Overlapping classes and outliers can signicantly decrease a classier performance. We adress here the problem of giving a classier the ability to reject some patterns either for ambiguity or for distance in order to improve its performance. Given a set of typicality degrees for a pattern to be classied, we use an operator based on triangular norms and a discrete Sugeno integral to quantify their blockwise similarities. We propose a new class-selective rejection scheme which uses this operator outputs. We present the resulting algorithm which allows to assign a pattern to zero, one or several classes, and show its eciency on real data sets.