Abstract: This paper proposes a novel image registration algorithm based on Zernike moments and multilevel feedforward neural networks. Low order Zernike moments are used as image global pattern features and the algorithm feeds them into multilevel feedforward neural networks to provide translation, rotation and scaling parameters. The second level feedforward neural network is utilized behind the first level neural network to improve parameter estimation accuracy. Experimental results show that the proposed algorithm can provide accurate registration and robust to noise attack.

Key words: Zernike moment, feedforward neural network, image registration, affine transform, parameter estimation

摘要： 提出一种基于Zernike矩和多级前馈神经网络的图像配准算法。利用低阶Zernike矩表征图像的全局几何特征，通过多级前馈神经网络学习图像所经历的旋转、缩放和平移等仿射变换参数，在一级前馈神经网络的基础上添加二级前馈网络，以提高参数估计精度。仿真结果表明，与基于DCT系数的神经网络算法相比，该算法旋转、缩放和平移估计精度较高，对噪声的鲁棒性较强。

关键词: Zernike矩, 前馈神经网络, 图像配准, 仿射变换, 参数估计

CLC Number: 

• TN911.73