【作者】 陈强；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2007， 博士

【摘要】 图像分割是将图像划分成多个具有相似特征的区域，并提出感兴趣目标的技术与过程，图像分割是图像理解和图像识别的前提，也可以把它看作为图像理解与识别的初级阶段。图像分割的大致发展过程为：早期的基于灰度和梯度的经典分割技术(如阈值法，边缘和区域技术)，八十年代的活动轮廓模型(如参数活动轮廓模型和几何活动轮廓模型)，最近的结合形状等先验知识的分割方法(如活动形状模型和活动表面模型)。从图像分割的发展过程可以看出：所使用的先验信息越来越多，所具有的智能化程度越来越高，分割能力越来越强。本文主要利用灰度、形状、纹理和运动等信息研究了几种图像分割方法。本文首先分析一种经典的基于灰度与边缘信息的图像分割方法——图像二值化，然后详细讨论形状先验知识与活动轮廓模型相结合的方法和纹理分割方法，最后利用运动信息对序列图像进行跟踪。本文试图将各种信息(如灰度、形状、纹理、运动等)综合在一个统一的图像分割框架下，使它具有较高的智能化程度，从而能够解决一些复杂的图像分割问题。为了验证本文方法的有效性和实用性，本文方法被应用于左心室核磁共振图像分割。本文所做的主要工作和研究成果如下：(1)综合考虑了边缘信息和灰度信息，提出了一种新的双阈值二值化方法。该方法首先通过在图像边缘附近寻找种子点，然后在由封闭轮廓划分的高阈值二值图像中进行种子填充得到初步的二值化结果，最后通过低阈值二值化图像的修补得到最后的二值化结果。实验结果表明此方法能够较好地解决低对比度和目标象素灰度不均匀图像的二值化问题，与传统的二值化方法相比，也与本文类似的结合边缘信息的二值化方法相比，本文的方法具有一定的优点。(2)给出了一种结合先验形状统计信息的Mumford-Shah模型的水平集实现方法。结合形状统计的图像分割方法主要包括先验形状模型的构造和形状能量项的构造。本文针对这两个主要方面做了如下两点工作：首先提出了一种简单可行的先验形状模型构造方法；再者则是重新构造了形状能量项，它综合考虑了全局和局部形状信息，且不含形状姿态参量，使曲面演化稳定可靠。(3)提出了一种基于方向局部方差的快速主动纹理分割方法。首先，采用方向局部方差得到一组特征图像，为了降低计算的复杂度，通过可分离性评价准则，选出可分离性较好的四个方向上的特征图像；然后采用非线性扩散滤波平滑待分割特征图像，以改善分割效果；最后通过变差框架将特征图像结合到基于水平集的无监督分割过程中，为了提高算法的运算速度，采用了直接计算能量泛函的方法求解变差公式，而不需要求解欧拉-拉格朗日方程。合成图像和真实图像的分割结果证明：该方法具有一定的通用性和较好的性能。(4)提出了三种基于目标轮廓的运动跟踪方法：●提出了一种新的基于边缘形状匹配的目标跟踪方法，主要利用图像的边缘信息来构造目标的形状相似度。为了保证跟踪精度和处理目标的局部变形，本文提出了一种新的基于权重窄带的形状匹配度计算方法。●提出了一种基于目标轮廓点匹配度的参数活动轮廓目标跟踪方法。基本思想是根据目标轮廓点构造匹配度图像，然后采用参数活动轮廓模型跟踪目标。本文构造了一种新的特征匹配方法和一种新的方向滤波方法。考虑到相邻两帧间的目标运动较小，本文将目标轮廓的演化限定在初始轮廓的窄带区域内，采用了窄带水平集的窄带构造方法。●将基于区域的方法与基于边界的方法相结合，提出了一种两阶段目标跟踪方法。首先采用基于核的目标跟踪方法快速定位目标区域，然后采用扩散snake进一步演化目标轮廓，精确定位目标边界。在初始目标定位阶段，为了能够有效地给出目标的初始跟踪位置，用kalman滤波对初始目标位置进行预测，同时通过Bhattacharyya系数进行进一步判断。在基于轮廓的精确定位阶段，采用了颜色空间成分阵列的方法生成目标特征图像，然后在特征图像上采用扩散snake进行边界演化，同时在演化过程中，为了保证目标轮廓向好的方向演化，对演化得到的目标区域进行相似性比较。(5)提出了两种左心室核磁共振图像的分割方法：●针对带标记线左心室核磁共振图像的特点，提出了一种基于分级处理的自动分割方法，主要由三部分组成：首先用数学形态学的方法实现左心室的自动定位；然后用k均值聚类、模板匹配和基于骨架的心肌形状恢复方法给出左心室的内外初始轮廓线；最后用改进的水平集方法对初始轮廓线进行演化而得到最终结果。●针对不带标记线左心室核磁共振图像外轮廓的特点，在变差框架下，通过将Mumford-Shah模型作用于目标特征空间，同时结合形状统计先验知识和边缘图像，本文提出了一种改进的结合形状统计的变差方法来分割核磁共振图像的左心室外轮廓。形状先验知识的引入较好地克服了左心室外轮廓的边界断裂现象；边缘图像可以增强弱边缘，提高分割精度：通过以心肌为目标区域建立目标特征图像，使得左心室内部区域“灰度”一致，从而方便了Mumford-Shah模型的应用。更多还原

【Abstract】 Image segmentation is a process of dividing an image into different regions such thateach region is, but the union of any two adjacent regions is not, homogeneous. Imagesegmentation is the first step in image understanding and pattern recognition. The basicdevelopment process of the image segmentation is as follows: the early classicalsegmentation methods based on image intensity and gradient, such as thresholding,contour-based, and region-based; the eighties active contour models, such as parametricactive contours and geometric active contours; the segmentation methods with priorknowledge, such as active shape model and active appearance model. From thedevelopment process, we can observe that the intelligence and ability of the imagesegmentation become better and better. In this paper, the following information is mainlyused to segment objects, such as intensity, shape, texture and motion. We first analyze aclassical image segmentation method, image binarization, and then discuss thesegmentation methods by combining shape priors and active contours and the texturesegmentation. Finally we give some object tracking methods. According to the research ofthis paper, we want to form a union image segmentation framework with various cues(such as intensity, shape, texture, motion etc), which has a high intelligent ability in orderto solve some complex image segmentation problems. This paper also shows oneapplication, namely the left ventricle magnetic resonance (MR) image segmentation. Theprimary work and remarks of this paper are as follows:(1) This paper presents a new double-threshold image binarization method based onthe edge and intensity information. We first find the seeds near the image edges, andpresent an edge connection method to close the image edges. Then, we use closed imageedges to partition the binarized image that is generalized using a high threshold, and obtainthe primary binarization result by filling the partitioned binarized image with the seeds.Finally, the final binarization result is obtained by remedying the primary binarizationresult with the binarized image generalized using a low threshold. Compared with theclassical binarization methods and the binarization methods similar with our method thatare based on the edge information, our method is effective on the binarization of imageswith low contrast and inhomogeneous object intensities.