Amrit Baskota
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Amrit Baskota has a total of 3 co-authors. Here are the top co-authors by number of shared publications:
- Abiskar Ghimire
School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology, Vellore, Tamil Nadu, India
1 shared publication - P. Baskaran
School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology, Vellore, Tamil Nadu, India
1 shared publication - Shubham Ghimire
School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology, Vellore, Tamil Nadu, India
1 shared publication
Latest Publications
- Open Access
- 16 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 179333-179350
Herbguard: An Ensemble Deep Learning Framework With Efficientnet and Vision Transformers for Fine-Grained Classification of Medicinal and Poisonous Plants
Classifying whether a plant is herbal or poisonous is a significant challenge for trekkers, hikers, and nature enthusiasts, particularly in remote areas where several unfamiliar plant species are encountered. Consumption or even close contact to some poisonous plants species may lead to serious health risks, highlighting the need of an intelligent and real-time classification system. In this study, we an approach based on deep-learning based to classify plants into several species under herbal and poisonous categories. The system employs Convolutional Neural Network (CNN) based EfficientNetV2-S, which is optimized for local feature extraction, and transformer-based ViT-Tiny, which is capable of capturing global dependencies across images. Both models are trained using a two-stage fine-tuning strategy with label smoothing, MixUp, and CutMix augmentations. Preprocessing steps include CLAHE-based contrast enhancement, HSV masking and GrabCut segmentation, that are applied to training images focus on relevant plant regions. The models are evaluated on 48 different plant species, consisting of 40 herbal and 8 poisonous species ultimately achieving species-level accuracies of 95.86% (EfficientNet) and 96.69% (ViT) on validation dataset. A soft-voting ensemble of the two models further improves species-level accuracy to 97.10% upon validation and 98.43% upon testing, while category-level accuracy remains consistently above 99.7% for all the models. These results demonstrate that combining convolutional and transformer-based approaches leads to a robust, highly accurate classification system that is capable of distinguishing varieties of medicinal and poisonous plants, offering a practical tool for safe-trekking, biodiversity monitoring, and herbal medicine research.
- 16 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 179333-179350
Open AccessHerbguard: An Ensemble Deep Learning Framework With Efficientnet and Vision Transformers for Fine-Grained Classification of Medicinal and Poisonous Plants
Classifying whether a plant is herbal or poisonous is a significant challenge for trekkers, hikers, and nature enthusiasts, particularly in remote areas where several unfamiliar plant species are encountered. Consumption or even close contact to some poisonous plants species may lead to serious health risks, highlighting the need of an intelligent and real-time classification system. In this study, we an approach based on deep-learning based to classify plants into several species under herbal and poisonous categories. The system employs Convolutional Neural Network (CNN) based EfficientNetV2-S, which is optimized for local feature extraction, and transformer-based ViT-Tiny, which is capable of capturing global dependencies across images. Both models are trained using a two-stage fine-tuning strategy with label smoothing, MixUp, and CutMix augmentations. Preprocessing steps include CLAHE-based contrast enhancement, HSV masking and GrabCut segmentation, that are applied to training images focus on relevant plant regions. The models are evaluated on 48 different plant species, consisting of 40 herbal and 8 poisonous species ultimately achieving species-level accuracies of 95.86% (EfficientNet) and 96.69% (ViT) on validation dataset. A soft-voting ensemble of the two models further improves species-level accuracy to 97.10% upon validation and 98.43% upon testing, while category-level accuracy remains consistently above 99.7% for all the models. These results demonstrate that combining convolutional and transformer-based approaches leads to a robust, highly accurate classification system that is capable of distinguishing varieties of medicinal and poisonous plants, offering a practical tool for safe-trekking, biodiversity monitoring, and herbal medicine research.