Single Image Super-resolution Based On Fractal Residual Network

In modern society,images are functioning as a vital information carrier and are widely used.However,in the actual imaging process,due to the hardware conditions and other factors,a lot of key information is lost,which makes it impossible to meet application requirements.Therefore,generating high-resolution images through related algorithms has a very important research significance in many fields such as medical imaging and video surveillance.So far,a large number of scholars have studied superresolution algorithms and achieved relatively good results.Based on convolutional neural networks,algorithms mainly use a large number of high-definition images to build a sample library,and use a convolutional neural network to study the mapping relationship between low-resolution images and high-resolution images.At present,most algorithms improve the depth of the network through raising the number of residual modules to promote network performance.However,if the network is too deep,it will lead to problems such as high frequency information loss,high training and running costs.In addition,the original feature information is insufficiently extracted,and the generated images lack real details,since a single network structure cannot make full use of the information of low-resolution images.Starting from the above problems,this thesis proposes a new super-resolution reconstruction method called Fractal Residual Network(FRN).The main contributes of FRN are listed as follows:(1)it uses fractal residual attention blocks to make full use of different hierarchical features to generate more refined features.(2)it introduces the channel attention mechanism to adaptively rescale the characteristics of each channel and increase the network's ability to discriminate and learn.(3)it combines local residuals and global residuals to compensate for information loss and diminish the difficulty in learning.The results of experiments are that this method is better than many other algorithms in reconstruction performance.This thesis makes improvements based on FRN and constructs the Mixed Attention Fractal Residual Network(MAFRN)so as to take advantage of the information with a high-frequency characteristic in the image and promote the capability of network using features in a selective way.In MAFRN,a new attention module is proposed.The module is composed of channel attention unit,spatial attention unit and a fusion mechanism combining the two.This module is introduced into the fractal residual attention block,which can strengthen the network's extraction of high-frequency information and further improve the reconstruction effect of the network.The experimental results imply that the MAFRN network recommended in this thesis can enhance the detailed information of the reconstructed image,and the reconstruction performance is better.