Identification of Indian Medicinal Plants from Leaves using Transfer Learning Approach

India is well-known for providing the optimal habitat for a wide variety of medicinal plants. Numerous parts of the plants are utilized as crucial elements for the creation of natural medications. Utilization of image processing and Computer Vision strategies for the distinguishing proof of the medicinal plants are extremely urgent as a large number of these plants are under extinction according to the IUCN (International Union for Conservation of Nature) records. Subsequently, the digitization of helpful therapeutic plants is urgent for the protection of biodiversity. This paper examines Convolutional Neural Network (CNN) based methodologies for distinguishing Indian leaf species. In recent times, numerous Deep Learning structures have been utilized in the distinguishing proof and characterization of a wide assortment of plants. This research chiefly centers around identifying the therapeutic plants that are accessible in rustic territories. To do so, three notable pre-trained CNN architectures namely ResNet101, InceptionV3, and VGG16 that were trained for the ImageNet database were chosen by actualizing the Transfer Learning technique. These models were inspected with their pre-trained weights for the Ayur Bharat dataset that was created utilizing 10 unique classes of medicinal plants that sum up to a total of 10000 images. The performance of these models was improved via the Canny edge detection method and a comparison was carried out between these three architectures that were trained, both without and with, preprocessing techniques. The best classification performance for the Ayur Bharat dataset was attained by the InceptionV3 architecture trained with the Canny edge detection pre-processing technique, with the finest validation accuracy and F1-score of 0.9732 and 0.9653 respectively.