geometric transformation between the map and the image 
using an algorithm. The parameters of this transformation 
are used to find the rotational differences, which in this case 
  
    
  
      
is -39.3925 in degrees to alter each edge pixel gradient 
direction. 
Figure 2. a) Histogram equalisation contrast stretched subscene of Farnborough image, b) Edge preserve smoothness of 
Farnborough subscene, c) Building region segmentation of Farnborough subscene, d) Edge enhancement of Farnborough 
subscene, e) Farnborough subscene - edges to the local maxima, f) Edge gradient direction of Farnborough subscene. 
The algorithms used above for the preparation of the image, 
resulted in a two frame output, the first is edge strength and 
the second edge gradient direction which is shown in Figure 
2(f). The image with these two frames, edge strength and its 
directions, is ready at this stage to be used as input for the 
matching with the map. 
3. MAP AND IMAGE MATCHING 
The matching routine owes much of its inspiration from 
Maitre et al, (1989), and the use of dynamic programming 
method for matching purpose is described by Newton et al, 
(1994). The core of the algorithm is a routine for matching 
map boundaries to image edges. Four control points are 
selected manually in well distributed pattern for an initial 
2D transformation between the map and the image. The 
906 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
matching routine with these control points project boundary 
pixels from the map into the image space to predict edge 
pixel position. Matching is defined by two components of a 
cost function: 
* The distance between predicted edge pixel and the edge 
pixel under consideration is the first cost component. 
* The difference in their gradient directions is the second 
component of cost. 
Only one edge pixel can be matched to one map boundary 
pixel, and the edge pixel under consideration with minimum 
cost is considered as the best match. 
The use of these dynamic programming method generated 
923 match points which are shown in Figure 3, and also 
  
  
  
  
uu d sn aun ud V s ud VAN V Mo aude FIN S fA ASU Qu e jy — C . 1)