Research On Image Super-Resolution Algorithm Based On Lightweight Neural Network

With the rapid development of computer technology,pictures have become the main media for information transmission.People are increasingly demanding image resolution.For example,in the field of medicine,high-resolution medical images can help physicians to effectively grasp the physical condition of patients.In the field of remote sensing images,high-resolution remote sensing images can help us respond effectively to special weather and geographic disasters.However,the limitations of imaging technology,the impact of external environmental conditions,and the loss of image transmission process all result in the degradation of the original high-resolution image to the low-resolution image.Therefore,image super-resolution algorithms have emerged.Image super-resolution algorithms are widely used in many fields such as medical images,public security,military and national defense.In recent years,based on the strong ability of Convolutional Neural Network(CNN)to extract effective abstract information in low-rate and high-resolution space,great progress has been made in single image Super-Resolution(SISR)algorithm research.However,the vast majority of single-image super-resolution networks achieve excellent super-resolution performance by continuously stacking network depths,but it also brings a huge amount of computing costs and parameters,which are difficult to apply on mobile devices with low computing performance and lack of flexibility.How to balance the relationship between network parameters and performance so that it can extract rich detail texture features from low-resolution images and rebuild high-resolution images has become a hot issue in single-image super-resolution research.In this paper,two lightweight super-resolution networks are proposed,which not only guarantee the super-resolution performance,but also greatly reduce the parameters and calculation costs.(1)An information distillation attention network based on residual structure is proposed to solve the problem of insufficient utilization of rough features in the existing lightweight image super-resolution network based on information distillation mechanism and ignoring the feature processing in the fusion process.The network consists of residual distillation attention module and reconstruction module.The Distillation Attention Module,Global Context Fusion Module and Multiple Feature Distillation Connection Module were embedded in the Residual Distillation Attention Module.The Distillation Attention Module captures the dependencies between pixels and filters redundant features to improve the sensitivity of the model to important features.The Global Context Module integrates distillation features at different levels to achieve better feature expression.The Multi-Feature Distillation Link Module connects features at multiple levels to enhance feature information transmission and compensate for information loss.The experimental results show that the lightweight network achieves good image super-resolution performance,while effectively reducing the computational load,parameters and reconstruction time.(2)At present,many image super-resolution networks handle different frequency characteristics in the same way,which is not flexible enough.Blind stacking of network depth results in redundant parameters and huge calculation costs,leaving room for model improvement.This paper presents a lightweight multibranch residual network.The network has three branches: the context information extraction branch,the detail texture branch and the edge extraction branch,which extract the context information,the detail texture and the edge part of the input low-resolution features respectively.First,the Laplace operator is combined with the void convolution to extract some edge features on the edge extraction branch,which not only enlarges the receptive field,but also significantly reduces the number of parameters and improves the model performance.A multiscale feature residual attention block is proposed on the detail texture extraction branch.It is used to extract and fuse multiscale features adaptively,and a symmetric structure is used to increase the convolution utilization by reusing the convolution layer,so as to deepen the network depth while reducing the amount of parameters and computational consumption.This paper holds that the same treatment of the characteristics of different channels may waste part of the network performance on low frequency characteristics,so the design of multi-scale channel attention blocks is flexible to extract the characteristics of different channels.Finally,the shallow residual network is used in the context information extraction branch to collect image context information to assist in the final image super-resolution reconstruction performance.Compared with other advanced networks,this lightweight network achieves a good balance in rebuilding performance and parameter amount.