IPASJ International Journal of Computer Science(IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 51






ABSTRACT
Image fusion is the method of combining significant information from two or more images into a single image. In this work,
image fusion using mean and variance is explored in transform domain. Discrete Cosine Transform (DCT), Discrete Wavelet
Transform (DWT) and Non Subsampled Contourlet Transform (NSCT) are used for image fusion. Feature Similarity (FSIM)
index is used to test the quality of the fused image. Experimental results indicate that the proposed methods are producing
better results than other methods available in the literature.
Keywords: Mean, Variance, DCT, DWT, NSCT
1. INTRODUCTION
Image fusion has become a regulation which demands more general formal solutions to different applications. Some
applications require both high spatial and high spectral content in a single image. However, the instruments are not
able of providing such content. One potential solution for this is image fusion. Image fusion methods can be
implemented in two domains - spatial domain and transform domain. The fusion methods such as averaging, high pass
filtering contrast, and variance based methods fall under spatial domain. The disadvantage with spatial methods is that
they produce spatial deformation in the fused image. This problem is well addressed in frequency domain. Image fusion
can be implemented in spatial domain by manipulating the pixel intensities. It can also be implemented by
manipulating the coefficients of transformed image using various transforms, viz., Discrete Fourier Transform (DFT),
Discrete Sine Transform (DST), Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT) and Non
Subsampled Contourlet Transform (NSCT) etc.

DCT has attracted widespread interest as a method of information coding [14 ]. ISO/CCITT Joint Photographic Experts
Group (JPEG) has selected DCT for its baseline coding technique . DCT provides better energy compaction for natural
images than other block based transforms. DWT has attracted widespread interest as a method of information coding
[1,6]. The ISO/CCITT Joint Photographic Experts Group (JPEG2000) has selected the DWT for its baseline coding
technique. DWT is used because of its ability to solve the blocking artifacts introduced by DHT and DCT based image
processing applications. The idea of wavelet transform is to convolve the image to be analyzed with several oscillatory
filter kernels representing different frequency bands, respectively. Hence, image is split into different frequency bands
by the DWT. The image is split into four bands if the level of decomposition is one. These bands are denoted by LL,
LH, HL, and HH. L indicates low pass filter and H indicates high pass filter. Low pass information is preserved in
the LL band. Hence, this band contains more redundancy than other bands. CT uses double filter bank structures for
obtaining sparse expansion of typical images having smooth contours [4,5 ]. In this double filter bank structure,
Laplacian Pyramid (LP) is used to capture the point discontinuities, and Directional Filter Bank (DFB) is used to link
these point discontinuities into linear structures. The NSCT expansion is collected of basis images leaning at various
directions in multiple scales, with bendable aspect ratios.

In this work DCT, DWT, and NSCT are considered to test image fusion algorithms. DCT, DWT, and NSCT are
reviewed in section 2. Proposed fusion algorithms are presented in section 3. Experimental results are discussed in
section 4. Concluding remarks are presented in section 5.
Image Fusion using various Transforms

G.Ramesh Babu
1
and K.Veera Swamy
2


1
P.G.Student, ECE Department, QIS College of Engineering &Technology,Ongole

2
Professor, ECE Department, QIS College of Engineering &Technology,Ongole
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 52

2. TRANSFORMS
a. 2D-DCT
An image block with a more changes in intensity values has a very random looking resulting DCT coefficient matrix,
while an image matrix of just no change in intensity values, has a resulting matrix of a large value for the first element
and zeroes for the other elements. The first element in DCT coefficient matrix[14] represents DC value of image
matrix. DCT is calculated for an image of size nxn is given below:

( ) ( ) ( ) ( )
( ) ( )
(

+
(

+
=

=
n
v y
n
u x
y x g v u
n
v u G
n
x
n
y
2
1 2
cos
2
1 2
cos ,
2
1
,
1
0
1
0

(1)

where
where,

( )

>
=
=
0 1
0
2
1
u
u
u

( )

>
=
=
0 1
0
2
1
v
v
v


IDCT is calculated using equation (2)


IDCT The inversis given as below.

