基于流形学习的正交稀疏保留投影  

Orthogonal Sparsity Preserving Projections Based on Manifold Learning

作　　者：刘茜[1,2] 荆晓远[1,2] 李文倩[1] 姚永芳[1] 

机构地区：[1]南京邮电大学自动化学院,江苏南京210003 [2]武汉大学软件工程国家重点实验室,湖北武汉430072

出　　处：《计算机技术与发展》2014年第7期34-37,42,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61073113;61272273);江苏省333工程(BRA2011175)

摘　　要：稀疏保留投影通过保留样本之间的全局稀疏重构关系来进行特征提取,获得了良好的分类效果。但是,稀疏保留投影得到的投影变换通常不是正交的,而且在实际应用中,正交性一直被认为有利于提高鉴别能力。另外,根据流形学习理论,局部流形结构比全局欧式结构更重要。因此,文中在稀疏保留投影中引入了流形结构保留和正交投影,提出了整体正交流形稀疏保留投影(HOMSPP)和迭代正交流形稀疏保留投影(IOMSPP)两种实现算法来实现人脸和掌纹图像的特征提取。Sparsity Preserving Projections （ SPP） extracts features by preserving the global sparse reconstruction relations among samples, which achieves favorable classification results. However,the obtained transformation of SPP usually is not orthogonal,while in real appli-cations,orthogonality is advantageous for classification in many scenarios. Besides,according to the manifold learning theory,local mani-fold structure is more important than global Euclidean structure. Therefore,in this paper,introduce manifold preserving and orthogonal transformation into SPP,and propose two novel approaches for face and palmprint image feature extraction,which are Holistic Orthogonal Manifold and Sparsity Preserving Projections （ HOMSPP） and Iterative Orthogonal Manifold and Sparsity Preserving Projections （ IOM-SPP） .

关 键 词：人工智能 人脸和掌纹图像特征提取 流形学习 正交稀疏保留投影 子空间学习 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]