Research On Deep Learning-Based Algorithm For Image Super-Resolution Reconstruction

Image super-resolution(SR)reconstruction technology enjoys a wide range of real-world applications,such as human life,biomedicine,communications,military,industrial,and others.Obtaining higher resolution images by upgrading hardware is more costly and subject to a variety of environmental distractions.Software-based super-resolution algorithms can reconstruct images of high resolution and quality with more detail and textures from images of low resolution and quality.These algorithms do not require updating or replacing image acquisition equipment and are not affected by environmental factors,resulting in better visual results and more valid information about the target object,while facilitating subsequent image analysis and processing.Focused on the in-depth study of deep learning-based image super-resolution algorithms,this paper proposes a novel deep learning-based image SR reconstruction algorithm by analyzing the advantages and disadvantages of existing SR algorithms based on deep learning.First of all,this paper proposes a lightweight U-shaped wide activation residual network(U-WARN)for fast and accurate image super-resolution,to address the problem of current image SR methods with a large model,heavy computation,and big memory storage.We propose a more efficient and lightweight wide activate residual block with the lightweight group convolution as the basic block for extracting features.Simultaneously,a two-branch Ushaped structure is proposed to distinguish between features of different depths,where a simple network is used to extract shallow feature information in the downward branch,and a complex network to extract deeper feature information in the upward branch,and then fuse the multi-level information by the step-bystep reverse fusion strategy.Moreover,a lightweight Non-Local module is proposed to model the global context information and further improve the performance of SR.Finally,a high-frequency loss function is designed to alleviate smoothing image details caused by pixel-wise loss.Secondly,a large number of ablation experiments are conducted on natural image datasets for comparative analysis to verify the effectiveness of the proposed modules and strategies in our proposed U-WARN.The extensive experiments show the U-WARN achieves a better trade-off between image SR performance and model complexity against other state-of-the-art SR methods.Our algorithm gets 38 FPS on the standard benchmark dataset B100 with a scale factor of 2,which is 21.5 times faster than RCAN,and 136 and 115 times smaller than it in terms of computational consumption and model size respectively,resulting in greatly benefiting applications for applications on low computing power or portable devices.At last,this paper investigates the application of image super-resolution in medicine.In clinical medicine applications,there is often costly in terms of time,safety,and cost to acquire high-resolution magnetic resonance(MR)images.However,current most SR algorithms applied to MR images do not work well.The proposed U-WARN in this paper can be successfully applied to MR images and can substantially improve the resolution of MR images with very little computational cost.Extensive experiments show our U-WARN achieves better performance on MR images with a very low number of parameters(only0.3K)and computational burden.At the same time,our algorithm facilitates practical deployment due to less model size(only 1.3M),faster inference speed,and lower memory storage.