基于Hausdorff距离和变参差分进化的SAR图像配准  

SAR image registration based on Hausdorff distance and variable parameter differential evolution

作　　者：黄玲 陈浩文 张伟[4] Huang Ling;Chen Haowen;Zhang Wei(Hunan Aircraft Maintenance Engineering Technology Research Center,Changsha 410124,China;School of Educational Science,Hunan Normal University,Changsha 410081,China;College of Computer Science and Electronic Engineering,Hunan University,Changsha 410012,China;School of Computer Science and Engineering,Central South University,Changsha 410082,China)

机构地区：[1]湖南省飞机维修工程技术研究中心,湖南长沙410124 [2]湖南师范大学教育科学学院,湖南长沙410081 [3]湖南大学信息科学与工程学院,湖南长沙410012 [4]中南大学计算机学院,湖南长沙410082

出　　处：《南京理工大学学报》2023年第3期337-342,372,共7页Journal of Nanjing University of Science and Technology

基　　金：湖南省职业教育教学改革研究项目(ZJGB2021286)。

摘　　要：为了提高合成孔径雷达(SAR)图像配准性能,提出了一种基于变参差分进化算法的图像配准方法。首先,获取了原始图像和待配准图像样本。然后,建立了基于差分进化(DE)算法的SAR图像配准模型。通过交叉、变异和选择操作等进化方式获得最优匹配特征点对,并根据特征点对获得图像配准结果。为了进一步优化DE算法在SAR图像配准中的性能,对差分缩放因子引入自适应策略。采用变参DE算法进行图像配准,弥补了因差分缩放因子设置不当而导致配准精度降低的缺点。在DE算法的适应度函数选择上,选取图像特征的Hausdorff距离函数作为相似性度量方法。实验结果显示,相比尺寸不变特征变换(SIFT)算法、卷积神经网络(CNN)算法,该文算法能够获得更高的配准准确度,接近100%,且配准稳定性更高,均方根误差(RMSE)均值为0.506 7。In order to improve the performance of synthetic aperture radar(SAR)image registration,an image registration method based on variable parameter differential evolution algorithm is proposed.Firstly,the original image and the image samples to be registered are obtained.Then,a SAR image registration model based on differential evolution(DE)algorithm is established.Through crossover,mutation and selection,the best matching feature point pairs are obtained,and the image registration results are obtained according to the feature point pairs.In order to further optimize the performance of DE algorithm in SAR image registration,an adaptive strategy for differential scaling factor is introduced.The variable parameter differential evolution algorithm is used for image registration,making up for the shortcoming of low registration accuracy caused by improper setting of differential scaling factor.In the selection of fitness function of DE algorithm,Hausdorff distance function of image features is selected as the similarity measure.The experimental results show that compared to the scale invariant feature transform(SIFT)algorithm and convolutional neural network(CNN)algorithm,the proposed algorithm can achieve higher registration accuracy,approaching 100%,and higher registration stability,with an root-mean-square error(RMSE)average of 0.5067.

关 键 词：HAUSDORFF距离 差分进化 合成孔径雷达 图像配准 差分缩放因子 适应度函数 尺寸不变特征变换 卷积神经网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]