鲁棒KAZE算法的SAR图像配准  

Robust SAR Image Registration Based on KAZE Algorithm

作　　者：吴鹏[1] 于秋则[1] 闵顺新 逄博[1] 

机构地区：[1]武汉大学电子信息学院,武汉430072

出　　处：《科学技术与工程》2016年第28期101-109,共9页Science Technology and Engineering

基　　金：国家863计划(2015AA7123050);国家自然科学基金(61174196);深圳市科技计划项目(合同编号:JCYJ20150513162829635)资助

摘　　要：针对传统的SIFT算法及其改进算法在多波段SAR图像匹配性能低下,对非线性扩散滤波的KAZE特征检测算法分别从构建尺度空间和特征点检测两个方面进行改进,来提高多波段SAR图像匹配性能。尺度空间构建方面:提出采用高斯引导滤波取代KAZE算法中的非线性扩散滤波,充分利用高斯引导滤波的算法实时性、尺度不变性和良好的边缘保持性来克服非线性扩散滤波构建尺度空间的算法耗时较高的问题。特征点检测方面:提出采用引入指数加权均值比例算子(ROEWA)结合多尺度Harris角点检测的改进方法,充分利用ROEWA算子抑制相干斑噪声和多尺度Harris算子计算简单、旋转、尺度不变等优点,克服了传统的梯度差分方法检测SAR图像特征点时在SAR图像强散射纹理区造成大量错误率。描述符构建阶段,在梯度比例图像上采用原始的KAZE算法相同原理构建MSURF描述符。最后通过最近邻匹配和一致性倒数(RANSAC)进行错误匹配剔除。实验结果表明本文提出的改进KAZE算法应用于SAR图像配准优于原始KAZE特征算法。To solve the traditional SIFT algorithm and its improved algorithm in multi-band characteristics of SAR image registration poor performance, a new algorithm which is based on the KAZE was presented. To improve KAZE algorithm from two aspects, one is scale space construction, the other is the feature point detection. On the aspect of construction scale space dimension, Gauss-Guided filter was used instead of nonlinear diffusion filter, re- al-time and edge-preserving Gauss-Guided filter was sed to solve nonlinear diffusion filter algorithm existing higher scale space consuming problems. On the aspect of feature point detection levels, introducing an exponentially weigh- ted average of the proportion of operator（ROEWA） was combined with to improve method of multiseale Harris cor- ner detection, making full use of ROEWA operator which suppressed speckle noise and multiscale Harris operator which is simple calculation, rotation and scale invariant, overcomes the traditional gradient of difference method to detect SAR image feature points error caused by a large number of feature points detection in high contrast areas. Local descriptors was built by using M-SURF descriptor which is equal to original KAZE way except using the gradi- ent of ratio replace the gradient of difference. Finally, the coarse match were finished by nearest neighbor algorithm and the false match were eliminated by RANSAC. Experimental results show that the improved KAZE algorithm, applied to SAR image registration, is better than the original KAZE feature algorithm.

关 键 词：多波段合成孔径雷达(SAR) 尺度空间 高斯引导滤法(KAZE) 引导滤波 指数加权均值比例算子(ROEWA) 多尺度Harris 最近邻匹配和一致性倒数(RANSAC) 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]