【作者】 向世明；

【导师】 李华；

【作者基本信息】 中国科学院研究生院（计算技术研究所） ， 计算机应用技术， 2004， 博士

【摘要】 纹理是计算机视觉的一个重要研究方向。与纹理相关的诸多问题一直没有得到有效的解决。纹理感知就是其中的一个基本问题之一。纹理感知研究可以帮助人类洞察基本视觉机理,也是理解、分析和应用纹理的基础。基于对纹理图像及其纹理特征的统计分析,本文从理论和实际应用的角度,对纹理感知以及与纹理感知相关的纹理分类与分割、纹理生成、基于纹理的运动分割等问题做一些新的探索。本文的主要贡献如下:(1)　将纹理视为基本视觉元素的线性组合,并通过拓扑图格独立分量分析(TICA)来学习这些基本元素。由于采用矩阵表示,基本元素可用于表达所有纹理图像。在此基础上,提出了一种提取视觉元的方法。视觉元是TICA学习得到的基本元素的代表,为研究纹理所包含的知觉元素提供了一种视觉线索。不同视觉元的个数被称为纹理图像的内在视觉维数。内在视觉维数越大,纹理图像呈现出更多的视觉模式。　(2)　通过TICA从观测纹理图像学习图像分离基,并将其转化为一组滤波器,实现了一种无监督纹理分割算法。对比实验分析表明,该方法可取得满意的分割结果。　(3)　开展了对纹理图像集合整体性质的研究工作。通过TICA对图像进行了整体分析,发现纹理图像与自然图像在整体属性上存在较大的差别。为“人类对于纹理识别中最重要的三个测度为周期性,方向性和随机性”这一结论给出了一个统计学习的证据。基于对纹理图像与自然图像之间既相区别又相联系的研究,提出了研究纹理图像知识的新方法,即纹理图像本体理论。分析了纹理本体的研究任务、技术路线以及与上层本体即图像本体之间的联系。(4)提出了一种新的标识纹理的方法。该方法充分利用“纹理是一种区域性特性”这一共识,使用探测点确定观测纹理所在的主导区域,采用谱聚类解决了其中的关键技术。　(5)　通过无监督纹理分割实验证明了ICA整合Gabor纹理特征的能力。整合后的特征具有更好的可分性。(6)　提出了采用控向金字塔作为桥梁传递统计信息算法,实现了有选择的统计信息传输,生成了新的纹理图像。实验结果验证了该方法的有效性。　(7)　提出了一种基于在线高斯混合模型和纹理支持的运动检测算法。由于采用纹理捕获序列图像结构特征,因而该算法对于亮度的局部变化具有较好的鲁棒性。更多还原

【Abstract】 Texture is one of the main research topics in computer vision. However, there exist many unsolved texture-related problems. One of those underlying problems is texture perception. The study of the visual perception of texture can help us to better understand the basic mechanisms of human vision and the impact of texture itself, and also illumine us to develop deeper researches on texture analysis and the corresponding applications. By virtue of statistical analyses on texture image and texture features, this thesis, in both the theoretical perspective and the practical perspective, probes into the visual perception of texture and such related problems as texture classification/segmentation, texture synthesis, and texture-based motion segmentation. The main contributions of this thesis are summarized as follows:(1) A texture in this thesis is treated as the superposition of the basic visual elements, which are learned from the observed texture image by the topographical independent component analysis (TICA). Those basic elements are expressed as matrices. The form of matrix for visual element has the ability to represent any type of natural textures. Based on the learned results, an approach of extracting visitons is proposed. Visitons are the representatives of those visual elements in the observed texture image, which provide a visual cue to research texture perception. Further, the number of visitons is defined as the intrinsic visual dimension. The larger the dimension is, the more the visual patterns emerge.(2) A new method for texture segmentation by using the learned TICA unmixing bases as filters is implemented. Experiments on mosaic texture images show the segmenting results are satisfactory.(3) This thesis develops the new research direction about global properties of textures. Based on the analysis of the global properties, this thesis finds that there exits distinct difference between texture images and natural images. It provides a proof in the statistical earning perspective for the fact that the most important measures for texture perception are periodicity, direction, and randomness. By virtue of the study on the relationships between natural images and texture images, ontology is introduced to analyze texture knowledge. The research tasks, the technical routes of texture ontology, and the relationship between its father ontology are also discussed.(4) A novel algorithm of identifying texture is proposed based on the commonly更多还原