Group Sparse Representation Based Dictionary Learning for SAR Image Despeckling

Since the sparse representation coefficients of synthetic aperture radar (SAR) images often appear in clusters with intrinsic structure, traditional sparse representation theory cannot capture this property. In this paper, the concept of group sparse representation (GSR) is utilized to exploit the intrinsic structure of SAR images. Different from traditional patch-based sparse representation theory, GSR is able to sparsely represent images in the domain of group which contains the image patches with similar structure. Based on the multiplicative speckle noise model, a novel dictionary learning algorithm based on GSR (GSR-DL) for SAR image despeckling is proposed. The proposed algorithm mainly consists of three steps. First, in order to realize the recovery of despeckled SAR image by the GSR model, a mean filter is included in the modeling process. Second, the proposed GSR-DL algorithm is used to calculate the optimal dictionary and group sparse representation coefficients. Third, the despeckled SAR image is reconstructed by the learned dictionary and coefficients. The experimental results on SAR images manifest that the proposed GSR-DL algorithm achieves a better performance than other state-of-the-art despeckling algorithms.