OPTIMIZING MEMORY HIERARCHY IN MACHINE VISION COMPUTERS

THIS WORK PROPOSES AN INVESTIGATION AND CHARACTERIZATION OF IMAGE PROCESSING ALGORITHMS AND THEIR RELATIONSHIP TO IMAGE REPRESENTATION SCHEMES. CLASSIC IMAGE PROCESSING FUNCTIONS MANIFEST SIGNIFICANT DIFFERENCES IN MEMORY ACCESS PATTERNS, I.E. SPATIAL AND TEMPORAL LOCALITY, FROM THE STANDARD 'GENERAL PURPOSE' ROUTINES WHICH HAVE HISTORICALLY BEEN USED AS THE METRICS OF OPTIMAL SYSTEM DESIGN. IMAGE PROCESSING ALGORITHMS HAVE NOT BEEN CHARACTERIZED AS A SEPARATE CLASS OF FUNCTIONS EXECUTING IN A UNIPROCESSOR SYSTEM. IT IS POSTULATED THAT IMPORTANT INCREASES IN PERFORMANCE CAN BE OBTAINED BY CUSTOMIZATION OF THE COMPUTING ARCHITECTURE. THIS ENHANCED PERFORMANCE WILL BE OBTAINED WITHOUT RESORTING TO HIGHLY PARALLEL SYSTEMS AND THE RESULTING PROBLEMS OF DECREASED MEAN TIME BETWEEN FAILURES AND HIGHLY COMPLEX PROGRAMMING ISSUES. ITEMS TO BE STUDIED ARE MEMORY HIERARCHY AND CACHING STRATEGIES AS WELL AS STORED IMAGE REPRESENTATION SCHEMES. USING STOCHASTIC MODELS OF BEHAVIOR AN OPTIMAL IMAGE PROCESSING ARCHITECTURE AND STORED IMAGE REPRESENTATION SCHEME WILL BE DESIGNED.