Shriniwas T. Budhewar / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol.

5 (3) , 2014, 3703-3707

Wavelet and Curvelet Transform based Image

Fusion Algorithm
Shriniwas T. Budhewar
Dept. of Electronics and Telecomm. Engineering,
Government college of Engineering,
Kathora naka, Amravati,
Maharashtra, India

Abstract: The prerequisite of more unblemished and realistic gives better space-frequency characteristics. Results have
Images has contributed the significant development in the revealed the improvement in SNR and spectral resolution
Image processing field. An image should encompass every fine over the spatial domain technique. Wavelet function
aspect of scene, but practically it is impossible to do so due to operates on image linearly, but the image is rarely a linear
optical limitations of Image acquiring devices. The solution to
this kind of problem is provided by Image fusion Technique.
structure [7] [3]
Image fusion can be regarded as the process of merging of two The problem of tracing of curve has led to development
or more images to get a synthetic image. Among the several of anisotropic, nonlinear transform viz. Curvelet. E. J.
techniques of image fusion, wavelet transform based Candes and D. L. Donoho proposed Curvelet Transform
algorithm is often practiced. The need for sparse theory in 2000. It is highly directional transform and is able
representation and anisotropic way of image decomposition of to trace the curve with minimum coefficient. Curvelet is
image for detection of curvature entity, has led the concept of characterized by three parameters scale, orientation, and
curvelet transform. In this paper, implementation of image translation. The disadvantage of this transform is missing
fusion algorithm using wavelet and curvelet transform has texture and contrast information. Nevertheless applications
been described and practical results are compared with
several algorithms.
of curvelet have been up surging widely. Wavelet
transform is challenged with representation of edges while
Keywords: Spatial and Transform domain techniques, curvelet can’t represent texture information clearly.
Wavelet and Curvelet transform, Image fusion, performance Previous literature shows the implementation of both
metrics. wavelet and curvelet transform. Theoretically continuous
transforms are defined but practically discrete algorithms
1. INTRODUCTION are implemented by using Matlab viz. discrete wavelet and
Image acquisition is usually accomplished by a device curvelet transform. [1]
focusing on particular portion of scene leaving other Including first chapter this paper contains five sections;
portion blurred. This happens due to practical limitations of second section deals with wavelet transform whereas
focal depth of devices. For the particular scene, two or curvelet transform has been discussed in third section.
more than two images, at different focus plane can be Image fusion and results are included in successive
acquired. These images are then merged to form a synthetic sections. Paper is concluded at the end.
image which is known as ‘multifocus image fusion’.
Several methods or algorithms of image fusion have been 2. REVIEW OF WAVELET TRANSFORM
advanced which can be broadly categorized as:[2] As discussed earlier, the wavelet transform best
A. Spatial domain based describes the space-frequency relationship of image.
B. Transform domain based Wavelet transform is composed two sub functions viz.
In spatial domain methods, spatial variables i.e. intensity scaling and translation (also called wavelet). In 1-D,
of pixels etc. are considered and manipulation is done with continuous wavelet transform can be defined as:[11]
them. The intensity values of pixels are calculated with τ
ψψ (τ, s) = x(t). ψ( ) dt (1)
√
some mathematical calculations e.g. In maximum select
Where ‘s’ is scaling parameter and ‘τ’ is translational
method intensity values of two or more images are
parameter. ‘ѱ(t)’ is basis function called ‘wavelet function’.
compared pertaining to their intensity values and whichever
The same equation holds true for image as image is two
is maximum that is used to assign the intensity of same
dimensional signal. For better application and ease,
pixel in the fused image. The point of caution is that the
discretization of wavelet is preferred so-called ‘Discrete
size of input images must be same. This method is simple
wavelet transform (DWT)’. It can be practically
but does not provide adequate solution as it is just
implemented by using Matlab.
approximation[6]. Second method one can implement by
The image can be treated as piece-wise constant
using selecting mean values of pixels. These spatial domain
functions space ‘Vj’ will include all piecewise constant
methods have many drawbacks such as reduced contrast,
functions defined on the interval [0; 1) with constant pieces
unsolicited artefacts. Another class of fusion techniques
over each of the ‘2j’ equal subintervals. We can define the
seeks out the involvement of mathematical transforms such
basis function for ‘Vj’ which is termed as ‘scaling
has wavelet, curvelet. Wavelet transform has been
function’. Mathematically, scaling function is given by:[5]
commonly practiced in the field of image processing as it

www.ijcsit.com 3703
 Shriniwas T. Budhewar / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 5 (3) , 2014, 3703-3707

