Multi-view Sparse Embedding Analysis Based Recognition

With the rapid development of information technology, multi-view data has commonly appeared in our daily life. On the one hand, it enables us to reveal the inherent property of pattern better by using the various data. On the other hand, it brings challenge to those conventional methods designed for single view data. How to fully explore the potential shared information in multi-view data, and utilize the data for recognition or classification tasks, is a hot research topic in the machine learning domain.Firstly, motivated by canonical correlation analysis, on the foundation of Sparse Embedding Analysis, we propose a novel approach named Multi-view Sparse Embedding Analysis(MvSEA). MvSEA not only explores the potential shared information that hides in multi-view data, but also considers the discriminative correlation information by maximizing the within-class correlation and simultaneously minimizing the between-class correlation from intra-view. The math model can be solved though iterative optimization algorithm provided by MvSEA.Secondly, we add the orthogonal contraints of transformation matrix, and thus Enhandced Multi-view Sparse Embedding Analysis(EMvSEA) is proposed. On the one hand, it preserves the reconstruction relationship in data. On the other hand, it can achieve the goal of reducing redundant information. With the aid of SVD, the math model can be solved though iterative optimization algorithm provided by EMvSEA, which enhances the performance of EMvSEA not at the cost of running time.Finally, in order to address the inseparable problem in samples, we extend it to the nonlinear space, and propose the Multi-view Kernel Sparse Embedding Analysis(EMvSEA). EMvSEA can explore the nonlinear structure of multi-view data to further enhance the performance of MvKSEA. The math model can be solved though iterative optimization algorithm provided by EMvSEA, which owns favaroble computation efficiency.The experimental results on the MFD, Multi-PIE and PolyU databases demonstrate that the proposed approaches can effectively solve the multi-view classification problem. Compared with the popular mulit-view classification methods, it improves the recognition performance.