Abstract: When images are aligned by non-uniform subsampling methods based on intensity probability distribution and gradient information of floating image, the subsampling strategies on pixels with the same intensity and gradient will change the subsampling results and then change registration process. The registration measure’s curved surfaces and convergence are compared by applying six subsampling strategies to rigid image registration under several subsampling rates. The results of tests show that the non-uniform stochastic subsampling strategy outperforms other subsampling strategies in convergence and anti-noise performance of the normalized mutual information.

Key words: image registration, normalized mutual information, non-uniform stochastic subsampling strategy

摘要： 使用基于浮动图像灰度概率分布和其梯度信息的非均匀抽样方法进行图像配准时，对同灰度和同梯度像素的抽样原则会影响到抽样结果和配准过程。通过对多种图像在多抽样率下进行刚体配准实验，从函数曲面和收敛性能方面，对比分析了6种数据抽样原则。实验结果表明，随机抽样原则的归一化互信息测度的收敛性能和抗噪声能力优于其他5种抽样原则。

关键词: 图像配准, 归一化互信息, 非均匀随机抽样原则