SAR despeckling via classification-based nonlocal and local sparse representation

Abstract

Nonlocally centralized sparse representation (CSR) is an effective approach for estimating original image in noise. In order to promote sparse coefficients more accurate than CSR, we present a new framework where another nonlocal sparsity constraint term is introduced to work with the original term alternatively. To gain the two nonlocal sparsity constraint terms, we classify the image into different types according to the statistical characteristics of speckle in SAR image firstly. Then, we choose the appropriate methods to denoise different types of image and utilize the projected coefficients of these denoising results to estimate the nonlocal sparsity constraints. Our method not only modifies the well-salgortudied method of CSR, but also applialgories to despeckle SAR image. Experimental results, carried out on both simulated SAR images and real SAR images, demonstrate that the proposed approach has a competitive despeckling performance in terms of both evaluation indicators and visual quality assessment.