基于DCNN的人脸识别技术在身份验证中的研究

发布时间：2018-01-13 10:22

本文关键词：基于DCNN的人脸识别技术在身份验证中的研究　出处：《郑州大学》2017年硕士论文　论文类型：学位论文

【摘要】：为了防范和查处招生考试中的替考舞弊行为,进一步保证考试的公平、公正,基于对河南省招生办公室现有的传统身份验证系统的分析,本文提出了基于DCNN的人脸识别技术在考生身份验证中的解决方案,并重点研究了DCNN模型的优化设计和阈值判定两个关键问题。依托真实的高招考试考生数据集,通过对考生数据集的规模和特点以及实际应用的分析,本文根据GoogLe Net的设计思想以及相关的DCNN优化方法,设计了一种更具表达能力且适用的变形网络结构GoogLeNet-D(Deformation),该模型共有34层约8.9M可训练参数,最终输出128维向量作为考生的人脸特征向量。为了评估GoogLe Net-D模型并设定合理的阈值判断考生是否为同一个人,本文鉴于考生数据集中正样本数量远远大于负样本数量的情况,在评估模型时直接采用真实的正样本对,将人脸查找/分类的精准率作为模型评估的依据。但采用该评估方法无法对阈值进行判定,因此本文进一步提出了一种直接、简单有效的定量确定阈值的算法,能够在计算精准率的同时确定阈值。实验表明,GoogLeNet-D具有良好的人脸特征表达能力,其使用2014年-2016年170万考生共10,406,024张人脸数据做训练,在20万考生1,022,031张人脸数据的测试集上取得了98.87%的人脸分类精准率,并得出该模型的最佳阈值0.35,目前该模型已成功应用于河南省高招考试及其他招生考试中。
[Abstract]:In order to prevent and investigate the fraud in the recruitment examination, and to ensure the fairness and fairness of the examination, based on the analysis of the existing traditional identity verification system of Henan enrollment Office. This paper proposes a solution of face recognition technology based on DCNN in examinee authentication. This paper focuses on the optimization design of DCNN model and the threshold decision. Relying on the real test data set, the paper analyzes the scale and characteristics of the test data set and its practical application. In this paper, according to the design idea of GoogLe Net and the related DCNN optimization method. In this paper, a more expressive and applicable textured network structure, Google LeNet-Dwe design, is designed. The model has about 8.9m trainable parameters in 34 layers. In order to evaluate the GoogLe Net-D model and set a reasonable threshold to judge whether the examinee is the same person. In view of the fact that the number of positive samples in the examinee data set is much larger than the number of negative samples, the real positive sample pairs are directly used in the evaluation of the model. The accuracy rate of face search / classification is taken as the basis of model evaluation, but the threshold can not be determined by this evaluation method. Therefore, this paper further proposes a direct, simple and effective algorithm for quantitative determination of threshold. Experiments show that Google LeNet-D has good facial feature expression ability, which uses 10% of 1.7 million candidates from 2014 to 2016. 406,024 face data are trained, and 98.87% face classification accuracy rate is obtained on the test set of 1,022,031 face data for 200,000 examinees. The optimal threshold value of the model is 0.35, which has been successfully applied to the high enrollment examination and other enrollment examinations in Henan Province.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】