基于超分辨率迁移学习的卫星遥感图像鲁棒车辆检测研究

Coming with the era of massive satellite remote sensing data, the research on vehicle detection in satellite remote sensing images has important applications in the fields of battlefield directing and intelligent transportation. However, the challenges in remote sensing imagery, such as the lack of labeling, the small object size, as well as the image noise, have been restricting the direct usage of traditional vehicle detection techniques in this field. In this project, we target at tackling above challenges towards robust vehicle detection in satellite remote sensing images. Our key innovation is a transfer learning framework, which "transfers" reliable cutting-edge vehicle detectors from aerial remote sensing images to satellite remote sensing images. Our project covers three seamlessly-coupled points: (1) a high-low resolution vehicle detector "transfer" framework; (2) a super-resolution reconstruction algorithm for the targeted low-resolution region via sparse coding; and (3) an robust vehicle detection via exemplar search. The output of our project is a robust vehicle detection framework for satellite remote sensing images. This framework has the potential to benefit the real-world vehicle analysis and surveillance applications used in military and transportation systems, push forward the theoretical breakthrough in remote sensing image understanding and analysis, as well as further advancing the theoretical development in the related fields.

伴随着海量卫星遥感数据时代的来临，面向卫星遥感图像的车辆目标检测在战场指挥与智能交通等领域具有重要应用。然而，卫星遥感图像的标注信息不足、小目标、及噪声干扰等问题，一直困扰着传统的车辆目标检测技术在遥感领域的直接应用。本项目针对这三个问题展开卫星遥感图像中的鲁棒车辆目标检测研究。拟建立基于迁移学习的一体化框架，探索将卫星遥感图像中的车辆目标检测“迁移”至航空遥感图像下进行解决，包含三个相互耦合研究点：(1) 基于高-低分辨率遥感图像的迁移学习框架; (2) 基于稀疏编码的目标区域超分辨率重构; (3) 基于样例搜索的鲁棒目标检测。本项目拟形成一套面向不同尺度卫星遥感图像的一体化鲁棒目标车辆检测框架，推动遥感图像内容理解与分析理论的突破，垫定车辆行为分析与监控等国防、交通重大应用的技术基础，并促进相应理论研究在相关学科应用领域的发展。

伴随着海量卫星遥感数据时代的来临，面向卫星遥感图像的车辆目标检测在战场指挥与智能交通等领域具有重要应用。然而，卫星遥感图像的标注信息不足、小目标、及噪声干扰等问题，一直困扰着传统的车辆目标检测技术在遥感领域的直接应用。本项目针对这三个问题展开卫星遥感图像中的鲁棒车辆目标检测研究。项目执行期间, 建立了基于迁移学习的一体化框架，完成了将卫星遥感图像中的车辆目标检测“迁移”至航空遥感图像下进行解决，包含三个相互耦合研究点：(1) 基于高-低分辨率遥感图像的迁移学习框架; (2) 基于稀疏编码的目标区域超分辨率重构; (3)基于样例搜索的鲁棒目标检测。本项目完成了一套面向不同尺度卫星遥感图像的一体化鲁棒目标车辆检测框架，并发表了多篇学术论文, 该研究成果推动遥感图像内容理解与分析理论的突破，垫定车辆行为分析与监控等国防、交通重大应用的技术基础，并促进相应理论研究在相关学科应用领域的发展。