Volume 2, Issue 2, February 2012 ISSN: 2277 128X

International Journal of Advanced Research in

Computer Science and Software Engineering
Research Paper
Available online at: www.ijarcsse.com
Combined Denoising and Fusion of
Multi Focus images

K. Kannan
*
S. Arumuga Perumal


Kamaraj College of Engineering and Technology S.T. Hindu College,




Abstract Due to the limitation of depth of field in CCD devices which are used to capture the image of the scene of
interest, it is impossible to get an image that contains all relevant objects in the scene in focus. A possible solution to
bring all objects in the scene in focus is image fusion. And also images are often corrupted by noise due to errors
generated in imaging sensors. So, it is important to eliminate noise in the images before some subsequent processing.
In this paper, it is proposed to denoise the images using Decision Based Algorithm for Removal of High-Density
Impulse Noises (DBAIN) method in spatial domain followed by fusion using Higher Density Discrete Wavelet
Transform (HDWT) to form the multi focused images.

Keywords Image Denoising, Image Fusion, DBAIN and HDWT.

I. INTRODUCTION
The driving forces in todays manufacturing
environments are quality improvement, cost reduction,
increased volume of production and shorter cycle
times of manufacturing. The quality of many raw
materials, parts, and products can be measured by
electrical or mechanical means or by visual inspection.
Inspection by eye is costly, subjective, qualitative,
inaccurate, eye-straining, and time consuming. For
high speed and real time applications, manual
inspection is not possible. Therefore, more and more
manufacturers look for fast, accurate, reliable, and
consistent machine vision system for automated visual
inspection of their products. Machine Vision involves
acquisition of the image of an object of interest
followed by processing and interpretation of this
image using computer for useful applications like
inspection of quality of raw materials, parts and
products in industries. The first step in machine vision
is to acquire the images of objects in the scene. When
the images of objects in the scene are taken from the
camera, only the objects within the depth of field of
camera are clearly focused and the other objects in the
scene will be blurred. To get the image in which every
object in the scene is focused, there is need to fuse the
images taken from the same view point with different
focus settings. The resulting composite image, called
as fused image, will contain clear images of all objects
in the scene. This is known as multi-focus image
fusion. There are a number of methods for multi-focus
image fusion. These methods are based on the
assumption that images are noise-free. However, in
practical situations, images are often corrupted by
noise in image acquisition or transmission. Thus, it is
necessary to investigate joint fusion and denoising of
noisy multi-focus images. In this paper, it is proposed
to denoise the images corrupted by Salt and Pepper
Noise using Decision Based Algorithm for Removal
of High-Density Impulse Noises (DBAIN) method in
spatial domain followed by fusion using Higher
Density Discrete Wavelet Transform (HDWT) to form
the multi focused images.
II. IMPULSE NOISE IN IMAGES
Digital Images are often corrupted by Impulse
noise which is always independent and uncorrelated to
the image pixels and is randomly distributed over the
image. So, unlike Gaussian noise, in impulse noise
corrupted image, not all the image pixels are noisy,
but small number of image pixels will be noisy and
the rest of pixels will be noise free. There are different
types of impulse noise namely fixed valued impulse
noise, also called as salt and pepper noise, and random
valued impulse noise. The salt and pepper type of
noise is easier to restore whereas random valued
impulse noise is somewhat difficult. In salt and pepper
type of noise the noisy pixels takes either salt value
(225) or pepper value (0) and it appears as black and
white spots on the images. If p is the total noise
density then salt noise and pepper noise will have a
noise density of p/2.This can be mathematically
expressed by
y
i,j
=
0 or 255 with probability p
x
i,j
with probability 1 p

(1)
Volume 2, issue 2, February 2012 www.ijarcsse.com
2012, IJARCSSE All Rights Reserved
where y
ij
represents the noisy image pixel, p is the
total noise density of impulse noise and x
ij
is the
uncorrupted image pixel. In case, if salt noise and
pepper noise may have different noise densities p1 and
p2, then the total noise density will be p=p1+ p2 .In
case of random valued impulse noise, noise can take
any gray level value from zero to 225. In this case,
also noise is randomly distributed over the entire
image and probability of occurrence of any gray level
value as noise will be same. This can be
mathematically expressed by,
y
i,j
=
n
i,j
with probability p
x
i,j
with probability 1 p
(2)
where n
ij
is the gray value of the noisy pixel.

