Support vector representation and discrimination machine

We address rejection-classification problems, which have been ignored in most prior work. For such a system, a high classification rate and a low false alarm rate are simultaneously desired. We first propose a one-class support vector representation machine (SVRM). The SVRM achieves a high test set detection rate by requiring a high training set detection rate; the SVRM reduces the false alarm rate by minimizing the upper bound of the decision region. The SVRM is then extended to a new support vector representation and discrimination machine (SVRDM) classifier to address multiple-class cases. We show that the SVRDM has a comparable discrimination performance to that of the traditional SVM, but has a much better rejection performance. We provide the theoretical basis for the superior rejection performance of our new SVRDM. We also present ways to select the parameters of the SVRDM, including new sigma parameter selection methods, which can handle complex multiple-cluster cases. We apply the SVRDM algorithm to several databases including synthetic data, the MSTAR synthetic aperture radar database, and the CMU PIE Face database with both pose and illumination variations. Both rejection and classification are considered. Comparison results of our SVRDM with other well-known classifiers such as support vector machines, distortion-invariant filters, feature space trajectory classifiers, eigenface and Fisherface classifiers are also presented. The proposed SVRDM is shown to outperform popular classifiers in various applications in terms of rejection and classification.