恶劣环境下遥感图像小目标鲁棒检测算法研究

Coming with the construction of national remote sensing satellite networks like BeiDou and Gaofen, the recognition of small targets like vehicles and ships for satellite remote sensing images is playing an increasingly important role in battlefield commanding and intelligent transportation. However, on one hand the observation quality is highly affected by the imaging hardware restriction and the severe weather condition. On the other hand, precise labels are severely insufficient for multi-resolution remote sensing images. Therefore, traditional objet detection approaches in computer vision cannot be directly applied into such real-world applications. This proposal studies the fundamental theory and key techniques for tiny object detection in satellite images under severe environment. In particular, this proposal presents a unified learning framework based on generative adversarial networks, which jointly optimizes both the haze removal and the tiny object detector training components, as well as simulating virtual objects to enlarge the training corpus by the generator network. Relying on the platform of the applicant, the theoretical achievements of this proposal would be directly deployed into practical requirements like Gaofen earth observation. Correspondingly, key breakthroughs can be expected in key techniques of sensitive object surveillance for satellite remote sensing, as well as theoretical research of computer vision and related fields.

伴随着北斗与高分等一系列国家级遥感卫星网络的构建，面向卫星遥感图像的车辆舰船等小目标检测在战场指挥与智能交通等领域扮演着愈发重要的作用。然而，一方面成像硬件限制和恶劣气象条件严重影响了目标观测质量，另一方面多分辨率遥感图像中精确的标注信息严重缺乏，因此传统的计算机视觉目标检测方法始终难以在该问题中进行实用化部署。本项目拟研究恶劣环境下的遥感图像小目标检测的基础理论与关键技术。项目拟提出一套基于生成对抗网络的一体化学习框架，协同优化雾霾干扰去除和小目标检测两个网络训练模块，并通过生成网络进行跨分辨率虚拟样本的自动生成。依托申请人科研平台，本项目的理论成果拟直接对接高分对地观测等业务需求，推动卫星遥感敏感目标监控等关键技术突破，并促进相应理论研究在计算机视觉等学科领域的发展。

伴随着北斗与高分等一系列国家级遥感卫星网络的构建，面向卫星遥感图像的车辆舰船等小目标检测在战场指挥与智能交通等领域扮演着愈发重要的作用。然而，一方面成像硬件限制和恶劣气象条件严重影响了目标观测质量，另一方面多分辨率遥感图像中精确的标注信息严重缺乏，因此传统的计算机视觉目标检测方法始终难以在该问题中进行实用化部署。本项目研究了恶劣环境下的遥感图像小目标检测的基础理论与关键技术，提出了一套基于生成对抗网络的一体化学习框架，协同优化雾霾干扰去除和小目标检测两个网络训练模块，并通过生成网络进行跨分辨率虚拟样本的自动生成。项目组在项目执行期间在计算机视觉、人工智能领域的权威期刊和顶级国际会议上共发表学术论文19篇，本项目的理论成果直接对接高分对地观测等业务需求，推动了卫星遥感敏感目标监控等关键技术突破，并促进了相应理论研究在计算机视觉等学科领域的发展。