∅ ( ) = ∅(2 ) 0 2 1 (2)
Where

1 0 1
∅( ) =
0

We can now define a new vector space ‘Wj’as the

orthogonal complement of ‘Vj’ in ‘Vj+1’. That is, ‘Wj’ is the
space of all functions in ‘Vj+1’ that are orthogonal to all
functions in ‘Vj’ under the chosen inner product.
Informally, ‘Wj’ contains the details in ‘Vj+1’ that cannot be
represented in’Vj’. A collection of linearly independent Figure 2 frequency components after decomposition
functions ѱjk(x) spanning ‘Wj’ are called wavelets.
General scaling function of DWT is given by:[9] In above pattern of image, texture information is
ѱ ( )= ѱ ( ) (3) present in LL function while edge or high frequency
information is present in rest of the bands.
For Image reconstruction inverse method of the
There are different functions which can be treated as subband coding is followed. The filter bank structure is
‘Mother wavelet’ function, Mother wavelets are orthogonal shown below in fig.3
functions which satisfies ‘Admissibility’ condition. It is
given by
1-D multiresolution wavelet decomposition can be
forthrightly stretched to two dimensions by introducing HH 2 H0
separable 2-D scaling and wavelet functions as the tensor
2 H0
products of their 1-D complements. They are given by:
HL 2 H1
ΦLL=φ(x).φ(y) ѱLH=ѱ(y).φ(x) synthesized
ѰHL=ѱ(x).φ(y) ѱHH=ѱ(x).ѱ(y) image
LH 2 H0
For implementation of DWT, ‘Subband coding ’ is 2 H1
preferred and it is straight-forwardly extended for 2D-
LL 2 H1
DWT. Subband coding uses low pass filter (H1) and high
pass filter (H0) firstly, subband coding is applied on rows
then on columns as shown in fig.1. After first level of Figure 3 synthesis filter bank for IDWT
decomposition, there will be four frequency bands, namely
Low-Low (LL), Low-High (LH), High-Low (HL) and In figure 3, upsampling or interpolation is carried
High-High (HH). The next level decomposition is just out and a pair of low pass (H0) and high pass (H1) filters are
applied to the LL band of the current decomposition stage, used which are same as that of in analysis.
which forms a recursive decomposition procedure. Thus, Matlab has provided the instruction for transform
N-level decomposition will finally have 3N+1 different and inverse transform. The low pass filtering coefficients
frequency bands, which include 3N high frequency bands are called ‘approximation’ and high pass filtering
and just one LL frequency band. After the first level the coefficients are called ‘detail’.
nature of image is shown in fig.2.[9]
REVIEW OF CURVELET TRANSFORM

E. J. Candes and D. L. Donoho put forth the theory of

Curvelet Transform theory in 2000. While dealing with the
curve, wavelet transform becomes inefficient as it is linear
function and decompose image in a isotropic manner. In
this type of decomposition, more wavelet coefficients and
more levels of decompositions are needed. Moreover, it
requires large amount of time to get fully decomposed
image.[10][12]
In curvelet transform, the basis function is in the
form of curve i.e. length2~width. It is multiscale transform
which operates on image in anisotropic way. Which gives
best approximation with less curvelet coefficient in reduced
amount of time compared to wavelet transform. The way of
Figure 1 Filter bank structure of the DWT Analysis approach is shown in fig.4[10]

www.ijcsit.com 3704
 Shriniwas T. Budhewar / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 5 (3) , 2014, 3703-3707

