Kondo, Hiroshi 
  
Fig.6 Histogram of Fig.5 
(T, : Threshold value) 
  
gation NN with training patterns are reliable and 
suitable. The output values are 0 — 1 through a 
logistic function. The outputs are set as 0.6 mm6ó, 
0.8mm 9 , 1.0mm $ , , and 3mm diameters of 
rounded opacities. The number of outputs are 18. The 
13 outputs for rounded opacities. The other 5 outputs 
for unnecessary parts like vessel and/or rib shades. 
Hidden layer has 171 neurons that is determined 
optimally. The input layer has 30^ 30 neurons i.e., 
30^ 30 pixel image picked up from bi-level image. 
4.2 Bi-leveling 
As an NN input the simplest figure is favorable for the pattern 
matching. Here we employ a bi-level image for our NN. The bi- 
level image is made from the moving normalized image Fig.5. 
Fig.6 is a histogram of Fig.5. From this figure we see that it is 
symmetric and so it is very easy to make a bi-level image. The 
threshold value for making a bi-level image is taken as a mid 
point i.e., valley between two major modes. Exactly we can 
determine the threshold value by using the peak score of Fig.6. 
Figure 7 shows a resulting bi-level image. Almost all rounded 
opacities and rib shades are detected so clearly. It means the 
moving normalization is effective as the pre-stage of a bi- 
leveling. 
4.3 Neural Network Approach 
NN has an excellent feature for pattern matching. Furthermore 
the pattern matching is done in a real time. We employ the back 
propagation NN with three layers. It requires the training 
patterns that is shown in the succeeding section. 
4.3.1 Back propagation Neural Network 
    
  
Thre 
e 
Output layer 
w^ s 
back 
propa 
Hidden layer 
v 
Input 
first 
Fig.8 Neural Network Structure are 
With the number of input layer N, the hidden layer unit H ; 1s expressed as 
av; 
1 zo IH 
H,- f, &à V, T 
&-1 
ceo 
(2) 
where I, in the i th input value of the input layer and W/" is the coefficient between the input layer and the hidden 
layer. f; () is defined as 
~~ 
db uh 
YC ya 
ur nez 
(T, : Threshold value) 
3) 
similarly with N, the number of hidden layer neurons and W;" the coefficient between a hidden layer and output one, 
the output O, is 
eue, 
O, = f; B "Ha 
where f? is a logistic function shown as 
(4) 
  
  
  
456 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.