Historical Texts; Neural Networks; OCR; Synthetic Data

We describe a development of a custom OCR system, which is designed specifically for a linguistic analysis of texts printed during the early modern period. This analysis requires precise detection of individual graphemes, and we, therefore, could not apply standard approaches that transcribe whole lines in an end-to-end fashion. We also describe our use of synthetically generated images, which allow us to avoid manual annotation of a large training set.

We describe a development of a custom OCR system, which is designed specifically for a linguistic analysis of texts printed during the early modern period. This analysis requires precise detection of individual graphemes, and we, therefore, could not apply standard approaches that transcribe whole lines in an end-to-end fashion. We also describe our use of synthetically generated images, which allow us to avoid manual annotation of a large training set.