电子学报 ›› 2012, Vol. 40 ›› Issue (10): 1965-1970.DOI: 10.3969/j.issn.0372-2112.2012.10.010

• 学术论文 • 上一篇    下一篇

姿态和光照可变条件下的仿射最小线性重构误差人脸识别算法

平强, 庄连生, 俞能海   

  1. 中国科学技术大学多媒体计算与通信教育部-微软重点实验室, 安徽合肥 230027
  • 收稿日期:2011-01-10 修回日期:2012-03-28 出版日期:2012-10-25
    • 作者简介:
    • 平 强 男.1986年10月生于江苏苏州.中国科学技术大学硕士研究生,研究方向为计算机视觉、人脸识别技术等. E-mail:pingq@mail.ustc.edu.cn庄连生 男.讲师,博士,1978年生于福建.主要研究方向为计算机视觉、机器学习等.俞能海 男.教授,博士.1964年生于安徽.主要研究方向为图像处理与视频通信、多媒体信息检索、多媒体内容安全等.
    • 基金资助:
    • 国家自然科学基金 (No.61103134,No.60933013); "新一代宽带无线移动通信网"国家科技重大专项 (No.2010ZX03004-003); 中央高校基本科研业务经费专项资金 (No.WK210023002,No.WK2101020003); 安徽省优秀青年人才基金 (No.BJ2101020001)

Affine Minimum Linear Reconstruction Error Face Recognition Under Varying Pose and Illumination

PING Qiang, ZHUANG Lian-sheng, YU Neng-hai   

  1. MOE-MS Key Laboratory of Multimedia Computing and Communication, University of Science and Technology of China, Hefei, Anhui 230027, China
  • Received:2011-01-10 Revised:2012-03-28 Online:2012-10-25 Published:2012-10-25
    • Supported by:
    • National Natural Science Foundation of China (No.61103134, No.60933013); National Science and Technology Major Special Project of National New Generation of Broadband Wireless Mobile Communication Networks (No.2010ZX03004-003); Special Fund of Fundamental Research Funds for the Central Universities (No.WK210023002, No.WK2101020003); Outstanding Young Talent Fund Project of Anhui Province (No.BJ2101020001)

摘要: 传统人脸识别算法通常把光照处理和姿态校正作为两个相对独立的处理过程,难以取得全局最优识别性能.针对该问题,本文根据人脸的非刚体特性,将仿射变换和分块思想融入线性重构模型中,提出了一种基于仿射最小线性重构误差(Affine Minimum Linear Reconstruction Error,AMLRE)的人脸识别算法,在处理光照问题的同时能够补偿姿态变化造成的局部区域对齐误差,以获得更好的全局识别性能.在公共数据集上的实验结果表明,本文提出的算法对光照和姿态有很好的鲁棒性,同时与现有的人脸识别算法相比,本文的算法具有更高的识别率.

关键词: 人脸识别, 线性重构, 仿射变换, Lucas-Kanade算法

Abstract: Traditional face recognition algorithms usually handle variations in illumination and pose independently.Therefore,it is difficult to obtain the global optimal recognition performance.To this end,we propose an affine minimum linear reconstruction error (AMLRE) algorithm based on the non-rigid characteristics of human faces in this paper,which combines an affine transformation model and the idea of patch with a linear reconstruction model. Our algorithm simultaneously handles illumination variations as well as compensates the local area alignment errors caused by pose variations,which achieves much better recognition performance.Comprehensive experiments on several public face datasets clearly demonstrate that our proposed algorithm is robust to both illumination and pose,and thus outperforms most state-of-the-art methods.

Key words: face recognition, linear reconstruction, affine transformation, lucas-kanade algorithm

中图分类号: