DCGAN-based deep learning approach for medicinal leaf identification

Abstract

Though visual identification of plants seems easier for the trained botanists or agriculturists, the automated identification of plants using leaf images still remains a challenging task. The proper identification of plants forms the most important phase as it leads to usage of plants for various purposes. In this paper, we have manually collected about 30 leaves per species belonging to five medicinal plant species. The dataset was created using the scans of the adaxial and abaxial sides of the leaves. As the small number of images makes it difficult for the Convolutional neural network to learn the features, we have augmented the dataset using Deep Convolutional Generative Adversarial Networks (DCGAN). This paper shows that the low-quality images obtained by the scanner could be effectively augmented using the DCGAN thus increasing the variance in the dataset. A comparison of proposed versions of deep learning models namely VGG16, ResNet50 and DenseNet 121 is presented. To validate the results obtained, 5-Fold-Cross validation was used.