Content-based systems retrieve images based on low-level features (color,texture) while the user usually seeks some objects from real world. As segmentation is never accurate, such systems do not allow the use of powerful feature during retrieval (shape, structure). We propose a new system that relies on a hierarchy of segmentations, in order to handle some artifacts. Besides, it allows using some object-related features for indexing (shape, structure). We also present some result on a 600 images database from Corel.