针对超分辨率图像重建中图像运动估计精度要求高，速度要求快的问题，对传统的基于Fourier-Mellin变换和Keren算法的运动估计方法做出以下改进：首先提取参考图像和待估计图像的边缘，从而避免了Fourier-Mellin变换的不足(对细节不明显的图像运动估计精度极差)；由于只是用Fourier-Mellin变换进行粗估计，对角度估计精度要求不高，只需小于1°，因此在进行对数极坐标变换时，可以减少角度坐标和对数坐标的采样点数，大幅缩小了矩阵大小，提高了运动估计速度；由于先用Fourier-Mellin对待估计图像进行粗估计，Keren算法可以避开复杂的金字塔计算而只需一层估计，减少了运动估计时间。在VC++中的仿真实验表明，该方法有效地结合了Fourier-Mellin变换和Keren算法的优点，同时又提高了运动估计速度。经测试，用未改进的算法对328×500像素大小的两幅图像进行运算估计需要3.53 s，而用改进的算法则只需要1.15 s，大大提高了运动估计速度。

Because Super-resolution image reconstruction requires high precision and high speed of motion estimation of image,so it is proposed to make three improvements with the traditional Fourier-Mellin transform and Keren algorithm of motion estimation：to avoid the shortcoming of Fourier-Mellin transformation (bad registration accuracy of image with poor details), edge detection of the image in advance is made; as Fourier-Mellin transformation is used for coarse estimation,the number of sampling points is reduced during the log polar transform with the angle accuracy below 1°, which greatly reduced the size of matrix, and increased the rate of registration; with Fourier-Mellin transformation being used for coarse estimation firstly, do not need to use the pyramid of Keren algorithm, but only need one layer for motion estimation so that to reduce the registration time. The simulation in VC++ shows that the method is effective in maintaining the advantages of Fourier-Mellin transform and Keren algorithm,at the same time improving the speed of motion estimation.Accordding to the experiment,with the traditional algorithm it needs 3.53 s to do motion estimation of image with 328 multiply 500 pixels,but needs 1.15 s with the improved algorithm which greatly improve the speed.