. Agaian, Multidimensional Discrete Unitary Trans-
forms: Representation, Partitioning and Algorithms,” Marcel Dekker Inc., New
York, 2003.
[49] A.M. Grigoryan and S.S. Agaian, “Transform-Based Image Enhancement Algo-
rithms with Performance Measure,” Advances in Imaging and Electron Physics,
Academic Press, vol. 130, pp. 165-242, May 2004.
[50] F.T. Arslan and A.M. Grigoryan, “Fast splitting α-rooting method of image en-
hancement: Tensor representation,” IEEE Trans. on Image Processing, vol. 15,
no. 11, pp. 3375-3384, Nov. 2006.
426 Image Processing

[51] A.M. Grigoryan and M.M. Grigoryan, Brief Notes in Advanced DSP: Fourier
Analysis with MATLAB, CRC Press, Taylor and Francis Group, Boca Raton,
Florida, 2009.
[52] A.M. Grigoryan and K. Naghdali, “On a method of paired representation:
Enhancement and decomposition by series direction images,” Journal of Ma-
thematical Imaging and Vision, vol. 34, no. 2, pp. 185-199, June 2009.
[53] A.M. Grigoryan and Nan Du, “2-D images in frequency-time representation:
Direction images and resolution map,” Journal of Electronic Imaging, vol. 19,
no. 3, (033012) pp. 1-14, July-September 2010.
[54] Nan Du and A.M. Grigoryan, “Paired transform method of image reconstruc-
tion from projections,” in Object Modeling, Algorithms and Applications, edited
by R.P. Barneva, V. Brimkov, et al., Research Publishing, June 18, 2010.
[55] A.M. Grigoryan, Multidimensional Discrete Unitary Transforms, book-chapter
19 (69 pages) in Transforms and Applications Handbook, 3rd ed., in the Electri-
cal Engineering Handbook series (Editor in Chief, Prof. Alexander Poularikas)
CRC Press, Taylor and Francis Group, Boca Raton, Florida, 2010.
[56] A.M. Grigoryan and Nan Du, “Principle of superposition by direction images,”
IEEE Trans. on Image Processing, vol. 20, no. 9, pp. 2531-2541, Sept. 2011.
[57] A.M. Grigoryan, “Comments on ‘The discrete periodic Radon transform,’”
IEEE Trans. on Signal Processing, vol. 58, no. 11, pp. 5962-5963, Nov 2010.
[58] T. Hsung, D.P.K. Lun, and W. Siu, “Reply to ‘Comments on The Discrete
Periodic Radon Transform,’” IEEE Trans. on Signal Processing, vol. 58, no. 11,
pp. 5963-5964, Nov 2010.
[59] A.M. Grigoryan and Nan Du, “Comments on ‘Generalised finite Radon trans-
form for N × N images,’” Image and Vision Computing, vol. 29, pp. 797-801,
November 2011.
Grigoryan 427

Appendix A

AN ALGORITHM OF THE TWO-DIMENSIONAL

FOURIER TRANSFORM1

GRIGORYAN A. M.

A new algorithm for calculating the two-dimensional Fourier transform

is proposed, by constructing the tensor of the third order. It is shown,
that such an algorithm requires operations of multiplication by 1.6-2
times fewer, than known algorithms and does not require an additional
temporary storage for saving intermediate results of calculations.

Consider an arbitrary array {fn,k } of the discrete signal, whose size for simplicity
of exposition will be considered equal, i.e., n, k = 1 : N, for an integer number N.
At each point (p, s) of the spectrum, the Fourier transform of the signal up to
the normalized factor equals
N X
X N
Fp,s = fn,k W np+ks , (A.1)
n=1 k=1

where W = WN = exp(2πi/N ).
For arbitrary p, s, t = 1 : N, define the sets

Vp,s,t = {(n, k); n, k = 1 : N, np + ks = t(modN )} (A.2)

and let X
fp,s,t = fn,k . (A.3)
Vp,s,t

Then, due to (A.2) and (A.3), at arbitrary point (p, s) of the spectrum we obtain
N
X
Fp,s = fp,s,t W t . (A.4)
t=1

Thus, by constructing the tensor of the 3rd order {fp,s,t ; p, s, t = 1 : N } by (A.3),

we represent the spectrum of the signal at each point (p, s) in the form of Fp,s =
(fp,s,1 , fp,s,2 , ..., fp,s,N ).
Such tensor representation is one-to-one and is characterized by the following
property. If for any integer l we denote ¯
l = l(modN ), then for any p, s and k
N
X
Fkp,ks = fp,s,t W tk . (A.5)
t=1

Indeed, due to definitions (A.2) and (A.3), as well the property of periodicity of
the discrete Fourier transform, we have
0 1
XN XN X XN X N
tk
fp,s,t W = @ fn,m A W tk = fn,m W (np+ms)k = Fkp,ks = Fkp,ks ,
t=1 t=1 Vp,s,t n=1 m=1

1 Vol. 27, no. 10, Izv. vuzov MV and SSO SSSR. Radioelectronica, 1984. UDK 517.443
428 Appendix A. Algorithm of the DFT

because, for arbitrary p, s, the sets Vp,s,t divide the domain of definition of the
[N
spectrum XN = {(n, m); n, m = 1 : N } in such a way, that XN = Vp,s,t and
t=1
Vp,s,t1 ∩ Vp,s,t2 = ∅ when t1 6= t2 .
Formula (A.5) means for fixed values of (p, s), from components fp,s,1 , . . . , fp,s,N
of the tensor, by means of the Fourier transform, one can obtain the spectrum at
N
points (kp, ks) which compose the cyclic group Tp,s with the generator (p, s), i.e.,
N
Tp,s = {(kp, ks); k = 1 : N }. (A.6)

