基于注意力机制与特征相关性的人脸表情识别  

Facial expression recognition based on attention mechanism and feature correlation

作　　者：兰凌强 刘淇缘 卢树华[1] LAN Lingqiang;LIU Qiyuan;LU Shuhua(College of Information and Cyber Security,People's Public Security University of China,Beijing 102600,China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京102600

出　　处：《北京航空航天大学学报》2022年第1期147-155,共9页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家重点研发计划(2016YFC0801005);中央高校基本科研业务经费项目(2019JKF225);公共安全行为科学实验室开放课题(2020SYS16)。

摘　　要：针对自然条件下人脸表情识别面临遮挡、光照、姿势变化等挑战,存在识别准确率低的问题,提出了一种新的深度学习网络模型用于人脸表情识别。以ResNet为基础网络,融合了瓶颈注意力机制及全局二阶池化层,其中瓶颈注意力机制专注于表情重要特征的提取,全局二阶池化层度量表情特征之间的相关性,在此基础上通过联合正则化策略,平衡和改善特征数据分布情况,提高表情识别准确率。所提方法在2个公开数据集FER2013和CK+进行了测试及验证,最高准确率分别达到了74.227%和95.8%,性能优于诸多现存的主流方法,表明所提模型具有较好的准确性和鲁棒性。There are many challenges including occlusion, illumination and posture variation in the facial expression recognition under natural conditions, leading to the low accuracy. This paper proposes a new deep learning network model for facial expression recognition. This model uses ResNet as backbone, and introduces the bottleneck attention module and the global second-order pooling layer. The bottleneck attention module focuses on the extraction of important expression features, and the global second-order pooling layer aims to measure the correlation among expression features. On this basis, the joint normalization strategies are used balance and improve the distribution of feature data, which improves the accuracy of expression recognition. The test and validation of the proposed method have been carried out on the two public datasets FER2013 and CK+, resulting in the highest accuracy rates of 74.227% and 95.8%, respectively. The performance is better than most of the latest facial expression recognition methods. The results indicate that this model has better accuracy and robustness.

关 键 词：表情识别 深度学习 瓶颈注意力机制 全局二阶池化层 联合正则化策略 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]