Canonical correlation analysis based on local sparse representation and linear discriminative analysis

The natural discriminating information contained in the data structure and class information of the datasets is very vital for the feature fusion. Then in order to utilize all the information, a canonical correlation analysis algorithm based on local sparse representation and linear discriminative analysis is proposed. Firstly, the local sparse representation method is utilized to obtain the sparse manifold reconstruction matrix with less computational complexity. Then, the united optimization is realized in the canonical correlation analysis scheme to constrain the sparse reconstructive relationship among each feature set with optimizing the combined discriminability and the feature correlation simultaneously, so that the discrimination capability of the feature extracted is increased. Finally, the simulation examples on artificial dataset, multiple feature database and facial databases are presented, and the experimental results show the effectiveness of the proposed method.