Research On Cross-view Classification Algorithms Based On Multi-view Subspace Learning

With the rapid development of information acquisition technology,we have the ability to capture objects from different views or with different sensors.Consequently,an object usually has multiple-view representations,and it is often referred to as multi-view data.Despite more information provided by multi-view data,it also poses the cross-view classification problem.Cross-view classification intends to classify objects when gallery and probe data come from different views.Although great progress has been made in theory research for this problem in recent years,there still remain difficulties and challenges.Existing cross-view classification algorithms are incapable of handling widely used nonlinear distributed data due to the large view discrepancy.Hence,this paper proposes two cross-view classification approaches to effectively handle nonlinearity,thus boosting the classification performance.This paper proposes a Multi-view Hybrid Embedding(MvHE)algorithm based on multi-view subspace learning(MvSL)considering the difficulty in handling discrepancy and nonlinerity simultaneously.To handle them separately,this paper divides cross-view classification problem into three subproblems with the divide-and-conquer strategy and builds each model for each problem.Compared with the state-of-the-art methods,the accuracy of our MvHE is improved obviously on four public cross-view classification datasets.Moreover,experimental results on noise data indicate the robustness of MvHE to outliers.We also present a Multi-view Common Component Discriminant Analysis(MvCCDA)algorithm based on MvSL considering that the samples in a subspace are difficult to preserve the local geometry of observed data.To make common components preserve the local discriminant geometric structure,this paper incorporates supervised and local geometric information into the process of common component learning with discriminant and local consistency regularizations.Our MvCCDA achieves a promising performance on four benchmark datasets.The comparison with the classical methods demonstrates the effectiveness of MvCCDA.