A true IDB Management System 
(IDBMS) is an IDB and a collec 
tion of tools to handle (store, de 
scribe, retrieve ...) these images. 
(see fig. 3). It is able to discern 
two qualitative IDBMS levels. 
At the lower level the images are 
managed by annotations 
(e.g. satellite, date, processing 
state). This information might be 
managed by conventional 
DBMS which are available on 
the market in a great number. 
Such low level IDBMS fulfil the 
functional requirements for 
many applications using image 
like information but the speed of 
information retrieval is not al 
ways satisfactory. In order to 
meet the efficency of the de 
scribed storage unit a specialized 
hardware was developed which performs searching operations in DBMS’s very effi 
ciently (see Chapter 3.1.). 
Higher level IDBMS manage semantic image information. The geometric structure of 
the objects in an image and their relations (e.g. neighbourhood) are stored in such an 
IDBMS (see fig. 4). An overview on this rather complex task may be found in /3/; 
advanced applications are explained in /3,4/. Research activities are concentrated on 
the introduction of background knowledge in the process of scene understanding and 
on developing adapted query language (Image Query Languages). At present we do 
some work in testing description methods well suited for non- exact object matching. 
3.1. Main Memory DBMS with hardware accelerator 
A relational DBMS was concepted based on the hardware architecture shown in fig. 5. 
A detailed description can be found in /5/. 
Due to the intended applications of the whole system the DBMS was concepted as an 
Main Memory DBMS,i.e. that all annotations are kept in the main memory during the 
working phase of the system. The external memory serves only as a backup medium. 
Main advantages are the fast data access and the independence of physical data dis 
tribution from the user access behaviour. 
The most important algebraic operator in a relational DBMS is SELECT. Presumed 
the data in a MMDBMS are organized in a sequential manner it is possible to con 
struct a rather simple but very efficient hardware to perform the SELECT operator. A 
simulation shows that such a functionally adapted hardware is faster by a factor 2-20 
compared to a pure software solution using well known index structures (this state 
ment must be not true for hash support) /6/. 
Data 
Structure 
Definition 
—I— 
Object 
Coding 
1-. 
Prototype 
Objects 
*|^ De; 
Scene 
Description 
Data Object 
Management 
Image to be 
analyzed 
Object 
Matching 
Segmented 
*IRaster Image 
Raster Image 
Coding 
Image 
Feature 
Selection 
Image Object 
Selection 
Feature + 
Object 
Management 
User Interface 
Fig. 4: IDBMS schema using semantic features