Writer Identification in Multiple Medieval Books: A Preliminary Study

One of the key areas of study in palaeography involves identifying the various scribes collaborating on a medieval book. Although digital technologies have allowed significant improvements in this field, it is far from being solved in the general case and is still an open issue. Very interesting results were obtained in the case of highly standardized handwriting and book typologies, where the analysis of some basic layout features regarding the organization of the page and its exploitation by the scribe allowed a high recognition rate. The main drawback of approaches based on layout features is that the results obtained from an ancient text are difficult to use in other texts, produced following different standards. Based on these considerations, we have developed a new approach that attempts to overcome the above-mentioned limitations. The basic idea is to exploit the knowledge of palaeographers who have identified, for each scribe, some letters or abbreviations that characterize them. In this preliminary study, we used two ancient manuscripts, the Avila Bible and the Trento Bible, and we considered the letter “a” as a reference symbol: such letter, according to the indications of the palaeographers, is one of the distinctive symbols able to characterize individual scribes and it is also widely present in all pages of text. A template matching technique was used to identify the occurrences of the character “a” on each page, and a Convolutional Neural Network (CNN) was used to train a classification system capable of attributing each occurrence of the character “a” to the corresponding scribe. Finally, we used a majority voting technique to assign the entire manuscript page to the scribe with the highest number of occurrences of the character “a” on that page. The experimental results obtained on both Bibles confirmed the effectiveness of our method, allowing us to correctly attribute to each scribe over 80% of the pages processed.