( ) ( ) ( ) ( )
( ) ( )
(

+
(

+
=

=
n
v y
n
u x
v u G v u
n
y x g
n
u
n
v
2
1 2
cos
2
1 2
cos ,
2
1
,
1
0
1
0

(2)

DCT gets its name from the fact that the basis matrix (
D
A ) is obtained as a function of cosines.
D
A of any order is
generated by using the equation (3).

( )
( )
( )

= =
+
= =
+
=
. 1 ,..., 1 , 0 , 1 ,..., 2 , 1
2
1 2
cos
2
1 ,..., 1 , 0 , 0
2
1 2
cos
1
,
n j n i
n
i j
n
n j i
n
i j
n
j i A
D

(3)




b. 2D-DWT

In DWT[1,6,7], an image can be processed by passing it through an analysis filter bank followed by a decimation
operation [ ]. This analysis filter bank, which consists of a low pass and a high pass filter at each decomposition stage,
is commonly used in image compression. When a image passes through these filters, it is split into two bands. The low
pass filter, which corresponds to an averaging operation, extracts the coarse information of the image. The high pass
filter, which corresponds to a differencing operation, extracts the detail information of the image. The output of the
filtering operations is then decimated by two. The mother wavelet is defined as
( ) |
.
|

\
|
=
a
b x
a
x
b a

1
,
(4)
where, a is scale coefficient and b is shift coefficient. All other basis functions are variations of the mother wavelet.
In two dimension DWT, a two dimensional scaling function, ( ) y x, and three two dimensional wavelets,
( ) ( ), , , , y x y x
V H
and ( ) y x
D
, , are required. Each is the product of a one dimensional scaling function and
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 53

corresponding wavelet . Excluding products that produce one dimensional results, like ( ) ( ) x x , the four
remaining products produce the separable scaling function
( ) ( ) ( ) y x y x = , (5)
and separable, directional sensitive wavelets
( ) ( ) ( ) y x y x
H
= , (6)
( ) ( ) ( ) y x y x
V
= , (7)
( ) ( ) ( ) y x y x
D
= , (8)
These wavelets measure functional variations, intensity or gray level variations for images, along different directions.
H
measures variations along columns (horizontal edges),
V
corresponds to variations along rows (vertical edges),
and
D
corresponds to variations along diagonals. Scaled and translated basis functions are as follows:
( ) ( ) m y m x y x
j j
j
m m j
= 2 , 2 2 ,
2
, ,
(9)
( ) { } D V H i m y m x
j j i
j
m m j
i
, , , 2 , 2 2
2
, , = = . (10)
where, index i identifies the directional wavelets. j is scaling function. DWT of image ( ) y x g , of size M X M
is then,
( ) ( ) ( ) y x y x g
MM
m m j
m m j
M
x
M
y
, ,
1
, ,
, .
1
0
1
0
0
0

=
= + (11)
( ) ( ) ( ) y x y x g
MM
m m j
i
m m j
M
x
M
y
i
, ,
1
, ,
, .
1
0
1
0

=
= + (12)
Inverse DWT is calculated using equations (11) and (12). Inverse DWT calculation is given in equation (13).
( ) ( ) ( )
( ) ( ) y x m m j
MM
y x m m j
MM
y x g
i
m m j
D V H i j j m m
i
m m j
m m
, , ,
1
, , ,
1
,
, ,
, ,
, , 0
0
0

=
+ +
+ =
(13)

The decomposition based on the scale is given in the Figure 1.










Figure 1 Decomposition in DWT domain


Low pass (L) filtered coefficients are stored on the left part of the matrix and high pass (H) filtered on the right.
Because of the decimation, the total size of the transformed image is same as the original image. The image is split into
four bands denoted by LL, HL, LH and HH after one level decomposition and as shown in Figure 2.







