Une méthode autonome de ciblage de l'optimisation d'un système de détection d'objets par analyse de la responsabilité 1 An autonomous method of optimization targeting for an object detection system based on responsability analysis

and key words Object extraction systems performances are not homogeneous over different corpora because objects can take many different aspects within such sets. An adaptation of these systems is thus required in order to maintain equal performances over every kind of object the system may be applied on. Focusing on the issue of parameters optimization, a method has been developed to restrict optimization to parameters of operators which compose the system, responsible for the different categories of errors produced by the system. Two stages are involved in our method. The first one is dedicated to the analysis of the system performances and leads to the extraction of the different error categories already mentionned. The second one relates to the analysis of the behavior of the different operators, leading to extract a single operator responsible for each error category. Experiments have been carried out over a video text detection system. Control, optimization, autonomy principle, responsibility diagnostic, object detection, video texts. traitement du signal 2007_volume 24_numero special 5 SCATI 353 1. Ce travail a ete realise dans le cadre d'une Bourse CIFRE entre l'Institut National de l'Audiovisuel et le laboratoire du LIRIS.