A Review of Artificial Intelligence-Based Handwriting Recognition Techniques for Ancient Sinhala and Tamil Scripts

Abstract

Ancient Sinhalese and Tamil palm-leaf manuscripts are an important part of the Sri Lankan cultural heritage, but their existence and accessibility are under a harsh siege due to their material decay, incomprehensible handwriting, and inconsistent script. The present paper is a systematic review of previously conducted research on AI based techniques of recognizing handwriting on historical manuscripts, particularly in Sinhalese and Tamil scripts. The review reports on the previous studies regarding datasets, preprocessing methods, deep learning architectures, evaluation metrics, and reported performance. It has been found that Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), CNN-LSTM-based Handwritten Text Recognition (HTR) models, Transformers, and Vision Transformers (ViT) are prevalent in the recent research in the field of historical script recognition. Nevertheless, the fact that there are no large, annotated collections of ancient Sinhala or Tamil manuscripts is the most debilitating constraint. The review also points out some major obstacles associated with work on manuscript degradation, script complexity, and scarcity of data, which explains why these scripts are not well represented in the global research on text recognition by hand. According to the gaps defined in this paper, the knowledge that has been identified is summarized and a structured line of research is planned to facilitate the further work of digitizing and preserving the Sri Lankan manuscript heritage with the help of AI.