3) 
er 
Liang-Chien Chen 
2.3 Edge Strength Calculation 
The proposed scheme combines segment information and edge strength to determine the accurate shape of 
buildings. A number of edge detectors such as Sobel operator (Gonzalez &Woods, 1992), LoG (Marr & Hildrith, 
1980), Canny operator (Canny, 1986), and MEDT (Deok, et. al., 1995) have been reported. We select Sobel 
operator for computational simplicity. Combining the gradients convoluted by a horizontal and a vertical operators 
to each spectral band image. The one with the highest strength among the 3 band images will be considered as the 
final edge strength. 
2.4 Detection for Buildings and Corners 
The results of segmentation and edge detection are combined in this stage. Two steps are processed. The first is to 
detect the preliminary corner points. Then a refinement procedure follows. 
2.4.1 Preliminary Corner Detection 
For a given window with n xn pixels, it moves along the border of a region segmented from $2.2 (fig. 3.a). Then 
we calculate the angl 0 between C and the outmost points E, and E, (fig. 3.b). If 0 is in a reasonable range, the 
point is assumed to be a comer candidate. In general, several points would be retained around a real comer (fig. 
3.c). Thus, we take average location for the preliminary corner points (fig.3.d). 
  
  
  
  
  
( a) ts in BR | ? ( b) 
  
  
  
  
  
  
  
  
   
(d) 
Fig.3. An Illustration of Preliminary Corner Detection 
(c) 
2.4.2 Detection for Precision Buildings and Corners 
We search for the edge line with the highest strength in the buffer zones for the possible lines connected by 
preliminary corners. Starting from the preliminary corners C, and C,, a window is selected (fig.4.a). Then the edge 
strengths of all the line combinations of the points in window C1 and window C2 are computed. The two linking 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 171