Parallel Techniques for Rule-Based Scene Interpretation

摘要

We consider a parallel, rule-based approach for learning and recognition of pattern and objects in scenes. Classification rules for pattern fragments are learned with objects presented in isolation and are based on unary features of pattern parts and binary features of part relations. These rules are then applied to scenes composed of multiple objects. We present an approach that solves, at the same time, evidence combination and consistency analysis of multiple rule instantiations. Finally, we introduce an extension of our approach to the learning of dynamic patterns.