Classification et extraction des documents complexes à partir des images issues d'un périphérique mobile : application aux documents d'identité

We propose in this paper a document image classification method. In contrary to most of existing systems, the proposed approach allow locating the document and recognizing its type simultaneously. First, a knowledge base of document models is created from reference images. Training images are not indispensable and though only one reference image is enough to create a document model. Then, key-points are extracted from the request image and matched against the models in the knowledge base. Once the document model is found, a more accurate matching is applied in order to extract the document from the image. The matching process is optimized in order to guarantee an execution time independent from the number of document models. Finally, the document quality is estimated in order to reject unkown documetns. Our system has been evaluated on a set of 2494 real documents MOTS-CLÉS : Classification, Document d’identité, Descripteurs locaux, SURF, Flann, KDTree.