S/N STREO MATCHING ALGORITHM USING NEURAL NETWORK BASED ON HOPFIELD MODEL 
nent 
has Riichi Nagura Nobuyuki Matsui Teppei Ogiyama 
e as Department of Computer Engineering 
tude Faculty of Engineering 
Himeji Institute of Technology 
2167, Syosya, Himeji-shi, 671-2201, Japan 
E-Mail: nagura@comp.emg.himeji-tech.ac.jp 
Commission VII, working group NO:1 
S/N KEY WORDS: Earth image observation, Stereo matching, Correlation analysis method, Neural network, 
the Hopfield model 
tion 
S. ABSTRACT 
sults: The correlation analysis method is well known for extracting height data from the stereoscopic pair 
fove images. The correlation analysis method, however, has some issues such as accuracy, although many 
the efforts have been expended in its improvement. In this paper, we propose a stereo matching using neural 
network(N.N.) method based on Hopfield model, and report the improvement of accuracy compare to the 
correlation analysis method. First, we indicate our method, and next, we report results of computer 
simulation in applying our method to some kinds of stereoscopic images. The results show our method is 
superior to the conventional correlation analysis method. 
and 
y of 1. INTRODUCTION numerical patterns, actual images obtained from } 
Sims. the second method, using our stereo matching } 
UNS There has been significant progress in based on Hopfield model. 
ition applications of earth observation from satellite. | 
land Particularly, demands for stereo imaging are 
. 78. increasing to obtain a global terrain map of the 2. CORRELATION ANALYSIS METHOD 
earth's surface. 1 
fight Stereo matching is the method extracting height 
orbit Stereo matching method is used to extract height data from the parallax that is calculated by i 
jural data from stereoscopic pair images obtained from searching the most resemble parts, matching 
different two viewpoints. There are two methods points, in stereoscopic pair images shown in | 
for taking stereoscopic pair images by a satellite. Fig.1. In the correlation analysis method, the 
1: The method of taking stereoscopic pair images correlation coefficient is described by 1 
from two different orbits, 2: The method of taking 
stereoscopic pair images at the different time N-1N-1 Um, n)- I T (m,n)- T| | 
deep from the same orbit. The former is adopted by C= 3S > ICT yD j 
Nota. SPOT in Europe, the latter is by MOMS in Oro cc 
atory Germany and by JERS-1 in Japan. In the first ] Nux 
ring method, there is a problem of observation where / = y 22 I (m, n) ; 
e of condition, which is the difference weather and so m=0 n=0 
on. Whereas, in the second method, it is possible ] ZH 
to take stereoscopic pair images on similar T= — >, > T (m, n) ; 
condition because the time difference is within 1 N° m=0n=0 
minute or so. #  N-IN-1 a2 
L2 3 3 mn)- 1]. 
m=0 n=0 
In this paper, we analyze images, for example 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 23