Enhancing Readability of Damaged Inscriptions with Synthetic Data

What is it about?

Ancient inscriptions are crucial windows into past cultures, but many have been worn down by time, leaving their texts faint or unreadable. One such example is the Panamuwa I inscription on the Hadad statue from the 8th century BCE, currently housed in Berlin. Over 30% of its letters are damaged or missing. Traditional restoration methods, like photogrammetry or Reflectance Transformation Imaging (RTI), are time-consuming and often insufficient when the inscriptions are severely degraded. To tackle this, the authors developed DeepHadad, a deep learning system designed to digitally restore eroded inscriptions. Rather than relying on rare examples of both damaged and undamaged versions of the same text—which don’t exist—the team created thousands of synthetic “damaged” images using realistic 3D models of Aramaic letters. These synthetic datasets mimic various types of deterioration such as erosion, cracking, and warping. The DeepHadad neural network was trained on these images to learn how to “reconstruct” damaged inscriptions by predicting what the original letters might have looked like. Using an advanced image-to-image translation model and 3D depth data (displacement maps), the system can enhance the legibility of severely worn texts while preserving historical accuracy. Tests show that DeepHadad outperforms existing methods and significantly improves readability, even on the real Hadad statue. Experts rated the restorations highly in terms of realism, historical accuracy, and readability. While challenges remain—especially with inscriptions whose styles weren’t part of the training set—this study marks a major step forward in using artificial intelligence for cultural heritage preservation. DeepHadad opens up new possibilities for recovering lost inscriptions and enriching our understanding of ancient texts. The code and tools are freely available for researchers, making this not only a technical breakthrough but also a valuable resource for historians and archaeologists.

Featured Image

Why is it important?

This research is important because it offers a groundbreaking solution to a long-standing problem in archaeology and epigraphy: the loss of legibility in ancient inscriptions due to erosion and damage. Inscriptions are irreplaceable historical sources, but once their texts are worn away, vital cultural and linguistic information can be lost forever. DeepHadad leverages cutting-edge artificial intelligence to digitally reconstruct these texts, using synthetic data to overcome the lack of training material—a key innovation. By enhancing readability while maintaining historical authenticity, the model enables more accurate interpretation of damaged inscriptions and reduces reliance on labor-intensive manual methods. This approach has broad implications: it can be applied to a wide range of inscribed materials across cultures and time periods, helping scholars recover and study texts previously considered illegible. As such, DeepHadad represents a powerful new tool for preserving and reinterpreting the written heritage of the ancient world.