【作者】 程起敏；

【导师】 徐冠华； 杨崇俊；

【作者基本信息】 中国科学院研究生院（遥感应用研究所） ， 地图学与地理信息系统， 2004， 博士

【摘要】 遥感影像数据作为数字地球各项重大计划建设中的基础数据，其快速浏览和高效检索是遥感影像信息提取和共享的重要手段；基于内容的图像检索技术作为从试图理解图像内容的角度有效管理和利用图像数据库中信息的手段，已经成为图像数据库、计算机视觉等领域的研究热点和未来信息高速公路、数字图书馆等重大项目中的关键技术，为解决大型遥感影像数据库的信息提取难题提供了新的契机。然而，遥感影像数据的多样性、复杂性和海量性无疑对基于内容的遥感影像库检索、特别是Web环境下的应用提出了巨大的挑战。基于内容的遥感影像库检索技术是遥感影像处理、图像数据库技术、计算机视觉、模式识别等领域相结合的国际前沿课题，对于促进遥感影像信息的提取和共享，具有十分重要的理论意义和实用价值。本文旨在针对基于内容的遥感影像库检索的关键技术，提出一些创新性思路和方法，并分别从理论和技术的角度对其价值和实用性予以分析和验证。本文的主要内容包括： ①系统地归纳和分析了目前基于内容的遥感影像检索的国内外相关研究计划及主要研究成果，总结了基于内容的遥感影像库检索涉及到的各项关键技术，分析了基于内容的图像检索技术在遥感领域的应用与在多媒体和医学等其它领域的应用相比所面临的困难，并指出了解决问题的出发点。 ②提出遥感影像纹理特征的多尺度描述以及合理的遥感影像数据分块组织策略，是支持基于纹理特征的遥感影像库检索的关键。从分析传统的遥感影像数据分块组织策略的弊端以及小波变换在遥感影像纹理特征提取中的意义入手，提出了基于Nona-tree结构和小波直方图创建同质纹理特征库的方法和检索实现方法，既利用了Nona-tree在分块组织遥感影像数据时能够在检索精度、检索效率和存储空间之间达到较好平衡的优势，同时利用了小波直方图提取遥感影像多尺度纹理特征的高效、快捷的特性，并通过实验进行了充分的验证。此外，本文将多进制小波理论引入遥感影像纹理特征描述，从理论上分析了多进制小波与常规的二进制小波相比在描述图像特征方面的优势，提出了采用多进制小波直方图和多进制快速小波直方图进行遥感影像纹理特征提取以及相似性度量的方法，通过实验对比了基于多进制小波直方图和二进制小波直方图的遥感影像纹理特征检索精度和检索效率。 ③结合目前基于形状的图像检索技术的研究现状以及遥感影像上目标识别的困难，探索性提出在相关领域技术发展的现有水平下实现基于目标形状的遥感影像检索的可行方案，避开了由于目前遥感影像目标全自动识别的困难而对遥感影像形状检索造成的影响，为利用基于形状的图像检索技术的现有理论和方法创造了条件，克服了模板匹配方法要求精确匹配及计算量大的局限性。从理论和技术的角度研究了在基于目标形状的遥感影像检索中，应用基于小波变换模极大值和多尺度形态学的边缘检测方法以及基于不变相对矩的轮廓特征描述方法的具体实现技术流程，通过实验充分验证了本文方法的可行性，并分析了检索性能。 ④从分析基于内容的图像检索技术在计算机科学中的抽象本质以及目前基于内容的图像检索技术中常用索引机制的局限性入手，强调了研究基于距离的度量空间高维索引结构的重要意义。对现有的各种度量空间高维索引结构做了分类，并从分析构造和搜索算法的角度研究了将典型的基于距离的度量空间高维索引结构VP一tree及其改进算法MVP一tree应用于遥感影像可视化特征相似性索引的具体实现方法，并通过实验对其性能进行了充分的验证和分析，为克服传统的降维技术以及多维索引结构用于可视化特征索引的不足提供了可靠的实践依据。最后进一步分析了改进广义MvP一tree性能的两个出发点，提出通过参数优化模型有利于改进广义MVP一tree的搜索性能，并提出基于遗传算法的性能改进思路。 ⑤设计了基于内容的遥感影像库检索原型系统的三层B/S模式体系结构和支持基于内容遥感影像库检索的层状数据模型，探讨了主要功能模块的实现方法并给出部分函数原型。 总结本文的研究工作，主要贡献及创新点可概括如下: ①提出了基于Nona一tree结构和小波直方图的遥感影像库纹理特征检索新方法，保证了影像检索在效率、精度和存储空间之间的平衡。 ②提出了采用多进制小波直方图技术提取遥感影像纹理特征的方法，为充分利用多进制小波变换在提取图像细节信息方面的优势提供了技术路线。 ③探索性开展了基于目标形状特征的遥感影像检索研究，从理论和技术的角度研究了基于小波变换和数学形态学的边缘检测算法和基于不变相对矩的轮廓特征提取方法，为有效开展基于目标形状特征的遥感影像检索提供了可行的技术方案。 ④将典型的基于距离的度量空间高维索引结构vP一tree及其改进算法MVP一t ree应用于遥感影像可视化特征的相似性索引;进一步提出通过参数优化模型提高广义MVP一tree搜索性能的思路。 本文在结论部分指出需要进一步研究的问题。更多还原

