Computer vision has made significant progress in recent years, thanks in part to developments in machine learning, and is now poised to become a standard part of the toolkit of all those who work with such ‘legacy’ media as printed books and illustrations. Computers can now reliably match the same printed page or illustration; visualise variant typesettings or images; and, through, improvements in OCR models, extract text. More challenging applications, such as segmenting pages into meaningful parts and classifying their content, are within reach. But we still have much to learn about the printed book – in particular, how page layout and such material dimensions as paper, ink and format structure their form and meaning.
This paper will offer an account of recent work in computer vision applied to print, largely using the example of the University of Oxford’s Visual Geometry Group (VGG), whose open-source tools are embedded in projects such as Bodleian Ballads Online; the 15C Illustration project; and the Traherne Digital Collator. It will assess how far ML-based approaches are able to classify the content of historical printed illustrations, using the British Library’s Million Images corpus as a case study, and give an account of the state of the art in OCR for pre-modern or other challenging typefaces. Lastly, it will offer some critical reflections on the ‘datafication’ of this iconic form of knowledge and cultural production, arguing that a historically-informed perspective on remediation has much to offer digital humanities’ practice across multiple media.
Keywords: printing, typography, computer vision