Bangladeshi Venomous and Nonvenomous Snakes Classification Using Vision Transformer with a Novel Dataset

In this study, we present a novel approach for classifying venomous and nonvenomous snakes of Bangladesh using the cutting-edge Vision Transformer (ViT) model. To the best of our knowledge, this is the first study in the area to use ViT for snake classification. ViT is a more efficient option than Convolutional Neural Networks (CNNs), which were the mainstay of earlier research due to their limited capacity to capture contextual features and long-range dependencies.We created a comprehensive dataset with 3,350 photos of nonvenomous snakes and 1,150 photos of venomous snakes that covers a range of orientations, lighting conditions, and environmental backgrounds in order to overcome the shortcomings of the existing datasets. Class disparities and fictitious backgrounds that were present in previous research are eliminated in this dataset.Our model achieves an impressive 98.8% classification accuracy by utilising ViT’s superior feature extraction capabilities, indicating its potential as a more accurate and dependable solution than CNN-based techniques. To further confirm the model’s resilience, we thoroughly assess its performance using F1-score, precision, and recall metrics.This study has important ecological and public health ramifications in addition to its technical contributions. In Bangladesh, efforts to prevent snakebite, educate the public, respond to medical emergencies, and conserve biodiversity can all benefit from accurate snake classification. In addition to providing a solid basis for further investigation in this uncharted area, our work demonstrates the usefulness of Vision Transformers in actual snake recognition tasks.