几何特征和属性标签驱动的人脸图像合成  

Face Image Synthesis Driven by Geometric Feature and Attribute Label

作　　者：代福芸 迟静[1,2] 任明国 张琪东 DAI Fu-yun;CHI Jing;REN Ming-guo;ZHANG Qi-dong(School of Computer Science and Technology,Shandong University of Finance and Economics,Jinan 250014,China;Shandong Provincial Key Laboratory of Digital Media Technology,Jinan 250014,China)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东省数字媒体技术重点实验室,济南250014

出　　处：《计算机科学》2022年第10期214-223,共10页Computer Science

基　　金：山东省重点研发计划(2019GSF109112);山东省自然科学基金(ZR2019MF016);山东省高等学校青创科技支持计划(2020KJN007);济南市“新高校20条”科研带头人工作室(2021GXRC092)。

摘　　要：针对当前人脸图像合成中存在的模样和表情合成的多样性不足、表情的真实感不高、合成效率低的问题,提出了一种引入面部几何特征和属性标签约束的新型人脸合成网络模型。该模型在给定一张源人脸图片、一张目标人脸图片和属性(如发色、性别、年龄等)标签的情况下,能够生成一张具有源人脸表情、目标人脸身份特征以及指定属性的高真实感的人脸图像。模型包括两部分:人脸特征点生成器(Facial Landmark Generator, FLMG)和几何-属性感知生成器(Geometry and Attribute Aware Generator, GAAG)。FLMG利用人脸几何特征点编码表情信息,并将源人脸的表情信息迁移到目标人脸的特征点上;GAAG结合FLMG中生成的特征点、给定的属性标签和目标人脸图片,生成一张具有指定模样和表情的人脸图片。文中还引入了一个新的软截断三元感知损失函数用于约束GAAG,该函数可使生成的人脸更好地保持目标人脸的身份特征且更加真实自然,同时可使GAAG模型以更快的速度收敛。实验结果表明,所提方法可合成模样和表情多样化的人脸图像,且人脸的外貌真实、表情自然。另外,所提网络模型只需训练一次即可实现任意不同表情之间的迁移,效率较高。Aiming at the problems in current face image synthesis, such as the lack of diversity of synthetic appearances and expressions, the low reality of the facial expressions and the low synthesis efficiency, this paper proposes a novel face synthesis network model driven by facial geometric feature and attribute label.Given a source face image, a target face image and the attribute(e.g.,hair color, gender, age) label, the new face synthesis model can generate a highly realistic face image which owns the expression of the source face, the identity of the target face and the specified attribute.The new model consists of two parts: facial landmark generator(FLMG) and geometry and attribute aware generator(GAAG).FLMG uses the facial geometric feature points to encode the expression information, and transfers the expression from the source to the target face in the form of feature points.Combining the transferred feature points, the specified attribute label and the target face image, GAAG generates a face image with specified appearance and expression.A novel soft margin triplet perception loss is introduced to GAAG,which can make the synthesized face more natural and keep the identity of the target face well, and makes the GAAG converge faster.Experimental results show that the face images generated by our approach have more diverse appearances and more realistic expressions.In addition, our model only needs to be trained once to realize the transfer between any arbitrary different expressions, so its efficiency is high.

关 键 词：人脸图像合成 表情迁移 人脸编辑 生成式对抗网络 

分 类 号：TP317[自动化与计算机技术—计算机软件与理论]