多模态耦合特征子空间正则的SVDD  

Coupled Feature Subspace Regularized Support Vector Data Description for Multimodal Data

作　　者：王闯 胡文军[1,2] 刘闯 王余波 WANG Chuang;HU Wenjun;LIU Chuang;WANG Yubo(School of Information Engineering,Huzhou University,Huzhou 313000,China;Zhejiang Province Key Laboratory of Smart Management&Application of Modern Agricultural Resources,Huzhou 313000,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000 [2]浙江省现代农业资源智慧管理与应用研究重点实验室,浙江湖州313000

出　　处：《湖州师范学院学报》2023年第8期51-61,共11页Journal of Huzhou University

基　　金：国家自然科学基金项目(61772198)。

摘　　要：针对传统支持向量数据描述(SVDD)方法无法用于多模态数据异常检测的问题,提出一种新颖的用于处理多模态数据的SVDD方法.该方法将多模态数据通过投影矩阵映射到公共低维子空间,再利用多模态图正则SVDD来保持模态内与模态间的结构关系,同时利用稀疏投影矩阵正则SVDD来降低原始空间中的特征耦合影响.该方法称为耦合特征子空间正则的支持向量数据描述(CFSR-SVDD).实验结果表明,所提出的方法在精度和稳定性上具有更好的优势.For the problem that the traditional support vector data description(SVDD) method cannot be used for anomaly detection with multi-modal data,a novel SVDD method for multi-modal data is proposed.It first maps multi-modal data to a common low-dimensional subspace by the projection matrix,and then uses the multi-modal graph to regularize the SVDD to maintain the structural relationship between modals.Meanwhile,the sparse projection matrix is used to regularize the SVDD to reduce the effect of feature coupling in the original space.This method is called coupled feature subspace regularized SVDD(CFSR-SVDD).The experimental results show that the proposed method has better accuracy and stability than other methods.

关 键 词：一类分类 多模态数据 支持向量数据描述 子空间学习 多模态图 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]