Génération d'images semi-synthétiques de documents anciens

This paper presents a software dedicated to semisynthetic old document image generation and its associated ground truth. This work is thus part of the data generation for document processing algorithm performance evaluation. This software allows a user to specify the content and layout of images to generate (fonts, illustrations, line spacing). Several degradation models were incorporated into the software. It is thus possible to generate images containing defects most commonly seen in old books (ink transparency, character degradation). By generating a huge variety of documents, it is possible to assess the robustness of an algorithm according to these degradations. MOTS-CLÉS : modèle de dégradation d’images de documents, génération de vérité terrain, évaluation de performance, bases de données synthétiques.