Image Processing Tensor Transform And Discrete

Tomography With Matlab Artyom M Grigoryan

download

https://ebookbell.com/product/image-processing-tensor-transform-
and-discrete-tomography-with-matlab-artyom-m-grigoryan-4421784

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

From Algebraic Structures To Tensors Digital Signal And Image

Processing Matrices And Tensors In Signal Processing Set 1st Edition
Grard Favier Editor

https://ebookbell.com/product/from-algebraic-structures-to-tensors-
digital-signal-and-image-processing-matrices-and-tensors-in-signal-
processing-set-1st-edition-grard-favier-editor-11363588

Tensors In Image Processing And Computer Vision Santiago Ajafernandez

Rodrigo De Luis Garca Dacheng Tao Xuelong Li

https://ebookbell.com/product/tensors-in-image-processing-and-
computer-vision-santiago-ajafernandez-rodrigo-de-luis-garca-dacheng-
tao-xuelong-li-36969238

Image Processing And Intelligent Computing Systems 1st Prateek Singhal

https://ebookbell.com/product/image-processing-and-intelligent-
computing-systems-1st-prateek-singhal-47425180

Image Processing Masterclass With Python First Sandipan Dey

https://ebookbell.com/product/image-processing-masterclass-with-
python-first-sandipan-dey-47616880
Image Processing With Imagej Second Edition 2nd Edition Jurjen Broeke

https://ebookbell.com/product/image-processing-with-imagej-second-
edition-2nd-edition-jurjen-broeke-50195496

Image Processing In Biological Science Proceedings Of A Conference

Held November 1966 Reprint 2020 Wilfrid J Dixon Editor Mary A B
Brazier Editor Bruce D Waxman Editor Diane M Ramsey Editor

https://ebookbell.com/product/image-processing-in-biological-science-
proceedings-of-a-conference-held-november-1966-reprint-2020-wilfrid-j-
dixon-editor-mary-a-b-brazier-editor-bruce-d-waxman-editor-diane-m-
ramsey-editor-51817642

Image Processing And Communications Challenges 2 Ryszard S Choras

https://ebookbell.com/product/image-processing-and-communications-
challenges-2-ryszard-s-choras-52873854

Image Processing And Machine Learning Volume 2 Advanced Topics In

Image Analysis And Machine Learning 1st Edition Erik Cuevas

https://ebookbell.com/product/image-processing-and-machine-learning-
volume-2-advanced-topics-in-image-analysis-and-machine-learning-1st-
edition-erik-cuevas-54730812

Image Processing And Machine Learning Volume 1 Foundations Of Image

Processing 1st Edition Cuevas

https://ebookbell.com/product/image-processing-and-machine-learning-
volume-1-foundations-of-image-processing-1st-edition-cuevas-54734782
 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
2 1 1 2 −2 + j 1 −3j 1 + 6j
{fn,m } = → {Fp,s } = .
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
f1,0 f1,1 f1,2 f1,3 2 0 1 1
f = = .
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).

 Another Random Document on
Scribd Without Any Related Topics
 PLEASE READ THIS BEFORE YOU DISTRIBUTE OR USE THIS WORK

To protect the Project Gutenberg™ mission of promoting the

free distribution of electronic works, by using or distributing this
work (or any other work associated in any way with the phrase
“Project Gutenberg”), you agree to comply with all the terms of
the Full Project Gutenberg™ License available with this file or
online at www.gutenberg.org/license.

Section 1. General Terms of Use and

Redistributing Project Gutenberg™
electronic works
1.A. By reading or using any part of this Project Gutenberg™
electronic work, you indicate that you have read, understand,
agree to and accept all the terms of this license and intellectual
property (trademark/copyright) agreement. If you do not agree
to abide by all the terms of this agreement, you must cease
using and return or destroy all copies of Project Gutenberg™
electronic works in your possession. If you paid a fee for
obtaining a copy of or access to a Project Gutenberg™
electronic work and you do not agree to be bound by the terms
of this agreement, you may obtain a refund from the person or
entity to whom you paid the fee as set forth in paragraph 1.E.8.

