The Research Of Image Restoration Method Based On Convolution Neural Network

Image restoration technology is widely used in various fields of society.The quality of image affects people's communication and information acquisition to a large extent in the information age of today.Therefore,the study to image restoration is still necessary.Image restoration is an ill-posed inverse problem in Mathematics.In general,there are two major approaches to solving ill-posed problems in the field of visual image:One is model-based optimization methods,and the other is discriminative learning methods.These two methods have their respective merits and drawbacks.The former is flexible for handling different inverse problems but it is always time-consuming.On the contrary,the latter has a fast testing speed but a model structure can only solve specific recovery problems.Based on such considerations,the article adopts variable splitting techniques to separate fidelity and regularization items in image restoration problems.And it also integrates the priors,which learned by discriminant learning methods,into the model-based optimization methods in order to get the effect of combining the advantages of the two methods.The article mainly studies image denoising,image deblurring and single image super resolution(SISR).Convolutional neural network(CNN)is used to learn a series of denoising models and the models are applied to the study of the latter two problems.Take into account its many advantages,we adopt CNN as discriminative learning method.First,it can learn a highly complex image constraint through a large amount of data,so that the image priors can be applied to the limit;Second,by using some techniques in CNN,the convergence process of training can be accelerated;Third,it can make use of computer hardware resources conveniently,such as GPU,which greatly improves the speed of dealing with problems.Moreover,many other excellent algorithms are selected,and the experimental results of these algorithms are compared with the methods proposed in this paper.Thus,the performance of the proposed method is explained in a more comprehensive and three-dimensional way.