An associative-processor-based mixed signal system for robust grayscale image recognition

Abstract

An associative-processor-based VLSI system architecture has been developed for robust grayscale image recognition. The system receives a 64/spl times/64 pels block of a gray scale image, extracting a feature vector from the image and recognizing the image by template matching. An analog associative processor is adopted as the template matching core because it features compact implementation as well as fast processing due to its fully parallel architecture. For generating feature vectors, dedicated digital CMOS circuits have been developed because of their versatility in the algorithm. The analysis of medical X-ray pictures (Cephalometric landmark identification by expert dentists) was taken as an exercise for the system, and intensive computer simulations have been conducted to optimize the recognition performance of the system. Although the entire system has not yet been implemented on a single chip, all the key sub circuits in the system were fabricated as test circuits and their correct functioning has been experimentally demonstrated. It is also shown by experiment that very low power operation of the template matching core is possible by operating the analog circuitry in the subthreshold regime without degrading recognition performance.