On Multi-view Subspace Learning And Its Application In Image Classification

In many real-world pattern recognition,computer vision and data visualization applications,the same object is usually represented by multiple high dimensional feature data,namely multi-view data.Multi-view data tends to contain more complementary information than that of single view data,which is more conducive to classification.Therefore,in recent years,multi-view learning technology has attracted a great deal of attention.In this paper,we focus on researching methods about multi-view subspace learning,and the main research contents as follows:(1)A novel multi-view local discriminant approach is proposed,namely random subspace based multi-view local discriminant projection(RSMLDP).On the one hand,RSMLDP solves the problem that the constructing nearest neighborhood graphs by using the original high dimensional data directly are easy to be affected by noise and thus cause the instability of classification performance.On the other hand,by utilizing the label information and local information from multi-view data,RSMLDP explores the discriminant features and intrinsic local geometrical structure information from multiple views.First,by employing random subspace technology,RSMLDP performs random feature selection on the original data.After many random feature selection procedures,RSMLDP fuses these feature subsets,and then constructs nearest neighborhood graphs on the fused low dimensional feature space.Then,RSMLDP learns discriminative transformations for each view,so that in the learned common subspace,neighborhood samples from same class are close to each other,while neighbor samples from different classes are as far as possible.Meanwhile,the global local structure information of samples is well preserved.Experimental results on MNIST digit image dataset,COIL-20 object dataset,Multi-PIE face dataset and large-scale Caltech-101 object dataset show the effectiveness of the proposed RSMLDP approach.(2)A novel multi-view discriminant and correlation analysis approach,namely multi-view local discriminant and canonical correlation analysis(MLDC~2A).MLDC~2A mainly concerns two problems.One is how to extract more comprehensive discriminant information and correlation information,and the other is how to utilize the local geometrical structure information from multi-view data to enhance the classification performances.To achieve these goals,MLDC~2A jointly learns a shared subspace for multi-view data.In the learned shared subspace,the margin of different classes from both intra-view and inter-views are maximized,while the distances of samples in the same class from both intra-view and inter-views are minimized.Meanwhile,the sum of pair correlations of transformed features from different views is maximized.By this way,MLDC~2A can exploit the discriminant features,canonical correlation features and local geometrical structure information from both intra-view and inter-views.The results on the four datasets demonstrate that the designed items can improve the classification effectively.(3)Two multi-view latent intact space learning approaches are proposed,including multi-view intact space and local correlation analysis(MISLC)and semi-supervised multi-view manifold discriminant intact space learning(SM~2DIS).MISLC is an unsupervised latent intact space learning approach,which aims to learn the latent intact feature representation for each original data of different views,and reveals the local relevance of all the feature representations in the learned latent intact space.In detail,if two different original data are in the same manifold,we hope that the linear correlation between the learned corresponding latent intact feature representations can be well revealed.To achieve this goal,first,MISLC designs local correlation item where each original data of multiple views is considered as a sample set and then constructing the neighbor graph based on different sample sets.Second,MISLC explores the correlation information by maximizing the correlations between neighborhood feature representations from the learned latent intact space.Meanwhile,by minimizing reconstruction error between each view data and its corresponding generated multi-view data point,MISLC can well represent the corresponding multiple views data in the original space.In many real-world applications,although it is difficult to get the label information of all samples,employing the information of partially labeled samples effectively can enhance the discriminant power of the learning model.SM~2DIS mainly concerns how to make full use of the label information of partially labeled samples and the information of unlabeled samples.SM~2DIS approach is a semi-supervised latent intact space learning model.It designs a semi-supervised multi-view manifold discriminant item,which can not only learn discriminant intact feature representations for the original data of different views,but also can extract the local geometric structure information of the data in the latent intact space.The learned discriminant information and local geometric structure information can well enhance the classification performance.To achieve this goal,SM~2DIS makes the generated multi-view data points from intra-class gather together,and simultaneously maximizes the separabity of the generated inter-class multi-view data points.SM~2DIS uses all of the view generation matrices to reconstruct the original multi-view data space,so each learned feature representation is intact.In addition,SM~2DIS uses all sample points in the original space(including labelled and unlabelled)to construct neighbor graphs to extract the manifold structure information in the learned latent intact space.The experiment on the four datasets show the the proposed MISLC approach can achieve a better classification performance than compared methods.Meanwhile,compared with other semi-supervised potential space learning methods,SM~2DIS approch can gain better classification and can be well applied to image classification in semi-supervised scenes.