GPS 
Proc 
gart 
ance 
43rd 
of 
ata. 
ight 
eek, 
de 
sche 
Computer Assisted Problem Solving In Image Analysis 
Fang Luo 
N.J. Mulder 
International Institute for Aerospace Survey and Earth Sciences (ITC), 
P.O.Box 6 , 7500AA Enschede, 
The Netherlands. 
ABSTRACT: 
The researcher / user of software for digital image analysis is confronted with huge libraries of subroutines. In order 
to solve a problem from a set of problems, it is, in general, not clear which subroutines should be selected, with 
what setting of the parameters, The authors have set out to structure and classify subroutines generally available 
in image processing and image analysis libraries as a first step in bottom up knowledge engineering. In order to 
reduce the redundancy in the large sets of subroutines, a virtual image analysis engine has to have a 
minimum(reduced) instruction set.Computer assisted problem analysis is approached in a top down manner. A 
PROLOG style specification language is developed, which allows goal directed programming. This means that the 
problems have to be specified in terms of relations between the components of a model.The language will check 
whether the number of constraints is sufficient, and if so, will solve the unknown(s). Often a search of problem 
space has to be performed where an optimisation criterion is required (cost function). The criterion used here is 
minimum cost/ maximum benefit of classification or parameter estimation. 
Key words : image processing, expert system, knowledge based system, image analysis, knowledge representation, 
reasoning. 
1 INTRODUCTION 
In the image process domain, a variety of image 
processing algorithms have been devised to facilitate 
image analysis. Various software packages for image 
processing include many techniques advanced in the 
history of digital image processing. These software 
packages can be used efficiently in problem solving by 
only a few experienced people; they offer many choices of 
subroutines and often require a large set of parameters to 
be defined by the user. Choosing subroutines and 
parameters may prove to be quite a complex task, 
although expert users of such packages may find it easy. 
One way to make these software packages more 
manageable and usable by a wider user community is to 
capture the knowledge of expert users in controlling these 
software systems. We can visualize this as an expert 
program that monitors the use of the software package, 
helps the user in understanding and controlling the 
package and also provides interpretation of the results 
produced by the user's interaction with the program. This 
expert system would, therefore, contain domain knowledge 
to be used in choosing the appropriate methods and 
techniques from the software package. 
The following problems are encountered in designing 
Such a system : 
1. Assessment of image quality. To measure the 
quality of an image 
is the first step in image analysis. The quality of an image 
is defined by its potential in providing information about a 
class of objects in the scene. One decision based on the 
evaluation of the quality is to model errors and artefacts 
and remove them as well as possible by inversion of the 
error model. 
2. Selection of appropriate procedures. There are 
509 
many different procedures (algorithms) for a specific 
image processing task. They are designed on the bases of 
different image models and computation schemes. One 
has to select appropriate procedures considering image 
quality, the purpose of image analysis and characteristics 
of the procedures. 
3. Determination of optimal parameters. Many 
procedures have adjustable parameters, performance is 
heavily dependent on the values of the parameters. 
4. Combination of primitive procedures. It is often 
necessary to combine many primitive procedures to 
perform a meaningful task. For example, a popular way of 
extracting regions from an image is to apply edge 
detection --» edge linking --»closed boundary detection. To 
attain. effective combinations, knowledge about image 
processing techniques is required. 
5. Trial and error experiments[13]. It is usually very 
hard to estimate a priori the performance of a procedure 
for a given image so that one has to repeat experiments 
by modifying parameters. The definition of a cost / 
performance criterion allows the use of numerical 
optimisation [14]. 
Recently, several knowledge based systems for image 
processing were developed to facilitate the development of 
image analysis processes. These incorporate knowledge 
engineering(KE) techniques to solve the above problems. 
Examples are: a consultation system for image 
processing[3], a knowledge-based program composition 
system[11,12] and a goal-directed image segmentation 
system[10,13]. Here we present a knowledge based 
method to provide users with the main functions: 
hypothesis generation from queries, the organisation of 
processing sequences [7] and the setting of parameters of 
subroutines.