基于特征自学习的高分光学遥感图像多类目标快速检测与识别

The existing sample learning based object detection and recognition methods face several key problems needed to be solved when processing massive high resolution remote sensing images: 1) A large number of training samples need to be labeled manually, which is time-consuming, laborious and low-efficient; 2) Image features are human-designed, without deeply mining intrinsic features from high resolution remote sensing images; 3) Most of the present approaches are customized for a particular type of object, which are difficult to scale up to deal with the simultaneous detection and recognition of a large number of object classes; 4) The detection and recognition speed is slow and is difficult to meet the requirements of high resolution remote sensing images processing. Aiming at the aforementioned problems and based on the applicant’s previous research work (research work have been published in international journals such as ISPRS JPRS, IJRS, etc.), this project will focus on the research of the following four key technologies: 1) Automatic labeling of large number of training samples based on semi-supervised learning; 2) Deeply mining and self-learning of intrinsic features of high resolution optical remote sensing images based on deep learning; 3) The construction of a multi-class object detection and recognition system for massive high resolution remote sensing images; 4) Visual saliency guided fast detection of potential object regions. 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.

已有的基于样本学习的目标检测与识别技术在处理海量高分遥感图像时面临着以下几个关键性的难点问题亟待解决：1）需要人们手工标注大量的训练样本，费时费力、效率低下；2）图像特征需要人工设计，没有充分挖掘高分遥感图像的内在信息；3）目标针对性强，难以直接移植到多类目标的检测和识别中；4）检测识别速度慢，难以满足高分遥感图像快速有效处理的要求。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ISPRS JPRS，IJRS等），重点研究以下四个关键技术：1）基于半监督学习的大量训练样本的自动标注；2）基于深度学习的高分光学遥感图像的内在本质特征的深度挖掘与自学习；3）面向海量高分遥感图像的多类目标检测与识别系统的构建；4）视觉显著性引导的目标潜在区域快速检测。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多类目标检测与识别问题，实现了基于半监督学习的正例标注和难例挖掘；设计了针对高分遥感图像特性的深度学习模型，实现了高分遥感图像高层特征的挖掘与自动提取算法；构建并验证了基于特征自学习的多类目标检测与识别系统。取得的代表性工作包括：1）提出基于模型优化集成的半监督学习模型；2）提出基于旋转不变卷积神经网络的高分遥感图像目标检测方法；3）提出基于费舍尔判别卷积神经网络的目标检测方法；4）提出基于判别地物目标引导的高分遥感图像场景语义理解模型；5）提出基于两阶段分步优化的目标模型共享与重构技术；6）提出基于弱监督学习和高级特征学习的高分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SCI国际期刊论文15篇（JCR 2区以上SCI论文10篇），国际会议论文6篇（含2篇CCF A类国际会议）；7篇论文入选ESI高被引论文，3篇论文入选ESI热点论文；申请国家发明专利1项，授权国家发明专利2项；协助培养博士生3名、硕士生4名；项目负责人2017年获“陕西省青年科技新星”称号；项目负责人获得2017年教育部自然科学二等奖（排名第三）和2017年陕西省高等学校科学技术奖一等奖（排名第五）。