A CONCEPTION FOR OBJECT-ORIENTED DESCRIPTION OF KNOWLEDGE* 
AND DATA IN SCENE ANALYSIS AND OBJECT RECOGNITION 
Yonglong Xu, Msc., Institute of Photogrammetry and Engineering Surveys, 
University of Hannover, Nienburgerstr. 1, 3000 Hannover 1, BRD 
Presented Paper to Commission Ill, ISPRS 
Abstract 
A conception for object-oriented description of knowledge and data is suggested, aiming at semiautomatic 
or full-automatic image analysis and object recognition and integration of the results into a GIS. The 
conception itself can also be treated as an integrated GIS model. 
An object is defined by its attributes, method or action slots and relations with other objects. The world is 
represented in the knowledge base by 2 object nets: a class hierarchy for semantics, a membership 
hierarchy for physical objects and geometric data which are stored in either raster or vector form. 
Photogrammetric data are also represented in the object form and stored in the data base. The object 
recognition is carried out by finding out instances of class objects from the data base and filling the 
membership hierarchy with them. Some preliminary results are reported, e.g. structures for knowledge and 
data description, mutual transformation of external and internal representations, the work in the low-level 
processing and middle-level processing and some useful algorithms. Several open problems e.g. knowledge 
acquisition, reasoning for object interpretation and resegmentation are discussed. 
KEY WORDS: object, object-oriented description, knowledge base, image analysis, object recognition. 
1. INTRODUCTION 
The development of photogrammetry and remote 
sensing is today characterised by more and more 
handling various kinds of digital information with 
computer-aided or automatic approaches. For 
efficient use and management of such information 
Photogrammetry and remote sensing should be 
integrated with geographic information 
systems(GIS), where their main task is the 
collection and updating of data for GIS from 
remotely sensed data/Willkomm, 1991/. This 
tendency is especially reflected on the recent 
EARSeL Workshop on Relationship of Remote 
Sensing and Geographic Information Systems. 
Although several interactive data acquisition 
systems are already developed by different 
companies, such as PHOCUS from the company 
ZEISS /Willkomm, 1991/, automatic image 
interpretation and identification by using image 
understanding techniques remain to be a problem 
/Li, 1991/. 
Since 1960's it has been tried to recognize objects 
on images. It's realized, that it's almost impossible 
to interpret complicated objects on images 
automatically, just using general image processing 
routines without introducing semantics and other 
background knowledge, which are usually owned 
and governed by human experts. With this thought 
Artificial Intelligence (Al) and Expert Systems (ES) 
were created as a new science to treat this kind of 
problems. At the beginning some initiators in this 
  
field were very optimistical, believing that really 
intelligent machines which could think and see 
could be invented in the near future/Zhang, 1986, 
1987/. But the practice has proved that the tasks 
are extremely hard and so far the progress has not 
been so great as expected. Nevertheless this hasn't 
disturbed further relevant researches in this science 
and reasonable use of ES techniques in different 
fields, considering that the above thought 
represents the correct guide direction. Instead a lot 
of Efforts have been made to apply Al and ES to 
solving problems, where human expertise is 
needed, not only in other scientific fields, but also in 
image analysis and understanding, for example in 
medicine/Niemann, 1987; Towers, 1988; Vernazza, 
1987/, photogrammetry and remote 
sensing/Brooks, 1983; Compbell, 1984; Lambird, 
1984; Mooneyhan, 1983; McKeown, 1984; Nagao, 
1980; Riekert, 1990/, and other  industral 
applications/Liedtke, 1989; Pentland, 1986/. Many 
interesting results have been reached, which show 
the good promise of these techniques and 
approaches. 
The first but decisive step for an ES is to design a 
suitable information representation structure. This is 
especially true for a vision ES. Considering the 
supposition that any conceptual or physical object 
can be approximated through a structural reduction 
of it into a set of simple but with each other 
associated elementary parts /Brooks, 1983; Fu, 
1987; Pan, 1990/, such a structure should be 
supported by facilities for data consistency, 
effective data updating, multi-inheritance, various 
* The work, supported financially by the Gottlieb Daimler and Karl Benz Foundation, was made under the 
supervision of Prof. Dr. mult. G. Konecny. 
220