陈少波,雷泽人.基于GA-RoI Transformer的遥感图像任意方向目标检测[J].中南民族大学学报自然科学版,2022,41(3):312-318
基于GA-RoI Transformer的遥感图像任意方向目标检测
Oriented object detection in remote sensing image based on GA-RoI Transformer
  
DOI:10.12130/znmdzk.20220309
中文关键词: 遥感图像  目标检测  导向锚  有向边界框
英文关键词: remote sensing image  object detection  guided anchoring  oriented bounding-box
基金项目:中央高校基本科研业务费专项资金资助项目(CZY18002)
作者单位
陈少波 中南民族大学 电子与信息工程学院武汉 430074 
雷泽人 中南民族大学 电子与信息工程学院武汉 430074 
摘要点击次数: 18
全文下载次数: 23
中文摘要:
      遥感图像中的目标多呈现出方向上的任意性,导致遥感图像中感兴趣目标的检测难度大大增加.现有主流目标检测方法都是基于水平候选锚框的,现有方法通过对锚框添加旋转角度来解决任意方向目标检测问题,但这使得候选锚框的数量激增,导致算法计算开销过大.提出了一种基于GA-RoI Transformer(Guided Anchoring-RoI Transformer)的遥感图像任意方向目标检测方法:在语义特征的引下预测出感兴趣目标的中心位置、尺度和长宽比,并将它们相结合得到水平的高质量锚框作为候选框;通过RRoI学习器将HRoI转换成RRoI;从RRoI中提取旋转不变特征,来促进后续的分类和回归任务.在DOTA数据集上进行了仿真,可达78.17%,高于基线5.1%;且对于拥有极端形状的物体目标,检测性能优于现在的很多方法.
英文摘要:
      Most of the objects in remote sensing images show arbitrariness in direction, which greatly increases the difficulty of detecting interested objects in remote sensing images. The existing mainstream object detection methods are all based on horizontal proposals. Using the idea of adding rotation angle to the anchor, many methods have been designed to solver the problem of oriented object detection; however, the number of candidate anchor is greatly increased in this type of method, and what follows is that the computational overhead will be huge. In this paper, A novel method for oriented object detection in remote sensing images based on GA-RoI Transformer (Guided Anchoring-RoI Transformer) is presented. Firstly, under the guidance of semantic features, center position、scale and aspect ratio of the target of interest are predicted, then the horizontal high-quality Anchores will be generated by combing the aforementioned prediction information. Secondly, Horizontal Region of Interest (HRoI) is transformed to Rotated Region of Interest (RRoI) by a RRoI learner. Finally, the rotation-invariant features are extracted from RRoI for boosting subsequent classi?cation and regression tasks. Experiments on DOTA are conducted, the proposed method get the of 78.17%, 5.1% higher than the baseline, and the detection performance is better than many existing methods for those extremely tall or wide objects.
查看全文   查看/发表评论  下载PDF阅读器
关闭