Classification of Herbal Leaves using EfficientNetB0

Abstract

The identification of herbal leaves remains a challenging task due to the high morphological and visual similarity among commonly used species, which often leads to misclassification when performed manually. This study addresses the challenge of identifying herbal leaves, namely Sauropus androgynus, Moringa oleifera, Orthosiphon aristatus, Syzygium polyanthum, and Piper betle, which are often difficult to distinguish due to high morphological and visual similarity.The proposed approach utilizes the EfficientNetB0 Convolutional Neural Network architecture and employs a two-stage fine-tuning strategy, combined with data augmentation, to enhance generalization performance. A total of 500 manually collected leaf images were used for training, resized to 224×224 pixels, and augmented through rotation and flipping. Model optimization was performed using the Adam and SGD optimizers. The trained model was evaluated on 235 previously unseen external images to assess robustness. The experimental results demonstrate that the proposed model achieved an overall classification accuracy of 88.94%, with particularly strong performance on leaf classes exhibiting distinctive morphological features, such as Orthosiphon aristatus, which obtained an F1-score of 0.96. However, the model exhibited limitations in distinguishing visually similar classes, especially between Moringa oleifera and Sauropus androgynus, both of which possess compound leaf structures, and performance degradation was observed under varying illumination conditions and complex backgrounds. The novelty of this study lies in the application of an EfficientNetB0-based fine-tuning strategy for multi-class herbal leaf classification using a limited, manually collected dataset, demonstrating its potential for deployment in mobile or other resource-constrained environments to support fast and reliable herbal plant identification.