A Distributed Problem Solving Approach To Knowledge Based Sonar Image Interpretation

A modular knowledge based system (KBS) architecture providing a suitable framework for the design and implementation of complex inkormation processing systems is presented, and applied to the task of detecting objects and shadows in sector scan sonar imagery. The architecture has been derived from the conception of a blackboard system, where processing activity is organized to resemble a group of experts reading and writing on a globally accessible blackboard. The realisation described here is cellular, such that increasing amounts of feedback within each module establishes a hierarchy of information processing units. Each module, or rational cell, contains a blackboard, an inference mechanism and a knowledge base, and can manipulate uncertain information and goals (or hypotheses) represented as symbols, propositions, numerical parameters and two dimensional images. Each of these types may be operated on by processes employing rule based or algorithmic expertise to provide competitive or co operative support for the presence of a hypothesis. Confidence attributes are also maintained to indicate their validity. For initial experiments, three type 2 cells have been interconnected using a pipeline approach and three networked workstations. It is believed, however, that a hybrid functionally accurate/co-operative and functionally accurate/autonomous approach is the most appropriate for this vision task. The sonar processing knowledge base has successfully been applied to a limited class of real images derived from a UDI AS360 sector scanning device.