Human-perception-like image recognition system based on the Associative Processor architecture

A human-perception-like image recognition system has been developed aiming at direct hardware implementation based on the Associative Processor architecture. The Principal Axis Projection (PAP) method [1] has been employed for representing images for processing using Associative Processor chips [2-4]. The PAP vectors very well preserve the human perception of similarity among images in the vector space and ideal for use in our system. In this study, the system has been applied to medical radiograph analysis as well as binary image recognition including corrupted handwritten patterns. By introducing two new techniques, namely, “winner score/pattern mapping” and “macro-scale matching”, the recognition performance has been drastically improved as compared to our previous work [1]. By utilizing the macro-scale matching technique, the percent correct for three important Cephalometric landmarks (Nasion, Orbitale and Sella) has been improved to be 97.5%, 97.5%, and 87.5 % from 87.5%, 85%, and 62.5%[1], respectively. In order to expedite the PAP vector generation processing, dedicated VLSI chips have been developed and their proper operation has been experimentally demonstrated.