We describe in this paper a high level recognition system. The system implements a new approach to model-based object recognition, fully exploiting compositionality of representations : from the analysis of the elementary signs in the image to the analysis and description of an object structure and, finally, to the interpretation of the scene. Perceptual reasoning, likewise the symbolic description of the scene are stated in the Situation Calculus. A description is a specification of an object in terms of its single components which, in turn, are specified using SymGeons, a generalization of parametric Geons. The cognitive process of recognition relies on SymGeons recognition. Here we extend the concepts of aspect graphs and hierarchical aspect graph to obtain a Bayes network integrating composition of aspects together with composition of features.