ten 
ich 
cal 
10n 
bel 
nce 
the 
ard 
ger 
ion 
ion 
een 
om 
for 
1 is 
vill 
the 
Kubik Kurt 
  
  
After analysing their parameters, most extracted regions can be eliminated, but some regions will be described incorrectly 
as houses. In order to eliminate these false areas, an analysis is made of the histogram of the orientations of edges of the 
regions. 
Figure 4 illustrates the procedure for distinguishing houses from other objects. There are eight compass directions of region 
edge pixels in this small test image. Table 1 shows that for regularly shaped houses, the histogram contains a greater 
number of edges whose orientations are mutually orthogonal such as here at 0° and 90'. This does not occur for all 
images. However, because the directions of the edge pixels of the houses in this test image are significantly different from 
those of trees, which are obviously not square, this method can assist in differentiating between houses and trees. 
  
  
  
  
  
  
  
  
  
  
  
  
0o* 450 90 0* 1350 1800 * 
832 325 331 371 883 
115 73 103 201 80 
2250 2700* | 3150 Sumof * Sum of other 
192 217 179 2263 1067 
153 187 235 485 662 
  
  
  
  
| | 
| à o dia 
  
(1) (2) (3) (4) Table 1 The number of different orientation pixels(h:house t:tree) 
(1)original images (2) edge magnitude images (3) edge orientation images (4) histogram of the orientation images 
Figure 4 Analysis the difference between two regions 
4.1.4 Texture analysis 
A variance filter is determined from a texture algorithm capable of distinguishing uniform intensity areas in images. Although 
the dark roofs have similar intensity to the ground cover in the image, they have different textures. The variance filter, which 
provides a measure of local homogeneity of the intensities in an image, can also be regarded as a non-linear non-directional 
edge detector (Wilson 1997). The variance filter involves replacing a central pixel value with the variance of a specified set of 
pixel values surrounding it in a window on the image, which does not need to be square. The variance of such a set is given as 
follows: 
i-— Ys. v, -— Ye. cy e Sy (D 
r-l c-l r=1 c=1 n nu c 
  
Where n x n is the total number of pixels in the window. w is the window in the image, x, is the value of the pixel at row r 
and column c in the windows. x is the mean of pixel values in the window. 
4.1.5 The morphological functions 
Morphological transformations are powerful tools for extracting image components, that are useful for representing and 
describing region shapes. Dilation combines two image sets using vector addition of set elements, while erosion combines two 
sets by vector subtraction. 
        
(4) 
(6) 7) 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 523