Multi-layer graph model based SAR image segmentation with geometric interaction prior

This letter presents a novel synthetic aperture radar (SAR) image segmentation method based on a single graph cut. Due to the speckle, the classical pixel-by-pixel Markov random field (MRF) based SAR image segmentation method may still result in salt and pepper label map. Compared to the classical MRF based SAR image segmentation method, our proposed method exploits geometric prior to describe the interior relationship between each part. The classical MRF based SAR image segmentation method only utilizes data term and smoothness term. With the added geometric prior, the proposed method can overcome the influence of speckle better than the classical MRF based method. In order to use a single graph-cut for optimization, a new multi-layer graph model is utilized to handle the data term, smoothness prior term and geometric prior term. Additionally, we can set different smoothness terms for special kind of ground objects via the multi-layer graph model. This setting can also contribute to better segmentation result. The experimental results obtained on real SAR images show that our approach works well.