Research And Application Of Lightweight Single Image Super-Resolution Method Based On Deep Learning

Image has become an important information carrier in modern life as it can show more vivid information.However,there are a large number of low-resolution images in our life,which cannot meet the actual needs due to their blurred details.Image super-resolution is a technology that can reconstruct low-resolution images into corresponding high-resolution images with better visual quality,which is of great significance to scenarios that require high quality images such as monitoring and medical imaging.Recently,image super-resolution methods based on deep learning have made a remarkable breakthrough in image reconstruction.However,most of these methods contain a huge number of parameters and computational complexity,which limits the application of these methods on devices with limited resources,such as tablets,smartphones.Research on lightweight image super-resolution methods based on deep learning is expected to solve the problem.The main work of this paper is as follows:A lightweight efficient residual feature fusion method for image super-resolution is proposed.Firstly,aiming at the problem that residual block is difficult to be used in lightweight network due to the large number of parameters and computational cost,this paper designs an efficient residual block by combining group convolution and pointwise convolution,which can effectively extract image feature information and has fewer parameters and computational cost.In addition,this paper improves the existing enhanced spatial attention block by removing redundant network layers as well as redesigning the network structure,and further proposes a lightweight spatial attention block with fewer parameters and computational cost,and a more obvious performance improvement effect.Based on the above work,an efficient residual feature fusion module is designed as the building block to extract deep features of images.In order to further improve the performance,this paper designs a variation-based channel attention block,which can effectively help the network to fuse global feature information,and introduces few additional parameters and computational cost.Finally,this paper proposes a novel lightweight network model.The model complexity and performance comparison experiments show that compared with the mainstream lightweight methods,the proposed method achieves better results in both objective evaluation indexes and visual quality,and meanwhile contains fewer parameters and computational cost.The corresponding ablation experiments also verify the effectiveness of the modules proposed in this paper.In order to apply the proposed lightweight image super-resolution method in our real life,this paper finally implements a super-resolution image restoration system.The system is deployed on personal computer,and can invoke different super-resolution methods to restore images by using local computing resources.