Using discrimination graphs to represent visual knowledge

Abstract

Most model-based vision systems use specialization hierarchies to represent interpretations that are hypothetical and ambiguous. Such hierarchies have the disadvantage that local interpretations possible for a single image feature cannot generally be captured in a single specialization hierarchy. This causes several problems. First, it is difficult to represent hypothetical and ambiguous interpretations along one knowledge representation dimension. Second, the number of hypotheses represented for a single image feature tends to be large. Third, in an interpretation graph competing hypotheses cannnot be represented in the domain of a single variable. As a result, a restructuring of the graph is required, whenever a hypothesis is invalidated. In this dissertation a better solution of the problems associated with specialization hierarchies is proposed. Classes of objects which have local features with similar appearance in the image are represented by discrimination graphs. Such graphs are directed and acyclic. Their leaves represent classes of elementary objects. All other nodes represent abstract (and sometimes unnatural) classes of objects, which intensionally represent the set of elementary object classes that descend from them. Rather than interpreting each image feature as an elementary object, we use the abstract class that represents the complete set of possible (elementary) objects. Following the principle of least commitment, the interpretation of each image feature is repeatedly forced into more restrictive classes as the context for the image feature is expanded, until the image no longer provides subclassification information. This approach is called discrimination vision. Our system has been implemented as a program for interpreting sketch maps. A hierarchical arc consistency algorithm has been used to deal with the inherently hierarchical discrimination graphs. Experimental data show that, for the domain implemented, this algorithm is more efficient than standard arc consistency algorithms.