For instance, for (p, s) = (1, 1), from components of the tensor f1,1,t , we obtain
the values of the spectrum of the signal at samples (1, 1), (2, 2), . . . , (N, N ).
As an example, let N be a power of two, then, by performing the fast Fourier
transform (FFT) over the sequence fp,s,1 , fp,s,2 , . . . , fp,s,N , we define due to (A.5)
N
the values of the spectrum in the group Tp,s , by performing about 0.5N log2 N (the
exact number 0.5N (log2 N − 2) + 1)2 operations of complex multiplications by non-
trivial rotation factors (powers of the exponent W ), i.e., an average for each sample
of the spectrum
N
vp,s = 0.5N log2 N/card Tp,s (A.7)
operations of multiplication; card means the cardinality (order) of the group.
N
If p, s and N do not have common dividers, then card Tp,s = N, therefore in
each point of form (kp, ks), for such p, s, the spectrum of the two-dimensional signal
is calculated by means of 0.5 log2 N operations of complex multiplication, i.e., the
same number as is necessary in the case of the one-dimensional Fourier transform.
It should be noted, that the existing algorithms of the two-dimensional FFT require
1.5 − 2 times more similar operations [1,2].
For N prime, the similar assertion is valid and the corresponding estimate vp,s
can be considered equal to (A.7), where N in the numerator of the formula is sub-
stituted by the nearest number equal to the power of two. Indeed, for such N,
the algorithm of FFT does not exist and, therefore, as done often in most practi-
cal problems, the artificial lengthening of the processed sequence is performed by
adding zeros to the minimum size which is equal to the power of two. Let us show
how to use formula (A.5) for effective, in the sense of high speed, calculation of the
complete spectrum of the two-dimensional signal.
It is obvious, that to do this it is necessary to select the values (p, s) of the samples
N
in an optimal way, such that the corresponding groups Tp,s cover the domain XN
with a minimum number of intersections. Indeed, for N equal to powers of 2, one
can show that there is no such covering of the spectral domain
[ N
XN = Tp,s , (A.8)
J

where J is a set in XN , that the conditions of independence are fulfilled, i.e.,

N
Tp,s ∩ TpN1 ,s1 = (N, N ) (A.9)

for arbitrary (p, s) and (p1 , s1 ) from the set J.

2 It should be 0.5N (log2 N − 3) + 2, as in the paired FFT. A.M. Grigoryan, Dec. 2010
Grigoryan 429

For prime N such covering of the domain of the spectrum exists; for instance,
as such a set one can take
Jp = {(1, p), p = 1 : N } ∪ (N, 1)
or (A.10)
Jp0 = {(p, 1), p = 1 : N } ∪ (1, N ).
Indeed, the following lemma is valid.
L e m m a . If the number N is prime, then for any l 6= N
N
X “ ” • N
X “ ” •
• N N • N N
XN = Tk,l \ (N, N ) +Tl,N = Tl,k \ (N, N ) +TN,l , (A.11)
k=1 k=1
•
where the signs • and + denote the union of disjoint sets.
P
P r o o f. For simplicity, we take l = 1, and let k1 and k2 , k1 6= k2 , be such that
TkN1 ,1 ∩TkN2 ,1 6= (N, N ), i.e. there exists (p0 , s0 ) = n1 (k1 , 1) = n2 (k2 , 1), which however
takes place only with fulfillment of the equalities n1 = n2 and n1 k1 = n2 k2 ( mod N ),
from which it follows that n1 (k1 −k2 ) = 0( mod N ). And, because by the assumption,
n1 6= N and N is prime, the last equality when k1 6= k2 could not be fulfilled.
Similarly, one can prove that for any k ∈ [1, N ], we have TkN1 ,1 ∩ T1,N N
= (N, N ).
N N
Then, because all sets Tk,1 \ (N, N ) and T1,N are contained in XN and together
have the cardinality equal N (N − 1) + N = N 2 = card XN , from here the validity
of the decomposition (A.11) follows. The lemma has been proved.
Thus, the existence of the optimal covering of the domain XN of the spectrum is
proved for N prime. By similar reasoning, it can be proved that for N equal powers
of two, the optimal covering of the domain XN of the spectrum does not exist, in
the sense of (A.8) and (A.9).
Therefore, as indicated above, for N, being a power of 2, in the covering (A.8)
it is necessary to pick up such generators (p, s), whose corresponding cyclic groups
intersect in minimum in the aggregate. One can prove that for such N it is enough
to take, in (A.8), the set equal
Jr = {(l, p), p = 1 − N } ∪ {(2k, l), k = 1 : N/2}
or (A.12)
Jr0 = {(p, l), p = 1 − N } ∪ {(l, 2k), k = 1 : N/2},

whose power card Jr = card Jr0 = 3N/2 ( l is prime).

Thus, due to (A.5), (A.8), (A.12), and (A.10), we find that the spectrum of the
original signal can be defined completely by means of 3N/2 one-dimensional Fourier
transforms, for N equal to a power of two, and for prime N, it is enough to perform
(N + 1) times such procedures.
The existing methods of calculation of the two-dimensional Fourier transform,
which are based on the separability of the transform, perform 2N such procedures [3].
Therefore, from the point of view of computation, the proposed tensor method of
calculation of the DFT gains in the number of the used one-dimensional FTs, when
compared with the existing two-dimensional FFT, and the gain equals, for N being
a power of two and prime number, respectively:
r p
γN = 2N/(3N/2) = 4/3 and γN = 2N/(N + 1), (A.13)
from which one can see that for N prime, the number of non-trivial complex mul-
tiplications required for calculating the complete spectrum in the proposed method
430 Appendix A. Algorithm of the DFT

is almost twice fewer than in the existent methods of the DFT, and 4/3 times fewer
for N equal the power of two.
At that, if for prime N the estimate of the relation in (A.13) is exact, for N
equals a power of two, the number of complex multiplications can be reduced in the
following way. Consider the groups covering the domain of the spectrum
N N N N N N
Tl,1 , Tl,2 , . . . , Tl,N , T2,l , T4,l , . . . , TN,l , (A.14)
r
when N = 2 , where r is an integer number, and l is any prime. It is not difficult to
show that for all p = 1 : N, the following is valid:
N N k k
Tl,p ∩ Tl,p+N/2k = {(n, m); ∃ k1 , n = 2 k1 l, m = 2 k1 p}, (A.15)

