Under the Bayesian restoration framework, this paper aims at the problem of the inadequate accuracy of the sparse solution under the traditional convex regularization constraint, which leads to the loss of texture details and excessive edge smoothing in the recovered image, a new method is proposed. Firstly, the internal nonlocal self-similarity of the degraded image and the external nonlocal self-similarity in the clean dataset are used to construct similar block groups, which are subject to structured sparse coding, and the sparse coefficients are constrained by a weighted l-p penalty function;Secondly, by combining the sparse coefficients obtained in the previous step, an alternating minimization optimization method is designed to iteratively resolve the image restoration equations for computing a reasonable estimation for the original restored image. Simulation experiments verify the correctness and effectiveness of the proposed scheme.

image denoising; group sparsity representation; dictionary learning; weighted l-p norm