Abstract:  In order to accurately and effectively identify locust species in the Ningxia desert grassland region, a locust target detection model YOLOv5s-CG was proposed based on the YOLOv5s network model in a complex context, by using Contextual Transformer Networks (CoTNet) in the backbone network to retain locust feature information in the complex context, while incorporating the global attention mechanism in the neck network to improve the feature fusion capability of the detection model. The experimental results showed that the model YOLOv5s-CG had a precision of 92.5% and a mean average precision of 93.2% in locust identification in Ningxia desert grassland, which was 4.8 percentage points and 5.3 percentage points higher than the original model, respectively. Compared with the Fast R-CNN, YOLOv3, YOLOv4,YOLOv5s, YOLOv6s and YOLOv7 models, the YOLOv5s-CG network model had better detection performance for locusts in Ningxia desert grassland. Finally, based on the YOLOv5s-CG, a locust identification APP system was developed to achieve online locust identification and detection in the complex background of the Ningxia desert grassland region, providing data support for the supervision, prevention, control and comprehensive management of locusts in the Ningxia desert grassland region.

Key words: locust identification, deep learning, self?attention mechanisms, global attention mechanism

摘要： 为准确有效识别宁夏荒漠草原地区蝗虫种类，基于YOLOv5s网络模型提出一种复杂背景下蝗虫目标检测模型YOLOv5s-CG，在主干网络使用CoTNet保留复杂背景下的蝗虫特征信息，同时在颈部网络中融入GAM全局注意力机制提高检测模型的特征融合能力。结果表明，在对宁夏荒漠草原蝗虫进行识别时，模型YOLOv5s-CG 精确率为92.5%，平均精度均值为93.2%，相比原始模型分别提高4.8个百分点和5.3个百分点，与 Fast R-CNN、YOLOv3、YOLOv4、YOLOv5s、YOLOv6s、YOLOv7模型相比，YOLOv5s-CG网络模型对宁夏荒漠草原蝗虫具有更好的检测性能。基于YOLOv5s-CG，开发宁夏荒漠草原蝗虫识别APP系统，实现宁夏荒漠草原地区复杂背景下的蝗虫在线识别检测，为宁夏荒漠草原地区蝗虫的监管防控和综合治理提供数据支持。

关键词: 蝗虫识别, 深度学习, 自注意力机制, 全局注意力机制