Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (23): 245-253.DOI: 10.3778/j.issn.1002-8331.2105-0450

• Graphics and Image Processing • Previous Articles     Next Articles

Marginal Manifold Embedding for Feature Extraction

GONG Sicong, XU Jie, WAN Minghua   

  1. 1.Faculty of Automation, Guangdong University of Technology, Guangzhou 510006, China
    2.Faculty?of?Information?Engineering, Nanjing Audit University, Nanjing 211815, China
  • Online:2022-12-01 Published:2022-12-01

边界流形嵌入在特征提取中的应用

龚思聪,徐洁,万鸣华   

  1. 1.广东工业大学 自动化学院,广州 510006
    2.南京审计大学 信息工程学院,南京 211815

Abstract: Marginal information of sample points is of great importance for classification. In view of fact that marginal information of sample points used in constructing the intrinsic graph and the penalty graph by marginal Fisher analysis(MFA) and locality sensitive discriminant analysis(LSDA) can’t sufficiently characterize the separability of different classes of sample points in some cases, a novel graph embedding dimensionality reduction algorithm called marginal manifold embedding(MME) is proposed. According to the label information of sample points, firstly, MME algorithm finds the nearest heterogeneous margin sub-manifold of each sample point,  then returns to its own category to find corresponding homogenous margin sub-manifold that is the nearest one to the heterogeneous margin sub-manifold, thus defining the homogeneous margin neighbourhood and the heterogeneous margin neighbourhood closely related to different classes of sample points. Finally, by maximizing the distance between all paired margin sub-manifolds, MME algorithm can obtain a low-dimensional space with more discriminant significance. At the same time, MME algorithm can trap outliers wandering in the margin into the marginal neighbourhood, which is helpful to reduce the negative influence of outliers on the algorithm. Experimental results on face datasets show that the low-dimensional features extracted by MME algorithm can improve the classification accuracy.

Key words: margin, classification, graph embedding, dimensionality reduction, outliers

摘要: 样本点的边界信息对于分类具有重要意义。针对于边界Fisher分析(MFA)和局部敏感判别分析(LSDA)构造本征图和惩罚图所利用的样本点边界信息,在一些情况下并不能很好地表征不同类样本点的可分性,提出了一种新的图嵌入降维算法——边界流形嵌入(MME)。MME算法根据样本点的标签信息,寻找距离每个样本点最近的异类边界子流形,再返回本类中寻找距离异类边界子流形最近的同类边界子流形,从而定义出不同类样本间密切联系的同类边界邻域和异类边界邻域。通过最大化所有成对的边界子流形之间的距离,MME算法可以得到更具有鉴别意义的低维特征空间。同时,MME算法能将徘徊在边界的离群点收入到边界邻域里,这对减弱离群点给算法带来的负面的影响有一定的帮助。在人脸数据库上的实验结果表明了MME算法提取的低维特征能够提升分类的准确率。

关键词: 边界, 分类, 图嵌入, 降维, 离群点