摘要： 保局投影(LPP)忽略了数据的类别标记信息且鲁棒性较差，为此，提出一种线性判别投影(LDP)算法。引入类间权重矩阵和类内权重矩阵，使各流形间的分离性最大，局部子流形的内在紧致性最小，同时通过一种鲁棒的类内处理方式使算法对outlier数据具有鲁棒性。在ORL、AR和Extended Yale B人脸数据集上进行实验，结果表明，与PCA、LDA、LPP、LSDA和LPDP算法相比，该算法的最佳平均识别率较高，分别可达95.3%、93.64%和96.28%，证明了算法的有效性和可靠性。

关键词: 降维, 流形学习, 判别投影, 有监督学习, 保局投影, 线性判别分析

Abstract: Because Locality Preserving Projection(LPP) ignores the label information of the data and it is lack of robustness, this paper proposes a Linear Discriminant Projection(LDP) algorithm. By introducing between-class weight matrix and within-class weight matrix, LDP maximizes the separability of different submanifolds and minimizes the compactness of local submanifolds. Moreover, LDP is robust to outlier data by a robust within-class processing way. Compared with PCA, LDA, LPP, LSDA, LPDP, the experimental results on ORL, AR and Extended Yale B face databases show that the best average recognition rates of LDP are higher, which can reach 95.3%, 93.64% and 96.28%, and this verifies the efficiency of the proposed algorithm.

Key words: dimensionality reduction, manifold learning, discriminant projection, supervised learning, Locality Preserving Projection (LPP), Linear Discriminant Analysis(LDA)

中图分类号: 

• TP18