2. DESIGN OF AN ADVISORY EXPERT 
SYSTEM FOR IMAGE ANALYSIS 
At ITC and UT(University of Twente), there are at least 
six librariesof subroutines for image processing available. 
Each of these has its own characteristics and advantages. 
We categorize these subroutines according to their 
processing functions. We make an expert program to 
organize (by grouping and classification) the subroutines, 
and select them using heuristic information. A PROLOG 
style specification language is being developed to handle 
the problem analysis (logic of selection) and the running of 
compiles subroutines. This provides the user with an 
advisor for planning a process sequence and for setting 
the parameters. 
There are many redundant procedures in the 
image processing libraries. We keep these redundant 
procedures but provide users with a minimum (reduced 
instruction) set of a non existing (virtual) machine . This 
requires the administration of equivalence or the 
reorganisation of subroutines into a more orthogonal set 
with appropriate parameters. 
An example of elements of a reduced instruction set 
for image analysis is based on the treatment of a complete 
image as one object in a register. Typical Operations are 
: Copy( from image register A with shift(dx,dy), multiply 
with a constant and accumulate result in image register B) 
, Pack bytes from( A,B,C) into X, sort X, generate 
Frequency of coincidence(X -» A,B,C), Select maximum 
frequency (likelihood). 
In order to select the best sequence of 
subroutines and their parameters we must define 
evaluation functions in order to enable standard 
optimisation software [14] to select procedures, using 
minimum cost / maximum benefit .The value of the 
evaluation function is the weighted sum of all factors (e.g. 
the intention of a user, length of path, time consumption, 
software uniform etc). 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Knowledge of 
Arnysis Processes 
Pan : Question 
Kien — 
Knowledge of Generation . 
Image Domein 
t 
K f Procedures & e | Answers 
etes Er LE 
a 
c 
e 
Library of Image 
Processing 
Procedures 
  
  
  
Figure 1. Architecture of the system 
Figure 1. illustrates the general architecture of the 
system. It consists of a reasoning engine, the knowledge 
about image processing techniques, a library of image 
processing procedures and a database of characteristics 
of the input and processed image data; procedures in the 
library are applied to analyze the image and the result is 
stored in the database. The reasoning engine uses the 
510 
knowledge about image processing techniques and 
characteristics of the image data for reasoning. 
To develop the expert system for image processing, 
it is very important to create a programming environment. 
As we know, there are two kinds of programming style. 
One is the imperative style which tells the machine exactly 
what to do,such as FORTRAN, C, etc. Another is the 
declarative style which only describes the domain problem 
and lets the machine take over the problem solving, such 
as PROLOG, LISP etc. So far PROLOG is a rather 
successful language in Al research. But it is not 
satisfactory enough to develop an expert system for image 
processing, because its computation capability is too low 
to meet the needs of computation in image processing. It 
is essential to develop an application language which has 
both powerful capability of describing and of solving 
problems. 
Although the construction of a general knowledge- 
based image processing scheme is a very long-range goal, 
the appropriate combination of state-of-the-art techniques 
can solve a class of problems in a specific domain. Under 
the given computational environment at ITC, we develop 
a knowledge based system which is able to automatically 
plan the processing sequence and select the arguments 
for the user, using PROLOG for the advisory/planning part 
and linking it with compiled subroutine libraries. In section 
3, we present a knowledge representation scheme in the 
system. In section 4, we describe the reasoning process. 
3. KNOWLEDGE REPRESENTATION 
To create programs that have "intelligent" qualities, it 
is necessary to develop techniques for representing 
knowledge. Unlike people, computers do not have the 
ability to acquire knowledge on their own. Any knowledge 
they contain about the world has been explicitly provided 
in the form of data and knowledge structures. 
Knowledge structures are usually closely tailored to 
specific problem areas which are called problem domains. 
The domain is defined by the set of relevant information 
required to solve a specific problem. For a case study we 
selected a software package named SPIDER which 
consists of over 400 FORTRAN subroutines for various 
image processing algorithms. 
The knowledge used by the system comes from 
three aspects: 
(1) The knowledge of image processing 
algorithms. 
This deals with the usage of algorithms, condition 
for arguments and the range of the parameters. We can 
find this sort of knowledge in the manual of a software 
package, 
First of all, we divided SPIDER subroutines into 
groups, and found their common parameters for each 
group. Figure 2. illustrates the relation between 
subroutines. From the figure, we find that frames are 
rather suitable to represent these subsets of procedures. 
Using figure 2. as a guide for frame layout, we can 
embed the usage of the algorithm, conditions for 
arg 
of di 
funct 
like « 
cate 
L 
slots 
impo 
slots 
that i 
abou 
A 
inforr 
usag