Interactively Matching Hand-Drawings Using Induction

摘要

In this paper we consider the task of matching patterns, as occur in hand-drawn symbols and schematic diagrams, by their parts and relationships. Of particular interest for computer vision is the integration of two approaches to the recognition by parts problem—graph matching and syntactic rule-based approaches. A new procedure is developed, named CLARET, which matches parts and relationships by tightly coupling the processes of matching and rule generation at run time. We have developed an interactive system for interpreting hand-drawn symbols and schematic drawings. The system operates invariant to rotation, scale, and position and projects images onto a drawing canvas. The procedure is analyzed for its ability to accommodate new symbols and answer orientation queries, and it is compared empirically with machine learning techniques.