i.e., intersections of these groups is a subgroup which consists of all elements with
coordinates multiple to 2k . Indeed, for arbitrary integer k1 , we have (2k k1 l, 2k k1 p) =
(2k k1 l, 2k k1 (p + N/2k )) and on the contrary, if for some k1 and k2 we have
(k2 l, k2 p) = (k1 l, k1 (p + N/2k )), then the following conditions should be fulfilled:
(k2 − k1 )l = 0(modN ), (k2 − k1 )p = k1 N/2k (modN ), and because L is prime, N is
the power of two, it follows from the first equality that k1 = k2 , and from the second
one that k1 is the number multiple to 2k .
Thus, the equality in (A.15) has been proved. For any k and m = 1 : r, we denote
N,m
by Tl,k the set
“ ”
N,m N N N
Tl,k = Tl,k \ Tl,k ∩ Tl,k+N/2m . (A.16)
N,1
Consider first the sets Tl,k , i.e., the m = 1 case. Due to (A.12) and (A.15), we
obtain
N/2 “ ” N/4
[ N N,1
[ “ N N,1
”
XN = Tl,k ∪ Tl,k+N/2 T2m,l ∪ T2m+N/2,l , (A.17)
k=1 m=1
N,m
and it is obvious that card Tl,k N
= (1 − 1/2m )card Tl,k .
N,1 N
The sets Tl,k , for k = 1 : N, consist of all odd elements of groups Tl,k and have
N,1
cardinality card Tl,k = 0, 5N.
In [3], the algorithms of the FFT with decimation in frequency are described,
which after the first stage of iteration, when the N -point transform is divided by
two N/2-point Fourier transforms, all values of the spectrum with odd and even
coordinates are defined separately. Therefore, using such algorithms of the FFT for
calculating the two-dimensional spectrum by formula (A.5), one can define the values
N,1
of the spectrum at the sets Tl,k , by performing twice fewer operations of complex
N
multiplication than for the corresponding group Tl,k .
Thus, because of the selected covering (A.17), we find that the number of opera-
tions of complex multiplication sufficient for calculating the two-dimensional Fourier
transform by the tensor method is no greater than the number

v1 = N/2v0 + N/2 1/2v0 + N/4v0 + N/4 1/2v0 = 9/8N v0 , (A.18)

where v0 is the number of operations of multiplication by non-trivial rotated factors,

which are necessary for the one-dimensional FFT with decimation by frequency,
equals, as was already mentioned above,

v0 = N/2(log2 N − 2) + 1. (A.19)
Grigoryan 431

Therefore, substituting (A.19) in (A.18), we have the following: v1 <

9/16N 2 (log2 N − 2).
Continuing with similar reasoning, by means of (A.15) one can also exclude
N N
intersections in the groups Tl,k , where k = 1 : N/2, and Tm,l , where m = 1 : N/2 − 1.
For instance, in the next stage, when m = 2, we obtain the decomposition
N/4 “ ” N/2
[ N N,2
[ N,1
XN = Tl,k ∪ Tl,k+N/4 Tl,k+N/2 ∪,
k=1 k=1

N/4 N/8
[ N,1
[ “ N N,2
”
T2m+N/2,l T2m,l ∪ T2m+N/4,l , (A.20)
m=1 m=1
N,2
in addition, due to (A.15), card Tl,k = 3N/4.
As in the m = 1 case, if the spectrum of the signal was defined in the group
N,1 N,2
Tl,k , then for calculating the spectrum in the set Tl,k , it is enough on the second
stage of iteration in the algorithm of the FFT with decimation by frequency, after
when two N/2-point Fourier transforms are divided by four N/4-point transforms,
to use only the first three N/4-point Fourier transforms. Therefore, to define the
N,2
spectrum in Tl,k , it is enough to perform 4/3 times fewer operations of the complex
N,1
multiplication than for the corresponding Tl,k .
Thus due to (A.20), on the second stage of improvement, when m = 2, the
number of operations of multiplication for calculating the DFT by the tensor method
will not be greater than the number

v2 = N/4v0 +N/4·3/4v0 +N/2·0.5v0 +N/8v0 +N/8·3/4v0 +N/4·0.5v0 = 33/32N v0 .

Similarly, on the third stage, when m = 3, we obtain the corresponding number

v3 = 129/128N v0 and so on; if continue with m = 4, 5, ..., r, for large N, we find
that the number of necessary operations of multiplication by the rotation factors in
the algorithm is approximated from above by the number v = (N + 1)/N N v0 =
(N + 1)v0 .
Because the known methods of two-dimensional FFT, which use the property of
separability of the two-dimensional transform, perform 2N one-dimensional FFTs,
the number of necessary operations of multiplications for them equals v = 2N v0.
It follows that the proposed algorithm of calculation of the DFT, by the
number of operations of multiplication, is more than χN times effective, where
χN = 2N/(N + 1), than the mentioned algorithms; therefore, for large N, we have
χN ≈ 2.
If we compare the proposed algorithm of the DFT with the known algorithm
of the two-dimensional FFT, which uses the two-dimensional “butterfly” (which
is similar to the one-dimensional “butterfly”; see, for instance [1]), the number of
multiplications for which, as is not difficult to show, equals v 0 = 0.25 N 2 (3 log2 N −
4) + 1, then the corresponding coefficient of effectiveness by the number of the
specified operations will be:

v0 1 N 2 (3 log2 N − 4) + 1
χ0N = = .
v 4 (N + 1)(N/2(log2 N − 2) + 1)

Hence, we obtain χ0N > 1.6 for large N ; for instance, for values of N equal 128 and
1014, we have χ0128 > 1.68 and χ01024 > 1.62.
432 Appendix A. Algorithm of the DFT

Thus, the proposed tensor algorithm of calculation of the two-dimensional Fou-

rier transform is more than 1.6 − 2 times as effective as all known algorithms of the
FFT, by the number of necessary operations of complex multiplication. It should
also be noted that, if in the one-dimensional FFTs ω0 = v0 /N operations of multipli-
cation on average are used for each sample of the spectrum, in the tensor algorithm
of the DFT in average per sample ω1 = v/N 2 operations of multiplication are used,
i.e., ω1 = (N + 1)/N 2 v0 = (N + 1)/N ω0 ; hence, ω1 > ω0 and, for large values of N,
we have ω1 ≈ ω0 .
Thus, in the proposed algorithm of calculation of the DFT, the number of ope-
rations that are used for each sample of the spectrum is the same as in the algorithm
of the one-dimensional Fourier transform, for any large size of the signal, N × N,
where N is a power of 2.
Since for primes N in the tensor method of the DFT, it is enough to perform (N +
1) one-dimensional Fourier transforms, in this case the number of multiplications on
average per sample also equals

