Classification of Antihypertensive Medicinal Plant Leaves Using MobileNetV2 with Cosine Annealing and L2 Regularization

Abstract

The identification of medicinal plants through leaf characteristics was considered an important approach for supporting the utilization of herbal plants, particularly those associated with hypertension management. However, similarities in visual characteristics among plant species and variations in image conditions presented challenges in the classification process. This study aimed to develop a deep learning model for classifying seven medicinal plant leaf species, namely Annona muricata, Apium graveolens, Centella asiatica, Moringa oleifera, Ocimum basilicum, Psidium guajava, and Syzygium polyanthum. The dataset consisted of 2,100 leaf images from seven classes used for model development, while an independently collected private dataset captured using mobile cameras was used as an independent test set to evaluate model generalization under real-world conditions. The experiments were conducted in three stages. First, four deep learning architectures, namely MobileNetV2, VGG16, ResNet50, and EfficientNetB0, were compared. Second, the effect of Cosine Annealing as a learning rate optimization strategy was evaluated. Third, several regularization methods were compared using MobileNetV2, including no regularization, L1, L2, and Dropout. The results showed that MobileNetV2 achieved the highest performance among the evaluated architectures, with a test accuracy of 94.60%. The application of Cosine Annealing increased the test accuracy to 96.19%. Furthermore, L2 regularization achieved the best overall performance, with training, validation, and test accuracies of 97.68%, 96.87%, and 97.14%, respectively. These results indicated that the combination of MobileNetV2, Cosine Annealing, and L2 regularization improved classification performance while helping to control overfitting.