Image Super-resolution Reconstruction Based On Residual Multi-attention Network With Skip Connection

With the improvement of the quality of life today,people’s demand for high-definition images has become higher,and the previous super-resolution reconstruction methods can no longer meet people’s demand for high-quality images.Therefore,image super-resolution(SR)reconstruction technology based on deep learning is proposed to improve image quality to meet the needs of different groups of people in many fields,and its application value is extremely wide,such as satellite monitoring,medical and military,digital media,remote sensing,etc.In recent years,deep learning technology has continued to develop,and Convolutional Neural Networks have achieved great success in SR reconstruction.This paper focuses on single image super-resolution reconstruction based on convolutional neural network.These algorithms use high-definition images to build a sample database,and use the representation capabilities of convolutional neural networks to learn the spatial mapping relationship between low-resolution and high-resolution images,and then use the learned spatial mapping relationship as a priori Knowledge to reconstruct SR images.As we all know,with the continuous increase of network depth,deeper networks will lead to the loss of high-frequency information,making it difficult for neural networks to be trained.In addition,when low-frequency information and high-frequency information are processed,the channel features of low-resolution images are treated equally,which will lead to the weakening of the representation ability of CNN,and the original feature information extracted will appear insufficient.The real details of the reconstructed image are lacking.Most SR technologies lack flexibility in processing low-resolution and high-resolution information,so it is difficult for the network in the low-resolution space to directly extract the output of the convolutional layer.Considering the above problems,the image super-resolution reconstruction algorithm based on the residual multi-attention network with skip connection is proposed to apply to the single-image super-resolution,and it has been studied in depth.This paper proposes a super-resolution reconstruction algorithm based on the residual multi-attention network with skip connection,and specifically introduce the advantages of the model and the problems it solves.In addition,the residual multi-attention structure in the skip connection is also proposed,which can not only make the number of layers of the network model reach a very deep level,but also can solve the uncertainty problem caused by the deepening of the network model.It can make the main network pass through the skip connection to bypass the rich low-frequency information,and concentrate on learning the high-frequency information.The model also introduces multiple attention blocks to solve the problem that the above features are not distinguished and are treated equally.The main task of the attention block introduced is to pay attention to what is meaningful and feature location information in the input image,so that the network model can adaptively redistribute weights to adjust image features and extract important information.The sub-pixel convolution operation is introduced in the image reconstruction part,which can reduce the complexity of the network and avoid the interference of human factors.In this paper,arbitrary magnification factors are used in the super-resolution reconstruction of the skip connection residual multi-attention mechanism.