作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (4): 228-235. doi: 10.19678/j.issn.1000-3428.0054581

• 图形图像处理 • 上一篇    下一篇

基于条件生成式对抗网络的面部表情迁移模型

陈军波a, 刘蓉a, 刘明b, 冯杨a   

  1. 华中师范大学 a. 物理科学与技术学院;b. 计算机学院, 武汉 430079
  • 收稿日期:2019-04-11 修回日期:2019-06-04 出版日期:2020-04-15 发布日期:2020-04-07
  • 作者简介:陈军波(1994-),男,硕士研究生,主研方向为图像处理、模式识别、智能信息处理;刘蓉,副教授、博士;刘明,教授、博士;冯杨,硕士研究生。
  • 基金资助:
    国家科技支撑计划"现代科技馆体系展品展示关键技术研究及创新平台构建应用示范"(2015BAK33B00)。

Facial Expression Transfer Model Based on Conditional Generative Adversarial Network

CHEN Junboa, LIU Ronga, LIU Mingb, FENG Yanga   

  1. a. College of Physical Science and Technology;b. School of Computer, Central China Normal University, Wuhan 430079, China
  • Received:2019-04-11 Revised:2019-06-04 Online:2020-04-15 Published:2020-04-07

摘要: 面部表情迁移是计算机视觉角色动画领域的关键技术,但现有面部表情迁移方法存在生成表情不自然、缺乏真实感、迁移模型复杂以及训练难度大等问题。为此,构建一种基于条件生成式对抗网络的面部表情迁移模型。通过设计域分类损失函数指定表情域条件,使单个生成器学习多个表情域之间的映射,同时利用模型生成器和判别器之间的条件约束与零和博弈,在仅训练一个生成器的情况下同时实现7种面部表情迁移。实验结果表明,该模型能够有效进行面部表情迁移并且鲁棒性较强,其生成的面部表情较StarGAN模型更自然、逼真。

关键词: 表情迁移, 条件生成式对抗网络, 域分类损失, 重构损失, 零和博弈

Abstract: Facial expression transfer is a key technology for character animation in computer vision,but existing facial expression transfer methods have some problems,such as unnatural expression generation,lack of realism,complex transfer model and difficulty in training.Therefore,a face expression transfer model based on conditional Generative Adversarial Network(GAN) is constructed.The condition of the expression domain is specified by the classification loss function of the design domain,so that a single generator can learn the mapping relations between multiple expression domains.Meanwhile,the conditional constraints and zero-sum game between the model generator and the discriminator are used to realize the transfer of five facial expressions by training only one generator.Experimental results show that,this model can effectively transfer facial expressions and has strong robustness.Facial expressions generated by the proposed model are more natural and realistic than the StarGAN model.

Key words: expression transfer, conditional Generative Adversarial Network(GAN), domain classification loss, reconstruction loss, zero-sum game

中图分类号: