摘要
针对无人机航拍图像中小目标样本多、拍摄目标背景复杂、可提取特征信息少的问题,提出一种改进YOLOv5s的轻量化无人机航拍小目标检测算法。首先,改进算法网络结构,增加两条特征信息传播路径,跨层级连接避免特征损失,同时同级前后连接补充特征信息,并在特征融合过程中加入空间注意力机制,提高模型对小目标区域的关注程度,保留充足的目标特征信息;其次,针对数据集的特点,将骨干网络中低层小目标检测层融入到特征金字塔网络和路径聚合网络结构中,增加一个检测极小目标的头部;最后,在预测过程中引入SIoU Loss定位损失函数,进一步加快模型收敛速度,提升模型检测能力及定位精度。将该算法在VisDrone2019数据集上进行实验,结果表明,改进后的模型mAP50达到了38.5%,较基线方法 YOLOv5s提高了5.9百分点,同时与主流的检测方法相比也取得更高的检测精度,对于小目标检测任务具有较好的性能。
Aiming at the problems of large number of small and medium-sized target samples,complex target background and little extracted feature information in UAV aerial photography images,an improved YOLOv5s lightweight UAV aerial photography small target detection algorithm was proposed.Firstly,the network structure of the algorithm is improved,and two feature information propagation paths are added to avoid feature loss through cross-level connection.At the same time,the feature information is supplemented by cross-level connection and spatial attention mechanism is added in the feature fusion process to improve the model's attention to small target regions and retain sufficient target feature information.Secondly,according to the characteristics of the data set,the low-level small target detection layer of the backbone network is integrated into the feature pyramid network and the path aggregation network structure,and a small target detection head is added.Finally,SIoU Loss is introduced into the prediction process to further accelerate the model convergence speed and improve the model detection ability and positioning accuracy.The proposed algorithm was tested on the VisDrone2019 dataset,which showed that the improved model mAP50 reached 38.5%,5.9 percentage points higher than that of the baseline method YOLOv5s,and also achieved higher detection accuracy than the that of mainstream detection methods,with better performance for small target detection tasks.
作者
魏雅丽
牛为华
WEI Ya-i;NIU Wei-hua(Department of Computer,North China Electric Power University,Baoding 071000,China)
出处
《计算机技术与发展》
2024年第2期53-59,共7页
Computer Technology and Development
基金
河北省重点研发计划项目(20310103D)。
