A comprehensive investigation into sclera biometrics: A novel dataset and performance study

Abstract

Ocular biometrics is the study of the applicability of various ocular modalities in different tasks, most prominently identity recognition. It can be useful in various applications such as surveillance systems, forensics or authentication systems. However, the existing datasets for ocular research are often inappropriate for the study of all three of the visible modalities – the sclera, the iris and the periocular region. In this work, we present a novel dataset of high-quality eye images captured in the visible spectrum appropriate for the study of all three modalities. We have performed an analysis of the covariance with several state-of-the-art recognition methods, studying the performance of the approaches not only on the dataset itself but also across different image resolutions and gaze directions. The results of this comprehensive study give insight not only into the general usability of our dataset, but also into the effects of different image resolutions and gaze directions on the accuracy of sclera-based recognition methods. Our experiments show that deep networks outperform handcrafted approaches in sclera recognition both in terms of overall performance as well as the robustness to lower resolutions and missing gaze directions.