Reconnaissance automatique de textes dans les vidéos à l’aide d’un OCR et de connaissances linguistiques
Résumé
Our work aims at helping multimedia content understanding by extracting textual clues embedded in digital video data. For this, we
developed a video Optical Character Recognition (OCR) system, specifically adapted to detect and recognize embedded texts. Based on a neural approach, our method outperforms related work especially in terms of robustness to style and size variability, to background complexity and to low resolution of the image. We also introduced a language model that drives several steps of the video OCR in order to remove ambiguities related to recognition and reduce segmentation errors. This approach has been evaluated on a database of French TV news videos and achieves a character recognition rate of 95%, which enables its incorporation in a video indexing system.