Segmentation-based detection of targets in foliage penetrating SAR images

Segmentation and labeling algorithms for foliage penetrating (FOPEN) ultra-wideband Synthetic Aperture Radar (UWB SAR) images are critical components in providing local context in automatic target recognition algorithms. We develop a statistical estimation-theoretic approach to segmenting and labeling the FOPEN images into foliage and non-foliage regions. The labeled maps enable the use of region-adaptive detectors, such as a constant false-alarm rate detector with region-dependent parameters. Segmentation of the images is achieved by performing a maximum a posteriori (MAP) estimate of the pixel labels. By modeling the conditional distribution with a Symmetric Alpha-Stable density and assuming a Markov random field model for the pixel labels, the resulting posterior probability density function is maximized by using simulated annealing to yield the MAP estimate.