[1]颜增显,孔 超,欧卫华*.基于多模态融合的人脸反欺骗算法研究[J].计算机技术与发展,2022,32(04):63-68.[doi:10. 3969 / j. issn. 1673-629X. 2022. 04. 011]
　YAN Zeng-xian,KONG Chao,OU Wei-hua*.Research of Face Anti-spoofing Algorithm Based on Multi-modal Fusion[J].,2022,32(04):63-68.[doi:10. 3969 / j. issn. 1673-629X. 2022. 04. 011]
点击复制

基于多模态融合的人脸反欺骗算法研究()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年04期
页码:
63-68
栏目:
图形与图像
出版日期:
2022-04-10

文章信息/Info

Title:
Research of Face Anti-spoofing Algorithm Based on Multi-modal Fusion
文章编号:
1673-629X(2022)04-0063-06
作者:
颜增显1 ; 孔 超2 ; 欧卫华2*
1. 广西现代职业技术学院,广西 河池 547000;
2. 贵州师范大学,贵州 贵阳 550025
Author(s):
YAN Zeng-xian1 ; KONG Chao2 ; OU Wei-hua2*
1. Guangxi Modern Polytechnic College,Hechi 547000,China;
2. Guizhou Normal University,Guiyang 550025,China
关键词:
人脸反欺骗; 多模态融合; 多模态共享分支; 多模态通道注意力融合; 多模态特征
Keywords:
face anti-spoofing; multi-modal fusion; multi-modal shared branch; multi-modal channel attention fusion; multi-modal features
分类号:
TP301. 6
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 04. 011
摘要:
人脸反欺骗技术可以准确判断捕获的人脸图像是真实人脸还是虚假人脸,是人脸识别系统安全的重要保障。 传统的人脸反欺骗方法主要是利用手工设计的特征,如 LBP、HoG、SIFT、SURF 和 DoG 来刻画真实人脸和虚假人脸之间的不同特征分布,但人工设计的特征难以适应无约束环境下 ( 如光照、背景的变化) 的人脸反欺骗问题。 鉴于此,该文提出一种多模态融合卷积神经网络模型,通过融合不同模态上的人脸特征来实现鲁棒的人脸反欺骗。 首先根据通道注意力网络设计了多模态共享分支网络来实现特征提取过程中不同模态间的信息交互,然后在通道注意力融合网络的基础上提出了多模态通道注意力融合网络来融合不同模态的特征,最后利用融合后的多模态特征进行分类。 在 CASIA-SURF 数据集上的大量实验结果表明,与主流的多模态人脸反欺骗方法( multi-scale fusion) 相比,该方法在 APCER 和 ACER 指标上分别降低了 1. 1% 和 0. 4% ,充分证明该方法可以有效融合不同模态的特征,提高模型的鲁棒性。
Abstract:
Face anti-spoofing technology can accurately determine whether the captured face image is a real face or a false face,which is an important security guarantee for face recognition system. Traditional face anti-spoofing methods mainly use hand -crafted features,such as LBP,HoG, SIFT,SURF and DoG,to characterize the differences of feature distributions between real faces and spoofing faces,but the features of artificial design is difficult to adapt to face anti-spoofing in unconstrained environment ( such as illumination and background change) . In view of this,we propose a multi - modal fusion convolutional neural network model to achieve robust face anti -spoofing by fusing features from different modalities. Firstly,according to the channel attention network,a multi - mode shared branch network is designed to realize the information interaction between different modalities in the process of feature extraction,then based on the channel attention fusion network,? ? a multi - modal channel attention fusion network is proposed to fuse the features of different modalities. Finally,the fused multi-modal features are used for classification. A large number of experimental results on CASIA-SURF datasets show that compared with the mainstream multi - modal face anti - spoof method ( multi - scale fusion) , the proposed method reduces APCER and ACER by 1. 1% and 0. 4% ,respectively. It is fully proved that the proposed method can effectively integrate the features of different modalities and improve the robustness of the model.

相似文献/References:

[1]邵曦,陶凯云. 基于音乐内容和歌词的音乐情感分类研究[J].计算机技术与发展,2015,25(08):184.
　SHAO Xi,TAO Kai-yun. Research on Music Emotion Classification Based on Music Content and Lyrics[J].,2015,25(04):184.
[2]于 翔,周 波.基于多模态融合的室内人体跟踪技术研究[J].计算机技术与发展,2023,33(02):38.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 006]
　YU Xiang,ZHOU Bo.Research on Indoor Human Tracking Technology Based on Multi-modal Fusion[J].,2023,33(04):38.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 006]
[3]段毛毛,连培榆,史海涛.动态视音场景下问答模型研究[J].计算机技术与发展,2024,34(03):163.[doi:10. 3969 / j. issn. 1673-629X. 2024. 03. 024]
　DUAN Mao-mao,LIAN Pei-yu,SHI Hai-tao.Research on Question and Answer Models in DynamicAudio-visual Scenarios[J].,2024,34(04):163.[doi:10. 3969 / j. issn. 1673-629X. 2024. 03. 024]

更新日期/Last Update: 2022-04-10