Investigation of OCR Model Performance and Post-OCR Correction Strategies: A Comparative Analysis

Sammanfattning

The act of digitizing text from images and documents is Optical Character Recognition (OCR). This thesis investigates the effectiveness of modern OCR models and post-OCR correction strategies when applied to Swedish historical newspapers from 1818-2018. The aim of the thesis is to provide a thorough understanding of current optimal digitization approaches, both in terms of the OCR process and the post-OCR process where erroneous OCR output is corrected. Two post-OCR approaches are evaluated: model transferability across different OCR engines and models trained on synthetic data. Using a dataset of 43,823 newspaper paragraphs, four OCR engines were evaluated: Tesseract, Kraken, DocTR and Idefics. These were selected to display a variety of OCR engines, which includes vision language models, models specifically trained on historical material, and bidirectional long short-term memory text recognition models. The results show that Tesseract version 5.5.0 overall performs best with 6.7% Character Error Rate (CER) and 22.4% Word Error Rate (WER), followed by Kraken German with 7.6% CER and 26.5% WER. Even though Kraken German consistently struggles with the Swedish character "å", it still outperformed Tesseract in one time period of the data. This, along with the 8-fold increase in CER between the modern period (1980-2018) and historical period (1818-1859), underscores the need for specialized and localized OCR models for Swedish historical material. The post-OCR model transferability assessment reveals that all OCR models achieve improved results when their outputs are corrected using matching post-OCR models. However some OCR models like DocTR are more compatible for cross-model post-OCR correction than others. DocTR achieved the lowest CER after post-OCR correction step, surpassing both Tesseract's OCR output and Tesseract's post-OCR corrected output. The synthetic data approach found that a combination of synthetic data from within the domain and outside the domain yielded the best post-OCR results, achieving a 15.9% reduction in word-level errors. However, this synthetic approach is still inferior to post-OCR using real training data, which achieves better results.