The problems and solution of deep learning based rice disease detection in natural scene images

Rice is the most widely grown crop on the planet. However, the increasing severity of rice diseases has resulted in significant economic losses for farmers. With the rapid development of mobile devices and mobile services, computing is playing an increasingly significant role in our daily lives. How to develop an intelligent diagnosis system for rice diseases based on mobile computing services and to bridge the gap between rice growers and plant diagnostic specialists is worthy of study. With the assistance of experts, we construct an image dataset of eleven types of rice diseases, as well as some healthy plants for handling false-positive results, and realize an intelligent diagnosis system for rice diseases by constructing state-of-the-art object detection algorithm Faster Regional Convolutional Neural Network (Faster RCNN) by fine tuning different base networks, namely Visual Geometry Group (VGG) 16, 19, Densely Connected Neural Networks (DenseNet), and Recurrent Neural Networks. The system is implemented through the WeChat applet on mobile devices, allowing users to upload images and receive diagnostic results and comments. The experimental results demonstrate that the recognition accuracy of rice diseases exceeds 61%, and that the prediction time consumption has been reduced by simplifying the structure of the base networks.