改进YOLOv5的国产光学影像辐射异常检测方法
Improving YOLOv5's domestic optical image radiation anomaly detection method
石一剑 1谭海 2钟旭辉1
作者信息
- 1. 辽宁工程技术大学测绘与地理科学学院,辽宁 阜新123000;自然资源部国土遥感卫星应用中心,北京100048
- 2. 自然资源部国土遥感卫星应用中心,北京100048
- 折叠
摘要
随着国产光学卫星数量不断增多,获取的卫星影像数据大规模增加,卫星获取并通过传感器校正处理的影像中存在大量的影像辐射异常问题,影像辐射质量是决定影像质量检验等级评价的重要因子,目前检查主要采用人机交互方式.针对目前光学影像质检存在影像辐射问题,本文提出了利用改进的YOLOv5深度学习网络对辐射异常区域进行目标识别的方法,将改进后的Light-BiFPN特征融合网络和ShuffleNetV2主干网络融入YOLOv5s.通过探索影像辐射异常原理,此网络能够精确判断辐射异常影像目标的范围,训练出的模型能通过锚框较好地检测出辐射问题的范围,为进一步模型的部署应用做好准备工作.
Abstract
With the continuous increase in the number of domestic optical satellites,the obtained satellite image data has shown a large-scale increase. There is a considerable proportion of image radiation anomalies in the images obtained by satellites and processed through sensor correction. Image radiation quality is an important factor determining the evaluation of image quality inspection level. Currently,its inspection mainly adopts human-computer interaction. In response to the current radiation problem in optical image quality inspection,an improved YOLOv5 deep learning network is proposed to identify targets in radiation abnormal areas. Integrate the improved light BiFPN feature fusion network and ShuffleNetV2 backbone network into YOLOv5s. By exploring the principle of image radiation anomalies,this network can accurately determine the range of targets in radiation anomaly images. The trained model can effectively detect the range of radiation issues through anchor frames,lay the foundation for further model deployment and application.
关键词
深度学习/轻量化/遥感图像/辐射异常/目标检测Key words
deep learning/light weight/remote sensing images/radiation anomaly/object detection引用本文复制引用
基金项目
高分辨率对地观测系统重大专项(42-Y30B04-9001-19121)
出版年
2024
NETL
NSTL
万方数据