Zhongliang Fu 
  
Weight W"; may be computed with d-rule learning algorithm: 
m m m m-1 
W"zW?"-th-dP?-y" (14) 
h in the equation is study speed. Commonly -0.01« h <0.3,d ; is an epoch error. 
For output layer, 
d'y") (15) 
T, y", is respectively the desired and actual output of ith neurone. 
For hidden layer, 
SHEAR ES APR) YHA (16) 
j 
To avoid that correction value waves, an addition value is added to each correction value. 
DW" (n+1)=DW;/ (n+1) +a :DW;" (n) (17) 
n in the equation (17) is iterative numbers.a is a positive impulse coefficient, a —0.9. 
The equation (16) shows, when y”=1 or 0, even if T, # y",A;" =0 make DW; equal to 0. For avoiding the case, 
when y=0 or 1, set y,=0.1 or 0.9. 
For improving the identification reliability, Reject identification is introduced. The rule is: 
a. All of network output value is less than V,=0.75; 
b. Hypo-maximum output value is great than a threshold V,=0.4; 
c. The difference between maximum and hypo-maximum is less than threshold V;=0.35. 
Output result of first sub-net in a compound network is less different with ideal pattern. But this doesn't affect 
right identification to pattern. It is for second sub-net may also tolerate error. 
6. EXPERIMENT AND CONCLUSION 
To make sure the validity and reliability of all of above algorithms, 28 images are acquired and processed. Fig. 4 
shows an original image and its process results. 
  
  
  
  
S 
qp ste 
EZ (=) 
eR 
al Er ES 
4 886500 
ve 
Fig.4a Original image Fig.4b Binary image 
  
  
A RRESGDO 
  
  
  
  
  
  
Fig.4c Horizontal project Fig 4d Locating character group 
  
310 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.