An image representation algorithm compatible with neural-associative-processor-based hardware recognition systems

A robust image representation algorithm compatible with the VLSI-matching-engine-based image recognition system has been developed. The spatial distributions of four-principal-direction edges in a 64 × 64-pels gray scale image are coded to form a 64-dimension feature vector. Since the 2D edge information is reduced to a feature vector by projecting edge flags to the principal directions, it is named the projected principal-edge distribution (PPED) representation. The PPED vectors very well preserve the human perception of similarity among images in the vector space, while achieving a substantial dimensionality reduction in the image data. The PPED algorithm has been applied to medical radiograph analysis, which was taken as a test vehicle for algorithm optimization. The robust nature of the PPED representation has been confirmed by the recognition results comparable to the diagnosis by experts having several years of experience in a university hospital. Dedicated digital VLSI circuits have been developed for PPED vector generation in order to expedite the processing. A test hardware recognition system was constructed using the vector generation circuits, where the analog neural associative processor chip developed in a separate project was employed as a vector-matching engine. As a result, a successful medical radiograph analysis has been experimentally demonstrated using the hardware system. Feasibility of a very low-power operation of the system has been also demonstrated.