Reserch On Single Image Superresolution Based On Mulyi-level Residual Attention Fusion Network

In the real world,people have higher and higher requirements for the quality of electronic images,which means that people expect to get high-resolution(HR)images.High-resolution simply that the image owns a large picture element density,so it contains rich detail information.To put it simply,SR is to reconstruct HR images from LR images through some technical means,including hardware or software operations.Image super-resolution has been widely used in transportation,security,medical and other industries,so the study of image super-resolution,which obtain more image details has important theoretical research significance and practical application value.Recently,we found that deep convolutional neural network plays a good role in improving the performance of super-resolution reconstruction.However,in standard convolutional neural networks(CNNs),each layer of neurons is artificially set as a fixed receptive field,while the size of the receptive field of visual cortex neurons is regulated by stimulation,which is rarely considered when constructing neural networks.In addition,in the learning process of super-resolution model,there will be one or more up-sampling operations,and each up-sampling will lose some local surface information.However,most existing methods only focus on high-level features while ignore the information contained in low-level features,resulting in image artifacts and affecting the quality of super-resolution.Aiming at the above problems,this paper mainly completes the following three projects.1.A new residual group is proposed in the Multi-level Feature Fusion Network(MLFFN),which add a selective kernel convolution(SKConv)to the residual group to extract the feature information of the receptive field at different scales,which can alleviate the problem of fixed receptive field caused by standard convolution;We propose a innovative multi-level feature fusion module with the purpose of fusing the appearance information of multi-level features,and selectively fuse the low-level up-sampling features with the high-level up-sampling features to improve the super-resolution effect by fusing the complementary information between multi-level feature information.2.In view of the improvement of MLFFN,a multi-scale residual-attention feature fusion module(MRAF)is proposed,which takes into account the global and local dependencies of multi-scale features,so as to extract more abundant multi-scale features and further alleviate the disadvantages caused by fixed receptive field.