STEREO MATCHING USING ARTIFICIAL NEURAL NETWORKS
Goung Loung
Zheng Tan
Dept. of Information & Control Engineering
Xi’ an Jiaotong University (710048) P.R.C.
Comm. III
Abstract:
It is essential to combine the low-level vision with the high-level vision in the stereo matching problem.
At the high-level vision,
ture. The algorithm at this
the stereo matching problem often attribute to the shape
level is robust but much more time is consumed. While at the low-level vision,
recognition of the fea-
the stereo matching problem often attribute to the area-based correlation algorithm. The reliability of the
result at this level is not satisfied.
In this paper, a neural network
petwork consists of feature detecting layer, pattern matching layer, stereo fusion layer,
is employeed to overcome the shortcoming of the traditional methods. This
compound dicision
layer. The output of the pattern mathcing layer is fed back to itself and the fusion layer is guided by the
pattern matching layer. The stereo mathing process is completed when the condition of the compound layer is
satisfied.
Key words: neural networks, stereo matching, fusion, pattern recognition.
1. INTRODUCTION
Stereo vision has a wide application auch as in
robotics, automatic surveillance, remote sensing,
medical imaging ete. The depth information in stereo
vision depends on the retinal disparity, which is the
difference between the location of the retinal image
points in the two eyes. The traditional method te
get the depth information from a pair of image is
that : area based correlation method and feature
based matching method. Each method has its shortecom-
ing. Some researchers attempt to integrate the two
methods but have not got a full success yet.
More and more evidence reveals that stereo vision is
a complex problem. Using a computer-generated stereo
random-dot stereogram, we can find that the forming
of stereo vision does not depend on the recognition,
or say the understanding of the object in the image.
The correspondence is completed point by point in
the stereo fusion process. But it is very difficult
to describe the fusion process in certain a mathema-
tical formuler. And it is unavoidable to fall in
local minimal point when a matching process is attr-
ibuted to a optimum problem such as correlation
which is employeed to solve the problem of similari-
ty between the two images. That is to say, Ît is net
necessary to form the stereo fusion on the basis of
the understanding of the objects in the image. While
on the other hand, we can find out the corresponding
points in this way : we view a certain point in the
left image and then we can throw the left image away
‚viewing the corresponding point in the right image,
and vice versa. Here we recognise the feature of the
point being viewed and there is no stereo fusion
acted in the process.
The stereo vision, or say the stereo matching can be
realised either at low-level vision or high-level
vision. At the low-level vision, the stereo matching
emphasize the parallism of the fusion process. And
it is meaningfull only when the fusion condition is
satisfied. That is to say the fusion activate only
when the points being viewd is close enough in posi-
tion to the candidate points. While at high-level
vision, stereo matching emphasize the recognition of
the feature point, the shape, the edge structure in
417
the image. The feature representd the salient point
consist of the high-frequency part in the image. The
other pointa consists of low-frequency part in the
image. The stereo fusion is formed mainly by the low
frenqueney part in the image and restricted by the
shape, the structure of the image. According to the
physiology, the high frequency part performs the ri-
valry while the low frequency part perform fusion.
If the matching problem is realised by the combinat-
ion of stereo fusion with the feature recognition,
it's no doubt that the robustness of the algorithm
will be improved and the computational amount will
be reduced.
In this paper, we present a way to solve the corres-
pondence problem by using the parallel mechanism and
the computational power offered by the artificial
neural networks. The neural networks consist of four
layer : feature detecting layer, pattern matching
layer, stereo fusion layer, the compound decision
layer. While the pattern matching layer is working,
stereo fusion layer acts simultaneously in each seg-
ment devided by the feature points. It act under the
guidance of the pattern matching layer. The compound
decision layer ensure the reliability of the result
and adjust the network to avoid it fall into a local
aatem point. The diagram of the neural network aee
g 1.
Feature Pattern
Image Detect- Match- |
o» =
|Decision
Layer
stereo
Fusion M
Layer
Fig. 1. The diagram of stereo
matching neural networks
Layer
| Output
