From MRFS to CNNS: A novel image restoration method

Abstract

Markov Random Fields (MRFs) are one of the most widely used probabilistic graphic model in image restoration. However, MRFs still require designing of clique potential function and lack of a canonical parameter learning method. To overcome these disadvantages of MRFs, we propose a novel image restoration architecture leveraging Convolutional Neural Networks (CNNs). The central point shown here is that CNNs can be viewed as generalized MRFs, which gives a novel explanation for CNNs's excellent performance from a statistical perspective. Furthermore, all ingredients for image restoration via CNNs are presented in this paper. Specifically, a learning framework and reconstruction method are constituted through minimizing KL-divergence and half-quadratic regularization respectively. Finally, simulations show that the proposed method, referred as Image Restoration based on CNNs (IR-CNNs), outperforms the state-of-the-art image restoration methods based on MRFs.