正则化保局鉴别分析方法

Regularized Locality Preserving Discriminant Analysis

提出一种正则化保局鉴别分析方法(RLPDA)并将其应用于人脸识别.受样本有限制约和大量噪声干扰,保局类内散布矩阵的零特征值及小特征值估计不准确,进而影响鉴别保局投影算法的性能.结合倒数谱模型对保局类内散布矩阵的特征值进行正则化,并利用正则化后的特征值对相应的特征空间加权,使人脸空间被保留,噪声空间被削弱,而零空间则被加强.通过分析鉴别信息在数据空间的分布可发现,RLPDA方法有效利用整个特征空间的鉴别信息,有利于提高算法的识别精度,同时从原理上回避小样本问题.在FERET和UMIST人脸数据库上的识别结果表明,RLPDA是一种有效的人脸特征提取方法.

A regularized locality preserving discriminant analysis （RLPDA） for face recognition is proposed. Affected by the small sample size （SSS） problem and noises, zero eigenvalues and small eigenvalues of locality preserving within-class scatter matrix are inadequate. It degrades the performance of discriminant locality preserving projections （DLPP）. In this paper, eigenvalues of locality preserving within-class scatter matrix are regularized by a reciprocal spectrum model, and the subspaces are weighted according to the regularized eigenvalues. Specifically, the face subspace is kept, the noise subspace is weakened, and the zero subspace is enhanced. Through the analysis of the distribution of discriminant information in data space, it is found that RLPDA utilizes the whole discriminant information. Hence, RLPDA improves the recognition accuracies and avoids the SSS problem in principal. The experimental results on FERET and UMIST face databases illustrate the effectiveness of the proposed RLPDA algorithm.