Vessel local characteristics such as noise, illumination, and direction vary significantly in a fundus image, making it difficult to segment the vessel tree structure as a whole. To facilitate vessel detection, an alternative procedure proposed here, whereby retinal vessels first classified into two categories, large and small. Then, for its unique characteristics, each group has been processed with its own enhancement and detection filter. The sensitivity of the proposed method is boosted by capturing tiny vessels through a directional filter bank followed by its associated triple-stick filtering. Additionally, the specificity of the proposed method is enhanced through noise suppression attributed largely to the proposed BM3D filtering and multi-scale line detection approach. As a result, the detection accuracy on the DRIVE, STARE, and CHASE DB1 datasets is significantly improved, with scores of 0.9610, 0.9586, and 0.9578, respectively.
