Optimization Framework for a Multiple Classifier System with Non-Registered Targets

Abstract

A fundamental problem facing the designers of automatic target recognition (ATR) systems is how to deal with out-of-library or non-registered targets. This research extends a mathematical programming framework that selects the optimal classifier ensemble and fusion method across multiple decision thresholds subject to classifier performance constraints. The extended formulation includes treatment of exemplars from target classes on which the ATR system is not trained (non-registered targets). Further, a multivariate Gaussian hidden Markov model (HMM) is developed and applied using real world synthetic aperture radar (SAR) data comprised of ten registered and five non-registered target classes. The framework is exercised in an experimental design across classifier fusion methods, prior probabilities of targets and non-targets, correlation between multiple sensor looks, and levels of target pose estimation error.