融合局部语义与全局信息的人脸表情识别  

Facial expression recognition based on fusion of local semantic and global information

作　　者：潘海鹏[1] 郝慧 苏雯 PAN Haipeng;HAO Hui;SU Wen(School of Mechanical Engineering&Automation,Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China)

机构地区：[1]浙江理工大学机械与自动控制学院,浙江杭州310018

出　　处：《光电子．激光》2022年第6期652-659,共8页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(62006209);浙江省自然科学基金(LQ20F020001);浙江理工大学科研启动基金(1802225-Y);浙江理工大学基本科研业务费专项资金(2020Q014)资助项目。

摘　　要：人脸表情识别在人机交互等人工智能领域发挥着重要作用,当前研究忽略了人脸的语义信息。本文提出了一种融合局部语义与全局信息的人脸表情识别网络,由两个分支组成:局部语义区域提取分支和局部-全局特征融合分支。首先利用人脸解析数据集训练语义分割网络得到人脸语义解析,通过迁移训练的方法得到人脸表情数据集的语义解析。在语义解析中获取对表情识别有意义的区域及其语义特征,并将局部语义特征与全局特征融合,构造语义局部特征。最后,融合语义局部特征与全局特征构成人脸表情的全局语义复合特征,并通过分类器分为7种基础表情之一。本文同时提出了解冻部分层训练策略,该训练策略使语义特征更适用于表情识别,减少语义信息冗余性。在两个公开数据集JAFFE和KDEF上的平均识别准确率分别达到了93.81%和88.78%,表现优于目前的深度学习方法和传统方法。实验结果证明了本文提出的融合局部语义和全局信息的网络能够很好地描述表情信息。Facial expression recognition plays an important role in artificial intelligence such as human-computer interaction.However,current researchers ignore the semantic information of human faces.In this paper,we propose a facial expression recognition network fusing local semantic and global information,which consists of two branches:the local semantic region extraction branch and the local-global feature fusion branch.Firstly,the face semantic parsing is achieved by training semantic segmentation network on face parsing dataset.The semantic parsing of facial expression dataset is obtained by transfer training.Then the meaningful regions and their semantic features are extracted and fused with the global features to obtain the semantic local features.Finally,the global semantic composite features of facial expressions are constructed by combining semantic local features with global features.They are classified into one of the 7 basic facial expressions by the classifier.We also propose a training strategy of unfreezing partial layers,which makes semantic features more suitable for facial expression recognition and reduces the redundancy of semantic information.The average recognition accuracy on two public datasets,JAFFE and KDEF,reaches 93.81%and 88.78%,respectively.The performance outperforms the current deep learning methods and traditional methods.The experimental results demonstrate that the network proposed can describe the expression information comprehensively by integrating local semantic and global information.

关 键 词：人脸表情识别 人脸解析 迁移学习 局部-全局特征融合 解冻部分层训练策略 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]