地球信息科学学报 ›› 2022, Vol. 24 ›› Issue (3): 572-582.doi: 10.12082/dqxxkx.2022.210369

• 遥感科学与应用技术 • 上一篇    下一篇

基于YOLOv5算法的飞机类型光学遥感识别

龙怡灿1,2(), 雷蓉1,*(), 董杨1, 李东子1, 赵琛琛3   

  1. 1.中国人民解放军战略支援部队信息工程大学,郑州 450001
    2.中国人民解放军75833部队,广州 510000
    3.中国人民解放军61363部队,西安 710054
  • 收稿日期:2021-07-03 修回日期:2021-08-24 出版日期:2022-03-25 发布日期:2022-05-25
  • 通讯作者: *雷 蓉(1974—),女,重庆人,教授,博士,主要从事数字摄影测量与遥感、航天遥感工程方向研究。 E-mail: leirong@163.com
  • 作者简介:龙怡灿(1990—),女,湖南保靖人,硕士生,主要从事数字摄影测量与遥感方向研究。E-mail: long_yican@126.com
  • 基金资助:
    高分遥感测绘应用示范系统(二期项目)(42-Y30B04-9001-19/21);国家自然科学基金项目(41401534);国家自然科学基金项目(41971427)

YOLOv5 based on Aircraft Type Detection from Remotely Sensed Optical Images

LONG Yican1,2(), LEI Rong1,*(), DONG Yang1, LI Dongzi1, ZHAO Chenchen3   

  1. 1. PLA Strategic Support Force Information Engineering University, Zhengzhou 450001, China
    2. Troops 75833, Guangzhou 510000, China
    3. Troops 61363, Xi'an 710054, China
  • Received:2021-07-03 Revised:2021-08-24 Online:2022-03-25 Published:2022-05-25
  • Supported by:
    The High Resolution Remote Sensing, Surveying and Mapping Application Demonstration System (Phase Ⅱ)(42-Y30B04-9001-19/21);National Natural Science Foundation of China(41401534);National Natural Science Foundation of China(41971427)

摘要:

飞机类型检测是遥感影像分析领域的研究热点,在机场监控和情报分析等应用中起着重要的作用。其中,深度学习方法作为遥感影像分析领域广泛应用的手段之一,在飞机类型检测任务中仍面临许多问题,如使用未公开的自制数据集、实验复现困难、无法验证泛化能力等。同时,光学遥感影像易受光照条件、云雨变化等因素影响,使检测任务更加困难。为了解决这些问题,本文首先利用MTARSI数据集对样本进行筛选,再结合Google Earth等开源方法收集飞机影像,采用随机旋转、改变亮度等方法构建新的飞机类型检测数据集。其次,采用YOLOv5作为基础网络框架,针对其多层卷积和池化操作可能会削弱或完全丢失飞机特征的问题,进行多尺度优化训练,有效检测飞机类型特征。最后,利用跨数据集验证模型的泛化能力。实验结果表明,本文方法能准确、有效地检测出光学遥感影像中的飞机的具体类型,具有较强的鲁棒性和泛化能力,跨数据集进行飞机类型检测正确率达到82.12%,可为智能化的飞机目标语义分析、星上应用等研究提供技术支撑。

关键词: 飞机检测, 类型检测, 遥感影像, 深度学习, 数据增广, YOLOv5, 多尺度优化, 多时相检测

Abstract:

Automatic detection of aircrafts plays an important role in airport monitoring and intelligence analysis. Currently, there are various and mature methods for aircraft detection, but the aircraft type detection is still facing many problems, such as the use of unpublicized self-made datasets, the difficulty to reproduce experiment, and the inability to verify generalization ability. Therefore, detecting aircraft type quickly and accurately is still a hotspot in the field of remote sensing image analysis. The traditional detection methods are complicated in processes, poor in robustness and generalization, and cannot detect the specific type of aircrafts. In recent years, deep learning methods have been widely applied in the field of computer vision. Compared with the two-stage algorithm, YOLO, as a one-stage algorithm, rejects the steps of multiple regression, includes only a convolution network, and regards the detection problem as the regression problem of image classification and candidate box parameters. However, multi-layer convolution and pooling may weaken or completely lose aircraft features, making it challenging to extract practical features. Meanwhile, remote sensing images are susceptible to light conditions, seasonal changes, cloud occlusion, noise, and other factors, which makes the detection task harder. In order to solve these problems, this paper firstly used the MTARSI dataset to screen samples and then collected aircraft images from open-source methods such as Google Earth using random rotation, changed brightness, added noise, and other methods to form a new aircraft type detection dataset. Secondly, multi-scale adjustment and training were carried out based on YOLOv5. Finally, an across-dataset was used to identify the aircraft in the optical remote sensing images, which could verify the model’s generalization ability. The experimental results show that the method can accurately and effectively detect the number, location, and type of aircraft in the optical remote sensing images and has strong robustness and generalization ability. The accuracy of type detection reached 82.12% in the across-dataset, which can provide technical support for intelligent aircraft semantic analysis and on-board application research.

Key words: aircraft detection, type detection, remote sensing image, deep learning, data augmentation, YOLOv5, multi-scale optimization, multitemporal detection