Medicinal Plant Leaf Classification: From Hand-Crafted Features to Real-Time Mobile Deep Learning

Abstract

Medicinal plant leaves are central to traditional healthcare in biodiversity-rich countries such as Indonesia, yet high visual similarity between species makes manual identification error-prone. Classification methods for medicinal and herbal leaves have shifted from hand-crafted feature classifiers to deep convolutional and attention-based architectures, and more recently to real-time, on-device deployment. This narrative review traces that progression and identifies open gaps. Twenty primary studies published between 2016 and 2026 were identified through Scopus, Google Scholar, ScienceDirect, and Springer and grouped into five categories: hand-crafted features, CNN transfer learning, attention-based fusion, YOLO-based detection, and mobile/edge deployment. Unlike earlier reviews, each study is also coded for whether its model was validated on physical mobile or embedded hardware. The review finds a clear shift toward deep transfer learning models, many reporting accuracies above 90%, and toward architectures designed for on-device use. However, only three of the nine medicinal-leaf-specific studies tested their models on real hardware; these rarely report standardized latency or frame-rate figures, and public benchmark datasets for Indonesian species remain scarce. These gaps motivate work that evaluates classification accuracy and on-device inference performance together.