基于多视图鉴别特征学习的分类算法

发布时间：2018-01-04 07:26

本文关键词：基于多视图鉴别特征学习的分类算法　出处：《中国矿业大学(北京)》2016年博士论文　论文类型：学位论文

【摘要】：模式识别、机器学习等交叉学科需要从观察到的数据中发现规律。最近的十几年来,互联网、通信等信息技术得到了革命性的发展,而信息技术的发展促使当今社会所产生的数据量极速增长,其中有很多数据能够以多种不同的形式进行表示。比如,在互联网中,每个Web网页能够表示为其所含文档和指向它的超链接;人脸识别领域中,可以对同一人脸图像样本提取出不同形态的特征形式,如Gabor特征,HOG特征,LBP特征,PCA特征分别用来描述人脸的方向尺度特征,边缘轮廓特征,局部像素灰度变化特征以及整体主要信息特征等。传统的基于单视图的分析算法,仅利用单一视图内的结构特性,没有利用视图间的关联、互补信息,多视图学习方法则尝试在不同的视图之间提取出相互关联、互补的特征,从而可以改善在数据集上的学习或分类效果。因此在最近的十几年以来,多视图特征学习在机器学习、数据挖掘和计算机视觉等领域受到了广泛的关注。本文以研究多视图数据的分类方法为主题,以提取数据中的鉴别特征为重点,从子空间鉴别特征提取、半监督流行学习和鉴别字典学习三个方面入手,做了一些创新工作,其主要内容包括:(1)以典型相关性分析(Correlation Canonical Analysis,CCA)为基础,分别对鉴别典型相关性分析(Discriminant Correlation Canonical Analysis,DCCA)、多视图鉴别分析(Multi-view Discriminant Analysis,MvDA)、增强组合特征鉴别相关性分析(Combined-Feature-Discriminability Enhanced Canonical Correlation Analysis,CECCA)等算法进行研究分析,提出二重鉴别相关性分析(Dual Discriminant Correlation Analysis,DDCA)方法。DDCA算法设计的模型具有两点优势:其一,能够在每个视图内借助于Fisher鉴别分析(Fisher Discriminant Analysis,FDA)寻找投影向量以保证样本的可分性;其二,能够在视图之间分析样本的鉴别相关性,即寻找投影向量使得样本之间的类内相关性最大,类间相关性最小。DDCA是一种有监督的特征提取方法,相比较于CCA能够有效利用样本的标签信息;此外,传统相关性分析方法由于自身模型的限制,仅适用于两个视图之间,而忽略了视图内部自身的信息,而DDCA在同一视图内部和不同视图之间均能够对数据进行分析,基于以上几点,DDCA有助于改善分类效果。(2)在半监督场景下,提取每个样本的多个视图特征有助于进一步挖掘样本多方面的信息,目前已有学者和研究人员们提出了一些有效的半监督多视图学习方法。尽管现有的半监督多视图特征学习方法已经取得了一定的效果,但是这些方法并不能很好的同时考虑到视图内和视图间的鉴别信息,而且如何有效地提取无标记样本的近邻结构信息,也具有较大的提升空间。本文提出了一种新的半监督多视图特征学习方法,即半监督双重视图特征学习(Semi-supervised Dual-view Feature Learning,SDvFL),该方法可以让有标记、相同视图的异类样本之间互相远离,同时无标记、相同视图的近邻外样本之间也要互相远离;有标记、不同视图的同类样本互相靠近,同时无标记、不同视图的近邻内样本之间也要互相靠近。通过这种方式,SDvFL能够有效地挖掘多视图数据中的信息。(3)在(2)的基础上,研究半监督情形下不同视图之间的相关性,为了挖掘不同视图学习得到的投影矩阵之间的关联,引入了视图一致性的概念,提出了半监督双重视图一致性特征学习方法(Semi-supervised Dual-view Consistency Feature Learning,SDvCFL)。SDvCFL方法考虑多视图中的样本特征描述的是同一个对象不同方面的特性,那么不同视图特征学习的投影矩阵之间应该有一定的联系,因此不同视图的结构信息都是类似的,可以考虑让实际求解得到的不同视图之间结构信息的差异最小化,本文中称之为“视图一致性”,即通过视图一致性来进一步约束原始样本结构信息的差异性。(4)稀疏表示及字典学习技术在模式识别领域已经取得广泛关注,本文在传统单视图字典学习的基础上提出一种针对于多视图数据的鉴别字典学习方法(Multi-view Discriminant Dictionary Learning,MDDL),MDDL模型能够学习出结构化的鉴别字典,该字典具有三点优势:其一,同类样本能够使用同类同视图的字典进行逼近;其二,某一类样本由不同类所有视图的字典表示残差较大;其三,引入重构系数鉴别项进一步加强字典的鉴别能力。(5)在(4)的基础上,进一步分析稀疏重构系数的性质,在有监督的情况下考虑重新定义系数鉴别项,新的鉴别项能够使有标记、相同视图的异类重构系数之间互相远离,同时无标记、相同视图的近邻外重构系数之间也要互相远离;有标记、不同视图的同类重构系数互相靠近,同时无标记、不同视图的近邻内重构系数之间也要互相靠近,基于此提出了近邻多视图鉴别字典学习方法(Neighbour Multi-view Discriminant Dictionary Learning,NMDDL)。NMDDL方法在保证字典近邻关系的基础上进一步提升字典的鉴别性,最终能够有助于改善分类效果。
[Abstract]:Pattern recognition, machine learning and other cross disciplinary rules need to find from the observed data. In recent years, the Internet, communications and information technology has revolutionized the development of the rapid development of information technology and the amount of data that generated in today's society the growth, there are a lot of data can be performed in many different forms said. For example, in the Internet, each Web page can be expressed as contained in the documents and hyperlinks pointing to it; in the face recognition, feature extraction can form different forms of the same sample face images, such as Gabor features, HOG features, LBP features, PCA features are used to describe the orientation of face scale feature, edge contour feature, local pixel gray change characteristic and the whole main feature of information. The traditional analysis algorithm based on single view, using only the structural characteristics within a single view, no The correlation between views, complementary information, multi view learning method to extract the correlation between different views, complementary characteristics, which can improve the data set on the learning or classification effect. So in recent years, multi view feature learning in machine learning, data mining and computer vision get extensive attention. The classification method of multi view data as the theme, to identify the feature extraction in the data as the focus, from the spatial feature extraction, semi supervised manifold learning and differential dictionary learning in three aspects, do some innovative work, the main contents include: (1) in a typical correlation analysis (Correlation Canonical, Analysis, CCA) based on canonical correlation analysis were identified (Discriminant Correlation Canonical Analysis, DCCA), Multi-v (differential analysis of multi view Iew Discriminant Analysis, MvDA), enhance the combined characteristics of differential correlation (Combined-Feature-Discriminability Enhanced Canonical Correlation Analysis CECCA) algorithm research and analysis, put forward analysis of double differential correlation (Dual Discriminant Correlation Analysis, DDCA) model.DDCA algorithm design has two advantages: