THE USE OF NEURAL NETWORKS FOR TARGET IDENTIFICATION

A RECENT SURVEY OF NEURAL NETWORK TECHNOLOGY BY ATLANTIC AEROSPACE PERSONNEL HAS IDENTIFIED THE NECESSITY OF USING A NUMBER OF NETWORK LAYERS TO ADDRESS TYPICAL PROBLEMS IN TARGET IDENTIFICATION. IN ADDITION TO UPDATING THIS SURVEY, THIS EFFORT WILL DEVELOP AND DEMONSTRATE A NEURAL NETWORK APPROACH TO TARGET RECOGNITION BASED ON A NEW CLASS OF DISTRIBUTED, ITERATIVE IMAGE OPERATORS MODELED AFTER THE PHYSIOLOGY OF HUMAN VISION, WHICH WILL COMBINE PORTIONS OF THE LOW AND MEDIUM LEVEL VISION TASKS INTO A SINGLE INTEGRAL NETWORK. THE GOAL FOR THIS RESEARCH IS TO DEVELOP AND EVALUATE A NEURAL NETWORK PARADIGM BY WHICH AN INITIAL SET OF SINGULARITIES IN THE VISUAL FIELD (JETS) DERIVED DIRECTLY FROM THE IMAGE DATA, IS REFINED TO BEST MATCH PRESCRIBED VISUAL AND GEOMETRIC CONSTRAINTS, VIA AN ITERATIVE NETWORK RELAXATION TECHNIQUE. THIS OPERATION WILL HAVE THE EFFECT OF EXTRACTING VISUALLY SIGNIFICANT FEATURES FROM IMAGERY AND INHERENTLY GROUPING THEM IN A MANNER NORMALLY ACHIEVED BY POST PROCESSING TECHNIQUES. WE BELIEVE THAT THIS APPROACH IS UNIQUE AND ADDRESSES A PRIMARY NEED IN THE APPLICATION OF NEURAL NETWORK TECHNOLOGY TO IMAGE ANALYSIS IDENTIFIED IN OUR PREVIOUS SURVEY, I.E., THE SELECTION OF APPROPRIATE IMAGE FEATURES TO SUPPORT TECOGNITION OBJECTIVES. WHILE THE PROPOSED TECHNIQUE IS EXPECTED TO RESULT IN FEWER NETWORK LAYERS FOR LOW AND MEDIUM LEVEL VISION TASKS, IT WILL STILL BE NECESSARY TO PROVIDE ADDITIONAL PROCESSING TO SUPPORT HIGHER LEVEL FUNCTIONS SUCH AS TARGET IDENTIFICATION. ACCORDINGLY, THE UPDATED NEURAL NETWORK SURVEY WILL SPECIALLY EMPHASIZE THOSE NETWORK PARADIGMS THAT ARE APPROPRIATE TO HIGH LEVEL VISION PROCESSING TASKS.