International Archives of the Photogrammetry, 
  
    
    
ES n P 2 
2¢ Fourier Texture Image 2d Symbiotic Texture Image 
Fig 2 Comparison of Texture Feature Image 
As figure 2 shown, the deviation feature image has high 
contrast between the residential area and the other image object 
while the Fourier texture image is simulate to the original image 
with low contrast between the residential area and the other 
image object area. The grey symbiotic matrix texture feature 
image has little effect to the original image for the grey 
symbiotic matrix texture feature has strong rule and Spatial 
distribution requirement while most of the original image pixels 
are mixed pixels, which affect feature spatial distribution rule. 
Form the above view effect of the texture images, the deviation 
texture feature analysis has better effect after feature 
transformation and the Fourier texture image has not obvious 
effect while the grey symbiotic matrix texture feature analysis 
has the worse effect. 
In order to prove the above view effect to the texture analysis 
result of different kind of texture feature images, the statistic 
data of the above images and the residential region image 
manually selected from the total images are calculated for the 
comparison of different texture results of contrast and some 
other statistic value variation, which can make the image 
segment easier. The statistic data such as the grey average, 
deviation, maximize and minimize gray value and entropy are 
calculated in different texture feature dimension, as Table | 
shown: 
Table 1 Comparison of the statistic data in the original image and different tex 
  
  
ation Imz En: 
Fourier Feature 
Image 
Symbiotic matrix | Total Area | 255 
Feature Image 
Table 1 also calculates the difference o 
as the difference of deviation, difference of average and 
difference of entropy between the total image region and the 
  
residential region image manually selected from the total image. 
These differences of the gray value statistic data show the 
degree of the difference between the residential area image and 
the total image and these differences are very important to the 
segment threshold value selection. And the image segmentation 
will be better if the difference of the gray value statistic is high. 
From the result of Table 1, the difference of the gray value 
statistic data between the total image region and the residential 
image region manually selected is very low (difference of 
deviation is 4.02, difference of average is 2.2 and difference of 
entropy is 0.026) and it means that it’s very difficult to extract 
residential area from the test image just using the grey 
Remote Sensing and Spatial Information Science 
   
Total Area 0 
     
  
histogram statistic data. After transferring the image to the 
deviation texture images, the difference of the gray value 
statistic data has been greatly increased (difference of deviation 
is 12.07, difference of average is 16.7 and difference of entropy 
is 0.75). Thus, for the deviation texture image, it’s easy to 
obtain the segmentation threshold for the residential area 
extraction just using the grey histogram statistic data. And the 
difference of the gray value statistic data to the Fourier texture 
image and grey symbiotic matrix texture feature image just 
increase a little and for the symbiotic matrix texture image, one 
of the differences becomes small. Thus, we can get the 
conclusion that the deviation texture analysis is better for the 
image segmentation while the Fourier texture transformation 
and the grey symbiotic matrix texture feature analysis have not 
obvious effect, which is coincide to the view effect analysis to 
above texture images. 
2.3 Self-adaptive segmentation threshold Using Gauss Blur 
The residential areas pixels in the original images are consist of 
mixed pixels with both light pixels and dark pixels. And the 
dark pixel in the residential area is simulated to the background 
pixel, thus it's very difficult to obtain the suitable threshold 
value directly using the original image to classify the residential 
area and the background. From the above texture analysis to the 
residential area images, even the deviation texture image 
increases the difference between the residential area pixels and 
the total image pixels, then some background object will be 
included in the residential area. And the interior composing of 
the residential area is very complex and the texture character 
changes are huge which make it very difficult to get the suitable 
segmentation threshold value and extract the residential area 
through one time region growing (Vairy M, 1995, Chang YL, 
1994). And the mixed pixel make the region grow difficult 
using the simple seed. This is the main error source for the 
residential area extraction. 
Considering the extraction of residential area focuses on the 
board, the details of the residential area are not very important 
for the extraction result while the inside details with large 
ture feature images 
  
  
statistic difference have great effect to the board extraction 
result of the residential area. Thus some methods are proposed 
to process the inside details in the residential area to ignore or 
blur them. And Gauss blur is a good solution. 
Figure 3 is the gauss blur results of the original image and 
deviation image. 
As Figure 3 shown, the gauss blur make the deviation texture 
image consistency grey distribution inside the residential area 
and high contrast to the background. Thus the self-adaptive 
threshold can be easily obtained and the extraction can be well 
done. And the statistic data of these images and the residential 
image region manually selected from the total images are also 
calculated as Table 2 shown. 
These differences of the gray value statistic data in the Table 2 
the show the degree of the difference between the residential 
1123 
5, Vol XXXV, Part B7. Istanbul 2004