【作者】 陈雁；

【导师】 龚育昌；

【作者基本信息】 中国科学技术大学 ， 计算机应用， 2010， 博士

【摘要】 图像分割是遥感图像处理的重要研究内容。可见光遥感图像的处理在军事和民用方面都具有广泛的应用,而对可见光遥感图像的分割是其中的重要的研究内容之一。现有的遥感目标提取方法大都是在红外图像或者SAR图像上对某一类特定目标进行的,而可见光遥感图像上的分割方法大都是针对中低分辨率图像上的地物地貌分类。将中低分辨率遥感图像上的地物分类看作是一种分割,对这种区域性目标的现有分割方法通常效率比较低、实时性比较差。而对中高分辨率上边界清晰的地物目标而言,目前的提取方法基本处于人工经验判读或人机交互的半自动处理阶段,需要解决地物目标提取的智能化与自动化。本文针对可见光遥感图像中感兴趣目标分割提取的几个关键问题进行研究,根据不同的分割需求,对中低分辨率的区域性目标或中高分辨率的弱小目标分别采用不同的分割或提取方法。论文中对实际遥感图像进行了实验,验证了所提出方法的有效性。本文在对已有的遥感图像分割算法进行充分研究与分析的基础上,所做的主要研究工作如下:(1)考虑中低分辨率可见光遥感图像。对其上的区域性目标的分割问题进行了深入研究。针对可见光遥感图像上的城区目标通常具有边界模糊、连通面积大而导致分割速度慢等特点,本文基于模糊集理论提出一种矢量模糊分割法,通过样板法与模糊统计法相结合的方法构造模糊隶属函数,并构造了一个模糊训练过程来验证方法的有效性。(2)研究了中高分辨率可见光遥感图像中感兴趣目标分割的自动化技术。现有的目标分割方法,要么是需要先验知识的有监督的自动分割,要么是无需先验知识但需要人工初始化的无监督分割。本文依据解决计算机视觉中的丢失数据问题的思路,对最大期望法的初始化方法进行改进,实现了迭代过程中参数的自动初始化,从而实现了感兴趣目标的自动分割。(3)当感兴趣目标处在一个比较复杂的背景下,需要引入额外的附加先验才能够实现提取。论文引入了目标的形状先验。对复杂背景下的目标以受云层遮挡的海上舰船为例,通过分析获得舰船目标的形状模板作为先验,基于水平集理论的形状距离表示,将目标与复杂背景放在一起构造模型,并构造对应的能量函数,在能量最小化的过程中实现感兴趣目标的分割提取。(4)最后,为了实现完整的可见光遥感图像上感兴趣目标的解译识别,在保证高准确率和低时间复杂度的同时,建立了一个感兴趣目标的人机交互分割提取系统。研究工作中的新贡献在于:基于贝叶斯准则提出了城区的矢量模糊分割方法,相较于传统的多尺度分割方法,该方法快速有效,且具有较高的准确率。对视觉上的参数估计提出了一种自动初始化方法,称之为方向标定法。使用该方法时,目标分割过程既无需初始化也无需人工参与,只依据图像自身的光谱属性和颜色属性通过自身的初始化与迭代来实现。实验验证了该方法的有效性和可靠性。针对有云层遮挡的海上舰船提出了一种云层舰船模型(cloud cover ship model),将感兴趣目标与背景作为一个整体来看待。根据可见光遥感图像上云层是否存在阴影建立不同的模型以及对应的能量函数。实验验证了该方法的有效性。综上,本文的研究工作首先从理论上进行分析,进而进行相应的算法设计,并通过实验验证了各个算法的有效性。更多还原

【Abstract】 Segmentation is plays a vital role in remote sensing image processing. In the meanwhile, optical remote sensing image plays an important role in both military and civil situations. Therefore it is of great value about the research on optical remote sensing image segmentation. Existing remote objects extraction research mostly focuses on some kind of special objects in infrared images or SAR images. But the segmentation of optical remote sensing image mostly focuses on surface features landscape classification of low-middle spatial resolution optical remote sensing images. Taking the surface features landscape classification of low-middle spatial resolution optical remote sensing images as segmentation, existing methods mostly cannot get satisfied efficiency and reliability. For the surface objects with clear border and high-middle spatial resolution, the existing extraction methods are mostly depends on human experience or semi-automatic human-computer interaction. So the intelligent and automatic methods are necessary. In this paper, we study some key issues on objects segmentation and extraction in optical remote sensing images. Based on different segmentation requirements, for the low-middle spatial regional objects or high-middle small objects, different segmentation or extraction methods are proposed here. Experiment result on the actual remote sensing images gives evidence of this method’s efficiency.In this dissertation, based on the existing segmentation research of the remote sensing image processing; our work are as follows:Firstly, low-middle spatial resolution optical remote sensing image is taken into account. We take urban as an example of area objects and focus on the segmentation. The urban objects usually have a fuzzy boundary and Connectivity large area, so some segmentation speed is very slow. In this dissertation, a vectorized fuzzy segmentation method is proposed based on fuzzy set theory. The fuzzy membership function is constructed based on the model method and the fuzzy statistical method. The effectiveness of our method is verified by a fuzzy training process. Experiment result shows more efficient and higher accuracy compared with the traditional method.Secondly, we consider the segmentation and extraction in middle-high resolution optical remote sensing images. The existing segmentation is either the supervised automatic segmentation with priori knowledge or the unsupervised segmentation with human initialization. A complete automization method is proposed, without initialization or human interaction. It is a data loss problem in computer vision. We do some work with the initialization of expectation maximization and propose a direction labeling method, which is used in the iterations of parameter estimization process. Experiment result shows the efficiency and reliability.Thirdly, shape prior is introduced when the object of interest is in complicated background, such as ship objects under cloud in sea background. A cloud cover ship model is proposed based on features of optical remote sensing images. Based on the prior of the ship’s shape template, the model is constructed by putting the object and complex background together. Then the energy functions are constructed, and ship extraction is completed until the corresponding energy has been minimized. Experiment result gives evidence of this model’s efficiency.Finally, in order to get an object interpretation in the whole optical remote sensing image, we propose a human interactive interpretation system for objects of interest, which assures high accuracy and low time complexity at the same time. Our contributions are as follows:A vectorization fuzzy segmentation method for urban area is proposed based on the Bayes Rule. Compared with traditional multi-scale segmentation, our method is more effective and more accurate.A direction labeling method with automatic initialization is proposed for the parameter estimation of the vision. The object segmentation process is realized with its own initialization and iteration. The process only needs the spectrum and color attributes of the image, without initialization and human interaction. Experiment result shows the efficiency and reliability.A cloud over ship model is proposed for the ship objects under cloud in sea background. And the model is constructed by putting the object and complex background together. Different energy functions are constructed according to whether there is a cloud shadow. Experiment result shows the efficiency.In this dissertation, some research on theory analysis is presented at the beginning, and then some corresponding algorithms are constructed. Experiment result shows the efficiency and reliability of these algorithms.更多还原