多视图特征学习方法研究
发布时间：2019-06-02 07:44
【摘要】：多视图数据从多个角度刻画同一物体,包含了比传统的单视图数据更加丰富的分类识别信息,因此近年来多视图学习技术成为了一个研究热点。不同视图数据往往存在一定的信息冗余,如何充分地从多视图数据中提取有用特征并且消除冗余是多视图学习技术的关键问题。针对该问题,本文对多视图学习技术进行了系统的研究,主要研究成果总结如下:一、提出了两个多视图鉴别分析方法,即组递归鉴别子空间学习(GRDSL)和不相关局部敏感多视图鉴别分析(ULSMDA)。GRDSL在数据层融合多视图数据,使用递归学习的方式将样本集分解成多个近似集和相应的差分集,并在每次递归的差分集中学习一个鉴别变换。GRDSL设计了递归终止准则以及投影向量选择规则,并能在理论上保证多个鉴别变换的正交性。通过自适应的递归学习过程,GRDSL可以从多视图数据中有效地学得充足的有用特征。ULSMDA联合学习了多个视图特定的投影变换,使得在投影空间中,原始近邻的同类样本相互聚集,而原始近邻的异类样本相互排斥。ULSMDA考虑了跨视图数据的一致性,并设计了不相关约束,用于减少变换间的冗余。ULSMDA充分地使用了多视图数据的局部结构信息用于不相关鉴别特征的学习。四个数据集上的实验结果表明了这两个方法的有效性。二、提出了三个多视图字典学习方法,即不相关多视图鉴别字典学习(UMD2L)、多视图低秩字典学习(MLDL)和多视图低秩共享结构化字典学习(MLS2DL)。通过使得字典原子与类别标记相对应,UMD2L从多视图数据中联合地学习多个结构化字典。UMD2L设计了不相关约束用于减少不同视图字典间的冗余。从增强字典鉴别能力以及消除冗余这两方面出发,UMD2L提升了多视图字典学习技术有用特征学习的能力。MLDL将低秩学习技术引入到多视图学习技术中,运用低秩矩阵恢复理论来解决噪声存在情况下的多视图字典学习问题。MLDL设计了结构化不相关约束,并为多视图字典学习技术提供了高效的基于联合表示的分类机制。MLDL为多视图字典学习技术提供了在噪声影响情况下充分学习有用特征的方案。MLS2DL关注视图间共享信息的挖掘,提出在学习多个视图特定的低秩结构化字典的同时对视图共享低秩结构化字典进行学习。MLS2DL为多视图字典学习技术提供了在消除视图间冗余信息的同时有效利用多视图有利相关性的方案。实验证明了相比于代表性的多视图子空间学习方法和多视图字典学习方法以及提出的GRDSL和ULSMDA方法,这三个方法可以获得更优的分类效果。三、提出了一个半监督多视图鉴别分析方法,即不相关半监督视图内和视图间流形鉴别学习(USI2MD)。USI2MD给出了半监督视图内和视图间流形鉴别学习机制,使用无标记样本的局部近邻结构,以及有标记样本视图内和视图间的鉴别信息来从多视图数据中提取特征。USI2MD设计了一个半监督不相关约束,用于减少半监督场景下不利的多视图特征相关性。USI2MD充分使用了视图内和视图间的有用信息用于半监督场景下不相关鉴别特征的学习。实验证实了该方法相对于代表性的半监督多视图子空间学习方法的有效性。
[Abstract]:Multi-view data depict the same object from multiple angles, which contains the classification identification information which is more abundant than the traditional single-view data, so the multi-view learning technology has become a hot topic in recent years. Different view data often have some information redundancy, how to extract useful features from multi-view data sufficiently and to eliminate redundancy is a key issue in multi-view learning technology. In view of this problem, this paper studies the multi-view learning technology, and the main research results are as follows:1. Two multi-view identification and analysis methods are put forward. I.e., group recursive authentication sub-space learning (grDSL) and non-relevant local-sensitive multi-view identification analysis (ULSDA). And a discrimination transformation is learned every time the differential concentration is recursive. GRDSL has designed the recursive termination criteria and the projection vector selection rules and can theoretically ensure the orthogonality of multiple discrimination transforms. Through the self-adaptive recursive learning process, the GRDSL can have sufficient useful features in the multi-view data efficiently. The ULSFDA combines multiple view-specific projection transformations so that in the projection space, the same samples of the original neighbors are aggregated with each other, while the heterogeneous samples of the original neighbors are mutually exclusive. ULSDA takes into account the consistency of cross-view data, and designs non-related constraints to reduce the redundancy between transforms. The ULSDA fully uses the local structure information of the multi-view data for learning of the non-related authentication features. The experimental results on the four data sets show the effectiveness of the two methods. Two, three multi-view dictionary learning methods (UMD2L), multi-view low-rank dictionary learning (MLDL) and multi-view low-rank shared structured dictionary learning (MLS2DL) are proposed. By making the dictionary atoms correspond to the category tags, the UMD2L jointly learns a plurality of structured dictionaries from the multi-view data. The UMD2L is designed with non-related constraints to reduce the redundancy between different view dictionaries. From the two aspects of enhancing the ability of the dictionary and eliminating the redundancy, the UMD2L improves the ability of the multi-view dictionary learning to study the useful features. MLDL introduces the low-rank learning technology into the multi-view learning technology, and uses the low-rank matrix recovery theory to solve the problem of multi-view dictionary learning in the case of noise. MLDL has designed a structured non-related constraint and provides a highly efficient joint-representation-based classification mechanism for multi-view dictionary learning technology. The MLDL is a multi-view dictionary learning technique that provides a scheme that fully learns useful features in the event of noise impact. The MLS2DL focuses on the mining of shared information among the views, and proposes to study the view-sharing low-rank structured dictionary while learning a plurality of view-specific low-rank structured dictionaries. The MLS2DL is a multi-view dictionary learning technique that provides a scheme to effectively utilize multi-view advantageous correlation while eliminating redundant information between views. The experiments prove that the three methods can obtain better classification effect than the representative multi-view subspace learning method and multi-view dictionary learning method and the proposed GRDSL and ULSDA method. in this paper, a semi-supervised multi-view identification and analysis method is proposed, that is, there is no relevant semi-supervised view and inter-view manifold discrimination learning (US2MD). US2MD gives a semi-supervised and inter-view manifold discrimination learning mechanism, And identifying information in the view of the tagged sample and the view to extract features from the multi-view data. US2MD designed a semi-supervised non-related constraint to reduce the negative multi-view feature dependency in a semi-supervised scenario. The US2MD fully uses the useful information in the view and between the views to be used to semi-monitor the learning of the non-relevant authentication features in the scene. The experiment verifies the effectiveness of the method with respect to the representative semi-supervised multi-view subspace learning method.
