Research On Fundus Images Super-resolution Reconstruction Algorithm Via Deep Learning

With the aging of population and the change of life style,fundus diseases are becoming more and more common.Fundus images are widely used in the screening and diagnosis of ophthalmic diseases due to their safety and cost-effectiveness.For better diagnosis,high resolution fundus images are essential.However,the acquisition of high resolution fundus images often requires high hardware equipment.Therefore,using super resolution technology to improve the resolution of fundus image is a good choice.In recent years,the use of deep learning to solve medical image super-resolution has become a research hotspot.Aiming at the existing problems of super-resolution algorithm based on deep learning,this paper carried out research from the perspective of fundus images.The main contributions of this paper are as follows:(1)This paper proposes a super resolution reconstruction algorithm of fundus images based on attention mechanism,aiming to effectively capture the dependence between low-resolution and high-resolution images and feature layer channels.In this paper,we first propose a kind of denset module with attention mechanism.In this module,iterative up-sampling and downsampling and attention mechanism are applied to the super-resolution task to capture the dependence between low resolution and high resolution while paying attention to the dependence between characteristic layer channels.In order to reduce the optimization difficulty,the algorithm combines the module by connecting local and global residuals.Experimental results show that,compared with the existing algorithms,the proposed algorithm can better capture the dependence between low-resolution and high-resolution images and feature layer channels,and obtain better reconstructed fundus images.(2)In this paper,a style-guided fundus image superresolution reconstruction algorithm is proposed,aiming at effectively increasing the network receptive field and the fusion of multiscale features,and fully mining the dependency relationship between the middle layer channels.For the above two purposes,this paper proposes a module with style guidance,which contains two branches.In order to improve the network receptive field,an improved U-Net backbone branch is proposed in this paper.This branch aims to increase the network receptive field through the lower material samples and promote the fusion of feature layers of different scales through dense connection.In order to fully explore the dependencies between channels in the middle layer,we design a mask branch that can propose feature layer style descriptors,which can enhance or suppress the output of the trunk branch at the channel level.In addition,the residual attention mechanism is used to effectively combine the two branches to improve the representational ability of the model.In this paper,the effectiveness and robustness of the model are verified by a large number of experiments.