【作者】 陆丽珍；

【导师】 刘南； 刘仁义；

【作者基本信息】 浙江大学 ， 地图学与地理信息系统， 2005， 博士

【摘要】 21世纪遥感技术、计算机技术、网络技术的快速发展,使得各领域研究者获取所需要的高精度、高分辨率、多时相遥感图像成为可能,但与之对应的却是遥感图像检索理论和技术的严重滞后。如何从海量遥感图像库中快速准确地检索到所需要的信息具有十分重要的意义。论文从遥感图像检索研究现状和存在的问题出发,发展了通用遥感图像概念模型URSICM,设计了面向对象的逻辑组织与数据存储方案,提出了融合颜色纹理特征CTFFBIR和基于GIS语义的遥感图像检索新方法,探讨了特征相似性检索的优化方案,设计并开发了原型系统RSIQuery,为遥感图像库的检索与管理提供新的思路。论文主要研究内容如下:(1)讨论了遥感图像库内容检索RSIDBCI的部分关键技术,包括:遥感图像数据的组织与管理方式、数据库索引机制、视觉特征描述与提取、相似性度量、相关反馈机制,以及检索算法评价等,并指出目前RSIDBCI面临的困难和存在的问题。(2)通过对遥感图像所表达信息的特点、所包含的内容,以及现有图像数据模型的特点和局限的分析,提出一种通用的遥感图像概念模型URSICM,该模型将遥感图像的元数据、原始像元信息、视觉特征、图像对象、语义内容等信息纳入一个统一框架,并探讨了基于URSICM的面向对象的逻辑模型以及数据组织与存储方案。(3)论述了图像分解的目的和意义,在分析四叉树和九叉树两种图像分解方法后,提出五叉树分解新方法,该方法整合四叉树和九叉树方法的优势,在子图像数目、以及查询图像与子图像的重叠率之间达到了一个较好的平衡。(4)分析了单类视觉特征检索的不足,并根据高分辨率卫星与航空影像的光谱特点,提出一种融合颜色和纹理特征的遥感图像检索CTFFBIR新方法。该方法在利用多通道2D Gabor滤波器与图像做卷积得滤波能量值基础上,提取各子图像滤波能量纹理特征,计算子图像的颜色均值和均方差,对查询图像和与其大小相当的数据库子图像进行线性加权颜色和纹理特征距离相似性测度,其中特征的权重值可由查询者设置,也可通过相关反馈进行调整。(5)为了提高CTFFBIR的检索效率,提出了基于聚类的子图像分类索引优化算法。该方法通过离线对数据库各子图像按26维颜色和纹理特征向量进行聚类,并按聚类结果对数据库子图像建立分类索引,从而极大地减少了在线检索的响应时间。(6)提出了一种动态相关反馈算法,该算法采用适当改进Rui的多层特征权重更新方法的思路:各维、各类特征的权重在检索过程中通过对检索结果中子图像的相似性评分进行更新:前一轮被标上“不相关”的图像,不参与后续轮次的相似性测度;被标上“极相关”的图像,在后续轮次中具有优先排序号。(7)提出了基于GIS语义的遥感图像检索GISSBIR方法,该方法通过直接借用GIS描述空间对象语义和空间关系的能力,检索出感兴趣对象,并用这些对象的最小边框读取对应的遥感图像数据空间范围,从而完成图像检索任务,为遥感图像库内容检索提供了一种可行的思路。GISSBIR研究主要侧重在以下两方面:一是为协调用户查询请求与系统之间的语义冲突,设计并构建了概念语义网络;二是为实现空间关系的检索,对Oracle Spatial的方向关系进行了扩展。(8)设计并实现遥感图像库内容检索RSIQuery原型系统,该系统以Oracle为数据容器,以VC++为开发环境,采用分布式C/S架构实现。从试验结果分析来看,CTFFBIR及其优化算法是有效的,GISSBIR思路是可行的。更多还原

