Support vector number reduction by extending iterative preimage addition using genetic algorithm-based preimage estimation

As the support vector (SV) number of a support vector machine (SVM) determines the execution speed of the testing phase, there have been diverse methods to reduce it. Although iterative preimage addition (IPA), belonging to the 'reduced set construction', is reported to be able to reduce a large portion of the SV number of a standard SVM when the kernel is a radial basis function (RBF), the fact that it cannot be applied to other types of kernels is a significant drawback. To address this problem, this paper proposes a novel genetic algorithm-based preimage estimation method and incorporates it into a conventional IPA such that all types of kernels can be handled. Experimental results indicate that the proposed method shows the equivalent performance to the conventional IPA when an RBF kernel is used. Furthermore, they show that the proposed method can reduce the SV number of the histogram of oriented gradient (HOG) feature-based pedestrian classifier using linear, quadratic, and sigmoid kernels by more than 99.5%.