Error | Cor. Method — — — Our Method | 
30 
30 F 
25 | 
20 L rmm 
15 | 
10 —_ 
| e^ 
0 eti tet ieri EUN 
60 55 50 45 40 35 30 25 20(dB) 
Fig.12 Relation between S/N and error ratio 
Fig.12 shows relation between S/N and error ratio. 
An error ratio is calculated by 
|Resultantmage- Truelmagg 
x100 (%) 
X - Y -MaximumimageLevel 
Error= 
  
(10) 
In Fig.12, the error ratio remains some value in 
spite of high signal to noise ratio. This reason is 
that the true image has 256 steps of image level, 
however, the resultant image's steps of image 
level are restricted in the max range of search. 
From these results, the degrading of resultant 
image is a little, and we can show that our 
method is superior to the simple correlation 
analysis method. 
6. CONCLUSION 
In this paper, we propose a new stereo matching 
method using N.N. based on Hopfield model for 
the accuracy improvement of the correlation 
analysis method and we confirmed its advantage. 
Neurons used in our N.N. have two outputs, and 
this characteristic produced desirable result. 
Actual images essentially include noise and non 
Lanmbertian scattering components, and our 
method is effective to these actual condition. 
One of the most important matter for our method 
is the amount of calculation. We take a simple 
algorithm as we mentioned in 4-4 for this matter, 
however, this is insufficient. To improve this 
matter, we are considering an algorithm that a 
(2) 
neuron with v,,. 
changing frequently will be 
chosen in high probability when we select a 
neuron for a transition of the network state. 
Values of W, and W, are issues as well. We 
determine these values by means of trial and 
error now. If we can find the most suitable values, 
the accuracy of our method will be improved 
moreover. 
An attitude fluctuation of a satellite influence on 
the accuracy for extracting height data. A satellite 
moves for taking two images. Therefore, a 
satellite observes two images of slightly different 
points if an attitude fluctuation of a satellite 
occurs. Our method, at this time, searches for 
matching points on the same y-coordinate. It is 
effective to construct N.N. that can search for 
matching points on the different y-coordinates for 
this issue. 
Reference 
(1)K.Adachi, R.Nagura, N.Matsui, "Improvement 
of the height measurement accuracy from 
satellite image including random noise’, 
Proceedings of the 22nd Japanese Conference 
on Remote Sensing (1997) 
(2)Hopfield, J.J., “Neural networks and physical 
systems with emergent collective 
computational abilities”, Proc. Natl. Sci. USA, 
79, 2552-2558 (1982) 
(3) N. M. Nasrabadi, W. Li, “Object Recognition by 
a Hopfield Neural Network”, IEEE, Trans. 
Syst., Man, Cybern., 21, pp.1523-1535 (1991) 
(4) N. Farhat, D. Psaltis, A. Prata, E.Paek, 
“Optical implementation of the Hopfield 
model”, Appl. Opt. 24. pp.1469-1475 (1985) 
(5) T. Poggio, V. Torre, C. Koch, “Computational 
vision and regularization theory”, Nature, 317, 
pp.314-319 (1985) 
30 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998