Research On Low-quality Image Super-resolution Reconstruction Algorithm Based On Deep Learning

Single image super-resolution reconstruction is a long-standing classic low-level computer vision task,which means to reconstruct a given low-resolution image into a corresponding high-resolution image.Because there is a many-to-one mapping relation between high-resolution images and low-resolution images,so image super-resolution is a sick inverse problem.To address this,the researchers give a variety of image superresolution reconstruction algorithms.These algorithms are based either on traditional methods or on deep learning.Thanks to the ultra-high data fitting ability of deep learning,deep learning has achieved unprecedented breakthroughs in image super-resolution tasks.However,the current deep learning-based image super-resolution reconstruction algorithms mostly overlook the feature correlations of adjacent layers,which makes feature maps not fully utilized and reduces the representation ability of convolutional neural network;On the other hand,the underutilization of feature maps leads to the model become more bloated,which restricts its use in mobile devices.To address the problems mentioned above,this paper makes abundant research and gives effective solutions respectively.The specific research contents are as follows:(1)An image super-resolution reconstruction algorithm by using difference value is designed to solve the problem of insufficient utilization of feature maps caused by ignoring the feature correlation between adjacent feature maps.Firstly,this paper proposes a difference value block that can extract the difference value information between adjacent feature maps.The extracted difference value can highlight which regions should be paid more attention to;Then,this paper design a difference value group based on the difference value block,the value group is designed to fully utilize the difference information extracted by the difference value block.The difference group integrates different difference information together to provide additional structural priors for image reconstruction.The reconstructed image contains more high-frequency information;Finally,the paper proposes a multipath supervised reconstruction block to supervise the reconstruction process.The proposed block can effectively solve the problem of gradient vanishing,making the model training more easy and stable.(2)A lightweight image super-resolution reconstruction algorithm by utilizing feature enhancement is designed.Firstly,a multi-convolutional feature enhancement block is designed.This block uses the convolutional layer of different convolutional kernels to obtain more information and take full advantage of the feature maps;Secondly,this paper design the channel spatial attention block via spatial dependence and channel dependence,the block can make the model quickly locate the information-rich regions in the network and reasonably allocate network resources;Then,using information distillation knowledge and the strategy of high-frequency information guiding low-frequency information,a crossfeature distillation block is proposed.This block can extract the high-frequency information in the network gradually;Finally,based on the above three blocks,the MCCB block is designed through effective combination.This module allows the role of each module to the largest.Moreover,this model use the depthwise separable convolution,so that the model can keep a relatively small amount of computation and facilitate its deployment in mobile terminals.