(2) We propose a level set implementation of the Mumford-Shah model integratingprior shape statistical knowledge. The statistical shape based approach to the image segmentation using levcl sets mainly consists of the constructions of the prior shape modeland the shape energy term. Aiming at these two parts, we mainly do two pieces of work: (1)A simple and feasible construction method of the prior shape model is proposed, which isbased on binary images; (2) A new construction method of shape energy term is presented,which considers the global and local shape information at the same time, and withoutintroducing pose parameters makes evolving surface stable.(3) This paper describes a fast and active texture segmentation approach based on theorientation and the local variance. First, a set of feature images are extracted using theorientation and the local variance. To reduce the computational complexity, a separabilitymeasurement method, which is used for selecting four feature images with goodseparability in four orientations, is proposed in this paper. To improve the segmentation,we adopt a nonlinear diffusion filtering to smooth the four feature images. Finally, avariational framework incorporating these features in a level set based, unsupervisedsegmentation process is adopted. To improve the computational speed, instead of solvingthe Euler-Lagrange equation, we calculate the energy, with level set representation, tosolve the variational framework. Segmentation results of various synthetic and realtextured images has demonstrated that our method has good performance and efficiency.(4) This paper presents three object tracking methods based on object contours:●This paper present an object tracking method using edge-based shape matching isproposed, which presents a new shape similarity measurement according to the imageedge. To guarantee the tracking precision and handle the local distortion of objects, wepresent a new shape similarity measurement method based on the edge image in theweighted narrow band.●A parametric active contour model is presented for object tracking based on matchingdegree image of object contour points. We first construct a matching degree imageaccording to object contour points, and track the object using parametric activecontours. This paper presents a new feature matching approach and a new directionalfilter. Assuming that the motion of objects is small in this paper, we constrain themotion of object contour within the contour vicinity defined by a band, which isconstructed by the generation method of narrow band of level set. Experimental resultsdemonstrate that our method can effectively track rigid and non-rigid objects.●This paper presents a two-stage object tracking method by combining region-basedmethod and contour-based method. First, a kernel-based method is adopted to locatethe object region. Then diffusion snake is used to evolve the object contour in order to improve the tracking precision. In the first object localization stage, the initial targetposition is predicted and evaluated by Kalman filter and Bhattacharyya coefficient,respectively. In the contour evolution stage, the active contour is evolved based on theobject feature image that is generated with the color space component array (CSCA).In the process of the evolution, similarities of the target region are compared to ensurethat the object contour evolves in the right way.(5) This paper presents two segmentation methods of left ventricle MR images:●According to the characteristics of tagged MR images, we introduced an automaticsegmentation method based on multistage hybrid processing. First the left ventricle islocated using morphological method, and then the inner and outer contours areinitialized using k-mean clustering, templet matching and the myocardium shaperestoration based on skeleton. At last, the initial contour lines are evolved usingimproved level set method to achieve object boundaries.●According to the characteristics of MR images without tag lines, we present animproved shape statistics variational approach for the outer contour segmentation ofleft ventricle MR images. We use the Mumford-Shah model in an object feature spaceand incorporate the shape statistics and an edge image to the variational framework.The introduction of shape statistics can improve the segmentation with brokenboundaries. The edge image can enhance the weak boundary and thus improve thesegmentation precision. The generation of the object feature image which hashomogenous "intensities" in the left ventricle facilitates the application of theMumford-Shah model.更多还原