Speeding Up SVM in test phase: Application to radar HRRP ATR

Abstract

In this paper, a simple method is proposed to reduce the number of support vectors (SVs) in the decision function. Because in practice the embedded data just lie into a subspace of the kernel-induced space, F, we can search a set of basis vectors (BVs) to express all the SVs according to the geometrical structure, the number of which is less than that of SVs. The experimental results show that our method can reduce the run-time complexity in SVM with the preservation of machine's generalization, especially for the data of large correlation coefficients among input samples. © Springer-Verlag Berlin Heidelberg 2006.