III. IMAGE FUSION TECHNIQUE
The image of a scene formed by an optical system
used in machine vision contains information about the
distance of objects in the scene. Objects at a particular
distance are focused whereas other objects are
defocused or blurred by different degrees depending
on their distance. This is shown in figure1. The point,
P, on the object plane is clearly focused and perfectly
imaged as point P, on the image plane. For a camera
with a lens of focal length 'f', the relation between the
position of a point close to the optical axis in the scene
and the position of its focused image is given by the
well known lens formula
V
+
U
=
f
1 1 1
(3)
where U is the distance of the object, and V is the
distance of the image.

Fig 1: Model of Lens System

Therefore, in the image formed by a camera, only
objects at a certain distance are in focus; other objects
are blurred by varying degrees depending on their
distance. Further, each lens must have a finite aperture
of diameter D, which can be used to estimate radius of
the blur circle induced as
V S
=
V
D

2R
(4)
where S is distance between sensor plane and the lens.
Then the depth of field of a lens system can be given
as,
near far
U U = DOF (5)
U
D
R
f
)
D
R
Uf(
= U
far
2
2 1

(6)
U
D
R
+ f
)
D
R
+ Uf(
= U
near
2
2 1
(7)
where U
near
and U
far
are the distances to the nearest
and farthest object planes with blur circles less than or
equal to the chosen R and U is the distance to the in-
focus object plane. Due to limited depth of field of
lens, it is possible to take clear image of the objects in
the scene which are in focus only. The remaining
objects in the scene will be out of focus. A possible
solution to bring clear images of all the objects in the
scene is to combine several pictures taken by the
camera with different focus points into a composite
image. The resulting composite image, called as fused
image, will contain clear images of all relevant objects
in the scene. This is known as multi-focus image
fusion. There are two approaches to image fusion,
namely Spatial Fusion and Transform fusion. In
Spatial fusion, the pixel values from the source images
are summed up and taken average to form the pixel of
the fused image at that location. Transform fusion uses
pyramid or wavelet transform for representing the
source image at multi scale [4, 5]. The most
commonly used wavelet transform is critically
sampled Discrete Wavelet Transform (DWT) which
can be implemented using perfectly reconstructed
Finite Impulse Response filter banks. But, critically
sampled DWT suffers from four shortcomings namely
Oscillations, Shift variance, poor directionality and
aliasing. Shift variance in critically sampled discrete
wavelet transform exists due to down sampling during
analysis and up sampling during synthesis [1, 2].
Improved performance can be found using an over
complete or redundant wavelet transform in variety of
signal and image processing applications. For
example, the Undecimated Discrete Wavelet
Transform (UDWT), which is expansive by the factor
3J+1, when J scales are implemented, shows improved
results due to its shift invariant property. Complex
Wavelet Transform (CWT) is also an alternate and
complex valued extension to DWT with limited
redundancy. CWT uses complex valued filtering that
decomposes the real or complex signal into real and
imaginary parts in transform domain [14-19]. It is
approximately shift invariant and directionally
selective in higher dimensions. It achieves this with a
redundancy factor of only 2
d
for d-dimensional
signals, which is lower than the UDWT. The double-
density discrete wavelet transform (DDWT) which
provides a compromise between the UDWT and the
critically-sampled DWT is two-times expansive,
regardless of the number of scales implemented. Even
so, the DDWT is approximately shift-invariant [10].
Like the CWT, the DDWT is redundant by a factor of
4 for two dimensions, independent of the number of
Volume 2, issue 2, February 2012 www.ijarcsse.com
2012, IJARCSSE All Rights Reserved
levels. These above said expansive transform do not
increase the sampling with respect to frequency or
scale. An expansive dyadic wavelet transform, namely
High Density Discrete Wavelet Transform (HDWT)
over samples both space and frequency by a factor two
[11]. This paper uses HDWT to fuse the denoised
images.

