Abstract: In order to address the problem of vascular branch breakage existing in the process of retinal vessel segmentation at present, a non-local Unet model (N-Unet) is proposed. The model retained the encoder-decoder symmetric structure, and introduced non-local blocks at the encoder stage, which made the model pay attention to non-local information in the process of feature extraction and better capture the relationship between non-adjacent pixels in the image. This model was evaluated on the public dataset DRIVE, and gained 0.952 3 accuracy, 0.802 1 sensitivity, 0.974 3 specificity, and 0.894 9 AUC, respectively. Experimental results show that this method performs well in solving the problem of branch breakage in the process of blood vessel segmentation, and has research significance.