In order to reduce missed and false deteetion of low contrast vessels in retinal vascular skeleton extraction, a mulli-scale retinalvascular skeleton extraction method based on principal curvature and principal direetion is proposed. Firstly, the principal curvature andprincipal direetion of each pixel in the retinal image is extracted after multi-scale Gaussian filtering. Secondly, the loeal maximum pointsin the direction of the maximum principal curvature are estracted at each scale, and the high eontrast vascular center pixel is selected asthe seed point by curvature threshold. Then, the minimum principal direetion is used to track and labeled the low eontrast vessels.Finally, the vascular skeleton estracted at multiple seales are fused. The proposed algorithm is evaluated on DRIVE training dataset,DRIVE test dataset and STARE dataset. The number of missed detection are 89, 97, and 106, respeetively. And the number of falsedeteetion are 99, 101, and 122, respectively, Experimental results show that the proposed method can extract small vascular skeletonswith low contrast, but there are a few miss deteetion for small blood vessels with contrast less than three gray levels, and a few falsedeteetion for fine stripe texture and lesions interference which adheres to blood vessels.