An AI-Driven Computer Vision Framework for Automated Detection, Segmentation, and Classification of Medicinal Plant Leaves

Abstract

Medicinal plants are a significant source of bioactive compounds, which are utilized in traditional health care systems, in the development of pharmaceuticals, and for the production of herbal medicine. Accurate species identification of medicinal plants continues to be difficult, as conventional manual identification relies on the expert knowledge of botanists and was often influenced by the leaf structures, light sources, background complexity, environmental conditions, and similarity between species. Because of these constraints, intelligent and automatic recognition systems that reliably identify a species in real-world conditions are needed. This study introduces a novel computer vision approach to the automatic detection, segmentation, and classification of medicinal plant leaves using the incorporation of advanced deep learning architectures in an integrated processing chain. The proposed model involves image preprocessing, automatic leaf detection with an advanced object detection model, semantic segmentation to extract the leaves accurately, deep feature learning with EfficientNetV2, and transformer networks for precise species recognition. Transfer learning and extensive data augmentation methods are used in order to increase model generalizability of changes in illumination, scale, orientation, occlusion, and background clutter. All the quantitative metrics are calculated and used to thoroughly evaluate the effectiveness of the framework, such as accuracy, precision, recall, F1-score, Intersection over Union (IoU), Dice Similarity Coefficient, mean Average Precision (mAP), Receiver Operating Characteristic (ROC), Area Under the Curve (AUC), inference time, etc. The superior performance in classification, stability of convergence during training, segmentation accuracy, localization of leaves, and excellent generalization ability across different medicinal plant species are demonstrated through experimental analysis. Explainable artificial intelligence using Grad-CAM was another method to improve model transparency by understanding the regions of the image that influenced the decision-making process.