一种适应于旋转、平移和大尺度变换的图像配准方法研究

Analysis of Image Registration Technique for Recovering Large Rotation,Translation and Scale Parameters

传统的基于Fourier-Mellin变换的图像配准算法对配准存在较小旋转、平移和尺度变换的图像对十分有效,但由于有限采样图像的旋转和缩放与离散傅立叶变换的不可交换性,以及采样和插值误差的存在,该算法对尺度变形较大的图像不再适用。本文深入分析了算法产生误差的根本原因,考察了误差对离散图像的傅立叶功率谱产生的影响,然后综合运用加窗和滤波等手段去除假象、改善频谱,改进后的算法可配准的尺度范围扩展到0.29～5,能满足绝大多数应用的需求。

The traditional image registration technique based on Fourier-Mellin transform was proved to be efficient for recovering small rotation, scale and translation parameters. But because of the interpolation error and the aliasing stems from the fact that the discrete-Fourier transform does not commute with the rotation and scaling of sampled-images, the technique is unsuitable for pairs of images subjected to large scale transform. This paper investigates factors that degrade the precision of the algorithm, and makes some compensation through certain of methods like windowing and filtering. The combination of techniques improves the performance and extends the scale range to 0.29 - 5, which meets the demands of most applications.