Automatic extraction of retinal blood vessel based on matched filtering and local entropy thresholding

Due to some tiny structures and blurred boundaries of retinal vessels, especially with the low illumination contrast and strong noise resulted from retinal image collecting, it is difficult to automatically extract vessels from retinal images. In this paper, a novel automatic extraction algorithm for retinal vessels is proposed. Firstly, a non-uniform illumination correction method is used together with CLAHE method for contrast enhancement, and the image reconstruction by linear opening is executed to remove poorly contrasted exudates and some small lesions. Then, the retinal vessel image is enhanced by matched filtering using 2D Gaussian kernel with four different values of a parameter to cover thick and thin retinal blood vessel structures. Finally, the local entropy thresholding is used for the final extraction of blood vessels. Experimental results from the publicly available DRIVE and STARE databases show the better comprehensive performance of the proposed algorithm on pathological and healthy retinal images.