Research On Lightweight Image Super-resolution Reconstruction Algorithm Based On Adaptive Feature Fusion

Image super-resolution reconstruction is a technology that obtains information from lowresolution images and generates corresponding high-resolution images.As an important image processing technology in the field of computer vision to optimize image quality,image superresolution reconstruction has important application significance in medical images,remote sensing images,and video surveillance.In recent years,with the rise of deep convolutional neural networks,image super-resolution reconstruction algorithms based on deep learning have gradually attracted more attention.Compared with traditional super-resolution technology,deep learning super-resolution reconstruction algorithms are not only simple in training methods,but also significantly improved model performance.However,gradually deepening convolutional neural networks often suffer from loss of shallow high-frequency details and require more computational costs.In this regard,this thesis further studies the multi-level feature information of the network,and conducts algorithm research and lightweight exploration of image superresolution reconstruction based on adaptive feature fusion,the main work is as follows:(1)Aiming at the problem that the image super-resolution reconstruction algorithm is difficult to fully capture the high-frequency detail information of the image,a global attentiongated residual memory network is proposed,which integrates the multi-level and multi-scale feature information of the network.First,the recursive residual memory module is designed in the nonlinear mapping stage of the network,which can not only reduce the amount of parameters,but also integrate multi-level features to strengthen the shallow high-frequency information of the network.Then,multiple multi-scale residual units and a global attentiongated mechanism are cascaded in the residual memory module to fuse the feature information of different scales of the network to strengthen the network feature learning ability.Finally,at the end of the network,the multi-scale up-sampling module makes full use of the features of each scale in the reconstruction stage to achieve super-resolution reconstruction of the image.The algorithm has been evaluated by a number of standard test data sets,and has achieved superior performance compared to other advanced algorithms on the indicators PSNR and SSIM.Especially on the test data set Manga109,the PSNR result reached 39.19 d B,which is0.32 d B higher than the super-resolution algorithm AWSRN.(2)In view of the large amount of parameters and calculations that generally exist in image super-resolution reconstruction algorithms,this thesis optimizes on the basis of the previous work and proposes a lightweight feature adaptive network based on deep attention.First,a lightweight adaptive residual unit is designed to learn the attention levels of different dimensions of the feature map to obtain richer feature information.Then,the channel groupwise enhance module is designed to group the feature maps according to the channel to learn,enhance the ability of sub-feature maps to learn independently,and make the network more robust.At the same time,the weighted global context module is used to adaptively learn the feature map space and channel information,and adjust the relationship between pixels.Finally,this thesis uses a new learning rate decay method and explores a new super-resolution training optimization strategy to achieve a significant improvement in model performance.Experiments show that the algorithm is not only lightweight in design,but also achieves superior superresolution reconstruction effects.Compared with the previous work,the parameter amount is reduced by 149 K,and the calculation amount is reduced by 278.4G.At the same time,the evaluation effect in most standard test sets is similar to the previous algorithm.