摘要
针对Softmax(柔性最大值)损失对特征只有可分性的不足,提出一种基于深度卷积神经网络的判别性人脸识别算法.该算法首先根据Softmax损失特征分布,在特征和权重向量间施加一个类内余弦相似性损失,使类内更加紧凑,类间尽可能分离;然后在Softmax损失基础上通过归一化特征来更好地模拟低质量人脸图像,并通过归一化权重来减轻类别不平衡,使与测试时的余弦相似性度量一致;最后联合归一化的Softmax损失和类内余弦相似性损失在预训练模型上进行微调.该算法在人脸识别基准测试集LFW(户外人脸标记)和YTF(You Tube人脸数据库)上分别取得了98.72%和93.38%的识别率,实验结果表明:在大规模人脸身份识别中,该算法提高了特征的判别性,增强了模型的泛化能力,能有效提高人脸识别率.
Aiming at the shortcoming of only separability of Softmax(flexible maximum)loss for feature,a discriminative face recognition algorithm based on deep convolutional neural network was proposed.Firstly,according to Softmax loss feature distribution,an intra-class cosine similarity loss was applied between feature and weight vector to make the intra-class more compact and the inter-class separate as much as possible.Then,on the basis of Softmax loss,the normalized features was used to better simulate the low-quality face image,and the normalized weights was used to mitigate the classes imbalance,which was consistent with the cosine similarity measure during the test.Finally,combined normalized Softmax loss and intra-class cosine similarity loss were fine-tuned on the pre-training model.The proposed algorithm achieved 98.72%and 93.38%recognition rates on the face recognition benchmark test set LFW(labeled faces in the wild)and YTF(YouTube face database),respectively.Experimental results show that the proposed algorithm improves the discrimination of features,enhances the generalization ability of the model,and can effectively improve the face recognition rate in large-scale face recognition.
作者
任克强
胡慧
REN Keqiang;HU Hui(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2019年第10期127-132,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61501210)
