摘要
针对无人机航拍图像中小目标样本多、可提取特征信息少等问题,提出一种基于改进YOLOv7的无人机航拍小目标检测算法。首先,将骨干网络中低层小目标检测层融入聚合网络结构中,增加一个检测极小目标的头部;其次,将通道-空间注意力模块加入主干网络的特征提取过程中,同时引入特征融合中改进原有连接处的特征融合方式,自适应生成各层级特征图输出权重来动态优化特征图的表达能力;最后,在预测过程中引入SIoU Loss定位损失函数,提升模型检测能力及定位精度。实验结果表明,改进后的模型mAP50达到了52.6%,较基线算法YOLOv7提高了2.8个百分点,与主流的检测算法相比也取得了更高的检测精度,对于小目标检测任务具有较好的性能。
To solve the problems of a large number of small target samples and inadequate extractable feature information in UAV aerial photography images,a small target detection algorithm based on the improved YOLOv7 is proposed.Firstly,the low-level small target detection layer in the backbone network is integrated into the aggregation network structure,and a header is added to detect extremely small targets.Secondly,channel-spatial attention modules are added to the feature extraction process of the backbone network.At the same time,the feature fusion mode of improving the original connection in feature fusion is introduced,and the output weight of the feature graphs of each level is generated adaptively to dynamically optimize the representation ability of the feature graphs.Finally,the positioning loss function of SIoU is introduced into the prediction process to improve the model’s detection ability and positioning accuracy.Experimental results show that the mAP50 of the improved model reaches 52.6%,which is 2.8 percentage points higher than that of the baseline YOLOv7 algorithm.The improved model also achieves higher detection accuracy than the mainstream detection methods,and has better performance in small target detection.
作者
牛为华
魏雅丽
NIU Weihua;WEI Yali(Department of Computer,North China Electric Power University,Baoding 071000,China;Engineering Research Center of Intelligent Computing for Complex Energy Systems,Ministry of Education,Baoding 071000,China)
出处
《电光与控制》
CSCD
北大核心
2024年第1期117-122,共6页
Electronics Optics & Control
基金
中国高校基本科研业务费专项资助项目(2017MS156)。