ω̃1 = (N + 1)ṽ0 /N 2 = (N + 1)/N ω̃0 (A.21)

where ṽ0 denotes the number of multiplications required for calculating the N -point
FT, and ω̃0 = ṽ0 /N. As already mentioned, one can consider ṽ0 to be equal to the
estimate in (A.19), for the corresponding nearest to the prime N size equal to a
power of two. According to the Lemma on covering of the spectrum domain, the
estimate in (A.21) cannot be improved, and therefore is the minimum.
In comparison with the existing algorithms of the two-dimensional FFT, the
proposed algorithm is also more effective by the volume of required temporary sto-
rage. Indeed, in the known algorithms of the two-dimensional FFT, it is necessary
to have additional temporary storage for constantly storing the intermediate results
of computing during the entire process of calculation. The volume of such storage
exceeds a few times the volume of the original signal, which essentially reduces the
possibility of using the algorithms of the two-dimensional FFT for the large size of
the signal. In the proposed tensor algorithm, such additional storage is not required,
because the values of the spectrum in each group of samples are calculated right
after defining the corresponding components of the tensor.
Thus, the existing methods of calculation of the two-dimensional Fourier trans-
form yield the tensor method by the main characteristics. In spite of the fact that
the sizes of the two-dimensional signal were assumed equal, one can easily extend
the tensor method of calculation of the spectra in the general case. Moreover, simi-
lar reasoning can be transferred to the case of the n-dimensional Fourier transform,
where n > 2, as well.

References

[1] Sereda L. A. An algorithm of fast calculation of the two-dimensional discrete

Fourier transform. Izv. vuzov SSSR, Radioelectronica, 1983, vol. 26, no. 7, pp. 18–22.
[2] Gold B., Raider C. Digital processing of signals / Tran. from English, M.: Sov.
radio, 1973, 367 p.
[3] Yaroslavsky L. P. Introduction to digital image processing. M.: Sov. radio, 1979,
312 p.
Grigoryan 433

Appendix B

NEW ALGORITHMS FOR COMPUTING DISCRETE

FOURIER TRANSFORMS1

GRIGORYAN A. M.
(Yerevan)

Effective methods of calculation of the multi-dimensional discrete Fourier

transform are proposed, which are based on its new representation.

In this work, a new general approach to consideration of the multi-dimensional

discrete Fourier transform (MDFT) is described, the main idea of which is a unique
possibility of presentation of each component of the transform by the corresponding
one-dimensional vector. This approach allows for performing separate calculations
of the MDFT in each disjoint with other groups of samples, on which the whole
domain of definition of the spectrum is divided, which allows for constructing the
effective algorithms of calculation of the MDFT by means of the minimum number
of one-dimensional DFTs.
The case of the two-dimensional Fourier transform is described in detail, and the
corresponding algorithms are compared with the most current algorithms of calcula-
tion of the two-dimensional DFT, which are based on the method of decimation by
Cooley-Tukey [1],[2], the polynomial transforms [3], and the operation of the two-
dimensional “butterfly.” As the particular case, the corresponding algorithm for the
one-dimensional Fourier transform is considered, too.

1. Vector representation of the spectrum of the MDFT

Consider an arbitrary array {fk1 ,...,kn } for the n-dimensional discrete signal, whose
size for simplicity of exposition will be considered equal, i.e., 1 ≤ ki ≤ N, i =
1, 2, ..., n, for an integer number N. Each spectral component at sample (p1 , ..., pn ),
1
where pi ∈ ZN = 1, 2, ..., N , i = 1, 2, ..., n, which up to the normalized factor equals
N
X N
X
Fp1 ,...,pn = ··· fk1 ,...,kn W k1 p1 +...+kn pn , (B.1)
k1 =1 kn =1

where W = WN = exp(2πi/N ), can be represented in the form of the N -dimensional

vector
F̄p1 ,...,pn = (fp1 ,...,pn ,1 , ..., fp1 ,...,pn ,N ), (B.2)
for which
N
X
Fp1,...,pn = fp1 ,...,pn ,t W t .
t=1

For that, as follows from (B.1), each component of the vector (B.2) must be

1 Journal of Computation Mathematics and Mathematical Physics, AS USSR. Volume 26,

No. 9, 1986. SCIENTIFIC REPORTS, UDK 517.97:537.812

434 Appendix B. New Algorithms of the DFT

calculated by summing values of the original signal at points of the corresponding

set:
( n
)
X
Vp1 ,...,pn ,t = (k1 , ..., kn ); 1 ≤ ki ≤ N, i = 1, 2, ..., n, ki pi = t mod N , (B.3)
i=1

i.e. X
fp1 ,...,pn ,t = fk1 ,...,kn . (B.4)
Vp1 ,...,pn,t

It is obvious that for a fixed value of point (p1 , ..., pn ), the sets (??) for different
n
t = 1, 2, ..., N together cover the whole domain of definition ZN of the original signal
without intersecting, therefore we have
N
X
Fkp1 ,...,kpn = fp1 ,...,pn ,t W kt (B.5)
t=1

for any integer k; here kpi = (kpi ) mod N. It follows from (B.5), that the Fourier
transform over the vector (B.2), which corresponds to the spectral component at
point (p1 , ..., pn ), defines the values of the n-dimensional spectrum at all points of
the group with generator (p1 , ..., pn ) :

Tp1 ,...,pn = {(kp1 , ..., kpn ); k = 1, 2, ..., N }. (B.6)

By selecting the minimum number of points of the spectrum, whose corres-

n
ponding groups (B.6) together would cover the whole domain ZN of definition of
the spectrum, one can construct an effective algorithm of calculation of the n-
dimensional MDFT. It should be noted that in the particular case, when n = 1,
1
formula (B.5) degenerates into the original formula (B.1) for odd points p (Tp = ZN ),
and therefore it is not very interesting, but for any n > 1 it is useful when calculating
the spectrum. Below, as an example, the n = 2 case is described.

