Accurate face localisation for faces under active near-IR illumination

In this paper we propose a novel approach to accurate face localisation for faces under near-infrared (near-IR) illumination. The circular shape of the bright pupils is a scale and rotation invariant feature which is exploited to quickly detect pupil candidates. As the first step of face localisation, a rule-based pupil detector is employed to find candidate pupil edges from the edge map. Candidate eye centres for each eye are selected from the neighborhood of corresponding pupil regions and sorted based on the similarity to eye templates. Two support vector machine (SVM) classifiers based on eye appearance are employed to validate those candidates for each eye individually. Finally candidates are further validated in pair by an SVM classifier based on global face appearance. In the experiment on a near-IR face database with 40 subjects and 48 images per subject, 96.5% images are accurately localised using the proposed approach.