数据与计算发展前沿 ›› 2023, Vol. 5 ›› Issue (2): 2-23.

CSTR: 32002.14.jfdc.CN10-1649/TP.2023.02.001

doi: 10.11871/jfdc.issn.2096-742X.2023.02.001

• 专刊:“人工智能&大数据”科研范式变革专刊(上) • 上一篇    下一篇

基于生成对抗网络的人脸年龄编辑方法综述

刘云帆1,2(),李琦2,孙哲南1,2,*(),谭铁牛1,2   

  1. 1.中国科学院大学,人工智能学院,北京 100049
    2.中国科学院自动化研究所,智能感知与计算研究中心,北京 100190
  • 收稿日期:2023-01-28 出版日期:2023-04-20 发布日期:2023-04-24
  • 通讯作者: 孙哲南
  • 作者简介:刘云帆,中国科学院大学人工智能学院,博士研究生,主要研究方向为生物特征识别、计算机视觉、人脸图像编辑。
    本文中负责文献整理和论文主体撰写。
    LIU Yunfan is currently a graduate student in School of Artificial Intelligence, Chinese Academy of Sciences, and he is now pursuing his Ph.D. degree. His research interests include biometrics, computer vision, and face image editing.
    His contributions to this paper are literature survey and man-uscript writing.
    E-mail:yunfan.liu@cripac.ia.ac.cn|孙哲南,中国科学院自动化研究所,研究员,博士生导师,中国科学院大学人工智能学院岗位教授,天津中科智能识别产业技术研究院院长,国际模式识别学会会士IAPR Fellow和生物特征识别技术委员会主席,担任国际期刊IEEE Transactions on Biometrics, Behaviour, and Identity Science编委。主要研究方向为生物特征识别、模式识别、计算机视觉。
    本文中负责写作指导以及论文最终审定。
    SUN Zhenan is currently a professor with the Institute of Automation, Chinese Academy of Sciences. He serves as an Associate Editor of the IEEE Transactions on Biometrics, Behaviour, and Identity Science. He is a fellow of the IAPR. His research interests include biometrics, pattern recognition, and computer vision.
    His contributions to this paper are writing instruction and manuscript reviewing.
    E-mail: znsun@nlpr.ia.ac.cn
  • 基金资助:
    国家自然科学基金面上项目(62076240);北京市自然科学基金会(4222054)

Face Age Editing Methods Based on Generative Adversarial Network: A Survey

LIU Yunfan1,2(),LI Qi2,SUN Zhenan1,2,*(),TAN Tieniu1,2   

  1. 1. School of Artificial Intellignence, University of Chinese Academy of Sciences, Beijing 100049, China
    2. Center for Research on Intelligent Perception and Computing, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
  • Received:2023-01-28 Online:2023-04-20 Published:2023-04-24
  • Contact: SUN Zhenan

摘要:

【目的】对基于生成对抗网络的人脸年龄编辑方法进行系统梳理和全面总结。【文献范围】本文从该领域发表在国内外一流会议期刊上的文章出发,按照引用关系对具有代表性的方法进行了检索,并对最终得到的众多研究工作进行了综述。【方法】本文从生成对抗网络内在机制的角度出发,按照隐向量求取和操控方式的不同,对现有方法进行了分类,并对各个模型之间的共性和差异进行了分析和比较。【结果】随着生成对抗网络的快速发展,人脸年龄编辑方法的各方面性能稳步提升,但是仍然面临着诸多问题和挑战。【局限】由于尚不存在一套通用的评价标准,因此本文暂时无法将所有涉及到的方法用一组统一的定量指标进行衡量和比较。【结论】虽然已经取得了显著的成果,但是现有方法在编辑过程的灵活性、极端情况的适应性以及输出结果的多样性方面还有亟待解决的问题。

关键词: 深度学习, 生成对抗网络, 图像转换, 人脸属性, 人脸年龄编辑

Abstract:

[Objective] This paper provides a comprehensive and systematic review of the face age editing methods based on generative adversarial network. [Coverage] This survey contains numerous studies including publications in top-tier international journals and conferences, as well as other representative works in this field. [Methods] From the perspective of the internal mechanism of generative adversarial networks, these studies are categorized according to how the latent vector is obtained and manipulated, and a brief analysis and comparison of the similarity and difference of specific models is provided. [Results] With the rapid development of generative adversarial networks, the performance of the face age editing method in all aspects has been continuously and steadily improved, while there are still challenges and problems in various aspects. [Limitations] Due to the lack of a developed and mature evaluation standard, it is temporarily impossible for this paper to objectively measure and compare the discussed methods with a set of unified quantitative metrics. [Conclusions] Although significant outcome has been achieved, existing methods still face unsolved problems, including the flexibility of the editing process, the adaptability to extreme editing cases, and the diversity of generation results.

Key words: deep learning, generative adversarial network, image translation, facial attribute, face age editing