Structured forest edge detectors for improved eyelid and iris segmentation

Abstract

Edge detection has long figured into iris segmentation algorithms, often providing a first-pass estimate of the inner and outer iris boundaries. Standard edge detectors, however, generally produce too many extraneous edges inside and outside the iris to be used for simple ellipse fitting to robustly find the iris boundaries. Solutions to this problem have been either to have additional filtering to select relevant edges or to design specialized edge detectors that highlight iris boundary edges and suppress irrelevant edge types. This description holds ever so more for eyelid boundaries, which are often very subtle. An edge detector that will trigger on an eyelid boundary will also likely trigger on almost any slight intensity gradient in an image. We seek to solve this problem by learning specialized edge detectors for each type of relevant boundary in an iris image. Using a fast Structured Random Forest approach developed for learning generalized edge detectors, we train detectors for the iris/sclera, iris/pupil, and eyelid boundaries. The results show that learned edge detectors should become part of the standard toolbox for iris segmentation and eyelid boundary detection.