for intermediate steps. The task must be decom-
posed into several processing steps. But due to the
variability of the sensor data of an object and the
aim of individual descriptions, the sequence of pro-
cessing as well as the transformation steps to be ap-
plied can not be fixed a priori. Therefore, we have
to deal with a search process where for each step the
following question must be answered: What is the
best transformation at the present state to achieve
the overall goal of the analysis process? The search
process must be guided and restricted by informa-
tion about the problem domain and the specific task.
This knowledge about objects, events, structural
properties, and constraints must be explicitly rep-
resented in such away that it can be efficiently used
for the interpretation process. À knowledge base for
object recognition and description tasks must cover
semantic models which enable to establish connec-
tions between numerical sensor data and symbolic
entities. Fig. 2 reflects the two main lines which
must be incorporated in the construction of seman-
tic models and a knowledge base.
physical environment -- time and space constraints
explicit representation of structural properties
Y
declarative and
procedural knowledge
À
explicit representation of rules, algorithms
|
rules, algorithms
À
i
symbolic world -- experience, and common sense knowledge
Figure 2: Knowledge for Object Recognition and
Description
The base line of knowledge based object interpre-
tation systems is given by a state search approach.
The initial state is given by some complex pattern
f(c) and the knowledge base. This covers the se-
mantic models of objects, procedures, and functions
which realize transformations between and inside
both the numerical and the symbolic world, and in-
ference processes. An inference process provides a
state transformation by generating new or manipu-
lating data. If data(i) denotes the knowledge base
and already achieved intermediate results of state ¢
and T — (T,,..., TN) is the set of transformations,
the complete interpretation process can be outlined
by a search tree as depicted in Fig. 3. The initial
state includes the input pattern and a final state its
description. In general, several transformations can
be applied to a certain state. They compete with
each other and the successful and optimal sequence
of transformations forms a path in the search tree.
714
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
I
A a. &
( Yan al )
A TES zT
Bf
Figure 3: Search Tree of an Interpretation Process
One of the major problems in dealing with such sys-
tems concerns their architecture and functional or-
ganization. A widely used bases is the decompo-
sition of knowledge based systems into functional
modules. An adaptation to pattern interpretation
task is shown in Fig. 4. A centralized control
module supervises the process by activating suit-
able transformations and methods with respect to
| USERINTERFACE |
L -— e:
À
2 Y
| | CONTROL ||
| | | | | |
| | METHODS | | KNOWLEDGE | | EXPLANATION | | LEARNING | |
| | | I ]
| RESULT |]
| X i. |
PUER et
input signal f(x) | description |
MÀ Á— LI IL——— ————— ——— ÀJ
Figure 4: Functional Modules of a Knowledge Based
System
the achieved intermediate results and the knowledge
base. Further modules for knowledge acquisition,
explanation, and user interfaces complete the archi-
tecture. An orthogonal viewpoint addresses the hi-
erarchy of processing steps. Fig. 5 shows this model.
Each level is characterized by the knowledge and
processes available and the results which can be
achieved at this processing step. Again, a central-
ized control module is used. It should be mentioned
that this model only determines hierarchies but not
necessarily the direction of information flow. Data
as well as expectation guided strategies can be im-
plemented.
Organization and use of knowledge is not reflected
by both approaches. It is often assumed that a cer-
tain problem domain forms a homogeneous uniform
type of knowledge or that it can be separated into
hierarchies or functional units. Contrarily, inquiries
on knowledge representation point out that different
types of knowledge must be distinguished. Except
declarative models and procedural knowledge, types
like definitions, descriptions, and constraints must
be separated. Furthermore, each such type provides
its own inferences and consequently its results. Each
le
le
| de:
{le
one
ion
age
for
are
of «
Str:
tril
pro
SCO.
on
hin
whi
ter:
Thi
for
scri
the
anc
SCO!
stri
