A method for error rejection in multiple classifier systems

In the literature, the introduction of the reject option in multiple classifier systems has been analysed only from the experimental point of view. Following a first theoretical analysis provided by the authors, we analyse, within the framework of the minimum risk theory, the problem of finding the best error-reject trade-off achievable by a linear combination of a given set of trained classifiers. An algorithm for computing the parameters of the linear combination and of the reject rule is then proposed. Experimental results on two data sets of remote-sensing images are reported.