Research On Image Super-resolution Reconstruction Algorithm Based On Deep Learnin

High-resolution(HR)images possess a higher pixel density,contains more information,which enables people to observe subtle changes in the target surfaces at a smaller size compared with low-resolution(LR)images.However,in practical application,limited by hardware conditions,the resolution of the image still needs to be improved.Super-resolution(SR)reconstruction technology enables HR images to be obtained from LR images using a softwarebased method.In recent years,with the development of deep learning,which has achieved excellent results in various fields,more and more people use methods based on convolutional neural network(CNN)for image SR reconstruction.This paper mainly studies the SR reconstruction method based on CNN.The main contents are as follows:(1)An asymmetric multi-scale super-resolution network(AMSSRN)is proposed.In this network,a residual multi-scale block(RMSB)and a residual multi-scale dilation block(RMSDB)are designed to extract both shallow and deep features from images.This asymmetric structure enables the deep features to be fully extracted while reducing the redundancy of the modules used to extract shallow features.In this work,a feature refinement fusion(FRF)module is also established,which can make full use of extracted features to improve network performance.Experimental results indicate that compared with enhanced deep super-resolution(EDSR)network,the proposed AMSSRN can efficiently reduce the redundancy of network parameters and enhance the feature extraction capability: the maximum increases in peak signal to noise ratio(PSNR)and the structural similarity(SSIM)index can reach 0.23 d B and 0.9782,respectively.(2)A lightweight adaptive information fusion attention network(AIFAN)is proposed.Specifically,a dual path block(DPB)is designed,which can adaptively aggregate multi-scale features from different paths through the adaptive path fusion block(APFB).To take full advantage of the hierarchical features,all DPBs are connected using a dense connection mode.Meanwhile,a receptive field attention(RFA)module is proposed to increase the receptive field of the network and force the network to pay attention to the region with more information in the feature.In addition,a hybrid upsampling module is designed.The adaptive multi-scale upsampling(AMSU)module and feature update upsampling block(FUUB)are combined in the module to obtain better reconstruction results.Experiments results show that AIFAN achieves a better balance between model size and performance.