Sparse Representation via Gaussian Scale Mixture Model for SAR Image Denoising

In order to further study the noise removal problem of SAR images, a new sparse representation denoising algorithm based on a Gaussian scale mixture model is proposed. We first transform multiplicative noise into the logarithmic domain, and introduce sparse representations with non-local similarity to capture essential features. In this algorithm, each sparse coefficient is modeled as a Gaussian distribution with a positive scaling variable, and imposes a sparse distribution prior. The variance of the sparse coefficient (Gaussian scalar multiplier field), as the latent variable, can be estimated jointly with the unknown sparse coefficient. The performance of the algorithm, which outperforms some other algorithms, is tested through experiments.