Toward Robust and Fast Two-Dimensional Linear Discriminant Analysis
This paper presents an approach toward robust and fast Two-Dimensional Linear Discriminant Analysis (2DLDA). 2DLDA is an extension of Linear Discriminant Analysis (LDA) for 2-dimensional objects such as images. Linear transformation matrices are iteratively calculated based on the eigenvectors of asymmetric matrices in 2DLDA. However, repeated calculation of eigenvectors of asymmetric matrices may lead to unstable performance. We propose to use simultaneous diagonalization of scatter matrices so that eigenvectors can be stably calculated. Furthermore, for fast calculation, we propose to use approximate decomposition of a scatter matrix based on its several leading eigenvectors. Preliminary experiments are conducted to investigate the effectiveness of our approach. Results are encouraging, and indicate that our approach can achieve comparative performance with the original 2DLDA with reduced computation time.
KeywordsLinear Discriminant Analysis Class Number Normalize Mutual Information Scatter Matrix Label Information
Unable to display preview. Download preview PDF.
- 1.Anderson, T.W.: An Introduction to Multivariate Statistical Analysis. Wiley-Interscience (2003)Google Scholar
- 2.Ding, C., Ye, J.: Two-dimensional singular value decomposition (2dsvd) for 2d maps and images. In: Proc. of SDM 2005, pp. 32–43 (2005)Google Scholar
- 4.Harville, D.A.: Matrix Algebra From a Statistican’s Perspective. Splinger (2008)Google Scholar
- 8.Voss, J.: Tagging, folksonomy & co – renaissance of manual indexing? In: Proc. 10th International Symposium for Information Science, pp. 234–254 (2007)Google Scholar
- 10.Ye, J., Janardan, R., Li, Q.: Two-dimensional linear discriminant analysis. In: Proc. of NIPS 2004, pp. 1569–1576 (2004)Google Scholar