【作者】 汪伟；

【导师】 杨新；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2007， 博士

【摘要】 利用计算机对遥感影像来进行分类是遥感数字影像处理的一个重要组成部分。近20多年来,出现了大量对遥感图像的地物目标进行分类识别的应用研究,而作为地物类别中的主要内容——人工地物区域的分类检测是其中一个重要组成部分,人工区域主要是指建筑物、道路、桥梁、和大型工程构筑物等。然而,由于人工区域的复杂性和多样性,提出正确、高效的分类算法具有相当的挑战性及研究价值。在人工地物与自然区域的分类检测中,特征提取和分类方法是两个重要的组成步骤,本文首先介绍了遥感图像分类的研究背景和国内外的研究现状,讨论了当前遥感图像的分类算法以及应用情况,以及对当前常用的特征提取算法进行了阐述和比较,进而参考图像处理和模式识别学科的最新发展引入了相应的特征提取和图像分类算法。本文将着重在以下几个方面展开研究工作:研究基于图像多尺度几何分析的遥感图像特征提取方法;研究基于图像偏微分方程的遥感图像分类方法;以及研究稀疏分类器于遥感图像分类中的应用。具体阐述如下:1.对于遥感图像中的特征提取进行了深入的研究,通过引入图像多尺度几何分析来对遥感图像进行最优逼近表示,图像多尺度几何分析相对于传统的小波分析更能够充分利用遥感图像本身所特有的几何特征来进行稀疏表征。文中首先引入Contourlet变换,并针对遥感图像所固有的特点提出了一种旋转不变特征的提取方法;接下来,针对遥感图像分析过程中,由于采样带来的信息丢失以及由此产生的Gibbs效应,本文引入了冗余无采样Contourlet变换来对遥感图像进行特征提取,并在对图像进行冗余Contourlet分解过程中提出了相应的基函数选择策略,进行自适应的遥感图像稀疏表征,优化了特征选择。2.针对人工地物与自然区域分类检测中的二分类和多分类问题,本文研究了图像偏微分方程在遥感图像分类中的应用,特别是基于水平集的几何曲线演化模型方法。针对二分类的人工区域和自然区域的检测问题,本文在经典的Chan-Vese两分模型基础上进行了改进,提出了相应的改进二分类模型;针对多类人工区域和自然区域的分类问题,本文将两分Chan-Vese模型进行拓展,得到图像多区域类别划分模型,避免了传统多分类模型中的耦合问题;而通过在演化模型中融合图像多尺度几何特征,可以得到理想的分类结果;在利用传统水平集方法来进行演化过程中,为了避免水平集收敛到局部最小的问题,本文采取了非传统的多分辨率处理方法,即通过在不同的演化阶段下对各分辨率特征进行分阶段处理,保证了水平集的正确演化,最终实现了对遥感图像的精确分类。3.由于遥感图像的复杂性,某些类别的区域特征会存在非线性分布问题,如果利用传统的Mumford-Shah分类模型及一些改进的模型则很难进行正确划分,而引入了稀疏分类器方法,提出相应的非线性映射Mumford-Shah分类模型,可以很好地解决这个问题。引入稀疏分类器方法的优点是可以充分利用大量的训练样本信息来提高对多区域类别的划分准确度;同时,利用稀疏分类器对原始遥感图像的非线性特征进行预处理后,对应的遥感图像会形成一个呈线性可分的类别归属度分布图。接下来利用二分和多分模型就可以完成类别的划分。关于非线性映射分类模型中稀疏分类器的选择问题,本文对SVM(Support Vector Machines)、RVM(Relevance Vector Machine)以及KMP(Kernel March Pursuit)等方法进行了理论分析与实验比对,最后选用了KMP方法。更多还原

【Abstract】 Classification of remote sensing image, which consists of assigning a label to each pixel of an observed image, has been one of key issues for remote sensing image analysis and understanding. Feature extraction and classification are two main steps in the classification procedure. In this paper, we concentrate on the following studies: feature extraction technology based on image multi-scale geometric analysis, remote sensing image classification method based on partial differential equations and the application of sparse classifier technology in the remote sensing image classification.Firstly, this paper presents the state of arts about aerial image classification, discusses the aerial image classification methods and the corresponding application fields. Then, the technologies of feature extraction are studied and each method is compared and evaluated. Finally, some algorithm of feature extraction and classification for remote sensing image are presented by considering the recent development and prograss of image processing and pattern recognition knowledge. In this thesis, we present some studies concentrated in the following topics:1. The wavelet transform is widely used in many fields, it can provide a very sparse representation for piecewise smooth 1-D signals but fail to do so for multi-dimensioned signals. Yet image multi-scale geometric analysis can extract the image’s intrinsic geometrical structure efficiently, it ensures the representation of the most distinguished features of the remote sensing image. The Contourlet Transform (CT) is firstly introduced into region classification in this paper to extract the rotationally invariant features. Then, the Non-Subsampled Contourlet Transform (NSCT) is also introduced which can avoids pseudo-Gibbs phenomena around singularities during the pre-process of remote sensing image denoising, owing to the properties of shift-invariant. NSCT also enriches the set of basis functions that makes it possible to extract some critical signal features. The optimization of basis selection is proposed in the NSCT to ensure the decomposition based on the maximum information content.2. We consider the remote sensing image classification as a partitioning problem. The partition is composed of homogeneous regions, namely the classes, separated by regularized interfaces. A novel method based on geometric contour model using level set evolution for partitioning of aerial image is presented. We modify the classical Chan-Vese model to deal with the two classes partition, i.e. the man-made objects detection. And then, an improved multi-region classification model was proposed based on Chan-Vese’s approach, which avoids the interactions between each level set function and speed up the curve evolution. By extending the improved models into vector image classification ones, these models could comprise the extracted features from image multi-scale geometric analysis, which will improve the classification result greatly. In order to avoid possible local minima in the level set evolution, we adjust the weighting coefficients of the multi-scale features in different evolution periods, instead of the classical technique which is only evolving in a multi-scale fashion.3. Some remote sensing images are so complicated that features in a certain class may be non-linearly distributed, and the traditional geometric contour models are only applicable to the linear feature partition problem. In order to achieve better classification results, the method of nonlinearly mapping extracted features to an easy classification space is presented in this paper. Consequentially, the sparse classifier is introduced to process these features, which is possible to classify the extracted features effectively. In our method, lots of training samples containing substantive information firstly yield the sparse classifier. Then pixels in the remote sensing image are labeled as different prediction values by the sparse classifier function. At last, the modified geometric contour model, which comprises the features of the prediction values, is built to deal with the non-linear situation. In the thesis, we also discuss each kind of sparse classifier method theoretically and demonstrate some fundamental experiments for comparison among them. According to the comparision results, the Kernel March Pursuit (KMP) approach is selected in our algorithm.更多还原