e to 
ıt of 
the 
the 
vj. k 
(7) 
s, (x, 
f E. 
| the 
its of 
| and 
The 
'"uron 
tput, 
lefine 
?uron 
i)Threshold: The threshold is of determining the 
output of neuron. Therefore, the first term of the 
right side of Eq.(7) can be represented by the 
threshold. That is, the threshold is given by the 
correlation coefficient with inverted sign: 
0, = -W,Cor(x, y, z) (8) 
xX, V,Z 
ii) Connection Weight: The connection 
weight is of preventing a neuron against a 
restriction from having large output. Therefore, 
the second term of the right side of Eq.(7) can be 
represented by connection weight. In Fig.4, the 
z-coordinate of a neuron represent the parallax. 
Therefore, we define connection weights to 
prevent neurons of which z-coordinates are 
irregular different from having large output. 
Concretely, the connection weight between a 
neuron at coordinate(x,,},,z,;) and a neuron at 
coordinate( X, , Y, , Z, ) is defined as 
= -Wja -zj if -xj «I and »" -y|sJ 
me 0 otherwise 
(9) 
4-4. Transition rule of Network State 
In our N.N., almost neuron-states are unchanged 
during iterations of the transition because our 
N.N. consists of a lot of neurons, For example, if 
we set X-256, Y=256, and z=15, we need 256 x 
256 X15 of neurons. Therefore, if we select a 
neuron at random, amount of calculation will be 
too much. 
To prevent this, we watch v . and if the 
x,y,z > 
number of change of v? reaches the set 
x3,»,2 
number, we do not calculate output of each 
neuron on the coordinate(x, y). However, this 
method cause ve , to be fixed error. Therefore, if 
(2) 
x»: the number of change 
there is a change of v 
of neighboring v sets 0. 
3,9,2 
5. EXPERIMENTS AND RESULTS 
In Fig.6 and Fig.9, we show input images used in 
our experiment. Fig.6(a), (b) show the numerical 
patterns for non Lambertian model. Fig.9(a), (b) 
show actual images of earth's surface from a 
satellite. Non Lambertian : model partially 
includes specular scattering components. This is 
the phenomenon observed at a building and a 
lake and so on. For non Lambertian models, the 
correlation analysis method searches miss 
matching points because the brightness levels are 
not equal although matching points. 
Fig.7, Fig.8, Fig.10 and Fig.ll show resultant 
images from experiments of the simple 
correlation analysis method and our method. The 
simple correlation method that are used in here is 
the simplest method, however, if we use other 
superior method, we can obtain images that are 
better than Fig.7(a), Fig.8(a), Fig10(a), Fig.11(a). 
However, the amount of calculation increase in 
that case. Fig.7(a), (b) show resultant images 
using numerical patterns without any noise for 
Lambertian model(Fig.2(1). Fig.8(a) (b) show 
resultant images using numerical patterns 
without any noise for non Lambertian model. 
Fig.10(a), (b) show resultant images using actual 
images of earth's surface from a satellite. 
Fig.11(a), (b) show resultant images using images 
that adds noise corresponding 30dB(S/N) to Fig.9 
(Fig.9 includes noise corresponding 
approximately 40dB(S/N)). In these images, the 
bright point shows the high point and the dark 
point shows the low point. We set the correlation 
window size 5, and set the comparison range of 
height 5X5 and set initial values of each neuron 
the values that are calculated by only the 
thresholds if the x- and y-coordinates of neurons 
are even, otherwise set others 0. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 27 
| 
| 
1 
3 
hi I 
L$ rand LSE Ss