56
P' = A(P)
P' = I>,+A
i
comparisons. A Pair of LANDSAT Thematic Mapper (TM) and
SPOT images are also evaluated (size256 X 256),in order to
test whether the merging method of this paper extend to actual
PAN and MS images.
Where, i is the number of decomposition levels and W i is the
responding wavelet plane. P r is the remaining approximation
part.
Because panchromatic band is rich in spatial information, and
structure spatial is mainly concentrated on wavelet planes. An
additive fusing algorithm can be deduced. It can be easy to
conclude that the wavelet planes are ranged from fine and
coarse, when the scale rises. And we need to extract finer detail
of PAN to Mulspectral bands, so a descent weight array A is
used to moderate different plane W. . (as shown in formula
12.)Finally, fusion result can be got by inverse ICA transform
(In formula 13).
/c; = X^ (+ /c,
i=l
C 2 = ^ J A i w i + IC 2 ( 12 )
j=i
c;=5>, + /c 3
¿=1
[«', G\ B'] = F' (ic;, IC[ ,/C 3 ') (13)
4. EXPERIMENTAL RESULTS AND COMPARISONS
4.1 Synthetic Datasets and Real MS-PAN Datasets
The main aim of this research is to determine the efficiency of
new algorithm based on ICA for merging images with a
particular resolution ratio. Due to the difficulties in obtaining
adequate imagery with particular ration, Yocky’s approach [9] is
employed to synthesize some MS-PAN datasets with particular
ratio. In this approach, a Landsat TM test image was available
in the three bands,i.e., B1 (green), B2 (red), and B3 (near
infrared). The image was used to synthesize a perfectly
a. original image b. synthesized P band
(512X512) (512X512)
Fig 2 Landsat-P synthesized from three Bl, B2, B3 as
P = (Bl+B2+B3)/3
overlapped panchromatic band at 20 m, which is shown in.
Then, the MS bands were decimated by two and used together
with the P at 20 m, to synthesize Bl, B2, and B3 back at 20 m.
there are several reasons to emphasis on the merging of
simulated Pand decimated multispectral bands. Firstly, all the
images are spatially coregistered, and any problem caused by
geometric corrections in fusing process can be avoided.
Secondly, we can get the true multispectral data for objective
4.2The Quality Analysis of the Fusion Image
We have adopted some quantification metrics to evaluate the
fusion quality, including entropy, mean, and standard
deviation, Average gradient. Among these metrics, entropy
explores the information changes, and an image has more
information when the entropy is bigger. And some other metrics,
such as mean, employed to evaluate the aberrance of the
spectral information. The mean calculates the degree of the
spectral information change. In our research, different
decomposition levels for Atrous wavelet have been tested.
The result for data set is show in Fig 3.
For visual analysis, we could find that our method can enhance
the image spatial resolution to a certain degree, which will
benefit those applications which are demanding strictly on the
details of an image, such as image interpretation, special
cartography, and photogrammetric survey, etc. With
decomposition level increasing, more panchromatic band
a. original MS bands b. 1 level
c.2 level d.3 level
Fig3 the fusion result with different decomposition
levels for synthetic data size(512 X 512)
information is injected into three multispectral bands and gray
levels of images become sharper. But on the other hand, the
spectral distortion became worse. It means that it may be have a
tradeoff between the spectral and spatial quality. So when we
must consider about the decomposition parameter by combining
with the post-processing purpose. We have also adopted some
quantification metrics to evaluate the fusion quality. The
statistical data is shown in table 1. It can be included from tables:
the information in both of the two datasets is increased in the
case of the injection of the information .Because the high
frequency information in the multispectral bands is substituted
selectively by the corresponding parts in the panchromatic band.
So we proposed an useful fusing algorithm.
Metrics
Band
Original
spectral
1 2
3
R
bands
149.73
149.99 149.88
145.93
Mean
G
138.87
139.01 139.12
139.23
B
116.28
115.92 115.75
118.49
Standard G
deviation g
R
Average G
gradient g
R
Entropy G
B
Tabled statis
5.COr
A new multispectr
provided by comb
The experiment re
spatial informatioi
spectral distortion
future, our work
fusion rule for infc
REFERENCE
[1] Choi, M. , 20
Approach to Imag
Transactions on G
[2] J. C and H. J. ,
adaptive algorithm
Processing, vol. 2 l
[3] J. C and H. J. .
problems statemen
[4] Zhang, ]L. Hui
Shape Index C
Classification of
Imagery, IEEE ’
Sensing vol. 44, p.
[5] Comon, P. 8, 1
concept?, Signal P
[6] Mi Chen, 200(
Image Data Fus
Analysis,Ph.D diss
[7] Hyvarinen, A.
http://www.cis.hut
[8] Hyvarinen, A
Independent Comp
p. 10.
[9] Yocky, D. , 1
image, merger
panchromatic data
Sensing, vol. 62, p.
