In many classification problems, overlapping classes and outliers can significantly decrease a classifier performance. In this paper, we introduce the possibility of a given classifier to reject patterns either for ambiguity or for distance. From a set of typicality degrees for a pattern to be classified, we propose to use fuzzy implications to quantify the similarity of the degrees. A class-selective scheme based on this new family is presented, and experimental results showing the efficiency of the proposed algorithm are given.