Assessing advanced handwritten text recognition engines for digitizing historical documents.

Area of Science:

• Digital Humanities
• Computer Science
• Information Science

Background:

• Digitizing historical documents is crucial for accessibility and research.
• Evaluating Handwritten Text Recognition (HTR) engine performance is essential for effective digital preservation.

Purpose of the Study:

• To critically assess and compare the performance of leading HTR engines.
• To evaluate their accuracy and efficiency across diverse historical document datasets.

Main Methods:

• Comparative analysis of five state-of-the-art HTR engines: PyLaia, HTR+, IDA, TrOCR-f, and Titan.
• Testing across datasets with varied scripts, styles, and historical orthography.

Main Results:

• Titan and TrOCR-f demonstrated superior out-of-the-box performance for Latin-script documents.
• PyLaia, IDA, and HTR+ showed strong performance on specific non-Latin scripts after fine-tuning.
• All engines benefit significantly from training and fine-tuning, including language model integration.

Conclusions:

• HTR engine selection depends on script and language requirements.
• Effective digitization necessitates tailored training, fine-tuning, and language model integration.
• This research provides insights for advancing HTR technology in digital humanities.