电子学报 ›› 2017, Vol. 45 ›› Issue (3): 619-625.DOI: 10.3969/j.issn.0372-2112.2017.03.017

• 学术论文 • 上一篇    下一篇

基于改进深层网络的人脸识别算法

李倩玉, 蒋建国, 齐美彬   

  1. 合肥工业大学计算机与信息学院, 安徽合肥 230009
  • 收稿日期:2015-06-24 修回日期:2016-01-14 出版日期:2017-03-25
    • 作者简介:
    • 李倩玉 女,1991年生于安徽合肥,合肥工业大学研究生,研究方向为数字图像处理与分析、人脸检测与识别.E-mail:qianyuli@mail.hfut.edu.cn;蒋建国 男,1955年生于安徽黄山,教授、博士生导师,合肥工业大学计算机与信息学院.研究方向为视频信息分析与处理、多媒体智能监控系统、分布式智能系统、DSP技术应用等.E-mail:jgjiang@hfut.edu.cn;齐美彬 男,1969年生于安徽池州,教授、硕士生导师,合肥工业大学计算机与信息学院.研究方向为运动目标检测与跟踪、视频编码等.E-mail:qimeibin@163.com
    • 基金资助:
    • 国家自然科学基金 (No.61371155); 安徽科技攻关项目 (No.1301b042023)

Face Recognition Algorithm Based on Improved Deep Networks

LI Qian-yu, JIANG Jian-guo, QI Mei-bin   

  1. School of Computer & Information, Hefei University of Technology, Hefei, Anhui 230009, China
  • Received:2015-06-24 Revised:2016-01-14 Online:2017-03-25 Published:2017-03-25
    • Supported by:
    • National Natural Science Foundation of China (No.61371155); Science and Technology Research and Development Program of Anhui Province (No.1301b042023)

摘要:

目前的人脸识别算法在其特征提取过程中采用手工设计(hand-crafted)特征或利用深度学习自动提取特征.本文提出一种基于改进深层网络自动提取特征的人脸识别算法,可以更准确地提取出目标的鉴别性特征.算法首先对图像进行ZCA(Zero-mean Component Analysis)白化等预处理,减小特征相关性,降低网络训练复杂度.然后,基于卷积、池化、多层稀疏自动编码器构建深层网络特征提取器.所使用的卷积核是通过单独的无监督学习获得的.此改进的深层网络通过预训练和微调,得到一个自动的深层特征提取器.最后,利用Softmax回归模型对提取的特征进行分类.本文算法在多个常用人脸库上进行了实验,表明了其在性能上比传统方法和普通深度学习方法都有所提高.

关键词: 人脸识别, 改进的深层网络, 卷积, 池化, 多层稀疏自动编码器

Abstract:

Current face recognition algorithms use hand-crafted features or extract features by deep learning.This paper presents a face recognition algorithm based on improved deep networks that can automatically extract the discriminative features of the target more accurately.Firstly,this algorithm uses ZCA(Zero-mean Component Analysis)whitening to preprocess the input images in order to reduce the correlation between features and the complexity of the training networks.Then,it organically combines convolution,pooling and stacked sparse autoencoder to get a deep network feature extractor.The convolution kernels are achieved through a separate unsupervised learning model.The improved deep networks get an automatic deep feature extractor through preliminary training and fine-tuning.Finally,the softmax regression model is used to classify the extracted features.This algorithm is tested on several commonly used face databases.It is indicated that the performance is better than the traditional methods and common deep learning methods.

Key words: face recognition, improved deep networks, convolution, pooling, stacked sparse autoencoder

中图分类号: