改进的Island损失函数在人脸表情识别上的应用  

Application of Improved Island Loss in Facial Expression Recognition

作　　者：张文萍 贾凯[1,2,4] 王宏玉 徐方[1,2,4] Zhang Wenping;Jia Kai;Wang Hongyu;Xu Fang(State Key Laboratory of Robotics,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169;University of Chinese Academy of Sciences,Beijing 100049;Shenyang SIASUN Robot&Automation Co.,Ltd.,Shenyang 110168)

机构地区：[1]中国科学院沈阳自动化研究所机器人学国家重点实验室,沈阳110016 [2]中国科学院机器人与智能制造创新研究院,沈阳110169 [3]中国科学院大学,北京100049 [4]沈阳新松机器人自动化股份有限公司,沈阳110168

出　　处：《计算机辅助设计与图形学学报》2020年第12期1910-1917,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2017YFB1304100);山东省重大科技创新工程项目(2019JZZY010128).

摘　　要：人脸表情识别在众多场景下都有应用,但是由于光照变化、面部外观改变和遮挡的影响,导致识别准确率下降.针对此现象,提出将改进的Island损失函数用于辅助表情识别.Island损失函数虽然能够提高测试性能,但是却降低了泛化性能,针对这一问题,改进的Island损失函数采用指数损失减小类内距离的方法提高泛化性能,采用余弦距离损失与欧几里得度量损失增大类间距离的方法进一步提高测试性能.结果证明,与Island损失函数相比,LCE损失函数与ECE损失函数均可以同时提高测试性能与泛化性能,ECE损失函数在CK+,JAFFE,SFEW,RAF和Older数据集上的平均泛化性能提高了1.05%.Facial expression recognition is widely used in many scenes,but the recognition accuracy is reduced due to the influence of illumination,facial appearance and occlusion.In view of the phenomenon,an improved Island loss was proposed to assist facial expression recognition.Although Island loss can improve test performance,it reduces generalization performance.To solve this problem,the improved Island loss used exponential loss to reduce intra-class distance to improve the generalization performance,and used cosine distance loss and Euclidean distance loss to increase inter-class distance to further improve the test performance.The experimental results indicate that LCE loss and ECE loss can both improve the test performance and generalization performance at the same time compared with Island loss.The average generalization performance of ECE loss on CK+,JAFFE,SFEW,RAF and Older datasets increases by 1.05%.

关 键 词：人脸表情识别 改进的Island损失函数 指数损失 类内距离 类间距离 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]