Face anti-spoofing with multi-color double-stream CNN

Abstract

Previous methods on face anti-spoofing rarely pay attention to the difference of multi-channel chrominance between genuine and fake faces, or only use hand crafted features, which cannot effectively fuse multi-channel chrominance information. This paper uses CNN (convolutional neural network) features instead of hand crafted features for face anti-spoofing. In order to fuse more discriminative chrominance information, this paper proposes a novel face anti-spoofing method based on a double-stream CNN. Through the jointly modeling of features from global face image and local patches, as well as integrating the features of two different color spaces, i.e. YCbCr and HSV, we explore the discriminative representation for face anti-spoofing. Extensive experiments on benchmarks including CASIA-FASD and Replay_Attack show that our method can achieve state-of-the-art performance. Specifically, 1.79% of EER (Equal Error Rate) on the CASIA-FASD, 0.29% of EER on the Replay_Attack database are achieved.