基于生成对抗网络的人脸属性合成技术综述  

Review of face attribute synthesis techniques based ongenerative adversarial network

作　　者：王健强 张珂[1,2] 李培杰 Wang Jianqiang;Zhang Ke;Li Peijie(Dept.of Electronic&Communication Engineering,North China Electric Power University,Baoding Hebei 071003,China;Hebei Key Laboratory of Power Internet of Things Technology,North China Electric Power University,Baoding Hebei 071003,China)

机构地区：[1]华北电力大学电子与通信工程系,河北保定071003 [2]华北电力大学河北省电力物联网技术重点实验室,河北保定071003

出　　处：《计算机应用研究》2025年第3期650-662,共13页Application Research of Computers

基　　金：国家自然科学基金资助项目(62076093,62206095,61871182);中央高校基本科研业务费专项资金资助项目(2023JG002,2022MS078,2023JC006);河北省省级科技计划资助项目(SZX2020034)。

摘　　要：人脸属性合成技术旨在保留人脸面部图像身份信息的情况下,根据指定目标重建人脸属性,从而在源图像上合成具有全新属性的人脸。计算机视觉技术的发展为人脸属性合成技术提供了全新的解决方案。为此,从人脸属性合成数据集、传统和生成对抗网络(GAN)的合成网络以及人脸语义方面综述了人脸属性合成技术的发展。首先分析了人脸属性合成领域中传统方法和主流的深度学习方法,探讨基于GAN方法的发展现状,将基于GAN的人脸属性合成模型划分为有监督、无监督以及半监督三种,将人脸属性划分年龄、表情和妆容三大类语义,并对多种合成模型进行深入研究。其次,对典型的损失函数进行分析和总结,同时介绍了常用人脸属性数据集以及评价指标。最后介绍现有人脸属性合成方法面临的问题,并对该领域未来的发展提出展望。Face attribute synthesis technology aims to reconstruct face attributes according to the specified target while retaining the identity information of face images.The development of computer vision technology has provided a new solution for face attribute synthesis technology.To this end,this paper focused on the face attribute synthesis dataset,traditional and GAN synthesis networks,and face semantics.It reviewed the development of facial attribute synthesis technology.First,this paper analyzed the traditional methods and mainstream deep learning methods in the field of FAS,explored the development status of GAN-based methods,divided GAN-based facial attribute synthesis models into supervised,unsupervised and semi-supervised,and divided facial attributes into three semantic categories of age,expression and makeup,and conducted in-depth research on multiple synthesis models.Secondly,this paper analyzed and summarized typical loss functions.At the same time,this paper introduced the commonly used facial attribute datasets and evaluation indicators.Finally,this paper introduced the problems with existing face attribute synthesis methods and proposed prospects for the future development of this field.

关 键 词：深度学习 生成对抗网络 人脸属性生成 人脸图像数据集 年龄 表情 妆容 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]