摘要
视频卫星能获得高空间分辨率的视频信息,为运动目标的检测和分析提供有效数据支撑.然而,由于卫星视频图像中目标像素比例低、纹理细节不清晰、背景复杂等缺点,从卫星视频中检测运动目标存在很大困难.为此,本文以YOLOv8为骨干网络,提出了一种基于注意力机制与可变卷积神经网络的卫星视频运动目标检测算法.首先,设计C2f-DCN模块替换原模型骨干网络中的C2f模块,以提高模型对不同尺度目标的特征提取能力.其次,在检测头前添加Shuffle Attention轻量级注意力机制,在保证模型计算速度的前提下增强重要特征,加强通道间信息沟通提高模型特征融合能力.最后,为了提高模型的学习能力和推理效率,采用Inner-CIoU损失函数,并引入辅助边界框概念来解决卫星视频图像中目标像素比例小的问题.利用SAT-MTB卫星视频影像数据集进行对比实验,实验结果表明本文算法的精确度、召回率、mAP50:95和F1分数分别为75.3%、62.8%、34.9%和68.48,相较于原始YOLOv8n网络,上述指标分别提高了11.6%、4.2%、3.0%和7.44,验证了本文方法的有效性和优越性.
Video satellites can obtain high spatial resolution video information,providing effective data support for the detection and analysis of moving targets.However,due to the disadvantages of low target pixel proportion,unclear texture details,and complex background in satellite video images,there are significant difficulties in detecting moving targets from satellite videos.Thus,on the basis of the backbone network of YOLOv8,this paper proposes a new detection method of motion targets in satellite video based on deformable convolutional neural network and Shuffle Attention.Firstly,a C2f-DCN module is designed to replace the C2f module in the original model backbone network for improving the model's ability of extracting features from targets with different types and different scales.Secondly,a lightweight Shuffle Attention mechanism is added in front of the detection head to strengthen important features while ensuring the computational speed of the model,enhancing information communication between channels,and improving the model's feature fusion ability.Finally,to improve the learning ability and inference efficiency of the model,the Inner-CIoU loss function is adopted,and the concept of auxiliary bounding boxes is introduced for solving the problem of small proportion of target pixels in satellite video images.Comparative experiments are conducted using the SAT-MTB satellite video image dataset,and the experimental results show that the accuracy,recall,mAP50:95,and F1 scores of the algorithm are 75.3%,62.8%,34.9%,and 68.48,respectively.Compared with the original YOLOv8n network,above indexes are improved by 11.6%,4.2%,3.0%,and 7.44.Thus,the effectiveness and superiority of the proposed method is verified.
作者
马洲俊
陈锦铭
刘浩林
张卡
Ma Zhoujun;Chen Jinming;Liu Haolin;Zhang Ka(State Grid Jiangsu Electric Power Co.,Ltd.,Nanjing 210019,China;Electric Power Research Institute,State Grid Jiangsu Electric Power Co.,Ltd.,Nanjing 211103,China;School of Geography,Nanjing Normal University,Nanjing 210023,China)
出处
《南京师大学报(自然科学版)》
北大核心
2025年第4期78-86,共9页
Journal of Nanjing Normal University(Natural Science Edition)
基金
国网江苏省电力有限公司科技项目(J2023121)。
