应用并联卷积神经网络的人脸防欺骗方法  

Face Anti-spoofing Algorithm Applying with Parallel Convolutional Neural Network

作　　者：李冰[1] 王宝亮[1] 由磊[1] 杨沫[1] 

机构地区：[1]天津大学电子信息工程学院,天津300072

出　　处：《小型微型计算机系统》2017年第10期2187-2191,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金青年项目(61202380)资助

摘　　要：针对人脸活体检测中人工提取的纹理特征不全面的问题,本文首次提出基于并联卷积神经网络(Parallel convolutional neural netw ork,P-CNN)和极限学习机(Extreme learning machine,ELM)的人脸防欺骗方法.算法采用SM QT-SNOW人脸检测器定位人脸,并加入人脸对齐算法优化人脸框,得到精准的人脸图像;并将人脸的灰度图和局部定向模式分别作为两个不同结构的网络的输入;然后采用主成分分析对每个网络的全连接层的输出分别降维后级联;最后将级联的特征向量送入ELM判定人脸的合法性.在NUAA和REPLAY-ATTACK数据库上实验,最高准确率分别为99.96%和99.98%,最高受试者工作特征曲线下方面积(AUC)均为1.实验结果表明算法相比其他方法,其特征维数小,准确率高以及应对不同介质攻击的泛化能力强.Facing the problem the hand-crafted texture features were not comprehensive for face liveness detection, a face anti-spoofing approach based on parallel convolutional neural network （ P-CNN ） and extreme learning machine （ ELM ） was proposed. The algorithm uses SMQT-SNOW face detector to localize face,combined with face alignment algorithm to refine face bounding box to obtain a precise face localization, and the input images of the two different structures of CNN are grayscale images and local directional patterns respectively. Then we use principal component analysis method to reduce feature dimensions of output by fully-connected layer in each CNN. Finally we feed the concatenated feature vectors into extreme learning machine to discriminate its validity of faces. It was conducted on NUAA and REPLAY-ATtACK database, the highest accuracy were 99.98% and 96.26% respectively and the highest Area Under Curve was both 1. Experimental results show that compared with other present methods, our algorithm achieves high accuracy, low feature dimensions and preferable generalization ability for detecting different medium attacks.

关 键 词：人脸防欺骗 并联卷积神经网络 主成分分析 极限学习机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]