Model-Based Fusion of Multiple Look SAR for Automatic Target Recognition

"In our Phase I study we developed and evaluated several initial approaches to fusing multiple-looks for SAR combat ID. Extensive evaluation of decision-level multi-look fusion showed there were significant performance benefits compared with single lookID. In Phase I of this SBIR, we developed and tested an initial, aspect-constrained hypothesis-level fusion, where pose parameters become part of the hypothesis being evaluated along with vehicle type. In Phase II, we propose to extend this byimplementing and testing aspect and translation-constrained hypothesis-level fusion. We will also implement and test feature-level fusion, where we accumulate evidence over regions of the model, thereby correctly accounting for model region visibilityacross the multiple views. As we increase the fidelity of the multi-look fusion approaches, we also require finer image registration requirements. To support accurate registration we propose to apply our hierarchical pixel/feature/region registrationalgorithms, which have proved to be effective on related applications. In order to analyze the performance tradeoffs of the multi-look techniques and understand their benefits and limitations, we will perform extensive analysis on available in-housemulti-look MSTAR 1' SAR imagery covering a broad range of operating conditions as well higher resolution 4" Lynx imagery which we have recently acquired."