Neural network design on the SRC-6 reconfigurable computer

Abstract

This thesis presents an approach to image classification via a Multi-Layer Perceptron (MLP) Artificial Neural Network (ANN) on the SRC-6 reconfigurable computer for use in classifying Low Probability of Intercept (LPI) radar emitters. The rationale behind the previously unexplored use of new reconfigurable computers combined with neural networks for this application is the potential for near real-time classification. Current potential near-peer competitors have access to LPI technology, so development of quick classification methods is crucial for ships to determine intent and to enable the possibility for self-defense against these types of emitters. The neural network, based on work conducted by Professor Phillip E. Pace of the Naval Postgraduate School (NPS), generates integer-cast weights by first using a sequential processor to conduct floating-point backpropagation to train the network on potential timefrequency images that allows generation of weights with lower overall Root Mean Squared (RMS) errors. The weights are then used in a parallel-processing reconfigurable computer for close to real-time classification. A second method of direct pixel comparison using Exclusive-Or (XOR) logic is presented as an alternative image classification method. Comparisons to similar representations in C++ are provided, for use in judging comparative error levels and timing between parallel and sequential processing methods.