基于Multitask-YOLO网络的卫星帆板ISAR图像快速分割
Fast Segmentation of Solar Panels in Satellite ISAR Images Using a Multitask-YOLO Network
姚雨晴 1汪玲 1王莲子 1张弓 1吴斌 2朱岱寅1
作者信息
- 1. 南京航空航天大学电子信息工程学院雷达成像与微波光子技术教育部重点实验室,南京 211106,中国
- 2. 上海宇航系统工程研究所,上海 201109,中国
- 折叠
摘要
随着空间技术的飞速发展,空间态势感知能力需求不断增加.与传统光学传感器相比,逆合成孔径雷达(Inverse synthetic aperture radar,ISAR)具有全天候、远距离高分辨率成像的能力,且成像不受光照条件的影响.此外,空间态势感知系统需要对周围航天器进行准确的评估,因此对空间目标部件识别能力的需求日益迫切.本文提出了一种基于YOLOv5结构的Multitask-YOLO网络,用于卫星ISAR图像中卫星帆板的识别和分割.首先,本文添加了分割解耦头来实现网络的分割功能.然后用空间金字塔池快速算法(Spatial pyramid pooling fast,SPPF)和距离交并比算法(Distance intersection over union,DIoU)代替原有结构,避免图像失真,加快收敛速度.通过在通道中引入注意机制,提高了分割和识别的准确性.最后使用模拟卫星的ISAR图像进行实验.结果表明,所提出的Multitask-YOLO网络高效、准确地实现了部件的识别和分割.与其他的识别和分割网络相比,该网络的平均精度(mean Average precision,mAP)和平均交并比(mean Intersection over union,mIoU)提高了约5%.此外,该网络的运行速度高达16.4 GFLOP,优于传统的多任务网络的性能.
Abstract
With the rapid development of space technology,the situation awareness ability of spacecraft is increased.As compared to the optical sensors,inverse synthetic aperture radars(ISARs)have the capability of high-resolution imaging in all day from far range regardless of the light condition.Furthermore,the component recognition is much desired by the accurate evaluation of the threat degree of surrounding spacecrafts.In this paper,we propose a multitask-you only look once(Multitask-YOLO)network based on the YOLOv5 structure for recognition and segmentation of solar panels of satellite ISAR images.Firstly,we add a segmentation decoupling head to introduce the function of segmentation.Then,the original structure is replaced with spatial pyramid pooling fast(SPPF)to avoid image distortion,and with distance intersection over union(DIoU)to speed up convergence.The accuracy of segmentation and recognition is improved by introducing an attention mechanism in the channels.We perform the experiments using simulated satellite ISAR images.The results show that the proposed Multitask-YOLO network achieves efficient and accurate component recognition and segmentation.As compared to typical recognition and segmentation networks,the proposed network exhibits an approximate 5%improvement in mean average precision(mAP)and mean intersection over union(mIoU).Moreover,it operates at a higher speed of 16.4 GFLOP,surpassing the performance of traditional multitask networks.
关键词
Multitask-YOLO/空间目标/逆合成孔径雷达图像/目标识别与分割Key words
Multitask-YOLO/space objects/inverse synthetic aperture radar(ISAR)images/target recognition and segmentation引用本文复制引用
基金项目
Shanghai Aerospace Science and Technology Innovation Foundation(SAST 2021-026)
Fund of Prospective Layout of Scientific Research for Nanjing University of Aeronautics and Astronautics(NUAA)()
出版年
2024
国家科技期刊平台
NSTL
万方数据