1.B. “Project Gutenberg” is a registered trademark. It may only

be used on or associated in any way with an electronic work by
people who agree to be bound by the terms of this agreement.
There are a few things that you can do with most Project
Gutenberg™ electronic works even without complying with the
full terms of this agreement. See paragraph 1.C below. There
are a lot of things you can do with Project Gutenberg™
electronic works if you follow the terms of this agreement and
help preserve free future access to Project Gutenberg™
electronic works. See paragraph 1.E below.
1.C. The Project Gutenberg Literary Archive Foundation (“the
Foundation” or PGLAF), owns a compilation copyright in the
collection of Project Gutenberg™ electronic works. Nearly all the
individual works in the collection are in the public domain in the
United States. If an individual work is unprotected by copyright
law in the United States and you are located in the United
States, we do not claim a right to prevent you from copying,
distributing, performing, displaying or creating derivative works
based on the work as long as all references to Project
Gutenberg are removed. Of course, we hope that you will
support the Project Gutenberg™ mission of promoting free
access to electronic works by freely sharing Project Gutenberg™
works in compliance with the terms of this agreement for
keeping the Project Gutenberg™ name associated with the
work. You can easily comply with the terms of this agreement
by keeping this work in the same format with its attached full
Project Gutenberg™ License when you share it without charge
with others.

1.D. The copyright laws of the place where you are located also
govern what you can do with this work. Copyright laws in most
countries are in a constant state of change. If you are outside
the United States, check the laws of your country in addition to
the terms of this agreement before downloading, copying,
displaying, performing, distributing or creating derivative works
based on this work or any other Project Gutenberg™ work. The
Foundation makes no representations concerning the copyright
status of any work in any country other than the United States.

1.E. Unless you have removed all references to Project

Gutenberg:

1.E.1. The following sentence, with active links to, or other

immediate access to, the full Project Gutenberg™ License must
appear prominently whenever any copy of a Project
Gutenberg™ work (any work on which the phrase “Project
Gutenberg” appears, or with which the phrase “Project
Gutenberg” is associated) is accessed, displayed, performed,
viewed, copied or distributed:

This eBook is for the use of anyone anywhere in

the United States and most other parts of the
world at no cost and with almost no restrictions
whatsoever. You may copy it, give it away or re-
use it under the terms of the Project Gutenberg
License included with this eBook or online at
www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of
the country where you are located before using
this eBook.

1.E.2. If an individual Project Gutenberg™ electronic work is

derived from texts not protected by U.S. copyright law (does not
contain a notice indicating that it is posted with permission of
the copyright holder), the work can be copied and distributed to
anyone in the United States without paying any fees or charges.
If you are redistributing or providing access to a work with the
phrase “Project Gutenberg” associated with or appearing on the
work, you must comply either with the requirements of
paragraphs 1.E.1 through 1.E.7 or obtain permission for the use
of the work and the Project Gutenberg™ trademark as set forth
in paragraphs 1.E.8 or 1.E.9.

1.E.3. If an individual Project Gutenberg™ electronic work is

posted with the permission of the copyright holder, your use and
distribution must comply with both paragraphs 1.E.1 through
1.E.7 and any additional terms imposed by the copyright holder.
Additional terms will be linked to the Project Gutenberg™
License for all works posted with the permission of the copyright
holder found at the beginning of this work.
 1.E.4. Do not unlink or detach or remove the full Project
Gutenberg™ License terms from this work, or any files
containing a part of this work or any other work associated with
Project Gutenberg™.

1.E.5. Do not copy, display, perform, distribute or redistribute

this electronic work, or any part of this electronic work, without
prominently displaying the sentence set forth in paragraph 1.E.1
with active links or immediate access to the full terms of the
Project Gutenberg™ License.

1.E.6. You may convert to and distribute this work in any binary,
compressed, marked up, nonproprietary or proprietary form,
including any word processing or hypertext form. However, if
you provide access to or distribute copies of a Project
Gutenberg™ work in a format other than “Plain Vanilla ASCII” or
other format used in the official version posted on the official
Project Gutenberg™ website (www.gutenberg.org), you must,
at no additional cost, fee or expense to the user, provide a copy,
a means of exporting a copy, or a means of obtaining a copy
upon request, of the work in its original “Plain Vanilla ASCII” or
other form. Any alternate format must include the full Project
Gutenberg™ License as specified in paragraph 1.E.1.

1.E.7. Do not charge a fee for access to, viewing, displaying,

performing, copying or distributing any Project Gutenberg™
works unless you comply with paragraph 1.E.8 or 1.E.9.

