【作者】 李朋杰；

【导师】 马华东；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2013， 博士

【摘要】 随着三维扫描设备和三维建模软件的发展,三维模型的数量快速增长并且广泛地应用于各个领域中,研究一套高效的三维模型检索系统势在必行。在三维模型检索系统中,特征提取算法是其关键技术。三维模型一般被划分为两种类型：刚体三维模型和非刚体三维模型,相应地形成了刚体三维模型库和非刚体三维模型库,它们分别应用于某些特定领域,然而更多的应用需要既包含刚体三维模型又包含非刚体三维模型的通用三维模型库。三维模型检索主要包括模型标准化预处理、特征提取和相似度计算三个步骤。本论文主要研究特征提取算法,从刚体三维模型特征提取、非刚体三维模型特征提取、通用三维模型特征提取三个方面展开并提出解决算法。具体问题包括：如何对有空洞的刚体三维模型进行特征提取；如何解决非刚体三维模型特征提取过程中尺度变化问题；如何提取特征来提高通用三维模型的检索准确率；本文的主要贡献如下：(1)对于刚体三维模型库检索,提出了融合特征提取算法,这种算法联合使用视图特征和函数变换特征,提高了刚体三维模型的检索准确率。视图特征是基于三维模型的投影图像来获取的,用来提取三维模型的外轮廓信息；函数变换特征用来提取三维模型的内部结构特征,通过径向积分函数变换和球面积分函数变换分别从径向和轴向来提取特征,这样可以充分描述三维模型,而且对于有空洞的三维模型也可以正确检索。(2)对于非刚体三维模型库检索,提出了多尺度局部特征提取算法。首先在多个尺度上提取三维模型的关键点,为了提高关键点的可靠性我们通过计算关键点的重复度来最终确定关键点的位置,在多个尺度上检测关键点是为了防止在固定尺度上关键点遗漏的问题,使用主轴曲率比例来自动选择多个尺度。然后,在关键点处提取热核信号特征,热核信号对于平移和旋转具有不变性,但是对于尺度变化是敏感的,我们提出把热核信号纳入到特征袋框架中。在特征袋框架中尺度问题转化成平移问题,通过直方图量化技术来解决平移问题,从而解决了热核信号的尺度敏感问题。(3)对于通用三维模型库检索,提出了基于拓扑和视图的特征提取算法。对于通用三维模型检索需要同时考虑刚体三维模型和非刚体三维模型的特点。我们联合使用拓扑特征和视图特征,拓扑特征用多分辨率Reeb图来表示,视图特征是从空间结构环图像获取；多分辨率Reeb图描述三维模型的整体拓扑结构,空间结构环图像描述三维模型的局部特征。我们从三维模型的关键拓扑点处捕获三维模型的二维图像,关键拓扑点是基于三维模型本身的形状结构获取的,克服了传统算法渲染图像时增加的额外约束条件；视图特征比起传统算法的低维度的几何特征可以更好地描述三维模型的局部信息。(4)为了验证本文算法有效性,我们设计并实现了三维模型检索原型系统。通过在该系统上的大量实验表明,本文提出的特征提取算法能够提高三维模型的检索准确率。更多还原

【Abstract】 With the development of3D scanners and modeling software, the number of3D models increases quickly and3D models play important role in many fields, so a high-efficiency3D model retrieval system is necessary. Feature extraction algorithm is the key technology in the3D model retrieval system.3D models are generally classified into two classes:rigid3D models and non-rigid3D models. The rigid3D model databases and non-rigid3D model databases are produced, and they are used in specific fields. There are many applications needing generic3D model databases, which contain both rigid and non-rigid3D models.3D model retrieval mainly contains three steps:model preprocessing, feature extraction and similarity computation. This thesis mainly focuses on the feature extraction, which is carried out through three aspects:rigid3D model feature extraction, non-rigid3D model feature extraction and the generic3D model feature extraction. Our study contains three aspects:how to extract the feature of rigid3D modes with cavity, how to make the feature of non-rigid3D model invariant to scale and how to improve the retrieval accuracy of generic3D model. The main contributions of this thesis are as follows:(1) For rigid3D model retrieval, we propose an integration feature extraction algorithm, which is based on view and transform-based features. The view-based features are from projection images, we extract the exterior contour information of3D models by using view features. We extract the interior structure features of3D models based on transformation. Through the radial integration transformation and the spherical integration transformation, we can extract features from both radial direction and axial direction to fully represent the3D model. We can also retrieve rigid3D model with cavity correctly.(2) For the non-rigid3D model retrieval, we propose a multi-scale local feature extraction algorithm. First, we detect keypoints at multi-scale. In order to improve the reliability of the keypoints, we compute the multiplicity of the keypoints, and we define the final keypoints according to their multiplicity. The purpose of keypoints detection at multi-scale is to aoid skipping any keypoints. We use principal axis curvature ratio to automaticlly select the suitable scales. Then, we extract heat kernel signature features from the keypoints. The heat kernel signatures are invariant to translation and rotation, but they are sensitive to scale variation. Finally, the heat kernel signatures are put into the bag-of-features framework. In the framework, the scale problem is changed into the translation problem, we use histogram quantization technology to solve it.(3) For the generic3D model retrieval, we propose a hybrid feature descriptor, which combines the topological features and the view-based features. We use the multiresolution Reeb graph to represent the topological features and the Spatial Structure Circular Descriptor (SSCD) images to represent the view-based features. The multiresolution reeb graph can represent the global features, and the SSCD images can represent the local features. We render images from topological points, which can overcome the constraint condition. Furthermore, the view-based features have better descriptive power than low-dimension geometrical features, which are used by traditional3D retrieval methods.(4) We design and implement the3D model retrieval system to verify our method. The experimental results show that our feature extraction methods can improve the retrieval accuracy of3D model.更多还原