摘要
在视频及非约束条件下获取的动态人脸受到姿态、表情和侧脸等复杂干扰因素,使其识别难度增大。针对上述问题,提出一种基于联合损失和恒等映射的动态人脸识别算法。以Resnet34为基础网络,联合SoftMax Loss,中心损失(Center Loss)和Joint Loss扩大人脸类间距;减小人脸类内距,同时在网络中引入深度残差恒等映射模块,进一步减小侧脸的干扰。在LFW,SLFW,YTF,MegaFace等数据集上,所提算法表现出更强的性能,并且在实验平台上能完成实时视频动态人脸识别的任务。
Dynamic face acquired under video and unconstrained conditions is subject to complex interference factors such as pose,expression and side face,which makes it more difficult to recognize.In order to solve the above problems,a dynamic face recognition algorithm based on Joint Loss and constant mapping is proposed.Resnet 34 is used as basic network.SoftMax Loss,Center Loss and Joint Loss are combined to expand the distance of face inter-classes and reduce the distance of face intra-classes.At the same time,deep residual constant mapping module is introduced into the network to further reduce the interference of side faces.On LFW,SLFW,YTF,MegaFace and other data sets,the algorithm shows stronger performance and can complete the task of real-time video dynamic face recognition on the experimental platform.
作者
刘成攀
吴斌
杨壮
LIU Chengpan;WU Bin;YANG Zhuang(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;Robot Technology Used for Special Environment Key Laboratory of Sichuan Province,Mianyang 621010,China)
出处
《传感器与微系统》
CSCD
北大核心
2021年第9期153-156,共4页
Transducer and Microsystem Technologies
关键词
非约束条件
动态人脸
联合损失
深度残差恒等映射模块
unconstrained condition
dynamic face
joint loss
deep residual constant mapping module