Image Restoration Based On Sparse Representation And Convolutional Neural Network

With the rise and development of internet and digital multimedia technologies,people's requirements on image quality are increasing day by day.High-quality images usually contain richer content and information and even bring viewer the beautiful feeling,while poor quality images not only lose a lot of information,but also bring discomfort to the viewer sometimes.However,in the process of digital image acquisition,compression,transmission and display,image degradation will inevitably occur due to interference from external environment,its own limitations of imaging equipment and man-made operational mistakes.Therefore,how to eliminate noise,effectively remove blur,improve resolution and repair missing parts from existing degraded images plays an import role in theoretical research and practical application and image restoration came into being.In this paper,the sparse representation prior based image restoration model and the residual learning based convolutional neural network are taken as the two start points.The three types of classical image restoration applications,such as image denoising,image deblurring and super-resolution reconstruction,are deeply studied.First of all,we briefly introduce and summarize the relevant theoretical knowledge of image restoration,which lays a solid theoretical foundation for the follow-up work.Then,based on the sparse representation prior model,we propose two similarity criteria for measuring the similarity between two image patches,which called PCA(Principal Component Analysis)-subspace Euclidean distance and SSIM distance respectively,to remedy the defect of Euclidean distance.In addition,the sparse representation based image restoration generally just considers the local sparsity of image and neglects the non-local self-similarity of image,so that the restored image deviates from the original image largely and the result of image restoration is not satisfactory.Therefore,we apply both the local sparsity and non-local self-similarity of image to dictionary learning and sparse coding coefficient estimation and propose a non-local sparse representation regularization model called NSRR.On the three types of classical image restoration tasks,a large number of experimental results show that the proposed algorithm achieves a higher value of evaluation index objectively,and the restored image is more consistent with the original image subjectively.Finally,we take the residual learning based convolutional neural networks model called DnCNN as the benchmark network and propose three key technologies to solve the problems existed in DnCNN.The three technologies are spatially related and learnable activation unit,formatted residual layer and cross-level loss function respectively,based on which the final convolutional neural networks for image restoration is given and its performance is verified by training a series of models.On the three types of classical image restoration tasks,a large number of experimental results show that the proposed algorithm achieves a higher value of evaluation index objectively,and the restored image is more consistent with the original image subjectively.