SAR Automatic Target Recognition Using Maximum Likelihood Template-based Classifiers

摘要

A review of several recently-developed maximum likelihood template-based automatic target recognition (ATR)algorithms for extended targets in synthetic aperture radar (SAR) imagery data is presented. The algorithms arebased on 'gradient' peaks, 'ceiling' peaks, edges, corners, shadows, and rectangular-fits. A weight-based Bayesianmaximum likelihood scheme to combine multiple template-based classifiers is presented. The feature weights arederived from prior recognition accuracies, i.e., confidence levels, achieved by individual template-based classifiers.Application of feature-based weights instead of target specific feature-based weights reduces the resulting ATRaccuracy by only a small amount. Preliminary results indicate that (1) the ceiling peaks provide the most target-discriminating power, (2) inclusion of more target discriminating features leads to higher classification accuracy.Dempster-Shaffer rule of combination is suggested as a potential alternative to the implemented Bayesian decisiontheory approach to resolve conflicting reports from multiple template-based classifiers.