Ayurvedic Plant Identification Using Transfer Learning

Abstract

Our healthy ancestors mostly depended on plants around them for their medical reasons. There are hundreds of species of plants around us in our environment which will serve us as home remedies for all most all the diseases and keeps us healthy when taken as daily consumption. Especially in Sri Lanka we have our own set of rare Ayurvedic herbs. But most of us unable to identify these plants due to lack of knowledge. There are existing applications which can identify plants with low prediction accuracies. Also, those applications are based on foreign plant data sets that do not include the valuable herbs and shrubs with medicinal qualities. Hence this research work proposes a mobile application which can identify Ayurvedic plants using Convolution Neural Network models (CNN) with Transfer learning. Three CNN architectures have been applied in this study experimentally to obtain the best accurate model. Ayurvedic plant dataset is trained on CNN from scratch and obtained about 70% of accuracy. Anyhow training a model from scratch is too costly. To overcome this challenge transfer learning has been applied in the Ayurvedic dataset. The pretrained deep learning models used in this study are Google-Net and Mobile-Net and the accuracies obtained are 80% and 93% respectively. Finally, the Mobile-Net, a small efficient convolution neural network which produces the highest accuracy is used to produce the final prediction model. This model is then used to build a mobile application in the Android platform with TensorFlow-Lite which can identify the Ayurvedic plants using the built-in camera module.