Denoising Ancient Epigraphical Estampages from India using Transfer Learning with Surrogate Benchmarking

Archaeological inscriptions in India are of immense historical importance offering conclusive evidence about the reign of various dynasties. Several obsolete languages, such as the Brahmi script, have survived as inscriptions on copper plates, slabs, pillars & other artifacts. Many ancient epigraphical carvings have deteriorated due to insidious exposure to natural elements, decay, deliberate destruction, and a multitude of other factors. To facilitate computational transcription, this paper proposes the use of transfer learning methods with models such as Autoencoder and U-Net for effectively denoising degraded epigraphical estampages, supplemented by a structured image processing framework for downstream character recognition. In the absence of definitive ground truth, we rely on surrogate benchmarking to quantitatively evaluate the effectiveness of the models on alternative datasets, particularly in the context of epigraphical denoising. The proposed U-Net Model is benchmarked on the publicly available SIDD denoising dataset, obtaining an SSIM score of 0.9559. Using surrogate benchmarking to evaluate the performance of the model on denoising epigraphical images, an SSIM score of 0.85 is obtained for the output of the U-Net model, demonstrating its ability to denoise epigraphical estampages. The proposed research lays down a framework for effectively denoising epigraphical estampages sourced from India.