一种应用于人脸识别的核正交等度规映射算法  

A kernel orthogonal isometric projection algorithm for face recognition

作　　者：王庆军[1] 张汝波 刘冠群 

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《光电子．激光》2010年第11期1702-1705,共4页Journal of Optoelectronics·Laser

基　　金：国家高技术研究发展计划(863)资助项目(2009AA04Z215);国家自然基金资助项目(60803036)

摘　　要：针对人脸识别中的特征提取,提出了一种新的核正交等度规映射(KOIsoP,kernel orthogonal isometric projection)人脸识别算法。首先用核方法提取人脸图像中的非线性信息,并将其投影在一个高维非线性空间,从而更好地提取人脸非线性流形结构信息。然后通过等度规映射做一线性映射得到基向量。最后正交化得到的基向量,使得算法更利于保留人脸非线性子流形空间与距离有关的结构信息和重构样本,以便获得更好的识别效果。ORL和PIE库上的人脸识别实验验证了算法的有效性。In this paper, a new feature extraction method for face recognition named as kernel orthogonal isometric projection （KOIsoP） is proposed.First,the face images are mapped into a high-dimensional nonlinear space by the kernel method to obtain the nonlinear information,which helps to extract the nonlinear manifold structure information,and then a linear transformation is performed by the isoprojection to get the basis vectors.Finally the basis vectors are orthogonalized,which can preserve the nonlinear manifold subspace in faces and the distance-based structure information,as well as reconstructed datasets,so the proposed algorithm can gain higher recognition rate with lower operation compexity.Experiments on ORL and PIE face database demonstrate the effectiveness of the proposed algorithm.

关 键 词：人脸识别 子空间 流形 核正交等度规映射(KOIsoP) 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]