In the diagnosis of glaucoma, segmentation of optic cup and optic disc based on digital fundus image is a common diagnostic method. In order to segment the cup and disc accurately, we proposed a segmentation method based on the improved U-Net. Compared with the traditional U-Net, a residual block was used to improve the down sampling part, and convolution part was used to improve the skip connection, so that the network could obtain more sufficient feature information. The Dice and IOU of the optic disc segmentation model and the optic cup segmentation model on DRISHTI-GS data set reached 98.3% and 97.2%, 93.2% and 88.5%.