Evaluation of Commercial OCR: A New Goal Directed Methodology for Video Documents

Texts embedded in video streams convey crucial information for documentation. Many text detection and recognition systems have been designed to automatically extract such documentary data from video streams. Most of the research teams involved argue that commercial OCR do not work properly on images extracted from a video stream. They thus concieve their own detection systems. Nevertheless, commercial OCR have never been evaluated on such corpora. This article details a new methodology to evaluate a commercial OCR on a video document. This methodology is goal directed: the system is penalized proportionally to TFIDF (Term Frequency Inverse Document Frequency) scores of texts [1]. We experiment our methodology on Abbyy FineReader 6.0.