ed DEM 
esolution simplifies 
le, can be detected 
1981), (Aviad and 
93), (Barzohar and 
88). In figure 10 
ponds to 25 cm) is 
dges due to the tex- 
e pixel corresponds 
lution more closely, 
iot suffice in many 
the segmentation. 
the raw shape of an 
fined resolution) is 
object and to distin- 
ts. 
segmentation: Seg- 
The results of this 
t refined resolution. 
rther refined resolu- 
traction of a road in 
re extracted as lines 
'dges are extracted. 
| of the road bound- 
. For final results a 
larks. 
  
twice refined 
in different resolu- 
in some case. This 
he gray value of the 
fore the road cannot 
road is defined only 
n. In this case the 
building original image 
      
anisotrope diffusion 
Figure 9: Part of an image with noise and some texture and the results of different filters for noise reduction 
texture of tiles on the roof 
  
  
low resolution 
Figure 10: Edge detection in high and low resolution 
interpretation process is very difficult because there is a very low 
hypotheses for a road because it could not be found in the initial 
resolution. 
  
More information on scale space and pyramids can be found in 
(Gauch and Pizer, 1993), (Lindeberg, 1991), (Lindeberg, 1993). 
5 OBJECT CLASSES 
As we have seen in figure 1 there are alot of different object classes 
inan aerial image. Ideally, one segmentation procedure should be 
used to extract primitives which are sufficient to recognize objects 
of all classes. Unfortunately, this is not the case. As we will see 
in the following subsection there exist specific procedures for a 
broader class of objects which can be processed in a similar way. 
But no procedure for all classes exists. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
5.1 Compact Artifical Objects 
Examples for this class are buildings, cars, trucks, and ships. All 
these objects are composed of more or less homogeneous areas 
with polyhydral borders. Therefore, models can be contructed us- 
ing descriptions of areas, lines, and points together with attributes 
(e.g., color or size) and relations between the primitives. The 
interpretation of objects is done by extracting similar primitives 
(area, edges, junctions) and matching these with the model after 
an optional grouping (Dolan and Weiss, 1989), (Lin et al. 1994), 
(Lu and Aggarwal, 1992), (Mohan and Nevatia, 1987), (Mohan 
and Nevatia, 1992), (Sarkar and Boyer, 1993). In figure 13 a 
building with extracted edges and an approximation of the con- 
tours can be seen. To ease the interpretation the lines are grouped 
     
& 
raw contours polygons 
Figure 13: Extraction of edges and polygon approximation 
(figure 14): In a first step all parallel lines are selected given a 
maximal distance and a maximal error for the angle. From these 
all those pairs are selected which enclose homogenous areas. 
169