Automated Identification of Medicinal Plants Through Image Processing and Transfer Learning

Abstract

This study presents an automated system for medicinal plant identificationusingdeep learning techniques, addressing the need for accurate classificationtosupport traditional healthcare, biodiversity conservation, and sustainablepractices. Five advanced CNN models—VGG19, MobileNetV2, ResNet50v2, DenseNet201, and Inceptionv3—were evaluated on a dataset of 6,890 images, divided into training (70%), validation (15%), and testing (15%) sets. The modelswere trained using standardized parameters, including a 224×224 imageresolution, batch size of 128, 50 epochs, and the Adam optimizer with a learningrate of 1e-4. DenseNet201 achieved the highest performance, with a testingaccuracy of 95.19% and an average AUC score of 0.998, demonstrating exceptional generalization and class discrimination, followed closely by ResNet50v2. Amobileapplication was developed to integrate the classification tool, enabling real-timeplant identification in field settings. The mobile app was optimized for efficiency, ensuring compatibility with low-resource environments and offline functionality. While DenseNet201's robustness made it the most suitable model for deployment, Inceptionv3 struggled with stability and precision, highlighting areas for potential refinement. This integrated mobile application offers a scalable and accessiblesolution for researchers, practitioners, and communities, facilitating the accurateidentification of medicinal plants in diverse ecological settings. Future workwill expand datasets, incorporate additional plant features, and further optimizethetool for real-world applications, ensuring its utility in both healthcareandconservation initiatives.