计算机与现代化

• 图像处理 • 上一篇    下一篇

基于DCNN的证件照人脸验证及应用研究

  

  1. (1.临沂大学信息科学与工程学院,山东临沂276000;2.临沂大学教务处,山东临沂276000)
  • 收稿日期:2019-06-25 出版日期:2020-03-03 发布日期:2020-03-03
  • 作者简介:李硕(1998-),男,山东郓城人,本科生,研究方向:深度学习,图形图像处理,E-mail: 2237885798@qq.com; 卞青山(1979-),男,硕士,研究方向:公共事业管理,E-mail: bianqingshan@lyu.edu.cn; 刘传文(1979-),男,本科,研究方向:数据库,图像图形处理,E-mail: liuchuanwen@lyu.edu.cn; 刘鸣涛(1982-),男,讲师,博士,研究方向:深度学习,人脸识别,图像处理,E-mail: liumingtao@lyu.edu.cn; 通信作者:张林涛(1985-),男,讲师,博士,研究方向:深度学习,人脸识别,E-mail: zhanglintao@lyu.edu.cn。
  • 基金资助:
    临沂大学博士启动研究基金资助项目(LYDX2016BS115,LYDX2016BS114) 

Face Verification and Application Research of ID Photos Based on DCNN

  1. (1. School of Information Science and Engineering, Linyi University, Linyi 276000, China;  
    2. Academic Affair Office, Linyi University, Linyi 276000, China)
  • Received:2019-06-25 Online:2020-03-03 Published:2020-03-03

摘要: 在不同证件审核场景中,由于存在年龄跨度、装扮及样本缺乏等因素的影响,使得现有方法难以适应不同证件照下的人脸识别,无法满足实际应用要求。为解决上述问题,提出一种基于深度卷积神经网络的不同证件照识别方法。该方法对VGG网络做出适应于不同证件照识别的改进,实现端到端的自主学习人脸特征,消除年龄跨度、装扮等因素的影响,并且可将训练参数减少为原网络结构的〖SX(〗1〖〗6〖SX)〗,使得在保证识别精度的同时,模型训练时间大幅减小。实验结果表明,该方法在高校毕业审核场景下的自建数据集和CAS-PEAL-R1公开数据集上训练后,验证准确率和召回率较原始方法分别提高了6.29个百分点和7个百分点,能够满足多种应用场景下的不同证件审核需求。

关键词: 人脸识别, 证件审核, 卷积神经网络, 人脸验证

Abstract: In different authentication scenarios, it is difficult to adapt the existing methods to face recognition under different authentication photos for the sake of the influence of age span, dress-up and lack of samples, which cannot conform to the practical application requirements. For the sake of solving the above problems, it puts forward a different identification method on the basis of the deep convolution neural network. This method makes the improvement of VGG network adapted to different document photo recognition, realizing end-to-end autonomous learning of face features, eliminating the influence of age span, dress-up and other factors. In addition, the method cuts down the trainable parameters to 1/6 of the original network structure, thus ensuring the identification accuracy while greatly reducing the training time of the model. According to the experimental results, after training on the self-built data set and CAS-PEAL-R1 public data set under the college graduation examination scene, the verification accuracy and recall rate of this method were 6.29 and 7 percentage points higher than the original method respectively, which can conform to the different document examination needs under various application scenarios.

Key words: face recognition, ID photo verification, convolutional neural network, face verification

中图分类号: