【作者】 娄联堂；

【导师】 丁明跃；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2005， 博士

【摘要】 图像分割是图像处理、计算机视觉、模式识别中的核心问题,对它们的发展有非常大的影响。目标轮廓提取是图像分割重要研究内容,在图像识别与图像分析中占有重要地位,已广范应用于军事、医学图像分析等许多领域,取得了令人瞩目的研究成果。该文针对目标轮廓提取方法及应用,从以下四个方面进行了研究:(1) 基于偏微分方程目标轮廓提取方法; (2) 基于量子力学中粒子运动规律的目标轮廓提取方法; (3) 具有仿射不变性的目标轮廓自动提取; (4) 将具有仿射不变性的目标轮廓自动提取方法应用于舰船打击效果评估。该文首先对目前常用的基于偏微分方程目标轮廓提取方法,如主动轮廓模型、目标轮廓能量全局最小主动轮廓模型、拓扑自适性Snake 模型,以及水平集分割方法等作了比较,指出了各自的优缺点。在此基础上,改进了基于最小作用曲面与鞍点的封闭轮廓曲线提取方法,提出了一种基于最小作用曲面及图像二分法的封闭轮廓提取方法,用一条直线将图像分成两幅小图像,以一种简单的方式解决了封闭轮廓曲线检测问题,避免了复杂的鞍点检测过程。同时,从理论上分析了基于最小作用曲面及图像二分法的封闭轮廓曲线提取方法的算法复杂性,证明了该方法的运行时间比基于最小作用曲面与鞍点的封闭轮廓曲线提取方法快。自Snake 模型出现以来,人们对其作了许多改进,提出了许多不同的目标轮廓提取方法。但是从本质上讲,它们均属于基于经典力学中能量最小方法或者与之等价的方法(如牛顿定律: 力的平衡,合力为零),其自身存在固有缺点。例如,图像含有噪声,而经典力学则没有考虑图像和目标轮廓的这种统计特性。为了解决这一问题,该文将量子力学中关于粒子运动规律引入目标轮廓提取。通过对经典力学与量子力学中粒子运动规律的分析和类比,提出了量子轮廓模型新概念,给出了预测粒子运动位置集合的两种方式:线模式与扇形模式; 讨论了粒子从一点运动到另一点的概率计算问题; 研究了量子轮廓模型中有分枝目标轮廓的提取、轮廓曲线的收敛性及轮廓曲线光滑性问题; 提出了多粒子量子轮廓模型; 最后分析了基于量子力学中粒子运动规律的目标轮廓提取方法的边缘检测定位性能及算法时间复杂性。实验结果表明,基于量子力学的目标轮廓提取方法具有定位精度高,运算速度快等优点。更多还原

【Abstract】 Image segmentation is a central problem of computer vision. Contour extraction is one of the most important aspects of image segmentation, extensively applied in image analysis, and has a great significance in both theoretical research and industrial applications. This dissertation focuses on contour extraction of objects and concentrates on the following aspects: (1) the approach of contour extraction based on partial differential equations (PDE); (2) contour extraction based on particle motion in quantum mechanics; (3) automatic extraction of affine invariant contour; (4) battle damage assessment (BDA) of naval vessel based on image understanding. Approaches of contour extraction based on PDE, e.g. active contour models, also known as snakes, are being extensively used to solve the problem of image segmentation. In this thesis, the contour extraction approaches based on PDE are studied systematically in theory. First, we compared the typical contour extraction approaches based on PDE such as active contour models, global minimum for active contour models, topologically adaptable sankes, and level set methods, etc., and their merits and demerits were pointed out. Secondly, the approach of closed contour extraction based on global minimum optimization and detection of saddle points is improved. An approach of closed contour extraction based on surface of minimal action and dichotomy of image was presented. It divides the original image into two small images in the detection of closed contour so that a complicated detection of saddle points was avoided. Thirdly, computational complexity of our approach is theoretically analyzed, and it was proved that our improved approach has a runtime less than original approach. Many improvements in active contour models (also known as Snakes) have been made since they were introduced first by Kass et al. (1987). Active contour models are an energy-minimizing spline guided by external constraint forces and influenced by image forces, whose physical background is the principle of minimum action or the force equilibrium in classical mechanics. However, the contour of object in a noisy image is often blurred but this image statistical property is not considered in the classical mechanics model. The author explained that it is a logical extention that the law of particle motion in quantum mechanics is applied to contour extraction of objects. From this point of view, a new model, quantum contour model, is proposed. The probability for particle moving from a point to another point is estimated. The relationship between active contour model and quantum contour model is addressed. The concept of multiple particle quantum contour models is given. Furthmore, the extraction of boundary with branches, the convergence of contour extraction and the smoothing of the extracted contour were addressed, which are three problems faced when we apply quantum contour model to contour extraction of objects. Edge detection localization and computational complexity are analyzed. Experiments with simulated and real images demonstrated that the quantum contour based approach could extract a close boundary quickly, accurately and robustly with a single initial point close to the boundary of the object of interest. An important problem in object recognition is the fact that an object can be seen from different viewpoints, resulting in different images. For near planar object, these deformations can be modeled approximately by an affine transformation. In the thesis, automatic extraction of affine invariant contour of objects was dealed with. First, six parameters of affine transformation are normalized according to the practical problem in applications. Then, the concept of shape-specific line is given. The properties of shape-specific points and shape-specific lines are discussed. Using shape-specific lines, the parameters of scale, rotation and translation transforms are computed. The experiments showed that the contour registration method by shape-specific lines is better than the method by shape-specific points. Finally, the values of the parameters of affine transformations are computed using a Genetic Algorithm while the energy of contour is referred to as a fitness function. Automatic extraction of affine invariant contour is performed. The applications of contour extraction of objects are considered, such as image understanding based BDA of naval vessels. Two modes of BDA are presented. The building and automatic intelligent maintenance of BDA knowledge database on naval vessels and analysis of the attacked target properties were done.The contour extraction of naval vessel and recognition of attacked target in BDA were investigated. Tne mathematics model of BDA was established. The relationship between BDA of part and whole was discussed. By using statist ical decision-making method, the database maintenance of BDA knowledge database beccomes a simple quadratic programming problem.更多还原