基于3D卷积神经网络的活体人脸检测  

3D Convolutional Neural Network Based on Face Anti-Spoofing

作　　者：甘俊英[1] 李山路 翟懿奎[1] 刘呈云 

机构地区：[1]五邑大学信息工程学院,广东江门529000

出　　处：《信号处理》2017年第11期1515-1522,共8页Journal of Signal Processing

基　　金：国家自然科学基金项目(61372193;61070167);广东高校优秀青年教师培训计划资助项目(SYQ2014001);广东省特色创新项目(2015KTSCX143;2015KTSCX145);广东省青年创新项目(2016KQNCX171)

摘　　要：非法入侵者通过伪装人脸骗取系统认证,给人脸认证系统带来了严重的威胁。因此,活体人脸检测成了人脸认证系统走向实用必须解决的一个重要课题。现有活体人脸检测方法多为基于照片的人脸攻击方面的研究成果,对于基于视频的人脸攻击,效果并不理想。3D卷积神经网络(Convolutional Neural Network,CNN)具有深度学习的特点,能自动学到图像的分布式特征表示;与2D卷积相比,它能学到连续视频帧的动作信息。本文结合3D卷积神经网络的特性,提出利用3D卷积实现视频人脸伪装检测。通过提取3D卷积神经网络最后全连接层学到的时间空间特征,训练SVM(Support Vector Machine)分类器,实现真实人脸和伪装人脸的分类。实验采用两个人脸伪装公开数据库Replay Attack和CASIA,实现多尺度内部数据库测试和交叉数据库测试。实验结果相对于纹理特征及2D卷积方法有较大提高,可应用于视频人脸攻击的活体人脸检测。Spoofing face can be used to deceive face authentication systems for illegal purposes, and thus it poses a serious threat to the face authentication system. Therefore, it has become an important subject that the face authentication system must be solved in the practical stage. Many existing literatures focus on the study of photo attack. For the video attack, however, the related research efforts remain to be improved. 3D convolution neural network has the characteristics of deep learning, which can automatically learn the representation of the distributed features of image. Compared with 2D convolution, it can learn the motion features of continuous video frames. In combination with 3D convolution, in this paper, 3D convolution neural network （CNN） is proposed to learn features from video frames to implement face anti-spoofing. After training the 3D convolution neural network, the SVM classifier is trained with the spatio-temporal features extracted by the last fully connected layer of the 3D convolutional neural network to carry out the classification of the real face and the disguised face. In the experiments, two face anti-spoofing public database ReplayAttack and CASIA are used to implement multi-scale internal database test and cross database test. Compared with other methods based on texture features and 2D convolntional neural network, the experimental results have been greatly improved. Which can be applied to living face detection of video attack.

关 键 词：3D卷积神经网络 活体人脸检测 人脸反伪装 社会安全 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]