Study Of Retinal Image Enhancement And Segmentation Methods

With the change of people’s lifestyle,various electronic products have penetrated into all aspects of life.Various unhealthy habits and excessive eye fatigue have caused more and more patients with eye diseases.Vision is an important way for people to obtain outside information.Once blinded,it is irreversible,which seriously affects people’s lives.Various eye diseases can cause corresponding changes in the fundus structure.Retinal images are the most direct and effective basis for doctors to diagnose eye diseases.Therefore,research on retinal image analysis is of great significance for the auxiliary diagnosis of eye diseases.Enhancement and segmentation of retinal image are prerequisites for retinal image analysis,and also classic puzzle in the field of image processing.So,the following research is performed on the retinal image.First,an optic disc segmentation algorithm based on ShuffleNet and attention is proposed to extract the retinal optic disc accurately.We use the basic unit of ShuffleNet to replace the convolution in U-Net,which makes the network take both "less parameters" and "high precision" into account.On this basis,the attention mechanism is used to enhance or suppress the low-level features in the identity connection of U-Net,to obtain better features for extracting optic disc and avoid introducing irrelevant noise.In addition,based on the ideas of average Hausdorff distance and weighted Hausdorff distance,a weighted average Hausdorff distance was proposed and used as constraint in the loss function,which restrict the training to get better segmentation result.Experiments on DRIONS-DB,RIM-ONE and DRISHTI GS1 datasets verify the effectiveness of the proposed network and constraints.Second,in order to denoise the retinal image,and facilitate doctors to diagnose.Prior to denoising,the fundus vessels were enhanced first,and an improved Frangi filtering algorithm was proposed to solve the problem of vascular rupture in traditional Frangi filtering.Then denoise the enhanced retinal image,and use Fista algorithm to solve the sparse constraint denoising model,so as to improve the denoising performance of sparse constraint.Finally,the DRIVE dataset is used to compare the sparse constraint models solved by the OMP and Fista algorithms,which verify that the proposed algorithm is more suitable for the enhancement and denoising of retinal images.Third,in order to segment fundus vessel automatically and improve the accuracy as much as possible,a recurrent residual convolutional network with multi-scale ideas and attention is proposed for fundus vessel segmentation.The network has made various improvements to UNet.First,the recurrent convolution is used to replace the convolution in standard U-Net,to allow the network not only learn features from previous layer,but also the current layer.Then use residual block in the convolution to avoid " gradient vanishing " and "degeneration ",which caused by the deepening due to the introduce of recurrent connection.Besides,use the attention to enhance and suppress features,and on this basis,add the idea of multiscale input and prediction,so that the network can obtain multiscale features as much as possible,and constrain the network training from multiple prediction.Finally,experiments on DRIVE,STARE and CHASE_DB1 datasets verify the effectiveness of the proposed network.This paper presents preliminary research on the problems of retinal optic disc segmentation,retinal vessel denoising and segmentation.The experiment results have proved that our algorithms can deal well with retinal optic disc segmentation,fundus vessel denoising and segmentation.