计算机应用 ›› 2017, Vol. 37 ›› Issue (2): 512-516.DOI: 10.11772/j.issn.1001-9081.2017.02.0512

• 人工智能 • 上一篇    下一篇

无监督局部特征学习的鲁棒性人脸识别

冯姝   

  1. 重庆大学 数学与统计学院, 重庆 401331
  • 收稿日期:2016-08-01 修回日期:2016-10-18 出版日期:2017-02-10 发布日期:2017-02-11
  • 通讯作者: 冯姝,fengshu@cqu.edu.cn
  • 作者简介:冯姝(1991-),女,山西人,硕士研究生,主要研究方向:机器学习、偏微分方程、图像处理。

Unsupervised local feature learning for robust face recognition

FENG Shu   

  1. College of Mathematics and Statistics, Chongqing University, Chongqing 401331, China
  • Received:2016-08-01 Revised:2016-10-18 Online:2017-02-10 Published:2017-02-11

摘要: 特征表示是人脸识别的关键问题,由于人脸图像在拍摄过程中受光照、遮挡、姿势等因素的影响,如何提取鲁棒的图像特征成了研究的重点。受卷积网络框架的启发,结合K-means算法在卷积滤波器学习中所具有的效果稳定、收敛速度快等优点,提出了一种简单有效的人脸识别方法,主要包含三个部分:卷积滤波器学习、非线性处理和空间平均值池化。具体而言,首先在训练图像中提取局部图像块,预处理后,使用K-means算法快速学习滤波器,每个滤波器与图像进行卷积运算;然后通过双曲正切函数对卷积图像进行非线性变换;最后利用空间平均值池化对图像特征进行去噪和降维。分类阶段仅采用简单的线性回归分类器。在AR和ExtendedYaleB数据集上的评估实验结果表明所提方法虽然简单却非常有效,而且对光照和遮挡表现出了强鲁棒性。

关键词: 人脸识别, 卷积网络框架, K均值, 空间平均值池化, 线性回归

Abstract: Image representation is a fundamental issue in face recognition, it is desired to design a discriminative and robust feature to alleviate the effect of illumination, occlusion and pose, etc. Motivated by the convolutional architecture and the advantages (stable result and fast convergence) of K-means algorithm in building filter bank, a very simple yet effective face recognition approach was presented. It consists of three main parts:convolutional filters leraning, nonlinear processing and spatial mean pooling. Firstly, K-means was employed based on preprocessed image patches to construct the convolution filters quickly. Each filter was convoluted with face image to extract sufficient and discriminative feature information. Secondly, the typical hyperbolic tangent function was applied for nonlinear projection on the convoluted features. Thirdly, spatial mean pooling was used to denoise and reduce the dimensions of the learned features. The classification stage only requires a novel linear regression classifier. The experimental results on two widely utlized databases such as AR and ExtendedYaleB demonstrate that the proposed method is simple and effective, and has strong robustness to illumination and occlusion.

Key words: face recognition, convolutional network architecture, K-means, spatial mean pooling, linear regression

中图分类号: