Foundation of a Knowledge Representation System for Image Understanding

摘要

The basis of a knowledge representation system is presented that is able to recognize real-world objects from partial information delivered by human or mechanical 'experts', each of which is assumed to have its own information-processing tasks. The system deals directly not with the real world, but with the outputs of the experts' processing, and consists of three component parts, a descriptional component, a category component, and a functional component, in each of which knowledge is structured and accessed from general to specific, making it possible to access exactly as much information as is desired at an appropriate level of detail. Uncertainties are dealt with through the use of possibility theory, which also provides a means for approximate pattern matching. The knowledge representation language includes the use of trivalent quantifiers, whose semantics is explained and elaborated. Areal coordinate systems based on hexagons and squares are examined as possible alternatives to the standard use of bands in image processing. (Author)