DTFBNet:一种面向智能终端的轻量级人脸识别新方法  

DTFBNet:the New Lightweight Face Recognition Method for Smart Terminals

作　　者：叶继华[1] 郭凤 黎欣 江蕗 江爱文[1] YE Jihua;GUO Feng;LI Xin;JIANG Lu;JIANG Aiwen(School of Computer and Information Engineering,Jiangxi Normal University,Nanchang Jiangxi 330022,China)

机构地区：[1]江西师范大学计算机信息工程学院,江西南昌330022

出　　处：《江西师范大学学报（自然科学版）》2022年第2期126-133,共8页Journal of Jiangxi Normal University(Natural Science Edition)

基　　金：国家自然科学基金(62167005,61966018);江西省教育厅重点科研课题(GJJ200302)资助项目.

摘　　要：对于智能终端资源不足的问题,目前有许多解决方法,但普遍存在依赖样本数据和参数量的问题.为此,该文先构造了一个深度卷积和传统卷积融合的模块DTFBlock(depthwise convolution and traditional convolution fusion Block);然后在该基础上提出了一种基于MobileFaceNet的改进方法DTFBNet,DTFBNet参数量较小,在网络的识别效果较好;最后,在面部识别数据集CASIA-Webface和LFW上进行实验,结果表明:该算法的最高准确率达到了99.40%,达到在同等参数量上具有竞争力的分类准确率.There are many solutions to the problem of insufficient intelligent terminal resources,which are dependent on sample data and the number of parameters.The depthwise convolution and traditional convolution fusion block(DTFBlock)is proposed.Hence,an improved method DTFBNet based on MobileFaceNet is proposed.The DTFBNet proposed in the paper has a smaller number of parameters and better network results.Experiments on face recognition datasets CASIA-Webface and LFW show that the highest accuracy rate of the algorithm proposed in this paper reaches 99.40%,which is already a competitive classification accuracy for the same parameter amount.

关 键 词：DTFBNet DTFBlock 融合损失 轻量级 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]