( ) m m j
H
, ,

+


( ) m m j
V
, ,

+


( ) m m j
D
, ,

+


( ) m m j , ,

+



( ) m m j , , 1 + +

IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 54


























Figure 2 Pyramidal decomposition

In the second level, the LL band is further decomposed into four bands. The same procedure is continued for further
decomposition levels. This process of filtering the image is called pyramidal decomposition of image. DWT performs
multiresolution image analysis [2,3 ]. Wavelet decomposition at second level is shown in Figure 3.










Figure 3 Pyramidal structure of wavelet decomposition

c. NSCT

The NSCT[10,11,12] is a shift-invariant version of the CT. The CT[4,5,8,9] employs the LP for multi-scale
decomposition, and the DFB for directional decomposition. To avoid aliasing of the CT the NSCT minimizes the down
sampling and the up sampling .It works using pyramids filter banks (NSPFBs) and the nonsubsampled directional filter
banks (NSDFBs). The NSPFB, employed by the NSCT, is a two-channel nonsubsampled filter bank (NFB). The filters
for next stage are obtained by upsampling the filters of the previous stage with the sampling matrix
L

2


2



L

H

2


2


L

2


H

2


H

Original
Image
LL
LH
HL
HH
Horizontal Vertical
LH2 HH2
HL2 LL2
HL1

LH1

HH1
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 55


(

= =
2 0
0 2
2I D
This gives the multi-scale property without the need of additional filter design. Illustrates the proposed NSPFB
decomposition with J=3 stages. The ideal frequency support of the lowpass filter at the j-th stage is the region [-( /2
j
),
( /2
j
)]
2
. Accordingly, the ideal support of the equivalent highpass filter is the complement of the lowpass, namely, the
region [-( /2
j-1
),( /2
j-1
)]
2
\[-( /2
j
), ( /2
j
)]
2
and the equivalent filters of a J -level cascading NSPFB is given by


(14)

Where H
0
(z) and H
1
(z) denote the lowpass filter and the corresponding highpass filter at the first stage, respectively.
After J-level NSCT decomposition, one lowpass subband image and

=
J
j
l
j
1
2 (l
j
denotes the direction decomposition
level at the j-th scale) bandpass directional subband images are obtained, all of which have the same size as the input
image.

3. PROPOSED FUSION ALGORITHMS

Block diagram for fusion process is shown in Figure 4.
















Figure 4 Fusion process

3.1 Fusion algorithm using 2D-DCT and Mean

1. Decompose source images into sub- blocks
2. Apply 2D-DCT for each sub-block using equation (1)
3. Compute mean for each sub-block
4. Compare mean values of each image sub-blocks
5. Choose the sub-block which has high value of mean
6. For all these sub-blocks compute 2D-IDCT using equation (2) to get the fused image

3.2 Fusion algorithm using 2D-DCT and Variance

1. Decompose source images into sub- blocks

+ =
s s
=
[
[

1 , ) (
1 ), ( ) (
) (
2
0
2
0
2
0
2
0
2
1
1
j n z H
j n z H z H
z H
n
j
I
n
j
I I
eq
n
j
j n

Source Image 1

Source Image 2

Transform

Transform

Fusion decision
map

Inverse
Transform
Fused
Image
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 56

2. Apply 2D-DCT for each sub-block using equation (1)
3. Compute variance for each sub-block
4. Compare variance values of each image sub-blocks
5. Choose the sub-block which has high value of variance
6. For all these sub-blocks compute 2D-IDCT using equation (2) to get the fused image

3.3 Fusion algorithm using 2D-DWT and Mean

1. Decompose source images into sub- blocks
2. Apply 2D-DWT for each sub-block using equations (11) and (12)
3. Compute mean for each sub-block
4. Compare mean values of each image sub-blocks
5. Choose the sub-block which has high value of mean
6. For all these sub-blocks compute 2D-IDWT using equation (13) to get the fused image

