An Ensemble Neural Network Model For Malayalam Character Recognition From Palm Leaf Manuscripts

Abstract

Palm leaf manuscripts (PLMs), crucial for ancient communication hold a wealth of information encompassing culture, art, literature, religion, and medicinal wisdom. Malayalam, Kerala's official language, significantly contributes to medical sciences, making palm scripts invaluable, especially in times of pandemics. This study introduces a ground-breaking model for automatic recognition of characters in Malayalam palm scripts. This is the first significant deep learning-based attempt, to our knowledge, to automate Malayalam character recognition in PLMs. The developed model is a fusion of fine-tuned Convolutional Neural Network (CNN) and Bi-directional Long Short-Term Memory (BiLSTM). Discriminative features were extracted from each character in the manuscript through multiple convolutional layers, and these feature vectors were then classified into their respective character classes using an ensemble deep learning model. The performance of the proposed method was evaluated using a self-generated dataset of old Malayalam PLMs from the period 1800 to 1908 AD. Overcoming challenges such as complex morphology, large character set, similar characters, and a unique writing style, the model achieved an impressive accuracy of 96.40%, outperforming state-of-the-art systems. Notably, the model obtained a negative predictive value (NPV) of 99.3%, positive predictive value (PPV) of 83.33%, sensitivity of 79.55%, specificity of 99.45% and F-Measure of 88.39%.Thus this advancement marks a significant milestone in automatic transcriptions providing a crucial tool for doctors and researchers.