基于网格分割和层级特征的三维模型检索方法研究

发布时间：2018-07-15 11:43

【摘要】：随着计算机软、硬件的更新换代、互联网的快速普及和计算机图形学理论的不断完善，在影视动漫、仿真、生物医学等领域，3D网格建模技术已经被广泛应用。以三维立体取代二维平面，用虚拟模拟现实的3D建模技术带领人们步入了立体世界。然而，建模高度逼真的3D网格模型十分费时费力，假如能高效复用网络资源中现存的3D网格模型，就可以大大减少新模型的建模工作，如何从海量3D网格模型资源中快速查找并精确的筛选出所需模型成为了必需解决的问题。因此，研究3D网格模型检索方法工作是一项具有重要现实意义的工作。基于内容的检索是当前三维模型检索的一个热点研究方向，该方法根据网格模型的材质、纹理、空间结构等信息计算并提取其形状特征作为该网格模型的唯一标识，然后计算目标模型与模型数据库中待查模型的特征差值，将差值最小的前N个模型作为结果输出，实现三维模型的检索。因此，三维模型检索的关键在于如何提取网格模型的形状特征。本文对检索系统结构进行了分析并对现有的特征提取算法进行了总结，并提出改进的检索方法。主要做了以下三方面工作： 1.分析了三维模型检索方法的研究背景及意义，介绍了模型检索的流程和系统的框架结构，总结了模型检索中的关键技术。当前常用的特征提取算法种类繁多，缺少统一分类标准，本文对其大致分类，并分别进行阐述。 2.针对现有检索算法在提取形状特征时仅整体计算模型信息、忽略模型局部信息、未充分利用网格模型特征点等问题，提出一种基于网格分割的三维模型检索方法，并将之应用于三维模型检索系统。首先，对多种信号值计算方法进行比较，获得稳定度较高的网格三角形平坦度，并将其作为高度函数应用到改进算法中，，然后，对网格模型进行预处理，采用多维标度分析MDS（multi-dimension scaling）描述模型姿势不变性并提取模型显著特征点，采用特征点作为种子点指导网格分割。分割结束后，为避免出现过分割区域，采用多轮动态加权完成从局部到全局的合并使分割结果更加合理。最后，对三维模型的局部信息进行特征提取建立特征树，比较树的匹配程度检索相似三维模型。通过对几种目标模型进行检索，分析检索方法的合理性与有效性，进一步改进算法思想、精练算法步骤，设计程序结构并编写算法。实验验证，算法较好地利用了模型局部信息，检索速度快，在相同查全率下具有较高的查准率。 3.目前多数三维模型检索算法只采用单一形状特征表述，然而单一形状特征描述能力终归有限，只能有针对性的描述网格模型的某些性质，并不能适应所有模型的检索，有一定的局限性。因此提出层级特征检索，研究多种特征按照层次结构进行匹配，同时结合用户反馈机制，动态的计算与模型数据库中模型匹配时的权值，利用用户反馈的方法，对训练中的特征权值进行动态调整，得到不同的阈值，最后在网格模型检索阶段，先利用第一类形状特征与阈值比较，再选择一个权值与第二特征进行比较计算，对模型数据库中模型进行对比，实现三维模型检索。
[Abstract]:With the updating and replacement of computer software and hardware, the rapid popularization of the Internet and the continuous improvement of the theory of computer graphics, 3D mesh modeling technology has been widely used in the fields of animation, animation, simulation, biomedicine and so on. In order to replace the two-dimensional plane with three-dimensional stereotyping, the 3D modeling technology of virtual mode pseudo reality has led people into the stereoscopic world. However, the highly realistic modeling of 3D grid model is very time-consuming and time-consuming. If we can reuse the existing 3D grid model in the network resources, it can greatly reduce the modeling work of the new model. It is a necessary problem to find out and select the required model from the massive 3D grid model resources. Therefore, research 3 D grid model retrieval is an important practical work.
Content based retrieval is a hot research direction in current 3D model retrieval. This method calculates and extracts its shape characteristics according to the material, texture and spatial structure of the grid model, and then calculates the difference between the target model and the model data base in the model data base, and minimization of the difference. The former N models are output as the result, and the retrieval of 3D models is realized. Therefore, the key of 3D model retrieval is how to extract the shape features of mesh models.
In this paper, the structure of the retrieval system is analyzed and the existing feature extraction algorithms are summarized, and the improved retrieval methods are put forward. The following three aspects are mainly done:
1. analyze the research background and significance of 3D model retrieval method, introduce the process of model retrieval and frame structure of the system, and summarize the key technologies in model retrieval. There are many kinds of feature extraction algorithms in common use and lack of unified classification standard.
2. a 3D model retrieval method based on grid segmentation is proposed and applied to 3D model retrieval system for the existing retrieval algorithms that only calculate model information, ignore the local information of the model and make full use of the feature points of the grid model.
First, a variety of method of signal value calculation is compared, and the grid flatness with high stability is obtained, and it is applied to the improved algorithm as a height function. Then, the mesh model is preprocessed, and the multidimensional scaling analysis MDS (multi-dimension scaling) is used to describe the invariance of the model posture and the significant feature points are extracted. The feature points are used as seed points to guide the mesh segmentation. After the segmentation, in order to avoid the segmentation area, the combination of multi wheel dynamic weighting from local to global makes the segmentation result more reasonable. Finally, the feature tree is extracted from the feature extraction of the local information of the 3D model, and the matching degree of the tree is compared to retrieve the similar 3D model. By retrieving several target models, analyzing the rationality and validity of the retrieval method, further improving the algorithm thought, refining the algorithm steps, designing the program structure and writing algorithms. The experiment proves that the algorithm makes good use of the local information of the model, the retrieval speed is fast, and the precision rate is higher under the same recall rate.
3. at present, most of the 3D model retrieval algorithms are only expressed in single shape features. However, the ability of single shape feature description is limited. It can only describe some properties of the grid model, and can not adapt to all models. Therefore, the hierarchical feature retrieval is proposed, and a variety of features are studied in accordance with hierarchical nodes. The structure is matched, and the user feedback mechanism is combined to dynamically calculate the weights of the model matching in the model database. The user feedback method is used to dynamically adjust the feature weights in the training, and different thresholds are obtained. Finally, in the grid model retrieval stage, the first type of shape features is compared with the threshold value, then one more choice is selected. The weights are compared with the second features, and the models in the model database are compared to realize the 3D model retrieval.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】