随着航空航天技术、传感器技术、计算机技术、网络技术的飞速发展，遥感图像的数据量也在急剧增加。因此，如何快速有效地管理和查询这些海量的遥感图像显得越来越重要，基于内容的遥感图像检索（Content-Based Remote Sensing Image Retrieval，CBRSIR）成为近几年来最活跃研究领域之一。本文首先研究了基于内容的遥感图像检索的国内外研究现状和进展，讨论了目前遥感图像检索存在的问题。然后，在总结现有研究成果的基础上，主要进行了如下研究：（1）详细介绍了基于内容的遥感图像检索的关键技术，包括：特征提取方法、相似性度量方法、评价标准和相关反馈机制等。（2）针对基于内容的遥感图像检索中视觉特征检索的现状，通过引入常规图像检索中的多特征融合技术，提出了融合多类特征（颜色、纹理、光谱）的遥感图像检索方法。（3）针对传统的基于支持向量机（Support Vector Machine，SVM）分类器的相关反馈策略在图像检索结果的排序上存在的问题，提出了改进的融合SVM分类器和特征相似性度量函数的相关反馈策略，在优化图像检索结果的排序的同时，进一步提高检索的查准率。（4）为了进一步提高基于传统SVM的相关反馈方法的性能，本文引入了合成核学习理论，并将之运用于传统SVM算法中，提出了一种基于面向数据的合成核支持向量机（Data Oriented Composite Kernel based Support Vector Machine，DOCKSVM）的相关反馈方法。（5）实现了一个功能较为完善的基于内容的遥感图像检索的原型系统。; With the rapid development of aerospace technology, sensor technology, computer technology and cyber technology, the amount of remote sensing image data has been increasing sharply. Therefore, how to effectively manage and retrieve such a huge amount of remote sensing images becomes a more and more important issue to be solved and content-based remote sensing image retrieval (CBRSIR) has been one of the most active research fields in recent years.This paper firstly studied the present state and development of CBRSIR at home and abroad, and discussed the problems existing in remote sensing image retrieval. Then, based on the actualities analyzed, several researches were conducted as follows:(1) Several key techniques were described in detail, including feature extraction, similarity measurement, evaluation criteria and relevance feedback mechanism.(2) Based on the actuality of visual feature based remote sensing image retrieval, this paper presented a multi-feature (color, texture, and spectrum) fused remote sensing image retrieval method by introducing the multi-feature fusion technique widely used in common image retrieval.(3) To solve the problems in ranking sequence using traditional support vector machine (SVM) classifier based relevance feedback strategy, this paper proposedan improved relevance feedback strategy using linear-weighted fusion of SVM classifier and feature similarity measurement function so as to optimize the ranking sequence in the retrieval results and further improve the retrieval precision.(4) To further boost the performance of traditional SVM based relevance feedback method, this paper adopted the composite kernel learning theory which was applied in traditional SVM algorithm, and put forward a data oriented composite kernel based support vector machine (DOCKSVM) based relevance feedback method.(5) A CBRSIR system having relatively perfect functions was developed in this paper.
修改评论