This paper presents a novel approach to the detection and recognition of qualitative parts like geons from real 2D intensity images. Previous works relied on semi-local properties of either line drawings or good region segmentation. Here, in the framework of Model-Based Optimisation, whole geons or substantial sub-parts are recognised by fitting parametric deformable contour models to the edge image by means of a Maximum A Posteriori estimation performed by Adaptive Simulated Annealing, accounting for image clutter and limited occlusions. A number of experiments, carried out both on synthetic and real edge images, are presented.