网格显著性检测中若干方法研究

发布时间：2018-06-17 23:05

本文选题：网格显著性 + 流形排序　；参考：《大连理工大学》2016年博士论文

【摘要】：随着扫描获取技术的发展和计算机处理能力的提高,三维几何模型成为新兴的数据类型,在娱乐、生物信息以及互联网有广泛的应用。数字几何处理是用计算机处理三维几何模型的技术,是计算机图形学的研究热点。本文对数字几何处理中网格显著性进行研究,主要工作如下：1.提出了基于描述子空间流形排序的网格显著性检测方法。首先,把网格过分割为超像素面片,借助局部中心对比机制计算每个超像素面片的局部对比值。局部对比值小的超像素面片认为是背景超像素面片,相反,则认为是前景超像素面片。其次,通过排序每个超像素面片与不显著的背景超像素面片的相关性得到显著图。最后,利用拉普拉斯算子得到光滑的显著图。与显著的前景超像素面片作为排序算法中的查询节点作对比,这样可以提高算法的鲁棒性,并且对查询节点的阈值不敏感。另外,考虑到三维模型的显著区域在空间域中是分散的,本文把流形结构加入到超像素面片的描述子空间中,在其中设计的排序算法更适合网格显著性。在大量的模型上与以前的方法进行对比,实验说明了本文算法的有效性和鲁棒性。2.提出了基于吸收马尔可夫链的网格显著性检测方法。本文利用特征方差得到不显著的区域,并考虑了背景信息和前景信息。首先,利用Ncuts算法对输入网格进行分块,根据Zernike系数将每块过分割成超像素面片,通过计算每块的特征方差来选择背景超像素面片。其次,把背景超像素面片复制为吸收马尔可夫链的吸收节点,计算每个节点的被吸收时间,得到了粗略的显著图。接着,从得到的显著图中提取前景节点,复制为吸收节点,相似的计算过程可以得到更好的显著结果,该过程抑制了背景区域并且有效的增强了显著的前景区域。最后,利用拉普拉斯算子得到光滑的显著结果。实验结果证实了本文的方法比以前方法具有优越性。3.视觉显著性可用来指导很多计算机图形学技术,比如简化、分割、光滑、视角选取。根据人类视觉变化的基本原则,本文利用局部对比机制来度量显著性。考虑到熵具有描述系统混乱程度的属性,采用熵值来刻画区域的局部变化。把法向量作为顶点描述子,计算每点的局部邻域内所有法向量的熵值,从而得到显著图。该方法简单、快速、有效并能产生好的结果。另外,本文还将显著性结果应用到了一些几何处理中。
[Abstract]:With the development of scanning acquisition technology and the improvement of computer processing ability, 3D geometry model has become a new type of data, and has been widely used in entertainment, biological information and the Internet. Digital geometry processing is the technology of computer processing 3D geometry model, and it is the research hotspot of computer graphics. In this paper, the significance of mesh in digital geometric processing is studied. The main work is as follows: 1. A mesh salience detection method based on description subspace manifold sorting is proposed. Firstly, the meshes are divided into superpixel slices, and the local contrast values of each superpixel surface are calculated by means of the local center comparison mechanism. The super-pixel slice with small local contrast value is considered as the background super-pixel surface, whereas the foreground super-pixel surface is considered as the background super-pixel surface. Secondly, the significance map is obtained by sorting the correlation between each super-pixel slice and the background super-pixel slice. Finally, the smooth salient graph is obtained by using Laplace operator. Compared with the significant foreground super-pixel slice as the query node in the sorting algorithm, the robustness of the algorithm can be improved and the threshold of the query node is insensitive. In addition considering that the significant region of the 3D model is decentralized in the spatial domain the manifold structure is added to the description subspace of the superpixel surface in which the sorting algorithm designed is more suitable for mesh saliency. Compared with the previous methods in a large number of models, the experimental results show the effectiveness and robustness of the proposed algorithm. A mesh salience detection method based on absorbing Markov chains is proposed. In this paper, we use the characteristic variance to obtain the region which is not significant, and consider the background information and foreground information. Firstly, the input meshes are divided into blocks by using Ncuts algorithm, and each over-pixel surface is divided into super-pixel slices according to Zernike coefficients, and the background super-pixel surface is selected by calculating the characteristic variance of each block. Secondly, the background super-pixel surface is copied as the absorbing node of the absorbing Markov chain, and the absorption time of each node is calculated, and a rough salient diagram is obtained. Then, the foreground node is extracted from the obtained salient map and copied to the absorbing node. The similar computing process can obtain better significant results, which can suppress the background area and effectively enhance the significant foreground region. Finally, by using Laplacian operator, a smooth and significant result is obtained. The experimental results show that the proposed method is superior to the previous method. Visual salience can be used to guide many computer graphics techniques, such as simplification, segmentation, smoothness, angle selection. According to the basic principle of human visual change, this paper uses local contrast mechanism to measure salience. Considering that entropy has the attribute of describing the degree of chaos of the system, the entropy value is used to describe the local variation of the region. The normal vector is regarded as the vertex descriptor, and the entropy of all normal vectors in the local neighborhood of each point is calculated, and the salient graph is obtained. The method is simple, fast, effective and can produce good results. In addition, the significant results are applied to some geometric processing.
【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】