Learning smooth dictionary for image denoising

Priors play an important role for most of image denoising approaches under the Bayesian framework. Dictionary learning can capture the sparseness prior and has been widely used for various applications in recent years. In the other aspect, TGV (Total Generalized Variation) can reduce the staircasing effects and preserve the geometric structures by the smoothness prior. In this paper, a new dictionary learning model is proposed to combine the above two priors by adding a second order TGV regularizer on each atom of the dictionary. The proposed model is applied to image denoising, and experiments demonstrate its effectiveness.