摘要
在图像分析与解译中,遥感影像目标检测是一项基础性的工作。针对遥感影像目标尺度多样和背景复杂等问题,提出了一种多尺度空洞卷积特征融合检测器(MDCF2Det)来实现遥感目标的精确检测。首先,改进原始特征金字塔网络(FPN),用空洞卷积代替普通卷积,增大感受野;其次,增加从输入节点到输出节点的跳跃连接操作以充分地利用不同层级的语义和位置信息;最后,为了抑制噪声并突出前景,在区域候选网络前增加多维注意力机制模块,从而实现更精确的遥感影像目标检测。在DOTA和RSOD数据集上进行了实验,所提算法的mean average precision(mAP)分别达到了92.95%和73.39%。实验结果表明,所提算法能够有效提升遥感影像目标检测精度。
Object detection in remote sensing images is a fundamental task in image analysis and interpretation.We proposed a Multiscale Dilated Convolution Feature Fusion Detector(MDCF2Det)to achieve precise object detection in remote sensing by addressing the problems of multiscale objects and the complexity of the background.To begin,we improve the original feature pyramid network by replacing the general convolution with the dilated convolution to increase the receptive field.Second,to take full advantage of different levels of semantic and location information,we add a skip connection operation from the input node to the output node.Finally,to suppress the noise and highlight the foreground,we add the multi-dimensional attention model before the regional proposal network,to achieve more accurate object detection in remote sensing images.Experiments are carried out on the DOTA and RSOD datasets,and the proposed algorithm’s mean average precision reaches 92.95%and 73.39%respectively.The results show that the proposed algorithm can significantly improve the object detection accuracy of remote sensing images.
作者
田婷婷
杨军
Tian Tingting;Yang Jun(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,Gansu,China;National-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring,Lanzhou 730070,Gansu,China;Gansu Provincial Engineering Laboratory for National Geographic State Monitoring,Lanzhou 730070,Gansu,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2022年第16期417-425,共9页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61862039)
甘肃省科技计划(20JR5RA429)
兰州市人才创新创业项目(2020-RC-22)
兰州交通大学天佑创新团队(TY202002)。
关键词
遥感
目标检测
多尺度
卷积神经网络
特征融合
remote sensing
object detection
multiscale
convolutional neural network
feature fusion