Random Subspace Based ECOC Classifier with Reject Option

Abstract

ECOC based multi-class classification has been a topic of research interests for many years. Yet most of the previous studies concentrated only on different coding and decoding strategies aiming at improvement over classification accuracies. In this paper, the classification reliability is addressed. By applying the Random Subspace method, a base classifier is created for each of the coding position. The improvement over classification accuracy on each of the coding position is achieved by a reject option and decision fusion. By rejection of those low-confidence samples, the systems reliability is enhanced. The performance of the proposed system was demonstrated by a vehicle classification example, showing promising results.