A new approach to combining outputs of multiple classifiers

This paper presents a novel method for multiple classifier fusion. The classifier combiner operates on the single classifier outputs, which consist of vectors of pairs (c, d), with c being a class name and d the confidence degree with which a pattern is recognized as belonging to class c. The main idea of the combiner is to exploit the knowledge of the statistical behavior of the single classifiers on the training set to re-calculate a global recognition confidence degree based on the a posteriori probability that the input pattern belongs to a given class conditioned by the specific responses of the classifiers. Applying the Bayes´s theorem we can also easily adapt our classifier combiner to a specific application. We compare our model with some popular techniques for classifier fusion on the Satimage and Phoneme data sets from. the database ELENA.. We show that our method is in most cases superior (or substantially equivalent) to the other techniques on both data sets.