2 - Utilisation de la transformée de Fourier-Mellin pour la reconnaissance de formes multi-orientées et multi-échelles : application à l'analyse automatique de documents techniques

In this paper, we propose an original methodology which allows the detection and the recognition of multi-oriented and multiscaled patterns. The supports on which the method is applied are technical documents representing the network of the French Telephone operator France Telecom. The adopted technique, based on the Fourier-Mellin Transform (FMT) is integrated in a global strategy that solves ambiguous situations, through the providing of contextual information. The strategy which is applied to solve the character and symbol classification problem can be divided into two stages. The first one consists in computing a set of invariant descriptors for each isolated pattern belonging to a characters layer detected thanks to a connected components extractor. The second stage, based on a filtering scheme, consists in detecting and recognising the shapes which are either interconnected or connected to any other object. The results of the application of this technique are very encouraging since the classification rate reaches excellent scores in comparison with classical techniques.