— OM peo V UI NUM 
wv 
|a 
is 
he 
act 
on 
ler 
be 
Sy, 
us 
data types and description of multirelations and 
interface to external environment. Not all available 
object-oriented systems or ES-shells fulfil all the 
requirements. This motivate us to develop our own 
system for object-oriented description of knowledge 
and data. It aims at semiautomatic or full- 
auotomatic analysis and recognition of objects on 
aerial images and integration of the results into 
GIS. An object is defined by its attributes, method 
or action slots and relations with other objects. The 
world knowledge is represented in the knowledge 
base (KB) by two object nets: a class hierarchy for 
object semantics, a membership hierarchy for 
physical objects and geometric data which are 
stored either in raster form or in vector form. 
Photogrammetric data are also represented in the 
object form and stored in the data base(DB). 
Objects on images are interpreted by building the 
connection between the KB and DB, determining 
instances of class objects from the DB and filling 
the membership hierarchy with them. The 
characteristics of the system are that any data are 
described by an uniform structure i.e. object; each 
object can have fathers of any number with different 
father-son relations and the level depth of each 
hierarchy has also no bounds, so long as there is 
still memory space. Dynamic updating and interface 
to C-functions and UNIX are embedded. This 
structure itself can be used as a integrated GIS 
model. The communication with an existing GIS can 
also be erected by a corresponding interface. 
In this paper the solution of image understanding 
using Al and ES will be briefly reviewed at first. 
Then the model of the object-oriented description 
will be presented, where its internal and external 
structures, characteristics and advantages will be 
explained. Afterwards the up to now obtained 
results and other necessarry preprocessing 
proceedures are concisely stated, e.g. mutual 
transformation of external and internal 
representations, image preprocessing, 
segmentation and its description. Finally several 
open problems e.g. knowledge acquisition, 
reasoning for object interpretation and 
resegmentation are mentioned, which are to be 
studied and worked out in future. 
2. SOLUTION OF IMAGE UNDERSTANDING 
USING Al AND ES 
Image understanding is just to recognize and 
describe concerned objects from input grey-level or 
iconic images. In general, the solution using ES 
consists of three phases in different levels (Fig. 2.1 
): in the low. level processing(LLP) iconic images 
are first .divided into in some sense meaningful 
segments - for the purpose here region-oriented 
segmentation methods are normally used; the 
segments (regions) are then described as primitives 
according to thier attributes and relations in the 
middle level processing(MLP) - all of these are 
main part of the DB; in the high level 
221 
processing(HLP), with the support of DB and the 
information in the KB, which is in advance acquired 
as the model knowledge for the interested objects, 
objects on images are finally interpreted by the 
reasoning mechanism through setting up the 
connection between the KB and DB and filling the 
membership hierarchy with the instances of the 
class hierarchy found in DB. As the result a more 
complete description of the objects is obtained. The 
system architecture in HLP is shown in Fig. 2.2, 
which has four components: reasoning machine, 
KB, DB and interface to GIS. 
Low Level Processing: 
Smoothing & Segmentation 
| 
  
  
       
     
Middle Level Processing: 
Description & set up DB 
| 
I Level Processing: 
   
Focusing & Recognition 
  
Fig. 2.1 solution for image analysis using ES 
To date, understanding primitive perception as the 
interpretation of sensory data by use of 
models(konwledge) of the world is the standard 
vision research paradigm /Pentland, 1986/. Since 
objects in the real world are modelled, a system 
built up on the above scheme is also model-based. 
Such a system is so called blackboard one, when 
special state areas are opend for recording 
processing states, which are modified by system 
actions, and only change in states can cause and 
start new actions of the system. If the actions are 
independent of each other and data-driven, the 
system is also named production system/Negoita, 
1985/. The great advantage of such a system is 
thier distinguished modularity. 
  
      
   
  
iu E INN Reasoning 
Interface to GIS 
    
  
Machine | 
aa] 
| 
| 
Knowledge Base | | Data Base | 
HAAS Lt x: | 
  
Fig. 2.2 knowledge based image analysis in HL 
3. THE MODEL OF THE OBJECT-ORIENTED 
DESCRIPTION