3.4 Fusion algorithm using 2D-DWT and Variance

1. Decompose source images into sub- blocks
2. Apply 2D-DWT for each sub-block using equations (11) and (12)
3. Compute variance for each sub-block
4. Compare variance values of each image sub-blocks
5. Choose the sub-block which has high value of variance
6. For all these sub-blocks compute 2D-IDWT using equation (13) to get the fused image

3.5 Fusion algorithm using NSCT and mean

1. Decompose source images into sub- blocks
2. Apply NSCT for each sub-block using procedure explained in section 2.3
3. Compute mean for each sub-block
4. Compare mean values of each image sub-blocks
5. Choose the sub-block which has high value of mean
6. For all these sub-blocks compute INSCT to get the fused image

3.6 Fusion algorithm using NSCT and Variance

1. Decompose source images into sub- blocks
2. Apply NSCT for each sub-block explained in section 2.3
3. Compute variance for each sub-block
4. Compare variance values of each image sub-blocks
5. Choose the sub-block which has high value of variance
6. For all these sub-blocks compute INSCT to get the fused image

4. EXPERIMENTAL RESULTS

Experiments are performed on clock image. Here, two metrics are used to test the performance of the proposed
algorithms. One metric is Peak Signal to Noise Ratio(PSNR) and another metric is Feature Similarity (FSIM) [ 13]
index . An often used global objective quality measure is the mean square error (MSE) defined as
| |
2
1
0
1
0
'
) , ( ) , (
*
1
=

=

=
n
x
m
y
y x g y x g
m n
MSE (15)
where, n and m is number of rows and columns in the image respectively. ) , ( y x g and
'
) , ( y x g are the pixel
values in the original and reconstructed image. The Peak Signal to Noise Ratio (PSNR in dB) [14] is calculated as
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 57

MSE
PSNR
2
10
255
log 10 =
(16)
where, the usable gray level values range from 0 to 255. Feature Similarity (FSIM) index can be measured as

O e
O e
=
x
m
x
m L
x PC
x PC x S
FSIM
) ( .
) ( ). (
(17)
where means the whole image spatial domain,
m
PC means Phase congruent structure,
L
S is the similarity depends
upon the gradient measure. Fused images are shown in Figure 5. The haar wavelet decomposition is used in the 2D-
DWT-based. In the NSCT-based decomposition two filters (directional filter and pyramidal filter) are used.



( (a a) ) ( (b b) ) ( (c c) )
Figure 5 The clock source images (size of 128128) and fusion results of different algorithms: (a) fused image using
2-D DCT method (b) fused image using 2-D DWT method (c) fused image using NSCT method


Experimental results are tabulated in Table 1. Results indicate that the DWT and NSCT are outperforming with
variance based fusion than other methods. However, DCT is giving better results with mean based fusion.

Table 1: Performance of proposed fusion algorithms
Method
FSIM for
Mean
FSIM for
Variance
PSNR for
Mean
PSNR for
Variance
2D-DCT

0.9595

0.9662

30.8623

30.3379
2D-DWT 0.9503 0.9703 30.5118 34.0510
NSCT 0.9405 0.9703 29.8426 34.0510


Experiments are also performed with medical images also. Results are shown in Figure 6.


Figure 6 (a) CT image (b) MRI image (c) fused image using NSCT method
5. CONCLUSIONS
In this paper mean and variance based fusion algorithms in DCT , DWT , and NSCT domains are proposed and
discussed. Variance+DWT and variance+NSCT stood better than other possibilities. Further, blocking artifacts are
eliminated in DWT and NSCT methods. Results are good with multi focus and medical images. These techniques will
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: [email protected]
Volume 2, Issue 1, January 2014 ISSN 2321-5992

