GlyptNet: Deep Learning Recognition and Translations of Ancient Tamil Stone inscriptions

Abstract

Old Tamil stone inscriptions are found to be worn out due to the passage of time, weather and natural growth and this makes them extremely hard to read using normal OCR systems. The letters are normally distorted, stained, or worn out. To address this issue, in this paper, a hybrid system is proposed that incorporates several kinds of image information such as 3D depth information, infrared images, and normal colour images to enhance the recovery and reading of ancient Tamil inscriptions used in South Indian temples. We restore the damaged surfaces first by using depth-based correction, infrared enhancement to detect any concealed letters and contrast enhancement algorithms to exploit erosion, lichen growth and uneven surfaces of stones. Once this is done, the better photos are subjected to a deep learning model that processes ancient Tamil characters. The model is a combination of sequence-based recognition algorithm, convolutional and recurrent networks in order to achieve the text without errors. The system was experimented on over 1,850 inscriptions on temples between the 12th and 17th centuries in Tamil Nadu. Experts were involved in the verification of each sample. These findings indicate that the character recognition accuracy is 92.4% which is 18.3 her higher than the traditional OCR systems that use colour images only. The system also was found to be good in semantic accuracy, particularly in the heavily damaged inscriptions where depth and infrared data were used to restore lost letters. The method promotes the cultural heritage by simplifying the process of reading and analysing old writings. It is also an effective and scalable device that allows access to damaged historical texts by historians and researchers. The methods presented in this paper can be modified in order to preserve and keep digitally ancient inscriptions and documents found in other cultures undergoing the same form of destruction.