视觉注意机制引导的高分辨率遥感图像检索研究

As so far, people's capacity to effectively organize and manage the remote sensing image data has been far behind the growth rate of the remote sensing image data, which has been a cutting-edge challenging problem for remote sensing applications. Aiming at the increasing requirements for effective query and retrieval of the high resolution optical remote sensing images, this project introduces the visual attention mechanism and the hierarchical processing mechanism for visual information of human beings into the remote sensing image analysis field, to study efficient methods for high resolution remote sensing image retrieval. The main research content of this project includes: (1) creating a novel multiscale object-oriented visual attention model for high resolution remote sensing images, to extract the object of interest from remote sensing images efficiently and precisely; (2) building a novel convolutional neural network-based model for object feature extraction and semantic description, to extract object features and semantic descriptors from remote sensing images accurately and completely; (3) designing a novel feature-level and semantic-level similarity transduction model for image similarity measurement, to measure the visual similarity between remote sensing images comprehensively and precisely. In a word, guiding by the visual attention mechanism, this project focuses on the accurate feature and semantic descriptions and similarity measurements for high resolution remote sensing images, and strives to provide a practical solution for accurate and efficient high resolution remote sensing image retrieval.

当前，人们对于遥感图像数据的有效组织与管理能力已远远滞后于图像数据量的增长速度，该矛盾已成为遥感应用领域国际前沿的挑战性难题。本项目针对高分辨率光学遥感图像的高效查询与检索需求，将人类的视觉注意机制和视觉信息分层处理机制引入遥感图像分析领域，开展高分遥感图像的高效检索方法研究。研究内容包括：（1）构建面向多尺度地物目标的高分遥感图像视觉注意模型，以高效准确地提取遥感图像的感兴趣目标；（2）构建基于深度卷积神经网络的目标特征提取和语义描述模型，以准确完备地获取遥感图像的目标特征和语义描述；（3）构建基于特征和语义相似性转导的图像相似性度量模型，以准确全面地度量遥感图像之间的视觉相似程度。本项目以视觉注意机制为引导，探索高分遥感图像中准确的特征和语义描述方法及相似性度量方法，力图为符合人类视觉感知特性的高分遥感图像检索提供一条现实的解决途径。

随着遥感对地观测技术的不断进步，人类可获取的高分遥感影像数据量已经呈现出指数级增长的态势，而人们对于高分遥感影像的有效组织与管理能力已远远滞后于影像数据的增长速度，“数据海量、信息匮乏、知识难觅”的矛盾愈发凸显。针对传统遥感图像分析方法在高分遥感图像检索应用中的局限，本项目将人类视觉系统的视觉注意机制引入高分遥感图像分析中，围绕遥感图像检索的两个核心问题——“特征提取”和“相似性度量”，开展视觉注意机制驱动的高分遥感图像检索研究。项目首先构建新的视觉注意模型实现对遥感图像中多尺度对象的视觉显著性计算，以提取遥感图像的感兴趣区域；进而驱动遥感图像中主要地物对象的视觉显著特征提取，构建高分遥感图像视觉显著特征提取框架，以获取遥感图像的准确特征描述；最终通过多元视觉显著特征的相似性融合，构建基于多特征相似性融合的高分遥感图像检索方法，充分挖掘高分图像库整体特征空间的流形结构以及不同类型视觉显著特征之间的互补特性，以获得符合人类视觉感知特性的图像检索结果。项目通过大量实验证实了相关模型和方法的有效性，为高分遥感图像的高效智能检索提供了一条可期的解决思路。