Clustering methods are useful tools for the unsupervised classification and analysis of the elements of a set or scene, e.g., a visual or auditory scene. Such methods can be seen as an integral part of cognition-like operations performed by artificial systems. The problematic is that usually no a priori information is available about the structure, the size or the number of classes. Therefore, unbiased methods that are able to provide a 'natural' classification are of interest. As it has been shown (Blatt, Wiseman, Domany), superparamagnetic clustering (SC) is a promising algorithm that comes close to an ideal unbiased method. SC gives the option of choosing different classes on different resolution levels. It, however, does not directly provide an intrinsic criterion for the choice of the 'most natural' levels, i.e. for finding the most natural classes.