A palimpsest is a distinctive category of historical manuscript from which the original text was erased, washed, or scraped away to allow the material to be reused for new artefacts. This practice was particularly common in ancient and medieval times when writing materials, such as parchment, were scarce and costly, making recycling a widespread cultural practice. Despite the efforts of scribes to obscure the original content, faint traces of the earlier writing often remained beneath the newer text. These fragments provide historians and scholars with invaluable insights, uncovering hidden layers of history that were not entirely eliminated.

 

The advent of Multispectral Imaging (MSI) has revolutionised the study of palimpsests by enabling the recovery of these hidden texts. By capturing data across multiple wavelengths, MSI can render the original undertext visible. However, while MSI is highly effective in this respect, the recovered content often remains difficult to interpret due to its overlap with the overwritten text. Efforts to separate or remove the overwritten content frequently result in the loss of portions of the undertext, mirroring the challenges encountered in the study of degraded manuscripts where sections of the document have faded or been damaged over time.

 

The main objectives of this project are to explore the feasibility of revealing undertexts in palimpsests and to integrate image inpainting techniques for reconstructing undertexts in MSI images using Generative AI. By applying undertext image inpainting, the project aims to produce reconstructions that are contextually clearer, more readable, and more useful for experts compared to the original MSI images. This approach has the potential to significantly enhance document preservation and content recovery, ensuring that palimpsests continue to serve as invaluable resources for cultural heritage preservation.

 

The results of this project, which aims to improve the readability of text in MSI images, particularly in palimpsests, are actively being used by members of the DeLiCaTe project (2022–2027) to reveal undertexts in Georgian and Armenian palimpsests, as well as by the Recovery of Writing in Large Collections project (2023–2025) to enhance text retrieval and develop methods using MSI. Furthermore, this research holds potential benefits for other stakeholders, including libraries, academic institutions, and scholars by providing advanced methods for the analysis and restoration of historical manuscripts.

 

Mahdi Champour (Jampour)

 

keywords: Generative AI, Palimpsests, Inpainting, Latent Diffusion Model, GANs, Deciphering Palimpsests, VAEs,