【学位授予单位】：南京邮电大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP181
本文编号：2490936
[Abstract]:Multi-view data depict the same object from multiple angles, which contains the classification identification information which is more abundant than the traditional single-view data, so the multi-view learning technology has become a hot topic in recent years. Different view data often have some information redundancy, how to extract useful features from multi-view data sufficiently and to eliminate redundancy is a key issue in multi-view learning technology. In view of this problem, this paper studies the multi-view learning technology, and the main research results are as follows:1. Two multi-view identification and analysis methods are put forward. I.e., group recursive authentication sub-space learning (grDSL) and non-relevant local-sensitive multi-view identification analysis (ULSDA). And a discrimination transformation is learned every time the differential concentration is recursive. GRDSL has designed the recursive termination criteria and the projection vector selection rules and can theoretically ensure the orthogonality of multiple discrimination transforms. Through the self-adaptive recursive learning process, the GRDSL can have sufficient useful features in the multi-view data efficiently. The ULSFDA combines multiple view-specific projection transformations so that in the projection space, the same samples of the original neighbors are aggregated with each other, while the heterogeneous samples of the original neighbors are mutually exclusive. ULSDA takes into account the consistency of cross-view data, and designs non-related constraints to reduce the redundancy between transforms. The ULSDA fully uses the local structure information of the multi-view data for learning of the non-related authentication features. The experimental results on the four data sets show the effectiveness of the two methods. Two, three multi-view dictionary learning methods (UMD2L), multi-view low-rank dictionary learning (MLDL) and multi-view low-rank shared structured dictionary learning (MLS2DL) are proposed. By making the dictionary atoms correspond to the category tags, the UMD2L jointly learns a plurality of structured dictionaries from the multi-view data. The UMD2L is designed with non-related constraints to reduce the redundancy between different view dictionaries. From the two aspects of enhancing the ability of the dictionary and eliminating the redundancy, the UMD2L improves the ability of the multi-view dictionary learning to study the useful features. MLDL introduces the low-rank learning technology into the multi-view learning technology, and uses the low-rank matrix recovery theory to solve the problem of multi-view dictionary learning in the case of noise. MLDL has designed a structured non-related constraint and provides a highly efficient joint-representation-based classification mechanism for multi-view dictionary learning technology. The MLDL is a multi-view dictionary learning technique that provides a scheme that fully learns useful features in the event of noise impact. The MLS2DL focuses on the mining of shared information among the views, and proposes to study the view-sharing low-rank structured dictionary while learning a plurality of view-specific low-rank structured dictionaries. The MLS2DL is a multi-view dictionary learning technique that provides a scheme to effectively utilize multi-view advantageous correlation while eliminating redundant information between views. The experiments prove that the three methods can obtain better classification effect than the representative multi-view subspace learning method and multi-view dictionary learning method and the proposed GRDSL and ULSDA method. in this paper, a semi-supervised multi-view identification and analysis method is proposed, that is, there is no relevant semi-supervised view and inter-view manifold discrimination learning (US2MD). US2MD gives a semi-supervised and inter-view manifold discrimination learning mechanism, And identifying information in the view of the tagged sample and the view to extract features from the multi-view data. US2MD designed a semi-supervised non-related constraint to reduce the negative multi-view feature dependency in a semi-supervised scenario. The US2MD fully uses the useful information in the view and between the views to be used to semi-monitor the learning of the non-relevant authentication features in the scene. The experiment verifies the effectiveness of the method with respect to the representative semi-supervised multi-view subspace learning method.
【学位授予单位】：南京邮电大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP181
【参考文献】
相关期刊论文 前1条
1 程圣军;刘家锋;黄庆成;唐降龙;;基于样本条件价值改进的Co-training算法[J];自动化学报;2013年10期
，本文编号：2490936
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2490936.html