IV. DBAIN ALGORITHM FOR IMAGE DENOISING
Decision Based Algorithm for Removal of High-
Density Impulse Noises (DBAIN) method in spatial
domain is a recently proposed algorithm to remove
salt and pepper noise. In DBAIN, each Pixel is
processed for de noising using a 3 X 3 window.
During processing if a pixel gray scale value is 0 or
255 then it is processed, else it is left unchanged. In
DBAIN, the corrupted pixel is replaced by the median
of the window. At higher noise densities, the median
itself will be noisy, and, the processing pixel will be
replaced by the neighbourhood processed pixel.
Step 1) A 2-D Window of size 33 is selected.
Assume the pixel to be processed is P(X, Y).
Step 2) the pixel values inside the window are sorted
and the first element of the window is the Minimum
value P
min
, the last element of the Window is the
Maximum value P
max
and the middle element of the
window is the median value P
med
.
Step 3) Case 1: The P(X, Y) is an uncorrupted pixel
if P
min
< P(X, Y) <P
max
, P
min
>0 and P
max
<255: the pixel
being processed is left unchanged .otherwise P(X, Y)
is a corrupted pixel. Case 2: If P(X, Y) is a corrupted
pixel, it is replaced by its median value if P
min
<P
med
<
P
max
and 0<P
med
<255. Case 3: If P
min
<P
med
< P
max
is not
satisfied or 255<P
med
=0,then P
med
is a noisy pixel .In
this case, the P(X,Y) is replaced by the value of
neighbourhood pixel value.
Step 4) Steps 1 to 3 are repeated until the processing is
completed for the entire image.

V. HIGHER DENSITY DWT
The higher density DWT is an expansive dyadic
wavelet transform that over samples both space and
frequency by a factor of two. Like DDWT, at each
scale of HDWT, there are twice as many coefficients
as the critically sampled DWT. However, HDWT also
has intermediate scales; it has one scale between each
pair of scales of the critically-sampled DWT. The
over complete wavelet basis associated with this
expansive transform has two generators,
i
(t), i = 1, 2.
The spectrum of the first wavelet
1
() is
concentrated between the spectrum of the second
wavelet
2
() and the spectrum of its dilated version

2
(2) In addition, the second wavelet is translated by
integer multiples of one half, rather than whole
integers. The transform can be implemented with
digital filter banks like the conventional DWT as
shown in the following figure 2. Similar to DDWT, it
uses three filters, one scaling and two wavelet filters.
However, one of the wavelet filters is band pass
instead of high-pass filters. And also the high pass
filter is not down sampled and up sampled during
analysis and synthesis. The analysis filter bank
structure of HDWT is shown in fig 2. Therefore, the
2-D form of the HDWT is 5-times expansive.

Fig 2: 2D Higher Density Discrete Wavelet Transform
VI. PROPOSED WORK
A pair of source images focusing different objects
in the scene of size 640 X 480 in GIF format is taken
for study and is assumed to be registered spatially. A
salt and pepper noise is added to each image in the
pair and DBAIN algorithm is applied to restore the
original image. Then, the images are transformed into
wavelet domain by taking HDWT transform. There
will be nine sub-bands for each image after HDWT
was taken namely LL ,LB, LH, BL, BB, BH, HL, HB,
HH, where L stands for low pass, H stands for high
pass, and B stands for band pass. To form the fused
coefficient map for sub-bands of LB, LH, BL, BB,
BH, HL, HB and HH, the wavelet coefficients from
the source images whose absolute value is maximum
was selected. For, LL sub band the fused coefficients
are calculated as follows.
- Canny edge detector is applied to the LL sub band
of the each source images.
- After the edge detection, region segmentation is
performed based on the edge information using
region labelling algorithm. In the labelled image,
zero corresponds to the edges and other different
value represents different regions in the image.
- The focus measure Spatial Frequency is
calculated as the activity level of region k in LL
sub band of the each source images using the
formulae [29 ],
Volume 2, issue 2, February 2012 www.ijarcsse.com
2012, IJARCSSE All Rights Reserved
( )
( )

M
= x
N
= y
M
= x
N
= y
2 2
y) f(x y) f(x,
MXN
= RF
and ) y f(x, y) f(x,
MXN
= RF
where
CF + RF = SF
2 1
2
1 2
2
1,
1
1
1

