【作者】 苏菱；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2010， 硕士

【摘要】 图像理解是当前计算机研究领域的热点。遥感图像解释作为图像理解在复杂场景中的应用研究,需要正确描述遥感图像中的内容,而基于知识的分割则是遥感图像解释系统的重要组成部分。SVM是基于统计学习理论的识别方法,其特点是通过在高维特征空间中构造判别函数来完成对样本点的分类,是图像理解中解决广义目标分类问题较为常见的分类模型。本文研究SVM分类方法并提出SVM多类分类模型的改进策略,旨在通过更为精确的模式分类实现一种适合城市遥感图像的分割方法,完成遥感图像解释的前期工作。本文的主要工作如下:(1)研究SVM的学习理论和算法原理,讨论SVM算法的核心问题,并通过总结SVM的拓展模型,掌握该方法的分类模型体系;(2)介绍SVM的多类分类模型构造思想,通过数据分类实验比较DTSVM和DAGSVM这两种Multi-SVM模型的分类表现;引入层次聚类的分类器设计方法和混合核函数的数据升维方式,构造基于层次聚类的DTSVM多类分类模型,并通过标准数据集上的实验验证了该分类模型的可靠性;(3)针对遥感图像的特点设计特征提取方法,对语义目标进行标记,形成特征集合,训练层次聚类DTSVM多类分类模型,实现基于DTSVM的遥感图像分割;(4)总结了图像理解在遥感图像解释上的发展,实现了一种基于知识的遥感图像分割方法,并以流行准则对分割结果做出评价;分割的结果包含语义,形成了可进一步进行遥感图像解释的信息。更多还原

【Abstract】 Image Understanding is the hotspot in computer research area. As the application research of Image Understanding on complex scenes, remote sensing Image Interpretation is aimed at correctly describing the contents in the image, and knowledge-based segmentation is an essential part of a remote sensing Image Interpretation system. As a recognition method based on Statistical Learning Theory, SVM classifies samples by constructing discriminant function in feature space, and have been a popular classifier for generalized object recognition. The paper studies the SVM methods and promotes Multi-SVM model so as to realize a remote sensing image segmentation method through classifying patterns, as primary work of remote sensing Image Interpretation.This paper includes the following contents:(1) We study the learning theory and classification mechanism of SVM, discuss the kernel points of the SVM algorithm, and summarize the modified SVM models, forming a comparatively complete system of SVM.(2) We introduce the construction mechanism of Multi-SVM, and compare DTSVM with DAGSVM through data classification experiments; we construct a Hierarchical Clustering DTSVM model by introducing hierarchical clustering for designing decision tree, and mixture of kernels for data mapping, and then verify the validation of this model through experiments on standard dataset, so as to lay a theoretical foundation for model selection.(3) We focus on the feature of remote sensing image and design the feature extraction method, annotate the semantic objects to form feature dataset, and accomplish the DTSVM based segmentation of remote sensing image through training Hierarchical Clustering DTSVM model.(4) We summarize the development of remote sensing Image Interpretation, realize knowledge-based segmentation for remote sensing image, and evaluate the results via popular criteria. The results involve semantic knowledge, and therefore provide information for further interpretation of remote sensing image.更多还原