【作者】 谢世朋；

【导师】 胡茂林；

【作者基本信息】 安徽大学 ， 应用数学， 2006， 硕士

【摘要】 本论文主要研究纹理图像的特征提取和分类。首先在第一章描述了与纹理相关的一些基本概念；第二章介绍了描述纹理特征的方法，接着在第三章讲述了基于滤波器组的纹理分类方法。该方法首先利用滤波器组对纹理进行滤波，纹理特征是用滤波器输出的统计值表示的；然后用这些特征向量进行纹理分类研究，分类主要利用了简单(naive)Bayes分类方法和最大加权相关树分类方法；最后通过实验验证了对纹理的分类，最大加权相关树分类方法的效果是较好的。 在第四章研究了基于局部仿射变换的稀疏纹理分类。以往对纹理的研究主要是针对2D纹理的研究的，一般是通过大量的训练样本来完成纹理特征的提取，本文在第四章介绍了通过仿射变换来提取稀疏纹理的立体信息，这样只需要较少的训练样本就可以完成3D纹理特征提取。实验应用CUReT数据做纹理数据集，运用贝叶斯分类器和改进的分类器对纹理进行分类。实验结果证明了改进的分类器对纹理分类的精度较好，但速度很慢。 在第五章提出了基于单应的稀疏纹理的3D表面特征描述方法，并进行相应的纹理分类研究。方法是利用随机抽样一致算法(RANdom Sample Consensus(RANSCA))自动计算两幅图像间的单应，并使用Levenberg-Marquardt算法最小化重投影误差估计与引导匹配的约束剔除初始匹配中的误匹配点，同时找到新的匹配对。提取两幅纹理的匹配点，不仅可以提取一般纹理的特征，而且可以提取包含立体信息的纹理特征，如由光照和视点的变化引起立体纹理的阴影变化等，通过匹配点的提取而不再需要大量的训练样本来进行纹理的特征提取。实验证明了该算法可以较准确和快速的进行纹理特征提取和分类，使得纹理分类工作变得有较强的可行性和实用性。 论文的创新点有：(1)基于最大加权依赖树的纹理分类；(2)基于仿射变换的3D稀疏纹理特征提取；(3)基于单应的3D纹理特征提取。更多还原

【Abstract】 In this paper, the classifications of textured materials are studied. Texture classification is of great challenge in general. In the first chapter, the background of this paper is described. In the second chapter, the general methods of feature measurement of texture are provided.In chapter three, firstly, the filter bank is exploited to generate features, and the filter responses are described as the feature of texture. Secondly, the two classification approaches: the Maximum Weight Dependence Trees and the Naive Bayes classification are explored. Finally, the experiments demonstrated that the method of Maximum Weight Dependence Trees (MWDT) had a better result. And MWDT can be a powerful tool for texture classification.In chapter four, the spare texture representation is described by the local affine regions. This is a significant difference from the traditional feature extraction framework, which only involves processing every pixel location in the 2D image. In this paper, a new texture representation is developed that is invariant to geometric transformation and it can be locally approximated by an affine model. Then, the two classification approaches: the Naive Bayes classification and the improved classification are explored. The proposed approach is evaluated in retrieval and classification tasks by the entire CUReT database.In chapter five, based on homography in computer vision, a geometric correction method of sparse texture is discussed, which used feature points of the texture. At the feature extraction stage, Using RANSAC (RANdom Sample Consensus) estimates the homography mapping, which doesn’t need lots of sample for feature measurement. Our experimental demonstrates that this chapter’s method is a powerful tool for texture classification.更多还原