【作者】 王新颖；

【导师】 王钲旋；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2009， 博士

【摘要】 三维模型检索作为多媒体信息检索领域的重要组成部分,已逐渐受到人们的广泛关注。现有的三维模型检索研究主要集中在基于内容的检索方面,特别是努力提高三维模型的形状特征描述能力。但是,模型的形状特征只体现了模型的物理信息,并不能很好的表达模型的语义信息,这导致基于内容检索的效果常常不能尽如人意。针对上述问题,本文在前人工作的基础上,就三维模型的语义检索方面,特别是如何跨越三维模型的低层特征与高层语义之间的语义鸿沟问题展开研究,主要工作如下:1.使用用户在检索过程中的相关反馈记录,实现了一种基于语义关系的三维模型自动标注及语义检索方法。2.提出了将三维模型之间的语义关系融合到模型内容特征的基于流形以及基于核的特征值维数缩减方法,通过特征值维数缩减算法来搭建三维模型的底层特征与高层语义之间的桥梁。3.实现了一种三维模型语义与特征的联合聚类方法,使得三维模型的特征聚类结果包含更多的语义信息。4.构建了一个描述三维模型语义关系的本体,并将该本体应用到三维模型的语义检索中,这种基于本体的检索方法在一定意义上有效地解决了语义鸿沟的问题。5.提出了一种适用于三维模型检索的半监督加权距离度量学习方法。算法改进了传统的相关成分分析距离度量方法对少量已知分类数据分类不稳定的缺陷,使得用户在标注极少量模型的情况下有效地提高检索效率。更多还原

【Abstract】 3D model is a natural and direct way to illustrate the objects in the real world. It has more details of visual perception than that of two-dimensional image,and more applicable for human visual perception and mode of thinking. With the proliferation of 3D model,it becomes an emergency task to obtain the desired models from the exist,which has great application value in manufacturing,military,virtual reality,simulation etc.. 3D model retrieval is the same as other multimedia retrieval technique. The most commonly used search methods are text-based retrieval and content-based retrieval.The method of text-based retrieval looks on 3D model as an object in the database. It describes the model with keywords and text. Although the text-based approach may be convenient for user,it relies on artificial annotation and has many shortcomings such as time-consuming and hard-sledding. Moreover,the retrieval results need match strictly in query keyword,so its performance is not always satisfying.Therefore,the content-based retrieval becomes research focus in 3D model retrieval,especially the shape-based retrieval. It is widely accepted that the key problem of shape-based retrieval is extracting model’s shape feature with good properties,such as rotationally invariant, scale invariant, affine invariant etc. Researches of this kind concentrate on improving the describing ability of the shape features.However,model’s shape feature only reflects its physics information and can not represent its semantics. Due to the influence of Semantic Gap,the shape-based method doesn’t perform quite well. The root of the problem is keyword and feature have not any meaning for computer,namely they do not have any semantics. So the retrieval results cannot satisfy the user’s intention. An effective way to improve retrieval efficiency is to research the method of semantic based retrieval. So how to span the semantic gap between low-level features and high-level semantic and use semantic concept to manage and visit 3D model data is a challenge research problem in the field of multimedia.To solve these problems,the thesis concentrates on 3D model semantic retrieval,especially on how to span the gap between low-level features and high-level semantic of 3D models. The contributions of the thesis are stated as follows:(1).The research of 3D model semantics recognition and automatic annotation is the current hot and difficult issues. In order to avoid the difficulty of trying to automatically annotate single 3D model the thesis use the relationship among 3D models to replace the traditional method of automatically annotating single 3D model. During the course of 3D model retrieval,the system pays more attention to the relativity among 3D models,not the semantic of a single model. From above point of view,according to the relevance feedback records the thesis explores a method of 3D model automatic annotation and semantic retrieval.(2).To solve the drawbacks of the traditional text-based method,current researches concentrate on the shape-based 3D model retrieval. However,the performance of the shape-based method is not satisfying because of the semantic gap,especially for the high-dimensional data. The traditional method of reducing dimension does not care of the semantic information among data,so the result of the retrieval could not have an essential advance. In order to holding the shape similarity and the semantic relationship of the 3D models during the course of reducing dimension. The paper explores several methods to merge 3D model’s semantic relationship and content features by manifold based and kernel based nonlinear dimensionality reduction methods. A bridge has thus been built between 3D model’s low-level features and high-level semantic.(3).A method of co-clustering 3D model features and semantic is presented. The thesis accomplish this by modelling user feedback logs and low-level features using a bipartite graph,and then use the spectral method to partition this bipartite graph. Our experiments demonstrate that incorporating semantic information achieves better 3D model clustering.(4).The paper uses image and other fields for reference,and explores to implement an ontology which can describe the semantic relationship among 3D models. After applying the ontology to 3D model semantic retrieval we can build a bridge between 3D model’s low-level features and high-level semantic to some extent. This kind of ontology-based 3D model retrieval system can use the semantic information of 3D models in knowledge base to inference and search target 3D models according to user’s retrieval condition. It can also infer 3D model’s high-level semantic information from its low-level features through interaction with users during the course of retrieval. All these work has effectively solved the problem of semantic gap in a certain sense.(5).This dissertation proposes a method of semi-supervised distance metric learning for 3D model retrieval. In the field of 3D model retrieval,the performance of the unsupervised similitude matching method is not satisfying because of the semantic gap. However,supervised classification learning method usually needs a lot of training set. In order to improve the effectiveness of retrieval by a small count of classification information,we introduce a semi-supervised weighted distance metric learning method for 3D model retrieval. The method use a graph based semi-supervised Label Propagation algorithm to increase the classification information which is provided by user and then adopt a method of improved weighted relevant component analysis to learn a Mahalanobis distance function. After that,we can use the Mahalanobis Distance metric function to retrieval 3D models. The proposed method improves the defects of the instability of the relevant component analysis. Experimental results on Princeton Shape Benchmark have shown the effectiveness of our proposed method.On the whole,the thesis pursues the researches around several topics in semantic-based 3D model retrieval. Some new methods proposed in this paper have academic significance and value of application for reducing the semantic gap between 3D model low-level features and high-level semantic. In the future works,in addition to perform in-depth research based on current works,the proposed method in this thesis will be integrated in the efficiency and applicable 3D model retrieval system.更多还原