SEMANTIC MODELS AND OBJECT RECOGNITION
IN COMPUTER VISION
G. Sagerer, F.Kummert, G. Socher
University of Bielefeld
Technische Fakultät, AG Angewandte Informatik
Postfach 100131, 33501 Bielefeld, Germany
KEY WORDS: Vision, Modeling, Knowledge Base, Image Understanding, Hybrid Systems, Semantic Net-
works
ABSTRACT:
Object recognition has a long history in pattern recognition and computer vision. A major problem addressed
is the development of models which are suitable for recognition and scene interpretation tasks. Two principal
paradigms are emphasized. On the one hand side statistical and neural models making use of representative
training samples optimizing parameters of decision functions. Contrarily, knowledge based techniques build
explicit representations by modeling the structure and the constraints associated with a specific task. The
main point of this paper is to show that both paradigms shall and can be incorporated to achieve efficient
problem solutions for complex problems. According to this goal the basic techniques for object recognition
and scene interpretation will be presented and discussed. Based on this evaluation a hybrid system has been
evolved which tries to combine the advantages of the fundamental paradigms. The system is derived from
the knowledge representation scheme of procedural semantic networks integrating the advantages of neural
network approaches for classification and scoring purposes. Thus, explicit semantic models are combined with
learning sample dependent analogous representations. One application of this environment the reconstruction
of three dimensional scenes illustrates that this approach is appropriate for complex tasks. Furthermore, the
accuracy of the results shows that hybrid and distributed modeling of objects and scenes is a powerful and
efficient technique for scene interpretation tasks.
1 Introduction
Object recognition and scene interpretation are
great challenges in various scientific fields. The de-
velopment of algorithms and system architectures
is mainly influenced by research activities in pat-
tern recognition, neural,computer science, and ar-
tificial intelligence. Although the problem is ad-
dressed by a large number of projects, there is not
a unique baseline algorithm or a general paradigm
to solve these complex tasks. According to the clas-
sical problem solving techniques of the three men-
tioned disciplines, approaches have been developed
for many applications. Additionally, the aspect of
suitable models for image processing, image under-
standing, or computer vision is increasingly empha-
sized. Two main streams are considered. Semantic
models are derived from representation techniques
mainly developed in artificial intelligence research.
Statistical and neural network based approaches are
studied mainly in the context of computer vision
providing algorithms for classification and localiza-
tion of objects. While semantic models emphasize
on structural properties, both analogous techniques
concentrate on a holistic viewpoint of an entire ob-
ject.
710
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
In this paper the cooperation and integration of
these basic techniques is proposed. In our opin-
ion, the development of systems which are able to
solve complex tasks requires the study of various
approaches and their use for subtasks within the
overall task of a system. In order to substanti-
ate this, we discuss the problem on three differ-
ent levels. First of all, the baselines of the general
paradigms are discussed. Of course it is not pos-
sible to give a complete review of all algorithms,
representation schemes, and languages. The pre-
sentation aims at the general ideas, advantages,
and methodology with respect to the task of ob-
ject recognition and scene interpretation. Detailed
descriptions on pattern recognition techniques are
given in [15, 25, 5, 7, 18], artificial neural networks
with different models and applications are discussed
in [22, 23, 12, 19, 29]. General knowledge repre-
sentation is addressed in [24, 21, 6, 27]. [2, 16, 3]
discussing semantic models for scene understand-
ing. Computer vision methodology is outlined in
[1, 2, 13]. Secondly, with respect to this discussions
a hybrid representation system for object recogni-
tion and scene interpretation is presented. It com-
bines and integrates semantic networks as a knowl-
edge based approach with artificial neural networks.
for
str
AT
the
SiO
pe
wh
the
tifi
pr
far
Th
ski
rec
of
int
Sit
sm
cifi
ele
thi
sib
on
for
acc
ex:
ac
res
ste
wil
SOC
Or
