<i>DeepHadad</i> : Enhancing Readability of Damaged Inscriptions with Synthetic Data

The deterioration of ancient inscriptions over centuries has resulted in irrevocable loss of vital written records, hampering epigraphic analysis, and creating significant gaps in historical knowledge. Factors such as eroded letters and physical damage often compromise the readability of these inscriptions. We present DeepHadad , a neural network trained on procedurally generated synthetic data that use displacement maps and image-to-image translation to digitally restore severely damaged ancient inscriptions to a more readable state. The network’s name is derived from the famous Panamuwa (I) inscription on the mid-8th century BCE Hadad statue, where, at certain places, only faint traces of letters remain on the damaged basalt statue. A key challenge in this work is the lack of well-preserved and damaged glyph pairs for training, as each glyph instance is unique and therefore not found in different states of erosion. We address this by generating synthetic training data through simulated erosion processes, enabling our neural network to successfully generalize to real data. By extracting and overlaying completion maps onto the 3D model, we significantly enhance the legibility of the barely recognizable Aramaic inscription on the Hadad statue. Quantitative and qualitative experiments confirm that our approach can recover textual content that would otherwise be lost or recoverable only through time-consuming manual work. This research opens a pioneering avenue for employing state-of-the-art AI to enrich the readability of ancient textual heritage. Our methodology facilitates a more comprehensive analysis of significant inscriptions and demonstrates the potential of AI-assistive technologies to advance the field of ancient restoration and epigraphic studies.