REAL-TIME RECEPTIVE FIELDS: THE GENERIC NEIGHBORHOOD OPERATION SCALE SPACE PROCESSOR (GNOSSP)

THE PROPOSED INNOVATION IS REAL-TIME HARDWARE FOR IMAGE PROCESSING, IMAGE UNDERSTANDING, AND MACHINE VISION THAT WILL ACHIEVE SIGNIFICANT PERFORMANCE INCREASES OVER CURRENT SYSTEMS BY IMPLEMENTING THE REDUCED SET OF OPERATIONS REQUIRED TO SUPPORT A SPECIFIC SET OF ALGORITHMS. THESE ALGORITHMS ARE THE "GENERIC NEIGHBORHOOD OPERATIONS" OF KOENDERINK AND VAN DOORN WHICH PROVIDE A FRAMEWORK FOR DESCRIBING THE LOCAL STRUCTURE OF OBJECTS IN IMAGES AT A POINT. THE GOAL OF PHASE I OF THIS PROJECT IS TO PERFORM THE TECHNICAL RESEARCH AND DETAILED ENGINEERING DESIGN OF A SYSTEM TO IMPLEMENT THIS THEORY OF GENERIC NEIGHBORHOOD OPERATORS IN REAL-TIME. THIS PROJECT IS CONCERNED WITH PERFORMING THE ENGINEERING DESIGN FOR REAL-TIME HARDWARE FOR IMAGE PROCESSING AND IMAGE UNDERSTANDING. IN ITS FINAL IMPLEMENTATION, IT WILL ACHIEVE ORDERS OF MAGNITUDE PERFORMANCE INCREASES OVER CURRENTLY AVAILABLE SYSTEMS BY IMPLEMENTING ONLY A REDUCED SET OF OPERATIONNS REQUIRED TO SUPPORT A SPECIFIC SET OF ALGORITHMS. IN PARTICULAR, IN ITS FINAL CONFIGURATION, IT MAY PROCESS AT UP TO 138 G OPERATION PER SECOND.