Research Of Image Super-resolution Based On Feature Fusion Attention Networks

Image super-resolution reconstruction refers to the process of restoring high-resolution images on the existing low-resolution images by means of software and hardware.It has been a hot issue in the field of image processing and computer vision.Recently,image super-resolution reconstruction methods based on deep learning have been emerging and the results are encouraging,Great progress has been made in this field under the continuous efforts of researchers.However,the current method still has many problems,such as:the network model is complex,the reference number is large,the practicability is low,the reconstructed image is too smooth,the detail information is lost seriously,and the like.To solve these problems,we proposed an image super-resolution algorithm based on feature fusion attention network.This algorithm aims to reduce the number of network parameters,improve the practicability of the network,and improve the ability of the network to filter and utilize features,so that the reconstruction results contain more texture details and have more information.In particular,the main contribution of this paper is as follows:(1)A multi-level feature fusion network structure is proposed.Past super resolution algorithms show that:network depth is one of the most important factors affecting network performance.With the deepening of the network,the sensory field will increase,and the extracted features of each network layer will be gradually ed.The features extracted from the network model at different depths all contain information that is conducive to super-resolution reconstruction.Most previous methods ignored how to make full use of these information.The feature fusion sub-network we proposed can make more efficient use of these information.(2)We proposed feature attention network structure.The focus of the super resolution reconstruction is to restore the high-frequency detail information such as the image texture edge.The previous method is usually treated in the same way as the extracted feature information,with no focus on the details of the image edge and the texture.We designed an attention module and introduced the attention mechanism to enhance the high-frequency information in the image reconstruction process,so as to obtain clearer details of the reconstructed image.(3)We used hierarchical parameter sharing and recursive structure of convolution block to reduce the number of parameters.All modules between multi-level subnetworks adopt the strategy of parameter sharing.At the same time,the residual convolution block in the feature extraction module of each level subnetwork adopts the recurrent structure,which further reduces the number of parameters.The two strategies cooperate with each other,which greatly reduces the number of parameters of our network model and makes the model structure more compact and the computing speed faster.In conclusion,based on the theoretical basis of convolutional neural network,we improved and innovated the image super-resolution method based on deep learning.We proposed a single image super resolution algorithm based on feature fusion attention network,and the research is carried out on this basis.The results show that compared with other super-resolution algorithms based on deep learning,the method presented in this paper is more effective in both subjective and objective aspects.