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 (5): 267-273,281. doi: 10.19678/j.issn.1000-3428.0054134

• 图形图像处理 • 上一篇    下一篇

标签引导的生成对抗网络人脸表情识别域适应方法

孙冬梅, 张飞飞, 毛启容   

  1. 江苏大学 计算机科学与通信工程学院, 江苏 镇江 212013
  • 收稿日期:2019-03-07 修回日期:2019-04-19 发布日期:2020-05-08
  • 作者简介:孙冬梅(1993-),女,硕士研究生,主研方向为计算机视觉、情感计算;张飞飞,博士研究生;毛启容,教授、博士、博士生导师。
  • 基金资助:
    国家自然科学基金(61672267,61672268);江苏省研究生科研创新项目(KYCX17_1811)。

Label-Guided Domain Adaptation Method in Generative Adversarial Network for Facial Expression Recognition

SUN Dongmei, ZHANG Feifei, MAO Qirong   

  1. School of Computer Science and Communication Engineering, Jiangsu University, Zhenjiang, Jiangsu 212013, China
  • Received:2019-03-07 Revised:2019-04-19 Published:2020-05-08

摘要: 传统的人脸表情识别方法主要针对实验室环境下的基本表情,难以应对现实场景中人类微妙和复杂的表情变化,并且目前自然环境人脸表情识别数据集普遍缺乏足够的训练数据。针对该问题,利用实验室环境下的数据库样本,提出以标签引导的生成对抗网络表情识别域适应方法。将情感标签作为辅助条件,训练生成对抗网络的生成模型,把实验室环境的数据库样本转化为类似自然环境数据库的样本,以扩充自然环境数据库,同时基于扩充的数据库样本训练基本分类器VGG、Resnet等,从而学习自然环境的数据库的情感特征。在RAF_DB等自然环境人脸表情数据库上的实验结果表明,与Boosting-POOF和PixelDA方法相比,该方法扩充得到的数据库可使人脸表情识别率取得6%~9%的提升。

关键词: 生成对抗网络, 情感标签, 人脸表情识别, 域适应, 自然环境, 数据库样本

Abstract: In Facial Expression Recognition(FER),sufficient training samples have significant influence on recognition results.To address insufficient database samples in natural environment,this paper proposes a Label-guided Domain Adaption method in Generative Adversarial Network(LDAGAN) for FER by using laboratory environment database samples.This method adopts the generation model of GAN and takes emotional labels as auxiliary condition.Then the method uses the laboratory environment database samples to generate samples similar to those of natural environment database,so as to build a bridge between laboratory environment database and natural environment database,enlarging the natural environment database.The samples assist in learning the emotional features of the natural environment database.Experimental results on the facial expression database of natural environment such as RAF_DB show that the proposed method achieves an improvement of 6% to 9% in FER accuracy compared with Boosting-POOF and PixelDA methods.

Key words: Generative Adversarial Network(GAN), emotional label, Facial Expression Recognition(FER), domain adaptation, natural environment, database sample

中图分类号: