【作者】 李嘉；

【导师】 王乘；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2009， 博士

【摘要】 多光谱遥感影像作为大幅面、多波段的对地观测数据，包含了丰富的光谱反射和辐射信息，为地物识别和判读提供了细致的诊断性依据，是地理信息系统和3S集成应用系统的重要基础数据。随着航空航天技术、传感器技术的飞速发展，人类可以更加方便快捷地获得大量对地实时观测数据，并将其广泛应用于各类地物、地貌和地质监测与规划等应用中。对地观测数据的急剧增长使遥感影像的快速浏览和高效检索成为一项繁琐、艰难的工作，甚至在一定程度上严重限制了遥感影像的共享与应用。基于内容的图像检索(CBIR)技术试图在计算机理解图像内容的基础上实现图像数据的合理组织和符合人类感知习惯的查询检索，为多光谱遥感影像的管理和检索提供了新的发展思路。然而多光谱遥感影像描述的地物光谱信息丰富、种类繁多、位置关系复杂、主题内容不明确，其视觉信息提取和数据组织方法与应用于普通图像、医学影像的基于内容检索系统有着较大差别。本文针对遥感影像特征提出了一套行之有效的CBIR系统技术方案，在影像自动分割、地物特征提取与匹配、高维视觉索引和数据存储模型等关键技术上提出了具有实用意义的创新，本文所讨论的各项关键技术可广泛应用于遥感影像的计算机辅助判读、地物识别、动态目标跟踪、灾害和变化监测等领域。本文的主要内容包括：①系统归纳和分析了目前CBIR国内外研究现状和主要科研成果，总结了其中所涉及到的各项关键技术，分析了遥感影像区别于普通图片、医学影像的不同之处以及由此带来的基于内容检索技术实现过程中的难点，并指出了解决问题的出发点和方法。②指出光谱纹理是多光谱遥感影像描述地物之间区别的重要视觉特征，是遥感影像自动分割最值得依赖的地物本质性特征。重点总结分析了当前基于多尺度小波和马尔可夫随机场模型(MRF)的纹理特征表达方法和遥感影像自动分割实现，并指出了其中存在的不足。为了兼顾像素之间和波段之间的光谱变化规律，将MRF模型应用于多光谱遥感影像的光谱纹理描述，通过实验分析，指出了有效的MRF模型光谱纹理特征描述方法。在此基础上，利用四叉树影像划分方法提出一种以检索为目的的遥感影像自动分割和地物对象一致性光谱纹理特征提取方法。针对纹理特征提取与纹理一致性假设存在的矛盾进一步提出利用主成份分析法(PCA)提高分割效率的方法。③详细总结了各类基于形状的图像检索技术，在分析各种对象形状特征描述方法特点和不足的基础上，指出区域形状描述方法相对于轮廓形状描述方法更适用于遥感地物对象。根据遥感影像光谱纹理自动分割结果的特征，在格网形状描述方法基础上提出一种采用特殊采样和编码方法的区域形状特征描述和相似性查询方法。该方法具有良好的尺寸、平移、旋转不变性，而且具有较高的特征提取效率和较大的特征压缩比，并利用实验充分验证了该方法的可行性和形状对象检索效率。④概括总结了应用于高维数据集的向量空间索引结构和度量空间索引结构，通过理论分析指出度量空间索引方法是摆脱高维数据索引“维度灾难”的有效途径，适用于高维视觉特征的遥感影像快速查询。在详细分析金字塔技术(PT)和iDistance索引机制和适用范围的基础上，将这两种优秀的度量空间高维索引方法结合起来，提出了一种能够根据高维数据分布特征进行空间划分的度量空间高维索引方法。该索引能够在一次查询处理中同时完成针对距离和空间方位的数据过滤操作，实验证明具有较高的剪枝效率。⑤针对遥感影像的大幅面、海量性特征，结合本文提出的遥感影像自动分割策略，选择无重叠的四叉树分块作为多光谱遥感影像数据的存储管理基本单元。为了克服传统四叉树空间索引在处理大范围邻域查找时I/O次数多、查询效率低的不足，依据Hilbert空间填充曲线规则对四叉树空间索引进行改造，提出了针对非满四叉树分块的Hilbert空间填充曲线生成方法和分块影像数据组织策略。实验证明这种数据组织策略具有较高的空间聚集特性，在处理大范围遥感影像同质区域数据查询时，具有较高的数据访问效率。总结本文研究工作，主要贡献及创新点可概括如下：①提出了一种以马尔可夫随机场模型为基础的地物光谱纹理特征描述和提取方法，并在此基础上设计实现了以检索为目的的遥感影像自动分割方法。②提出以索引为目的的地物形状特征描述和提取方法，该特征描述具有平移、旋转和尺寸不变等特性，而且抗噪能力强、相似性匹配计算简便。③提出了一种能够充分反映数据分布特征的度量空间高维索引结构，可高效处理高维数据的k-NN查询，该索引结构广泛适用于具有空间聚集特性的高维数据集。④提出了以图像分析为基础的遥感影像组织策略，并利用Hilbert空间填充曲线特征对传统四叉树空间索引进行改进，有效提高了同质纹理区域和空间相邻地物影像的数据访问效率。文章在结论部分指出需要进一步深入研究的问题。更多还原

