关键词: 马铃薯疫病, YOLOv5s, 损失函数, SimAM注意力, 轻量化网络

Abstract: Rapid and accurate identification of crops in the early stage of epidemic disease is an important link to reduce the economic loss of crops. Aiming at the actual production environment, the traditional image processing algorithms can not accurately identify the leaf surface texture features and determine the type of disease, and the YOLOv5s model has a large number of parameters, and the recognition effect is poor in complex environments and other problems, this paper proposes an integrated and improved method for potato disease detection and identification. In this paper,  the number of parameters is reduced by replacing the lightweight network of YOLOv5s, the fusion ability of different feature layers of the model is enhanced by using the weighted bidirectional feature pyramid network (BiFPN), and the attention mechanism module SimAM is increased, the ability of the YOLO algorithm is enhanced to extract the key information by using the GSConv convolution, and finally,  the regression accuracy is improved by introducing the SIOU loss function. Under the same experimental conditions, compared with the original model of YOLOv5s, YOLOv7—tiny, Faster R—CNN and other models, the accuracy rate, recall rate, and average recognition accuracy of this proposed method reaches  97.7%, 95.9% and 95.4%, respectively. The proposed algorithm not only improves the accuracy and average accuracy, but also reaches the speed of 144.93 frames/s, which meets the requirement of potato blight detection.

Key words:  , potato blight detection, YOLOv5s, loss function, SimAM attention, lightweight network