《复杂流场特征提取与可视化方法研究》由会员分享，可在线阅读，更多相关《复杂流场特征提取与可视化方法研究（150页珍藏版）》请在金锄头文库上搜索。
1、国防科学技术大学 博士学位论文 复杂流场特征提取与可视化方法研究 姓名：徐华勋 申请学位级别：博士 专业：计算机科学与技术 指导教师：李思昆 2011-03 国防科学技术大学研究生院博士学位论文 第 i 页 摘 要 流场可视化是可视化研究的一个重要方向，在科学计算和工程分析中占据着 非常重要的地位。流场具有典型拓扑特征，并且这些特征具有普遍性。特征可视 化通过提取流场特征，使用户能忽略大部分冗余、不感兴趣的数据，减少了可视 化映射的数据量，同时保持了信息的准确性，具有独特的优势。 如何描述和提取流场特征信息是流场特征可视化需要解决的首要问题，虽已 有较多的研究成果，但目前已有的复杂流场特征描述
2、理论与提取方法仍存在通用 性和扩展性较差问题。三维复杂流场特征纹理可视化可准确地表现特征在三维空 间的连续变化特性，但由于存在显示遮挡问题，使用户难以发现和分析流场内隐 藏的特征性质。本文针对复杂流场特征可视化存在的问题，对流场特征描述、提 取及可视化方法进行较深入研究。基于提取的特征结构，采用纹理可视化方法得 到了高质量的流场表现，取得的研究成果包括如下几个方面： （1）针对已有特征提取方法对拓扑图中重要的构成要素临界点特征无法 合理界定区域范围的问题， 本文提出了一种基于拓扑分析的 2D 流场临界点特征区 域界定方法，在不需要用户选择的条件下，可自动合理地界定和填充各种类型临 界点的特征区
3、域范围。实验结果表明该方法能在定性描述流场全局拓扑结构的同 时，也对流场特征的定量属性如强度及其变化进行刻画，具有较好的应用意义。 （2）针对已有特征提取方法通用性和扩展性较差问题，结合神经网络强大的 非线性映射能力，提出了一种基于 BP（Back Propagation）神经网络的智能流场特 征提取方法。为提高提取方法的性能，设计了基于 GPU 的 BP 特征检测与识别算 法。实验结果表明，本文提出的算法的通用性较好，能有效提取典型的 2D/3D 流 场特征。同时，算法对于新的用户感兴趣的局部流场特征也能有效进行提取，扩 展性较好。 （3）针对复杂 3D 流场特征精确描述与提取困难问题，本文
4、采用模糊理论对 流场典型特征区域进行了统一描述，提出了一种基于模糊理论的流场特征区域提 取算法 FRFE。理论分析表明 FRFE 算法在最小平方和准则下是对流场区域的最优 模糊划分，并具有良好的划分性质。实验表明，与传统特征提取涡旋特征方法相 比，FRFE 算法提取的涡旋特征区域更为准确。 （4）针对传统的菜单式交互方式和确定的特征提取算法难以保证对复杂流场 特征所具有的不确定性进行有效描述问题， 在已提出的 FRFE 算法基础上， 提出了 一种模糊特征描述语言 FFDL 和基于 FFDL 的交互式模糊特征提取算法。 实验结果 证明该方法可实现对三维流场特征的交互式模糊提取，并可实现与研究人员
5、和工 程人员知识和经验的灵活结合。 国防科学技术大学研究生院博士学位论文 第 ii 页 （5）针对 3D 流场纹理可视化方法面临的遮挡问题，提出了一种基于流场特 征模糊提取的自适应稀疏纹理绘制方法。同时为提高稀疏纹理的方向感，提出了 两种冷暖光照方法。实验表明，通过生成与流场特征重要性成正比的 LIC 卷积纹 理，可有效改善 3D 流场的遮挡现象。同时，提出的冷暖光照方法可有效解决纹理 方法对流场具体方向表现上的缺陷。 主题词：流场可视化，特征提取，纹理可视化，拓扑可视化，BP 网络，模糊理论， 非真实感光照， GPU 绘制 国防科学技术大学研究生院博士学位论文 第 iii 页 ABSTRAC
6、T Flow visualization is an important topic in scientific visualization, and plays an important role in scientific computation and engineering analysis. There are some typical and universal feature structures in flow field. Feature-based approaches have been developed to extract flow features to elim
7、inate the redundant and uninterested data, which can reduce the final rendering data while maintaining the correct information. This advantage makes feature-based approaches an excellent method compared with others to visualize flow field How to describe and extract feature structures is the main pr
8、oblem in visualizing flow fields, and there are lots of papers try to solve this problem. However, the existing methods mainly aim at a certain class of features, such as vortices, which lack universality and extensibility. Texture-based methods can represent the continuous variability very well, bu
9、t it is difficult for users to discover and analyze the inner feature structures due to the occlusion problem. To solve this problem, this paper investigates the related theories about flow feature description, extraction and visualization. Based on the extracted features, this paper adopts texture-
10、based methods and obtains a high quality flows representation. The main research achievements are detailed as follows: (1) It is incapable of depicting the scope of the flow features for the topology-based methods, whereas the existing feature analysis methods are difficult to define the feature bou
11、ndary for the vector direction of the flows. To solve this problem, this paper presents a feature extraction approach based on these two kinds of methods, which can define the feature scope in 2D flow fields reasonably without the users selection. The experiments show that this approach can define t
12、he scope reasonably and describe the trend of flow features in the time-dependent fields very well. (2) To solve the universality and extensibility deficiency of the existing feature extraction methods, this paper proposes an intelligent feature extraction approach based on the back propagation netw
13、ork, which can make full use of the powerful non-linear description ability of the network. The proposed approach is implemented on GPU to improve the performance. The experiment results show that the approach proposed has favorable universality and extensibility, which can extract the typical and n
14、ew-come local features very well. (3) To solve the problem of describing and extracting the 3D flow features accurately, this paper proposes a universal description approach based on the fuzzy theory. Based on this description, an extracted algorithm called FRFE is presented. This algorithm is prove
15、d to be optimal in the minimum square sum rule. The experiments show that the FRFE algorithm can extract the feature structures more accurately than the traditional methods. 国防科学技术大学研究生院博士学位论文 第 iv 页 (4) Traditional menu-based interaction and certain feature extraction methods cant describe the unce
16、rtainty of the complex flow features. To solve this problem, this paper presents a fuzzy feature define language (FFDL) based on the FRFE algorithm. Furthermore, an interactive fuzzy feature extraction algorithm is proposed with the aid of the FFDL. The experiments show that this method can extract the feature structures fuzzily and can utilize the users knowledge flexibly as well. (5) To solve the occlusion problem in the texture-based method when visualizing
