Model Driven Deep Neural Network For Image Restoration

Images provide a large part of information acquired by people.High quality images will provide more accurate information.However,due to the limitation in the process of image acquisition,compression,transmission and decompression,quality of image sometimes is objectively low.Meanwhile,with the development of high technology,people need more and more high-resolution images to match with the multimedia,so image restoration plays an important role in people's life.The traditional model-based method includes the data items and prior iterms,which can be combined to restore the details of the orignal image on the basis of approximating the whole image.Generally,the prior information of the image is extracted manually,which is designed by observating the image features and has a relatively intuitive mathematical expression,but there is a problem of single prior information.The image restoration method based on deep learning relies on the drive of big data to realize the end-to-end mapping between low-quailty images and high-quality images,which is largely dependent on the design of the network,and with the increasing of the convolutional network layer,there is little improvement.In this paper,combining the two methods,we proposed an algrithom—denoising prior driven deep neural network for image restoration method.The innovation points of this paper are as follows:1.With the powerful feature extraction ability of deep learning,the deep neural network based denoising prior is learned,which comprehensively covers the prior information of the image and provides more accurate constraints for the model-based image restoration method.The structurally symmetric U-Net is adopted as the network of denoising operator learning,and its strcture of “down-sampling and up-sampling” also greatly facillitates the extraction of image features.It not only overcomes the limatation of single image prior,but also use deep neural network to learn denoising operator,which lays a foundation for the model expansion to be an end-to-end network.2.Embed the deep learning-based denoising operator into the iterative process derived form the image degradation model,so that after the initial denosing by the denoising operator,the image can still be further accurately restored by combining the image degration process.At the same time,the whole process of alternating iteration is expanded into a nonoverlapping process to realize end-to-end network training,so that the parameters that need to be manually adjudted in the network can be adjusted adaptively through network training,so as to achieve the optimal overall results.3.When solving the sub-objective function in the process of alternating optimization,this paper adopts single-step gradient descent to replace the original closed-form solution,which greatly reduces the operation complexity of the whole model and makes the network training more simple.At the same time,the rationality of the substitution process and the convergence of the model are proved mathenatically.The algorithm proposed in the paper provides a good basis for the combination of model-based and deep learning-based image restoration algorithm,and the results achieved are also the state-of-the-art in the same period.