孪生网络模型在多人种人脸认证中的性能研究  

Research of Siamese Network Based Model for Face Verification Under Multi Races

作　　者：赵淑欢 葛佳琦 刘文 刘帅奇[1,2] ZHAO Shu-huan;GE Jia-qi;LIU Wen;LIU Shuai-qi(College of Electronic and Information Engineering,Hebei University,Baoding 071002,China;Machine Vision Technology Innovation Center of Hebei Province,Baoding 071000,China)

机构地区：[1]河北大学电子信息工程学院,河北保定071002 [2]河北省机器视觉技术创新中心,河北保定071000

出　　处：《小型微型计算机系统》2022年第2期362-366,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61572063,61401308)资助;河北省自然科学基金项目(F2019201151,F2018210148,F2020201025)资助;河北省高等学校科学技术研究项目(QN2016085,QN2017306,BJ2020030)资助;河北大学校长基金项目(XZJJ201909)资助;河北大学高层次人才科研启动经费项目(2014-303,8012605)资助;河北大学高性能计算平台支持;河北省创新技术中心开放课题项目(2018HBMV01,2018HBMV02)资助;广东省数字信号与图象处理技术重点实验室开放基金项目(2020GDDSIPL-04)资助。

摘　　要：深度学习网络在模式识别领域性能优异,但需要大量有标注样本对网络进行训练,而对于人脸认证情况下训练样本有限,且已有网络模型在不同人种间的性能差异大,往往会导致人脸认证失效.针对以上问题,本文首先在几种预训练深度网模型上构造孪生网络,并设计相似度度量网络;其次,选用多人种的人脸数据库(Racial Faces in-the-Wild,RFW)中不同人种构造正负样本对作为训练集,扩展数据分布,提高模型泛化能力,且在训练过程中采用循环训策略练提高模型的稳定性;最后,在测试集上采用ROC曲线及AUC值对不同模型性能进行评估,测试集与训练集样本无身份交叉,根据实验结果得出结论:VGG16作为主干网络的孪生网络模型更适合多人种人脸认证问题,PCANet提取的特征对African地区的人脸认证更有效.Deep learning neural networks achieve significant performance in pattern recognition,but training a deep model need huge labeled samples.In face verification,the number of labeled training sample is limited and the performance of existing models in different races changes greatly.To address the above problems,we construct Siamese networks based on pre-trained deep models with reasonable designed similarity network.Then,the training sample pairs are sampled from RFW dataset with various races to expend data distribution and enhance the generalization ability of the proposed model.In the training stage,we adopt cycle strategy to increase the robustness.Finally,the evaluation of the proposed model is implemented on the RFW without overlapped identification with training set by ROC curves and AUC values.The conclusions are as follows:VGG16 based model is more reasonable for multi-race face verification than Inception,DenseNet and ResNet.Besides,the features extracted by PCANet is more robust for African face images.

关 键 词：多人种人脸认证 孪生网络 预训练网络 VGG16 PCANet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]