2. Algorithm of computation of the two-dimensional DFT

For the spectrum of the two-dimensional discrete signal {fk1 ,k2 }, due to (B.5), at
any point (p1 , p2 ) with an integer k, we have
N
X
Fkp1 ,kp2 = fp1,p2 ,t W kt . (B.7)
t=1

Consider the cases when the size of the signal N is an arbitrary prime number or
power of two. It is not difficult to show that for N prime, the totality of groups (B.6),
which corresponds to the points of the set Jp = {(1, 1), (1, 2), ...., (1, N ), (N, 1)},
2
cover the whole domain ZN in optimal way, in the sense that all covering groups in
such totality intersect between themselves only at the point (N, N ). In the case when
N equals a power of two, such optimal covering does not exist. Indeed, one can be
easily convinced that in any totality of groups (B.6) that entirely cover the domain
of definition of the spectrum, there will, without fail, be intersections in the even
points as well. Moreover, each such point belongs to at least two different groups of
that totality, besides each covering totality contains no less than 3N/2 groups (B.6),
because for powers of two N one can fulfill the covering of the domain of definition
Grigoryan 435

of the spectrum, for instance, by the totality of groups corresponding to points of

the set Jr = {(1, 1), (1, 2), ...., (1, N ), (2, 1), (4, 1), (6, 1), ..., (N, 1)}.
Thus, the two-dimensional (N × N )-point DFT can be calculated completely by
formula (B.7), by means of N + 1 and 3N/2 one-dimensional N -point DFTs, for N
equal to arbitrary prime numbers and powers of two, respectively.
Such algorithms are more effective than the known algorithms from [1],[2], be-
cause the latter perform the Fourier transform first by all rows, and then by the
columns of the obtained signal, i.e., perform 2N indicated transforms. From here
it follows that the proposed algorithms perform 2 or 4/3 times fewer operations of
multiplication, if N is prime or a power of two, respectively. Further, the calcula-
tion at each group of points (B.6) is performed by formula (B.7) independently of
calculation at other groups; therefore, the large additional storage for saving inter-
mediate results is not required in the proposed algorithms. It also is obvious that
when computer-aided programming such algorithms, all accesses to memory, where
the original signal is stored, can be fulfilled only by rows and columns.
It should be noted that the analogous result for prime N has been obtained in
the reducible algorithm [3] of calculation of the two-dimensional DFT, by means of
more complicated polynomial transformations.
In the case when N is a power of 2, in such algorithm the calculation of the
(N × N )-point DFT is fulfilled by means of the polynomial transforms and not only
3N/2 reducible N -point DFTs, but also (N/2 × N/2)-point DFT, which in its turn
can be calculated in a similar way by 3N/4 reducible N/2-point DFTs and one
(N/4 × N/4)-point DFT, and so on.
The proposed algorithm of the two-dimensional DFT for N being powers of
two is also more effective, by the number of operations of multiplication, than the
known algorithms which use the operation of the two-dimensional butterfly (see, for
instance, [4]), where v̄N,N = (N 2 /4)(3 log2 N − 4) + 1 such operations are required.
Indeed, because the N -point DFT, when fulfilled by the algorithm of the FFT,
requires (N/2)(3 log2 N − 3) + 2 non-trivial complex operations of multiplication,
then the proposed algorithm of calculation of the two-dimensional DFT requires
vN,N = 3/4N 2 (3 log2 N − 3) + 3N such operations. Hence, for N ≥ 8, we have
v̄N,N − vN,N = 5/4N 2 − 3N > 0.
Similar conclusions are valid for the algorithms of calculation of the n-
dimensional DFT, which are based on (B.5) for the n > 2 case, too. As already
mentioned, for N equal to a power of two, groups (B.6), which together cover the
n
domain of definition of the spectrum ZN , define in their intersections all even points
of the spectrum, the number of which equals (N/2)n . In this connection, the es-
sential reduction of operations of multiplication at such points can be achieved by
removing the repeated calculations of the MDFT by formula (B.5).
For instance, in the two-dimensional case, such improvement of the described
algorithm is reduced to the fact that the (N × N )-point DFT is calculated by means
of 3N/2 incomplete N -point DFT, which decreases the number of operations of
multiplication almost by 1.5 times.
Below, a more effective algorithm of calculation of the n-dimensional DFT is
proposed for the case when N is a power of two.
436 Appendix B. New Algorithms of the DFT

3. Paired representation of the spectrum of the MDFT

Consider the vector representation of the arbitrary component of the spectrum,
which was defined in §1. Because for all t ∈ {1, 2, ..., N/2}, the following takes place
W t = −W t+N/2 , one can represent uniquely the component (B.1) at each point
(p1 , ..., pn ) by the N/2-dimensional vector
F̄p0 1 ,...,pn = (fp0 1,...,pn ,1 , ..., fp0 1 ,...,pn ,N/2 ), (B.8)
where each component is calculated from (B.4):
fp0 1 ,...,pn ,t = fp1 ,...,pn ,t − fp1 ,...,pn ,t+N/2 , (B.9)
where t = 1, 2, ..., N/2, and for which
N/2
X
Fp1,...,pn = fp0 1 ,...,pn ,t W t .
t=1

Moreover, as follows from (B.5), for an arbitrary odd value of k the following
takes place
N/2
X 0
Fkp1 ,...,kpn = fp1 ,...,pn ,t W kt .
t=1
By opening here the exponential factors for odd values of k = 2m − 1, where m is
any integer, we obtain
N/2
X` ´ mt
Fkp1,...,kpn = −t
fp0 1 ,...,pn ,t WN WN/2 . (B.10)
t=1

