International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
processing procedure based on Retinex-Wavelet (RW) ima 
representation model can be illustrated as figure 5: 
> 
ge 
e 
As figure 5 shown, one of the key technique of RW model is to 
estimate the constant part image /(x,v). D. J. Jobson, 2002, 
propose a statistic model based on grey distribution through the 
research of the statistics of visual representation to estimate 
I(x,y). And iteration algorithms using a constant as an initial 
Section 3.3 and Section3.3. 
  
G(x,y) ce Estimate 
  
    
    
  
  
  
    
The Constant 
  
  
  
Image Processing 
To r(x,y) in 
  
  
  
  
  
  
g Gy) — 
mm) J fx) 
  
  
  
value to /(x.y) with probability model constraint to the r(x.y) Wavelet Wavelet Domain 
are proposed in many color enhancement research. To different Transform 
image quality processing, the estimation of /(x,v) will be 
difference. This will be discussed in the following Section 3.2, Figure 5 RW Imaging Model 
2.2 Image Dodging in Retinex Wavelet Domain 
8 GJ) 
Gauss 
Smooth 
  
r(x,y) 
  
  
   
    
  
     
foy) 
  
  
  
High pass 
Filtering 
  
  
  
Wavelet! 
Transform; 
——1- 
  
  
  
| Low pass 
d Ll Inverse 
Filtering 
Wavelet 
  
  
  
  
  
Figure 6 Dodging Processing Based on RW Model 
In the aero-photograph, the distribution of intensity in the focus 
plane is not well-proportioned and gradually minish from centre 
to side when light get through the field lens of the huge camera. 
At the same time, building shadow and the factory fog affect 
the local contrast in the image. Anymore, the discord of 
illumination of the coteua in the side of exposed to the sun and 
back to the sun makes different contrast and grey distribution in 
the image (As figure 8 shown). All these in the image are 
showed as light decreasing around, asymmetry of intensity and 
huc, declination of contrast. To improve quality of these images 
is called image dodging and it can be classified into image 
enhancement. 
For image dodging, the constant image /(x,y) can be taken as a 
blur Mask in traditional photograph copy. Thus we can taken 
the following assumptions to calculate reflectance image r(x. v): 
(1) The illumination is its spatial smoothness. 
(2) The reflectance image r = g- [ can be assumed to have a 
high prior probability. One of the simplest prior functions used 
for natural images assigns high probability to spatially smooth 
Images. 
(3) We can assume that the illumination continues smoothly 
as a constant beyond the image boundaries and we can take a 
Gauss filtering to obtain an initial value. 
hus. we can calculate reflectance image r(x,Y): 
rx, y)=g(x, y)-g(x,y)e mx.,y,o) (7) 
m(x.y,0)=exp((x" + y*)/o") (8) 
Where ois a constant which controls the extent of mask M(X.y) 
ind it should restrict r(x,v) to a high prior probability , and e 
represents spatial convolution. An high pass to LL and low pass 
filtering HH to reflectance image r(xv) in wavelet 
domain(frequency domain) can achieve the aim of dodging to 
remove center/surround light and contrast distribution degrade. 
Figure 6 gives the dodging processing flow. 
\ detail image dodging processing sample based on RW model 
analvzed as figure 6 shown. The sample image is an area 
degradation image with  center/surround  un-even light 
listribution, which is caused by imaging sensor such as camera 
lens’ distortion as figure 7(a) shown. 
In the sample dodging processing, the illumination is obtained 
using an r=3 gauss blur. Figure 7(b) is the final dodging result 
based on RW model to the sample image. It has a better view 
effect with even light and contrast distribution. 
Image illumination problems are common existed in all kinds of 
imaging system. Figure 8 is the dodging result of some Mars 
images from the recent landed “Courage” cameras (Left 
navigation camera non-linear full frame EDR acquired on Sol 9 
of Spirit's mission to Gusev Crater at approximately 13:18:01 
Mars local solar time). 
2.3 Image Restoration in Retinex Wavelet Domain 
The purpose of image restoration is to compensate for defects, 
which degrade an image such as edge blur. This is different to 
image enhancement, which is mainly to remove additional 
imaging defects such as the over exposure to obtain the best 
view effects. The image decomposition in equation (4) for 
image enhancement is always understood with illumination 
model while for image restoration, the imaging focus model is 
more suitable. Thus, the estimation is different in Section 3.2. 
Bit slicing can segment image in 8-fixed focal panel image to 
differ noise, as equation (10). And bit synthesis can combine 
any bit image together to obtain variance focus panel image. In 
this paper, the bit slicing and bit operation is adopted to 
estimate /(x,y), which is taken as imagery on focal length panel. 
With transcendental cognition of satellite on-orbit focus, we can 
use the following model to judge whether the synthetic image 
with different bit pancl images is the imagery on focus as 
equation (9): 
o 
= x] (9) 
8, 
ge, =} va, 1=01,47 (10) 
8, =2&it=0,1.7 (11) 
964 
Sn E iPnpsamm ——R— 
Intern 
uer 
Ce 
fil 
pre 
Fi; 
co 
tak 
im 
sh 
du 
RY 
Th 
the 
im 
cle 
Re 
pre 
T: 
Meat 
im 
wi 
fig 
Foi 
im 
tha 
dif 
wit 
clo 
13( 
In 
RV