计算机工程与应用 ›› 2010, Vol. 46 ›› Issue (13): 18-22.DOI: 10.3778/j.issn.1002-8331.2010.13.006

• 博士论坛 • 上一篇    下一篇

基于光照线性子空间的人脸图像识别研究

刘中华1,2,史恒亮1,金 忠1

  

  1. 1.南京理工大学 计算机科学与技术学院,南京 210094
    2.河南科技大学 电子信息工程学院,河南 洛阳 471003
  • 收稿日期:2010-01-19 修回日期:2010-03-09 出版日期:2010-05-01 发布日期:2010-05-01
  • 通讯作者: 刘中华

Face recognition based on illumination linear subspaces

LIU Zhong-hua1,2,SHI Heng-liang1,JIN Zhong1   

  1. 1.School of Computer Science & Technology,Nanjing University of Science and Technology,Nanjing 210094,China
    2.Electronic Information Engineering College,Henan University of Science and Technology,Luoyang,Henan 471003,China
  • Received:2010-01-19 Revised:2010-03-09 Online:2010-05-01 Published:2010-05-01
  • Contact: LIU Zhong-hua

摘要: 以前的研究已经证明,在变化光照条件,甚至多个光源和阴影存在的情况下,对象(尤其是人脸)所形成的图像能够被低维线性子空间有效地表示。Lee证明了存在9个光源方向的统一配置,任一对象在这些统一配置光照方向下的9幅图像所组成的线性子空间,能够很好地表示该对象所有光照情况,并且达到了很好的识别性能。但它要求每一对象在统一配置光照方向下的9幅图像作为训练集,这一要求限制了它在实际中的应用。此外,商图像方法中简单的三维点光源模型无法很好地近似任意光照情况,因此,提出一种基于九维线性子空间的改进的商图像方法,并利用改进后的商图像方法合成对象在统一配置光源方向下的9幅图像,克服了Lee所提方法的不足。该文方法较好地满足了商图像方法的理论前提,从而达到了较好的图像合成和人脸识别性能。

关键词: 光照线性子空间, 商图像, 人脸识别, 图像合成

Abstract: Previous work has demonstrated that the image variation of many objects(human faces in particular) under variable lighting can be effectively modeled by low-dimensional linear spaces.Lee has proved that there exists configuration of 9 point light source directions,such that,by taking 9 images of an object under these single sources,the resulting subspace is an effective representation for recognition under a wide range of lighting conditions.The method proposed by Lee achieves better recognition performance.However,it requires the 9 images of each object illuminated by lights from the universal configuration to act as training samples,which restricts its applications in face recognition.In addition,3-d subspace in quotient image method doesn’t suffice to represent all images of an object under varying illuminations.Therefore,an improved method based on image on 9-d linear subspace constructed using 9 images of object under the configuration is presented to overcome these deficiencies.And 9 images of another object under the configuration are generated.The suggested method can meet the theoretical prerequisites of quotient image method,and achieve better image synthesis and face recognition results.

Key words: illumination linear subspace, quotient image, face recognition, image synthesis

中图分类号: