Resumen
Buckwheat is an important cereal crop with high nutritional and health value. Buckwheat disease greatly affects the quality and yield of buckwheat. The real-time monitoring of disease is an essential part of ensuring the development of the buckwheat industry. In this research work, we proposed an automated way to identify buckwheat diseases. It was achieved by integrating a convolutional neural network (CNN) with the image processing technology. Firstly, the proposed approach would detect the buckwheat disease area accurately. Then, to improve the accuracy of classification, a two-level inception structure was added to the traditional convolutional neural network for accurate feature extraction. It also helps to handle low-quality image problems, which includes complex imaging environment and leaf crossing in sampling buckwheat image, etc. At the same time, instead of the traditional convolution, the convolution based on cosine similarity was adopted to reduce the influence of uneven illumination during the imaging. The experiment proved that the revised convolution enabled better feature extraction within samples with uneven illumination. Finally, the experiment results showed that the accuracy, recall, and F1-measure of the disease detection reached 97.54, 96.38, and 97.82%, respectively. For identifying disease categories, the mean values of precision, recall, and F1-measure were 84.86, 85.78, and 85.4%. Our method has provided important technical support for realizing the automatic recognition of buckwheat diseases.