A third way to design pattern classifiers with reject options

Abstract

The design of a rejection-based classifier can be made according to two well-identified strategies operating in two sequential steps: the accept-first strategy and the reject-first one. The first one is the most usual. Recently, we have proposed a general class of the latter classifiers using fuzzy XOR operators based on dual triples (t-norm, t-conorm, complement) (2001). In this paper, we investigate a new approach. It consists in starting with testing for ambiguity rejection, and if needed, testing for either exclusive classification or distance rejection. For this purpose, we define a new operator called the fuzzy OR-2 allowing us to propose a new class of classifiers.