Chapter 17 - A Distributed Problem Solving Architecture for Knowledge Based Vision

Abstract

This paper reports ongoing work investigating the use of a hybrid distributed problem solving (DPS) approach to a knowledge based vision task. A hierarchical architecture is employed using a number of interconnected processing agents termed rational cells. Different levels in the hierarchy are arranged to correspond to different levels of processing in the vision pyramid. It is argued that the approach differs from other hierarchical architectures for computer vision because of the way that mutually inconsistent and incorrect hypotheses can be represented throughout the architecture (so called functionally accurate processing), and the way these hypotheses are exchanged when cells interact. Horizontal interaction between cells at a given level in the hierarchy should be cooperative, involving the exchange of inconsistent and possibly incorrect partial hypotheses to resolve uncertainty resulting from disagreement and lack of information. Vertical interaction, however, need not be cooperative, because the vision pyramid contains distinct levels (i.e., is horizontally stratified). Nearly autonomous interaction is sufficient, involving the exchange of locally complete solutions (i.e., complete and correct from an individual cell's point of view). Although solutions passed vertically may be inconsistent when combined at the new level, cooperative interaction at the new level can resolve this. The application focus for the work is the automatic detection and classification of objects and shadows contained in image data derived from an active sector scanning sonar system. This knowledge based vision application is part of a much larger system with distributed problem solving tasks, involving an unmanned, free-swimming submersible vehicle currently being implemented for offshore energy exploration and production activities. After some introduction, the paper discusses the internal structure of the rational cell, the hybrid DPS architecture for knowledge based vision, and some preliminary results investigating processor loading and communication bandwidth for a distributed implementation of a sonar interpretation knowledge base.