An analysis of knowledge representation schemes for high level vision

Abstract

This paper analyses the criteria necessary for a knowledge representation (KR) language for implementing high level vision (HLV) recognition systems. We show the importance of introducing a specific KR language for specification, and possibly for implementation of HLV systems. In particular, we examine the adequacy, tractability and suitability of implementing a HLV system using logic, the KR language most commonly used in areas of Artificial Intelligence isomorphic to HLV. In addition, we use this analysis of classical logic to identify the criteria necessary for any HLV KR language. Logic is seen to be at least as good a language for specification of HLV systems as any other KR language. However, using evidence obtained from an object recognition system implemented using propositional logic, evidence which is supported by theoretical analyses, we argue that classical logic is an inadequate KR language for implementing HLV systems. It cannot identify preferred interpretations, and is computationally intractable, even for simple propositional languages.