Study and Analysis ofPCA, DCT & DWT based

Image Fusion Techniques

Mr. Rajenda Pandit Desale

Prof. Sarita V. Verma

Department ofElectronics and CommunicationEngg.,

Department ofElectronics and CommunicationEngg.,

Marathwada Institute ofTechnology,

Marathwada Institute ofTechnology,

Aurangabad (Maharashtra),India.

Aurangabad (Maharashtra), India.

desale.rajendra@gmail.com

sarita verma04@yahoo.com

Abstraet- Image Fusion is

a process of combining the relevant

information from a set of images, into a single image, wherein

the resultant fused image will be more informative and complete
than any of the input images.
This paper discusses the Formulation, Process Flow Diagrams
and algorithms of PCA (principal Component Analysis), DCT
(Discrete

Cosine

Transform)

and

DWT

(Discrete

The covariance of Y, i. e., cov(Y) is a diagonal and inverse of

I
T
V is equivalent to its transpose (V =V ). Using Matrix
Algebra,
T
cov(Y) = E{yy }

(1)

Wavelet

(2)

Transform) based image fusion techniques. The results are also

presented in table & picture format for comparative analysis of

(3)

above techniques.
The PCA & DCT are conventional fusion techniques with
many drawbacks, whereas DWT based techniques are more
favorable as they provides better results for image fusion.

cov(Y) = VT conv(X)V

(4)

MuItiplying both sides of equation (4) by V, we get,

In this paper, two algorithms based on DWT are proposed,

these

are,

pixel averaging &

maximum

pixel

replacement

approach.

Keywords- image fusion; principal eomponent analysis;

diserete eosine transform; diserete wavelet transform
I.

Vcov(Y) = VVT conv(X) V = conv(X) V

Substituting equation (4) into the equation

(5)

gives,

[AIV I,A2V2, . . . . . . ,A..iVdl

INTRODUCTION

Any piece of information makes sense only when it is able

to convey the content across. The c1arity of information is

quality of information from a set of images. By the process of
image fusion the good information from each of the given

Ai Vi = cov(X) Vi

images is fused together to form a resuItant image whose

the images that would make the good information in each of

the image prominent. The fused image is constructed by

B.

(7)

Process Flow Diagram

The input images are

arran ged

in two column vectors and

their empirical means are subtracted. The resulting vector has

a dimension of n x

combining magnified information from the input images.

2,

where n is length of the each image

vector.

PRINCIPAL COMPONENT ANALYSIS

Compute the eigenvector and eigenvalues for this resuIting

Principal components are linear combinations of optimally

weighted observed variables.

vector are computed and the eigenvectors corresponding to

the larger eigenvalue obtained.

In PCA,the number of components extracted is equal to the

number of observed variables being analyzed.
However, in most PCA

analysis, only

the

first

few

components has the largest possible variance, so only these

Image
I1

first few components are retained, interpreted & used for

subsequent analysis.
A.

Formulation
Let X be a d-dimensional random vector and assume it to

have zero empirical mean. The projection matrix V would be

T
such that Y =V X with the following constraints,

978-1-4673-4862-1/13/$31.00 2013

IEEE

(6)

Where,i =1,2,... ,d and Vi is an eigenvector of cov(X ).

quality is superior to any of the input images.

This is achieved by applying a sequence of operations on

[conv(X) VI,cov(X) V 2 , ,cov(X)V d l

This could be rewritten asb

important. Image Fusion is a mechanism to improve the

11.

(5)

Image
12
Figure 1. Image Fusion Process using PCA

2013

The normalized components PI and P2 are computed from

the obtained eigenvector.
The fused image is given by equation,

(8)
PCA Algorithm

C.

Let the source images (images to be fused) be arranged in

two-column vectors. The steps followed to project this data
into 2-D subspaces are:
1. Organize the data into column vectors. The resuIting
matrix Z is of dimension 2 x n.
2. Compute the empirical mean along each column. The
empirical mean vector Me has a dimension of 1 x 2.
3. Subtract the empirical mean vector Me from each column
of the data matrix S. The resulting matrix X is of dimension 2
x n.
4. Find the covariance matrix C of X i.e. C=XXT mean of
expectation = cov(X)
5. Compute the eigenvectors V and eigenvalue D of C and
sort them by decreasing eigenvalue. Both V and D are of
dimension 2 x 2.
6. Consider the fIrst column of V which corresponds to
larger eigenvalue to compute PI and P2 as,
_

P1-

V(1)
LV

DISCRETE

( nl, n2)cos

Where,

a(kl)

,1

(12)

N2 - 1

(13)

'\'Nl':-O1 Lm2
'\'N2':-O1 a (kl ) a(k2)X(kl, k2)cos (TI(2n1+1)kl)
Lml
2Nl
COS

(TI(2n2+l)k2 )

B.

