ISPRS IGU CIG Table of contents Authors index Search Exit

SIPT UCI ACSG Table des matières Index des auteurs Recherches Sortir

IMAGE FUSION USING WAVELET TRANSFORM

Zhu Shu-long a, b

a
Department of Remote Sensing Information Engineering, Zhengzhou Institute of Surveying & Mapping, Zhengzhou, China,
Postcode: 450052 -zsl1964@public2.zz.ha.cn
b
LIESMARS of Wuhan University, Wuhan, China, Postcode: 430079

Commission IV, Working Group IV/7

KEY WORDS: Multi-source images, Fusion, Wavelet transform, Edge processing, Wavelet based fusion, Fused Images

ABSTRACT:

Extracting more information from multi-source images is an attractive thing in remotely sensed image processing, which is recently
called the image fusion. In the paper, the image fusion algorithm based on wavelet transform is proposed to improve the geometric
resolution of the images, in which two images to be processed are firstly decomposed into sub-images with different frequency, and
then the information fusion is performed using these images under the certain criterion, and finally these sub-images are
reconstructed into the result image with plentiful information.

1. THE INTRODUCTION plentiful information. Since the geometric resolution of the

image depends on the high-frequency information in it,
For the remotely-sensed images, some have good spectral therefore this image fusion algorithm can acquire good results.
information, and the others have high geometric resolution,
how to integrate the information of these two kinds of images
into one kinds of images is a very attractive thing in image 2. THE ALGORITHM OF WAVELET TRANSFORM
processing, which is also called the image fusion. For the
purpose of realization of this task, we often need some 2.1 Mallat Algorithm of Wavelet Transform
algorithms to fuse the information of these two kinds of images,
and however we find few such algorithms. In recent ten years, a
Suppose that {V j , j ∈ Z } is a multi-resolution Analysis in
new mathematical theory called "Wavelet theory" ( Charles K.
L 2 (R ) ，ϕ ( x ) is the scaling function of subspace V 0 ，W
Chui, 1992 ) has gradually been used in the fields of graphics j

and imagery, and been proved to be an effective tool to process is the orthogonal complement of V j with respect to V j + 1 ，
the signals in multi-scale spaces ( Thierry Ranchin et al. ). In i.e. V j +1 =V j + W j
， ψ ( x ) is wavelet function of
this paper , the image fusion algorithm based on wavelet
subspace W 0 . If a signal f ( x ) ∈ V j +1
，then it can be
transform is proposed to improve the geometric resolution of
the images, in which two images to be processed are firstly expressed as :

decomposed into sub-images with the same resolution at the

same levels and fifferent resolution among different levels, and f (x ) = ∑c n
n ϕ
j +1
j +1 , n
(1)

then the information fusion is performed using high-frequency

sub-images under the "gradient" criterion, and finally these
sub-images are reconstructed into the result image with

Symposium on Geospatial Theory, Processing and Applications,

Symposium sur la théorie, les traitements et les applications des données Géospatiales, Ottawa 2002
since V j +1 = V j + W j , then f ( x ) can also be expressed
~
as: Where h n = h − n , g~ n = g − n , when the signal is finite,
~
hn = h1−n , g~n = g1−n . The formula (6) are the

f (x ) = ∑c n
n
j
ϕ j, n + ∑d n
j
ψ j ,n
(2) reconstruction formula of signal. The formulas (4)、(5) and (6)

are called Mallat algorithm ( S. Mallat, 1989 ) .

Combining formula (1) with (2), we get ： j +1

For two-dimensional signal c , its decomposition formulas

are :
∑c
n
n
j +1
ϕ j +1, n
=
∑cn
n
j
ϕ j, n + ∑d n
j
ψ j ,n
(3)

1
c mj , n =
2
∑
k ,l ∈ Z
c kj ,+l 1 h k − 2 m h l − 2 n

since ϕ is orthogonal with respect to different j and k ， 1

∑
j ,k
d mj 1,n = c kj ,+l 1 h k − 2 m g l − 2 n
if two sides of formula (3) are multiplied by ϕ and then 2 k ,l ∈ Z
j ,k
1
integrated with respect to x , we can obtained : d mj 2,n =
2
∑
k ,l ∈ Z
c kj ,+l 1 g k − 2 m h l − 2 n (7)

