北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2017, Vol. 40 ›› Issue (1): 84-88,110.doi: 10.13190/j.jbupt.2017.01.015

• 研究报告 • 上一篇    下一篇

基于深度卷积神经网络的跨年龄人脸识别

李亚1, 王广润2, 王青2   

  1. 1. 广州大学 计算机科学与教育软件学院, 广州 510006;
    2. 中山大学 数据科学与计算机学院, 广州 510006
  • 收稿日期:2016-03-01 出版日期:2017-02-28 发布日期:2017-03-14
  • 作者简介:李亚(1980-),女,讲师,E-mail:liya@gzhu.edu.cn;王青(1973-),男,副教授.
  • 基金资助:
    广州市属高校科研项目(1201620302);广东省科技计划项目(2013B010406005,2015B010128009)

A Deep Joint Learning Approach for Age Invariant Face Verification

LI Ya1, WANG Guang-run2, WANG Qing2   

  1. 1. School of Computer Science and Educational Software, Guangzhou University, Guangzhou 510006, China;
    2. School of Data and Computer Science, Sun Yat-sen University, Guangzhou 510006, China
  • Received:2016-03-01 Online:2017-02-28 Published:2017-03-14

摘要: 提出了一种应用于跨年龄人脸识别的联合学习方法,该方法由深度卷积神经网络构建而成,能在特征学习的同时学习到最优的测度函数,从而避免不合适的固定阈值所带来的匹配错误。针对有限的内存、过拟合和计算复杂性高的问题,在模型训练过程中采用了多种新颖和有效的训练策略。实验证实了该联合学习方法的有效性,在公开数据库MORPH-II上的识别正确率达到了93.6%。

关键词: 人脸比对, 人脸识别, 跨年龄, 深度卷积神经网络, 联合学习

Abstract: A joint learning approach (JLA) based on deep convolutional neural network (CNN) for age-invariant face verification was proposed. Feature representation, distance metric and decision function can be learned simultaneously thereafter. Comparing with traditional approaches, it uses fix threshold, so the match errors caused by unfit threshold can be avoided. Some strategies to overcome insufficient memory capacity, prevent over-fitting and reduce computational cost were also introduced. Experiment demonstrates the effectiveness of this approach; the rank-1 recognition accuracy is improved to 93.6% on the MORPH-II database.

Key words: face verification, face recognition, age invariant, deep convolutional neural network, joint learning

中图分类号: