Image Processing Reference
Most of the data handled in fusion in the fields of signal and image process-
ing is analog or digital. Analog descriptions require a complete description of the
world. Switching over to logical representations, which are cheaper and more com-
pact, requires converting analog representations into symbolic representations.
Symbolic representations have requirements on several levels:
- the ontological level: all of the important concepts have to be taken into account;
- the epistemic level: we should not have to express what is not known;
- the computational level: the representation should allow for an efficient compu-
tation of the properties expressed.
The first two levels induce constraints on the language and the third on the infer-
The knowledge representation (symbolic) community focuses on non-monotonic
reasoning, automatic reasoning, logic descriptions, subjective representations (prefer-
ences, wishes, etc.), ontologies, etc. [REI 91]. Closer to what concerns us, it also fo-
cuses on learning, the integration and fusion of knowledge bases, decision and diagno-
sis, temporal and spatial reasoning, the representation of actions and planning. There
are definitely directions to explore in this direction for the fusion that concerns us.
5.5. Knowledge-based systems
The evolution of image processing, from the lowest level of signal processing to the
interpretation of complex scenes, quite naturally leads the user to taking into account
knowledge beyond merely the image signal. As a consequence, this brings knowl-
edge management techniques in contact with image processing. These are methods
developed in artificial intelligence and which have been referred to as: expert systems,
knowledge-based systems (KBS), multi-agent systems, etc.
These techniques have had a strong influence on the development of image pro-
cessing techniques. In particular, they contributed to important projects that were part
of European programs. We suggest discussing here the lessons that can be learned
from these methods.
The objectives of these methods are as follows:
- representing knowledge in a declarative and not just procedural way as in the
most common image processing algorithms;
- separating knowledge into different categories:
- factual knowledge separated from operational knowledge,
- particular knowledge separated from general knowledge;
- using the same knowledge to achieve different objectives;