电子学报 ›› 2016, Vol. 44 ›› Issue (8): 1940-1946.DOI: 10.3969/j.issn.0372-2112.2016.08.024

• 学术论文 • 上一篇    下一篇

光照健壮人脸识别的低秩相对梯度直方图特征

杨利平, 李武   

  1. 重庆大学光电技术及系统教育部重点实验室, 重庆 400044
  • 收稿日期:2015-02-13 修回日期:2015-05-07 出版日期:2016-08-25 发布日期:2016-08-25
  • 作者简介:杨利平 男,1981年生于内蒙古鄂尔多斯.重庆大学光电工程学院副教授.研究方向为模式识别、图像复原、声音信号处理.E-mail:yanglp@cqu.edu.cn;李武 男,1989年生于四川成都.硕士研究生,研究方向为模式识别、图像处理.
  • 基金资助:
    中央高校基本科研业务费专项(No.106112015CDJXY120013)

Low-Rank Relative Gradient Histogram Features for Illumination-Robust Face Recognition

YANG Li-ping, LI Wu   

  1. Key Laboratory of Optoelectronic Technology and Systems(Chongqing University), Ministry of Education, Chongqing 400044, China
  • Received:2015-02-13 Revised:2015-05-07 Online:2016-08-25 Published:2016-08-25

摘要: 为了进一步提升人脸梯度特征的光照健壮性,本文结合低秩分解能有效分离图像本质特征和噪声的特性,提出了一种光照健壮的低秩相对梯度直方图特征提取方法.首先,通过对人脸图像进行相对梯度运算获得了图像的相对梯度幅值图像和各像素的梯度方向信息.然后,为了去除相对梯度图像中由于非均匀光照而引入的光照边缘误差,利用低秩分解将相对梯度图像分解为低秩分量和稀疏噪声分量之和.最后,结合人脸图像的梯度方向信息对相对梯度图像的低秩分量进行离散化、滤波和局部二值模式编码形成了人脸的低秩相对梯度直方图特征.在经典的FERET子集以及具有代表性的YaleB和PIE光照子集上的实验显示:低秩相对梯度直方图特征的人脸识别性能显著优于相对梯度直方图特征、方向梯度幅值模式特征和图像低秩特征等方法的性能;在YaleB子集上,低秩相对梯度直方图特征的人脸识别精度比相对梯度直方图特征的人脸识别精度高至少4%.实验结果证明,低秩相对梯度直方图特征对光照变化,尤其是非均匀光照变化的人脸识别具有很强的健壮性.

关键词: 人脸识别, 低秩分解, 相对梯度, 光照健壮性

Abstract: To further enhance the illumination robustness of facial gradient features,an illumination robust low-rank relative gradient histogram feature extraction method (LR-RGHF) is proposed based on the property that low-rank decomposition can separate the intrinsic characteristic and sparse noise of an image effectively.In the first place,the relative gradient magnitude image and gradient direction information of each pixel is obtained by doing relative gradient operation on face image.In the next place,in order to remove the edge error caused by uneven illumination distribution,the relative gradient magnitude image is decomposed into its low-rank component and sparse noise component using low-rank decomposition.Finally,the low-rank component of relative gradient magnitude image is decomposed into several sub images according to its gradient direction information,each of the sub images is then filtered and encoded using local binary pattern to form the LR-RGHF.Experimental on the classical FERET subset and representative illumination subsets,YaleB and PIE subsets,illustrated that the recognition performance of LR-RGHF outperforms the recognition performances of relative gradient histogram feature (RGHF),patterns of oriented edge magnitudes (POEM) and low rank facial feature.On YaleB subset,the recognition accuracies of LR-RGHF are at least 4% higher than RGHF.Experimental results demonstrate that LR-RGHF reveals strong illumination robustness for face recognition,especially for uneven illumination distribution.

Key words: face recognition, low-rank decomposition, relative gradient, illumination robustness

中图分类号: