To findout how many clusters exist in a sample set is an old yet unsolved problem in unsupervised clustering. This problem inevitably occurs in region merging/growing, a well studied and popular technique in image segmentation. Region merging usually needs a stop criterion. The stop criterion is not automatically determined and often has to be set manually to arrive at a sensible segmentation, which is rather difficult for natural images. To address this problem, we present a robust stop criterion that is based on a novel distinctness predicate for adjacent regions. The predicate discerns distinct regions by examining the evidence of the boundary between neighboring regions. Requiring that every region should be distinct from each other, the proposed method is able to choose a stop point where a natural partition is most likely. Under a region merging framework, we demonstrate the effectiveness of the stop criterion using two merging criterion: one based on optimizing a global functional, and another based on a local criterion. Experimental results and comparison are given at the end.