1.E.8. You may charge a reasonable fee for copies of or

providing access to or distributing Project Gutenberg™
electronic works provided that:

• You pay a royalty fee of 20% of the gross profits you derive
from the use of Project Gutenberg™ works calculated using the
method you already use to calculate your applicable taxes. The
fee is owed to the owner of the Project Gutenberg™ trademark,
 but he has agreed to donate royalties under this paragraph to
the Project Gutenberg Literary Archive Foundation. Royalty
payments must be paid within 60 days following each date on
which you prepare (or are legally required to prepare) your
periodic tax returns. Royalty payments should be clearly marked
as such and sent to the Project Gutenberg Literary Archive
Foundation at the address specified in Section 4, “Information
about donations to the Project Gutenberg Literary Archive
Foundation.”

• You provide a full refund of any money paid by a user who

notifies you in writing (or by e-mail) within 30 days of receipt
that s/he does not agree to the terms of the full Project
Gutenberg™ License. You must require such a user to return or
destroy all copies of the works possessed in a physical medium
and discontinue all use of and all access to other copies of
Project Gutenberg™ works.

• You provide, in accordance with paragraph 1.F.3, a full refund of

any money paid for a work or a replacement copy, if a defect in
the electronic work is discovered and reported to you within 90
days of receipt of the work.

• You comply with all other terms of this agreement for free
distribution of Project Gutenberg™ works.

1.E.9. If you wish to charge a fee or distribute a Project

Gutenberg™ electronic work or group of works on different
terms than are set forth in this agreement, you must obtain
permission in writing from the Project Gutenberg Literary
Archive Foundation, the manager of the Project Gutenberg™
trademark. Contact the Foundation as set forth in Section 3
below.

1.F.
1.F.1. Project Gutenberg volunteers and employees expend
considerable effort to identify, do copyright research on,
transcribe and proofread works not protected by U.S. copyright
law in creating the Project Gutenberg™ collection. Despite these
efforts, Project Gutenberg™ electronic works, and the medium
on which they may be stored, may contain “Defects,” such as,
but not limited to, incomplete, inaccurate or corrupt data,
transcription errors, a copyright or other intellectual property
infringement, a defective or damaged disk or other medium, a
computer virus, or computer codes that damage or cannot be
read by your equipment.

1.F.2. LIMITED WARRANTY, DISCLAIMER OF DAMAGES - Except

for the “Right of Replacement or Refund” described in
paragraph 1.F.3, the Project Gutenberg Literary Archive
Foundation, the owner of the Project Gutenberg™ trademark,
and any other party distributing a Project Gutenberg™ electronic
work under this agreement, disclaim all liability to you for
damages, costs and expenses, including legal fees. YOU AGREE
THAT YOU HAVE NO REMEDIES FOR NEGLIGENCE, STRICT
LIABILITY, BREACH OF WARRANTY OR BREACH OF CONTRACT
EXCEPT THOSE PROVIDED IN PARAGRAPH 1.F.3. YOU AGREE
THAT THE FOUNDATION, THE TRADEMARK OWNER, AND ANY
DISTRIBUTOR UNDER THIS AGREEMENT WILL NOT BE LIABLE
TO YOU FOR ACTUAL, DIRECT, INDIRECT, CONSEQUENTIAL,
PUNITIVE OR INCIDENTAL DAMAGES EVEN IF YOU GIVE
NOTICE OF THE POSSIBILITY OF SUCH DAMAGE.

1.F.3. LIMITED RIGHT OF REPLACEMENT OR REFUND - If you

discover a defect in this electronic work within 90 days of
receiving it, you can receive a refund of the money (if any) you
paid for it by sending a written explanation to the person you
received the work from. If you received the work on a physical
medium, you must return the medium with your written
explanation. The person or entity that provided you with the
defective work may elect to provide a replacement copy in lieu
of a refund. If you received the work electronically, the person
or entity providing it to you may choose to give you a second
opportunity to receive the work electronically in lieu of a refund.
If the second copy is also defective, you may demand a refund
in writing without further opportunities to fix the problem.

1.F.4. Except for the limited right of replacement or refund set

forth in paragraph 1.F.3, this work is provided to you ‘AS-IS’,
WITH NO OTHER WARRANTIES OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR ANY PURPOSE.

1.F.5. Some states do not allow disclaimers of certain implied

warranties or the exclusion or limitation of certain types of
damages. If any disclaimer or limitation set forth in this
agreement violates the law of the state applicable to this
agreement, the agreement shall be interpreted to make the
maximum disclaimer or limitation permitted by the applicable
state law. The invalidity or unenforceability of any provision of
this agreement shall not void the remaining provisions.

