Image Denoising Using Expected Patch Log Likelihood and Hyper-laplacian Regularization

Image denoising is a fundamental problem of the field of computer vision and image processing. To decrease the computational complexity, many patch-based priors have been presented in image denoising methods, which have achieved the state-of-the-art performances. However the formation of global prior from local patch interaction is still challenging for patch-based image denoising methods. This paper proposed a new image denoising method by introducing a hyper-laplacian regularization term into the expected patch log likelihood (EPLL) framework. In EPLL framework, the patch-based prior can be used in global image denoising process by increasing penalty parameter with iteration. Meanwhile gradients in the whole image satisfy heavy tailed distribution, which can be well depicted by hyper-laplacian. The unified framework could benefit from the local patch-based prior and non-local gradient sparsity prior. The proposed method is extensively evaluated on Berkeley Segmentation Dataset (BSD), and has the comparable performances with the state-of-the-art image denoising methods.