基于多任务学习和知识图谱的面部表情识别  

Facial Expression Recognition Based on Multi-Task Learning and Knowledge Graph

作　　者：陈龙 张水平[1,2] 王海晖 陈言璞[1,2] CHEN Long;ZHANG Shuiping;WANG Haihui;CHEN Yanpu(School of Computer Science and Engineering,Wuhan Institute of Technology,Wuhan 430205,China;Hubei Key Laboratory of Intelligent Robot(Wuhan Institute of Technology),Wuhan 430205,China)

机构地区：[1]武汉工程大学计算机科学与工程学院,湖北武汉430205 [2]智能机器人湖北省重点实验室(武汉工程大学),湖北武汉430205

出　　处：《武汉工程大学学报》2021年第6期681-688,共8页Journal of Wuhan Institute of Technology

基　　金：湖北省教育厅科学技术计划青年人才项目(Q20191514);武汉工程大学科学研究基金(16QD25,20QD32);湖北省大学生创新创业训练计划项目(S202110490040)。

摘　　要：针对面部表情分类的模型中参数较复杂、识别准确率较低的问题,提出了一种基于知识图谱辅助识别的多任务学习算法模型(MLAM),该模型由基于深度学习的识别模块与知识图谱嵌入模块两部分构成。首先从输入的数据中提取潜在的人脸局部表情特征,通过知识图谱实现局部表情和个体的复杂交互;然后在MLAM模型中设计一个交叉压缩单元,关联这两个独立模块,自动学习局部表情和实体特征的高级交互,并在这两个任务之间传递交叉知识转移;最后,在FER2013和CK+的数据集上对比了同类算法,实验结果表明,该模型在上述数据集上分别得到了0.69和0.99的识别率,提高了面部表情识别准确率。Aiming at the complicated parameters and low recognition accuracy rate in classification models of facial expression,this paper proposes a multi-task learning algorithm model(MLAM)based on knowledge graph assisted recognition,which consists of recognition module and knowledge graph embedding module based on deep learning.First,the potential facial expression features from the input data were extracted,and the complex interaction between the local expression and the individual through the knowledge graph was realized;then a cross&compress unit was designed in the MLAM,these two independent modules were associated,and the local expression and the advanced interaction of entity features were automatically learned and cross-knowledge between these two tasks was transferred.Finally,similar algorithms were compared on the FER2013 and CK+data sets.The results showed that the model obtains recognition rates of 68.85%and 99.16%respectively on the above data sets,which improves the accuracy of facial expression recognition.

关 键 词：表情识别 多任务学习 知识图谱 交叉压缩单元 

分 类 号：TN911.72[电子电信—通信与信息系统]