The BIR database – Identifying typographic emphasis in list-like historical documents

Layout analysis and optical character recognition have become traditional tasks for processing historical prints, but are now insufficient. Additional information is found in typographic emphasis, such as bold and italic letters. They carry semantic meaning (titles, emphasis...) and also outline the structure of the page (entries, sub-parts...). Retrieving such data is therefore crucial for information extraction and automatic document structuring. In this paper, we introduce the Bold-Italic-Regular (BIR) database, which contains 285 pages of scanned, list-like historical prints that have been annotated at word level with bold and italic emphasis. Baseline results are provided for word detection and style classification using state-of-the-art deep neural network models, highlighting promising possibilities, such as near-human performance for isolated word classification, but also demonstrating limitations for the task at hand.