Application of MobileNet-v1 for Potato Plant Disease Detection Using Transfer Learning

Infectious diseases have troubled farmers continuously by spreading throughout crops. Thus, a proper identification of such disease is obligatory for the timely treatment which can save the money and efforts of small-scale farmers. The recent advancement in deep learning has provided a way to contribute to the sector of agriculture. In this paper deep learning based MobileNet architecture is employed to identify potato plant lesion characteristic. The application of transfer learning is accomplished by freezing the base layers and training only top 23 layers containing the added classifier layer. The model is then trained further to improve performance. The frozen layer weights of this pretrained model remained constant during training while the top layer weights are constrained by fine tuning to quit generalize feature map and get associated with specific features of new dataset. This enhances the model performances and gives 99.83 % accuracy in the image classification on the leaves of potato plant into the categories of infected disease. The experimental results demonstrate the feasibility of this procedure on portable devices.