关键词: 人脸反欺骗, 多模态特征, 多输入卷积神经网络, 模型鲁棒性, 融合

Abstract: Most of the previous face anti-spoofing methods employed the manually extracted features or only face features on a single modality, and rarely noticed the differences in multi-channel chroma. Therefore, the face anti-spoofing model was so low in robustness that it could not effectively distinguish between real and fake faces. In view of this, a convolutional neural network (CNN), which substituted the hand-crafted features extraction, was utilized as a feature extractor, and an effective multi-input CNN model was proposed to fuse face features on multiple modalities to achieve more robust face anti-spoofing. Through the joint modeling of features concerning two different color features (i.e. HSV and YCbCr) of the local face patches, as well as the temporal feature, the optimal robust representation of face anti-spoofing was explored. A large number of experiments on two benchmarks of REPLAY_ATTACK and CASIA-FASD show that this method can attain the most advanced performance. Specifically, 0.23% error rate (ERR) and 0.49% half total error rate (HTER) were obtained on REPLAY_ATTACK, and 1.76% error rate and 3.05% half total error rate were yielded on the CASIA-FASD database.

Key words: face anti-spoofing, multimodal feature, multi-input convolutional neural network, model robustness, fusion