计算机与现代化 ›› 2023, Vol. 0 ›› Issue (11): 89-94.doi: 10.3969/j.issn.1006-2475.2023.11.014

• 图像处理 • 上一篇    下一篇

基于沙漏结构与注意力机制的轻量级人脸表情识别方法

  

  1. (广东工业大学计算机学院,广东 广州 510006)
  • 出版日期:2023-11-29 发布日期:2023-11-29
  • 作者简介:罗明杰(1999—),男,湖南耒阳人,硕士研究生,研究方向:深度学习,人脸表情识别,E-mail: 1873044943@qq.com;通信作者:冯开平(1963—),男,教授,硕士生导师,研究方向:AR/VR,计算机图形学,E-mail: 461090628@qq.com。
  • 基金资助:
    教育部人文社科研究规划基金资助项目(20YJAZH073)

Lightweight Facial Expression Recognition Method Based on Sandglass Structure and Attention Mechanism

  1. (School of Computers, Guangdong University of Technology, Guangzhou 510006, China)
  • Online:2023-11-29 Published:2023-11-29

摘要: 摘要:人脸表情检测分类是人机交互领域的一个挑战性任务。为了解决当前表情识别模型参数量大、分类准确率低等问题,提出一种基于沙漏结构与注意力机制的轻量级人脸表情识别方法。首先利用改进的沙漏结构构建轻量级主干特征提取网络;然后设计一个新颖的特征融合注意模块,融合Focus池化特征以提取关键的细节信息,同时嵌入轻量级ECA注意力机制,强化关键表情特征以提升模型的特征表达能力;最后采取Random Erasing、Dropout等多种训练策略以缓解轻量级网络过拟合现象,从而提升模型的泛化性能。在2个经典表情数据集FER2013和CK+上进行测试实验,识别率分别达到了71.72%、95.96%,同时参数量仅约为1×106。

关键词: 关键词:表情识别, 沙漏结构, 注意力机制, 轻量化

Abstract: Abstract: Facial expression detection and classification is a challenging task in the field of human-computer interaction. In order to solve the problems of large parameters and low classification accuracy in current facial expression recognition models, a lightweight facial expression recognition method based on sandglass structure and attention mechanism is proposed. First, the improved sandglass structure is used to build a lightweight backbone feature extraction network. Then a novel feature fusion attention module is designed. Focus pooled features are fused to extract key details, and lightweight ECA attention mechanism is embedded to strengthen key expression features to improve the feature expression ability of the model. Finally, various training strategies such as Random Erasing and Dropout are adopted to alleviate the over fitting phenomenon of lightweight networks, so as to improve the generalization performance of the model. Testing experiments were conducted on two classical expression datasets FER2013 and CK+, and the recognition rates reached 71.72% and 95.96% respectively, while the number of parameters is only about 1×106.

Key words: Key words: expression recognition, sandglass, attention mechanism, lightweight

中图分类号: