Abstract :

We expose a recently introduced method for computing geometric structures in a digital image, without any a priori information. According to a basic principle of perception due to Helmholtz, an observed geometric structure is perceptually ``meaningful'' if its number of occurences would be very small in a random situation : geometric structures are characterized as large deviations from randomness. This leads us to define and compute ``partial gestalts'' (in computer vision, features) like alignments, edges, clusters, groups of objects similar for some quality in an image by a parameter-free method. Maximal meaningful objects are defined, computed, and the results compared with the ones obtained by classical algorithms. A discussion ensues : are the partial gestalt (or feature) detectors enough to build up computer vision algorithms ? We show by experiments that this is rather an illusion : the ``conflicts'' between gestalt laws that were discussed in a phenomenological framework by the gestaltists indeed arise in well chosen images and lead to wrong (but explainable) detections. We show how no ``good'' feature detector can work if not applied simultaneously and in conflict with all other detectors. We are led to the conclusion that no correct image analysis can be obtained if the gestalt conflicts and the subsequent masking phenomena are not adressed.