Kubik Kurt 
  
such as houses, and trees. When houses and trees exist in the images, the disparity values of these areas are locally larger, often 
with discontinuities occurring in the disparity values at the edges of the features. Hence, edge extraction methods, in which 
relatively distinct changes in grey level properties between two regions in an image are located by changes in local 
derivatives, are used to define the discontinuities in the disparities values by treating the disparity map as an image. Common 
methods used to calculate these derivatives are the gradient and Laplacian operators (Gonzalez & Woods 1992). The Sobel 
gradient operator has the advantage of providing both a differencing and a smoothing effect. Since the derivatives enhance 
noise, the smoothing effects are a particularly attractive feature of this operator. 
4 PROCESSING OF SIGLE IMAGE 
4.1 Single Image Processing for House Extraction 
The left image is processed to separate house and tree areas. Figure 3 illustrates the implementation steps. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Single tep 1 Single y Sobel edge detection Step2 
Image | Region Number of regions L_ Image 
growing |} Single 
: Image . 
Li > va —| Dilation || Erosio 1 
: ilter 
Region to — Tokenshape v | 
token | 
M Analyze histogram of 
Tokenfiter [23 the directions of Ni 
Tokenshape | w Tokenfilter 
  
  
  
  
  
  
  
  
  
| | 
Y House area 
Figure 3 A single image processing procedure of Part 2 in Figure 1 
4.1.4 Dynamic region growing 
A dynamic region growing technique is used to determine homogeneous areas in the images in which the intensities of the 
pixel values are within a given threshold value. The threshold is modified dynamically according to the mean and standard 
deviation of the pixels in the region while it is being grown (KB Vision 1996). The adaptive threshold will never be larger 
than the pre-defined threshold (7), but may be smaller. It is th = (1-min(0.8, standard-deviation/mean))*7. The first region 
is chosen at the lower left corner of the image and processed until that region can no longer be grown. The next region 
starts at a pixel that has not been incorporated into the previous region. This process continues until all pixels have been 
grouped into separate regions which represent homogeneous areas in the input image. 
4.1.2 Analysing region parameters 
The number of regions in the image are calculated in the Figure 3. Regions are represented by tokens which also 
describe the features of that region (KBVision 1996). The task “Region to token” implements the transformation from 
region to token. “Tokenshape” is used to calculate a series of feature values for the regions, as follows: 
——) 4)log_h_to_w=log,, ( height 
2height + 2width width 
"Tokenfilter" in Figure 3 is used to filter out extracted regions which are not houses, based on the five feature values for 
every region. For each image, the minimum sizes of houses, perimeter and pixel count can be defined. Intensity mean is 
based on a special case and helps to extract the houses which have a bright roof. However, houses with dark roofs will 
have similar grey values as the ground cover, so it is difficult to locate them. Step 2 in Figure 3 can be used to recognize the 
area of the dark roof house as described below. 
perimeter 
  
l)perimeter, 2)intensity mean, 3)br to perimeter- ( ) S)pixel count. 
4.1.3 Analysing the histogram of orientations of edges 
  
522 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.