The description of the proposed system and approach is made using a concrete case study in real-world digitisation scenarios. All required steps are accessible and controllable through a graphical user interface, including the preparation of training data, OCR engine-specific training and dictionary creation, as well as assessment of the impact of the training by performing OCR and evaluating its results. This paper describes such a fully integrated OCR engine training approach using the Aletheia document analysis system. Through a graphical user interface (GUI), OCR engine training should be a straightforward process, allowing anyone to optimise OCR performance, even for smaller volumes of content in several different circumstances. Ideally, an efficient and effective training approach should involve an integrated system where the sequence of data preparation and training steps flows seamlessly and real progress (effectiveness) is evaluated objectively. With the exception of simply adding a set of unseen symbols to commercially available engines (not real training), the process involves several steps for extensive data preparation, running training scripts, evaluating the performance of the newly trained engine, and repeating the cycle incrementally until sufficient performance improvements are made.
Training OCR engines are currently a time-consuming, non-trivial, and disjointed process requiring expert knowledge.
#TESSERACT OCR DOWNLOAD TRAIN DATA MANUAL#
Even in cases where OCR performs well, training can result in meaningful increases of recognition accuracy-a small percentage of quality increase over a large collection of documents can mean significant savings in manual error correction. In such cases, training OCR engines become important in order to recognise those rarer/historic fonts and languages. Some systems allow adjustments via recognition parameters, but this has typically no major impact on results. However, for the multitude of historical documents and for documents written in the many smaller languages in the world, out-of-the-box OCR engines do not perform optimally or even not at all. On average, for most use-cases involving relatively simple (no complex backgrounds, no scanning artefacts) modern material, out-of-the-box OCR engines perform very well as they are configured to recognise text written in the most common fonts in the most popular languages. charities, community enterprises) to individuals undertaking small projects. libraries, archives) to medium-sized operations (e.g. Document digitisation is an everyday continuing activity at all scales, ranging from the very large content holding institutions (e.g.
0 kommentar(er)
