Statistics of local surface curvatures for mis-localized iris detection

Eye detection is a hot research topic in computer vision for its wide applications in human-computer interaction, face and iris recognition, etc. However, robust eye detection is still a grand challenge due to the numerous appearance variations of eye images in real-world applications. In this paper, we present a novel local surface curvature analysis method to deal with this problem. Firstly, by regarding an eye image as a 2D surface in 3D space, we propose to use the histogram of local surface curvatures as the general representation of eye pattern. Then, a SVM classifier is employed for eye detection using the histogram vectors of eye and non-eye samples. Extensive experiments are performed and the results show that the proposed method achieves state-of-the-art performance in eye detection. In particular, it is more efficient in mistakenly localized iris detection.