Research On Super-resolution Reconstruction Algorithm For Retinal OCT Images

As the population ages and people’s lifestyles change,vision loss and fundus diseases are becoming more common.Fundus images are widely used for screening and diagnosis of ophthalmic diseases because of their safety and low cost.A high-resolution image is necessary to facilitate efficient and accurate diagnosis.The acquisition of highresolution images has high requirements on the device hardware,so the problem of image resolution from the perspective of deep learning has become a hot research topic nowadays.This thesis will analyze the features of retinal OCT images specifically,make the collected OCT images into a dataset,and design two reconstruction algorithms,EMAFFN and MARN,for the problems of inadequate feature extraction and large number of parameters that are easy to occur in super-resolution reconstruction network and the features of retinal OCT images,and build a super-resolution reconstruction system for retinal OCT images,as follows:(1)In view of the problems of inadequate feature extraction and large number of parameters in the super-resolution reconstruction network,an efficient multiattentive feature fusion network EMAFFN is proposed.This network uses the optimization methods of progressive feature fusion block PFFB,efficient multiattentive block EMAB,and multi-scale perceptual field block RFB＿x,which can enhance the feature transfer in the deep-level network,strengthen the network’s extraction of high-frequency information capability,and the retention of extracted deep-level feature information.Experiments show that EMAFFN has the highest average PSNR value of 37.93 d B and SSIM value of 0.9609 on the Set5 dataset;it works better on the OCT image dataset,and thus has certain generalization property.(2)In view of the characteristics of excessive small feature information,poor circulation of high and low dimensional information,and large differences in feature information on each channel in retinal OCT images,the multi-attentive residual network MARN is proposed for the improvement of EMAFFN.the PFFB in EMAFFN is replaced by the residual group RG to further enhance the communication of high and low dimensional feature information in OCT images;the multilayer residual block MRB and efficient bottleneck attention EBA can optimize the problem that EMAB focuses too much on high-frequency features of the image and enhance the extraction of lowdimensional feature information in OCT images.Compared with before optimization,the PSNR and SSIM values of MARN on the public dataset Set5 are improved by 0.32% and0.02%,respectively,and the parameter volume is reduced by 0.27M;it can better recover the texture contour of the lesion location on the retinal OCT image dataset,which has certain advantages compared with many models.(3)The reconstruction system was designed,including the OCT image acquisition system and the reconstruction software interface.The OCT images were acquired by the OCT image acquisition system,and the acquired images were reconstructed by superresolution using the system software.Figure [58] Table [10] Reference [81]...