first, to within each view by Fisher discriminant analysis (Fisher Discriminant Analysis, FDA) for projection vector to ensure sample separability; second, to analyze the correlation between the sample identification in the view, namely for projection vector so that the sample within class correlation maximum between class correlation, the minimum.DDCA is an extraction method of supervised feature, compared to CCA can effectively use the sample in addition, the traditional information label; correlation analysis method due to its limited model System applies only to between two views, while ignoring the views of internal information, and between DDCA within the same view and different views are able to analyze the data, based on the above points, DDCA helps to improve the classification results. (2) in the semi supervised setting, multiple view feature extraction each sample is helpful to mining the samples more information, at present, scholars and researchers have proposed some effective semi supervised multi view learning methods. Although the existing semi supervised multi view feature learning method has achieved certain effect, but at the same time, these methods are not good considering the identification information intra view and inter view, and how to effectively extract the local structure information of unlabeled samples, but also to be improved. This paper presents a new semi supervised multi view feature learning method, namely semi supervision Du dual view features (Semi-supervised Dual-view Feature Learning study, SDvFL), the method can make a mark, the same view of the heterogeneous sample away from each other, and no mark, also want to move away from each other between the same view neighbor samples; marked, not with the view of similar samples close to each other, and have no marks, close to each other between different views within the nearest neighbor sample. In this way, SDvFL can effectively mine the multi view data. (3) in (2) on the basis of the correlation between different views of the semi supervised case, in order to link between the projection matrix obtained by different mining view learning, introduces the concept of view consistency, this paper presents a semi supervised feature dual view consistency learning method (Semi-supervised Dual-view Consistency Feature Learning, SDvCFL).SDvCFL method considering multi view in Sample characteristics described are the same object characteristics in different aspects, there should be some relationship between the projection matrix so different views of learning, so the structure information of different views are similar, can consider the differences between different views make the actual obtained structure information minimization, this paper called "view consistency, difference through the view consistency to further constrain the original sample structure. (4) sparse representation and dictionary learning technology has been widely concerned in the field of pattern recognition, based on the traditional visual chart based on dictionary learning is proposed for a differential multi view data dictionary learning method (Multi-view Discriminant Dictionary Learning, MDDL), MDDL model can learn to identify the structured dictionary, the dictionary has three advantages: first, the same sample can use the same With the view of the dictionary is approaching; second, a class of samples by different views all dictionary representation of the residuals greatly; third, the reconstruction coefficient identification to further strengthen the ability to identify the dictionary. (5) in (4) on the basis of further analysis of the sparse coefficients, in the supervised case consider the re definition of coefficient identification, identification of new can be marked, the same view between heterogeneous coefficients away from each other, and no mark, also away from each other between the neighbor reconstruction coefficients of the same view; marked, similar reconstruction coefficients of different views towards each other, and no mark, to close to each other between different views within the nearest neighbor reconstruction coefficient based on the nearest neighbor multi view differential dictionary learning method is proposed (Neighbour Multi-view Discriminant Dictionary Learning, NMDDL.NMDDL) method in order to ensure close neighbor dictionary On the basis of the system, the differentiation of the dictionary can be further promoted, and the classification effect can be improved in the end.

【学位授予单位】：中国矿业大学(北京)
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】