Efficient kernel discriminative common vectors for classification

摘要

Kernel discriminant analysis (KDA) which operates in the reproducing kernel Hilbert space (RKHS) is a very popular approach to dimensionality reduction. Kernel discriminative common vectors (KDCV) shares the same modified Fisher linear discriminant criterion with KDA and guarantees a 100 % recognition rate for the training set samples as well as favorable generalization performance. However, KDCV has the disadvantage of high computational complexity in both the training and the testing stage. This paper attempts to improve the computation efficiency of KDCV by two strategies. First, the Cholesky decomposition is introduced to obtain the projection matrix instead of eigen-decomposition. Second, we replace the matrix operation with vector operation in the testing process which reduces the computational complexity. Extensive experiments on COIL images dataset, ORL faces dataset, PIE faces dataset, and USPS handwritten digits dataset demonstrate that the proposed algorithm is more efficient than the traditional KDCV algorithm without loss of accuracy.