In this paper, a hierarchical image matting model is proposed to extract blood vessels from fundus images. More specifically, a hierarchical strategy utilizing the continuity and extendibility of retinal blood vessels is integrated into the image matting model for blood vessel segmentation. Normally the matting models require the user specified trimap, which separates the input image into three regions manually: the foreground, background and unknown regions. However, since creating a user specified trimap is a tedious and time-consuming task, region features of blood vessels are used to generate the trimap automatically in this paper. The proposed model has low computational complexity and outperforms many other state-ofart supervised and unsupervised methods in terms of accuracy, which achieves a vessel segmentation accuracy of 96:0%, 95:7% and 95:1% in an average time of 10:72s, 15:74s and 50:71s on images from three publicly available fundus image datasets DRIVE, STARE, and CHASE DB1, respectively.