Sparse method for image denoising based on clustering

Non-local means and sparse models are two important denoising methods attracting many attentions during the recent years. The non-local means method uses the weighted average of the pixels that share similar neighborhoods as the denoised result, and the sparse denoising method recovers the non-noisy components of an image by a sparse representation with a few atoms in a dictionary. Based on these two denoising methods, we propose a clustering based sparse model for image denoising, which first partitions the image patches according to their similarities, and then uses l1/l2 norm regularization to make the similar image patches in the same cluster share the same sparse structure when they are represented by overcompleted dictionary. For dictionary selection, two dictionaries Discrete Cosin Transformation (DCT) dictionary and bi-orthogonal wavelet dictionary are chosen to represent both smooth components and detail components of the image. Experiment results show that the proposed method has better performance of image denoising compared with some traditional sparse denoising methods.