结合多模态和注意力机制的人脸活体检测算法  

Face Anti-Spoofing Algorithm Combined with Multi-Modal and Attention Mechanisms

作　　者：周航 蔡茂国[1] 唐剑兰 徐翔 ZHOU Hang;CAI Mao-guo;TANG Jian-lan;XU Xiang(College of Electronics and Information Engineering,Shenzhen University,Shenzhen Guangdong 518000,China)

机构地区：[1]深圳大学电子与通信工程学院,广东深圳518000

出　　处：《计算机仿真》2023年第9期202-207,共6页Computer Simulation

基　　金：国家自然科学基金(61872244)。

摘　　要：针对常见人脸活体检测算法从单一RGB模态图像中提取的特征较为单一等问题,提出一种结合注意力机制的多模态双流活体检测算法SE-FeatherNet。首先,基于改进的FeatherNet网络,分别从Intel RealSense300相机拍摄的Depth、近红外(Infrared Radiation, IR)图像中提取特征,然后将获取的特征图叠加在一起进行特征层融合,最后从融合的特征图中继续提取特征,并加入自注意力机制。针对单一评价标准存在偶然性的问题,使用多指标评价标准来验证模型的准确性。仿真结果表明,所提算法在CASIA-SURF数据集中的等错误率(equal error rate, EER)为1.341%,半错误率(half total error rate, HTER)为1.537%,真正类率TPR@FRR=10e-2为96.99%,并且参数量仅仅为0.58M。融合多模态的特征信息可以获取更低的错误率,改进的网络保证了算法的高效性和时效性以满足边缘设备算力有限的需求。Aiming for the characteristics extracted from a single RGB modal image for common human face antispoofing algorithms,a multi-modal dual flow living detection algorithm SE-FeatherNet combined with attention mechanism is proposed in this paper.First,based on the improved FeatherNet network,depth and near infrared Features were extracted from infrared radiation(IR)images,and then the obtained feature maps were superimposed for feature layer fusion.Finally,the features were extracted from the fused feature map and self attention mechanism was added.To address the issue of randomness in a single evaluation criterion,multiple indicator evaluation criteria were used to verify the accuracy of the model.The simulation experiment results show that,the equal error rate(EER)of the proposed algorithm in the CASIA-SURF data set is 1.341%,the half total error rate(HTER)is 1.537%,and the true class rate is TPR@FRR=10e-2 is 96.99%,and the parameter amount is only 0.58M.The fusion of multi-modal feature information can obtain lower error rate.The improved network ensures that the efficiency and timeliness of the algorithm can meet the needs of limited computing power of edge devices.

关 键 词：多模态 深度神经网络 特征融合 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]