A Method for Rapid Feature Extraction Based on KMSE

Kernel minimum squared error (KMSE) is a mathematically tractable feature extraction method in comparison with other nonlinear methods. However, as a kernel method, it also suffers from the drawback that the classification efficiency decreases as the size of the training samples increases. In order to improve the KMSE-based classification efficiency, we propose to approximate the discriminant vector in the feature space using a certain linear combination of some samples selected from the set of the training samples. Based on this idea, we develop an algorithm to determine the samples, and a linear combination of which can approximately express the genuine discriminant vector. Experiments on benchmark dataset illustrate the effectiveness of the improvements.