Comparaison de deux familles complètes de descripteurs de formes pour l’indexation de bases d’objets 2D à niveaux de gris

Abstract

In this work, we propose the use of two sets of 2D shape descriptors for grey-level object content-based retrieval and image indexing. These families are invariant under planar similarities and complete. The completeness property ensures that an object is identified in a unique way up to a similarity transformation and guarantees a perfect discrimination between shapes. This reveals to be interesting when image databases are either homogeneous or large. The first set of invariant descriptors is extracted from the analytical Fourier-Mellin transform, while the second one comes from the complex moment image representation. The similarity between shapes is estimated according to the distances induced by the two invariant descriptions. We tested and compared these families on two real grey-level object databases for content-based image retrieval. Retrieval results show that the invariant descriptors produce a reliable shape description, discriminating and stable to small shape variations. The completeness is an interesting feature that offers an extensive and flexible representation of shapes.