摘要
当前大多数方法需要对人脸进行对齐等预处理,这不仅影响验证流程的连续性,还严重影响人脸验证的效率。本文设计了两种神经网络模型及三个阶段式的训练验证构架以及基于深度特征与SIFT特征相结合的高效的非对齐人脸验证方法:方法利用卷积神经网络的池化层中间结果同步生成SIFT特征描述符从粗粒度到细粒度进行多级联的非对齐的人脸验证,这极大的提高人脸验证的速度及准确度;在训练阶段提出了使用三元组样本作为输入,Triplet loss作为损失函数有效提高不同人之间的区分度提高人脸验证的准确率;本文根据不同应用场景设计了两种深度学习架构适应小型及大型设备的需要。本方法经过在Web-face数据集训练及在LFW,YOUTUBE等数据集上验证,结果表明该方法具有良好的性能。
At present,most methods need to process face alignment in advance,which not only verifies the continuity of the process,but also seriously affects the efficiency of face verification.In order to solve the above problems,this paper proposes a framework of unaligned face verification based on deep learning.In the process of face verification,SIFT descriptors of face feature is generated synchronously by making full use of the pooling results of deep learning,and these SIFT descriptors are used to cascade verify the unaligned face from coarse granularity to fine granularity,which will greatly improve the speed and accuracy of face verification.In the training stage,utilizing three-tuple samples as input,Triplet loss as loss function to improve the accuracy of face verification.In this paper,according to different application scenarios,two kinds of CNN are designed to meet the needs of smartphone and large devices.After a long time training on WEB-face dataset and experimenting on the LFW,YOUTUBE datasets,the results show that the proposed method has the superior performance compared with some state-of-the-art methods.
作者
朱梅
陈易萍
姚毅
樊中奎
李渤
ZHU Mei;CHEN Yiping;YAO Yi;FAN Zhongkui;LI Bo(School of Software Engineering,JiangXi University of Science and Technology,Nanchang 330013,China;Applied Science College,Jiangxi University of Science and Technology,Jiangxi Ganzhou 341000,China;Nanchang Academy of Agricultural Sciences,Nanchang 330200,China)
出处
《南昌大学学报（理科版）》
CAS
北大核心
2020年第5期496-503,共8页
Journal of Nanchang University(Natural Science)
基金
江西现代农业科研协同创新专项项目(JXXTCXQN201906)
江西理工大学科研基金(204201900010)。
