Fred Garber (Committee Member), Mark Oxley (Committee Member), Michael Raymer (Committee Member), Brian Rigling (Advisor), Mateen Rizki (Committee Member)
Doctor of Philosophy (PhD)
The difficulty of designing automatic target recognition (ATR) systems is that there are many sources of potential variation in the data, often referred to as operating conditions (OCs). Typical evaluation methodologies rely on empirical simulations on fixed datasets, where it can be difficult to fully sample the variations the algorithm will see in application. Here we focus on analytic performance prediction approaches to quantization based algorithms where the sources of variation are assumed conditionally independent. We have focused on three algorithms in particular: multinomial pattern matching (MPM), quantized grayscale matching (QGM), and a quantized mean-squared error approach (QMSE). The first two are known as model-based ATR algorithms and assume that in-class images are the result of realizations of a statistical model with class-conditional parametrizations. The last is a template- based algorithm which assumes a deterministic "mean" image is available with which to compare candidate targets. We then study performance prediction approaches to these algorithms under a baseline AWGN noise case applicable to both SAR and IR imagery, an occlusion case applicable to IR imagery, and an individual point response (IPR) case applicable to SAR imagery.
Department or Program
Ph.D. in Engineering
Copyright, all rights reserved. My ETD will be available under the "Fair Use" terms of copyright law.