基于自步式弱监督学习的高分遥感细粒度目标检测与识别

The existing deep learning based object detection and recognition methods face two key problems needed to be solved when performing fine-grained object detection and recognition from massive optical high resolution remote sensing images under weakly supervised settings: 1) The lack of efficient supervision information of weakly supervised learning makes the object detection and recognition system be very sensitive to initialization and very difficult to achieve robust training; 2) Both the complex backgrounds and subtle differences between the fine-grained object categories cause serious interference to accurately and effectively extract discriminative key features of objects. Aiming at the aforementioned problems and based on the applicant’s previous research work (research work have been published in international journals such as TGRS, etc.), this project will focus on the research of the following three key technologies: 1) Automatic initial labeling of training samples based on spectral rotation co-clustering; 2) Discriminative deep feature learning for fine-grained objects constrained by hierarchical structure information of category labels; 3) Effective training of object detection and recognition system based on self-paced learning. This research is expected to significantly improve the utilization efficiency and automatic processing of high resolution remote sensing images, having an important value for both civil and military applications.

已有的基于深度学习的目标检测与识别技术在弱监督条件下处理面向海量高分光学遥感图像的细粒度目标检测与识别时面临着以下两个关键性的难点问题亟待解决：1）弱监督学习框架下有效监督信息的缺失，使得目标检测与识别系统对初始化非常敏感，很难实现系统的鲁棒训练；2）高分遥感图像的复杂背景以及细粒度目标类别之间的细微差异，给精确、有效地提取更具判别性的目标关键特征造成严重的干扰。本项目针对上述问题并结合申请人的前期研究工作（研究成果发表在遥感图像处理领域重要国际期刊，如TGRS等），重点研究以下三个关键技术：1）基于谱旋转联合聚类的训练样例初始标注；2）融合类别标签层级结构信息的细粒度目标判别性深度特征学习；3）基于自步学习的目标检测与识别系统训练。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。

本项目围绕高分光学遥感图像弱监督细粒度目标检测与识别问题，实现了弱监督条件下训练样本的初始标注；设计了目标类别层级结构引导的判别式深度学习模型，实现了高分遥感图像细粒度目标的判别性特征学习与自动提取；构建并验证了基于自步式弱监督学习的细粒度目标检测识别系统。取得的代表性工作包括：1）提出基于部件卷积神经网络的细粒度目标识别方法；2）提出基于动态课程学习的弱监督遥感图像目标检测方法；3）提出基于渐进式实例精炼的弱监督遥感图像目标检测方法；4）提出基于多重上下文感知的弱监督遥感图像目标检测方法。本项目研究产生了一批高水平研究成果，发表JCR 2区以上SCI国际期刊论文8篇，国际会议论文4篇；1篇论文荣获陕西省自然科学论文二等奖，1篇论文入选2018中国百篇最具影响国际学术论文；申请国家发明专利3项；协助培养博士生4名、硕士生4名。