基于判别和旋转不变卷积神经网络的高分遥感图像目标检测

The existing deep learning based object detection methods face several key problems needed to be solved when processing massive high resolution remote sensing images: 1) There is a lack of a large-scale high resolution remote sensing image benchmark, which severely limits the application of deep learning based methods to remote sensing big data; 2) There exist high within-class diversity and between-class similarity owing to the large number of object classes in remote sensing images, which increases the difficulty of object detection; 3) The objects in remote sensing images generally have arbitrary orientations and so how to effectively achieve rotation invariance is a big challenge. Aiming at the aforementioned problems and based on the applicant’s previous research work (research work have been published in conference such as CVPR and international journals such as Proceedings of the IEEE, IEEE TGRS, etc.), this project will first propose a large-scale, publicly available remote sensing image benchmark for object detection, and then focus on the researches of the following key technologies, including the design of discriminative convolutional neural networks, the design of rotation invariant convolutional neural networks, deep learning model optimization and multi-task learning collaborative optimization, to promote the state-of-the-art performance of object detection. This project is expected to significantly improve the utilization efficiency and automatic processing level of high resolution remote sensing images, and hence being very valuable for both civil and military applications.

基于深度学习技术的目标检测算法在处理高分遥感图像时面临着以下几个亟待解决的难点问题：1）缺乏大规模的高分遥感图像数据集，制约了深度学习技术在遥感大数据上的应用；2）遥感图像可识别的地物目标种类繁多，存在着目标的类内多样性和类间相似性问题，增加了目标检测的难度；3）遥感图像中的地物目标存在方向的多变性，如何实现旋转不变性是一个急需解决的问题。针对上述问题并结合申请人的前期研究工作（研究成果发表在领域顶级会议如CVPR和顶级国际期刊如Proceedings of the IEEE、IEEE TGRS等），本项目拟构建一个大规模、公开的高分遥感图像目标检测数据库，并深入研究判别式卷积神经网络构建、旋转不变卷积神经网络构建、深度学习模型优化与多任务学习协同优化等关键技术，提高高分遥感图像目标检测的技术水平。本项目的研究成果能显著提高高分遥感图像的利用效率和自动化处理水平，具有重要的民用和军事价值。

针对高分遥感图像目标检测面临的大规模数据集稀缺、类内多样性大、类间相似性高、目标方向多变等挑战，本项目从多个层面开展创新研究。取得的代表性工作包括：1）建立了大规模公开的光学遥感图像目标检测数据库DIOR；2）提出了一种基于旋转不变和费舍尔判别卷积神经网络的图像目标检测方法；3）提出了一种旋转不敏感、上下文特征增强的高分辨率遥感图像目标检测方法；4）提出了一种新颖的基于部件定位网络的细粒度目标识别方法；5）提出了基于动态特征融合网络的高分辨率遥感图像目标检测方法；6）提出了一种简单、有效的有向目标检测框架Oriented R-CNN；7）提出了多元信息监督的遥感图像有向目标检测方法；8）提出了高质量目标候选框生成与选择的弱监督目标检测方法。..本项目研究产生了一批高水平研究成果，发表高水平学术论文22篇，其中领域重要的国际期刊论文15篇，包括IEEE TPAMI在内的IEEE汇刊系列12篇，本领域重要国际会议论文4篇，中文期刊论文3篇；申请国家发明专利5项；培养博士生6名、硕士生8名；项目负责人获得2021年IEEE Transactions on Circuits and Systems for Video Technology最佳论文奖（IEEE TCSVT Best Paper Award）、2021年IEEE地球科学与遥感学会最有影响力论文奖（IEEE GRSS Highest Impact Paper Award）、2021年人工智能学会技术发明奖一等奖（R4）和2020年测绘科学技术奖一等奖（R6）。