Authors

Abstract

To accurately detect various kinds of locusts in real-time and make locust detection more universal, A locust data set that contains all kinds of locusts was created through the Internet crawler and public dataset IP102, and a locust target detection algorithm YOLOv7-MobileNetV3-CA.was proposed in this paper, Firstly, to reduce the size of model parameters, the backbone of YOLOv7 was replaced by MobileNetV3, Secondly, a CA (Coordinate Attention) attention mechanism was added to further improve the detection accuracy of locusts. after feature enhancement. The experiment showed that the precision of locusts was 95.96%, the recall rate was 92%, the AP was 95.74%, and the F1 was 0.92. Compared with YOLOv7, the model size was reduced by 27%, and the AP was improved by 4.48%. Compared with YOLOv4, YOLOv4 MobileNetV3, YOLOv5, and SSD algorithms, AP has improved by 51.16%, 26.81%, 11.9%, and 11.75%, respectively. Experiments have shown that this algorithm performs well in detecting locusts of different scales, scenes, and types, and can provide reference for real-time locust detection.

Abstract in Chinese

为了能实时准确地检测各类蝗虫目标，使得蝗虫检测更具有普适性，本文通过互联网爬虫及公有数据集IP102形成蝗虫数据集，提出了YOLOv7-MobileNetV3-CA的蝗虫目标检测算法。首先，为了降低模型参数量，使用MobileNetV3替换YOLOv7骨干网。其次，在特征加强后加入了CA（Coordinate Attention）注意力机制，以进一步提高蝗虫的检测精度。实验表明，蝗虫的检测精确率为95.96%，召回率92%，mAP为95.74%，F1为0.92，与YOLOv7相比，模型大小降低27%，mAP提高了4.48%。与YOLOv4、YOLOv4-MobileNetV3、YOLOv5、SSD算法相比，mAP分别提高了51.16%、26.81%、11.9%、11.75%。试验表明本算法对不同尺度、不同场景及不同种类的蝗虫检测效果较好，可以为蝗虫实时检测提供参考。