Abstract:
Accurate identification of Sri Lankan Ayurvedic medicinal plants, such as Osbeckia octandra,
Atalantia ceylanica, Coccinia grandis, and Murraya koenigii, is challenging due to visual
similarities, complex natural backgrounds, and the absence of specialized datasets. This limits
their safe and effective use for managing chronic diseases like diabetes, high cholesterol, and
NAFLD, particularly for non-specialists.
This research proposes AayuFind which is a deep learning-based ensemble system integrating
UNet for precise leaf segmentation and an ensemble model of ResNet-18, EfficientNet-B0, and
a custom CNN for classification. A custom dataset of 2,400 high-resolution images, captured
in natural settings and annotated with LabelMe, was developed. Data augmentation techniques,
including rotation and brightness adjustments, enhance model robustness, while an Agile
methodology with iterative prototyping ensures user-centric design and performance
optimization.
The prototype achieved a segmentation and a classification accuracy of approximately 99.32%
on a test set images, validated through metrics like precision, recall, F1-score, and AUC-ROC.
These results demonstrate AayuFind’s potential to reliably identify Ayurvedic plants in real-
world conditions, supporting healthcare practitioners and non-specialists.
