ADAPTIVE IMAGE ENCODING AND CLASSIFICATION USING NEURAL NETWORKS

WE INTEND TO EXPLORE THE VIABILITY OF NEURAL NETWORKS FOR IMAGE COMPRESSION AND PATTERN CLASSIFICATION IN AN INTEGRATED SYSTEM. THIS IS A NEW APPROACH TO IMAGE COMPRESSION THAT HAS SEVERAL ADVANTAGES OVER STANDARD APPROACHES. SO FAR, NEURAL NETWORKS HAVE BEEN SHOWN TO BE COMPARABLE TO STANDARD TECHNIQUES (COTTRELL, MUNRO & ZIPSER,1987) FOR THE TASK OF IMAGE COMPRESSION. THEY SHOWED THAT A 3 LAYER BACK PROPAGATION NETWORK (RUMELHART, HINTON & WILLIAMS, 1986) WITH ESSENTIALLY NO TUNING COULD ACHIEVE LEVELS OF COMPRESSION ON THE ORDER OF 1 BIT PER PIXEL (BPP). WE INTEND TO EXTEND THAT WORK IN AN ATTEMPT TO ACHIEVE COMPRESSION RATES BELOW 1 BPP AND INVESTIGATE ITS USEFULNESS IN NASA APPLICATIONS. IF THE ALGORITHM CAN BE EMBEDDED IN HARDWARE, THE POTENTIAL ADVANTAGES FOR SPACE-BASED APPLICATIONS ARE: A COMPRESSION DEVICE THAT ADAPTS TO THE CURRENT ENVIRONMENTAL AND HARDWARE CONDITIONS AND OPERATES IN REAL TIME, LESS SENSITIVITY TO CHANNEL ERRORS, A RECONFIGURABLE PATTERN CLASSIFIER THAT CAN BE TRAINED IN SITU.