【作者】 郭竞；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2013， 博士

【摘要】 随着三维模型数据的不断增加及其应用领域的不断扩大,三维模型语义检索作为更智能化的检索方式,越来越受到人们的关注。三维模型语义检索是在三维模型语义知识提取、标注的基础上,在检索过程中利用语义知识查找模型的方式。语义检索方式对三维模型的检索和重用起到关键作用。针对基于语义检索方式的知识数据库的建立、知识提取、语义度量等关键问题,在建立三维模型本体知识网的基础上,围绕基于语义的三维模型检索展开研究,具体内容如下：1.针对检索中的三维模型几何特征提取问题,提出了相对角度直方图特征,并在该特征的基础上提出了基于相对角度直方图聚类的三维模型检索方法。该方法简单、有效,且满足旋转、平移、缩放不变性,具有较强的鲁棒性。在实验中该算法平均准确率可达85%以上,与其它统计特征提取方法进行比较而言,有效提高了三维模型的检索效率。2.针对三维模型检索中的无监督自动分类问题,根据实际采样到的三维模型数据所在的低维流形在局部是线性的、且每个采样点可以用它的近邻点线性表示的思想,最终将非线性问题线性化。在无监督的鉴别映射方法和局部保持映射方法的基础上,提出了一种基于半监督正交局部保持映射的三维模型分类算法。该算法利用大规模三维模型的流形结构,将观测数据映射为低维数据,提高三维模型的正确识别率。实验结果表明该算法能够有效地解决大规模三维模型的无监督分类问题。3.针对三维模型检索中有监督自动分类问题,基于隐马尔科夫模型的图像分类基础上,提出了一种基于二维隐马尔科夫模型的三维模型自动分类方法。该方法根据有监督机器学习原理,通过对少量已知样本进行学习,构建二维隐马尔科夫模型,以实现对未知样本的自动分类。实验表明该方法能够有效解决三维模型细分类问题。4.针对三维模型语义网构建问题,定义了基于三维模型应用本体的语义网,并在此基础上定义了三维模型应用本体描述方法和分层的语义网快速搜索策略。根据三维模型知识数据库的层级结构,在WordNet基础上进行扩展,采用树形拓扑结构构建了三维模型本体语义网。实验结果表明,该方法更符合三维模型数据库的组织结构,在三维模型语义检索中提高了检索速度和准确性。5.针对三维模型语义相关性度量问题,提出了基于三维模型本体属性的快速语义度量方法。该方法从人类认知学的角度,利用深度函数和广度函数度量三维模型的相关性。实验结果表明,该方法克服了传统语义度量计算中复杂度高,在三维模型检索中反馈结果不理想的问题；返回的结果相关性更高,更符合人的普遍认知。基于上述理论基础,设计和实现了三维模型语义检索系统。该系统具有良好的可扩展性,为深入开展基于语义的检索研究奠定基础。更多还原

【Abstract】 With rapid increases in3D models and continuous expansion, the semantic-based3D model retrieval as one of intelligent search methods attracts more and more attentions. Semantic-based retrieval is a model-searching way by using semantic knowledge based on knowledge extraction and semantic relatedness measure. Semantic-based retrieval plays a key role in3D model retrieval and reuse. As for the key problems of the semantic knowledge database creation, knowledge extraction, semantic metrics, based on the establishment of three-dimensional model, the semantic retrieval methods are studied as follows:1. Centering on the problem of how to extract geometric information feature of3D model, a feature extracting method is proposed based on the statistical characteristic--relative Angle histogram. The method satisfies invariance of the rotation, translation, scaling and has strong robustness. Comparing with other classical statistical feature extraction methods, this method has high retrieval recall and precision rates, and the average accuracy rate is over85%, which greatly improves the efficiency of three-dimensional model retrieval. The experimental results show that the proposed method is effective and feasible for semantic retrieval.2. Aiming at the problem of automatic classification of the three dimensionality (3D) models, according to the fact that the actual obtained data is always local linear in a low-dimensional manifold and each sample point can be represented with its neighbors, based on the UDP and LPP algorithms, a SSOLPP method is proposed and is applied to the3D-model-automatic-classification. The method makes use of the manifold structure of the large and high dimensionality data. The original data are projected to a low-dimensionality subspace by using the proposed method. In the low-dimensionality subspace, the within-class data are near to each other and the between-class data are far from each other. The experimental results on a real database show the effectiveness and feasibleness of the proposed method. 3. Aiming at the problem of supervised automatic classification, a kind of automatic classification method is proposed by using two-dimensional Hidden Markov Models. In the method, two-dimensional Hidden Markov Models are constructed by prior knowledge based on the machine learning theory. The experimental performance provide evidences that the proposed method can effectively improve the classification efficiency and accuracy of3D models repository.4. Aiming at the problem of build semantic web, the3D model ontology is defined and a layered semantic web search strategy is described. The ontology semantic web is constructed by using the hierarchical structure of knowledge database and based on WordNet extensions. The experimental results show that the method is more in line with the organizational structure of a3D model database and improves the recall ratio and precision of semantic retrieval and retrieval speed.5. Facing with the problem of semantic relatedness measurement for3D model ontology, a quick semantic-measurement method is proposed based on the features of the3D model. The method makes use of the human cognitive approach and the depth and breadth function to determine the relevance of the model. The experimental results show the propoed method has higher correlation with3D models and thus more consistent with general human perception. Based on the above analysis, a three-dimensional model of the semantic retrieval system and semantic search platform are designed and implemented. The platform has good scalability, which carry out the basis for the study of semantic retrieval.更多还原