基于流形学习的正交稀疏保留投影鉴别分析  

Orthogonal Sparsity Preserving Discriminant Analysis Based on Manifold Learning

作　　者：凌若冰[1] 荆晓远[1] 吴飞[2] 姚永芳[1] 李文倩[1] 

机构地区：[1]南京邮电大学自动化学院,江苏南京210023 [2]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2015年第1期66-69,73,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61272273);江苏省普通高校研究生科研创新计划(CXLX13_465)

摘　　要：稀疏保留投影(SPP)是一种保留样本间的稀疏重构关系的特征提取方法。但是根据流形学习理论,考虑局部流形结构比考虑全局欧氏结构更重要。此外,SPP得到的不是一组正交的投影向量,特征间存在冗余信息。为解决该问题,文中提出一种改进的稀疏保留投影算法,在SPP中引入有监督的流形学习,使得所得投影空间正交,并用迭代的方式求解最优投影变换,称为基于流形学习的迭代正交稀疏保留鉴别分析(MLIOSDA)。同时提出一种终止准则终止迭代。在CASPEAL人脸数据库和Poly U掌纹数据库的实验结果表明,文中提出的方法与一些相关方法相比有效地提高了识别结果。Sparsity Preserving Projections （ SPP） is an effective feature extraction method,which can preserve the sparse reconstruction re-lations among samples. However,according to the manifold learning theory,the local manifold structure of samples is more important than the global Euclidean structure of samples. SPP cannot get a set of orthogonal projection vectors,and thus there exists redundant informa-tion among the obtained features. To address these problems of SPP,propose a novel approach called Manifold Learning based Iterative Orthogonal Sparsity preserving Discriminant Analysis （ MLIOSDA） ,which introduces the idea of manifold learning into SPP and obtains orthogonal projection space. Obtain optimal projection vectors in an iterative manner. Also provide a terminating criterion to finish the it-eration. Experimental results on CAS-PEAL and PolyU databases demonstrate that the proposed approach can effectively improve the rec-ognition results compared with some related methods.

关 键 词：特征提取 流形学习 稀疏保留投影 正交 鉴别 终止准则 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]