Volume 2 Issue 1 January 2014 Page 58

be useful for Doctors for easy diagnosis. The same techniques can be extended to all other transforms available in the
literature.
Acknowledgements
Authors would like to thank AICTE authorities for providing facilities to carry this work under Research Promotion
Scheme. Authors utilized the equipment and software purchased under RPS(Ref.No.:20/AICTE/RIFD/RPS(Policy-
III)55/2012-13) .
References
[1] G. Pajares, J.M. de la Cruz, A wavelet-based image fusion tutorial, Pattern Recognition 37 (9) (2004) 18551872.
[2] H.A. Eltoukhy, S. Kavusi, A computationally efficient algorithm for multi-focus image reconstruction,
Proceedings of SPIE Electronic Imaging (June 2003) 332341.
[3] Z. Zhang, R.S. Blum, A categorization of multiscale-decomposition based image fusion schemes with a
performance study for a digital camera application, Proceedings of IEEE 87 (8) (1999) 13151326.
[4] P.J. Burt, E.H. Adelson, The Laplacian pyramid as a compact image code, IEEE Transactions on Communications
31 (4) (1983) 532540.
[5] A. Toet, Image fusion by a ratio of low-pass pyramid, Pattern Recognition Letters 9 (4) (1989) 245253.
[6] H. Li, B.S. Manjunath, S.K. Mitra, Multisensor image fusion using the wavelet transform, Graphical Models and
Image Processing 57 (3) (1995) 235245.
[7] I. De, B. Chanda, A simple and efficient algorithm for multifocus image fusion using morphological wavelets,
Signal Processing 86 (5) (2006) 924936.
[8] M.N. Do, M. Vetterli, The contourlet transform: an efficient directional multiresolution image representation,
IEEE Transactions on Image Processing 14 (12) (2005) 20912106.
[9] R.H. Bamberger, M.J.T. Smith, A filter bank for the directional decomposition of images: theory and design,
IEEE Transactions on Signal Processing 40 (4) (1992) 882893.
[10] A.L. da Cunha, J.P. Zhou, M.N. Do, The nonsubsampled contourlet transform: theory, design, and applications,
IEEE Transactions on Image Processing 15 (10) (2006) 30893101.
[11] A.L. da Cunha, J.P. Zhou, M.N. Do, Nonsubsampled contourlet transform: filter design and application in
denoising, in: IEEE International Conference on Image Processing, Genoa, Italy, 1114 September, 2005,
pp.749752.
[12] J.P. Zhou, A.L. da Cunha, M.N. Do, Nonsubsampled contourlet transform: construction and application in
enhancement, in: IEEE International Conference on Image Processing, Genoa, Italy, 1114 September, 2005,
pp.469476.
[13] Lian Zhang and X.Mou, FSIM: A Feature Similarity Index for Image Quality Assesment, IEEE Transactions on
Image Processing , 20 (8) (2011) 23782386.
[14] G.K. Wallace, The JPEG still picture compression standard, IEEE Transactions on Consumer Electronics 38 (1)
(1992) xviii-xxxiv.

AUTHORS

G.Ramesh Babu received the B.Tech degree in Electronics and communication engineering from Prakasam
Engineering College in 2011. He joined M.Tech through Gate Score. Now, He is pursuing M.Tech in
Digital Electronics and Communication Systems in QIS College of Engineering and Technology, Ongole.


K.Veera Swamy received his M.Tech and Ph.D from JNTU, Hyderabad. He has 16 years teaching
experience. He published 49 research papers in International, National journals and presented in reputed
conferences (International Journals-12, National Journals-01, International conferences-21, National
Conferences-15). One of the reviewers for the prestigious Journals IET Image Processing, UK and
International Journal of Computer Theory and Engineering, IACSIT. He executed MODROBS projects
worth Rs.14.5 Lakhs in the area of Antennas. Presently, he is executing Research Project worth Rs 12.5 Lakhs in the
area Development of medical image segmentation algorithms for easy diagnosis sponsored by AICTE.