基于L1范数的二维局部保留映射  

Two-dimensional Locality Preserving Projections Based on L1-norm

作　　者：邢红杰[1] 赵浩鑫[1] 

机构地区：[1]河北大学数学与计算机学院 河北省机器学习与计算智能重点实验室,保定071002

出　　处：《计算机科学》2012年第5期201-204,238,共5页Computer Science

基　　金：国家自然科学基金(60903089,61073121);中国博士后科学基金(20080440820);河北省自然科学基金(F2009000231)资助

摘　　要：提出了一种基于L1范数的二维局部保留映射(two-dimensional locality preserving projections based on L1-norm,2DLPP-L1)特征提取方法。与传统的基于L2范数的二维局部保留映射(2DLPP)相比,所提方法有两个优点。首先,由于L1范数对噪声不敏感,因此它具有更强的抗噪声能力;其次,它不需要进行特征值分解。在两个人脸数据库和一个手写数字数据集上的实验结果表明,当训练集中有噪声时,所提的2DLPP-L1能够取得优于传统2DLPP的分类性能。This paper presented a method of two-dimensional locality preserving projection based on L1-norm（2DLPP-L1）.The proposed approach has two advantages compared with the conventional L2-norm based two-dimensional localitypreserving projection（2DLPP）.Firstly,it is more robust against outliers because L1-norm is insensitive to noises.Moreover,it does not require the eigenvalue decomposition.Experiments on two face databases and one hand-written digit dataset illustrate that compared with 2DLPP,the proposed method exhibits better performance when there are outliers in training sets.

关 键 词：特征提取 L1范数 局部保留映射 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]