3-D TARGET MODELING REPRESENTATION AND PERCEPTION

CURRENTLY, CONSIDERABLE RESEARCH IN MACHINE VISION IS BEING DIRECTED TOWARD PROBLEMS IN PERCEPTUAL ORGANIZATION, WHICH INVOLVED HOW OBJECTS SUCH AS JUNCTIONS, LINES, AND REGIONS, EXTRACTED BY LOCAL PROCESSES, ARE COMBINED TO PRODUCE COHERENT PERCEPTIONS OF OBJECTS AND EVENTS. WE PROPOSE TO DEVELOP PERCEPTUAL ORGANIZATION AS A RECURSIVE ENERGY MINIMIZATION PROCESS USING A HOPFIELD NEURAL NETWORK APPROACH. OUR APPROACH IS BASED UPON THE INITIAL EXTRACTION OF SYMBOLIC TOKENS FROM AN IMAGE OR TIME-SEQUENCE OF IMAGES. THESE ARE THEN REDUCED TO PRIMITIVE TOKENS WHICH CORRESPOND TO THE NODES OF THE NETWORK. THE LINKS BETWEEN PRIMITIVE TOKEN NODES ARE THE VARIABLES OF THE HOPFIELD ENERGY FUNCTIONAL. A LOCAL OR GLOBAL ENERGY MINIMIZATION PROCEDURE IS THEN USED TO FORM STABLE PERCEPTUAL GROUPS OVER THIS NETWORK BY EXTRACTING LINKED CLUSTERS OF TOKENS. HIERARCHICAL PROCESSING IS POSSIBLE BY CREATING LINKS BETWEEN THESE STABLE CLUSTERS AND THEN APPLYING THE MINIMIZATION GROUPING PROCESS TO PRODUCE LARGER COHERENT STRUCTURES. THIS RECURSIVE COMBINATION OF PERCEPTUAL GROUPS CORRESPONDS TO A HIERARCHY OF PROCESSING PERFORMED BY THE NETWORK. OUR WORK WILL INVOLVE DESIGN, IMPLEMENTATION, AND EXPERIMENTAL ANALYSIS ON BOTH ARTIFICIAL AND REAL IMAGERY.