Multiple Reject Thresholds for Improving Classification Reliability

In pattern recognition systems, Chow's rule is commonly used to reach a trade-off between error and reject probabilities. In this paper, we investigate the effects of estimate errors affecting the a posteriori probabilities on the optimality of Chow's rule. We show that the optimal error-reject trade-off is not provided by Chow's rule if the a posteriori probabilities are affected by errors. The use of multiple reject thresholds related to the data classes is then proposed. The authors have proved in another work that the reject rule based on such thresholds provides a better error-reject trade-off than in Chow's rule. Reported results on the classification of multisensor remote-sensing images point out the advantages of the proposed reject rule.