基于面部生理信号的人脸活体检测

Face Liveness Detection Base on Facial Signal

针对现有人脸活体检测算法生理特征鲁棒性差的问题,提出一种新的基于面部生理信号的人脸活体检测算法。首先,利用远程光体积变化描记术从面部图像中提取由心脏跳动引起的面部生理信号;之后,对原始的面部信号使用带通滤波进行去噪,再进行快速傅里叶变换将信号转换至频域,提取新的频谱特征;最后,采用机器学习模型实现二分类,区分真实人脸和虚假人脸。结果表明,在Replay-Attack数据库上,应对打印、屏显两种欺诈攻击类型,最佳准确率达到99.15%。该算法在屏显攻击场景下展现出优越的性能,为进一步深化人脸反欺诈算法的研究奠定了坚实基础。

In view of the poor robustness of existing face detection algorithms,a new face detection algorithm based on facial physiological signals is proposed in this paper.First,remote photoplethysmography is used to extract facial physiological signals caused by heart beats from facial images.After that,the original facial signal is denoised by band-pass filtering,and then the signal is converted to frequency domain by fast Fourier transform to extract new spectral features.Finally,machine learning model is used to realize binary classification and distinguish real faces from false faces.The results show that on Replay-Attack database,the best accuracy rate is 99.15% against two types of fraud attacks:print and screen display.The algorithm shows excellent performance in the screen attack scenario,which lays a solid foundation for further deepening the research of face anti-fraud algorithm.