Learning Model Structures from Images (Special Issue on Computer Vision)

概要

論文の詳細を見る
Based on a newly proposed notion of relational network, a novel learning mechanism for model acquisition is developed. This new mechanism explicitly deals with both qualitative and quantitative relations between parts of an object. Qualitative relations are mirrored in the topology of the network. Quantitative relations appear in the form of generalized predicates, that is, predicates that are graded in their validity over a certain range. Starting from a decomposition of binary objects into meaningful parts, first a description of the decomposition in terms of relational networks is obtained. Based on the description of two or more instances of the same concept, generalizations are obtained by first finding matchings between instances. Generalizing itself proceeds on two levels: the topological and the predicate level. Topological generalization is achieved by a simple rule-based graph generalizer. Generalization of the predicates uses some ideas from MYCIN. After successful generalization, the system attempts to derive a simple and coarse description of the achieved result in terms of near natural language. Several examples underline the validity of relational networks and illustrate the performance of the proposed system.
社団法人電子情報通信学会の論文
1994-11-25