三通道多姿态面部正面化方法  

Three-channel multi-pose face frontalization method

作　　者：高峰 张元[1] 谢剑斌 闫玮[2] 郭锐[1] GAO Feng;ZHANG Yuan;XIE Jian-bin;YAN Wei;GUO Rui(School of Data Science and Technology,North University of China,Taiyuan 030051,China;School of Electronic Science,National University of Defense Technology,Changsha 410000,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051 [2]国防科技大学电子科学学院,湖南长沙410000

出　　处：《计算机工程与设计》2024年第5期1490-1498,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62106238);山西省回国留学人员科研基金项目(2020-113);山西省科技成果转化引导专项基金项目(202104021301055)。

摘　　要：针对现有面部正面化网络在复杂环境下难以保留面部显著性特征的问题,提出一种三通道(局部、半全局和全局)面部正面化方法。在TP-GAN算法的原有框架基础上设计半全局网络,融合全局网络和局部网络之间的依赖关系,使生成的正面化图像的分布与真实面部图像更接近;在半全局网络中设计多时空深度注意力模块,促进网络学习到更多面部显著性特征;将所提方法应用于CAS-PEAL-R1数据集和自建数据集,采用Rank-1指标进行评估。实验结果表明,所提方法在所有角度下的Rank-1平均准确率为99.40%,验证添加了多时空深度注意力模块的半全局网络可以有效保留面部显著特征,提高面部匹配准确率。Aiming at the problem that the existing face frontalization networks cannot preserve the salient features of faces in complex environments,a three-channel(local,semi-global and global)face frontalization method was proposed.A semi-global network was designed on the basis of the original framework of the TP-GAN algorithm,and the dependencies between the global network and the local network were fused,so that the distribution of the generated frontal image was closer to the real face image.In the semi-global network,a multi-temporal and depth attention module was designed in the network to facilitate the network to learn more facial saliency features.The proposed method was applied to the CAS-PEAL-R1 dataset and the self-built dataset,and the Rank-1 indicator was used for evaluation.Results of experiments show that the average Rank-1 accuracy of proposed method in all angles is 99.40%,which verifies that the semi-global network with multi-temporal depth attention module can effectively retain the salient features of faces and improve the accuracy of face matching.

关 键 词：多姿态面部 面部正面化 生成对抗网络 三通道网络 半全局网络 注意力模块 显著特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]