Digital Lunch Seminar Series: Enhancing the Readability of Palimpsests Using Generative AI
Foto: Esfandiari/UHH
When: Mon, 01.07.2024 12:00 PM until 1:00 PM
Where: Digital
Enhancing the Readability of Palimpsests Using Generative AI
Hussein Mohammed and Jost Gippert
Palimpsests are manuscripts that have been scraped or washed for reuse, typically as another document. Recovering the undertext of these manuscripts is of significant interest to scholars in the humanities. Therefore, scholars often employ the multispectral imaging (MSI) technique to render the undertext visible in palimpsests. Nevertheless, this approach might not be sufficient in many cases, given that the undertext in resulting images remains obscured by the overtext. Recent advances in the field of generative artificial intelligence present unprecedented opportunities to discern patterns in highly complex visual data and reconstruct them accordingly. Hence, we framed this challenge as an inpainting task in computer vision, aiming to enhance the readability of the undertext through generative image inpainting. To achieve this objective, we have developed a novel method for generating synthetic multispectral palimpsest images, thereby providing a substantial number of training examples without requiring manual annotation. Furthermore, we employed this dataset in fine-tuning a generative inpainting approach to improve the legibility of palimpsested undertext. The efficacy of this methodology is demonstrated using RGB and MSI images of Georgian-Albanian palimpsests from Mt Sinai.