The　automatic　segmentation　of　retinal　vessels　plays　an　important　role　in　the　early　screening　of　eye　diseases.　However,　pathological　retinal　images　are　difficult　for　us　to　segment　the　vessels.　In　this　paper,　we　regard　the　vessels　segmentation　task　as　a　multi-label　problem　and　combine　the　preprocessed　method　Gaussian　matched　filter　with　a　new　U-shaped　fully　convolutional　neural　network　called　U-net　to　generate　a　blood　vessels　segmentation　framework.　The　output　of　this　model　can　distinguish　the　vessels　from　background　although　in　the　inadequate　contrast　regions　and　pathological　regions.　The　proposed　method　is　tested　on　a　publicly　available　dataset　of　DRIVE.　Sensitivity,　Specificity,　Accuracy　and　Precision　are　used　to　evaluate　our　method,　and　the　average　classification　accuracy　is　0.9636　on　the　dataset　of　DRIVE.　Performance　results　show　that　our　method　outperforms　the　state-of-the-art　method　for　automatic　retinal　blood　segmentation.