抄録
Effective detection of rice diseases is an important subject to improve the yield and quality of rice. Convolution neural network (CNN) is widely used in plant disease detection, but a large number of training samples are needed to build the model. In this paper, a biphasic method based on CNN is proposed, which can simplify training samples. This method takes into account a variety of rice diseases and a prediction accuracy of 88.9%. Using this method can effectively establish a rice disease dataset, accurate detection, and disease classification. There are three novelties in this paper: (1) a dual-phase approach capable of learning from a small rice grain disease dataset has been proposed; (2) a smart segmentation procedure has been proposed which is capable of handling heterogeneous backgrounds prevalent in plant disease image datasets collected in real-life scenarios; (3) experimental comparison has been provided with straightforward use of improved CNN architectures on the small rice grain dataset to show the effectiveness of the proposed approach.
