Aioanei, Andrei C.Klein, JonathanKlein, Konstantin M.Hunziker-Rodewald, Regine R.Michels, Dominik L.Corsini, MassimilianoFerdani, DanieleKuijper, ArjanKutlu, Hasan2024-09-152024-09-152024978-3-03868-248-62312-6124https://doi.org/10.2312/gch.20241242https://diglib.eg.org/handle/10.2312/gch20241242We present DeepHadad, a novel deep learning approach to improve the readability of severely damaged ancient Northwest Semitic inscriptions. By leveraging concepts of displacement maps and image-to-image translation, DeepHadad effectively recovers text from barely recognizable inscriptions, such as the one on the Hadad statue. A main challenge is the lack of pairs of well-preserved and damaged glyphs as training data since each available glyph instance has a unique shape and is not available in different states of erosion. We overcome this issue by generating synthetic training data through a simulated erosion process, on which we then train a neural network that successfully generalizes to real data. We demonstrate significant improvements in readability and historical authenticity compared to existing methods, opening new avenues for AI-assisted epigraphic analysis.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Mesh geometry models; Reconstruction; Neural networks; Applied computing → Arts and humanitiesComputing methodologies → Mesh geometry modelsReconstructionNeural networksApplied computing → Arts and humanities10.2312/gch.202412426 pages