Multi-view Dimensionality Reduction And Its Application In Hyperspectral Image Classification

With the development of information technology,the same object can be described by multi-feature or multi-source data,which is called multi-view data.Multi-view data can reflect different characteristics of object.However,in many practical applications,multiview data has high dimension.On the one hand,directly using multi-view data will lead to high storage and transmission costs as well as low data mining efficiency.On the other hand,computation complexity will grow exponentially as dimension increases,which will cause the “curse of dimensionality”problem.Dimensionality reduction of multi-view data has become one of the hot research directions in machine learning field.Existing multi-view dimensionality reduction methods have the following shortcomings:(1)Most methods focus on learning a low-dimensional common space,which can't contain common information and complementary information of multi-view data simultaneously;(2)Existing methods do not realise discriminative locality alignment on multi-view data and the discriminant of the low-dimensional space needs to be improved;(3)In practice,the problem of insufficient labeled samples is existed.It is necessary to establish multi-view dimensionality reduction methods which are more suitable for practical applications.In order to solve the above problems,this thesis focuses on multi-view dimensionality reduction and its application in hyperspectral image classification based on recently proposed multi-view subspace learning and dimensionality reduction methods.The main research contents and works of this thesis are listed as below:1.Aiming to solve the problem of maintaining both common information and complementary information of multi-view data in low-dimensional space,a local neighbor alignment based unsupervised multi-view dimensionality reduction method is proposed.The low-dimensional space is treated as the source space of multi-view data to learn a lowdimensional sufficient space which maintain the common information and complementary information simultaneously.Secondly,as different view has different neighboring relationship,the common neighbors or complementary neighbors of multi-view data are maintained in the low-dimensional space.The discriminant of the low-dimensional space is improved by local neighbor alignment.Experiments on “Swiss Roll”,objects,indoor scenes and face images indicate the effectiveness of the proposed method.2.Aiming to solve the problem of improving the discriminant of the low-dimensional space,a supervised multi-view dimensionality reduction method is proposed by combining the low-dimensional sufficient space with discriminative locality alignment.The discriminant of the multi-view sufficient space is enhanced by constraining the local within-class neighbors to locate as close as possible and the between-class neighbors to locate as far as possible.Experiments on multiple databases demonstrate that the proposed method obtains good classification results with a small number of local neighbors.3.Aiming to solve the problem of dimensionality reduction with insufficient labeled samples,a transductive learning based semi-supervised multi-view dimensionality reduction method is proposed as transductive learning can use the information of unlabeled samples.Furthermore,the discriminant of the low-dimensional space is improved by constraining labeled samples to locate around their class centers.In particular,hyperspectral image is one kind of typical high-dimensional data and it is very difficult to acquire its label information.Hyperspectral image classification with insufficient labeled samples is a problem that needs to be further solved.In this paper,transductive learning based semi-supervised multi-view dimensionality reduction method is applied in hyperspectral image classification.Classification experiments on different hyperspectral image databases demonstrate the effectiveness of the proposed method.4.Different from transductive learning that can only predict unlabeled samples,(pure)semi-supervised learning can predict new samples.Another method is proposed to solve the problem of insufficient labeled samples by using(pure)semi-supervised learning.In the new method,spatial information is utilized to learn the neighbor relationship as well as choosing unlabeled samples.Hyperspectral image is one kind of high-dimensional data,spatial information is very important in its classification.The proposed method is applied to hyperspectral image classification.Experiments on different databases show the effectiveness and robustness of the proposed method.