Discriminative Color Space Learning for Face Anti-Spoofing via Convolutional Neural Networks

Abstract

Face spoofing detection is gaining an increasing attention in the biometric research. Various approaches have been proposed in the literatures. In these methods, the color variation of facial regions, caused by the defect of medium of fake face, is a vitally important clue. The traditional color spaces (e.g. RGB, HSV and YCbCr) are used in many spoofing detection approaches, however, it is not very discriminative to distinguish real and fake faces in these existing color spaces. So, in this paper, we propose a novel method to learn a new color space, which is suitable for face anti-spoofing and can be discriminative between real and fake faces. Different from other color learning methods, our novel method is based on convolutional neural networks and can nonlinearly project the real and fake face images into a distinguishable color s-pace. Extensive experiments are conducted on two publicly available databases, showing very interesting performance compared to other existing color spaces and state-of-the-art.