A Novel Application of Multiscale Low-Rank Image Decomposition for Optic Disc Localization

Abstract

We propose optic disc localization using a multiscale low-rank image decomposition technique, which separates the optic disc from other features on the retina. By decomposing the input image as the sum of $L$ images, each at a specific scale and minimizing the rank of these blocks to emphasize the details at that scale. At one of the scales, the optic disc is isolated. Thereafter, we employ blob detection algorithms to localize the optic disc. The proposed method has been tested on seven fundus image databases amounting to a total of 2035 images, achieving an accuracy of 100%, 100%, 96.15%, 97.75%, 98.15%, 96.85% and 96.41% on DRIVE, DRIONS-DB, DIARETDBO, DIARETDB1, IDRiD-1, IDRiD-2, and MESSIDOR databases, respectively. Synthetically generated images with features at different scales were used to test the rate of convergence of the method as well.