摘要
为获取亚像元级目标运动参数,提出一种基于相位相关分析的图像配准方法。首先讨论了目标局部运动和全局运动的目标参数估计问题,通过图像减影运算和模块匹配方法实现粗配准,从全景图像中分离目标信息和背景信息,计算目标中心坐标,获取像元级运动参数;然后采用相位相关图像配准方法实现精配准,利用傅里叶变换的平移特性,对产生平移的目标图像,通过求解归一化的互功率谱的傅立叶逆变换,得到二维脉冲函数,其峰值对应图像位移,由此获取亚像元级位移量。在实验室通过自准直光学系统获取光斑运动图像,使用Leica经纬仪标定光斑运动参数精度。结果表明,该方法效果显著,最大配准误差为0.156,标准差为0.091,配准精度优于1/10像元。
A high accuracy sub-pixel-level motion estimation of algorithm based on phase-only correlation was proposed.This algorithm consists of two steps: one for pixel-level correspondence estimation and the other for sub-pixel-level correspondence estimation.In pixel-level estimation,the method of image subtract or method of template matching are used to find the target image in the original input images for two classes of motion,namely local motion and global motion,then the pixel-level image shift enables to be detected coarsely by calculating the coordinate of center of the target image.In sub-pixel-level estimation,the method of image registration based on phase-only correlation(POC) is employed to detect sub-pixel-level displacement,which applies the translation characteristics of Fourier transform to search translation between images.The 2-D inverse Fourier transform for the normalized cross spectrum is a 2-D impulse function,the impulse enables to determine the translation displacement from the location of the correlation peak,so as to obtain finely image displacement.The method described in this paper has been applied to the experimental of registration of light-spot from autocollimator.The result shows that this method can detect sub-pixel-level shift parameters accurately and efficiently,which makes possible to estimate the displacement with 0.1-pixel accuracy,the maximum error of registration is only 0.156-pixel,and the root of mean square error of image registration is better than 0.091-pixel.
出处
《液晶与显示》
CAS
CSCD
北大核心
2011年第6期858-862,共5页
Chinese Journal of Liquid Crystals and Displays
基金
973国家重点基础研究发展计划(No.2009CB72400607)
国家高技术研究发展计划(863计划)项目(No.2008AA121803)
关键词
相位相关
亚像元
图像配准
运动估计
phase-only correlation
sub-pixel-level
image registration
motion estimation