Hyper parameter estimation in MRF-based SAR chip image segmentation

SAR chip image segmentation is a key step in SAR automatic target recognition (ATR). Aiming at this challenging task, many available methods are proposed. Among these methods, MRF-based segmentation is the most popular one. MRF-based method for SAR chip image segmentation is more like an inverse-filter which tries to smooth noise. The hyper parameter determines its capability of noise reduction. The larger the hyper parameter, the more noise is smoothed. However, when the hyper parameter gets large, some interested regions may be regarded as noise and be smoothed. In order to smooth all noise and preserve interested regions, the hyper parameter should be carefully selected. In this paper, by rewriting the regularization term in MRF-based SAR chip image segmentation, we find its similarity with total variation (TV) filtering. Moreover, some sort of analytic expression of TV regularization hyper parameter has been derived. According to this similarity, we are convinced that these analytic results of hyper parameter estimation in TV filtering may be carefully extended to MRF-based SAR chip image segmentation. A core concept in these expressions is ‘scale’ which refers to the area-perimeter ratio of interested regions. However, when it comes to MRF-based SAR chip image segmentation, the ‘scale’ has to be redefined. In our study, we redefine the ‘scale’ and illustrate scales of some canonical shapes. Finally, these formulations of MRF hyper parameter are validated by simulated data.