Retinal vascular image segmentation is an important and difficult task in medical image analysis, and it is difficult for conventional methods to detect the small and dense vascular structures effectively. To solve this problem, a novel high-precision retinal vascular segmentation method that combines Swin transformer block (STB) and full-scale attention skip connection technology (FSASC) was proposed. By constructing a U-shaped encoder-decoder network, the proposed method realizes self-attention from local to global, so that our model can pay more attention to the key vascular features. FSASC technology is used for fusing different features, which provides a simple and powerful mechanism for our model to learn multi-scale semantic and spatial information. The proposed method was tested by using open datasets DRIVE and STARE. The experimental results show that the proposed method can achieve high-quality and high-precision segmentation for retinal vascular structures. Compared with Unet and other methods, the proposed method has better performance in both the detailed feature segmentation and the segmentation accuracy.