计算机应用

• 人工智能与仿真 •    下一篇

基于对抗生成网络和网络集成的面部表情识别方法EE-GAN

杨鼎康,黄帅,王顺利,翟鹏,李一丹,张立华   

  1. 复旦大学
  • 收稿日期:2021-05-18 修回日期:2021-06-13 发布日期:2021-11-09 出版日期:2021-11-09
  • 通讯作者: 杨鼎康

EE-GAN:Facial expression recognition method based on generative adversarial network and network integration

  • Received:2021-05-18 Revised:2021-06-13 Online:2021-11-09 Published:2021-11-09

摘要:

摘 要: 现实生活场景差异大,人类在不同场景中表现的情感也不尽相同,导致获取到的情感数据集标签分布不均衡,这一问题成为了面部表情识别中的一项关键挑战。传统方法多采用模型预训练和特征工程来增强与表情相关特征的表示能力,但没有考虑不同特征表达之间的互补性,限制了模型的泛化性和鲁棒性。为此,本文提出了一种包含网络集成模型Ens-Net的端到端深度学习框架EE-GAN,该方法一方面考虑了多个异质网络获得的不同深度和区域的特征,实现不同语义、不同层次的特征融合,并通过网络集成以提高模型的学习能力;另一方面,基于对抗生成网络生成具有特定表情标签的面部图像,在进行数据增强的同时,达到平衡表情标签数据分布的目的。通过在CK +,FER2013和JAFFE数据集上的定性和定量评估表明,本文方法优于现有的基于视图学习的方法和传统的单一结构模型,验证了本文方法的有效性。

Abstract:

Abstract: Because there are many differences and changes between various daily life scenarios, the uneven distribution of emotion labels in the datasets have become a key challenge in facial expression recognition. Most traditional methods utilizemodel pre-training and feature selection mechanisms to enhance the ability to express features related to expressions. Different from the existing methods, we propose an end-to-end deep learning model EE-GAN, which includes the network integration module Ens-Net. On the one hand, our model takes the characteristics of different depths and regions into consideration. We obtain it through multiple heterogeneous networks, then achieve different semantic and multi-level feature fusion, through network integration, which improve the learning ability of the model. On the other hand, we perform data enhancement, using the generative adversarial network to generate facial images with specific expression labels, to achieve the purpose of balancing the distribution of data. In the experiments, we perform the qualitative and quantitative evaluations on the CK+, FER2013 and JAFFE datasets. The results show that this method is superior to the existing view-based learning methods and the traditional single structure models. At the same time, it also verifies the effectiveness of the method we proposed.

中图分类号: