电子学报 ›› 2006, Vol. 34 ›› Issue (10): 1900-1905.

• 论文 • 上一篇    下一篇

一种基于小波分解及鲁棒估计的ICA算法及其在人脸识别中的应用

安高云, 阮秋琦   

  1. 北京交通大学信息科学研究所,北京 100044
  • 收稿日期:2005-08-26 修回日期:2006-03-03 出版日期:2006-10-25 发布日期:2006-10-25

A Novel ICA Algorithm for Face Recognition Based on Wavelet Decomposition and Robust Estimation

AN Gao-yun, RUAN Qiu-qi   

  1. Institute of Information Science,Beijing Jiaotong University,Beijing 100044,China
  • Received:2005-08-26 Revised:2006-03-03 Online:2006-10-25 Published:2006-10-25

摘要: 鲁棒主分量分析(RPCA)模型在选取幅度参数时,忽略了各变量独有的统计特性.为克服RPCA模型的这一不足,本文提出了通用鲁棒主分量分析(GRPCA)模型,采用M估计器(M-Estimator)为每个变量估计符合其自身统计特性的幅度参数,以提高模型的鲁棒性和通用性,并在此基础上提出了一种集成小波分解、鲁棒估计及独立分量分析的WR-ICA人脸识别算法.WR-ICA对人脸识别中的多种外部干扰(残缺人脸图像、化妆及遮挡等)都表现出很好的鲁棒性.理论分析和实验结果证实了WR-ICA的有效性,采用Cos距离作相似性度量时,WR-ICA的平均识别率达到99.44%.

关键词: 人脸识别, 主分量分析, 独立分量分析, 小波分解, 鲁棒估计

Abstract: In the robust principal component analysis(RPCA)model,the statistical properties of every variable are neglected when scale parameters are chosen for them.In order to overcome this drawback,a generalized robust principal component analysis(GRPCA)model was proposed in this paper,a M-estimator was adopted to estimate robust scale parameters for every variable according to their statistical properties.Then,a new independent component analysis algorithm for face recognition based on wavelet decomposition and robust estimation (WR-ICA) was proposed.WR-ICA is robust to many types of outliers (incomplete face image,making up,occlusion,etc).The validity of WR-ICA is confirmed by theory analysis and experimental data,with Cos distance as similarity measurement,the average recognition rate of WR-ICA is 99.44%.

Key words: face recognition, principal component analysis, independent component analysis, wavelet decomposition, robust estimation

中图分类号: