一种用于特征提取的保局判别分析算法

Locality Preserving Discriminant Analysis Algorithm for Feature Extraction

针对保局投影(LPP)为无监督算法的局限,提出了一种新的监督版的LPP,即保局判别分析(LPDA)算法。LPDA吸收了流形学习算法与最大边界准则(MMC)的共同特点,可以将高维的人脸数据投影到低维子空间,具有能处理新样本与无小样本问题的优点。与现有的多种经典相关方法相比,从Yale,UMIST及MIT 3个人脸数据库的实验结果表明,提出的LPDA算法在降维的同时提取了用于人脸识别的更有效的特征,人脸图像识别性能较好,具有较强的判别分析能力。

To address the limitation that locality preserving projection（LPP） algorithm belongs to unsupervised,a novel approach,named as locality preserving discriminant analysis（LPDA） was proposed.LPDA algorithm absorbs the common characteristics of the manifold learning algorithm and maximum margin criterion（MMC）,and can project the high-dimensional face data into the low-dimensional subspace.The new sample can be processed and the small sample size problem can be prevented.Compared with several classical and related methods,the experimental results from Yale,UMIST and MIT face databases show that LPDA algorithm can extract the more efficient features for face recognition while the dimensionality is reduced,and obtains much higher recognition accuracies and stronger power of classification.