基于双注意力模型和迁移学习的Apex帧微表情识别

PDF下载HTML

Title:

Apex frame microexpression recognition based on dual attention model and transfer learning

作者:

徐玮¹, 郑豪², 杨种学²

1. 广西师范大学 计算机科学与工程学院/软件学院，广西 桂林 541004;
2. 南京晓庄学院 信息工程学院，江苏 南京 211171

Author(s):

XU Wei¹, ZHENG Hao², YANG Zhongxue²

1. School of Computer Science and Engineering/School of Software, Guangxi Normal University, Guilin 541004, China;
2. School of Information and Engineering, Nanjing Xiaozhuang University, Nanjing 211171, China

关键词:

微表情识别; 深度学习; Apex帧; 双注意力模型; ResNet18网络; Focal Loss函数; 宏表情; 迁移学习

Keywords:

microexpression recognition; deep learning; Apex frame; dual attention model; ResNet18 network; Focal Loss function; macroexpression; transfer learning

分类号:

TP391

DOI:

10.11992/tis.202010031

摘要:

微表情具有持续时间短、强度低的特点，其识别准确率普遍不高。针对该问题提出了一种改进的深度学习识别方法，该方法取微表情视频序列中的Apex帧，采用集成空间、通道双注意力模块的ResNet18网络，引入Focal Loss函数解决微表情数据样本不平衡的问题，并将宏表情识别领域的先验知识迁移到微表情识别领域，以提高识别效果。在CASME II微表情数据集上使用“留一交叉验证法”进行实验，结果表明本文方法相比一些现有的方法识别准确率及 ${F_1}$ 值更高。

Abstract:

Microexpression is of short duration and low intensity, and its recognition accuracy is generally not high. To address this problem, an improved deep learning recognition method is proposed. This method takes Apex frames in the microexpression video sequence and adopts the ResNet18 network integrating spatial and channel dual attention modules. Moreover, the method introduces the Focal Loss function to solve the imbalance of microexpression data samples and transfers the prior knowledge in the field of macroexpression recognition to the field of microexpression recognition to improve the recognition effect. Experiments were performed on the CASME II microexpression dataset using the “leave one out-cross validation” method. The results show that the method presented in this paper has a higher recognition accuracy and F₁ value than other existing methods.