Research On Single Image Super-resolution Reconstruction Algorithm Based On Deep Learning

Image super-resolution(SR)reconstruction is based on Image processing and signal processing.The technology of converting existing Low resolution(LR)images with less detail information into High resolution(HR)images with more texture details.In the field of digital image,compared with the hardware scheme,the software scheme is cheaper and easier to promote and implement,and the reconstructed image quality is close to the ideal image.Therefore,it has not only academic research value,but also a wide range of practical application prospects.With the development of machine learning in the field of computer vision,deep learning has been widely applied to image super-resolution reconstruction tasks,and achieved good reconstruction results compared with traditional algorithms.Nowadays,the image super-resolution method based on deep learning has gradually become the mainstream.Based on this,the main research work of this paper is as follows:(1)The image super-resolution reconstruction method based on deep learning is sorted and analyzed according to network structure design,data set and evaluation indexes.(2)A super-resolution reconstruction model based on residual network and dense network is proposed.Aiming at the problem that the linear network is difficult to make the network layer deeper,residual connection and dense connection are used to enhance the network depth,strengthen the feature propagation and reduce the training difficulty.Firstly,feature learning is carried out by using two sub-networks composed of residual block and dense block respectively,and then residual learning is carried out by the two features,and the backward transmission is continued.Then,the learned features are introduced into the feature fusion layer for feature fusion,and the high frequency information is learned adaptively to improve the reconstruction effect.(3)An image super-resolution reconstruction model based on multi-scale residual channel attention network is proposed.This network model is based on the Multi Scale Residual Channel Attention Block(MSRCAB)and adopts a three-channel setting.Three different convolution kernels of 3×3,5×5 and 7×7 are used for feature extraction and learning in each Channel.It not only expands the width of the network,but also improves the receptive field and enhances the nonlinearity of the network.Moreover,multi-scale convolution can extract more image details.GELU activation function was used in MSRCAB,and perceptual loss was introduced into the use of loss function for optimization.Finally,the effect of reconstruction was improved through channel attention and residual connection.(4)A super-resolution reconstruction model of generative adversarial Networks images combined with attention mechanism is proposed.In order to solve the problems of poor image reconstruction effect and unclear texture,the generator and discriminator are improved and optimized simultaneously based on SRGAN.First,the dual attention mechanism is introduced at the generator end and BN layer is removed,and the dual attention mechanism module based on channel attention and spatial attention is constructed.The image reconstruction makes full use of the features of each level,so that the features of each layer can be effectively utilized.Removing BN layer can improve the flexibility of network,reduce the computational cost of model,and improve the efficiency of model reconstruction.Then,based on the WGAN idea of Wasserstein distance at the discriminator end,Wasserstein distance is used to measure the antagonism loss between the real image and the generated image,making the training process of the whole network more stable.