Process Flow Diagram

2N2

For O:s k1:s N1-1

(TI(2n1+l)kl )
2Nl

(10)

Kl
kl

Nl - 1

(11)

(14)

and O:s K2:s N2-1

Images to be fused are divided into non-overlapping blocks

of size NxN as shown in Fig-2. DCT coeffIcients are
computed for each block and fusion rules are applied to get
fused DCT coeffIcients. mCT is then applied on the fused
coeffIcients to produce the fused image/block.
First
Image

DCT

Second
Image

mCT

Fused

... Image

DCT

Figure 2. Image Fusion Process using DCT

The following fusion rule is used for image fusion process,

Let the XI be the DCT coeffIcients of image block from
image 11 and similarly let X2 be the DCT coeffIcients of image
block from image Iz. Assume the image block is of size N x N
and Xf be the fused DCT coeffIcients.
Here, all DCT coefficients from both image blocks are
averaged to get fused DCT coeffIcients. It is very simple and
basic image fusion technique in DCT domain.
(15)
Where k1,k2= 0,1,2,...., N-1
IV.
A.

{ ,
{2
2
..jN1
Nl

k2

Similarly, the 2D inverse discrete eosine transform is

defIned as,

for OklNl-1 and 0 K2N2-1

2N2

,1

(9)

x(k1,k2)=a(kl )a(k2)

(TI(2n2+l)k2 )

Kl & K2 are discrete frequency variables.

COSINE TRANSFORM

Formulation

Fusion
Rule

The 2D discrete eosine transform of an image or 2D signal

x(nl,n2) of size NI xN2 is defIned as,

"'Nl-l "'N2-l
Lml=O Lm2=O

K2

$,

V(2)

A DCT is used to express a sequence of fInite data points in

terms of a sum of eosine functions oscillating at different
frequencies.
Discrete eosine transform (DCT) is an important transform
extensively used in digital image processing. Large DCT
coeffIcients are concentrated in the low frequency region;
hence, it is known to have excellent energy compactness
properties.

cos

a(k2)

1
,fN2

and P2-"fV
III.

A.

International Conference on Signal Processing, Image Processing and Pattern Recognition [ICSIPRl

DISCRETE

WAVELET TRANSFORM

Process Flow Diagram

Wavelet transform is fIrst performed on each source images

to generate a fusion decision map based on a set of fusion
rwes.
The fused wavelet coefficient map can be constructed from
the wavelet coeffIcients of the source images according to the
fusion decision map. Finally the fused image is obtained by
performing the inverse wavelet transform.

2013

International Conference on Signal Processing, Image Processing and Pattern Recognition [ICSIPRl
Algorithms

C.

First
Image

Second
Image

Coeff.
Mapl

Coeff.
Map2

/'

Fused
Coeff.
Map

Fused
Image

---+

IDWT
DWT
Figure 3. Image Fusion Process using DWT

In discrete wavelet transform (DWT) decomposition, the

filters are specially designed so that successive layers of the
pyramid only include details which are not already available
at the preceding levels.

Row wise

Input Image
..
HP

filtering

Row wise LP filtering

..

..

Column wise down

sampling

Column wise down

sampling

..

..

Column
wise HP
filtering

..

Column
wise LP
filtering

Column
wise HP
filtering

..

..

..

Row wise
down
sampling

Row wise
down
sampling
(HL)

Row wise
down
sampling
(LH)

(HR)

C. i. image fusion using Maximum Pixel replac ement

g)Take the pixel having the maximum value of

bands i.e. HHa and HHb and put in HHn.
h)Take the pixel having the maximum value of
bands i.e. HLa and HLb, and put in HLn.
i)Take the pixel having the maximum value of
bands i.e. LHa and LHb, and put in LHn.
j)Take the pixel having the maximum value of
bands i.e. LLa and LLb, and put in LLn.
i) Thus we will get HHn,HLn,LHn and LLn
coefficients.
j) Take Inverse Discrete Wavelet Transform.
k) Obtain the fused Image and Display.

the two

DWT Decomposition

The algorithm of image fusion using DWT has following

common steps applicable to proposed methods of fusion.
a) Accept the two input images.
b) Resize both the images to 256 x 256.
c) Convert to Gray scale image.
d) Convert to double precision format.
e) Take Discrete Wavelet Transform of both the images.
t) Let for first image fOUT bands be HHa, HLa, LHa, LLa
and for second image be HHb HLb, LHb, LLb.
,

