摘要
针对航拍图像小目标数量多、尺度变化大的情况,对YOLOv5s结构做了4点改进,提出了LL-YOLOv5s网络,主要包括删除主干网络的下采样32倍特征层和仅使用2个检测头等改进措施,使模型专注于小目标的检测。在DOTA-v1.5航拍数据集中,改进后的LL-YOLOv5s的mAP提高了2.9个百分点。提出了一种简单高效的自注意力模块,即qv自注意力模块,将其添加到LL-YOLOv5s的第1个检测头之前,在增加极少计算量的基础上,mAP进一步提高了0.9个百分点。实验发现,将qv自注意力模块和卷积层连用,mAP又提高了0.4个百分点;相比于YOLOv5s,改进模型的参数量大大减小,运算量少量增加,检测精度明显提升。
In view of the large number of small targets with wide scale scope in aerial images,the YOLOv5s structure is improved from four aspects,and the LL-YOLOv5s network is proposed.The improvement measures mainly include deleting the feature layer of 32 times undersampling in the backbone network and using only two detection heads,so that the model can focus on the detection of small targets.The LL-YOLOv5s improves the mAP by 2.9 percentage points on the DOTA-v1.5 aerial dataset.Then,a simple and efficient self-attention module called qv self-attention module is proposed,which is added to the position before the first detection head of LL-YOLOv5s,and the mAP is further improved by 0.9 percentage points at the cost of adding a small amount of calculation.It is found that the combination of qv self-attention module with convolution layer further improves the mAP by 0.4 percentage points.Compared with YOLOv5s,the improved model greatly reduces the number of parameters and improves the detection accuracy significantly at the cost of adding a small amount of calculation.
作者
唐田尧
石永康
王浩然
吕玉龙
TANG Tianyao;SHI Yongkang;WANG Haoran;LYU Yulong(School of Intelligent Manufacturing Modern Industry,Xinjiang University,Urumqi 830000,China)
出处
《电光与控制》
CSCD
北大核心
2024年第5期89-94,共6页
Electronics Optics & Control
基金
国家自然科学基金(51965056)
自治区高层次人才项目(100400027)。