1.F.6. INDEMNITY - You agree to indemnify and hold the

Foundation, the trademark owner, any agent or employee of the
Foundation, anyone providing copies of Project Gutenberg™
electronic works in accordance with this agreement, and any
volunteers associated with the production, promotion and
distribution of Project Gutenberg™ electronic works, harmless
from all liability, costs and expenses, including legal fees, that
arise directly or indirectly from any of the following which you
do or cause to occur: (a) distribution of this or any Project
Gutenberg™ work, (b) alteration, modification, or additions or
deletions to any Project Gutenberg™ work, and (c) any Defect
you cause.
 Section 2. Information about the Mission
of Project Gutenberg™
Project Gutenberg™ is synonymous with the free distribution of
electronic works in formats readable by the widest variety of
computers including obsolete, old, middle-aged and new
computers. It exists because of the efforts of hundreds of
volunteers and donations from people in all walks of life.

Volunteers and financial support to provide volunteers with the

assistance they need are critical to reaching Project
Gutenberg™’s goals and ensuring that the Project Gutenberg™
collection will remain freely available for generations to come. In
2001, the Project Gutenberg Literary Archive Foundation was
created to provide a secure and permanent future for Project
Gutenberg™ and future generations. To learn more about the
Project Gutenberg Literary Archive Foundation and how your
efforts and donations can help, see Sections 3 and 4 and the
Foundation information page at www.gutenberg.org.

Section 3. Information about the Project

Gutenberg Literary Archive Foundation
The Project Gutenberg Literary Archive Foundation is a non-
profit 501(c)(3) educational corporation organized under the
laws of the state of Mississippi and granted tax exempt status
by the Internal Revenue Service. The Foundation’s EIN or
federal tax identification number is 64-6221541. Contributions
to the Project Gutenberg Literary Archive Foundation are tax
deductible to the full extent permitted by U.S. federal laws and
your state’s laws.

The Foundation’s business office is located at 809 North 1500

West, Salt Lake City, UT 84116, (801) 596-1887. Email contact
links and up to date contact information can be found at the
Foundation’s website and official page at
www.gutenberg.org/contact

Section 4. Information about Donations to

the Project Gutenberg Literary Archive
Foundation
Project Gutenberg™ depends upon and cannot survive without
widespread public support and donations to carry out its mission
of increasing the number of public domain and licensed works
that can be freely distributed in machine-readable form
accessible by the widest array of equipment including outdated
equipment. Many small donations ($1 to $5,000) are particularly
important to maintaining tax exempt status with the IRS.

The Foundation is committed to complying with the laws

regulating charities and charitable donations in all 50 states of
the United States. Compliance requirements are not uniform
and it takes a considerable effort, much paperwork and many
fees to meet and keep up with these requirements. We do not
solicit donations in locations where we have not received written
confirmation of compliance. To SEND DONATIONS or determine
the status of compliance for any particular state visit
www.gutenberg.org/donate.

While we cannot and do not solicit contributions from states

where we have not met the solicitation requirements, we know
of no prohibition against accepting unsolicited donations from
donors in such states who approach us with offers to donate.

International donations are gratefully accepted, but we cannot

make any statements concerning tax treatment of donations
received from outside the United States. U.S. laws alone swamp
our small staff.
Please check the Project Gutenberg web pages for current
donation methods and addresses. Donations are accepted in a
number of other ways including checks, online payments and
credit card donations. To donate, please visit:
www.gutenberg.org/donate.

Section 5. General Information About

Project Gutenberg™ electronic works
Professor Michael S. Hart was the originator of the Project
Gutenberg™ concept of a library of electronic works that could
be freely shared with anyone. For forty years, he produced and
distributed Project Gutenberg™ eBooks with only a loose
network of volunteer support.

Project Gutenberg™ eBooks are often created from several

printed editions, all of which are confirmed as not protected by
copyright in the U.S. unless a copyright notice is included. Thus,
we do not necessarily keep eBooks in compliance with any
particular paper edition.

Most people start at our website which has the main PG search
facility: www.gutenberg.org.

This website includes information about Project Gutenberg™,

including how to make donations to the Project Gutenberg
Literary Archive Foundation, how to help produce our new
eBooks, and how to subscribe to our email newsletter to hear
about new eBooks.
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com