The Impact of Template Matching Threshold on Scribe Identification in Medieval Books

Palaeography is the study of ancient writing. Paleographers often work on the identification of the different scribes who contributed to the writing of a medieval manuscript in order to infer more general information about it that is useful for its historical and cultural contextualization. Digital technologies have made significant progress in recognising standardised handwriting and manuscripts by analysing layout features like page organisation and spacing. However, this approach remains an open challenge, as it lacks generalizability across texts with varying writing conventions. To address this issue, in a previous paper, we proposed a method that draws on the expertise of paleographers, who have identified specific letters and abbreviations that distinguish individual scribes. In particular, we used the letter “a” as the reference symbol because, according to paleographers, this letter serves as a key identifier and appears frequently in texts. We used a Template Matching (TM) technique to identify occurrences of the letter on each page, followed by a Convolutional Neural Network (CNN) to classify each occurrence and assign it to the corresponding scribe. Finally, a majority voting technique determined the lead scribe for each manuscript page based on the most frequent occurrences of the letter “a”. Although this preliminary analysis produced good results, the TM method revealed some challenges, particularly in setting the threshold for visual similarity between characters. In some cases, this required manual cleanup of the segmented characters before CNN training. To address these limitations and study whether the CNN can identify scribes also using dirty letters, we studied the previously proposed system, varying the similarity threshold.