Thus, by fulfilling the N/2-point Fourier transform over the vector

−N/2
−1
G(F )p1 ,...,pn = (fp0 1 ,...,pn ,1 WN , ..., fp0 1 ,...,pn ,N/2 WN ), (B.11)
we obtain all values of the spectrum of the signal at points composing the odd half
of the corresponding group (B.6):
Tp0 1 ,...,pn = {((2m − 1)p1 , ..., (2m − 1)pn ); m = 1, 2, ..., N/2}. (B.12)
For an example, for the point (1, ..., 1) by the Fourier transform over the com-
ponents of the vector G(F )1,...,1 the spectral components at the points (1, ..., 1),
(3, ..., 3), ..., (N − 1, ..., N − 1) are defined.
By using formula (B.10), we construct the algorithm for computing the complete
n-dimensional DFT. As in the algorithms based on the method of computation of
the MDFT by (B.5), it is obvious one need select a minimum number of points
n
whose corresponding sets (B.12) together cover the entire domain ZN of definition
of the spectrum.
Unlike the method described in §1 for calculation of the MDFT, this method
of calculation, which is based on the representation of each spectral component in
form of (B.8), which we call the paired representation, makes it possible to cover
the domain of definition of the spectrum, for N equal powers of two, optimally, i.e.,
such that the following conditions will be fulfilled
n
[ 0
ZN = Tp1 ,...,pn , Tp0 1 ,...,pn ∩ Tg01 ,...,gn = ∅, (B.13)
J
Grigoryan 437
n
where the union is performed by points of some subset J in ZN , and the condition
of non-intersection is fulfilled for any different points of this set.
We show this with examples with the two-dimensional and one-dimensional
DFTs, whose corresponding algorithms are described below. More general cases of
the MDFT are considered in a similar way.

4. Paired algorithm of the two-dimensional DFT

Considering the two-dimensional case, we take the set J in (B.13) equal to

J 2 = JN ∪ 2JN/2 ∪ 4JN/4 ∪ . . . ∪ N J1 , (B.14)

where kJN/k = {(kp1 , kp2 ); (p1 , p2 ) ∈ JN/k }, for all k = 2r , r = 0, 1, ..., log2 N,
and JN/k are the above defined sets {(1, p2 ), p2 = 1, 2, ..., N/k} ∪ {(2p1 , 1), p1 =
1, 2, ..., N/(2k)} all powers of two k from 1 to N/2; for k = N we have J1 = {1, 1}. It
is easy to show that in set (B.14) when k1 6= k2 will be JN/k1 ∩ JN/k2 6= ∅, and thus
J 2 satisfies the condition of optimality (B.13) of covering the domain of definition
of the spectrum. Further, the cardinalities of sets (B.12) in (B.13) for such J equal
card Tp0 1 ,p2 = 0.5N/2r , if (p1 , p2 ) ∈ 2r JN/2r , where r = 0, 1, ..., log2 N, and for the
sets in (B.14) we have card JN/2r = 1.5N/2r . Therefore, it follows easily from (B.14),
card J 2 = 3N − 2.
Thus, in the optimal covering (B.13), which unites 3N − 2 sets (B.12), 3N/2
of them contain N/2 elements each, 3N/4 sets contain N/2 elements each, and so
on. Therefore, to define the complete spectrum of the two-dimensional signal of
size N × N, it is enough to fulfill 3N/2 one-dimensional N/2-point DFTs, 3N/4
one-dimensional N/4-point DFTs, and so on.
Indeed, let the arbitrary point (p1 , p2 ) from J 2 belong to the set 2r JN/2r
for some r ∈ [0, log2 N ], i.e., g.c.d.(p1 , p2 ) = 2r . Due to definitions (B.3),(B.4)
and (B.9), it is obvious, that in the corresponding vector G(F )p1 ,p2 , all com-
ponents with numbers not multiple to the number 2r equal zero. There-
fore, one can associate uniquely to such vector the N/2r+1 -dimensional vector
r
G(F )p1 ,p2 = (gp1 ,p2 ,1 , gp1 ,p2 ,2 , ..., gp1 ,p2 ,N/2r+1 ), wherein the components gp1 ,p2 ,t =
0 −t r+1
fp1 ,p2 ,2r t WN/2r for all t = 1 − N/2 . At that, for any values of m, the following
equality is valid

N/2 N/2r+1
X` −t ´ mt
X mt
fp0 1,p2 ,t WN WN/2 = gp1 ,p2 ,t WN/2r+1 . (B.15)
t=1 t=1

Therefore the N/2-point DFT of the vector F̄p0 1 ,p2 , for arbitrary point (p1 , p2 ) ∈
2r JN/2r , where r ∈ {0, 1, ..., log2 N }, is equivalent to fulfillment of the N/2r+1 -point
r
DFT over the corresponding vector G(F )p1 ,p2 . It follows from here, that the results
of 3N − 2 one-dimensional DFTs of order N/2, which are sufficient for calculating
the complete two-dimensional DFT by the proposed algorithm with formula (B.10),
are equivalent in union to the results of fulfillment of the mentioned DFTs: 3N/2 of
order N/2 each, 3N/4 of order N/4 each, and so on.
0
Calculate the volume vN,N of non-trivial operations of multiplication by the
rotated exponential factors, which are necessary for fulfillment of the proposed algo-
rithm of calculation of the (N × N )-point DFT. If we denote by vM (with arbitrary
438 Appendix B. New Algorithms of the DFT

power of two M ) the volume of such operations for the M -point DFT, then it is
obvious, that from (B.10) and (B.15) we obtain
0
vN,N = (3N/2)(vN/2 + N/2 − 2) + (3N/4)(vN/4 + N/4 − 2) + ...
log 2 N −2
X (B.16)
... + 3 · 2(v2 + 2 − 2) = 3N 2−r vN/2r + N 2 − 6N + 8.
r=1

This formula of estimation of the number of non-trivial operations of multipli-

cation, required for calculation of the two-dimensional DFT by means of the one-
dimensional transforms, due to optimality of the covering (B.13) is exact, i.e., can
not be improved. Therefore if one has minimum estimates of such operations for
the one-dimensional DFT, one can define by means of (B.16) a minimum number
of complex operations for the two-dimensional DFT, too. If we estimate all vN/2r
by means of the one-dimensional algorithms from [1],[2], for which these estimates
equal, as already mentioned, (N/2r+1 )[log2 (N/2r ) − 3] + 2, then it is not difficult to
obtain 0 1
2 log 2 N −2
0 N @log2 N −
X r
vN,N = − 1A − 8(log2 N − 1),
2 r=1
4r
from which it follows, that for N > 8
0
vN,N ≈ (N 2 /2)(log2 N − 5/2) − 8(log2 N − 1). (B.17)
When compared with the algorithm described in §1 with formula of calculation
(B.5), which performs 3N/2 one-dimensional N -point DFTs by means of vN,N =
0
(3N/2)vN operations of multiplication, we find that vN,N > 1.5vN,N . Thus, by the
number of operations of multiplication the described here algorithm, which we call
paired, is effective more than 1.5 times the algorithm in §1.

