Analyse et représentation ensembles-échelle d'une image

We propose a multiscale approach to energy minimization-based image segmentation. Our goal is to compute a complete series of ordered partitions which approximates the minima of a functional depending on a real parameter which acts as a scale parameter. In order to capture all these solutions, we introduce the scale-sets representation, which is the region-oriented equivalent of the scale-space representation. We end up with an efficient parameter-free algorithm, called scale climbing, which produces a scale-sets description of an image whose cuts correspond to a whole family of solutions, from fine to coarse.