Linear boundary discriminant analysis

In this paper, we propose a new discriminant analysis, called linear boundary discriminant analysis (LBDA), which increases class separability by reflecting the different significances of non-boundary and boundary patterns. This is achieved by defining two novel scatter matrices and solving the eigenproblem on the criterion described by these scatter matrices. As a result, the classification performance using the extracted features can be improved. This effectiveness of the LBDA is theoretically explained by reformulating the scatter matrices in pairwise form. Experiments are conducted to show the performance of LBDA, and the results show that LBDA can perform better than other algorithms in most cases.