基于视觉特性的图形图像分割算法研究

发布时间：2018-06-04 05:17

本文选题：网格分割 + 网格谱聚类　；参考：《吉林大学》2016年博士论文

【摘要】：得益于计算机科学和相关数学理论的进步与完善,图形图像处理已成为当今非常活跃的研究方向之一。其中,图形图像分割问题一直是该领域中的重要研究课题。经过多年的研究与发展,图形图像分割技术已被广泛应用于计算机动画、医学影像处理、虚拟现实、计算可视化等多个领域。在许多图像处理工作中,都需要对图像中的某些区域进行提取,我们可以借助图像分割技术对像素进行划分,将目标区域从背景中分离出来。由于图像分割的结果对后续的视觉任务有直接影响,这使得图像分割成为从底层图像处理进入到图像识别与理解的关键步骤。人类可以准确的将图像中的目标区域分离出来,但对于计算机这却并不是一件容易的事情。多数情况下受图像自身质量、以及图像内容的复杂性和多样性等因素的影响,使计算机很难按照人的理解对图像进行分割。目前的图像分割方法多是以图像中各区域的相似性,或特征差异度作为判断准则,将图像分成互不相交的若干区域,却很少把人类视觉特性应用于分割的过程中,致使产生的分割结果通常与人的视觉感知相差甚远。因此,如何将视觉特性与图像分割技术相结合,产生更符合人类视觉感知的分割结果,仍然是图像处理及相关领域中值得深入研究的课题。随着数据获取设备的进步以及建模技术的不断发展,三维图形数据已经成为一种新的数字媒体表示形式,对于三维模型的分析与处理也成为计算机图形学领域的研究热点。与图像分割问题一样,人们仍然希望可以借助人类视觉特性及相关理论对网格模型进行“有意义”的分割,得到多个具有视觉意义或物理意义的部件,以便于从更高层次上对模型进行理解。但通常情况下,人们对于“有意义”部件的定义是非常主观的,而且在不同的应用背景下,对于“有意义”分割的定义也有所差异。此外与二维图像数据相比较,三维模型除了几何属性外,还包含复杂的空间信息与拓扑信息,这使得三维模型的分割问题更具挑战性。因此,如何利用人类视觉特性,产生更加符合视觉感知的网格分割结果仍是值得进一步研究的课题。鉴于如何产生更加符合视觉感知的分割结果是图像分割与三维模型分割共同关注的问题之一,本文从人类视觉特性的角度出发,对图像分割与网格分割问题进行研究,并分别提出新的图像分割算法与网格分割算法。本文主要研究工作包括以下几点：(1)提出一种符合人类视觉特性的图像自适应阈值分割方法(Visual consistent adaptive thresholding method, VCA method)。传统阈值分割方法在分割过程中只考虑了图像灰度特性与空间信息,而忽略了视觉对于分割结果的影响。与传统阈值分割方法不同,我们的方法将阈值选择过程与人类视觉特性相融合,提出一种视觉一致的自适应图像阂值分割方法。首先根据像素的灰度信息构建两幅子图；然后根据人类视觉特性定义目标函数,定量刻画图像中的视觉信息；通过对目标函数优化求解,得到每幅子图的全局最优阈值；最后再利用图像的局部特性,进行局部自适应阈值操作得到最终的阂值分割结果。由于在阈值分割的过程中,我们利用人类视觉特性对前景与背景进行自动分离,使得分割后的二值图像获得了较好的视觉效果,其整体的视觉质量更符合人类视觉感知。(2)提出一种新的选取网格模型关键点的方法,我们称之为种子点,并在此基础上提出一种有意义的网格分割方法。首先找到网格模型中的尖锐特征区域,选出每个区域中最为显著的网格顶点构建候选点集合；用于分割的种子点是特征点集合的子集,通过最大化顶点集合之间的差异度对特征点集合进行筛选,从而得到网格模型的种子点集合；在此基础上,利用种子点集合对网格模型进行分割；根据视觉理论中的最小准则可知,人们通常将模型中的凹区域看成潜在的分割边界,为此我们利用网格模型的几何属性定义网格顶点间的距离函数,该函数由弧长,角距离和修正项三部分组成；最后通过对网格顶点进行聚类,得到视觉上有意义的分割结果。(3)提出一种基于视觉显著性与谱聚类的网格分割方法。我们将三维模型在原空间中的分割问题转化为谱空间的聚类问题。通过将视觉显著性与谱聚类过程相结合,生成有视觉意义的网格分割结果。首先根据视觉理论中的最小值规则制定多个判断准则以确定网格凹区域；然后根据网格的显著性来刻画顶点间的关联度,从而定义出网格模型的Laplacian矩阵；通过计算矩阵的特征向量,我们可以对原网格模型进行k维谱空间嵌入,从而将模型在原空间域中的分割问题转化为谱空间的聚类问题：最后通过分析网格的显著性确定每一类的初始聚类中心,并利用高斯混合模型(Gaussian Mixture Model, GMM)聚类方法对嵌入空间的网格顶点进行聚类,最终得到有视觉意义的网格分割结果。实验结果表明该算法可以得到视觉上有意义的分割结果,特别是对于凹凸特征明显,以及具有核心部件和分支结构的模型,该方法可以产生较好的视觉结果。
[Abstract]:Because of the progress and perfection of computer science and related mathematics theory, graphic image processing has become one of the most active research directions. The image segmentation problem has always been an important research topic in this field. After years of research and development, graphic image segmentation technology has been widely used in computer animation, Medical image processing, virtual reality, computing visualization and many other fields. In many image processing work, some areas of the image need to be extracted. We can divide the pixels by image segmentation technology and separate the target area from the background. This makes the image segmentation a key step in the image recognition and understanding from the underlying image processing. Human can accurately separate the target area from the image, but it is not an easy thing for the computer. In most cases, the quality of the image itself, and the complexity and diversity of the image content are in most cases. The influence of other factors makes it difficult for the computer to divide the image according to human understanding. At present, the image segmentation method is mostly based on the similarity of each region in the image, or the difference of feature as the criterion, and divides the image into several regions which are not intersected with each other, but rarely applies the human visual consciousness to the process of segmentation. The segmentation results are usually far from the human visual perception. Therefore, how to combine the visual characteristics with the image segmentation technology to produce the segmentation results more consistent with the human visual perception is still a subject worth studying in the image processing and related fields. With the progress of data acquisition and the continuous development of modeling technology, three Graphic data has become a new form of digital media representation, and the analysis and processing of 3D model has become a hot topic in the field of computer graphics. Like image segmentation, people still want to use human visual characteristics and related theories to make "meaningful" segmentation of the grid model, and get many of them. A component with visual meaning or physical meaning to facilitate understanding of the model at a higher level. However, in general, the definition of a "meaningful" component is very subjective, and the definition of "meaningful" segmentation is also different in different application backgrounds. In addition, compared