Detection of Glasses in Near-Infrared Ocular Images

Abstract

Eyeglasses change the appearance and visual perception of facial images. Moreover, under objective metrics, glasses generally deteriorate the sample quality of near-infrared ocular images and as a consequence can worsen the biometric performance of iris recognition systems. Automatic detection of glasses is therefore one of the prerequisites for a sufficient quality, interactive sample acquisition process in an automatic iris recognition system. In this paper, three approaches (i.e. a statistical method, a deep learning based method and an algorithmic method based on detection of edges and reflections) for automatic detection of glasses in near-infrared iris images are presented. Those approaches are evaluated using cross-validation on the CASIA-IrisV4-Thousand dataset, which contains 20000 images from 1000 subjects. Individually, they are capable of correctly classifying 95-98% of images, while a majority vote based fusion of the three approaches achieves a correct classification rate (CCR) of 99.54%.