This work is motivated by the control of industrial crystallisation processes, which involve the knowledge of the crystals distribution over time. For this purpose, an in situ camera in batch crystallisers provide 2D images of the projected crystals population in real time such as Figure 1 (a).In order to characterize the geometry of the overlapped crystals, advanced image processing [1, 2] or stochastic geometry [3] can be used. The proposed approach in this study is based on community detection (as done in social networks [4]). One can see on the images that the borders of individual crystals is quite visible and such a data Figure 1(b) looks like a tessellation of the union of crystals. In-deed, for instance in Figure 1 (b-c), four polygonal crystals are recognizable from a human visual perception (Gestalt theory [5]). It involves the use of topological and geometrical relationship between tessellation cells to make the decision to put the cells in the same crystal or not. Our objective is to provide an automatic detection of the crystals based on such ”community membership” rules.