新型融合注意力机制的遮挡面部表情识别框架  

A novel framework for occluded facial expression recognition by integrating attention mechanism

作　　者：张本文 高瑞玮 乔少杰 ZHANG Benwen;GAO Ruiwei;QIAO Shaojie(School of Science and Engineering,Sichuan Minzu College,Kangding 626001,China;School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]四川民族学院理工学院,四川康定626001 [2]成都信息工程大学软件工程学院,成都610225

出　　处：《重庆理工大学学报（自然科学）》2023年第9期217-226,共10页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金项目(62272066,61962006);四川省科技计划资助项目(2021JDJQ0021,2022YFG0186,2022NSFSC0511,2023YFG0027);教育部人文社会科学研究规划基金项目(22YJAZH088);成都市“揭榜挂帅”科技项目(2022-JB00-00002-GX,2021-JB00-00025-GX);中国电子科技集团公司第五十四研究所高校合作课题(SKX212010057);成都信息工程大学国家智能社会治理实验基地开放课题(ZNZL2023B05);成都信息工程大学科技创新能力提升计划项目(KYTD202222);宜宾市引进高层次人才项目(2022YG02);区块链数据管理教育部工程研究中心开放基金。

摘　　要：现有基于深度学习的面部表情识别模型不能有效地应对面部遮挡部分的干扰,无法准确捕捉面部未遮挡部分的特征,会导致识别准确率降低。为此,提出一种新型融合注意力机制的遮挡面部表情识别框架FER-AM(facial expression recognition framework based on attention mechanism),应用局部特征网络提取面部表情的局部关键特征,设计全局特征网络学习整个面部表情中的互补信息,采用注意力机制处理面部遮挡部分如眼镜、口罩和围巾等。在RAF-DB、AffectNet、CK+(Cohn Kanade)及FED-RO数据集进行大量实验,结果表明:FER-AM的7种表情分类性能均优于基于深度学习的代表性人脸面部表情识别模型,识别准确率达到88.1%。Traditional facial expression recognition technologies rely heavily on manually formulated feature extraction rules,while deep learning-based facial expression recognition technologies can automatically perform the operations of feature extraction,feature selection and feature classification.However,for the faces with occluded parts,the existing facial expression recognition models based on deep learning cannot effectively deal with the interference of the occluded part of the face,and cannot accurately capture the features of facial unobstructed parts,thus leading to the degradation of recognition accuracy.To solve the aforementioned problems,a novel occluded facial expression recognition framework by integrating attention mechanism called FER-AM(facial expression recognition framework based on attention mechanism)is proposed,the local feature network is used to extract the local key features of facial expressions,and the global feature network is designed to learn the complementary information in the whole face,and the attention mechanism can effectively deal with facial occluded parts,such as glasses,masks and scarves.A large number of experiments are conducted on RAF-DB,AffectNet,CK+(Cohn Kanade)and FED-RO data sets,and the results show that the seven expression classification performance of FER-AM is better than the representative facial expression recognition models based on deep learning,and the recognition accuracy can reach 88.1%.

关 键 词：遮挡面部表情识别 特征提取 特征分类 深度学习 注意力机制 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]