Research On Single Image Super-resolution Technology Based On Deep Learning

High-resolution(HR)images have more pixels and thus contain richer information,and have extensive demand and market value in many fields.In actual scenes,due to the limitation of shooting hardware and environment,it is more difficult to obtain high-resolution images.The birth of superresolution technology that restores low-resolution(LR)images to high-resolution images aims to solve this pain point.In recent years,with the development of deep learning technology,research work on the integration of image super-resolution technology and deep learning has gradually become the mainstream and achieved fruitful research results.However,due to the ill-posed nature of learning the mapping from low-resolution images to high-resolution images,the obtained solution is not unique,and the solution space is very large,so it is difficult to find a solution with excellent performance.On the other hand,most of the current super-resolution models lack the ability to distinguish different feature information,and the feature information of the original low-resolution image is lost in the process of deep neural network propagation.In response to these problems,this paper conducts in-depth research on single image super-resolution(SISR)technology based on deep learning.The main work is as follows:(1)A convolutional neural network structure(CNN)based on dual regression and dense residual learning is proposed.First,the network introduces additional constraints on the basis of the original task through the dual regression task,reducing the LR to Mapping space for HR images.Secondly,a dense residual learning module is designed,which integrates global features to enhance the mobility of feature information,and introduces a residual structure to speed up training.The effectiveness of the network structure is verified by comparative experiments.(2)A residual module that fuses channel and spatial attention mechanisms is proposed,and it is introduced into the network structure based on dual regression and dense residual learning.This module can distinguish the feature information during feature extraction,and enhance the network to the feature map.ability to learn high-frequency information.Comparative experiments prove that,compared with the state of the arts(SOTA)model,the network has certain advantages both in terms of objective indicators and subjective comparison.(3)A generative adversarial networks(GAN)based on a dynamic second-order channel attention mechanism is proposed.First,a dynamic second-order channel attention module is designed to maximize the attention module through dynamic adjustment.the regulating effect.Second,the network adopts the method of adversarial training and utilizes the perceptual loss as the optimization goal.Experiments prove that the texture details of the image generated by the network are more in line with human visual perception.