Abstract: Automatic blood vessel segmentation in images can help accelerating diagnosis and improving the diagnostic performance of less specialized physicians. Because of low contrast between vessels and background, noise, nonuniform illumination in the retinal image, it is hard to completely and accurately extract the vascular structure. This paper presents a new method for vessel segmentation based on pixel specificity. Each pixel is assigned a pixel specificity computed from lines passing through it with different directions, by statistical analysis, an adaptive threshold for pixel specificity can be determined, then the whole image is binarized based on the threshold, those pixels whose specificity values are greater than the threshold are segmented as vessels, otherwise marked as background. The segmentation result based on global threshold is not satisfactory, some true vessel pixels are missed and some background pixels are misclassified. To make the result more precisely, local adjustment is performed to refine the segmentation, i.e. find more vessel pixels, and remove the noise. The image is divided into 16×16 sub-images, within each sub-image, the local segmentation result is corrected by gradient descent local adaptation where a more suitable threshold can be determined based on the local context. All the sub-images are combined together to get the whole segmentation result. Postprocessing is performed by noise filtering to remove the discrete false vessel pixels, collective awareness to search the missing vessel pixels and broken vessel segments linking to make up the gaps, the vascular structure of the image is finally segmented. The method was evaluated on the publicly available DRIVE database, and the performance on the images is better than other existing solutions in literature. The effectiveness and robustness with different image conditions make this blood vessel segmentation method suitable for retinal images analyses such as automated screening for early diabetic retinopathy detection.