摘要： 传统的二维保局投影(2DLPP)算法未考虑样本邻域间局部信息,并且所提取的特征矩阵分量间存在相关性。针对该问题,提出基于大间距准则的最小相关性监督2DLPP 算法。引入类间局部散度矩阵和类内局部散度矩阵,最大化带权的散度矩阵迹差,以增大样本类间散度,减小样本类内散度,从而更好地刻画数据的流形结构。计算所提取特征矩阵各分量间的协方差矩阵,通过最小相关性分析,减少特征信息的冗余。在Yale 和ORL 人脸库上进行仿真实验,结果显示,当训练样本数为5 时,该算法的最高识别率分别为92. 5% 和96. 2% ,与传统2DLPP 算法、二维主成分分析法、二维线性判别分析法和二维大间距准则法相比,识别率均有所提高。同时对不同训练样本数下识别率均值和方差进行分析,验证了算法的稳定性。

关键词: 流形学习, 最大间距准则, 散度矩阵, 二维保局投影, 最小相关性, 人脸识别

Abstract: Two-dimensional Locality Preserving Projection(2DLPP) ignores the face sample local information between neighborhood and the correlation between the extracted feature matrix component problems. Aiming at this problem,the minimum correlated supervision 2DLPP algorithm based on Maximum Margin Criterion(MMC) is proposed. Between class local scatter matrix and within class local scatter matrix are brought in,which maximize the trace difference of scatter matrix to increase the sample’s between-class scatter and decrease within-class scatter,then manifold structure of data can be characterized better. It calculates the covariance matrix of extracted feature matrix, reduces the feature redundant. Experiments on Yale and ORL face database are done,when the train sample number is 5,the result shows that the highest recognition rates are 92. 5% and 96. 2% ,the recognition rate is higher than traditional 2DLPP algorithm,Twodimension Principal Component Analysis(2DPCA) algorithm,Two-dimension Linear Discriminate Analysis(2DLDA) algorithm and Two-dimension Maximum Margin Criterion(2DMCC) algorithm. It also analyses the mean and variance of recognition rate to prove the stability of the improved algorithm.

Key words: manifold learning, Maximum Margin Criterion ( MMC ), divergence matrix, Two-dimensional Locality Preserving Projection(2DLPP), minimum correlation, face recognition

中图分类号: 

• TP18