Parallelism in a Semantic Network for Image Understanding

The use of multiprocessor systems in the field of image analysis and understanding is motivated by the huge amount of data and the need for extremly fast processing that is necessary to achieve practical computer vision systems. The development of parallel algorithms for the knowledge-based interpretation of complex scences (patterns) must consider aspects of knowledge representation as well as it’s efficient use. It is complicated by special problems of symbolic image processing, like unreliable segmentation results, variable data dependencies, or the need for integration of different levels of representation.This paper introduces a parallel control algorithm for an image understanding system that uses semantic networks for the representation of task-specific knowledge. The algorithm is based on an explicit representation of all necessary inferences in a single data flowchart. In a bottom-up instantiation the sensor data are used for the computation of competing instances for each element of the knowledge base. A top-down optimization is used for the iterative improvement and selection of hypotheses.