Stefan Growe
scription of the scene observed in one or more images, sometimes from different sensors. Information about object states
and its possible changes over time can be integrated within the semantic net in form of a state transition graph. This tempo-
ral knowledge is used for the interpretation of multitemporal images to improve the explanation of land-use changes or
to detect complex patterns due to a typical behaviour over time observed in the data set.
The image analysis is controlled by a rule-based inference engine, which documents competing scene interpretations in
the leaf nodes of a search tree. To optimize the path through this search tree the alternatives are judged and the most prom-
ising one is investigated first. For the comparison of the intermediate interpretation results a common judgement calculus
is needed which evaluates to which degree the measured object properties match to the expectations derived from the
knowledge base.
In this contribution a probabilistic judgement calculus for the AIDA system is presented which is based on Bayesian net-
works. For the interpretation of multitemporal images it is shown that the approach causes a more efficient search
compared to an existing judgement approach, if additional information about the probabilities of events is provided. The
paper is organized as follows: After a brief introduction in the AIDA system the representation and use of temporal knowl- |
edge is described. Thereafter a short excursion into the theory of Bayesian networks is given followed by a discussion
how it is used to judge a scene interpretation represented by a semantic net. Finally results are shown for the detection
of an industrial fairground from a set of multitemporal images.
2. SYSTEM OVERVIEW
The architecture of the knowledge based image interpretation system AIDA has already been described in numerous pub-
lications (e.g. (Tónjes et al., 1999), (Tónjes, 1999b), (Liedtke et al., 1997)). For this reason only a short introduction is
given here. The knowledge about expected scene objects is defined prior to the analysis in a separate knowledge base.
By exchanging the knowledge base the system can easily be adapted to varying application tasks without modifying the
interpretation module itself. This flexibility is the main advantage of a knowledge based approach. From the prior knowl-
edge, hypotheses about the appearance of the scene objects are generated which are verified in the sensor data. Additional
domain specific knowledge like GIS data (geographic information system) can be used to strengthen the interpretation
process. An image processing module extracts features that meet the constraints given by the expectations. It returns the
found primitives — like line segments — to the interpretation module which assigns a semantic meaning to them, e.g. road
or river. The system finally generates a symbolic description of the observed scene. In the following, the knowledge repre-
sentation and the control scheme of AIDA is described briefly.
2.1 Knowledge Representation
The knowledge base is formulated by a semantic net. The nodes of the net, called concepts, represent generic prototypes
of the expected scene objects, like roads, rivers or buildings. Realizations of the concepts detected in the scene during
analysis are documented in the semantic net by new nodes called instances. The process of their generation is named
instantiation. While an object is modelled by only one concept in the knowledge base, there might exist several instances
of this object in the scene. During interpretation four different states of object recognition are distinguished: hypotheses,
partial instances, complete instances and missing instances. The object properties are described by attributes attached
to the concepts. Attributes possess a value derived from measurements in the data and an expected range of values which
mirrors the expert knowledge. The expectations are restricted consecutively during analysis due to the current intermedi-
ate results. Computation functions are used to determine the attribute values and ranges from the sensor data or other
instances at run—time.
The nodes are connected by edges to form a semantic network. The edges represent the structural, topological and tempo-
ral relations between the objects. The specialization of objects is described by the is—a relation along which the more spe-
cial concept inherits all properties of the more general one. The decomposition of objects in their components is repre-
sented by the part—of link. Via the concrete—of link (abbreviated con—of) an abstract description is transformed into its
more concrete representation in the data. For example the symbolic term “road” is connected to the primitive “line” to
define its geometrical appearance in the image. The concrete—of relation structures the knowledge base into different con-
ceptual layers like for example a symbolic layer, a geometry layer, and a material layer. Topological relations provide
information about the kind and the properties of neighbouring objects. Therefore, the class of attributed relations (attr—
rel) is introduced. In contrast to other relations, this one may possess attributes, which are used to constrain the properties
of the connected nodes. For example, a topological relation close—to can be generated which restricts the position of an
object to its immediate neighbourhood. The initial concepts which can be extracted directly from the data are connected
via the data—of link to the primitives segmented by image processing algorithms. Especially for the representation of tem-
poral knowledge the temporal relation is introduced which describes temporal changes of objects. In Chapter 3 the analy-
sis of multitemporal images is discussed in detail.
For the efficient representation of multiple relations, the minimum and maximum number of edges can be defined for
a relation. The minimum quantity describes the number of obligatory relations and the difference to the maximum quanti-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 343