with the two-dimensional image data, three In addition to geometric properties, dimensional models also contain complex spatial and topological information, which makes the segmentation of 3D models more challenging. Therefore, how to make use of human visual characteristics to produce mesh segmentation results more consistent with visual perception is still a subject worthy of further study. The segmentation results are one of the issues of common concern for image segmentation and 3D model segmentation. From the perspective of human visual characteristics, this paper studies the problem of image segmentation and mesh segmentation, and proposes new image segmentation algorithms and mesh segmentation algorithms. The main research work includes the following points: (1) a conformation is proposed. The image adaptive threshold segmentation method (Visual consistent adaptive thresholding method, VCA method) for human visual characteristics. The traditional threshold segmentation method only takes into account the image grayscale characteristics and spatial information in the segmentation process, but neglects the effect of vision on the segmentation results. Our method is different from the traditional threshold segmentation method. Combining the threshold selection process with the human visual characteristics, a vision consistent adaptive image segmentation method is proposed. First, two subgraphs are constructed according to the pixel gray information, then the target function is defined according to the human visual characteristics, and the visual information in the image is quantitatively depicted. The global optimal threshold of the amplitude subgraph; finally, using the local characteristics of the image, the final threshold segmentation results are obtained by local adaptive threshold operation. In the process of threshold segmentation, we use the human visual characteristics to automatically separate the foreground and the background, and make the two value images after the score cut better visual effect, The visual quality of the whole is more in line with human visual perception. (2) a new method of selecting key points of the grid model is proposed. We call it the seed point. On this basis, we propose a meaningful mesh segmentation method. First, we find the sharp feature area in the grid model, and select the most significant grid vertex in each area. The seed point set for the segmentation is a subset of the set of feature points. The seed point set of the grid model is obtained by selecting the set of the feature points by the difference degree of the maximum vertex sets. On this basis, the mesh model is segmented by the seed set. According to the minimum criterion in the visual theory, It is known that the concave region in the model is usually considered as a potential segmentation boundary, so we use the geometric attributes of the grid model to define the distance function between the vertices of the grid. This function is composed of three parts: arc length, angular distance and correction term. Finally, by clustering the vertices of the grid, the visual meaningful segmentation results are obtained. (3) proposed A mesh segmentation method based on visual significance and spectral clustering. We transform the segmentation problem in the original space into the clustering problem in the spectral space. By combining the visual significance with the spectral clustering process, the results of the mesh segmentation with visual meaning are generated. The criterion is determined to determine the concave area of the grid, and then the correlation degree between the vertices is depicted according to the significance of the grid, and the Laplacian matrix of the mesh model is defined. By calculating the eigenvectors of the matrix, we can embed the K dimensional space of the original mesh model to transform the segmentation problem into the spectrum in the original space domain. Clustering problem of space: finally, the initial clustering center of each class is determined by analyzing the saliency of the grid, and the Gauss hybrid model (Gaussian Mixture Model, GMM) clustering method is used to cluster the mesh vertices of the embedded space. Finally, the results of the mesh segmentation with visual sense are obtained. The experimental results show that the algorithm can be viewed. The results of meaningful segmentation, especially the obvious characteristics of concave and convex, and the model with core components and branch structures, can produce better visual results.
【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】