Multi-dictionary learning with superpixel-based clustering for SAR Image despeckling

In this paper, a despeckling method is proposed to reduce the speckle noise from synthetic aperture radar images, which is based on the sparse model. First, we use logarithmic function to transform speckle noise to additive noise in order to simple the despeckling process, while the image is transformed to logarithmic domain. Then, simple linear iterative clustering, which has the merits of quick computation speed and high accuracy among superpixel algorithms, is employed to segment the log-intensity image. Considering the difference of texture and edge, the transformed image is distinguished into different regions by using fuzzy c-means to analyze superpixels. Final, the output image is obtained by multi-dictionary learning, which homogeneous regions of original image are represented by corresponding dictionary. From simulated and actual experiments, the proposed method can provide the better performance than other competing methods.