Image restoration using spatially variant hyper-Laplacian prior

The heavy-tailed hyper-Laplacian prior has been successfully applied in image restoration tasks, in which the unified distribution is adopted for the whole image. However, the gradient distribution of natural image is reasonably assumed to be spatially variant, e.g., gradient distribution of the region with less texture is more heavy-tailed. In this paper, we propose to model the gradient distribution of natural images as spatially variant hyper-Laplacian. The proposed model adapts the hyper-Laplacian parameters to each pixel. Within the maximum a posterior model, we address the problem using an alternating optimization method. Also the proposed model is free from tedious tuning trade-off parameters. Compared with the deconvolution algorithm using uniform hyper-Laplacian prior, the experimental results validate that the proposed model can achieve better restoration results in terms of visual quality and quantitative indicators.