基于DCNN的人脸识别技术在考生身份验证中的应用研究  

The Research of Face Recognition Technology Based on DCNN in Candidate Authentication

作　　者：景晨凯 宋涛[1,2] 庄雷 刘刚[2] 王乐[2] 李兵奎[1] 

机构地区：[1]郑州大学信息工程学院,郑州450001 [2]河南省招生办公室,郑州450046

出　　处：《河南大学学报（自然科学版）》2017年第6期699-707,共9页Journal of Henan University:Natural Science

基　　金：国家自然科学基金项目(61379079);河南省国际合作项目(152102410021)

摘　　要：考试的公平、公正、安全和秩序是全社会关注的焦点,尤其是备受关注的高招考试.近些年来DCNN(deep convolution neural networks)算法促进了人脸识别技术的实际应用,因此若在考生身份验证中采用人脸识别技术将进一步保证考试公平,降低人工成本.但是针对具体的应用,DCNN算法需要做相应的改变.依托真实的考生数据集以及应用场景,基于GoogLeNet设计了一种更具表达能力更适用的网络结构GoogLeNet-D;因将人脸查询/分类精准率作为模型评估的方法,所以没有判定阈值,为了设定合适的阈值判断考生是否为同一个人,提出了一种直接、简单有效的定量确定阈值算法,能够在计算准确率的同时确定阈值.最终利用该阈值判定算法,在2014-2016年170万考生共10 406 024张人脸数据集上选取出基于GoogLeNet-D训练的最优模型,其在20万人1 022 031张人脸的测试集上取得了98.87%的人脸分类精准率,同时得到了该模型的最佳阈值为0.35.Fairness,justice,security and order are the focus of the whole society,especially in the national college entrance examination which is extremely concerned with.In recent years,DCNN（deep convolution neural networks）algorithm has improved the application of face recognition technology.Using face recognition technology,the examine identity authentication will further ensure the fairness of the exam and reduce labor costs.However,when it comes to the specific application,the DCNN model needs to make corresponding changes.This paper designs a more expressive and applicable network structure based on the design idea of GoogLeNet,which is based on the analysis of real candidates＇data sets and application scenarios.This paper uses the face query/classification precision as a method of model evaluation,so there is no threshold.In order to evaluate the GoogLeNet-D model and set a appropriate threshold to determine whether the candidate is the same person,this paper proposes a direct,simple and effective algorithm to quantify the threshold,and can determine the threshold when calculating the precision.Finally,using the threshold decision algorithm,we selected the optimal model based on GoogLeNet-D training in 10 406 024 face data sets of 1.7 million candidates in 2014-2016.The precision of the face classification was 98.87%in 200 thousand people 1 022 031 face test set,and the optimal threshold of the model was 0.35.

关 键 词：深度卷积神经网络 人脸识别 身份验证 GoogLeNet-D 阈值判定 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]