PDF《清华大学》

清华大学综合论文训练题目: 基于光场的 3D 对象检索系统的 实现与研究 系别:自动化系 专业:自动化 姓名:王泉 指导教师:戴琼海 教授

2010 年6月16 日 关于学位论文使用授权的说明 本人完全了解清华大学有关保留、使用学位论文的规定,即:学校有权保留 学位论文的复印件,允许该论文被查阅和借阅;

学校可以公布该论文的全部或部 分内容,可以采用影印、缩印或其他复制手段保存该论文.

(涉密的学位论文在解密后应遵守此规定) 签名: 导师签名: 日期: I 中文摘要 利用光场将 3D 对象表示为一系列多角度、多光照条件下的二维视图,能够 很好地利用现有的基于内容的图像检索技术来实现基于内容的 3D 对象检索.随 着光场数据量的不断积累,人们开始需要能够通过光场数据来对 3D 对象进行检 索.因此,基于光场的 3D 对象检索正逐渐成为学术界研究的重点话题. 之前的研究中已有人利用对象的形状特征建立过一套基于光场的 3D 对象检 索系统,并能得到较好的检索结果.然而,检索的准确率和计算的复杂度往往还 不能令用户满意.因此,本文实现了一个集成了多种特征和多种距离度量算法的 基于光场的 3D 对象检索系统,并研究了多种提高检索准确率以及降低计算复杂 度的方法. 首先,本文介绍了一种基于实物的光场数据库以及我们实验室搭建的一套基 于3D 模型的光场数据库.在这两个数据库中,本文讨论并比较了三种用于描述 3D 对象视图的底层视觉特征, 以及四种用于衡量 3D 对象相似度的距离度量算法. 其次,我们通过实现一种按类进行标注的人机交互方式来提高检索的准确率,并 研究了一种通过综合利用多种特征来提高检索准确率的多特征选择方法.然后, 本文讨论了通过动态聚类将对象的视图数压缩以减小检索计算复杂度的方法.此外,我们提出了一种通过将数据库中的对象进行分级聚类并找出待查询对象的相 似集来优化检索结果的方法.最后,本文介绍实验中我们搭建的 3D 对象检索软 件,并描述了该软件的安装、使用和扩展方法. 关键词:光场;

基于内容的 3D 对象检索;

特征;

距离度量;

聚类 II ABSTRACT By representing 3D objects as a series of 2D images obtained from multi-views and multiple light conditions with light field, we can make very good use of present Content-Based Image Retrieval methods for Content-Based 3D Object Retrieval. As the light-field datas on the Web keep increasing, people demand that light-field datas could be used for 3D object retrieval. Therefore, light-field-based 3D object retrieval is becoming more and more important in academics. In previous studies, a light-field-based 3D object retrieval system has been established based on shape features, which also showed good performance. However, the precision of the retrieval results and the time cost of the retrieval are still unsatisfactory. Therefore, this article implemented a light-field-based 3D object retrieval system which integrated multiple features and multiple distances, and studied some promising approaches which help to improve the precision and lower the time cost. Firstly, this article introduces a real-object-based light field database, and a 3D-model-based light field database that we established. Within these two databases, this article discusses three kinds of low-level visual features which are used to describe 3D object images, and compares four different distances which are used to measure the similarity between 3D objects. Secondly, we implement a human-computer interaction method in which the user can mark the retrieval results by classes to improve the precision of the retrieval, and studies a multi-feature selection approach which makes use of different features to improve the precision. Further, this article attempts to compress the number of object views by performing dynamic clustering to lower the time cost of the retrieval. Besides, we propose one approach which can help to optimize the retrieval results by performing hierarchical clustering on the whole database and finding out the similar object set for the input object. Finally, this article introduces the 3D object retrieval software that we established, and describes the installation, use and expansion of this software. III Keywords: light field;