Abstract: According to the error and the interpolation artifact using mutual information directly for image registration，a computational method on frequency-domain for the estimation of the mutual-information is proposed combining the image spectral characteristics and the gradient ascent algorithm which is improved by bringing in the idea of annealing is used to find the peak of mutual information alternately.The correlation-length distance is used to estimate the region of the optimal parameters，and this algorithm can be initialized close to the optimal registration parameters.The results of the experiment indicate that it has obvious convergence and stability compared with traditional methods for multispectral remote sensing image.

摘要： 针对直接利用互信息进行图像配准存在的误差和插值假象问题，结合图像的频谱特性提出了基于频域的互信息计算方法，引入退火的思想改进了梯度上升法，利用它迭代搜索互信息最大值，使用相关长度估算最佳参数域，使得参数初始化更接近于最大值。实验结果表明，该方法对于多谱段遥感图像，较之传统方法具有明显的收敛性和稳定性。