The fusion rwes play a very important role during the

fusion process
B.

the two
the two
the two
as new

image Fusion using Pixels Averaging

C.2.

g) Take the average of pixels of the two bands i.e. HHa and
HHb and put in HHn.
h) Take the average of pixels of the two bands i.e. HLa and
HLb, and put in HLn.
i) Take the average of pixels of the two bands i.e. LHa and
LHb, and put in LHn.
j) Take the average of pixels of the two bands i.e. LLa and
LLb, and put in LLn.
i) Thus we will get HHn,HLn,LHn and LLn as new
coefficients.
j) Take Inverse Discrete Wavelet Transform.
k) Obtain the fused Image and Display.
,

Column
wise LP
filtering
..

Row wise
down
sampling
(LL)

Figure 4. DWT Decomposition

V. RESULTS

The DWT decomposition uses a cascade of special low

pass and high-pass filters and a sub-sampling operation. The
outputs from 2D-DWT are fOUT images having size equal to
half the size of the original image. So from first input image
we will get HHa, HLa, LHa, LLa images and from second
input image we will get HHb HLb, LHb, LLb images. LH
means that low-pass filter is applied along x and followed by
high pass filter along y. The LL image contains the
approximation coefficients. LH image contains the horizontal
detail coefficients. HL image contains the vertical detail
coefficients, HH contains the diagonal detail coefficients. The
wavelet transform can be performed for multiple levels. The
next level of decomposition is performed using only the LL
image. The result is fOUT sub-images each of size equal to half
the LL image size.

Following images are results of fusion process with PCA,

DCT & DWT fusion techniques,

Figure 5(a). First Input Image

Figure 5(b). Second Input Image

2013

International Conference on Signal Processing, Image Processing and Pattern Recognition [ICSIPR]

Figure 5(c). Fused Image Using PCA

Figure 5(d). Fused Image Using DCT

Figure 5(e). Fused Image Using

Figure 5(t). Fused Image Using

DWT(Pixel Averaging)

DWT(Maximum Pixel Replacement)

Following Table demonstrates the various quality measures for different image fusion techniques.
TABLE I
Fusion Techniques & Their Quality Measures
Quality Measures

Fusion Methods

Peak Signal
to Noise
Ratio

Mean Square
Error

Normalized
Absolute
Error

Maximum
Difference

Average
Difference

Normalized
CrossCorrelation

Structural
Content

PCA

30.7729

54.4246

0.0463

67.9043

-0. I728

0.9989

0.9991

DCT

30.9663

52.0536

0.0441

73.6046

-0.1718

I.OO I 2

0.9945

Pixel
Averaging

78.9555

8.2704e-004

0.0460

0.2647

-6.737ge-004

0.9989

0.9991

Maximum Pixel
Replacement

75.0723

0.0020

0.0695

0.3539

-0.0114

I.O I 69

0.9602

DWT

CONCLUSION

From above table, it can be concluded that, PCA & DCT

based image fusion technique can be used for applications
which does not require high quality & precision. Whereas
DWT based fusion techniques provide us good quality fused
images than PCA & DCT based techniques.
REFERENCES
[I]

[2]

V.P.S. Naidu and J.R. Raol, "Pixel-level Image Fusion using wavelets
and Principal Component Analysis ," Defence Science Journal, VoI. 58,
No. 3, May 2008, pp. 338-352 02008,DESIDOC.
Nirosha Joshitha J, R. Medona Selin,"lmage Fusion using PCA in
Multifeature Based Palmprint Recognition,"lnternational Journal of
Soft Computing and Engineering (IJSCE) ISSN: 2231-2307, Volume-2,
Issue-2, May 20 I 2.

[3]

[4]

[5]

VPS Naidu,"Discrete Cosine Transform based Image Fusion

techniques," Journal of Communication, Navigation and Signal
Processing (January 2012) VoI. I, No. I, pp. 35-45.
Yong Yang,Dong Sun Park,Shuying Huang, and Nini Rao," Medical
Image Fusion via an EffectiveWavelet-Based Approach," Hindawi
Publishing Corporation EURASIP Journal on Advances in Signal
Processing
Volume 20 I 0, Article
ID 57934 I, 13 pages
doi:10.1155/2010/57934I.
M.A.Berbar,S.F.Gahe,N.A.IsmaiII,"lmage
Fusion
Using
Multi
Decomposition Levels of Discret Transform," 8 2003 The Institution
of Electrical Engineers. Printed and published by lEE, Michael Faraday
House, Six HiIIs Way, Stevenage, SG I 2AY.