3 IMAGE FUSION
The transform based methods follow a generic
method which is shown in fig 6

Figure 4 comparison of wavelet and curvelet transform

Figure 6 generic method of transform based fusion
Curvelet functions are characterized by scale,
orientation and translational parameters, values of which Transformation is applied on images to be fused to
are adaptably defined. The curvelet transform has four transform them into frequency or space-frequency domain.
stages:[10] The coefficients obtained during this process are used for
the fusion. Fusion rule is nothing but the criteria for the
I. Sub-band decomposition: selection or impositions of the coefficients of fused image,
In this step, the image is divided into individual e.g. maximum coefficient selection, minimum coefficient
subband frequencies. Mathematically, it is given selection etc. after fusion reverse transformation is applied
by: to get image back into spatial domain. The algorithm for
f α (P0f, ∆1f,∆2f, K) (4) the image fusion is given in fig.8. for the experimentation
where ‘f’ is image matrix or function, ‘P0’is low the raw images are derived from single image, by blurring
pass filter, ∆1,∆2,................ are band pass filters. either left or right side of image. Left blurred image can be
II. Smooth partitioning: treated as right focused image and vice versa. These two
A grid of dyadic squares is defined: images can be fused and the results can be validated by
Q( s ,k1 ,k 2 ) = [
2s
k1
][
, k21 +s 1 × k2
2s
(5)
]
, k22 +s 1 ∈ Q s comparing it with original image.
For the fusion, the selected image is shown in
Qs – all the dyadic squares of the grid. fig.7 (a). the right and left focused images which are
Let w be a smooth windowing function with obtained by blurring the part of image are shown in fig (b)
‘main’ support of size 2-s×2-s. For each and (c) respectively. Note that in fig (b),larger, square
square,‘wQ’ is a displacement of w localized near alarm clock is clearly visible while in fig (c) smaller, round
Q. Multiplying Δs f with wQ (∀Q∈Qs) produces a clock is clearly visible. These two images form the set of
smooth dissection of the function into ‘squares’. images to be fused.
The windowing function w is a nonnegative
smooth function.
III. Renormalization:
Renormalization is centering each dyadic square
to the unit square [0,1]×[0,1]. For each Q, the
operator TQ is defined as:
( ) (
TQ f ( x1 , x2 ) = 2 s f 2 s x1 − k1 , 2 s x2 − k 2
(6)
) (a) (b) (c)
Figure 7 original and partially focused images
IV. The Ridgelet Transform:
Ridgelet are an orthonormal set {ρλ} for L2(ℜ2) which
divides the frequency domain to dyadic coronae |ξ|∈[2s,
2s+1]. In the angular direction, samples the s-th corona at
least 2s times. It is shown in fig.5

-s
2

Figure 5 Ridgelet function

Curvelet transform can be implemented by the

frequency wrapping algorithm given by E. J. Candes and D.
Figure 8 Image fusion algorithm
L. Donoho. It is available at http://www.curvelet.org.[1]

www.ijcsit.com 3705
 Shriniwas T. Budhewar / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 5 (3) , 2014, 3703-3707

4 RESULTS
For above shown images, fusion is carried out by
using both the transforms and the resulting images are
compared. For validation of the results statistical
parameters are taken into account. Five performance
metrics are calculated for the original, and fused images
which are shown in fig. 9, viz. PSNR, standard deviation,
cross correlation, Entropy.[4][8]

Curvelet fused image Wavelet fused image

Figure 9 Transform based fusion

Five performance metrics are detailed below:

I. Peak signal to noise ratio (PSNR):
Peak signal to noise ratio is defined as ‘ratio between
the maximum possible power of a signal and the power
of corrupting noise that affects the fidelity of it’s
representation’.
 ( L − 1) 2   L −1 
PSNR = 10 log10   = 20 log10  
 MSE   RMSE  (7)

Where ‘L’ is maximum intensity level possible.

II. Standard deviation:

Standard deviation is a measure of contrast in an
image.larger is standard deviation, the higher is
contrast.
N −1 M −1

  ( A ( i, j ) − m )
2
L −1

(r − m ) p ( rk ) =
2 i =0 j =0
σ= σ = 2
k Figure 10 Performance metrics
k =0 NM (8)

