An automated segmentation scheme for urban scenes characterization on SPOT images

Proposes an adaptive and automated segmentation scheme to be applied on SPOT images describing urban scenes. Our algorithm is intended to provide segmented images preserving subtle details (i.e. streets) while showing a low incidence of isolated pixels and well-defined edges. The proposed method performs the segmentation task in three main stages: (a) a non-contextual segmentation stage by means of a maximum-likelihood classifier, based on a Bayesian model, (b) a segmentation stage, based on geometric properties for street detection, and (c) a contextual segmentation stage, based on Markov random fields theory. Results provided by the proposed segmentation scheme show a good estimation of streets and edges, as well as a low incidence of isolated pixels, resulting in a segmented image showing homogeneous regions and preservation of subtle details.