Bed Ny i) 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
The chosen rotation vector F (r) is described as the result of 
the gray value together with the radius. Actually, the target 
image and the reference image are not obtained under the same 
conditions. Subject to weather, light, the sensor and impact of 
other factors, the target window and the reference window 
usually have gray change, noise, light and contrast difference. 
So, for the vector F(r), which has the rotation invariance, is still 
sensitive to these changes, as showed in Figure 2: 
  
Figure 2 a:F(r) = 34345, b:F(r) = 12432 
As we can see, if gray-scale between left image and right 
image has obvious difference, F(r) is definitely different. 
To reduce the gray-scale impact on the rotation vector, Wallis 
filler is applied. Wallis filter is used to enhance the image 
contrast and reduce the noise, and especially it can be used to 
enhance the textures in the images which are very weak in the 
original image (Zhang Li, Zhang 1999). Wallis filter can adjust 
the color difference between images. It adjusts the gray mean 
and variance to a given mean and variance. The formula of 
Wallis filtering is as follows. 
g(x,y) =m, *tv.(g.(x, y) " m, )/ v, 
The m, is the aim image's mean, v, is aim image's variance, 
m, is current image's mean, v, is current image's variance, 
g(x,y) is the gray value of current image, g(x,y) is the gray 
value after Wallis filtering operation. The mean value m and 
variance value v of the image is computed with formulas below. 
; JO g(x,y)!=7 
f(x. y) 9: 
f(x. 3) li a: se) 
=F f(x) 
x=0 y=0 
Y^ *j 
mz 
  
à > 
sa -nmy *, 
0 = 
w*h 
   
Wallis filtering result is showed in Figure 3: 
be udis 
vas 
    
    
Figure 3 a:F(r) = 211890, b:F(r) = 212432 
After the filtering processing, we can obtain the new value of 
F(r).It can be used in the relate-computing. 
In finding the correspondence in the target image, the F (r) is 
calculated by the same method for every image feature. F (r) can 
be computed for all the image features firstly, and then do the 
  
correspondence search. In this way, multi-to-multi search 
process can be simplified to a linear search process. As showed 
in Figure 4, it reflects the reference rotation vector and the 
target image rotation vector's distribution. 
           
dues 
Reference 
d. 
HON 
4 
F(r) 
Target Wa value 
Figure 4 Target Feather and Target Image with the F(r) curve 
Target Image 
  
  
  
Broken line X coordinate 
The feature in the target image has more than one strongly 
related position, all these positions should be recorded as the 
correspondence and considered together. After completing the 
collection of all potential corresponding points, Generalized 
Hough Transform (GHT) is applied to eliminate the gross errors. 
All points’ right image coordinate x and y will subtract the left 
image coordinate, and the highest group of dx and dy will 
denote the correspondences. The points over trinal root mean 
square (RMS) will be removed. 
3. EXPERIMENTS AND ANALYSIS 
To verify the validity of the matching algorithm, we selected 
three sets of data. The detailed experimental procedure is as 
follows: 
1) Firsty, Wallis filter is applied. 
2) Then 10*10 Harris feature points are extracted from the 
left image, each feature point contains a window with size 
331; 
3) Calculates the rotation vector of these feature points, 
taking 15 as the radius of rotation vector. 
4) Then, calculate the rotation vector of all pixels in the right 
image with the same method after edge extraction. 
5) According to the similarity measure function, find 
potential feature points’ correspondences. Thus, we get the 
same points both appcaring in left and right images. 
6) Analyzing the statistics of the shift which based on the 
Generalized Hough Transform (GHT), the peak of votes is used 
to establish the match point. 
Following are the test data and some process results. The first 
image set has a 90 ° rotation angle, the second set is 30 °, the 
third one is 45 °. 
Test data 1 
p Area covered by 
trees. 
Size: 2048*2048 
  
119