In order to decrease the number of local extrema caused by over subsampling of conventional subsampling methods in intensity-based medical image registration,a new subsampling method based on intensity probability distribution and graclient information of floating image is proposed. The function curves and convergence are compared by applying five subsampling methods to the rigid registration of brain images. The results of tests show that the new subsampling method outperforms other subsampling methods in less extrema and convergence performance of the normalized mutual information.