(8)

The approximation sub band of the fused image F
is taken from the approximation sub band wavelet
coefficients of source images whose activity
measure in the particular region is high.
VII. EVALUATION CRITERIA
There are four evaluation measures are used in this
paper, as follows. The Root Mean Square Error
(RMSE) between the reference image R and fused
image F is given by [8],
2
1 1
2
N
j)] F(i, j) [R(i,
= RMSE
N
= i
N
j=

(9)

The Peak Signal to Noise Ratio (PSNR) between
the reference image R and fused image F is given
by [8],
( ) MSE = PSNR / 255 10log
10
(10)

Quality index of the reference image (R) and fused
image (F) is given by [7],
) + )( b + (a
ab
= QI
b a
ab
2 2
2 2
4
(11)
The maximum value Q=1 is achieved when two
images are identical, where a & b are mean of
images,
ab
o
be covariance of R & F,
2
a
o
,
2
b
o

be the variance of image R, F. The Normalized
Weighted Performance Metric (NWPM) which is
given in the equation as [6],
+
+
=
B
ij
A
ij j i
B
ij
AF
ij
A
ij
AF
ij j i
W W
W Q W Q
NWPM
,
,
(12)

VII. RESULTS
The performance of DBAIN algorithm for
Denoising and HDWT for image fusion is measured in
terms of RMSE, PSNR, QI & NWPM and the results
are tabulated in table I. The results of combined
Denoising and Image fusion are shown in figure 3.
From the table, it is inferred that the proposed method
for combined Denoising and Image fusion out
performs the existing methods.

Fig 3: Results of Combined Denoising and Fusion
A & B. Input Images C &d. Noisy Images
E & F. Denoised Images G. Reference Image
H, I & J. Fused Images with Noise Density of 0.2, 0.4 & 0.6

TABLE I
RESULTS OF COMBINED DENOISING AND FUSION
Noise Density 0 0.2 0.4 0.6
RMSE 2.5737 4.0551 7.4751 10.9786
PSNR 39.6779 35.729 30.6039 27.7622
QI 0.9988 0.9966 0.9875 0.9729
NWPM 0.7009 0.6969 0.6246 0.5872

VIII. CONCLUSION
This paper presents new method for combined
Denoising and fusion of multi focus images HDWT
and their performance is compared in terms of various
performance measures like RMSE, PSNR, QI and
NWPM. Both DBAIN algorithms for image
Denoising and HDWT for image fusion provide very
good results both quantitatively & qualitatively. Hence
using the above proposed method, one can enhance
the image with high geometric resolution and better
visual quality.

REFERENCES
[1] S. Mallat, Wavelet Tour of Signal Processing, New York,
Academic Press, 1998.
[2] S. Mallat, A theory for Multiresolution signal decomposition:
Volume 2, issue 2, February 2012 www.ijarcsse.com
2012, IJARCSSE All Rights Reserved
The wavelet representation, IEEE transaction pattern anal.
Machine Intell., vol. 11, no. 7, pp 674-693, July 1989.
[3] Rick S. Blum and Yang Jin zhong, Image Fusion Methods
and Apparatus, US Patent, WO/2006/017233, 2006.
[4] P.J. Burt and E. Adelson, The Laplcian Pyramid as a Image
Codec, IEEE Transactions on Communications, Vol. Com-
31, No.4, pp. 532-540,1983.
[5] P. J. Burt and R. J. Kolczynski, 1993, Enhanced image
capture through image fusion, proceedings of the 4
th

