【作者】 王晓峰；

【导师】 黄德双；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 在信息社会里,图像已成为人类获取和交换信息的重要途径,而利用计算机进行数字图像处理是为了对图像中的目标进行分析,从而获得目标的客观信息并建立对图像的相关描述。图像分割一直是数字图像处理领域中最为基础和重要的问题,它是对图像进行视觉分析和模式识别的基本前提。所谓图像分割指的是根据灰度、颜色、纹理和形状等特征把图像划分成若干互不交迭的区域,并使这些特征在同一区域内呈现出相似性,而在不同区域间呈现出明显的差异性。近年来,水平集方法已经成为图像分割领域的一个研究热点,并在处理图像分割问题时表现出了良好的性能。相较于传统的图像分割方法,水平集方法有着显著的优点:用隐式表达的演化曲线(面)可以很自然地改变其拓扑结构,因此可以分割图像中具有复杂形状的目标对象;避免了对闭合曲线(面)演化过程的跟踪,将曲线(面)的演化转化成一个纯粹的偏微分方程求解问题;其有着较强的数学背景作为理论支撑,较为容易扩展到高维情况。因此,对其进行研究是非常有必要的。同时,水平集方法仍然处于发展阶段,其理论和应用方面的研究都有待于进一步深化和完善。在此背景下,本论文对水平集方法及其在图像分割中的应用和进一步扩展进行深入的研究,在基于局部信息的混合型水平集模型、基于多层水平集框架的多相图像分割、基于水平集方法的密度聚类框架、基于先验信息的植物叶片图像分割几个方面提出了有效的算法。本文的主要工作概况如下:(1)提出了一种新的基于局部信息的Local Chan-Vese(LCV)模型。通过使用局部图像信息,该模型可以在较少的迭代次数内分割灰度不均匀图像。在规则化项中引入能量惩罚项,使得水平集函数在演化过程中很自然地保持为近似的符号距离函数。此外,给出了一个基于演化曲线长度变化的水平集演化终止准则。最后,构造了一个新的扩展型结构张量,将其与LCV模型相结合,可以分割灰度均匀或者不均匀的纹理图像。在一系列人工和真实图像上的实验证明了LCV模型的有效性和鲁棒性。通过与Chan-Vese模型和Local Binary Fitting模型进行实验对比,显示出LCV模型可以在较少的迭代次数内分割灰度均匀或不均匀的普通与纹理图像,并且对于初始轮廓的位置和演化参数的选择不敏感。(2)通过在水平集方法中引入一种图像层的概念,构建了一种新的多层水平集分割框架。与传统的多水平集分割不同,多层水平集框架仅使用一个水平集函数,并且以一种层级演化的方式来进行多相图像分割。为了保证收敛的速度,提出了一种参数自适应更新方案。此外,定义了单图像层上和全局上的水平集演化终止准则,整个演化过程中无需任何人工干涉。在人工和真实图像上的实验结果表明了多层水平集框架的有效性,与传统的多相Chan-Vese模型相比,多层水平集框架具有较低的计算复杂度和更快的收敛速度。(3)提出使用水平集演化来逼近聚类中心的思想,并构建了一种基于水平集方法的密度聚类框架,从而成功地将图像分割方法扩展至密度聚类领域。与传统水平集方法不同,借助于数据空间的特性,水平集初始轮廓可被自动创建。演化过程中,不同类型的轮廓会以不同的方式包围各个聚类中心。为了得到包围聚类中心的最优的水平集边界,给出了演化轮廓记录集的动态更新准则。此外,还提出了一种有效的数据空间中的水平集演化终止准则。最后,在水平集边界的基础上设计了一种新的水平集密度,以用于在聚类过程中替代传统的概率密度。在人工和真实数据集上的实验结果表明,所提出的密度聚类框架可以有效地处理聚类中心较为接近的数据集,并进行离群点检测。通过与其它密度聚类算法的实验对比,显示出该聚类框架可以避免过拟合现象,并能解决聚类边界点与噪声或离群点的易混淆问题。(4)提出了两种有效的基于先验信息的植物叶片图像分割方案。两种方案的共同特点在于,分割过程需要分为预分割和正式分割两步来进行。第一种方案是基于水平集演化方式,使用叶片的近似对称性作为先验信息。第二种分割方案是基于形态学处理中的分水岭算法,使用叶片的形状大小作为先验信息。实验表明,对存在交叠或枝叶干扰情况的真实叶片图像,上述两种方案均能产生正确的分割。更多还原

【Abstract】 In information society, image has become an important way in which people can acquire and exchange information. So, the purpose of digital image processing on computer is to analyze the existing objects in images and acquire the essential information about the objects and give the related descriptions of image. Image segmentation has always been a most fundamental and important problem in the field of digital image processing. It is also the fundamental premise for the visual analysis and pattern recognition on the images. Generally speaking, image segmentation is to divide one image into some non-overlapping regions according to the intensity, color, texture and shape features. The segmentation result should make these features be homogeneous in the same region and obviously distinct in different regions.Recently, level set method has become a research hotspot in the field of image segmentation and achieved a good performance while addressing the image segmentation problem. Compared with the traditional image segmentation methods, level set method has some distinct advantages. First, it can segment the object with complicated shape in image since the evolving curve (surface) implicitly represented by the zero level set can naturally change its topological structure. Second, it can avoid the track procedure for the closed evolving curve (surface) and further transform the evolution problem of curve (surface) to the numerical solution to partial differential equation. Finally, it is theoretically supported by some strong mathematical backgrounds and can be easily extended to high dimensional case. Thus, it is very necessary to make a deep study on level set method. However, level set method is still staying in the developing stage, and the investigation of its theory and application should be enhanced and improved. In this thesis, the level set methods with their applications in image segmentation and further extensions have been deeply investigated. Some efficient algorithms have been proposed such as the hybrid level set model based on local information, multi-layer level set framework for multi-phase image segmentation, density-based clustering framework by using level set method, and prior information based plant leaf image segmentation methods.The main works in this thesis can be summarized as follows:(1) We proposed a new local Chan-Vese (LCV) model based on local statistical information. By incorporating