Retinal Vessel Segmentation Network Based on Improved U-Net

Abstract

To solve the difficult segmentation problem in the division of retinal vascular due to blurred vessel contours and complex vessel peripheral details. In this article propose a double residence network DRANet. It contains an attention structure. DRANet is an improvement to the U-Net network. In case of gradient inflation and gradient vanishing due to network deepening, a double residual module is inserted in the encode and decode stages. The Double Residual Spatial Attention Structure (DRSA) is added at the bottom of the networks to capture more signatures. At the same time, DropBlock is introduced to alleviate the overfitting of the network. Besides, a channel attention block has been added to the skip link to suppress the expression of unnecessary features and produce higher accuracy. Achieved better segmentation results on the communal retinal blood vessel data sets DRIVE and CHASE_DB1.