This paper describes how the formation of particle systems can be investigated by using computer vision techniques, namely effective exact dilations and multiscale skeletonization. By assuming that all particles started to grow at the same time and had the same growth rate, a hypothesis that can be validated later, it becomes possible to infer the shape and distribution of the particles in the initial system configuration. Multiscale skeletons are obtained for each individual particle, and a suitable overall spatial scale, i.e. those best approximating the initial configuration, is determined as that immediately before the smallest particle fades. The thus obtained skeletons provide an estimate of the shape and position of the initial particles before uniform expansion. By using exact dilations, a generalized Dirichlet tessellation reconstruction of the particle system is obtained from these skeletons, and the similarity between this reconstruction and the original system can be used to validate the assumptions about the growth conditions. The proposed methodology is illustrated with respect to KCl polycrystalline thin films.