A sequential scheduling approach to combining multiple object classifiers using cross-entropy

A method for multiple classifier selection and combination is presented. Classifiers are selected sequentially on-line based on a context specific (data driven) formulation of classifier optimality. A finite subset of a large (or infinite) set of classifiers is used for classification resulting not only in a computational saving, but a boost in classification performance. Experiments were carried out using single class binary classifiers on multi-class classification problems. Classifier outputs are combined using a Bayesian approach and results show a significant improvement in classification accuracy over the AdaBoost.MH method.