About Segmentation Step in Content-based Image Retrieval Systems

Despite of the hope arised a few years ago, Content Based Image Retrieval - CBIR - systems has not reached the initial goal, ie to manage and search images in database: we are unable to link the semantic sens of an image to numerical values. However, some members of the community have begun the necessary introspection. The analyze of each step of the feature extraction will allow us to overcome actual problematics and to take the right path in the future. In this context, we propose in this paper to discuss about a low-level tool frequently used: the segmentation step. In the general context of scene images, we evaluate the stability of some classical algorithms using a basic protocol. The quite inefficiency of all approaches let us conclude to the necessity to use meta-data and any other collected informations during this first segmentation step.