Robust iris segmentation based on learned boundary detectors

Abstract

Iris segmentation aims to isolate the valid iris texture regions useful for personal identification from the background of an iris image. Most state-of-the-art iris segmentation methods are based on edge information. However, generic edge detection methods may generate a large number of noisy edge points which can mislead iris localization. Therefore a robust iris segmentation method based on specific edge detectors is proposed in this paper. Firstly, a set of visual features including intensity, gradient, texture and structure information is used to characterize the edge points on iris boundaries. Secondly, AdaBoost is employed to learn six class-specific boundary detectors for localization of left/right pupillary boundaries, left/right limbic boundaries and upper/lower eyelids respectively. Thirdly, inner and outer boundaries of the iris ring are localized using weighted Hough transforms based on the output of the corresponding detectors. Finally, the edge points on the eyelids are detected and fitted as parabolas by robust least squares fitting. Extensive experiments on the challenging CASIA-Iris-Thousand iris image database demonstrate the effectiveness of the proposed iris segmentation method.