Graph-based Multi-view Semi-supervised Feature Extraction And Application For Image Classification

The multi-view learning has caused wide attention of researchers in various areas of researching and applying over the recent years.Traditional supervised multi-view learning uses only a small number of labeled samples among the training samples in its learning process.On the contrary,traditional unsupervised multi-view learning utilizes a large number of unlabeled samples.Compared with both kinds of methods,how to apply semi-supervised learning to multi-view learning and use labeled samples and unlabeled samples in the training set simultaneously has become a hotspot in the field of machine learning.The semi-supervised learning theory is used to extract features with the projection matrix of each view.Thus,Multi-view Semi-supervised Feature Extraction(MvSFE)is proposed.MvSFE benefits from multi-view learning and graph-based semi-supervised learning,which effectively utilizes a large amount of unlabeled data and complementary information from different views,making extracted features more effective and more robust.Discriminant Multi-view Semi-supervised Analysis(DMvSA)is proposed,and it not only applies graph-based semi-supervised method to multi-view,but also takes discriminant information into consideration.The final projection matrix makes samples from the same class closer and samples from different classes away from each other.Therefore,it is effective to extract discriminant information in views.Thus it can improve the effectiveness of classification algorithm.With statistical uncorrelated constraints,redundant information between different views can be removed,and a new sample projection matrix can be obtained.Thus,Discriminant Multi-view Semi-supervised Statistical Uncorrelated Analysis(DMvS2UA)is proposed.Redundant information is reduced as much as possible between views after samples from different views are projected to the subspace.Thus,it can optimize the performance of the algorithm and further improve the classification effect.Experimental verification on multiple databases and comparison with mainstream multi-view feature recognition methods show that proposed methods can effectively extract multi-view features,and have better recognition performance.