Accurate retinal vascular segmentation supports the treatment of diseases such as diabetes and hypertension. Because of the complex vascular structure of the eye, the complexity of the pathological features leads to many limitations in the accuracy and speed of vascular segmentation. To overcome this problem, an improved U-net segmentation method is proposed, which replaces the convolution module in the U-net network decoder and encoder with a residual module, using a non-local attention module to connect the encoder and decoder. The network model enhances the correlation of pixel information and the ability to extract features without increasing the number of parameters. Finally, the DRIVE dataset was used for comparison and evaluation with the original U-net network, and the model achieved 0.9679、0.9896、0.8245 and 0.8281 of feature detection accuracy, specificity, sensitivity and Dice coefficient on the test set, respectively. The experimental results demonstrate that the proposed network model can perform accurate vascular segmentation of the retina.