Massively Parallel Image Recognition Systems for Remotely Sensed Data

摘要

The design of classical vision systems is based on serially piecing together individual algorithms, each of which is intended to solve a specific part of the vision problem under a given set of assumptions. This has met with poor performance and low information throughput. The design process for each of these algorithms is usually disjoint, and ignores the system integration process. Speculation was made on a new philosophy for the design of vision systems that uses highly parallel and simple elements that are easily integrated. To improve the performance of the simple algorithms involved, heavy use of closed feedback loops is made through the system. These loops have self-correcting capabilities in different time scales. This work suggests a simple system that uses several of these concepts to perform multiple object recognition in noisy conditions. Contents-- Design Procedure for Resonating Algorithms; Special Purpose Feature Based System: Segmentation Algorithm, Problem Formulation; Feature Extraction with Connectist Models: Fourier Descriptor Method, Adaptive Methods for Feature Extraction; Bidirectional Associative Memories. (fr)