5. Algorithm of the one-dimensional DFT

Consider the n = 1 case, which corresponds to the one-dimensional Fourier trans-
form. The set J in (B.13), by which one must construct the algorithm based on
the paired representation of the spectrum, it obviously can be taken as equal to
J 1 = {1, 2, 4, 8, ..., N } with cardinality log2 N + 1. This set represents the pro-
jection of the set J 2 defined above on the one-dimensional domain ZN 1
. At that,
0 0 r
T1 = {1, 3, ..., N −1}, ..., Tm = {m, 3m, ..., N −m}, for m = 2 , r = 1, 2, ..., log2 N −1,
0
and TN = {N }.
Similar to the n = 2 case, we find that for calculating the N -point DFT it is
enough to fulfill one N/2-point DFT over the vector G(F )1 , one N/4-point DFT
2r+1
over the vector G(F )21 , and so on, one N/2r -point DFT over the vector G(F )2r ,
where r = 1, 2, ..., log2 N − 1; here,
2r+1 −N/2r+1
“ ”
G(F )2r = f20 r ,1 WN/2 −1 0 −2 0
r , f2r ,2 WN/2r , . . . , f2r ,N/2r WN/2r .

At that their components are calculated easily; for example, for r = 0, 1, we have
0 0
f1,t = ft − ft+N/2 , t = 1, 2, ..., N/2, f2,t = ft + ft+N/2 − ft+3N/4 − ft+N/4 , t =
1, 2, ..., N/4. Consequently, the number of non-trivial operations of multiplication in
this algorithm is calculated by the following recurrent formula for N > 8 :
0 0 0
vN = (vN/2 + N/2 − 2) + (vN/4 + N/4 − 2) + · · · + (v80 + 6) + v80 ;
Grigoryan 439

from here the estimate for multiplications follows:

0 0
vN = 2vN/2 + N/2 − 2 = (N/2)(log2 N − 3) + 2.

Thus, the same estimate of operations of multiplication has been obtained as in

the algorithms from [1],[2], which was used above when deriving estimate (B.17) for
the two-dimensional DFT.
It should be noted that the described method of calculation of the one-
dimensional DFT, which is based on the paired representation of the latter, allows
for constructing algorithms for calculation of the DFT with a minimum number of
the same transforms but smaller orders for any orders of the transform, which are
even numbers. Indeed, the existence of the necessary optimal covering of the domain
of definition of the spectrum in form (B.12) for any even order N can be proved by
the simple construction of sets J in (B.13). And such set J for any even N can be
taken as equal to J = {1, 2, 4, 8, ..., 2r }, where 2r is the greatest power of two, which
is the division of N. It is clear that in the general n > 2 case, when calculating the
n-dimensional DFT by the method based on the paired representation, the algo-
rithms using the minimal number of the one-dimensional DFTs can be constructed
similar to the algorithms described above for the n = 1, 2 cases, and for any orders
whose at least one dimension is the even number.
It should be noted in conclusion that with corresponding organization of the
procedure of fulfillment of the paired representation of the necessary spectral com-
ponents of the signal by the specialized processors, one can realize the process of
calculation of the MDFT in a time period that is close to the time of fulfillment of
the N/2-point DFT. Indeed, the process of calculation of the complete MDFT by
the mentioned algorithm is divided into separate procedures of calculation of the
spectrum for each disjoint group of samples, which all together define the whole do-
main of definition of the spectrum, and the maximum order of the one-dimensional
DFT by means of which the values of the spectrum of the original signal are defined
at each such group equals N/2.

References

[1] Rabiner L., Gold B. Theory and application of the digital signal processing. M.:
Mir, 1978.
[2] Yaroslavsky L. P. Introduction to digital image processing. M.: Sov. radio, 1979.
[3] Nussbaumer, H. J. Calculation of two-dimensional convolutions and discrete
Fourier transforms, in Fast algorithms in digital image processing. M.: Radio and
communication, 1984, pp. 43-88.
[4] Sereda L. A. An algorithm of fast calculation of the two-dimensional discrete
Fourier transform. Izv. vuzov SSSR, Radioelectronica, 1983, vol. 26, no. 7, pp. 18–22.
This page intentionally left blank
 176 200

175 180 FT(0)=4388.5

174 160

140
173
120
172
100
171
80
170
60
169 40
168 20

167 0
0 50 100 150 200 250 0 50 100 150 200 250

(a ) o rigina l (b ) 1 -D signal (c ) 1 -D DF T

1 176
original signal
175
0.95
174
0.9
173

0.85 172

0.8 171
new signal
170
0.75
169
0.7
168

0.65 167
0 50 100 150 200 250 0 50 100 150 200 250

(d ) f actors (e ) n ew signal

250

200

150

100

50

0
0 50 100 150 200 250

(f ) f requency-poi nt s (g ) e nhanced

FIGURE 1.4: Fast transform-based method of image enhancement. (a) The

original image, (b) the image-signal, (c) the magnitude spectrum of the signal,
(d) factors, (e) the processed signal, (f) marked locations of 256 frequency-
points, at which the 2-D DFT was changed, and (g) the image enhanced by
one signal.
 (0,0)

T1,2

T1,1

FIGURE 1.6: Torus of the lattice 20 × 20 with two spirals corresponding to

the groups T1,1 and T1,2 .

10−point DFT composition: component F(1).

8

6
9
6
imaginary part y(n) of the signal z(n)=x(n)+jy(n)

4
2

2
F(0)
7

0
1
F(1)
5 0
−2

3
−4

4
−6
8

−8
−8 −6 −4 −2 0 2 4 6 8

FIGURE 1.15: Calculation of the second component of the DFT, F1 .

 10−point DFT composition: F(p), p=0:9.
8
4

6
7

imaginary part y(n) of the signal z(n)=x(n)+jy(n) 2

4
9

5
2
1 F(0)

F(8) F(9) F(4)

0 F(3)
F(2) F(5) 8
F(6)
F(1)
0
F(7)
−2
3

−4
6

−6

−8
−8 −6 −4 −2 0 2 4 6 8

FIGURE 1.16: Result of ten rotations of the signal-data.

