Target characterization and matching in synthetic aperture radar imagery

Synthetic Aperture Radar (SAR) target characterization is a challenging area of focus in the field of computer vision. Typical image processing techniques are not directly applicable to SAR imagery which lack the typical features present in visual images. This thesis begins by developing the necessary tools for analyzing SAR images. These include image registration with respect to various operating parameters and the modeling of noise. Multiple algorithms are presented for segmenting targets from surrounding clutter. These methods are statistical in nature which rely on Constant False Alarm Rate detection, or on a histogram based approach. An algorithm for the determination of target aspect is presented which exhibits robustness to significant errors in the target segmentation process. Feature extraction is then developed by the use of topographic features which are invariant to monotonic intensity transformations. These features are shown to be more stable than other features presented in the literature. Target classification is accomplished by formulating the matching of previously developed features in the framework of a nonlinear objective function. The optimization of this objective yields a recognition engine for the SAR automatic target recognition system which is robust to spurious, missing and migrated features. Experimental results are presented for simulated as well as real SAR imagery.