Reconnaissance et extraction de documents. Une application industrielle à la détection de documents semi-structurés

This article deals with the problem of recognition of semi-structured documents image. The aim is to detect a document and to extract the region of interest containing it. Initially, an exemple of document is given by the user and a set of interest points are extracted from this query image. In a second step, a set of interest points is extracted from each image to analyse and is matched with the set of the query image. This matching is used to calculate the geometric transformation (translation, rotation, zoom) allowing the registration between the query image and the analysed image. Two main proposals have been made to make this technique usable for documents image matching : the selection of interest points and the adaptation of RANSAC. MOTS-CLÉS : comparaison d’images de documents, points d’intérêt, FLANN, SURF, RANSAC.