A Comprehensive Authenticated Dataset of Brahmi Script Characters for Recognition

The Brahmi script is among the most ancient and historically prominent systems of writing in South Asia, with applications that warrant corresponding modern Indic scripts. Digitally recognizing Brahmi inscriptions holds the key to epigraphical studies, historical research, and cultural preservation. The efforts remained always undermined for the absence of a greater and authentic dataset. To meet this requirement, a new dataset is introduced, comprising 348 different classes of characters from the Brahmi script. The dataset contains a total of 7880 character images, each such image extracted from letter carvings over stone at seven salient archaeological sites in India: Naneghat Caves, Junnar Caves, Bhaje Caves, Karle Caves, Bedse Caves, Nashik Caves, and Pitalkhora Caves. The dataset design was an iterative and exhaustive process of image acquisition, segmentation of characters, and applications of preprocessing to increase visibility and recognizability of characters therein. The usability and robustness of the dataset were tested using four different deep learning models: a custom Deep Convolutional Neural Network (Deep CNN), InceptionNet, ResNet50, and XceptionNet. The best recognition accuracy of 98.07% was attained with a loss of 0.0958 using ResNet50, which shows the adequacy and quality of the dataset for the training of advanced machine learning models. This will help researchers by providing a valuable resource and accelerating the development of Brahmi script character recognition.