基于相位一致性计算和RS-GLOH描述子的SAR和光学图像匹配

SAR and optical image matching based on phase consistency calculation and RS-GLOH descriptor

为解决合成孔径雷达(SAR)和光学图像之间由于非线性辐射和几何差异导致的匹配困难问题,提出一种新的基于特征的SAR和光学图像匹配算法。主要分为3个部分:特征提取、特征描述和特征匹配。在特征提取阶段,针对两者间的非线性辐射差异,使用基于相位一致性的方法得到SAR和光学图像的最小矩图和最大矩图,通过极值点检测与非极大值抑制提取出分布均匀且位置对应性好的最小矩点和最大矩点;在特征描述阶段,针对两者间的非线性几何差异,提出一种RS-GLOH描述子构建方法,主要使用斑点噪声抑制效果好的ROEWA算子和Sobel算子分别计算SAR和光学图像的梯度幅值和方向信息,然后建立梯度定位和方向直方图对特征点进行描述;在特征匹配阶段,结合最近邻距离比和快速抽样一致性分别匹配最小矩点和最大矩点。实验结果表明,本文方法具有旋转和尺度不变性,并且相比于PSO-SIFT、SAR-SIFT和OS-SIFT方法,本文方法在5组不同区域的SAR和光学图像对中可以得到更多的正确匹配点对数和更低的均方根误差。

In order to solve the difficulty in matching between synthetic aperture radar(SAR)and optical images due to nonlinear radiation differences and geometric differences,a new feature-based matching algorithm for SAR and optical images is proposed in this paper.This algorithm is mainly divided into three stages:feature extraction,feature description and feature matching.In the feature extraction stage,according to the nonlinear radiation difference between the two kinds of image,the minimum and maximum moment maps of SAR and optical images are obtained using the method based on phase consistency,and the minimum moment point and maximum moment point are obtained through extreme point detection and non-maximum suppression.In the feature description stage,in view of the nonlinear geometric difference between the two images,the ROEWA operator with good speckle noise suppression effects and Sobel operator are used to calculate the gradient amplitude and direction information of the SAR and optical images,respectively.Then gradient location orientation histogram(GLOH)is set up to describe the feature points(RS-GLOH).In the feature matching stage,method of combining the nearest neighbor matching(NNDR)and fast sampling consistency(FSC)is used to match the minimum and maximum moments respectively point.Experimental results show that the proposed method has excellent rotation and scale invariance.Compared with PSO-SIFT,SAR-SIFT,and OS-SIFT,more correct matching points and lower root mean square error(RMSE)can be obtained in 5 sets of SAR and optical image pairs in different regions.