【作者】 冷彪；

【导师】 覃征；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2009， 博士

【摘要】 在基于内容的三维模型检索领域,特征描述算法和相关性反馈算法是其关键技术。虽然特征描述算法能够提取不同的底层特征信息,但是在模型特征提取过程中存在不少亟待解决的问题。三维模型相关性反馈算法虽然能获取用户的语义知识,提高检索效果,但普遍存在样本筛选等问题。针对上述问题,本文进行了深入的研究。论文工作包括:1.提出了一种基于形状视觉的三维模型特征描述算法MATE。首先给出了一种改进的PCA方法;然后提出了一种等角距的傅立叶描述算子;接着介绍了一种改进的深度缓存算子;最后合成三种不同的算子形成MATE算法。与现有六种特征描述算法相比,MATE算法不仅在多种标准评价方式下获得很好的检索效果,而且还较好地处理了检索性能与特征维度之间的权衡问题。2.给出了一种基于先验知识的特征向量合成算法。利用查询模型计算各种特征向量的先验知识,然后动态地分配不同特征向量的权重合成一种特征向量。实验结果表明该算法能较好地利用各种特征向量的不同优势,其检索性能明显优于现有的两种三维模型特征向量选择算法和特征向量合成算法。3.提出了一种基于多层次的三维模型相关性反馈算法。采用多种特征向量描述模型不同底层特征信息的特点,并准确获取用户的检索需求。实验结果表明该算法能较快缩小用户高层语义知识与模型底层特征信息之间的差距,显著提高了三维模型的检索效果。与现有算法相比,该算法在多种标准评价方式下都具有明显的优势,仅用两轮相关性反馈就能获得较好的检索效果。4.建立了一种基于长期学习机制的SVM active learning相关性反馈算法。采用主动学习机制返回最具信息量模型,保存用户对所有模型的检索记录和标注信息,并利用拉普拉斯特征映射法挖掘历史检索过程中隐藏在模型之间的语义知识,最终实现模型在语义空间进行相似度匹配检索。与现有几种三维模型相关性反馈算法相比,该算法不仅能准确获取用户的语义知识,而且还能显著提高检索效果,仅通过一至两轮相关性反馈就能获得非常理想的检索结果。更多还原

【Abstract】 In the field of content-based 3D model retrieval, feature descriptors and relevance feedback algorithms are significant technologies. Although feature descriptors are able to extract various low-level characteristics, a few problems in the procedure of feature extraction need to be improved. 3D model relevance feedback methods can acquire users’semantic knowledge and improve retrieval effectiveness, but the troubles of sample filtering and others universally exist in such approaches. In order to deal with the difficulties discussed above, this dissertation systematically investigates on the area and proposes a series of algorithms to solve the problems.The main work in this dissertation includes:1. A visual based 3D model feature descriptor MATE is presented. This algorithm firstly proposes a modified PCA method, and then introduces an adjacent anlge distance Fourier descriptor, presents an improved depth buffer descriptor in succession, and finally combines three different kinds of descriptors. Compared with several 3D model descriptors, MATE not only acquires better retrieval effectiveness with a few standard evaluation methods, but also solves the tradeoff trouble between retrieval effectiveness and feature dimension.2. A prior knowledge based feature vectors combination method for 3D model retrieval is proposed. It calculates prior konwlege of different feature vectors using query model, and then dynamically allocates weight for feature vectors to combine a feature vector. Experimental results show that this method is able to make use of various advantages of feature vectors, and its retrieval effectiveness is obviously better than two state-of-the-art algorithms.3. A powerful relevance feedback mechanism for content-based 3D model retrieval is introduced. This approach utilizes several feature vectors to describe different low-level feature information, and acquires users’retrieval requirement accurately. Experimental results illustrates that this method quickly narrows the gap between low-level feature information and high-level semantic knowledge, and significantly imporves 3D model retrieval effectiveness. Compared with several algorithms, this method obtains evident superiority with several standard evaluation approaches, and achieves preferable retrieval effectiveness with two rounds of relevance feedback.4. A long-term learning mechanism based SVM active learning relevance feedback algorithm is presented. This method returns the most informative models with active learning mechanism, preserves retrieval record and marking information of users, mines semantic konwlege hidden among different models by Laplacian Eigenmaps, and finally retrieves models with similarity measuring in semantic space. Compared with relevance feedback algorithms in 3D model area, this method not only accurately acquires users’semantic konwlege, but also significantly improves 3D model retrieval effectiveness. It achieves perfect retrieval result only with one or two rounds of relevance feedback.更多还原