Attention-based Deep Learning Methods For Image Super-resolution Reconstruction

Single image super resolution(SISR)is the method of reconstructing an accurate highresolution(HR)image given its low-resolution(LR)counterpart,and usually take the way to reconstruct image into large scale.Recently,various super resolution methods are prone to generate images with blurs and artifacts,especially in large scaling factor.Most of the current reconstruction methods do not distinguish the features of LR images,resulting in a failure to recover high frequency details such as tiny textures.To tackle these problems,this paper studies the methods of super resolutions based on deep learning with attention mechanism,including:(1)This paper has carried out in-depth research on the traditional algorithms and deep learning methods of SR,analyzing the losses of the super-resolution reconstruction networks.Experiments show that the L1 norm can effectively reduce the reconstruction artifacts of the image smooth region compared to other loss functions,and the implementation complexity is low.Therefore,the L1 norm will be used in this paper based on the experiments.(2)In order to make better use of high frequency features of the images,this paper introduces a residual attention mechanism by adding the attention module to the backbone network in a stacking way,which can strengthen the ability of extracting the high frequency features without weakening the ability of the backbone network.The extraction of high frequency features enhances the performance of the network.Experiments show that the attention mechanism proposed in this paper can effectively locate the high-frequency feature position of the images and reduce the blurs and artifacts in reconstructed images.(3)With moderate parameters,in order to improve the performance of the reconstruction network in large scaling factor,this paper proposes an attention-based back-projection network based on projection unit and attention mechanism.This paper proposes an LR feature extraction module in the projection unit,which improves the ability to extract image features in the network forward reconstruction process.To reduce the number of network parameters and resource consumption,a simple attention mechanism module with deep convolution is proposed,which helps the network locate the spatial position of the high-frequency features of the images.To further reduce the network resource consumption,this paper uses the residual connection instead of the dense connection commonly used between the projection units.Experiments show that in the case of moderate number of the projection units,the residual connection can be used to obtain the performance of densely connected networks.With moderate depth,not only does it show good performance in large scaling factor,but also shows good denoising ability.