Abstract: Aiming at the problem of Scale Invariant Feature Transform(SIFT) achieving low precision when registating SAR images,an image registration method based on improved SIFT is proposed.Based on the construction of SIFT descriptors and the Multi-scale Autoconvolution(MSA) affine invariant moments of the region around keypoints,Canonical Correlation Analysis(CCA) based fusion method is adopted to fuse them together.The control points are rough matched by using threshold and the distance as well as gray correlation around the matched points are used to construct the similarity matrix.The Singular Value Decomposition(SVD) method is subsequently adopted to implement image registration precisely.The parameters of affine transformation are calculated and the images are registrated.Experimental results show that the registration results of this method is better than SIFT method and achieves precision in sub-pixel level.

Key words: SAR image registration, Scale Invariant Feature Transform(SIFT), Multi-scale Autoconvolution (MSA), Canonical Correlation Analysis(CCA), Singular Value Decomposition(SVD)

摘要： 针对尺度不变特征变换(SIFT)配准方法在处理SAR图像时精度不高的问题,提出一种基于改进SIFT的精确配准方法。在提取关键点SIFT描述子及其邻域多尺度自卷积矩不变特征的基础上,利用基于典型相关分析的融合算法对SIFT与矩不变特征进行融合,形成新的关键点描述子,使用阈值实现粗匹配,并结合关键点的距离与邻域灰度相关性构建相似矩阵,采用奇异值分解方法精确确定匹配点对,求出仿射变换模型参数,从而完成图像配准。实验结果表明,该方法的配准结果优于SIFT方法,且配准精度达到亚像素级。

关键词: SAR图像配准, 尺度不变特征变换, 多尺度自卷积, 典型相关分析, 奇异值分解

CLC Number: 

• TP751