【作者】 张佩；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2014， 硕士

【摘要】 社会在进步，科技也在不断发展，人们已经不满足于二维图像带来的视觉感官。三维模型采集与存储技术、渲染技术以及三维模型在网络上的传播技术使得越来越多的人关注三维模型的发展，带来了“三维模型大爆炸”时代。三维模型的应用已经渗透到各行各业，工业产品的设计大都是依赖于三维模型的蓝本然后再到实物生产；医学设备成像也依赖三维建模为医生提供更多的诊断条件。只有大力发展三维模型检索技术才能满足以上如此广泛的需求。三维模型检索包括两个子过程：特征提取和相似度计算。其中，主要挑战就是特征提取，应该选择一个什么样的特征对三维模型的内容进行描述。非刚体三维模型的出现加大了检索的难度，因为非刚体模型能够承受的弯曲力度比较大、变换形式更多，导致非刚体模型具有非常丰富的可变性。本文以非刚体模型的特征提取为主要研究对象，一方面针对热核特征描述三维模型时存在的缺陷，提出了改进方式。另一方面提出了一种新的基于特征显著性的检索方法。在非刚体模型的特征提取算法中，利用拉普拉斯算子的特征系统对三维模型进行描述的方法获得了广泛应用。热核特征就属于这个类别，通过分析发现它对于尺寸变化没有不变性，可以通过傅里叶变换，消除尺寸变化的因子，获得了尺寸变化前后一致的结果。本文针对热核特征提出了基于顶点影响力的改进方法。一个顶点周围的连接通路越多、越复杂，该点的鲁棒性就越强，一些形变不会对该点周围的整体结构带来大的影响，则这些顶点就应该分配较大的权值，通过这种方式，进一步提高了检索准确率。进一步研究发现，不管是热核特征还是改进后的热核特征，都只关注模型的宏观几何结构这一种形状特性，对模型描述不够全面。本文提出了一种基于特征显著性的检索方法，该方法采用波核特征提取微观几何结构，补充丢失的信息。在进行相似性比较时，进一步考虑了不同特征在识别时的重要程度，本文利用特征与类别之间的相关度对特征的重要程度进行量化，构建显著性特征向量，从而突出具有高区分度的特征。实验结果表明该方法不仅有较高的检索准确率，而且对于多种变换都有较强的鲁棒性。更多还原

【Abstract】 With the development of computer vision and computer graphics, people do notsatisfied with the visual senses in2D world. More and more people are abstract bythe3D model, because of the model acquisition, storage technology, rendering andtransforming easily in the network. It brings a3D model of the big bang era.Three-dimensional model has penetrated into all walks of life, mostly industrialproduct design relies on a three-dimensional model of the blueprint and then to thekind of production; Medical imaging equipment is also dependent three-dimensionalmodeling to provide more diagnostic criteria for doctors. Only to developthree-dimensional technology greatly can meet needs of over such a wide range.Three-dimensional model retrieval consists of two sub-processes: featureextraction and similarity calculation. Among them, the main challenge is to featureextraction, namely choose what kind of features to describe the contents of athree-dimensional model. Non-rigid three-dimensional model appears to increasethe retrieval difficult, because non-rigid model has extensive variability because itcan withstand large bending strength and broad transformation category. In thispaper we focus on feature extraction of non-rigid model. On one hand for the defectof heat kernel signature in description of three-dimensional models, we put forwardthe improvement way. On the other hand, puts forward a new retrieval method basedon feature saliency.In the feature extraction algorithm for non-rigid model, using theLaplace-Beltrami operators on three-dimensional model to describe has a wide rangof applications. We first found that heat kernel did not invariant to scale changes. Inthis paper we use Fourier transform to eliminate this factor, after using our method,it keep consistent to scale changing. In addition, this paper also improves the heatkernel by using weighted-based vertex influence method. When the connection patharound the vertex is sufficiently complex, the vertex will be more robustness totransformation, and will not bring a big impact on the overall structure of thesurrounding vertex. These vertices are defined as influence vertices. Theweighted-based influence method improved the experimental result.Further research found that regardless of heat kernel or improved ones justconcerns the macro geometry structure, the model description is not comprehensiveenough. Therefore, this paper presents a retrieval method based on saliency features.This method uses wave kernel signatures to extract micro geometric structure, andmake up for the lost information. During the comparison, we considered theimportance of the different features. We quantify the weight values, by compare the correlation between different features, and build the significant feature vector.Experimental results show that the method not only has higher retrieval accuracy,but also for a variety of transformation has strong robustness.更多还原