A convolution neural network based variational restoration model

Regularized restoration models in classical Sobolev space and BV space have been widely discussed in image processing area. The advantage and disadvantage of the regularized restoration models in W^1,2(Ω) and BV (Ω) are obvious. Choosing proper space for solution is an effective method for overcoming the ill-posedness and obtaining a more accurate restoration solution. For the obvious difference between different parts of an image, the regularized term in restoration models should be designed adaptively. In this paper, a regularized restoration model based on the convolution neural network is proposed. Using the special convolution neural network, the image or the image patches can be classified. Depending on the results of classified, a proper space and formulation of regularized term can be decided. Experiments show that the restoration results of this method based on the convolution neural network are superior to those of the traditional models in only one space. This method of choosing restoration model based on convolution neural network can give consideration to both maintaining the edge-texture and de-noising. And the cost of computation is very low.