| In this dissertation we present two feature-based pattern classifiers with application to synthetic aperture radar, a template-based geometric hashing classifier and a model-based Bayesian feature classifier.; First, we develop a classifier that exploits geometric locations of SAR scattering centers. The classifier is an extension of a geometric hashing method for optical images. We synthesize a hash table that can be efficiently searched for matching patterns. The hashing classifier can be used as a final classifier, and also as an index stage to generate a hypothesis list for a final classifier. We present classification results using features extracted from X-band SAR images of four vehicles.; Second, we present model-based Bayesian feature matchers using attributed scattering center features. The matcher attempts to find the Bayes optimal match between an extracted feature vector and a set of predicted feature vectors, and incorporates uncertainty in both predicted and extracted feature vectors. We implement the matcher, and conduct a number of classification performance estimation experiments on a ten class problem using measured X-band SAR images. We compare performance versus number of feature attributes, uncertainty on the attributes, and correlation on attribute uncertainty. We also consider classification performance as a function of resolution. Finally, we present an analytical performance model that predicts classification performance of the Bayes matcher under some simplifying assumptions. |
