Multi-view Discriminant Analysis Based Feature Extraction

Today, the information technology develop increasingly. We can acquire more and more multi-view data for patten recognition and machine learning research. Compared to the traditional single-view feature extract algorithm, a multiview algorithm can better showcase the essence of a sample. How to explore the information within the view, and find the potential relationship between the views to aid in classification and idenfication has become the hotspot of machine learning.Firstly, foucusing on the issue that traditional mutil-view algorithm don't effectively utilize the discriminant information in the projection vector. We proposed uncorrelated constrains multi-view discriminant analysis. Reconstruct the between-class scatter matrices and within-class scatter matrices to form the linear discriminant analysis. Then, project all the multi-view samples to a common subspace which the same class samples are clustered, and different classes samples are scattered.Secondly, on the basis of multi-view discriminant analysis, we introduced the thought of semi-supervised learning. General multi-view discriminant analysis is supervised and therefore can't directly use the data without label. In this paper, we proposed Semi-supervised Multi-view discriminant analysis, which can utilize both of data with and withion label. The data with label can be used to calculate the between-class scatter matrices, at the same time, the data without the label can be used to form the nearest-neighbor structure.Finally, take into accont the original non-linear and uneven-distributioned data, we proposed the weighted fisher criterion kernel MvDA.The experiment results on the Multi-feature database, AR color face database and PolyU palmprint database indicate that our proposed method can better improve the multi-view discrimination problem, compared with the famous mulit-view classification methods.