【作者】 徐玉华；

【导师】 张崇巍；

【作者基本信息】 合肥工业大学 ， 电力电子与电力传动， 2011， 博士

【摘要】 图像配准是图像处理领域中的一个基础问题，在机器人定位、遥感成像、医学图像分析、图像拼接、目标识别和定位、产品质量检测、导航制导等方面得到了广泛的应用。在过去的三十多年中，广大研究人员提出了大量的图像配准算法，但到目前为止，高精度的图像匹配和大变形的图像配准仍然是研究的热点。本文主要对有高精度要求的图像匹配方法及其应用和存在较大形变的图像的配准做了研究，所涉及的图像类型包括二维光强度图像和二维距离图像。完成了以下几个方面的工作：1、探讨了基于形状的高精度的模板匹配方法。采用基于Facet模型的亚像素边缘检测算法对形状模板中的边缘特征点进行亚像素定位，用ICP(Iterative Closest Point)算法精化基于形状的模板匹配结果，取得了与Halcon里面精度最高的模板匹配算法相当的精度。2、研究了基于形状的模板匹配方法在复杂多变环境下的圆测量中的应用。提出一种鲁棒的圆测量算法，采用基于形状的模板匹配方法对圆进行粗定位，再结合预先给定的圆的尺寸信息，确定圆的轮廓所处的环形区域。在局部的环形区域内自动确定Canny算子的双阈值，使得自适应阈值问题变得容易，不需要复杂的理论支持即可达到很好的效果。对提取出的边缘点用基于Facet模型的方法进行亚像素定位。用一种基于梯度方向的去噪算法去除边缘点中的噪声点。最后，用RANSAC算法估计圆的参数。实验结果表明，本文提出的圆测量算法的鲁棒性优于Halcon里面的圆测量算法，且在实际系统中得到了成功的应用。3、研究了基于Log-Polar变换和特征点的图像配准。针对传统的基于Log-Polar变换的配准方法存在的问题，提出一种结合Harris角点特征和Log-Polar变换的图像配准方法。对以待配准的两个Harris角点特征为中心的两个圆形图像窗口进行Log-Polar变换，再用NCC计算两幅Log-Polar图像的相关度作为这两个特征点的相似度。针对NCC计算量大的问题，一方面采用一维投影的方法对NCC进行加速；另一方面，采用Sum Table优化NCC的分母计算，用Jensen不等式构建NCC的终止条件，提前结束不能达到当前已经获得的最大匹配得分的NCC计算。实验结果表明，本文的配准方法的配准能力强于David Lowe的基于SIFT特征的匹配方法。为了配准更大变形的图像，从图像中提取Harris-Affine特征，把仿射不变的椭圆区域归一化为圆形区域，再进行Log-Polar变换，用NCC计算Log-Polar图像的相关度。实验结果表明，该方法比Mikolajczyk的方法获得了更多的正确的匹配对。4、研究了二维距离图像的配准问题及其在移动机器人位姿估计中的应用。针对ICP(Iterative Closest Point)算法在环境存在严重遮挡的情况下容易出现局部最小值的问题，对CP(Closest Point)规则进行了修改，提出双向最近点（DCP, Dual Closest Point）规则。DCP规则包含两次CP规则对应，使计算量增加了一倍。为了降低算法的复杂度，继而提出基于聚类的迭代双向最近点(IDCP BoC)算法。IDCP BoC对扫描数据进行聚类，在聚类的基础上进行数据精简。在相邻两次迭代的残差之差小于某个阈值之前，用精简数据进行迭代以提高计算速度，之后再改用非精简数据迭代以保证精度。实验结果表明，IDCP BoC算法能够有效避免ICP算法易陷入局部最小值的问题。更多还原

【Abstract】 Image registration is a basic problem in image processing, which finds it’s applications invarious fields like robot localization, remote sensing, medical image analysis, image mosaicing,target recognition and localization, quality control, navigation and guidance. Despite comprehensiveresearch spanning over thirty years, to register images precisely and to register images in thepresence of large deformations are still important research topics.In this dissertation, image registration methods which require high precision and to registerimages in the presence of large deformations are studied. The involved images include twodimensional light intensity images and range images. The accomplished works can be summarizedas follows:1. A shape-based template matching method with high precision is investigated. Subpixeledge detecting algorithm based on facet model is adopted to locate the feature points in shape modeland ICP algorithm is used to refine matching result. Comparisons with Halcon’s method showingthat the proposed method has comparable precision.2. Application of the shape-based template matching method for circle detection in highlyvariable environment is investigated. A robust method for circle detection is proposed. To begin with,the shape-based template matching method is adopted to locate the circle roughly. Then, an annulararea containing the circle’s contour can be defined with the rough location obtained by the matchingmethod and the size of the circle given in advance. So we can determine the double-threshold forCanny edge detector in the local annular area which is easy. To get subpixel location for Canny edgepoints, a facet model based method is used. Furthermore, to kick out the outliers, a de-noisingalgorithm based on gradient direction is developed. Finally, RANSAC algorithm is used to estimatethe circle parameters. Experimental results demonstrate that the method proposed outperformHalcon’s which is based on Hough transform.3. Image registration based on log-polar transform and feature points is explored. Aiming atthe problems of Zokai’s method, a registration method based on log-polar transform and Harris corner is proposed. Correlation between log-polar transforms of two circular windows whose centersare two Harris corners is calculated by NCC (Normalized Cross Correlation), which is considered asthe similarity of the two corners. But NCC is computationally expensive. On the one hand, onedimensional projection is utilized to speed up the process of NCC. On the other hand, sum table isused to optimize the calculation of the denominator of NCC and Jensen inequality is used toconstruct a termination condition of NCC. Experimental results show that the proposed methodoutperform David Lowe’s which is based on SIFT (Scale Invariance Feature Transform) features. Toregister images in the presence of larger deformations, Harris-Affine features are extracted from theimages. Harris-Affine features’ elliptical regions are normalized to circular regions, and thenlog-polar transform and NCC are used to calculate the similarity of the two features. Experimentalresults show that the method obtained much more correct match pairs than Mikolajczyk’s method.4. Registration method of2D range images and its application in pose estimation for mobilerobot is studied. To overcome the problem of local extrema existing in iterative closest point (ICP)algorithm when severe occlusions occur, the closest point (CP) rule is modified and dual closestpoint (DCP) rule is proposed. DCP rule contains twice CP correspondences so that computationcomplexity is doubled. To decrease the computation complexity, iterative dual closest point based onclustering (IDCP BoC) is proposed. Scan range points are divided into clusters and then a procedureof reducing the number of points is conducted. The reduced data set is used for iterative computationbefore the error of two consecutive iterations’ residual errors less than a preset threshold to speed upthe algorithm after that the data set which is not reduced is used to guarantee the accuracy.Experimental results show that IDCP BoC can avoid the problem of local extrema effectively.更多还原