【Abstract】 Remote Sensing Data are basic data in digital earth development and its quick browsing and efficient retrieval are important means of remote sensing information extraction and sharing. As an effective means of manage and utilize image database information according to comprehension of images themselves, content-based image retrieval (CBIR) has become one of the most active researches in image databases, computer vision etc. and a key technology of information high way and digital library. CBIR provides new chance to solve the problem of information extraction from large remote sensing image database. However, the diversity and complexity of remote sensing image and the enormous data volume as well are big challenge of valid retrieval from remote sensing image databases, especially under web environment. Content-based retrieval of remote sensing database is a hot topic by integrating multiple disciplines including remote sensing image processing, image databases, computer vision and pattern recognition etc. and has gotten international considerable attention. Research on it has important meaning in theory and practice for promoting remote sensing information acquisition and sharing. This paper intends to put forward some new thoughts and methods on key technologies for content-based retrieval of remote sensing image database and to validate its efficiency and practicability through theory and practice. Main research and concrete work include five aspects:1. Firstly, domestic and foreign research projects and state-of-the-arts of content-based retrieval from remote sensing image databases are induced systematically. Then concerned key technologies are summarized and main obstacles compared with other applications of content-based image retrieval are analyzed. Based on them, key aspects to solution are pointed out.2. Secondly, this paper points out multi-scale texture feature extraction and reasonable strategies of block-based data organization are two important areas to supporttexture-based retrieval from remote sensing image databases. The limitation of traditional remote sensing data organization and the significance of wavelet transform in texture feature extraction of remote sensing images are analyzed and based on them, a method of creating homogeneous texture feature databases and retrieval implementation by integrating Nona-tree data structure and wavelet histogram technology is presented. It can take advantage of the powerful ability of Nona-tree, which can reach balance among precision, efficiency and storage, and that of wavelet histogram, which can extract texture feature with high efficiency. Experimental results are given to prove its efficiency. Besides, this paper introduces M-band wavelet theory to texture feature representation and analyzes its advantage compared with traditional two-band wavelet transform in theory. Feature extraction and similarity calculation by adopting M-band wavelet histogram technologies are presented in detail. Retrieval performance based on 2-band wavelet histogram and M-band wavelet histogram are tested and compared.3. Thirdly, by considering the difficulty of shape-based retrieval at present and automatic man-made objects extraction synthetically, this paper explores the feasible strategy of shape-based retrieval from remote sensing images at current level. This strategy avoids the impact due to the big difficult of automatic manmade object discrimination in current state and overcomes the limits of exact matching and huge computational volume aroused by template matching. This paper studies the concrete implementation flow of contour-based retrieval based on wavelet transform modulus maxima (WTMM), multi-scale morphology and invariant relative moments. Also, experimental results are given to validate the feasibility of our strategy and the corresponding retrieval performance is analyzed.4. Fourthly, by analyzing the abstract essence of CBIR in computer science and current solution of visual feature index are analyzed. Further更多还原