【作者】 王佐成；

【导师】 李永树；

【作者基本信息】 西南交通大学 ， 地图制图学与地理信息工程， 2007， 博士

【摘要】 在遥感图像分类中，引入纹理特征能够很好地提高分类精度，已经得到研究人员的高度重视。另外，针对遥感图像的数据量大，更新快的特点，近年来兴起的数据挖掘技术为处理这些海量的遥感数据提供了新的技术手段和方法。纹理特征可看作图像中反复出现的局部模式和它们的排列规则，而关联规则数据挖掘方法够挖掘大型数据库中的频繁模式，这成为遥感数据处理与数据挖掘相结合的一个切入点。在图像中，纹理特征由于各种因素的影响，其模糊性和随机性表现尤为突出，采用模糊方法来处理图像模糊问题具有广阔的应用前景。基于模糊集的云理论为研究图像模糊纹理提供了新的理论依据，将其引入到遥感图像处理领域是一种创造性的应用。本文以遥感图像纹理特征作为研究对象，将其作为图像分类特征值进行遥感图像监督和非监督分类研究。在分类中重点研究了两种目前较为新颖的方法技术：一是基于数据挖掘图像纹理联合关联规则的遥感图像监督分类；另一个是基于模糊纹理特征矢量云的遥感图像非监督分类，最后用对象云模型对图像分类区域的模糊表达进行了研究。论文创新点在于：1．针对图像纹理数据挖掘的关键问题，提出适合纹理图像数据挖掘的方法体系，包括图像表示模型、基于象素特征图像纹理关联规则定义、纹理图像挖掘预处理、纹理图像模板统计挖掘算法。该方法体系涵盖了图像数据挖掘的各个环节，能够较充分挖掘出纹理图像中的关联规则。2．提出将遥感图像纹理特征联合关联规则作为纹理图像的特征表达。根据图像模板统计挖掘方法挖掘出系列频繁模式，实验证明通过这些频繁模式，能够较好表达图像的纹理特征。因此我们将其引入到图像监督分类算法中，通过建立纹理图像样本区域的纹理联合关联规则，构造模糊分类器，对纹理图像进行监督分类。实验证明该算法时间复杂度低，对于海量遥感数据的快速处理具有重要理论和应用价值。3．针对遥感图像的模糊性和随机不确定性的特点，创造性地将云模糊理论引入遥感图像处理领域，借助云模糊理论能够将定性论域的模糊性和随机性完全集成到一起构成定性和定量相互间映射的特性，来处理遥感图像的模糊和随机性。通过对纹理统计方法中纹理描述符的相关性分析，抽取最能表达某种纹理的遥感图像纹理描述符，基于遥感图像纹理特征，在图像微窗口进行多维云数字特征生成，构建纹理特征多维矢量云来表达遥感图像纹理特征。在此基础上，我们提出纹理特征矢量云距离计算方法，采用模糊聚类算法对遥感图像进行非监督分类，实验证明该方法能够提高图像分类的精度，并且算法收敛速度快。4．针对图像分类所获取的区域具有不确定边界，提出采用对象云模型来表达遥感图像上模糊空间分类区域(对象)。借助形态学中腐蚀的方法，获得空间对象区域的核心部分，将这部分最能够代表对象特征的区域作为云核，根据生成的云核，构建云滴隶属度数字特征，得到对象云及其数字特征表示。最后用对象云相似性对该表达方法进行分析验证，证明对象云能够很好地表达图像中的模糊对象区域。更多还原

【Abstract】 For better precision in classification of remote sensing (RS) image, many researchers pay attention to the texture of image. For the sake of dealing with the magnitude and updating quickly remote sensing image data, data mining technology emerging in recent years has become a new technique and method in RS image processing. Texture feature can be seen as local patterns and arrangement of the patterns. Association rules mining can mine the frequent patterns from large database. It is the cut-in of data mining and RS image processing. Because of many factors, the fuzziness and randomness of texture feature in image become significant. Dealing with the fuzzy RS image by fuzzy theory has a good future. Cloud theory based on fuzzy sets gives a new idea for processing fuzzy RS image. It is a creative application in RS image processing.The paper studies the supervised and unsupervised classification of RS image based on texture feature. Two new techniques were adopted in classification of RS image. One is data mining, supervised classification based on combined texture association rules. The other is cloud theory, unsupervised classification based on fuzzy texture feature vector cloud and representation of fuzzy classification region based on cloud model. The innovations of the paper were described as follows.1. Proposing the architecture of data mining of texture image, which is one of the most important matters in image data mining, including the representation model of image, concept of association rule based on pixel, pretreatment of texture image, data mining technique based on mask of texture image counting. The architecture includes all the steps of image data mining, and can mine the association rules of texture image.2. Proposing the combined texture association rules and representation of texture image based on combined texture association rules. By data mining technique based on mask of texture image counting, we can get the frequent patterns of texture image. Experiments validate that the frequent patterns can represent the texture feature of image perfectly. So we can accomplish supervised classification by frequent patterns that were mined from samples of texture image and fuzzy classifier of image. Experiments testify that the algorithm has important theory and application value by lower time complexity and better classification result.3. Aiming at the fuzziness and randomness of RS image, the paper introduces the cloud theory into RS image processing in a creative way. The digital characteristics of clouds well integrate the fuzziness and randomness of linguistic terms in a unified way and map the quantitative and qualitative concepts. We adopt the cloud theory to accomplish vagueness and randomness handling of RS image. After correlativity analysis of texture statistical parameters in Grey Level Co-occurrence Matrix (GLCM), we can abstract a few texture statistical parameters that can best represent the texture feature. Based on the abstracted texture statistical parameters, texture multi-dimensions cloud model can be constructed in micro-windows of image, which can represent the texture feature of RS image. At last, the method of counting the distance between the texture feature vector clouds was proposed and used in unsupervised classification of RS image. Experiments testified that the method can get better precision and the algorithm Convergence speed is quick.4. According to the fuzziness of boundary region from image classification, the representation of the fuzzy region (objects) from image classification was proposed. By morphological erosion, the core part (Cloud-core) of spatial region in classification image can be got. The Cloud-core can represent the characteristics of spatial region in classification image. Based on the Cloud-core, we can get the membership of Cloud-drop and accomplish the digital characteristics of Object-cloud. At last, the similarities between Object-clouds have been proposed and testified the method. Experiments validated that Object-cloud could represent the fuzzy region in image perfectly.更多还原