【作者】 潘翔；

【导师】 叶修梓；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2005， 博士

【摘要】 随着激光扫描技术的发展以及计算机性能的提高，三维模型不仅在数目方面快速增长，而且其应用领域也越来越广泛，如工业产品的模型设计、虚拟现实、3D游戏和模拟仿真等。特别是在互联网上，有大量共享的三维模型。研究和开发三维模型搜索引擎帮助用户快速、准确地找到自己所需的三维模型，是一个迫切需要解决的问题。论文针对三维模型形状分析和检索这一问题展开研究，主要工作包括以下几个方面： 在模型预处理方面。针对主向量分析用于姿态调整时产生的第一和第二主方向二义性问题，提出采用视图对称性来提高姿态调整的稳定性。首先采用主向量分析计算给定模型的三个主方向，得到模型在第一和第二主方向构成平面上的投影视图。对视图进行对称性分析，通过对称性进一步确定模型的第一主方向。实验表明，采用视图对称性可以有效地提高姿态调整的稳定性。 在基于直方图的特征描述方面。采用空间点极径和法向定义旋转不变的几何信号：径向夹角。以点极径和径向夹角作为基本几何信号构造直方图。径向夹角克服了直接采用法向构造高斯图像存在对模型姿态敏感的问题。提取的径向夹角直方图不仅具有旋转不变性，而且算法快速有效。采用直方图描述保证提取的特征在噪声干扰或是多分辨率描述下有比较好的稳定性。实验比较表明，和类似直方图描述相比较，径向夹角直方图可以得到更好的检索性能。论文还讨论了表面点采样结果对检索结果的影响，提出采用体素化方法对采样点进一步均匀化，用以提高检索准确率。 在基于矩描述的三维形状分析方面。以离散正交Krawtchouk多项式为基函数，定义用于形状描述的三维Krawtchouk矩。三维Krawtchouk矩能够对模型进行多分辨率描述。针对正交多项式的高计算复杂度问题，构造索引表用于加速特征提取过程。论文还讨论了不同阶下的矩对检索性能的影响以及体模型分辨率设置问题。在实验部分，我们和三维几何矩，三维Zernike矩进行检索性能比较，可以发现，采用三维Krawtchouk矩在检索准确率方面有明显的改进。在三维模型的结构特征提取方面。我们引入三维分割技术，从分割结果中提取结构特征。论文首先讨论了传统分水岭分割算法存在的问题，定义了一种称之为平坦度的几何信号用于描述模型表面特征。并提出分步合并算法用于解决过合并问题。然后从分割结果中构造拓扑连接图，采用图匹配算法计算不同模型之间的相似度。进一步的，论文针对拓扑连接图用于相似度计算存在的局限性，提出从两个模型的分割结果中构造二分图，把相似度计算转化为二分图的最大匹配问题。 单一特征都只是描述三维形状的某一方面几何特点，因此采用单一特征难以得到理想的检索准确率。采用多特征描述可以进行优势互补，从而有效地提高检索准确率。本文提出基于用户反馈的多特征联合。系统根据用户反馈自动更新权值，从而使不同的特征描述构成一个最优组合。论文进一步讨论在多特征描述下如何有效地提高检索效率，提出采用多层次检索结构，并定义自适应闭值。关键词:三维模型检索，对称性，径向夹角直方图，三维Kra叭chouk矩，分割结构特征，二分图，多层次检索结构，反馈更多还原

【Abstract】 Recent development in modeling and digitizing techniques has led to a rapid increase of 3D models, and more and more 3D free models can be accessed from Internet or from other resources. 3D model has been widely used in e-business, virtual environment and product design etc. An urgent problem is how to help people find their desirable 3D models accurately and efficiently from the model databases or from the web. Text-based retrieval has its limitation due to the difficulty of describing the shape content by text. Content-based 3D retrieval aiming to retrieve 3D model by shape content has become a hot research topic. The research of this thesis is about content-based 3D model retrieval, mainly 3D shape analysis. The main work can be concluded as following:In 3D model normalization, pose estimation is often performed by Principal Component Analysis(PCA). The algorithm, however, will appear to be unstable for some models. Post-processing using view symmetry is presented to improve the robustness of the PCA based pose estimation. The given model is first normalized by PCA algorithm. The main view of the normalized model is obtained by projection technology. If the view has a strong symmetry along the center line, we adjust the pose result by using the symmetry attribute.In histogram-based shape representation, Radius-Angle Histogram(RAH) is proposed to describe shape contents and used for shape retrieval. The RAH shape descriptor first uses a series of concentric spheres to capture the point distribution information of the given model. Then for points in each concentric sphere, a Radian Normal Angle is computed to extract the local geometry features. Finally, the Radius-Angle Histogram is constructed by using the extracted shape signatures. The proposed shape representation remains invariant under rotation. It can be generated from the given 3D model efficiently and easily as well. Performance comparisons for the shape benchmark database have proven that the proposed algorithm can achieve better retrieving performance than other similar histogram-based shape representations. The thesis also discusses the point sampling result’s affect on the final retrieving precision. The voxelization is used to make the sampled point more even over the surface, and better retrieving precision can be achieved by this process.In moment-based shape representation, we develop a new kind of 3D moment called 3D Krawtchouk moment. It is induced from discrete orthogonal polynomials defined by Krawtchouk.It can provide a multi-resolution representation for the given 3D model. The index table technology is also used to speed up the computation process of 3D Krawtchouk moment. In experiment, a comparison among different kinds of 3D moments, such as Geometrical moment, Zernike moments, has shown 3D Krawtchouk moment can achieve better retrieving performance.To extract structure feature for 3D retrieval, we decompose the given 3D model into some meaningful patches by segmentation technology. To achieve robust segmentation result, flatness measure for mesh faces is defined and used for 3D model segmentation. A two-stage merging strategy is presented to avoid over-merging, a problem often occur in traditional segmentation algorithms. The topological connection graph is constructed from the segmentation result, and used for calculating similarity between 3D models. The thesis further discusses the limitation of the topological connection graph based similarity calculation. Then Bipartite Graph is introduced to compute the similarity between models.Any shape feature, however, can work well for certain sets of models, and no single one can work well for all cases. To improve the retrieving precision greatly, we combine multiple shape features for 3D retrieval. The feedback technology is used to make an optimal combination on different shape features. On the other hand, a multi-stage retrieving architecture is developed to improve the retrieving efficiency based on hybrid shape features. Adaptive threshold is defined to assure that the retrieving system更多还原