【作者】 魏宁；

【导师】 万毅；

【作者基本信息】 兰州大学 ， 无线电物理， 2009， 博士

【摘要】 图像匹配算法的研究是模式识别学科研究的一个分支,该问题的核心是找到两幅图像在空间上或几何上的最佳映射函数,使得通过该映射函数能够将两幅图像相似的部分尽可能地进行重叠。在很多的研究领域,图像匹配算法的准确性及计算效率都对研究的最终结果起到了至关重要的作用。图像匹配由于其应用背景的广泛性和数据获得的多样性,导致了各种匹配算法的适用范围也具有一定的局限性。在图像的获取方式中,有一类是先获取物体的一系列投影数据然后通过这些数据还原出物体内部的时域图像,现代影像诊断医学中CT(Computerizedtomography)、PET(Positron Emission Tomography)等都是通过类似的方法得到人体的内部图像。对于此类图像的匹配目前是在恢复后的时域图像中完成的。本文对目前常见的投影模型如RT(Radon Transform)变换、DBT(DivergentBeam Transfrom)变换,以及在模式识别中常用的HT(Hough Transform)变换从性质和实现上进行了研究,并分析了它们之间的转换关系。在此基础上,本文对此类途径获取的图像提出了一种新的图像匹配思路:直接在投影域利用投影数据来对时域图像进行匹配,这样避免了对已经存在一定误差的时域图像来进行匹配。根据本文所调查的资料,由于变换参数之间所存在的复杂依赖关系,目前还没有研究能够完全在投影域中完成对存在仿射变换图像的匹配。在本文中提出了一种基于投影变换的快速图像匹配算法,该算法完全在投影域分离了参数之间的依赖关系,可以直接对所获得的投影数据使用基于FFT的运算和最小二乘方法就能完成对时域中两幅完全重叠且存在仿射变换图像的匹配,算法对在该背景下获得图像的匹配具有很高的计算效率与对噪声的鲁棒性。通过数值仿真说明,该算法与目前最新的类似研究相比(如[113])在计算效率上有了大幅的提高。该算法同样适用于普通图像的匹配,与目前常见的基于FFT的快速匹配算法相比在效率和准确性上也都具有一定的优势。为了保证所提出的快速算法对参数估计的准确性,本文同时还提出了一种改进的相位相关算法,该算法克服了传统方法对低频信号匹配误差较大的缺陷,并且在实际应用中获得了较好的匹配效果。论文还对目前经典的图像匹配算法从基于图像全局灰度值的和基于图像局部特征的角度进行了系统的总结,分析了各种算.法所具有的优势和局限性,并介绍了目前应用最为广泛的基于图像特征点的SIFT(Scale Invariant FeatureTransfrom)算法,以及一种基于图像全局灰度值一阶矩的快速匹配方法。更多还原

【Abstract】 The study of image registration algorithm is one of the branches in the field of Pattern recognition. The core of the problem is to find the best mapping function between two images where one image is a transformed version of the other. This mapping function needs to make the similar part of the two images overlapped as much as possible. The efficiency and accuracy of image registration algorithm plays an important role in the final results in many fields. Since image registration has an extensive application area with widely varied data sources, so far different methods are somewhat limited only to certain application problems.Among the image acquisition approaches, one of them is to acquire a series of projecting data at first, then we could recover the image of the inner part of the objects through these data. Both CT(Computerized Tomography) and PET(Positron Emission Tomography) in modern diagnostic medical imaging obtain the image of inner human body through similar methods. Registration of this kind of images are accomplished in the recovered time domain images. This thesis focuses on the study of the nature and realization of some common projecting models, such as RT(Radon Transform), DBT(Divergent Beam Transform) and HT(Hough Transform) which is frequently used in the pattern recognition research. The work in this thesis also includes the analysis of the switching relationships between them and makes a summary of fast algorithms for RT. On the basis of these, the thesis puts forward a new idea for registrating the images obtained from projection data. The idea is registrating the images only from their projections. In this way, we can avoid registrating images which already contain some errors when recovered from projection. So far no results have been obtained that solves the recognition problem completely in the projection domain due to coupling of transform parameters. We develop a novel algorithm based on this new idea for such parameter decoupling in the projection domain. The algorithm can accomplish the registration of two images which have affine tranform relationship through using FFT and least squares method on their projecting data. Simulation shows that the proposed algorithm has a high computational efficiency and is robust to noise. Compared with the state-of-the-art approaches such as [113], the proposed algorithm has a big improvement, in computational efficiency. The algorithm is also suitable for the registration of ordinary images. It also has some advantages in computation efficiency and accuracy compared to the image registration algorithms based on FFT. For the accuracy of the parameter estimation in the proposed fast algorithm, the thesis also puts forward an improved Phase Correlation method. The method overcomes the defect of the traditional method which has a big error in low-frequency signal registration. The improved Phase Correlation method achieves quite good registration effects in the proposed fast algorithm.This thesis systematically makes a summery of the classical image registration algorithms from the point of view of the image global gray value and the image local feature. We also analyze the advantages and disadvantages of various algorithms. And an introduction of currently extensively used SIFT(Scale Invariant Feature Transform) method which is based on the image local feature points and a fast registration algorithm which is based on the first-order moment of the images is made.更多还原