Segmenting multi-component microanalytical images consists in trying to find zones of the specimen with approximate homogeneous composition, representing different chemical phases. This can be done through pixel clustering. We first highlight some limitations of classical clustering algorithms (C-means and fuzzy C-means). Then, we describe a new algorithm we have contributed to develop: the Parzen-watersheds algorithm. This algorithm is based on the estimation of the probability density function of the whole data set in the feature space (through the Parzen approach) and its partitioning using a method inherited from mathematical morphology: the watersheds method. Next, we introduce a fuzzy version of this approach, where the pixels are characterized by their grades of membership to the different classes. Finally, we show how the definition of the grades of membership can be used to improve the results of clustering, through probabilistic relaxation in the image space. The different methods presented are illustrated through an example in the field of electron energy loss mapping, where four elemental maps are concentrated in a single chemical phase map.