1
c kj = ∑c n
j +1
ϕ j +1,n ,ϕ j ,k =
1
∑h n−2k c nj + 1 (4) d mj 3,n =
2
∑
k , l∈ Z
c kj ,+l 1 g k − 2 m g l − 2 n
n 2 n

Using the same method, we also have: and its reconstruction formula is ：

1
d kj =
2
∑ g n − 2 k c nj + 1 (5)
n
1 ~ ~ ~ ~ 
cmj+,n1 =  ∑ckj,l h2k−mh2l−n + ∑dkj,1l h2k−mg~2l−n + ∑dkj,2l g~2k−mh2l−n + ∑dkj,3l g~2k−mg~2l−n 
2k,l∈Z k,l∈Z k,l∈Z K,L∈z 
The formulas (4) and (5) are the decomposition formula of

signal, where c j
is an approximation of c j +1
, and d j
is (8)

j +1
the detailed part of c .

where c j
is an approximation of c j +1
, d j1
、d j2
and
If having the decomposed signals {c n
j
, n ∈ Z } and j3
are the detailed parts of c j +1
. Fig. 1 illustrates the
d
{d n
j
, n ∈ Z }，then two sides of formula (3) are multiplied Mallat algorithm of 2D images intuitively.
by ϕ j +1,k
, and then integrated with respect to x , we can

obtained :

j +1
forward transform cj d j1
c
backward transform
c nj + 1 = ∑ c nj ϕ d j2 d j3
k
j ,k ,ϕ j +1, n + ∑ k
d kj ψ j ,k ,ϕ j +1,n

1 1 Figure 1. The Mallat algorithm of 2D image

=
2
∑ n
c kj h n − 2 k +
2
d kj g n − 2 k

1  ~ 
=  ∑ c kj h 2 k − n + ∑d j
g~ 2 k − n 

2  k
k
 2.2 Edge Processing for Wavelet Transform

(6)
Since our image fusion algorithm need to use the forward and
the inverse algorithm of wavelet transform, in order to
reconstruct the signals precisely, the image edges must be
processed under the process of wavelet transform, otherwise
some information will be lost in this operation.
forward wavelet transform, which have the same resolution at
the same levels and fifferent resolution among different levels;
and (3) information fusion is performed based on the
high-frequency subimages of decomposed images; and finally
the result image is obtained using inverse wavelet transform.

Assuming that the low-pass filter and the corresponding Let A( x, y ) and B ( x, y ) be the images to be fused，the

high-pass filter are four coefficient Daubechies' ( I. Daubechies, decomposed low-frequency sub-images of A( x, y ) and
1988), i.e. H = (h0 , h1 , h2 , h3 ) and
B ( x, y ) be respectively lAJ ( x, y ) and lB J ( x, y )（ J
，the decomposed high-frequency
is the parameter of resolution）
G = ( g − 2 , g −1 , g 0 , g 1 ) respectively, the samples of the
sub-images of A( x, y ) and B ( x, y ) be
original discrete signal are C 0 = (c00 , c10 ,...,c M0 −1 ) , then
respectively hA ( x, y ) and hB j ( x, y ) （ j is the
k k
j
the image edges are processed according the following precise
parameter of resolution and j = 1,2, L , J . for every j ,
formulas ( Zhu Shu-long and Geng zhe-Xun, 1999 ) in our k = 1,2,3 ） ， the gradient image generated from
experiments: hA kj ( x, y ) and hB kj ( x, y ) be respectively
GAkj ( x, y ) and GB kj ( x, y ) ， then the fused
For left image edges, we have：
high-frequency sub-images F j ( x , y ) are：
k

if GA j ( x, y ) > GB j ( x , y )
k k

c −01 = c −02 = 0
 F jk ( x, y ) = hA kj ( x, y ) ,
 1 h2 1
if GA j ( x, y ) < GB j ( x , y )
k k
c −1 = h d 0

