with 
out- 
utput 
late 
(4) 
(6) 
or: 
(6) 
(7) 
ork 
est 
ted 
est 
ing 
ge, 
urs 
ge. 
eft 
in 
eds 
age 
la- 
Ing 
Ing 
to 
Il 
|1- 
et- 
the 
salient points at the edge should act simutaneoualy 
in the matching process of two adjacent epipolar 
lines. This can be controlled by software in the 
simulation process and can be performed in the real 
neural network. 8o that the structure information of 
the image act in hidden way in one dimension in the 
matehing proceas. 
As opposed in the feature detecting layer, the feat- 
ure of one interest point consist of six parameters. 
The feature of an interest point is denoted by Fij, 
where 1 denote the type of the feature, j denote the 
input position of the interest points. The number of 
the interest point in left Image and right image is 
Nl, Nr respectively. The model for pattern matching 
layer see Fig. 3. 
Fi Fa Fa Bu BsR3 Fa Fais fui Fa Fa 
HU Hl 
    
Left 
Fig. 8. The Model for Pattern 
Matching Layer 
The intersection denoted by circles represent the 
possible matching element. They occur at intersection 
which bring the same type of feature together. All 
matching elements corresponding to a common position 
make a group for that position. And mutual inhibitory 
connections are defined between these groups in the 
same way as for the stereo matching network. Here, & 
group of feature in solid circle inhibit the other 
group of feature in the two oblique direction along 
the line of vision. 
When using the neural network, the pattern matching 
problem can be formulated as the minimisation of a 
cost function ( constrained optimisation ). The cost 
function we adopted for the solution of the pattern 
matehing problem is as follows : 
E= - (1/2) à! > A x Z rus Vikm Vjin 
Fh Bat Vin © 
whero M is the feature number of one interest point, 
Where Vikm and Vjjg represent the binary state of 
ik and jl! neurons respectively, which can be either 
1(active) or O(inaetive). Tijkim ia the interconnee- 
tion strength between the two neurons, Iikm la the 
initial input to each neuron. A change in the state 
of neuron ik by AVikm cause an energy change of 
A Bik. 
AB = - | zi 3 RIAN Viku Yj im 
419 
+ Iikm J AVikm (9) 
The equation above describing the Dynamics of the 
network was shown by Hopfield to be always negative 
with a stochastic updating rule. 
| A 
Vik > 0 if [ N Ag uu Vikn Vj in 
+ Hk 1»? 0 
> | Jr 
Vik = 1 if [ > Y PETS Vikm Vi im 
+ Iikm 1 « 0 
| Air 
no change if I ON z PRIE Vikm Vj in 
+ Iikm 1 * 0 
The interconnection of the neural cell is indicated 
as Fig. 4. 
  
  
   
  
  
  
Right 
Image 
Feature 
l Ie eee uer! spe tmu E. 
I 
Ln 
kr | 79m 
| Ii | | in 
! | 
i J Left Image 
Feature 
Fig. 4. Interconnection of 
neural cell 
The deformation of the cost function for the stereo 
corresponding given below is minimized : 
Ie À 4 
Cikjin Pus P] * Bi Aa. d ue: 
, x 20 oh as? (10) 
The first term in (10) represent the degree of comp- 
atibility of a match between a pair of points (k I) 
in the right image, while the second end third terms 
tend to enforce the uniquess constraint where the 
probabilities in each epipolar line should add up to 
1. The compatibility measure is given by 
X=W Ad] * W ADI 
where Ad is the differendee in the disparities 
of the matched points pairs (i,k) and (jJ, D. AD is