Research On Single Image Super-resolution Based On Deep Learning

With the development of scientific technology and the improvement of people's material life,people are eager to get clearer images.However,constrained by the hardware performance and data transmission bandwidth,it is often only possible to obtain a low-resolution image,which cannot meet the demand.Therefore,using super-resolution reconstruction technology to improve the image resolution has great values both on research and practice.Thinking from a mathematical perspective,the super-resolution reconstruction is an ill-posed problem.Based on the deficiencies of the FSRCNN,this article proposes a deep learning superresolution method on the basis of local structure self-similarity and encoder-decoder structure,which optimizes the FSRCNN from the perspective of image feature extraction and network structure.The innovative work of this article mainly includes the following two points: 1.Research on single image super-resolution reconstruction algorithm based on space-time conversion and deep convolutional neural network..In view of the fact that the FSRCNN only considers the similarity of adjacent pixels,but ignores the structural information in the image.This article proposes an optimization method based on space-time conversion and deep convolution network.Firstly,this method constructs image groups from gray level,structure and hash features respectively to extract the local structure self-similarity information in low-resolution images.Moreover,using image block groups with structural similarity to simulate the image sequences.Finally,the super-resolution reconstruction is completed by fusing similar image blocks with sequence super-resolution reconstruction method.2.Research on super-resolution reconstruction algorithm based on encoder-decoder structure.In the view of the fact that the traditional single image super-resolution reconstruction algorithm based on deep learning,fixes the spatial resolution of feature map,which not only restricts the field of perception of feature map,but also inhibits the learning ability and feature expression ability of network for multi-scale information.This article proposes a deep learning super-resolution reconstruction algorithm based on encoder-decoder structure on the basis of research one.The proposed algorithm divides the network into two parts: coding network and decoding network.The coding network solves the limitation of spatial resolution of feature map by using sampling convolution,which improves the perception field of feature map,and maks the network model extract features in multiple scale spaces.Then,the decoder network completes the super-resolution reconstruction step-by-step,and use the feedback upsampling module to optimize the upscaling process.Finally,by using the bypass connection method,the shallow features in the network are reused to avoid the loss of information,and solve the problem that network depth is limited by the super-resolution reconstruction factor.The experimental results indicate that,compared with FSRCNN algorithm,the proposed deep learning super-resolution reconstruction algorithm combining local structure selfsimilarity features and coder-decoder structure can reconstruct more detailed texture under the same dataset and training parameters.On the PSNR index,it can increase by 0.85 dB on average.