Group Sparse Representation for SAR Image Segmentation with Local Linear Constraint

Abstract

Feature plays a vital role in process of synthetic aperture radar(SAR) image segmentation which affects interpretation and understanding. Although sparse Bayesian learning (SBL) provides discriminant feature benefit for segmentation, it suffers from a flaw that operation of raster-scanning image to vector destroys two-dimension structure and losses information. Therefore, group sparse Bayesian learning with local linear constraint is proposed, where the dictionary is construct by multi-scale analysis and learned with neighbor information considered. In this method, the SAR image is primarily decomposed by wavelet in order to obtain multi-scale information. After that, learning features and resolve sparse coefficient alternately by optimizing algorithm. Then, reconstruction error between pixel's feature and reconstruct signal is utilized for image segmentation operation. The experiments with synthetic and real images demonstrate that the proposed method compared with wavelet and gray level co-occurrence matrix features is more effective. Moreover, the learned feature is improved compared with unlearned feature.