摘要
为了提升人脸识别系统判别图像真实性的能力,针对较难检测到未知的活体人脸攻击问题,提出了一种采用超复数小波生成对抗网络的活体人脸检测算法。采用4个不同类型的数据集,随机选择3个作为训练集,另一个作为测试集,形成训练时未知的活体人脸。训练集视为3个源域,输入到超复数小波生成对抗网络中,使一个特征生成器与3个判别器进行对抗,当特征生成器成功欺骗过3个判别器时,形成具有3个源域共享且区别于3个源域的特征空间,能够检测到不同于源域的人脸特征。在判别器上设置了域间和域内的三元组约束函数,以此提高判别器的性能,将超复数小波的细节子带图与卷积网络联合,学习图像多个方向的细节纹理特征,用来提升判别器鉴定活性人脸特征的能力。由于真假人脸的远程光电体积描记术和深度图都具有较大的差异,所以将其嵌入到特征空间中,增强生成特征空间检测人脸特征的泛化性能,形成通用的特征空间。在该特征空间中使用测试集进行判别分类,得到真假人脸识别结果。实验结果表明,在CASIA-FASD、Replay-Attack和NUAA数据集上,所提算法的接受者操作特性曲线下的面积分别为84.65%、86.06%、91.21%,半错误率分别为24.05%、21.05%、15.01%,均高于对比算法的结果。
Aiming at the problem that the existing anti-spoofing detection algorithms difficultly detect unknown domain attacks,this paper proposes a face anti-spoofing detection algorithm using generative adversarial networks with hypercomplex wavelet transform to improve the ability of face recognition system to determine the face anti-spoofing.Four different types of datasets are adopted,in which three datasets are randomly selected as the training ones,and the remainder as the test dataset to serve for unknown face anti-spoofing detection during training.The training datasets are regarded as three source domain datasets,which are input into the generation network to make a feature generator against three discriminators.When this feature generator successfully deceives the three discriminators,a feature space sharing three source domains and being different from these domains is formed to detect the characteristics of the unknown domain data.The triple constraint functions inter-class and intra-class are set up to improve the performance of the discriminator,and the detailed subbands of the hypercomplex wavelet transform and the convolution network are combined to learn the detailed features in multiple directions of images.Then the depth map and remote photo plethysmography signal are embedded in the feature space to enhance the generalization performance of the generated feature space for the living face features.The test dataset for discriminative classification in the feature space is used to obtain live/fake results.The results show that on the CASIA-FASD,Replay-Attack and NUAA datasets,the proposed algorithm gets the AUC of 84.65%,86.06%and 91.21%,the HTER of 24.05%,21.05%and 15.01%,which are higher than those of the comparative algorithms.
作者
李策
李兰
宣树星
杨静
杜少毅
LI Ce;LI Lan;XUAN Shuxing;YANG Jing;DU Shaoyi(College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;College of Automation Science and Engineering,Xi’an Jiaotong University,Xi’an 710049,China;College of Artificial Intelligence,Xi’an Jiaotong University,Xi’an 710049,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2021年第5期113-122,共10页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(61971343,61866022)
陕西省重点研发计划高校联合项目(2020GXLH-Y-008)。
关键词
活体人脸检测
超复数小波
生成对抗网络
face anti-spoofing detection
hypercomplex wavelet
generative adversarial networks
作者简介
李策(1974-),男,教授;通信作者:杜少毅,男,教授。
