MODEL BASED SYNTHETIC DISCRIMINANT FUNCTIONS FOR PATTERN RECOGNITION

THE SYNTHETIC DISCRIMINANT FUNCTION (SDF) OFFERS REAL TIME DISCRIMINATION OF HIGH-VALUE TARGETS FROM OTHER OBJECTS AND CLUTTER. GIVEN THE ARMY MICOM MISSION, SDFS PROMISE GREATER PERFORMANCE OVER RETICLE-BASEDOR SIMPLER IMAGING SEEKER DESIGNS. AN SDF MUST BOTH DISCRIMINATE EFFECTIVELY AND LOCATE TARGET (CENTROID) POSITION ACCURATELY. JOINT OPTIMIZATION OF THESE TWO OBJECTIVES WILL BE EXPLORED BY TESTING THREE TYPES OF SDFS. THE SYNTHESIS OF THESE SDFS WILL BE FROM MODEL GENERATED IMAGE VECTORS, RATHER THAN FROM FIELD DATA. THE PHASE I VALIDATION OF THE MODEL BASED APPROACH TO SDF SYNTHESIS WILL MOTIVATE FOLLOW-ON EFFORTS TO EXPLOIT ITS MAJOR ADVANTAGES. THE USE OF A SYNTHETIC TARGET IMAGE GENERATION MODEL, E.G., AERODYNE'S SPIRITS AS THE FOUNDATION OF MODEL BASED SDF SYNTHESIS WOULD ALLOW GENERATION OF TRAINING VECTORS FOR ARBITRARY ASPECT, RANGE, ILLUMINATION, AND TARGET OPERATING CONDITION. THIS ENABLES GENERATION OF AN OPTIMAL REPRESENTATIVE TRAINING SET, ELIMINATING THE NEED TO ACQUIRE COSTLY MEASURED TARGET DATA VECTORS. SUCH MODEL WOULD BE COUPLED WITH A PROPOSED OPTIMAL TRAINING VECTOR SELECTION TECHNIQUE, USING TRADITIONAL METHODS OR MASSIVELY PARALLEL HOPFIELD NEURAL NETWORKS TO SOLVE AN EQUIVALENT GRAPH BIPARTITIONING PROBLEM. THE TECHNIQUE IS PARTICULARLY EFFECTIVE FOR LARGE SETS OF CANDIDATE VECTORS.