Use of a Pyramid Processor in Intermediate-Level Vision (Invited)

Whereas low-level vision consists of filtering and other imageto-image operations, and high-level vision involves matching and inference of symbolic, relational structures such as graphs and frames, the problem area known as "intermediatelevel vision" requires the extraction of features and symbols from a t w e dimensional array of pixels. Conventional serial architectures do not have enough parallelism to handle lowand intermediatelevel tasks very rapidly; at the same time, standard mesh architectures do well with low-level operations but are inefficient for intermediate and high-level problems. A pyramidal architecture, consisting of a hierarchy of meshes, can be more effective at intermediatelevel problems than these architectures. Two powerful algorithm design methodologies are explored: bottomup (reduction) methods, and top-down (refinement) methods. Feature extraction and feature selection are complementary a p proaches to intermediatelevel vision and they are easily imple mented with these methodologies. This is illustrated with a hierarchical Hough transformation and top-down feature-point localization. The programming of pyramid machines poses an interesting challenge; a pyramid-oriented iconic dataflow language, promises to make certain algorithms easy to express.