Image Super-Resolution Based On Convolutional Sparse Coding And Convolutional Neural Network

Convolutional Sparse Coding(CSC)has been attracting growing attention from researchers in recent years,for making full use of image global correlation to improve performance on various computer vision applications.However,very few studies focus on solving CSC based image Super Resolution(SR)problem.As a consequence,there is no significant progress in this area over a period of time.In this thesis,we exploit the natural connection between CSC and Convolutional Neural Network(CNN)to address CSC based SR.Specifically,the Convolutional Iterative Soft Threshold Algorithm(CISTA),which can be implemented using convolutional neural network,is firstly deduced to solve the problem of convolutional sparse coding efficiently.And then a novel CSC based framework for image SR is introduced.The main idea of this framework is that the low-resolution image features share the same convolutional sparse codes with the high-resolution image features.Based on the framework,two models for image SR are proposed,which are categorized as the pre-upsampling model CRNet-A(CSC and Residual Learning based Network,CRNet)and the post-upsampling model CRNet-B according to the size of input image,for the former takes the image pre-processed by the Bicubic algorithm as input,while the latter directly processes the low-resolution image directly.As the main component of both models is the CISTA module,in which recursive learning is used,this paper also compares the differences of both models with other super-resolution methods based on recursive learning.In addition,residual learning is adopted to effectively avoids the gradient vanishing or explosion problems during the training process.Finally,the effectiveness of the proposed methods is tested via a large number of experimental results.Compared with the previous CSC based image SR method,the proposed method significantly surpasses the former in PSNR,which demonstrates the effectiveness of the proposed method in learning the convolutional sparse coding.Furthermore,the PSNR and SSIM are obviously improved when compared with stateof-the-art SR methods,which shows the superiority of the proposed method.