Comparison between Constrained Mutual Subspace Method and Orthogonal Mutual Subspace Method – From the viewpoint of orthogonalization of subspaces –

摘要

This paper compares the performances between constrained mutual subspace method (CMSM), orthogonalmutual subspace method (OMSM), and also between their nonlinear extensions, namely kernel CMSM(KCMSM) and kernel OMSM (KOMSM). Although the princeples of the feature extraction in these methods aredifferent, their effectiveness are commonly derived from the orthogonalization of subspace, which is widely used tomeasure the performance of subspace-based methods. CMSM makes the relation between class subspaces similarto orthogonal relation by projecting the class subspaces onto the generalized difference subspaces. KCMSM is alsobased on this projection in the nonlinear feature space. On the other hand, OMSM orthogonalizes class subspacesdirectly by whitening the distribution of the class subspaces. KOMSM also utilizes this orthogonalization method inthe nonlinear feature space. From the experimental results, the performances of both the kernel methods (KCMSMand KOMSM) are found to be very high as compared to their linear methods (CMSM and OMSM) and theirperformances levels are well in the same order in spite of their different principles of orthogonalization.