融合判别深度特征和物体属性特征的高分遥感图像场景分类

Scene classification of high resolution remote sensing images is a crucial technology faced for information mining and extraction from remote sensing big data. The existing deep learning based scene classification algorithms are faced with several key problems needed to be solved when processing massive high resolution remote sensing images, such as the small scale of the publicly available high resolution remote sensing image benchmark datasets, the high within-class diversity and between-class similarity held by many remote sensing image scene classes, and there existing a large number of scene classes dominated by salient objects. Aiming at the aforementioned problems, this project will first propose a large-scale high resolution remote sensing image benchmark for scene classification, and then focus on the researches of the following key technologies, including the training of deep metric learning-based convolutional neural networks and discriminative deep feature extraction, unsupervised training of salient object detection models and object attribute feature extraction, and the fusion of discriminative deep features and object attribute features based on metric learning, to improve the state-of-the-art performance of scene understanding. The applicant has published more than 30 papers in top conferences such as CVPR and ACM MM and international journals such as IEEE TCSVT, IEEE TCYB, IEEE TGRS, etc., which has formed solid foundation for this project.

高分遥感图像场景分类，是遥感大数据信息挖掘与提取的一项关键技术。已有的基于深度学习的场景分类算法在处理高分遥感图像时面临着一些亟待解决的难点问题，如公开的高分遥感图像数据集规模较小、高分遥感图像场景类别存在着高度的类内差异性和类间相似性、高分遥感图像存在大量的以显著物体为主导的语义场景类别等。针对上述问题，本项目拟构建一个大规模的高分遥感图像场景分类数据库，并深入研究深度度量卷积神经网络训练与判别深度特征提取、无监督显著物体检测模型训练及物体属性特征提取、基于度量学习的判别深度特征与物体属性特征融合等关键技术，提高高分遥感图像场景理解的技术水平。本项目申请人近几年在领域顶级会议如CVPR、ACM MM和重要国际期刊如IEEE TCSVT、IEEE TCYB、IEEE TGRS等发表30余篇论文，具有良好的研究基础。

针对高分辨率遥感图像场景分类面临的挑战和难点问题，本项目从多个方面开展创新研究。取得的代表性工作包括：1）针对高分遥感图像场景识别中的类内多样性和类间相似性问题，提出了一种基于深度度量卷积神经网络的图像分类方法；2）针对高光谱图像分类面临的“同物异谱”和“同谱异物”现象，提出了基于深度迁移学习和度量学习的高光谱图像分类方法；3）提出了一种和全局特征互补的基于区域响应排序的特征表达和特征学习方法；4）提出一种基于流形空间约束的小样本遥感图像分类方法；5）提出一种基于旋转不变潜在语义模型的遥感图像目标检测方法；6）系统综述了基于深度学习的高分辨率遥感图像场景分类的研究进展；7）针对样本数据稀缺问题，提出了一种基于双原型网络的小样本遥感场景分类算法；8）针对遥感图像小样本分类任务面临的深度学习模型泛化能力差的难题，提出了判别学习自适应分类网络。..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15篇，其中领域重要的国际期刊论文11篇，本领域重要国际会议论文4篇；申请国家发明专利7项；培养博士后1名、博士生4名、硕士生8名。