S1,1,7

S1,1,3

−5
10

10
5
5
0
0
−5
−5

−10 −10

FIGURE 2.4: The net with knots of the grid 32 × 32 in the 3-D space with
locus of two spirals S1,1,3 and S1,1,7 .
 20 Re g’
T
3,1
10

−10
f’ T
−20 3,1

0 50 100

(a) (b)
700 200
600

500 100

400
0
300

200 200

100 100

0 200 250
0 50 100 0 50 100 150
0
(c) (d)

FIGURE 2.18: (a) Tree image 256 × 256, (b) the paired splitting-signal fT3,1
′

and the real part of the modified signal, (c) the 1-D DFT of the modified
signal, and (d) the arrangement of values of the 1-D DFT at frequency-points
′
of the subset T3,1 . (The 1-D DFT is shown in the absolute scale.)

l(0) l(4) l(7) l(8) l(9) l(12)

0
l(16)
0.25

0.25

0.375

3
y and m

0.5

4
l(20)
0.625

0.75

0.875

1
0 1 2 3 4 5 6 7

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1

x and n

FIGURE 3.22: Mask of the paired function χ′2,1,0 (n,m) and rays.
 image f(x,y) with 7 rectangles discrete image f(n,m)
1

10
0.8

20
0.6 1
3 30
72

0.4 40
4
5
50
0.2
6
60
0
0 0.2 0.4 0.6 0.8 1 10 20 30 40 50 60
intensities: 5 11 12 8 3 3 9 grid 64x64

FIGURE 4.2: Seven color rectangles and the discrete gray-scale image com-
posed by these rectangles.

10 rectangles and grid 128x128 reconstructed image 128x128

1

20
0.8
5 4
40
0.6 9 2
7 60
8
10
0.4 6 1 80

0.2 100
3

120
0
0 0.2 0.4 0.6 0.8 1 20 40 60 80 100 120
intensities: 8 10 16 9 16 9 15 4 11 8 cpu time/rect=0.40 min

FIGURE 4.6: Ten random rectangles (left) and the reconstructed image
(right).
 10 rectangles and grid 256x256 reconstructed image 256x256
1

0.8 1 8 50

9
0.6 100
5
3
0.4 10 150
4

6 7
0.2 200
2

0 250
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 4 7 11 6 14 9 15 5 11 1 cpu time/rect=4.56 min

FIGURE 4.7: Ten random rectangles (left) and the reconstructed image
(right).

random rectangles and grid 257x257 reconstructed image 257x257

1

0.8 50
2
8 100
0.6
7

0.4 1 4 150
6
5 3

0.2 9 200
10

250
0
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 1 11 3 5 5 8 4 10 3 13 cpu time/rect=1.86 min

FIGURE 5.37: Ten colored rectangles and the discrete gray-scale image of
size 257 × 257, which is composed by these rectangles.
 random rectangles and grid 257x257 reconstructed image 257x257
1

9
0.8 6 50

5 1
0.6 100
28
10
4
0.4 150
7
3

0.2 200

250
0
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 10 4 15 12 5 11 16 7 9 17 cpu time/rect=1.80 min

FIGURE 5.41: Image with ten rectangles and its reconstruction 257 × 257.

G−particle (1,1) G−particle (2,1)

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 0 0.5 1

FIGURE 6.13: The field functions of two G-particles (1, 1) and (2, 1).

UH−square of number 26 central square

1 0.06

0.04
0.5
0.02

0 0

−0.02
−0.5
−0.04

−1 −0.06
−1 −0.5 0 0.5 1 −0.05 0 0.05

FIGURE 6.14: Map of G-rays in the UH-square of number 26 and its central
component.
 10 rectangles on 128x128 128x128: Covering by 192 projections
1

20
0.8 3
2 7 40
0.6
10 60

5 6
0.4 80
1 9
4
8
0.2 100

120
0
0 0.2 0.4 0.6 0.8 1 20 40 60 80 100 120
(a) (b)

FIGURE 6.18: (a) Ten rectangles and the lattice 128 × 128 and (b) the
reconstructed image.

7 rectangles and grid 256x256 reconstructed image 256x256

1

0.8 50
6

0.6 100
5
2
0.4 7 150

3 1
4 200
0.2

0 250
0 0.2 0.4 0.6 0.8 1 50 100 150 200 250
intensities: 4 2 14 11 9 13 5 cpu time/rect=4.90 min

FIGURE 7.1: Image with seven random rectangles and the reconstructed
discrete image 256 × 256.
 Electrical Engineering

Image Processing
Tensor Transform and Discrete
Tomography with MATLAB®
Image
Focusing on mathematical methods in computer tomography, Image
Processing: Tensor Transform and Discrete Tomography with
MATLAB® introduces novel approaches to help in solving the problem
of image reconstruction on the Cartesian lattice. Specifically, it
Processing
discusses methods of image processing along parallel rays to more Tensor Transform and Discrete
quickly and accurately reconstruct images from a finite number of
projections, thereby avoiding overradiation of the body during a
computed tomography (CT) scan. Tomography with MATLAB®
The book presents several new ideas, concepts, and methods, many
of which have not been published elsewhere. New concepts include
methods of transferring the geometry of rays from the plane to the
Cartesian lattice, the point map of projections, the particle and its X-ray set
field function, and the statistical model of averaging. The authors
supply numerous examples, MATLAB®-based programs, end-of-chapter
problems, and experimental results of implementation.
image f(n,m)
The main approach for image reconstruction proposed by the authors
differs from existing methods of back-projection, iterative reconstruction,
and Fourier and Radon filtering. In this book, the authors explain how
to process each projection by a system of linear equations, or linear
convolutions, to calculate the corresponding part of the 2-D tensor
or paired transform of the discrete image. They then describe how projection
to calculate the inverse transform to obtain the reconstruction. The ø=26.565º
proposed models for image reconstruction from projections are simple
and result in more accurate reconstructions.

Introducing a new theory and methods of image reconstruction, this

book provides a solid grounding for those interested in further research
and in obtaining new results. It encourages readers to develop effective detectors
applications of these methods in CT.

K14749
ISBN: 978-1-4665-0994-8
90000

9 781466 509948
Artyom M. Grigoryan n Merughan M. Grigoryan

K14749_Cover_mech.indd 1 8/7/12 10:48 AM