A Fully Automated Approach for Enhancing Ancient Tamil Inscriptions Using Deep Learning

Ancient Tamil inscriptions hold vast historical knowledge but are often difficult to interpret due to deterioration, posing challenges for archaeologists. Various image processing methods have been proposed to enhance inscription readability, but they require manual tuning, making them not fully automatic, and are unable to restore sections that are damaged or eroded. This study introduces a novel, fully automatic approach that utilizes a Fully Convolutional Neural Network (FCN) for the enhancement of Tamil inscription images. The FCN is trained to generate improved images for enhanced readability, using an annotated dataset consisting of original Tamil inscription images and their corresponding text-segmented masks, annotated by a human expert. To address the challenge of imbalanced foreground and background pixels, where foreground pixels corresponding to annotated text regions are fewer than the background pixels, a class-weighted cross-entropy loss function is employed to train the parameters of the FCN model. The proposed method demonstrates promising applicability, achieving a balanced accuracy of 85% on a dataset comprising 25 Tamil inscription images. This automated approach represents a significant advancement in overcoming the challenges associated with the restoration and interpretation of ancient Tamil inscriptions in the field of archaeology.