This project develops a novel approach that takes one handwritten page and finds the other pages in a collection written by the same scribe, ranked by how similar their handwriting is. The current state-of-the-art methods first convert every page into a black-and-white image, which separates the ink from the writing surface, and analyse only this simplified version. This removes stains, uneven lighting and the texture of the paper. It also removes information in the writing itself, such as variations in ink intensity, and on damaged or faded material the conversion is often unreliable. The aim of the project is a method that works directly on the original image while improving accuracy.

The method is trained in two stages. First, a small neural network learns to describe handwriting from black-and-white images, without being told which scribe wrote which page. Second, a larger network receives the original image and learns to reproduce the descriptions of the small network. Black-and-white images are therefore needed only for training, not for the collections being searched. To evaluate the method, we measure how highly the pages by the same scribe are ranked within public benchmark collections of historical documents; this is the standard way of comparing such methods. All models, including the leading published system, are evaluated with the same procedure, so that differences in the results come from the models and not from the evaluation, and each difference is tested statistically. On these benchmarks, the method achieves the highest accuracy reported so far. The trained model is also tested, without further training, on collections it has never seen, such as modern handwriting and papyri, to check how well it works beyond its training material. The project also develops tools that apply the trained model to new collections of page images, for example to search for pages by the same scribe, to group pages by similarity of handwriting, to assign pages to known scribes from a few examples, or to compare two sets of pages.

Hussein Mohammed

Starting date: September 2025