F jk ( x, y ) = hB kj ( x, y ) ,
1

(9) and the fused low-frequency sub-image FJ ( x, y ) are：

FJ ( x, y ) = k1 ⋅ lAJ ( x, y ) + k 2 ⋅lB J ( x, y )
For right image edges, we have：
where k1 and k 2 are given parameters, if image B is

fused into image A , then k1 > k 2 . If the decomposed

c M
0 1
high-frequency sub-images hA j ( x, y ) of A( x, y ) is
= 0, dM /2 = 0 k

 0 h2 0 h3 2
replaced by be respectively F j ( x, y ) , the decomposed
k
c M +1 = − h c M −1 + h c M − 2

lAJ ( x, y ) of A( x, y ) is
0 0
low-frequency sub-image

replaced by be respectively FJ ( x, y ) , and FJ ( x, y ) and

(10)

F jk ( x, y ) are used to reconstruct the image A′( x, y )

If the image edges are processed according (9) and (10), then
using the inverse wavelet transform, then A′( x, y ) not only
the original image can be restored after wavelet forward
transform and then wavelet backward transform. has the low-frequency information of A( x, y ) and

B( x, y ) , but also has the high-frequency information of

A( x, y ) and B( x, y ) at the same time. This algorithm
3. IMAGE FUSION ALGORITHM BASED WAVELET shows that the high-frequency information fusion between two
TRANSFORM
images is completed under the "gradient" criterion.

The main idea of our algorithm is that: (1) the two images to be
For multi-spectral images, we can use the same method
processed are resampled to the one with the same size; and (2)
proposed above to process every band images respectively. For
they are respectively decomposed into the subimages using
color images, we can convert the color images from RGB ( i.e.
Red, Green, Blue ) space to HIS ( i.e. Hue, Intensity,
Saturation ) space, and then only process the intensity 5. THE EXPERIMENTS
component of HIS space using the same method proposed
above, and finally acquired the fused the images. In our experiments, two images are used to verify our algorithm,
the first is the image composed of SPOT multi-spectral images
with the resolution of 20 m ( as shown in Fig.2 ), the second is
4. THE ASSESSMENT FOR THE FUSED IMAGES SPOT panchromatic image with the resolution 10 m ( as shown
in Fig.3 ). In order to fully utilize the spectral information of
The index "Average gradient" is applied to measure the detailed the former and geometric information of the latter, the image
information in the images. The formula of average gradient "g" fusion algorithm proposed above is used to introduce the details
can be denoted as follows: of the latter into the former so that the result image ( as shown
in Fig.4 ) has not only good spectral information but also good
geometric information. The table 1 illustrates the average
 ∂f (x, y)   ∂f (x, y) 
2 2
M−1N−1
1
g= ∑∑   +
(M −1) ⋅ (N −1) x=1 y=1  ∂x   ∂y 
 (11) gradient "g" in these three images quantitatively, in which the
average gradient of the result image is obviously greater than
that of the original image 1.

Where f ( x, y ) is image function, M and N are the

Now this algorithm has been a practicable function in our
length and width of the image respectively. Generally, the
Image Processing System ( IPS ). The further research is to
larger g is, the clearer the image is. For color images, average adapt this algorithm to the large-data images.

gradient "g" can be calculated from the "intensity" images.

Fig. 2. The original image 1 with Fig. 3. The original image 2 with Fig. 4. The result image generated by
the resolution 20 m the resolution 10 m our algorithm

Table 1 the image and its average gradient

REFERENCES:
Image Name The average gradient
the original image 1 with the Thierry Ranchin et al., Efficient data fusion using wavelet
4.78
resolution 20 m transform: the case of SPOT satellite images, in
the original image 2 with the Mathematical imaging: Wavelet Application in Signal and
12.33
resolution 10 m
Image Processing, Andrew F. Laine, Editor, Proceeding of
the result image generated by
11.16 SPIE 2034, pp171-179.
our algorithm
Charles K. Chui, 1992, An Introduction to Wavelets,
Academic Press, Inc., America.

S. Mallat, 1989, A theory for multiresolution signal

decomposition: The wavelet representation, IEEE
Transaction on PAMI, Vol.11, No.7, pp.674-693.

I. Daubechies, 1988, Orthonormal bases of compactly

supported wavelets, Communication on Pure and Applied
Mathematics, Vol. XLI, pp.909-996.

Zhu Shu-long and Geng zhe-Xun, 1999, The Wavelets and Its
Applications to Graphics and Imagery, The PLA Press, P.R.C.