有局部差异及全局性的半监督正交保局投影算法  

A Semi-Supervised Orthogonal Locality Preserving Projection Algorithm Based on Diversity and Global Information

作　　者：梁兴柱[1] 林玉荣[2] 林玉娥[1] 

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]哈尔滨工业大学航天学院,黑龙江哈尔滨150001

出　　处：《武汉大学学报（理学版）》2013年第6期558-562,共5页Journal of Wuhan University:Natural Science Edition

基　　金：国家自然科学基金(60975009);安徽省自然科学基金(1208085QF123);安徽理工大学青年教师科学研究基金(2012QNZ10)资助项目

摘　　要：结合局部结构及差异信息的有监督投影算法是一种有效的特征提取算法,但只能处理有类别标签样本,忽略了样本全局信息的作用,故本文提出了一种有局部差异及全局性的半监督正交保局投影算法.该算法的思想是利用有标签和无标签样本来构造准则函数,在保持数据的局部几何结构的同时,最大化样本的全局与近邻样本的差异信息,因此该算法不但能够揭示样本的全局结构而且可有效地防止过学习问题,同时为了进一步增强算法的识别性能对所求取的投影矩阵进行了正交化约束,最后给出了该算法的非线性拓展.人脸库上的实验结果表明所提方法是可行的和有效的.Supervised local structure and diversity projection is an effective method for feature extraction,but the method can only deal with labeled samples and ignore global information of samples.Then,this paper proposes a novel method called semi-supervised orthogonal locality preserving projection algorithm based on local diversity and global information.By making use of both labeled and unlabeled data the criterion function of the algorithm is established,which efficiently preserves the local differences and simultaneously maximize the global structure.So the algorithm can not only reveal the global structure and but also effectively avoid the data over-fitting problem.In order to further enhance the recognition,projection matrix extracted is orthogonal.Finally,the algorithm of nonlinear expansion is given.Experimental results on face databases demonstrate that the proposed method is effective and feasible.

关 键 词：局部结构 差异 特征提取 全局信息 半监督 过学习问题 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]