the local image information into the proposed model, the images with intensity inhomogeneity can be efficiently segmented in few iterations. During the evolution process, the level set function can naturally maintain an approximate signed distance function by introducing a penalizing energy into the regularization term. Moreover, a termination criterion based on the length change of the evolving curve is proposed to ensure that the evolving curve can automatically stop on the true boundaries of objects. Particularly, an extended structure tensor (EST) was constructed for texture image segmentation. Combining the EST with the proposed LCV model, the texture image can be efficiently segmented no matter whether it presents intensity inhomogeneity or not. Finally, the experiments on some synthetic and real images have demonstrated the efficiency and robustness of our model. And the comparisons with the Chan-Vese (CV) model and local binary fitting (LBF) model also show that our LCV model can segment images with few iterations and be less sensitive to the location of initial contour and the selection of governing parameters.(2) By introducing a conception of image layer into the level set method, we proposed a new multi-layer level set framework. Different from traditional multiple level set segmentation schemes, the proposed multi-layer level set framework employs only one level set function with a hierarchical form to segment the multi-phase images. To keep the convergence speed, an adaptable evolution parameter update scheme was proposed. In addition, we also gave the termination criteria for level set evolution on single image layer and global evolution. It should be emphasized that no manual interventions are needed in the whole evolution process. Finally, the experiments on some synthetic and real images have demonstrated the efficiency of our multi-layer level set framework. And the comparisons with multi-phase Chan-Vese method also show that the proposed framework has a less time-consuming computation and much faster convergence.(3) We proposed finding the approximations of cluster centers through the level set evolution and constructed a density-based clustering framework by using level set method. Our framework can successfully extend image segmentation method to density-based clustering field. Unlike traditional level set methods, our level set evolution scheme can automatically compute the initial boundaries based on the characteristic of data space. In the evolution process, different types of contours would evolve in different ways to surround each cluster centers. To obtain the optimized level set boundary (LSB) surrounding the corresponding cluster center after the evolution process, the evolving boundary record (EBR) update criterion was defined. In addition, a termination criterion was presented to stop the evolution process when no more cluster centers can be found. Finally, a new level set density (LSD) was computed according to the level set boundary for clustering instead of traditional probability density. The experiments on some synthetic and real datasets show that the proposed framework works well while clustering the dataset with near cluster centers and detecting the outliers. The comparisons with some other density-based clustering methods further show that the proposed framework can successfully avoid the overfitting phenomenon and solve the confusion problem of cluster boundary points and outliers.(4) We proposed two efficient plant leaf image segmentation schemes based on the prior information. The common feature for two schemes is that the segmentation process can be divided into pre-segmentation procedure and formal segmentation procedure. The first segmentation scheme is based on level set evolution which uses the approximate symmetry of leaf as prior information. The second segmentation scheme is based on watershed algorithm in mathematical morphology which adopts the size of leaf as prior information. The experiments on some real leaf images show that two segmentation schemes can both achieve success while segmenting the leaf images with overlapping phenomenon and interference produced by branches and non-target leaves.更多还原