The structure and morphology of retinal blood vessels are essential for computer-aided systems to diagnose ophthalmological diseases. To solve the problem of low precision of tiny blood vessel segmentation, we propose an improved U-Net algorithm combining residual dense block and three-terminal attention gate block. First of all, we combine the residual block with the dense block to make full use of the features of each layer and improve the ability to extract the characteristics of tiny blood vessels. In the decoding stage, we introduce a three-terminal attention gate block. And the spatial attention mechanism is used to adaptively correct the features, suppress background noise and highlight the target area. At the same time, we use high-level semantic features to improve the segmentation effect of tiny blood vessels through multi-scale feature fusion. Finally, to obtain the multi-scale features of blood vessels, we introduce deformable convolution into the network structure and increase the receptive field without increasing the parameters. The experimental results based on the DRIVE and STARE data sets show that the sensitivity, specificity, accuracy and AUC (Area Under Curve) of the proposed network are 81.26%/82.57%, 98.20%/98.37%, 96.70%/97.51% and 98.12%/98.41%, which are better than existing advanced algorithms.