The identification and classification of medicinal plants is crucial for traditional medicine, medical applications, and plant variety preservation. However, doing this by hand requires specialized expertise, is time-consuming, and can result in errors. This highlights the demand for an automated, accurate, and efficient approach to differentiate between plant species and ensure the quality and safety of medicinal leaves. In this study, we address this medicinal plants, aiming to develop a more generalizable classification model. Pre-processing techniques are applied to emphasize the key features within the images. Several advanced CNNs like DenseNet121, InceptionV3, Xception, and MobileNetV2 were evaluated, with MobileNetV2 selected for its high accuracy and lightweight design. Enhancements such as dense layers, global average pooling, regularization, and self-attention were added to improve performance. The proposed model was tested on a dataset of Indian medicinal leaves and achieved excellent results: 98.03% precision, 98.07% recall, 98.45% accuracy, and a 99.89% F1 score. These results show that the model performs better than existing methods and is highly effective for both medical and industrial applications.