The invention provides a retinal vessel segmentation method based on an improved U-Net network. Image enhancement is performed on a color eye fundus image, so that the contrast ratio between a blood vessel and a background in the image is improved, and a training data set is amplified. A U-Net encoder-decoder structure is used as a basic segmentation framework, a dense convolution block and a CDBR layer structure are designed to replace a traditional convolution block, learning of multi-scale feature information is achieved, and the feature extraction capacity of the model is improved. Meanwhile, an attention mechanism is introduced at a jump connection part of the model, so that the model is enabled to allocate weights again, the importance degree of a feature channel is adjusted, noise is suppressed, the problem of blood vessel information loss in an up-sampling process at a decoder end is solved, and a GAB-D2BUNet network model is constructed based on the above technologies. According to the method, an internationally disclosed retina fundus blood vessel data set DRIVE is adopted for training, and finally the optimal segmentation model is reserved to verify the segmentation performance of the model. The retina fundus blood vessel segmentation method achieves the task of accurately segmenting the retina fundus blood vessel, and has better segmentation performance.