Vision-based iris and pupil detection

Abstract

Vision-based iris/pupil localisation is a necessary component in robust eye-gaze tracking as well as biometrics. However, this localisation becomes challenging when hindering situations are involved. Such cases include uneven lighting conditions containing specular reflections and shadows as well as low resolution images, non-frontal head poses and iris occlusion by the eyelids. The objective of this project entails an extensive review of the available literatures pertaining to iris/pupil centre localisation to identify the state-of-the-art. The preferred methods are implemented using MATLAB and their performance is tested and compared to determine the best method. The methods selected from the literature review are categorised into appearance-, shape-, and feature-based approaches in the following manner. To begin with, the appearance based method of Cristina and Camilleri [19] was chosen which utilises Bayes’ classification. For the shape-based approach, the Circular Hough Transform [9] was chosen and is performed after Canny edge detection on the result of the Bayes’ classification of Cristina and Camilleri [19]. Then, to examine the feature-based approach, the Fast Radial Symmetry Transform proposed by Loy and Zelinsky [13] was implemented. To improve the robustness of the methods, the specular reflections within the iris were removed as they may hinder the iris centre localisation process. This reflection removal was implemented based on the modified bilinear interpolation method proposed by He et al. [18]. The testing of the methods was carried out on the FERET image database [46], which was selected due to the wide-view face images containing both frontal-facing and tilted subjects, the diversity of subjects as well as the varying lighting conditions and backgrounds. These characteristics imply a high level of complexity which would serve to adequately test the aforementioned methods. It was deduced that all the methods produced relatively accurate results, with the Fast Radial Symmetry Transform together with the reflection removal yielding the best results, represented by an average mean square error of 5.16.