processing step 
  
Gaussian convolution : 22.0 
computation of zero crossings 10.9 
edge streak generation 4.7 
unselection of insignificant edge streaks 1.7 
corner Gerection d 2.0 
eneration of edge 
Sieighbourhood graph 36.3 
  
X | processing 1 image 77.6 
  
processing 2 images 155.2 
  
prediction Q 
propagation  . >. 
establishing point correspondence 1 
  
  
  
  
  
Table 7.1. Computation times for an image of 230 x 230 pixels 
on a Sun sparc IPX station. 
8.1 Future Improvements 
The generation of edge streaks as well as the 
detection of chain-code corners proved to be rather 
sensitive to changes in the image intensity function. Thus, 
the path of an edge streak as well as the detectability and 
detection of a corner at a junction in the image is 
uncertain. This problem can be substantially reduced by 
an extension of the corner detection towards the more 
general idea of junction nodes in the graph structure. 
Junction nodes are points where several edge segments 
can begin or end. Edge streak corners are considered 
being a type of junction nodes. The other type is a T- 
junction, where one edge streak begins or ends on 
another edge streak. 
    
Fig. 7.5. Matched segments 
The concept of the matching method is the 
establishment of matches by comparing descriptive and 
relational parameters. This way of matching works well 
as long as the parameters are distinguishable. As soon as 
several edge segments are equal in several parameters 
(e.g. in an image of a brick wall), minor disturbances of 
noise or texture edges in the graph structure weaken the 
relational information which is one of the few 
distinguishable parameters in the propagation. The 
traversal of the graph using noise edge pairs prevents the 
algorithm from finding correct matches. The problem 
can be solved with the help of an evaluation of disparity 
gradients. It is assumed that disparity changes slowly 
when surfaces are continuous. This property is also valid 
for one side of occluding edges. It can be used as a 
relational parameter during the propagation step. 
Disparity gradients between the match to be established 
and the last match of class I among the predecessor nodes 
in the propagation tree should be small in order to 
support the evidence for a correct match. For the 
disparities in y-direction this results in an integration of 
epipolar geometry into the match function, as y- 
disparities should be zero while she epipolar lines are 
horizontal. 
References 
Ayache, N. and Faverjon, B. [1987], "Efficient Registration of 
Stereo Images by Matching Graph Descriptions of Edge 
Segments", International Jounal of Computer Vision, pp. 
107-131. 
Berzins, V. [1984], "Accuracy of Laplacian Edge Detectors", 
Computer Vision, Graphics and Image Processing, Vol. 
27, pp. 195-210. 
Beus, H..L. and Tiu, S. S. H. [1987], "An Improved Corner 
Detection Algorithm Based on Chain-Coded Plane 
  
316 
Ft 
Fr 
Hu