Nonlinear Boost

In this paper, we propose the post-classification scheme that is useful for improving weak-hypothesis combination of AdaBoost. The post-classification scheme allows the weak hypotheses to be combined nonlinearly, and can be shown to have a generally better performance than the original linear-combination approach in either theory or practice. The post-classification scheme provides a general perspective on combining weak hypotheses, in which many existing boosting methods can be treated as its sub-cases. By using support vector machine (SVM) as post-classifier, we have shown that AdaBoost and SVM can be effectively combined in the post-classification scheme to achieve an effective classification algorithm. Experiment results show that the proposed nonlinear boost approach can significantly improve the performance of AdaBoost.