【作者】 王松涛；

【导师】 潘昊；

【作者基本信息】 武汉理工大学 ， 计算机科学与技术， 2011， 硕士

【摘要】 点模式匹配是计算机视觉和模式识别中基础而重要的问题,也是目前各领域关注的一个热点。它是一个普遍存在的问题,不仅限于计算机视觉领域,在航位与姿态估计、目标识别、遥感图像配准、医学图像配准、计算生物和化学等方面都有广泛的应用。而谱图理论应用于该课题的研究具有计算效率高、效果较好的优势。本文通过分析Scott和Longuet-Higgins, Shapiro和Brady两种点模式匹配谱图分析法的不足,提出了一种新的点模式匹配的谱图分析方法。方法的核心部分是构造了一种新的邻近矩阵,这种方法在图像较大幅度仿射变换下和点抖动的情况下得到的结果要好于前面两种方法,在算法的复杂度上具有谱图法的优点,效率比近几年提出的一些需要迭代的方法高。新方法经过了较为全面的检验,包括在合成数据和真实数据上的测试,都取得了较好的效果。本文在总结以往文献中谱图理论匹配算法的基础上,做出了一些改进和提炼,提出了一个算法框架,算法分为两个步骤,第一步是选择距离函数的组合,框架本身提供一组改进的距离函数,也可以根据应用需要另外加入函数。最终得到的距离函数为所选函数的乘积,这样的乘积可以突出适合特定匹配需求的距离函数的作用。第二步是构建距离函数的高斯加权邻近矩阵或者Laplacian矩阵,作SVD分解,得到匹配关系。这个框架的优点可根据具体应用的需要生成多种算法,具有较好的灵活性和适应性,在实现上,每一步产生的代码可以复用,易于编程实现。更多还原

【Abstract】 Point pattern matching is an essential part of many computer version or pattern recognition tasks; it has very wide range of applications, and is currently the hotpot of various studies. Now, Point pattern matching is widely used in stereovision, autonomous navigation, object recognition and tracking, medical imaging analysis, remote imaging registration, drug design and DNA sequences prediction, etc. Spectral graph theory is applied to this problem have some advantages, such as high efficiency and good effect.This work analyzed the drawback of Scott and Longuet-Higgins method and Shapiro and Brady method, proposed a novel approach of spectral graph method. The most important part of this new method is constructing a new-type proximity matrix. This method does pretty well in the condition of large scale transformation and point-jitter of images, and has the advantages of graph spectral method: algorithmic complexity is better than iterative method proposed in last two decades. The new method proposed in this work were tested extensively in heterogeneous case, include synthetic data and real image taken from benchmark of computer version.This work summarizes the spectral graph methods about the matching algorithm in previous papers, and made some improvements and refinement, proposed an algorithm framework. The algorithm framework is divided into two steps, the first step is to choose a combination of distance function, the framework itself provides a set of improved distance function, and additional functions can also be added depending on the application requirement. The resulting distance function is the product of the selected function, this product give prominence to the distance functions that are appropriate for the particular application. The second step is to build the Gaussian weighted proximity matrix or Laplacian matrix for singular value decomposition. The advantages of this framework is algorithms can be generated according to the specific needs of applications, the framework have good flexibility in implementation, in each step, the generated code can be re-used, reduce the programming time.更多还原