can 
| fit 
line 
sti- 
rich 
nal. 
ing 
not 
sed 
‘his 
2 of 
ons 
ect 
ods 
the 
ty 
ich 
res 
Ti) 
  
  
  
  
a | b 
d 
elf 
  
  
  
  
Figure 7: The line extraction from the first image 
(ZHENG and HAHN, 1990). The weight function p(r;) 
can be, for instance, the Cauchy function defined by 
1 
N) = mr, 
where c, is the standard deviation of the fit and can 
also be estimated. 
For the inference and reasoning process of a higher 
abstraction level during feature extraction, it is de- 
sired to know something about the quality of featu- 
res extracted from low level processing. Many feature 
extraction algorithms, however, lack a detailed and 
comprehensive description of extracted features. Ac- 
tually, after parameter estimation, it is also possible 
to estimate the posteriori accuracy of the estimation. 
As a measure for the global fittingness of data to a 
model, the estimate 
=D (16) 
n—u 
can be used, where wj is the weight and u is the num- 
ber of the unknown parameters. Besides, the poste- 
riori accuracies of the parameters in €) can also be 
estimated and they are denoted by (09, Jp, 0¢, 0a, 75). 
871 
  
  
  
  
  
  
  
  
Figure 8: The line extraction from the second image 
Now, the line support region illustrated in Figure 5 
is used to estimate the parameters in model I. The 
results can be stored in a form of knowledge represen- 
tation known as frame (MINSKY, 1975) (cf. Tabular 
1). It is to say that the accuracy of the line position is 
about 0.4 pixel (subpixel accuracy) and the accuracy 
of the line orientation is about 0.002 radian. The line 
estimated in this way is shown in Figure 5a using the 
white line. 
6 Experimental Results 
The algorithm described in the previous sections was 
applied to two aerial images. Due to the limited space 
of this paper, many interesting intermediate results 
can not be discussed in this section. Here we just il- 
lustrate the lines found by the algorithm. 
The first image (cf. Figure 7a) shows an aerial scene 
with a house and fence on rolling terrain. Due to sha- 
dows and poor contrast, the roof borders get frag- 
mented. The image was used to train the net illustra- 
ted in Figure 6. After training, the net can automa- 
tically group image pixels into line support regions