An Approach to the Classification of Ancient Indian Scripts Using the CNN Model

This research aims to investigate the performance of the convolutional neural network (CNN) for the classification of four rare ancient Indian scripts, namely, Devanagari, Nandinagari Newari, and Sharda. Script classification is an incredibly difficult task, especially in India where there are 14 Indian scripts used to write 22 official languages. Our work limits itself to the study of these four Indian ancient scripts to have a scope for developing an efficient classification model for script identification. In the context of our awareness of this work, earlier researchers have made significantly less effort to identify ancient Indian scripts: Devanagari, Nandinagari, Newari, and Sharda. Researchers have never implemented any deep learning techniques for the recognition of such ancient scripts. Hence our proposed approach is the first-time implementation of a CNN-based model for the multiclass classification of these ancient Indian scripts. For experimentation, we use the self-created dataset of IGNCA manuscripts, which consists of four rare Indian ancient scripts: Devanagari, Nandinagari, Newari, and Sharda. In this study, we accept the challenges of data collection, creation of our dataset, exploration of their distinguishable features, preprocessing, and classification of these Indian ancient scripts. The proposed convolutional neural network (CNN) model achieves classification accuracy of 99.03% and 98.71% on the training set and the testing set, respectively.