【Abstract】 As large-scale, multiband earth observation data, multispectral remote sensing image(MRSI) which includes rich information of reflection and radiation that provides the finediagnostic foundation for the object discrimination and interpretation is the GIS and 3Sintegration application system’s important element data. With the fast developing ofaeronautics, astronautics and sensor technology, now we could acquire plenty of real-timeearth observation data more conveniently, and widely use these data in the application ofdiverse earth objects, terrain and geology’s inspection and layout. Along with sharpincrease of earth observation data’s amount, the fast browse and efficient retrieval ofMRSI becomes a burdensome and tough work, which, to some extent, severely restrictsthe share and application of the RS images. The Content-Based Image Retrieval (CBIR)technology tries to implement the image data’s rational organization and query retrievalcorresponding to our human being’s perception habit, and supplies a new developmentdirection for the management and retrieval of MRSI. However, the spectral information ofearth objects based on the description of MRSI is ample, of great variety and also hascomplex positional relations, unclear theme content, based on these characters, the visualinformation retrieval and data organization method aimed at the MRSI has a big differencewith the content-based retrieval system used in common images and medical pictures.This paper advances a suit of effective technical scenario for the content-basedretrieval system according to the RS image’s characters, advances the practical innovationin the key techniques of image autonomous segmentation, earth object character retrievaland matching, data storage model and high dimension visualization indexing. The keytechniques discussed in this paper could be widely used in the area of computer aided discrimination of RS image, earth object recognition, dynamic target tracking, disaster andtransformation surveillance and so on. The main content in this paper includes:1. Firstly, this paper systemic conclude and analyse currently CBIR research statusand mainly achievement, summarize the related key technology of CBIR technology,analyse the difference between RS image and common image, medical picture and thedifficulty bring forwarded because of this in the implementation process of CBIRtechnology, and indicate the jumping-off place of trouble shooting.2. Secondly, this paper indicate that spectral texture is important visual feature usedto distinguish the earth object according to MRSI description, and is the RS imageautonomous segmentation’s most worthwhile reliant essential character of earth object.Specifically summarize and analyse currently texture feature express method andimplementation of RS image automatic segmentation which based on the multi-scalewavelet and Markov Random Field (MRF) model, then indicate the disadvantage of it. Toboth consider the spectral varying rule of interpixel and interband, we use the MRF modelto give the spectral texture description of MRSI, according to analysis of experiment result,indicate an effective MRF model spectral texture character description method. Based onthis, utilize Quad-tree image partition methodology to bring forward a method of RSimage automatic segmentation and earth object’s consistency spectral texture characterextraction with the aim of retrieval, and advance a method to improve the automaticsegmentation’s efficiency using principal component analysis.3. Thirdly, this paper particularly summarizes each kind of shape-based imageretrieval technology, on the basis of analyse each object shape character descriptionmethod’s advantage and disadvantage, indicate that region-shape description method ismore suitable for the RS earth object than contour-shape description. Then according tothe character of RS image spectral texture automatic segmentation result and on the basisof grid shape description method, advance an adaptable method of region shape character description and comparability searching. This method has favorable scale, translation androtation invariance, and also has high character efficiency computing efficiency bigcharacter descriptor compression ratio, we use the experiment demonstrate the feasibilityof this method and the shape object’s retrieval efficiency.4. Fourthly, this paper summarizes the vector space indexing machenism and metricspace indexing structure which applied to high dimension data set, through academicanalysis we indicate that the metric special indexing method is an efficient way to get ridof high-dimension data indexing’s "curse of Dimensionality", and apply to the RS imagefast searching of high-dimension visual feature. Then based on the detailed analysis of PTand iDistance’s indexing mechanism and suitable area, we combine this two excellentmetric special indexing method, put forward a metric special high-dimensional indexingmethod which could process space partition according to the high-dimension datadistribution feature. Then demonstrate through academic analysis and experiment that thisindexing could accomplish the distance and space direction aimed data filter operation inone query processing, and has good filtering efficiency5. Finally, according to the large-scale, magnanimity feature of RS image and thestrategy of RS image texture automatic segmentation proposed in this paper, theun-overloaded quad-tree-based blocks are used as the elements of MSRI’s storagemanagement. To overcome the flaw of conditional quad-tree special indexing’s large I/Otime and low query efficiency in processing large scale neighbor searching, we modify thequad-tree space indexing based on the Hilbert spce filling curve law, and then advance theHilbert space filling curve building method in allusion to the partition of un-filledquad-tree and block image data’s organization strategy. The experiment proved that thiskind of data organization strategy has favorable space congregate character, and has betterdata accessing efficiency when process large scale RS image’s homogeneous region dataquery. To sum up, the primary contribution and innovation in this paper are as follows:1. Firstly, a new feature descriptor of RS spectral texture is proposed which isextracted by MRF model. Based on this feature descriptor this paper further design andimplemented a RS image automatic segmentation method with the aim of retrieval.2. Secondly, a new shape feature descriptor and extraction method are proposed,which have characters of scale invariance, translation invariance and rotation invariance,and have strong antinoise ability. And the calculation of the similarity match is relativelysmall.3. Thirdly, this paper proposes a new high-dimensional indexing structure for metricspace which could sufficiently reflect the data distribution character, and could process thehigh-dimensional data’s k-NN query with high efficiency, this kind of indexing could bebroadly applied to the high-dimension data set that owns the space congregate character.4. Finally, a new organization strategy of RS image based on image analysis isproposed. This paper utilize the Hilbert space filling curve character to reconstruct thequad-tree space indexing, and effectively improve the data accessing efficiency ofhomogeneous texture region and images of space neighbor earth object.In the conclusion, the successive problems and destinations are pointed out.更多还原