Medicinal leaf classification is highly essential in the identification and location of medicinal plants that would eventually guarantee the discovery and development boosting health advancements and plant conservations. The paper aims to solve a medicinal plant classification problem using the lightweight deep learning network MobileNet architecture for leaves against the categories. The model will be designed to make use of the concept of transfer learning in leveraging pre-trained weights derived from large datasets such as ImageNet, which enables the fine-tuning for unique features in medicinal plant leaves. Consequently, this will reduce training time and computational resources while offering better accuracy. The best hyperparameters of learning rate, batch size, and optimizer have been obtained by hyperparameter optimization using grid search. Extensive comparisons between the models that used pre-trained weights against the models initialized with random weights have indicated that transfer learning has ensured the models converge faster with improved accuracy to show their potency, especially on limited and domain-specific datasets, placing them as one of the key techniques toward research improvement for medicinal plant classification.