nt 
In the 
lation 
points 
t data 
ye set 
the correlation window size to be as small as 
possible under the condition that we can remove 
miss matching points. 
3. HOPFIELD MODEL 
A basic model of neuron is shown in Fig.3. A 
neuron calculates its output by a certain function 
from a difference between the sum of input(^w" is 
the connection weight on each input) and 
threshold( 6 ). The step function or the sigmoid 
function is used as well, namely, 
1 if Xw;x,-0>0 
  
Output = (2) 
0 other 
or 
Output = Lu ; 
1.0— es|- dy W,X, - 2 (3) 
wherea is constant. 
Hopfield model is one of symmetrical 
interconnected N.N.. The energy of network is 
defined by 
E- NE +3 6x. @ 
i. sj i 
This is decreased by transition of the network 
state. The transition rules of network state are as 
follows: 
(1) Select a neuron. 
(2) Calculate this neuron's output from its input. 
Finally the energy reaches the minimum, and 
any neuron's output is not changed. Using this 
characteristic, we can solve a given problem by 
assigning an estimation function of its problem to 
the energy of network. 
W2* X2 
Input 2 > e Output 
Wn* Xn 
Fig.3 Model of neuron 
4. CONSTRUCTION OF N.N. 
It is difficult to solve a given problem by the 
traditional Hopfield model. This reason are as 
follows: 
(1) If we use the step function, Eq.(2), for a neuron 
output, its N.N. tends to fall into local minima 
owing to radical changes of its state. 
(2) If we use the sigmoid function, Eq.(3), each 
neuron output is often the same value. 
Therefore, in this paper, we construct our 
improved N.N. based on Hopfield model. 
We use a function that adds the relation between 
neighboring pixels to correlation value as the 
estimation function. We use a difference of height 
for the relation between neighboring pixels. 
4-1. Network Structure 
It is necessary, first, to define the network 
structure to solve a problem by Hopfield model. 
We define the network as shown in Fig.4. 
  
  
  
  
  
sa O y 
0000 — bei 
» X 
Fig.4 Network structure 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 25 
] 
1 
1H 
4 
| 
N 
i 
à 
j 
ia 
| 
Le 
ei 
d 
] 
ld