A generalised K-L expansion method which can deal with small sample size and high-dimensional problems

摘要

The K-L expansion method, which is able to extract the discriminatory information contained in class-mean vectors, is generalised, in this paper, to make it suitable for solving small sample size problems. We further investigate, theoretically, how to reduce the method's computational complexity in high-dimensional cases. As a result, a simple and efficient GKLE algorithm is developed. We test our method on the ORL face image database and the NUST603 handwritten Chinese character database, and our experimental results demonstrate that GKLE outperforms the existing techniques of PCA, PCA plus LDA, and Direct LDA.