关键词: 无人机图像, SimAM, GSconv, 小目标检测头

Abstract: Given the many problems with small targets in images in drone scenes，this paper proposed an improved YOLOv5（I-YOLOv5）model based on the traditional YOLOv5 model. Firstly，a small-target detection head was added to improve the network′s ability to represent small targets. Secondly，the SimAM attention mechanism was added to make the network more focused on small-target objects. Thirdly，it changed the coupled detection head in YOLOv5 was changed to a decoupled detection head to speed up model training and effectively improve model accuracy. Finally，the CBS structure in the neck of the original model was modified to the GSconv structure to reduce model parameters and improve model accuracy. On the Visdrone2019 data set，the I-YOLOv5 model outperformed the original YOLOv5 model by 6.6% and 4.2% in mAP50 and mAP50∶95. This confirms that our proposed model has certain advancements in the field of small-target UAV images.

Key words: UAV images, SimAM, GSconv, small-target detection head