Adaptive Context-based Configuration for Model-Based SAR ATR

Conventional ATRs are trained and configured off-line. Re-training and reconfiguring for new conditions and new targets are labor-intensive procedures, and there is no mechanism other than robustness to interpolate between the trained conditions. We propose to improve conventional model-based SAR ATR performance under EOCs by incorporating context-dependent learning to adaptively configure the ATR. Learning tasks include model characteristics-such as uncertainty distributions and scattering presence/detectability-as well as algorithmic decisions-such as feature selection, matching metric selection, and hypothesis search rules. Both the model knowledge as well as the algorithmic controls are parameterized by situation-dependent data/scene/image descriptors. In the phase I effort we will initiate an investigation into the following four aspects of this problem: 1) context classification, 2) knowledge representation, 3) the learning metric, and 4) the statistical search space learning mechanism. We anticipate that our enhancements to the MSTAR ATR will enable high recognition performance in untrained conditions and improve performance for trained conditions. We then plan to conduct a comparison between the identification performance of our prototype adaptive model-based ATR and a conventional baseline (manually optimized) ATR on synthetic and actual SAR data representing a realistic range of EOCs.