Geometrical Primitives for the Classification of Images Containing Structural Cartographic Objects

The new generation satellites provide an important quantity of high resolution images. There is a growing need for automated systems to detect cartographic objects on these images. A sub-task of this objective, proposed by CNES (French National Space Agency), is to classify images containing different categories of cartographic objects. The main challenge of the problem, compared with other object detection tasks, is the high intra-class variability of the target classes due to the vagueness of their definitions. We propose a classification system that exploits mainly the structural properties of the images that seem to be the most plausible cues for discriminating them. A region-based method and an edge-based method are used in parallel to extract the geometrical primitives in images. A feature vector is calculated from the primitives and their perceptual groupings, by the accumulation of combinations of their geometrical, relational and spatial attributes. A multi-class Adaboost classifier is trained using the feature vector. The main contribution of this paper is the use of structural shape attributes in a statistical learning method framework. We tested our method on CNES dataset prepared for the ROBIN competition and we obtained promising results.