合成孔径雷达(SAR)图像分割是SAR图像应用中非常重要的一个环节，但由于SAR图像中相干斑噪声的存在和特征选择不当，使得传统方法不能很好的对SAR图像进行分割。因此，将机器学习领域中新的研究成果应用到SAR图像分割中并构造有效的分类器具有重要的意义。 本文基于支持向量机(SVM)对SAR图像的特征提取和分割方法进行了较为深入的研究。主要内容如下： 首先，对SAR图像分割的背景及意义进行了介绍，分析了SAR图像分割的研究现状与发展趋势；对统计学习理论与支持向量机的基本理论进行了阐述；对合成孔径雷达和其中存在的相干斑噪声进行了研究，为图像处理提供了理论依据。 其次，本文结合小波变换在提取纹理特征、图像去噪方面的显著效果和支持向量机分类方法的优势，实现了一种基于支持向量机的单目标SAR图像分割方法。该方法先利用小波变换提取样本点周围的纹理特征，再对图像进行去噪处理；用小波能量特征及其加权平均值、去噪后样本点的灰度值和它的八邻域灰度值共同构成样本点的综合特征，利用归一化后的特征向量来训练SVM；然后利用训练得到的SVM对图像的每个像素点进行分类，从而实现SAR图像的单目标分割。实验表明，该方法对于单目标SAR图像进行分割，具有较好的分割效果。 最后，探索了一种基于支持向量机的多目标SAR图像分割方法。在人工提取样本后，将综合特征作为特征向量来训练得到多类支持向量机，利用该SVM实现多目标SAR图像的分割。其中，在由两类分类推广到多类分类时采用改进的“一对一方法”。对一幅含有多目标的SAR图像进行分割实验，结果表明该方法相对传统的分割方法能取得较好的分割结果。

Image segmentation is a key step in the application of synthetic aperture radar (SAR) image. However, because of the existing of speckles and unsuitable feature extraction, SAR image can not be segmentalized well by using traditional methods. So it is important to apply the new research of machine learning theory to SAR image segmentation and construct the effective classifier. The method of SAR image feature extraction and segmentation, which is based on support vector machine (SVM), is researched deeply in this thesis. The main contents and contributions are as follows: Firstly, the study background, significance, research status and development trend of SAR image segmentation are introduced. And the statistical learning theory, support vector machines, synthetic aperture radar and speckle noise in SAR image are studied. This provides theoretical basis for image processing. Secondly, according to the remarkable results of wavelet transform on texture feature extraction and image filtration as well as the advantages of SVM classification, a new single-target SAR image segmentation method based on support vector machine is proposed. The procedures of the method is as follow: First, texture feature of sample points is extracted by wavelet transform method. Second, image preprocessing is performed by using wavelet filtering method. Third, the comprehensive feature of sample points is constructed by wavelet energy features, weighted mean value of wavelet energy features, the gray values of the sample points which is denoising, and the gray values of eight-neighborhood. Fourth, a SVM classifier is designed and trained by using normalized feature vectors. At last, the testing sets of SAR image are sorted by trained SVM so that the single-target SAR image can be segmentalized. With the experiment result, the method is proved an efficient one of single-target SAR image segmentation. Finally, a new multi-target SAR image segmentation method based on support vector machine is proposed. Samples obtaining is processed by artificial choice, then comprehensive characteristics is regarded as characteristic vector to train support vector machine, this obtained SVM is used to realize multi-target SAR image segmentation. Among them, “one-against-one method” is adopted in the process of two kinds classification extend to various kinds classification. Eventually, through a multi-target SAR image segmentation experiment, the proposed method outperforms the previous traditional algorithm.