International Conference on Computer Vision, pp. 173-182.
[6] C.S. Xydeas and V. Petrovic, Objective Image Fusion
Performance Measure, Electronics Letter, Vol.36, N0.4, pp.
308-309, 2000.
[7] Zhou Wang and Alan C. Bovik, 2002, A Universal Image
Quality Index, IEEE Signal Processing Letters, Vol. 9, No.3,
pp. 81-84.
[8] Marta Mrak, Sonja Grgic and Mislav Grgic, 2003, Picture
Quality Measures in Image Compression Systems,
Proceedings of EUROCON 2003, pp 233-237.
[9] Selesnick, I.W.,Hilbert transform pairs of wavelet bases,
IEEE Signal Processing Letters, 8(6):170173, June 2001.
[10] I.W.Selesnick, The Double-Density Dual-Tree DWT, IEEE
Trans. on Signal Processing, 2004, 52(5), 1304-1314.
[11] W. Selesnick, A higher-density discrete wavelet transform,
IEEE Trans. on Signal Processing, 54(8):3039-3048, August
2006.
[12] Krista Amolins, Yun Zhang and Peter Dare, Wavelet Based
Image fusion techniques- An introduction, Review and
Comparison, ISPRS Journal of Photogrammetry & Remote
Sensing 62 (2007), 249-263.
[13] Gonzalo Pajares and Jesus Manuel de la Cruz, A Wavelet
based Image fusion Tutorial, Pattern Recognition 37(2004),
1855-1872.
[14] Bayram, I. and I. W. Selesnick, On the dual-tree complex
wavelet packet and M-band transforms, IEEE Trans. on
Signal Processing, 56(6):2298-2310, June 2008.
[15] N. G. Kingsbury, Image processing with complex wavelets,
Philos. Trans. R. Soc. London A, Math. Phys. Sci.,
357(1760):25432560, September 1999.
[16] N. G. Kingsbury, Complex wavelets for shift invariant
analysis and filtering of signals Journal of Appl. and Comp.
Harmonic Analysis, 10(3):234 253, May 2001.
[17] I.W.Selesnick. Smooth wavelet tight frames with zero
moments, Journal of Appl. Comput. Harmon. Anal. 2001,
10(2), 163-181.
[18] N.G. Kingsbury, Complex wavelets for shift invariant
analysis and filtering of signals Applied Computational
Harmonic analysis, vol. 10, no.3, pp 234-253, May 2001.
[19] W. Selesnick, R. G. Baraniuk, and N. G. Kingsbury, The
dual-tree complex wavelet transforms - A coherent framework
for Multiscale signal and image processing IEEE Signal
Processing Magazine, 22(6):123151, November 2005.
[20] R. Yu and H. Ozkaramanli, Hilbert transform pairs of
orthogonal wavelet bases: Necessary and sufficient
conditions IEEE Trans. Signal Processing, 53(12):4723
4725, December 2005
[21] N.G. Kingsbury, The dual tree complex wavelet transform: A
technique for shift invariance and directional filters in proc.
of 8
th
IEEE DSP Workshop, Utah, paper no. 86 August 9-12,
1998.
[22] Zhu Shu-long, Image Fusion using Wavelet Transform,
Symposium on Geospatial Theory, Processing and
Applications, pp. 5-9, 2004.
[23] H. Li, B.S. Manjunath, and S.K. Mitra, Multi-sensor image
fusion using the wavelet transform, Proceedings of the
conference on Graphical Models and Image Processing, pp.
235245, 1995.
[24] N.G. Kingsbury and J.F. A. Magarey, Wavelet transforms in
image processing in proc. of I European Conference on
Signal Analysis and Prediction, Prague, pp 23-34, June 24-27,
1997.
[25] N. G. Kingsbury, The dual-tree complex wavelet transform:
A new efficient tool for image restoration and enhancement,
In Proceedings of European Signal Processing Conference
(EUSIPCO), 1998.
[26] Z. Zhang and R. Blum, A categorization of Multiscale
decomposition based image fusion schemes with a
performance study for a digital camera application, proc. of
the IEEE, pages 1315-1328, August 1999.
[27] N. G. Kingsbury, A dual-tree complex wavelet transform
with improved orthogonality and symmetry properties, In
Proceedings of IEEE International Conference on Image
Processing (ICIP), 2000.
[28] N.G. Kingsbury, A dual tree complex wavelet transform with
improved orthogonality and symmetry properties in proc. of
IEEE International Conference on Image processing, Canada,
vol.2, pp 375-378, September 10-13, 2000.
[29] Li, S., Kwok, J.T., Wang, Y., 2001, Combination of Images
with diverse focuses using the spatial frequency, Int.fusion,
2,169-176.