oard System
. T he model of
oresents a hier-
n a blackboard
lative memory.
> model for the
fficiency of the
constant refer-
ducing interim
on. Hence fol-
visual recogni-
tion of a sym-
Y
nition of three-
es (Fig 1). It
ction of points
n, which is es-
st to other re-
pipolar geome-
approach we
nation such as
5], digital map
5], etc. Due
stablishing hy-
entation of ob-
m up. As on
ot be decided
ect, alternative
s) are pursued
/ by regarding
t, i.e. prepro-
ations between
ouring objects
uring in space.
the scene level
image).
For the consideration of different recognition tasks it
is particularly convenient to use a uniform framework
for the analysis. The software tools and special hard-
ware which have already been developed can be used
for an analysis. For the description of an object model
by a production net we use modular semantics. This
description is translated into independent knowledge
sources (processing modules), which may be reused for
other analysis tasks. For the evaluation we assume that
the number of primitive objects describing the image
content can possibly be very high. The analysis con-
cept is designed in such a way that in principle it al-
lows a parallelization of processing. Furthermore, the
evaluation of several information sources (e.g. multi-
sensor images, spatial image sequences) can easily be
integrated.
Fig. 1: Section of ISPRS-Test dataset FLAT
a) left image, b) right image
3 BLACKBOARD-BASED PRODUCTION
SYSTEM BPI
For structure analysis of complex scenes the black-
board-based production system for image understand-
ing (BPI) [Liitjen, 1986] is used as framework.
3.1 Production system
A production system typically consists of three basic
components: a database, a set of production rules and
a control unit. The knowledge about object structures
is represented by a set of production rules. A produc-
tion rule, or production, is a statement in the form:
IF condition holds, THEN action is appropriate.
The execution of action will result in a change of the
data contained in the data base of a production sys-
tem. A control unit controls the overall system activity
and has the special task of deciding which production
(with satisfied condition part) to fire next.
The process of building up more complex structures
from less complex structures, using such productions,
can be described by a rewriting system. With reference
to formal languages the rewriting system may be de-
termined by a Grammar G. Such a formal grammar is
defined by a 4-tuple
G = (S, V4, Ve,.P.),
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
where S is a set of start symbols (target objects), V,
is a set of non-terminal symbols (partial objects) V; is
a set of terminal symbols (primitive objects) and P is
a set of rewriting rules (productions). Attributes are
assigned to the objects, which represent certain struc-
tures. The productions determine how a given set of
objects is transferred into a set of more complex ob-
Jects.
In analogy to string grammars we may say that an im-
age content is parsed (bottom up) by the process of
image analysis. Instead of examining concatenation as
is done by parsers for string grammars, we examine the
topologic or geometric relation of objects in the condi-
tion part of a production. Therefore, a production rule
may be written in the form:
Pi XANY OS Z
This means that, if an object of type X and an object
of type Y fulfil the relation ©, then an object specific
generative function > is carried out which produces
an object of type Z. Here the productions describe the
part_of relations.
Starting with primitive objects, a target object can be
composed step by step using the productions repeat-
edly. Similar to tabular parsing methods (e.g. in Aho
& Ullman, 1972) all interim results (partial objects) re-
main stored in the database during the analysis.
The general interaction of productions and the step-
wise transfer of objects into objects of a higher ab-
straction level can be depicted by a production net
(Fig. 3). The compositions for the actual objects (in-
stances) are recorded with the aid of pointers and can
be illustrated by a derivation graph.
After the analysis, derivation graphs of the target ob-
jects can be constructed and used to explain the re-
sults. Thus, the subset of primitives, which represents
the target objects can be determined. If we compare
this subset with the set of primitives, we may say that
the production net acts like a filter.
3.2 Blackboard Architecture
In the BPl-System the productions are implemented in
a blackboard architecture (Newell, 1962; Nii, 1986).
Generally, a blackboard architecture consists of a global
database (blackboard) and a set of knowledge sources,
which communicate only via the blackboard. In BPI
the global database is stored in an associative memory.
Knowledge sources are constructed as independent ob-
ject specific processing modules, which examine a con-
dition and execute an action of a production.
Systems with a blackboard architecture are essentially
data driven. One or more hypotheses ” part_of a more
complex object” are attached to an object. An object-
hypothesis pair (processing element) triggers a process-
ing module to verify the hypothesis. The hypotheses
arise from the production net, so that the analysis pro-
ESSENCE RARE
