关键词: 水稻叶片病害, 卷积神经网络, YOLOv5, ResNet, 优化器

Abstract: Aiming at the problems of high computational complexity and slow computational speed of YOLOv5 model in rice disease leaf detection, an improved method for YOLOv5 model identification and detection of rice disease leaf based on SPP-x was proposed. Firstly, three MaxPool layers of different sizes (5×5, 9×9, 13×13) in the SPP module of the original BackBone network were replaced with three 5×5 MaxPool layers of the same size, and then the output feature dimension was adjusted by a 1×1 convolutional layer. Then the optimizer in the YOLOv5 network was replaced with Adam. Thus, a new YOLOv5 network structure was constructed. By comparing the convergence rate of SGD and Adam optimizer on the training set, the results showed that the operation time of the improved SPP-x module was only 50% of the original SPP, and the calculation accuracy reached 97%. The two indexes of mAP_0.5 and mAP_0.5:0.95 converged to 0.983 and 0.822, respectively. The experiment found that the detection speed of single image of YOLOv5 model improved SPP-x was 0.34 s, and the effect was good, which could effectively assist rice disease recognition.

Key words: rice leaf diseases, convolutional neural network, YOLOv5, ResNet, optimizer