Classifying Medieval Manuscripts by Pen and Support

We present a machine-learning approach for classifying medieval Hebrew manuscripts by two key material attributes: writing support (whether the substrate is paper or parchment) and writing implement (quill pen vs. reed calamus). Our work contributes to the emerging field of computational codicology, offering tools to aid paleographers in the large-scale analysis of digitized manuscripts. Our datasets—derived from existing digitized repositories— have been carefully annotated and balanced to capture the range of material and stylistic variation. For both classification tasks, we employ convolutional neural networks tailored to their respective challenges: identifying broad substrate textures and capturing fine-grained stroke morphology. The support classifier achieved an accuracy of 91% and demonstrated reliable performance even on visually ambiguous examples. Likewise, the implement classifier was 91.5% accurate. These findings show that computational analysis can aid and, in some cases, surpass manual paleographic methods in analyzing historical manuscripts. This work highlights the potential of computational tools to assist scholars in large-scale analysis of digitized corpora, aiding manuscript dating, provenance research, and the study of scribal practices.