Blood vessels in fundus images contain large number of important diagnosis information, while it is too tedious for clinicians to distinguish. This paper proposes a method to segment blood vessels efficiently. Firstly, non-uniform enhancement was used to balance the non-uniform background induced by light. Then the image was divided into grids. Secondly, optimal threshold was selected for each grid. Finally, blood vessels were segmented in gray images. The experimental results show that the extracted blood vessels have well continuity and accurate center-line which provides useful information support to clinicians.