Plant Disease Recognition Using Optimized Deep Convolutional Neural Networks

Abstract

In this paper, the problem of recognizing the plant’s diseases and pests using deep learning methods has been addressed. This work can be implemented on a client-side or integrated with IoT concept, in order to be employed efficiently in smart farms. Nearly $40\%$ of global crop yields each year are lost due to pests. By considering the global population growth, the agricultural food will run out of its resources very soon and this will endanger the lives of many people. A pretrained EfficientNet deep neural network architecture with student noise has been optimized, both in volume and the parameter number, and has been involved in this setup. Two different approaches have been adopted. First, achieving the highest accuracy of recognition using the optimum algorithms in development step. Second, preparation of the system as a microservice model in order to be integrated with other services in a smart agriculture deployment. Using an efficient number of parameters and inference time, it has become doable to implement this system as a service in a real world scenario. The dataset used in the training step is the plant village data. By implementing the model on this dataset, we could achieve the accuracy of $99.69\%$ on test data, $99.85\%$ on validation data, and $99.78\%$ on training data, which is remarkably competitive with the state-of-the-art.