A Unified Deep Learning Framework for Medicinal Leaf Detection, Segmentation, and Classification

Abstract

Medicinal plants play a vital role in traditional and modern medicine and have significant applications in the pharmaceutical, herbal, and nutraceutical industries. Manual identification can be difficult and time-consuming and requires expertise to make it an accurate identification because the leaves of the medicinal plants have a similar morphology and can have different phytochemical content and medicinal properties. Medicinal plant recognition was a three-part task involving leaf detection, segmentation, and classification that was performed sequentially in this work. In this research, the three tasks of recognition (detection, segmentation, and classification) are integrated into a single multi-task deep learning framework for medicinal plant recognition, which would enhance the recognition accuracy, computational efficiency, and robustness in various environmental conditions. Medicinal leaf images are first normalized and augmented to enhance image quality and diversify the dataset. The localization network based on deep learning was used to accurately locate the leaf region, and the segmentation network based on attention was used to accurately extract the boundaries of the leaves with minimal interference from the background. The segmented leaf images are then classified using a deep convolutional network that was able to learn very discriminative features such as leaf venation, texture, leaf shape, and leaf margin features. The wide-ranging experiments are performed on benchmark medicinal plant datasets, and the proposed framework was analyzed using classification accuracy, precision, recall, F1-score, Intersection over Union (IoU), Dice score, training time, inference time, and confidence score analysis. The experimental results show that the proposed framework has an accuracy of 99.1%, which was higher than the conventional CNN, ResNet50, EfficientNetV2, and YOLOv11 models, while also lowering the computational complexity and enhancing the segmentation quality. The integrated learning strategy successfully improves the feature representation, decreases the classification error rate, and achieves good generalization performance under different imaging conditions. The proposed framework ensures automated recognition of medicinal plants with reliability and efficiency and has the potential to be widely used in various fields such as herbal medicine authentication, pharmaceutical quality assessment, biodiversity conservation, precision agriculture, intelligent botanical information systems, and mobile healthcare technologies.