摘要
为解决在卫星遥感图像的多尺度目标检测中出现的背景混乱、小目标检测精度低、漏检率高等问题,提出一种用于卫星遥感图像的多尺度目标检测算法。在主干网络中使用通道和空间注意力模块,并重新设计特征融合网络,实现上采样-下采样-上采样的多重融合,并在其中加入通道权重参数,让网络更加关注重要的层次,实现不同层次特征信息的充分利用,使细节特征信息得到增强。在DIOR数据集中的实验结果表明,所提算法不仅显著提升对小目标的检测效果,而且提高对复杂场景中目标的检测精度,与YOLOv5m相比,对部分较小或者复杂的目标检测效果提升明显,精度提升4.5个百分点以上,整体精度提升3.1个百分点。
A multiscale object detection algorithm for satellite remote-sensing images is proposed to solve the problems of background confusion,low precision of small object detection,and high miss rate in multiscale object detection.The channel and spatial attention module is used in the backbone network,and the feature fusion network is redesigned to realize the multiple fusion of up-down-up sampling.The channel weight parameter is added to enable the network to pay more attention to critical channels,fully utilize different feature information levels,and enhance the detailed feature information.In a DIOR dataset,not only the detection effect of small objects but also the detection accuracy of objects in complex scenes is improved.Compared with that using YOLOv5m,the detection effect of some small or complex objects is improved significantly,the accuracy is improved by more than 4.5 percentage points,and the overall accuracy is improved by 3.1 percentage points.
作者
项建弘
陈振兴
王霖郁
Xiang Jianhong;Chen Zhenxing;Wang Linyu(College of Information&Communication Engineering,Harbin Engineering University,Harbin 150001,Heilongjiang,China;Key Laboratory of Advanced Ship Communication and Information Technology,Harbin Engineering University,Harbin 150001,Heilongjiang,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第2期342-351,共10页
Laser & Optoelectronics Progress
基金
国防通信抗干扰重点实验室(9140C020201120C02002)。
关键词
遥感
神经网络
多尺度目标检测
注意力机制
通道权重
特征融合
remote sensing
neural network
multiscale object detection
attention mechanism
channel weight
feature fusion
