Statistical Characterization of Clutter for Qualification of ATR Performance

The ability to characterize clutter and predict the performance of Automatic Target Detection and Recognition algorithms would be a revolutionary tool for analysis to be adaptively changed to maintain a constant false alarm rate aiding in the design of new detection and classification algorithms, and facilitating the tactical decision making processes of situational awareness, change detection, target identification, battle damage assessment, and retargeting. Our proposed research will investigate new methods of characterizing SAR cluster using wavelet and noparametric bootstrap methods. Wavelet analysis techniques such as wavelet--domain Markov random fields can model correlation in images across different resolutions. Bootstrap methods adaptively estimate the underlying probability distribution of clutter in the non-Gaussian colored-noise environment typical of SAR imagery. We will develop efficient methods of predicting the performance of ATR algorithms in the presence of clutter (i.e., probability of detection, false alarm rate, probability of classification), based on bootstrap methods and the new clutter models. The statistical methodology for clutter modeling and ATR performance prediction will be implemented in a cross-platform software environment suitable for deployment across a wide range of applications.