本文为印度国家技术研究所（作者：VEDAVRATH LAKIDE）的硕士论文，共64页。
本文的主要工作是设计一种新的合成孔径雷达（SAR）图像分类方法。SAR图像分类在国民经济和军事领域有着广泛的应用，分析人员试图通过使用视觉解释的元素对SAR图像中的特征进行分类，以识别代表各种特征或感兴趣土地覆盖类别的同质像素组。SAR图像与光学图像（即照片）完全不同，它们的视觉解释也并不简单。因此，有必要设计一种新的分类策略。
一般的图像分类程序可以分为两大类：基于监督和无监督分类的方法。基于无监督学习的SAR图像分类通常需要对一些指标进行优化。局部优化技术经常会失败，因为这些度量函数对于转换参数来说通常是非凸的、不规则的，因此通常需要采用全局方法。
本文将合成孔径雷达图像分类问题作为一个优化问题，在文献中对各种聚类技术进行了研究。本文研究了一种基于进化的随机优化技术，即粒子群优化（PSO）技术对合成孔径雷达图像进行分类。该技术包括三个主要过程：首先，对合成孔径雷达图像中的每个区域选取训练样本；其次，利用粒子群算法对这些样本进行训练，得到每个区域的聚类中心；最后，对合成孔径雷达图像进行聚类中心分类。为了验证该方法的有效性，将分类后的SAR图像与K均值算法、模糊C均值算法等聚类技术进行了比较。在分类精度和计算复杂度方面，粒子群算法的性能优于其它方法，最后，采用不同的合成孔径雷达图像对结果进行了验证。
The prime objective of this thesis work isto devise novel methodologies for classification of Synthetic Aperture Radar(SAR) images. Classification of SAR images has extensive applications innational economy and military field. An analyst attempts to classify featuresin an SAR image by using the elements of visual interpretation to identifyhomogeneous groups of pixels that represents various features or land coverclasses of interest. The SAR images, totally different from optical images,i.e. photographs, and their visual interpretation is not straightforward.Therefore, there is need to devise novel strategies for classification of SAR images.Common classification procedures can be broken down into two broad subdivisionsbased on the method used such as supervised classification and unsupervisedclassification. SAR image classification based on unsupervised learning usuallyrequires optimization of some metrics. Local optimization techniques frequentlyfail because functions of these metrics with respect to transformationparameters are generally nonconvex and irregular and, therefore, global methodsare often required. In this thesis, SAR image classification problem isconsidered as an optimization problem various clustering techniques areaddressed in literature for SAR image classification. This thesis focuses on anevolutionary based stochastic optimization technique that is Particle SwarmOptimization (PSO) technique for classification of SAR images. This techniquecomposes of three main processes: firstly, selecting training samples for everyregion in the SAR image. Secondly, training these samples using PSO, and obtaincluster center of every region. Finally, the classification of SAR image withrespect to cluster center is obtained. To show the effectiveness of thisapproach, classified SAR images are obtained and compared with other clusteringtechniques such as K-means algorithm and Fuzzy Cmeans algorithm (FCM). Theperformance of PSO is found to be superior than other techniques in terms ofclassification accuracy and computational complexity. The result is validatedwith various SAR images.
1 引言
2 K均值与模糊C均值算法的研究背景
3 基于粒子群优化技术的合成孔径雷达图像分类
4 图像分类精度评估
5 仿真与结果
6 结论
