基于语义分割的活体检测算法  

Face anti-spoofing algorithm based on semantic segmentation

作　　者：林云 孙晓刚[1] 姜尧岗 康鑫 解至煊 钟勇[1] LIN Yun;SUN Xiao-gang;JIANG Yao-gang;KANG Xin;XIE Zhi-xuan;ZHONG Yong(Chengdu Institute of Computer Applications,Chinese Academy of Sciences,Chengdu 610041,China;University of Chinese Academy of Science,Beijing 100049,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学,北京100049

出　　处：《吉林大学学报（工学版）》2020年第3期1040-1046,共7页Journal of Jilin University:Engineering and Technology Edition

基　　金：四川省重点研发计划项目(2018GZ0231).

摘　　要：为了提高人脸识别系统的安全性,防止手机、照片中的人脸图像伪装攻击,提出了一种基于语义分割的活体检测算法。首先,通过全卷积神经网络(FCN)对局部人脸区域进行语义分割,并提出了一种带方向的卷积核对网络进行优化。其次,训练了一个快速分类器对语义分割网络的结果进行分类。最后,对深度学习网络进行串联,形成一个端到端的活体识别框架。试验结果表明:本文检测算法在Casia活体数据集和私有数据集上表现突出,在实际项目中泛化能力突出。In order to improve the security of the face recognition system and prevent from fake face attack in mobile phones and photos,a face anti-spoofing algorithm based on semantic segmentation is proposed.First,the local face region is semantically segmented by full convolutional neural network(FCNN),and the convolution kernel with direction is proposed to optimize the network.Secondly,a fast classifier is trained to classify the results of the semantic segmentation network.Finally,the deep learning network is connected in series to form an end-to-end face anti-spoofing framework.The test results show that the performance of the proposed algorithm is outstanding in the Casia live dataset and private dataset collected by CBPM-XINDA,and the generalization ability in the actual project is outstanding.The main work includes the production of data sets and the selection of features(using the combination of deep learning and automatic learning features and manual selection of features),and proposes a method for optimizing convolution kernels and concatenation of multiple networks.

关 键 词：计算机应用 活体检测 深度学习 防欺诈 人脸识别 视频欺诈 图片欺诈 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]