基于微小运动的人脸活体检测算法  

FACE ANTI-SPOOFING ALGORITHM BASED ON TINY MOTION

作　　者：崔家礼[1] 郭华[1] 贾瑞明[1] Cui Jiali;Guo Hua;Jia Ruiming(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《计算机应用与软件》2024年第7期150-158,191,共10页Computer Applications and Software

基　　金：国家重点研发计划项目(2017YFB0802300)。

摘　　要：针对人脸识别过程中容易受到打印攻击、视频重放攻击的问题,提出一种利用微小运动的人脸活体检测算法。该算法对视频进行运动放大处理以增强微小运动;使用运动强度和运动方向描述微小运动,生成两种运动特征图并进行融合;使用基于注意力机制的VGG16网络进行真伪判别。在Replay-attack数据集上半错误率(HTER)为1.35%,在CASIA FASD数据集上等错误率(EER)为2.5%,证明了微小运动对人脸防伪的有效性。Aiming at printing attacks and video replay attacks in face recognition,this paper proposes a face anti-spoofing algorithm based on tiny motion.We perform motion amplification algorithm to the video to make tiny motion more obvious.Two motion feature maps were created and fused by using motion intensity and motion direction to describe the tiny motion.A VGG16 network based on attention mechanism was adopted to distinguish the real faces and fake faces.The half total error rate(HTER)of our algorithm on the Replay-attack dataset is 1.35%and the equal error rate(EER)on the CASIA FASD dataset is 2.5%,which proves that the micro motion is effective to prevent facial counterfeiting.

关 键 词：人脸活体检测 运动放大 面部运动 运动特征图 VGG16 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]