Research On Image Super-Resolution Method Based On Feature Refinement Network

In contemporary society,images,as the main medium of information transmission and emotional expression,have important application value.However,due to limitations of factors such as hardware acquisition equipment and network transmission media,people often only obtain images with lower resolutions.Super-resolution reconstruction technology can restore these low-resolution images to high-resolution images,thereby improving the clarity and utilization value of images.This technology is now widely used in many fields such as smart security,medical imaging and photo restoration.In recent years,with the development of deep learning technology,many models with excellent reconstruction performance have emerged in the field of image super-resolution.However,these models usually have a complex architecture and a large number of parameters.In practical applications,they are limited by computing resources and physical memory,making it difficult to run efficiently.The research of this thesis focuses on the realization of high-quality image super-resolution reconstruction under a smallscale model structure.The main research contents are as follows:(1)A feature refinement network image super-resolution method based on asymmetric convolution is proposed.In this thesis,using the knowledge of information distillation,a feature refinement block is designed as the basic module of the model.This module uses channel selection at different levels to extract refined features with different receptive fields and fuse them.This method can achieve comprehensive feature extraction and has a small amount of parameters.In addition,this thesis designs a residual asymmetric convolution block to improve the diversity of extracted features.In order to further enhance the expressive ability of features,this thesis also constructs a dual-attention mechanism module that integrates channel attention and spatial attention,which can better help the model focus on the information-rich part of the feature.Through the experimental comparison with other super-resolution algorithms,it is proved that the method proposed in this study can effectively balance the network size and reconstruction performance.(2)A feature refinement network image super-resolution method based on depthwise separable convolution is proposed.In order to further reduce the amount of network parameters,this method uses depthwise separable convolution instead of standard convolution based on the previous method.A lightweight residual block is designed based on depthwise separable convolution and residual connections to improve the efficiency of refined feature extraction.In addition,this thesis also designs a lightweight dual-attention block,which has a lower parameter amount and can effectively enhance the network representation ability.By improving the feature refining method and using 1×1 convolution to reduce the feature dimension to replace the channel selection operation of the previous method,this method can extract refined features more effectively.Finally,considering that the depthwise separable convolution will affect the reconstruction accuracy,this thesis expands the depth and number of channels of the network to improve the reconstruction performance.The experimental results show that,compared with other super-resolution methods,the method proposed in this study performs better in image reconstruction,and has a smaller scale in terms of model parameters.