Advanced Algorithms for Exploitation of Space-Based Imagery

Physical Sciences Inc. and its subcontractor, SAIC-Boulder, propose to implement and evaluate two novel pattern recognition methods to mitigate cluster noise and enhance the contrast between targets and backgrounds in an automatic target recognition (ATR) applications. The proposed methods are relevant to subpixel target detection using hyperspectral data and are compatible with real-time implementation on airborne and spaceborne operating platforms. Our approach is to integrate a statistically-robust blind source separation algorithm for spectral signature recognition with a Bayesian Evidential Reasoning framework to enable context-based false alarm mitigation. The Bayes net will be used to reduce false rate by evaluating the spectral characteristics of the region surrounding the potential target detect. The proposed spectral pattern recognition approach is robust with respect to the form of clutter noise in the data, i.e., non-Gaussian noise statistics, and will enable modeling of Receiver Operating Characteristic (ROC) curves for any user-specified clutter noise distribution. The Phase I program will involve benchmark testing of the proposed methods using real and synthetic data sets and will enable recommendation of an ATR approach to be implemented in hardware in Phase II.