ADVANCED AI TECHNIQUES FOR MULTIACUITY ATR

A FUNDAMENTAL PROBLEM IN AUTOMATIC TARGET RECOGNITION (ATR) IS THE OVERWHELMING AMOUNT OF SENSOR DATA WHICH MUST BE PROCESSED. MANY ADVANCED ATR ALGORITHMS WHICH REDUCE THE FALSE ALARM RATE CANNOT BE EXECUTED IN REAL TIME DUE TO COMPUTATIONAL COMPLEXITY. THIS LIMITS OVERALL SYSTEM PERFORMANCE. IN ATR, THE FEATURES THAT MUST BE RESOLVED ARE LOCALIZED WITHIN THE FIELD-OF-VIEW (FOV) OF THE SENSOR. UNIFORMLY SAMPLING WITHIN THE FOV IS THUS INAPPROPRIATE; REGIONS WITH LITTLE OR NO RELEVANCE TO THE TASK ARE SAMPLED AT THE SAME RESOLUTION AS KEY FEATAURES, OCCUPYING VALUABLE SIGNAL BANDWIDTH AND COMPUTATIONAL RESOURCES, AND INCREASING SYSTEM LATENCIES. A NEW CLASS OF MACHINE VISION SYSTEMS, CALLED FOVEAL SYSTEMS, IS PROPOSED FOR THE AUTOMATIC RECOGNITION OF STRATEGIC TARGETS. FOVEAL SYSTEMS FEATURES IMAGING SENSORS AND SIGNAL PROCESSING WITH GRADED ACUITY ANALOGOUS TO BIOLOGICAL VISION. FOVEAL SYSTEMS OPERATE MUCH MORE EFFICIENTLY BECAUSE RESOLUTION IS TREATED AS A DYNAMICALLY ALLOCATABLE RESOURCE. THE DEVELOPMENT AND ANALYSIS OF SELECTED ADVANCED FOVEAL ATR TECHNIQUES IS PROPOSED. THE TECHNIQUES INCLUDE HIERARACHICAL KNOWLEDGE REPRESENTATION AND DATA STRUCTURE/ MULTIPROCESSOR DESIGN, ADVANCED GAZE CONTROL, AND VARIABLE ACUITY FEATURE EXTRACTION, INCLUDING NEURAL NET BASED MULTIRESOLUTION CLASSIFIERS. OVERALL SYSTEM PERFORMANCE WILL BE COMPUTED ANALYTICALLY AND THROUGH SIMULATIONS USING ACTUAL IMAGING SENSOR DATA.