Patch-based image denoising: Probability distribution estimation vs. sparsity prior

Patch-based image denoising can be interpreted under the Bayesian framework which incorporates the image formation model and a prior image distribution. In the sparsity approach, the prior is often assumed to obey an arbitrarily chosen distribution. Our motivation is to estimate the probability directly from the distribution of image patches extracted from good quality images, thanks to a given dictionary and the redundancy of information between many images. In this paper, we provide a scheme to estimate the probability distribution and also an optimized algorithm for denoising. We demonstrate that using the estimated probability distribution as the image prior is more efficient than the state-of-the-art sparsity models for noise removal.