基于尺度不变特征和互信息的遥感图像自动配准

发布时间：2018-05-23 09:04

本文选题：图像配准 + 互信息　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：近年来，对于图像配准的研究受到了广泛的关注。图像配准能够实现不同拍摄条件下获得的图像间的匹配，已被应用于诸多领域，如医学图像分析中的疾病状态监测、计算机视觉中的三维图像重建，遥感图像中的变换检测。这些应用对图像配准提出了很高的要求，，如亚像素的精确度、全自动化和实时性。目前还没有一种完美的配准方法能够满足所有的要求，特别是在遥感图像处理领域。遥感图像受传感器机理、光照变换等因素的影响，图像间灰度差异较大，增加了图像配准的难度。针对遥感图像的特点，本文基于尺度不变特征（SIFT）和互信息（MI）提出了一种新图像配准策略。具体内容阐述如下： (1)提出一种基于尺度不变性的SIFT特征错误匹配滤除方法。该方法计算待配准图像中两两特征间距离与参考图像中相应特征间距离的比值，形成尺度直方图，直方图的波峰对应两幅图像的尺度比，直方图中边缘点被认为是错误配准并滤除。尺度直方图利用了SIFT的尺度不变性，获得更加准确的滤除结果。 (2)提出一种新的最大化互信息初始解选择策略。该策略获得SIFT匹配，并滤除其中的错误匹配，然后利用最小二乘法估计出配准参数。通过该策略获得的配准参数在最优解的附近，因此将其作为最大化互信息的初始解能够实现最大化互信息全自动化，降低互信息迭代次数，加快搜索算法收敛。 (3)利用Levenberg Marquardt最大化互信息实现遥感图像配准。Levenberg Marquardt算法结合了梯度下降法的鲁棒性和牛顿法的快速收敛性。利用该算法的优越性能，能够快速的获得精确配准参数。
[Abstract]:In recent years, the research on image registration has received extensive attention. Image registration has been applied to many fields, such as disease state monitoring in medical image analysis, 3D image reconstruction in computer vision and transform detection in remote sensing image. These applications require very high image registration, such as subpixel accuracy, full automation and real-time. At present, there is no perfect registration method to meet all the requirements, especially in the field of remote sensing image processing. Remote sensing images are affected by sensor mechanism, illumination transformation and other factors. The difference of gray level between images increases the difficulty of image registration. According to the characteristics of remote sensing images, a new image registration strategy based on scale-invariant features (sift) and mutual information (MII) is proposed in this paper. The details are as follows: A scale invariant based SIFT feature matching filtering method is proposed. This method calculates the ratio of the distance between the pairwise features in the image to be registered and the distance between the corresponding features in the reference image, and forms the scale histogram, and the wave peak of the histogram corresponds to the scale ratio of the two images. Edge points in a histogram are considered to be mismatched and filtered. The scale histogram utilizes the scale invariance of SIFT to obtain more accurate filtering results. A new strategy of maximizing mutual information initial solution selection is proposed. The SIFT matching is obtained, the error matching is filtered out, and the registration parameters are estimated by the least square method. The registration parameters obtained by this strategy are near the optimal solution. Therefore, as the initial solution of maximizing mutual information, the algorithm can realize the full automation of maximization of mutual information, reduce the number of iterations of mutual information, and accelerate the convergence of the search algorithm. 3) using Levenberg / Marquardt to maximize mutual information to realize remote sensing image registration. Levenberg Marquardt algorithm combines the robustness of gradient descent method and the fast convergence of Newton method. By using the superior performance of the algorithm, accurate registration parameters can be obtained quickly.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】