| YOLOv5-LR：一种遥感影像旋转目标检测模型 |
| |
| 引用本文： | 高明明, 李沅洲, 马雷, 南敬昌, 周芊邑. YOLOv5-LR：一种遥感影像旋转目标检测模型[J]. 红外技术, 2024, 46(1): 43-51. |
| |
| 作者姓名： | 高明明 李沅洲 马雷 南敬昌 周芊邑 |
| |
| 作者单位： | 1.辽宁工程技术大学 电子与信息工程学院, 辽宁 葫芦岛 125105;2.中国科学院自动化研究所, 北京 100190;3.辽宁工程技术大学 电气与控制工程学院, 辽宁 葫芦岛 125105 |
| |
| 基金项目： | 国家自然科学基金青年科学基金（61701211）；辽宁省应用基础研究计划项目（2022JH2/101300275）；辽宁省应用基础研究计划项目（22-1083）；北京市科技计划项目（Z201100005820010） |
| |
| 摘 要： | 真实遥感图像中，目标呈现任意方向分布的特点，原始YOLOv5网络存在难以准确表达目标的位置和范围、以及检测速度一般的问题。针对上述问题，提出一种遥感影像旋转目标检测模型YOLOv5-Left-Rotation，首先利用Transformer自注意力机制，让模型更加注意感兴趣的目标，并且在图像预处理过程中采用Mosaic数据增强，对后处理过程使用改进后的非极大值抑制算法Non-Maximum Suppression。其次，引入角度损失函数，增加网络的输出维度，得到旋转矩形的预测框。最后，在网络模型的浅层阶段，增加滑动窗口分支，来提高大尺寸遥感稀疏目标的检测效率。实验数据集为自制飞机数据集CASIA-plane78和公开的舰船数据集HRSC2016，结果表明，改进旋转目标检测算法相比于原始YOLOv5网络的平均精度提升了3.175%，在吉林一号某星推扫出的大尺寸多光谱影像中推理速度提升了13.6%，能够尽可能地减少冗余背景信息，更加准确检测出光学遥感图像中排列密集、分布无规律的感兴趣目标的区域。
|
| 关 键 词： | 遥感图像 滑动窗口 注意力机制 旋转目标检测 YOLOv5 |
| 收稿时间： | 2022-11-18 |
| 修稿时间： | 2022-12-30 |
Multi-oriented rotation-equivariant network for object detection on remote sensing images |
| |
| GAO Mingming, LI Yuanzhou, MA Lei, NAN Jingchang, ZHOU Qianyi. YOLOv5-LR: A Rotating Object Detection Model for Remote Sensing Images[J]. Infrared Technology , 2024, 46(1): 43-51. |
| |
| Authors: | GAO Mingming LI Yuanzhou MA Lei NAN Jingchang ZHOU Qianyi |
| |
| Affiliation: | 1.College of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China;2.Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China;3.College of Electrical and Control Engineering, Liaoning Technical University, Huludao 125105, China |
| |
| Abstract: | In a real remote sensing image, the target is distributed in any direction and it is difficult for the original YOLOv5 network to accurately express the location and range of the target and the detection speed is moderate. To solve these problems, a remote sensing image rotating target detection model, YOLOv5-Left-Rotation, was proposed. First, the transformer self-attention mechanism was used to make the model pay more attention to the targets of interest. In addition, Mosaic data were enhanced in the image preprocessing, and the improved Non-Maximum Suppression algorithm was used in post-processing. Second, an angle loss function was introduced to increase the output dimensions of the network, and the prediction box of the rotating rectangle was obtained. Finally, in the shallow stage of the network model, a sliding window branch was added to improve the detection efficiency of large-sized remote sensing sparse targets. The experimental datasets were the self-made aircraft dataset CASIA-plane78 and the public ship dataset HRSC2016. The results show that the average accuracy of the improved rotating target detection algorithm is improved by 3.175% compared with that of the original model, and the reasoning speed is improved by 13.6% in a large multispectral image swept by a Jilin-1 satellite. It can optimally reduce the redundant background information and more accurately detect the densely arranged and irregularly distributed areas of objects of interest in optical remote sensing images. |
| |
| Keywords: | remote sensing images sliding window attention mechanism rotating object detection YOLOv5 |
|
| 点击此处可从《红外技术》浏览原始摘要信息 |
|
点击此处可从《红外技术》下载全文 |
|
