A Comparative Study on Plant Classification Performance using Deep Learning Optimizers

Abstract

Recently, many Deep Learning architectures have been employed in the identification and classification of a wide variety of plants. This research mainly focuses on classifying the medicinal plants that are available in rural areas. To do so, six well-known pre-trained Convolutional Neural Networks (CNN) namely Dense121, InceptionV3, VGG16, Xception, VGG19, and MobileNet, that were trained for the ImageNet dataset, were chosen by implementing Transfer Learning concept. These models were examined with their pre-trained weights for the Rural Medicinal Plant (RMP) dataset that was created using 8 different classes of medicinal plants that sum up to a total of 16000 images. The performance of these models was improved by training through two state-of-the-art Deep Learning optimizers namely, Stochastic Gradient Descent (SGD) and Adam. These models were trained using Keras with a TensorFlow backend. A comparative evaluation was made for these models to identify the model that attains the best classification. The research concluded that for RMP dataset, the MobileNet architecture, in which the training performance was improved with the SGD optimizer is the best suited model to classify medicinal plants and thus proves the novelty of this research. Therefore, the proposed model can be used by traditional medicine practitioners for the identification and classification of medicinal plants.