Neumann - 9
information about the roles which some object may possibly play. If an object
participates in many part-of relationships, i.e. if the role is 1-to-many, there
are many aggregate candidates and the inference power may be small just
as in situations where partial evidence is inconclusive. But if the role singles
out only one aggregate, this information may speed up further processing
considerably.
Aggregate consistent with partial evidence
The usefulness of hypotheses-generating inference services can be further
improved by providing a ranking between hypotheses by means of a
measure of likelihood. Carried out with all consequences, this would be
equivalent to providing a probabilistic framework for classification. How this
can be done in a manner consistent with a terminological system, is a topic
of ongoing research [Jaeger 94].
As a less ambitious step into this direction we suggest to exploit A-box
statistics. The A-box of a terminological system contains assertions about
individual objects, in this case about the objects encountered in concrete
image interpretation tasks. In the absence of other sources of statistical
information, A-box statistics provide a valid basis for generating expectations
about the contents of new images. In particular, given partial evidence for an
unknown object, a likelihood ranking of candidate hypotheses can be
provided based on past experience.
Note that hypotheses ranking is nothing out of the ordinary for any odd
image interpretation system. By suggesting this service for image
interpretation in a terminological framework, we try to provide minimal
services which such a framework must offer to be competitive.
4. Conclusions
In this contribution we have taken a close look at terminological systems and
their use for image interpretation. Illustrated by an extended example taken
from [Lange and Schroder 94] we have shown that fairly complex changes
in aerial images can be precisely defined and automatically recognised
using the object classifier of the terminological system. Thus, in principle, the
construction of image interpretation systems can be greatly facilitated by
employing standardised knowledge representation and reasoning services