Vision Transformer-based Approach for Classification of Buddhist and Non-Buddhist Heritage Sites in Taxila

Preserving cultural heritage requires accurate classification of diverse artefacts to enable efficient analysis. In this paper, we present a unique dataset curated to classify cultural heritage sites in Taxila city, Pakistan, which was the capital of the ancient Gandhara civilization. The target is to classify Buddhist heritage sites. The dataset consists of different categories including images of Takshasila University, Sirkap, Giri Fort, Taxila Ruins, Jaulian Buddhist Monastery, Dharmarajika Stupa and Monastery, and non-Buddhist heritage sites. The dataset was captured with varying illumination, angles, and lighting to ensure the model was well-generalized. For non-Buddhist data samples, we included images of old buildings and a few heritage sites that are not from Buddhist civilization. After capturing the dataset, we used a vision transformer (ViT) to classify images. We also compared the performance of ViT with conventional convolutional neural networks such as ResNet and EfficientNet. The experimental results show that ViT outperforms CNNs in classifying cultural heritage sites by achieving 100% accuracy in test sample.