A dense cascade convolution and self attention feature aggregation network was constructed for the segmentation of retinal vascular images, addressing the difficulties in segmenting small blood vessels and the occurrence of fractures during the segmentation process. The network utilizes multi-scale dense convolution combined with self attention mechanism; To better extract complex feature information of retinal small blood vessels, a dense aggregation module is constructed as the backbone network of the U-shaped network; Embedding self attention patches and multi-scale aggregation modules at the bottom layer of the network to enhance receptive fields and obtain high-dimensional semantic feature information; The feature aggregation module is used in the skip connection part of the model to improve the segmentation accuracy of the model. The experimental results show that on the DRIVE public dataset, the F1 score of the network reaches 83.19%, the accuracy ACC score reaches 97.11%, and the AUC value reaches 98.94%; On the CHASE-DB1 and STARE datasets, compared with Unet, DUNet, SA Unet, and FR Unet networks, the AUC index of this network has achieved the best results so far. Using this network for retinal vessel segmentation, the accuracy and robustness of segmentation have been improved to varying degrees, achieving excellent results in small vessel segmentation and its generalization ability.