【Abstract】 The fast development of remote sensing technology, computer technology and internet technology in 21st century is now making it possible for researchers from various disciplines to acquire needed multi-temporal remote sensing images with both high accuracy and high resolution. In contrast with this, there is a significant lag of the theory of remote sensing image retrieval as well as its relevant technologies. How to effectively obtain the needed information from massive remote sensing image-base is, therefore, of great importance. Starting from the status quo and some existing problems of remote sensing image retrieval, this dissertation aims to: (1)bring forward a universal remote sensing image concept model (URSICM), design a logical object-oriented organization and a data storage schema based on URSICM, (2)raise a new approach of the color and texture fused features based remote sensing image retrieval (CTFFBIR) and discuss the optimized algorithms of URSICM, (3)raise an approach of GIS semantics-based remote sensing image retrieval, (4) design and develop the prototype system RSIQuey, and (5)inject some new ideas into the retrieval and management of remote sensing image base.Some main research contents are listed below:1. The research discusses some critical technologies of remote sensing image -base content retrieval(RSIBCR), including the organization and management of remote sensing image data, the database indexing mechanism, the description and extraction of low-level vision features, the assessment of similarity between, the mechanism of relevance feedback and the evaluation of indexing algorithm. Some confronted difficulties and existing problems have also been identified.2. By analyzing the information’s characteristics, the contents provided by remote sensing image and the characteristics and limits of current image data models, this dissertation intends to introduce a universal concept model of remote sensing image data URSICM, which integrates the metadata of remote sensing images, raw pixel information, vision features, image objects and semantic contents into a unified frame, and discusses the URSICM-based logical object-oriented model, data organization and storage schema.3. The research work demonstrates the purpose and significance of image decomposition. After examining the image decomposition methods of Quad-Tree and Nona-Tree, a method of Quin-Tree is put forward. This new method takes the advantages of both Quad-Tree and Nona-Tree, and reaches a more satisfactory balance between the quantity of sub-images and the overlapping ratio of query images and sub-images.4. This dissertation also seeks to analyze the shortcomings of single type vision feature based retrieval and develops a new remote sensing image retrieval method of integrating color and texture fused features CTFFBIR, according to the spectral characteristics of satellite and aviation imagery with high resolution. Based on the use of filtering power values as the convolution of multi-channel 2D Gabor filters and the images, this new method exacts each sub-image’s texture feature of filtering power, calculates average and mean square deviation of the color value of the subs-image, takes linear weighted similarity assessment from color and texture features’ distance with query image and sub-images holding the same size from the database. In particular, the weighted feature value could be set by people sending the query request, and adjusted by relevant feedbacks.5. In order to improve the retrieval efficiency of CTFFBIR, a cluster-based sub—image classifying indexing algorithm of optimization has been developed. This algorithm would largely reduce the response time of on-line qeuery by clustering each database’s sub-images using their 26 dimensional feature of color and texture features and constructing the classifying index of database sub-images in use of the results of clustering.6. An algorithm of dynamic relevance feedbacks is presented by the dissertation. The algorithm improves the Rui method of refreshing multi-level feature weights. The weight of different dimension and different class will be refreshed by evaluating the similarities of sub-images from indexing results during the process of image indexing. Images that have been labeled as uncorrelated in the previous round will not enter into the next similarity assessment, while the ones having been labeled as highly correlated will receive a number of high priority in the next round.7. The dissertation also presents a GIS semantics-based method of remote sensing image retrieval(GISSBIR). This method is able to search for objects of interests by directly making use of GIS’s capacity of describing spatial features and spatial relationships. By using the minimum border of these objects, the relevant geospatial scope of these remote sensing image data will be obtained, so that users can complete the task of indexing and produce an applicable alternative of remote sensing image database content indexing. GISSBIR puts special emphasis upon the following two questions: (1) constructing a conceptual semantic network to solve the semantic contradictions between user query request and the system, and (2) developing the retrieval of spatial relationships and making extensions for Oracle Spatial’s direction relationship model.8. The dissertation describes the design and implementation of the prototype system. Takes Oracle and Microsoft Visual C++ as its data container and developing platform, this system adapts a distributed architecture of Client/Server. By analyzing the experiment results, CTFFBIR and its optimum algorithm are essentially effective, and the thinking path is accessible.更多还原