III. Entropy: Measure of information present in the 5. CONCLUSION

image. Note that all the parameters except RMS are higher for
= ∑ ( ). log ( ) (9) transform based methods. Moreover from the graph it is
clear that transform domain methods prove to be efficient
over the spatial domain methods. Less value of RMS for
IV Root mean square error: the transform domain methods clearly indicates that the
Root mean squared error is the measure of fused image is void of artefacts. Higher value of PSNR
differences between the reference images and fused image. shows that image is less prone to noise compared to spatial
N −1 M −1
1 domain methods. Entropy is approximately same for all the
  ( A ( i, j ) − B ( i, j ) )
2
RMSE = MSE = methods. Cross correlation coefficient value approximates
NM i =0 j =0
(10) to 1, which represents the degree of similarity to that of
original image.
V Cross correlation: When the standard deviation value of images of
Cross correlation is used to find out the curvelet and wavelet are compared, it is higher for the
similarities between fused image and registered image. wavelet transform. These indicates that wavelet transform
For the fused image the above parameters are is efficient in representing the contrast information. The
calculated and same is done with spatial domain techniques same can be confirmed by using visual inspection of the
shown in graph. Based on the values the graphs are drawn fused images.(fig.9). the edges are more sharp for curvelet
which are shown in fig. 10 based image than wavelet based image while contrast for
wavelet is better than curvelet based method. These proves

www.ijcsit.com 3706
 Shriniwas T. Budhewar / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 5 (3) , 2014, 3703-3707

that curvelet transform can represent the curves efficiently

than wavelet transform and wavelet has better capability to
represent texture, contrast information than curvelet.
From above discussion it is clear that transform
domain methods proved to be efficient over spatial domain
methods. In spite of certain disadvantages for transform
based methods as explained earlier, they are popularly used
in the field of image processing. One can use two or more
transform to counter correct the limitations of
transform[12].

REFERENCES
[1] CurveLab, [Online]. Available: http://www.curvelet.org.
[2] K. Shivsubramani, K. P. Soman, "Implementation and Comparative
Study of Image Fusion Algorithms," International Journal of
Computer Applications, vol. 9, no. 2, p. (0975 – 8887), Nov 2010.
[3] Mathworks corporation, "Matlab and Simulink,"
[Online].Available:
http://www.mathworks.in/matlabcentral/answers/.
[4] M. Deshmukh and U. Bhosale, "Image Fusion and Image Quality
Assessment of Fused Images," International Journal of Image
Processing (IJIP), vol. 4, no. 5.
[5] A. Saha , G. Bhatnagar and J. W. Q. M., "Mutual spectral residual
approach for multifocus image fusion," Digital Signal Processing,
ELSEVIER, vol. 23, p. 1121–1135, March 2013.
[6] D. k. Sahu and M. P. Parsai, "Different Image Fusion Techniques –
A Critical Review," International Journal of Modern Engineering
Research, vol. 2, no. 5, pp. 4298-4301, sep-oct 2012.
[7] R. Sharma and K. Rani, "Study of Different Image fusion
Algorithm," International Journal of Emerging Technology and
Advanced Engineering, vol. 3, no. 5, 2013.
[8] M. A. P. A. Dr.S.S.Bedi, "Image Fusion Techniques and Quality
Assessment Parameters for Clinical Diagnosis: A Review,"
International Journal of Advanced Research in Computer and
Communication Engineering, vol. 2, no. 2, pp. 2319-5940, Feb
2013.
[9] R. C. Gonzalez and R. E. Woods, "Wavelets and Multiresolution
processing," in Digital Image Processing, New Jersey, Prentice
Hall, 2002, pp. 350-386.
[10] Jianwei Ma and . G. Plonka, "The Curvelet Transform : A review
of recent applications," in IEEE SIGNAL PROCESSING
MAGAZINE, 2010.
[11] R. Polikar, "The Wavelet Tutorial by Robi Polikar,"
2001.[Online].Available:
http://users.rowan.edu/~polikar/WAVELETS/WTpart1.html.
[12] B. Y. Shutao Li, "Multifocus image fusion by combining curvelet
and wavelet transform," Pattern Recognition Letters, ELSEVIER,
no. 28, p. 1295–1301, 2008. D. Hall and J. Llinas, "An
introduction to multisensor data fusion", Proceedings IEEE, Vol.
85(1), pp. 6-23, 1997

AUTHOR

Shriniwas Budhewar received the B.E. in

Electronics and Telecomm. Engineering from Pad.
Dr. D. Y. Patil Institute of Engineering and
Technology, Pimpri, Pune in 2011. Presently, he is
pursuing his M-Tech. in Electronic system and
Communication from Govt. college of Engineering,
Amravati. His area of interest includes Digital Image and signal
processing, mathematical transforms. Recently he is working on
the project ‘Curvelet and wavelet transform based image fusion.’

www.ijcsit.com 3707