相似度距离耦合角度径向变换的图像配准算法

Image Registration Algorithm Based on Similarity Distance Coupled with Radial Angle Transformation

目的为了解决当前图像配准算法因利用l1距离或l2距离相似度测量手段来完成图像特征点匹配,使其忽略了相位信息,难以有效消除高斯噪声的影响,使其配准精度与效率不佳不足的问题。方法提出最优相似度距离耦合角度径向变换的抗噪图像配准算法。首先引入角度径向变换,以降低算法复杂度,快速提取图像的特征点。然后联合图像的幅度和相位信息,基于欧式距离测度,定义最优相似度距离测量模型,通过求解其全局最小值,对特征点完成匹配,提高算法的抗噪性能。最后将图像分割为内点与外点,择取6个内点,通过计算其变换矩的几何配准误差,改进随机样本一致策略,对匹配进行提纯,消除误配。结果仿真实验结果显示,与当前基于l1距离或l2距离相似度测量的图像配准技术相比,该算法具有更强的抗高斯噪声性能和更高的匹配精度,且算法时耗最短。结论所提算法能够精确完成图像特征配准。

The work aims to solve the defects such as low registration accuracy and poor efficiency of current image registration algorithm due to the application of l_1 distance or l_2 distance similarity measuring means to complete image feature matching, causing it to ignore the phase information and making it hard to effectively eliminate the effect of Gaussian noise. The anti-noise image registration algorithm based on optimal similarity distance coupled with radial angle transformation was proposed. Firstly, image feature point was fast extracted by introducing the radial angle transformation to reduce the complexity of the algorithm; then the optimal similarity distance measurement model was defined by combining the image amplitude and phase information and based on the measurement of Euclidean distance to improve the anti-noise performance to complete the feature points matching through solving its global minimum. Finally, the image was divided into inner points and outer points. 6 inner points were selected to improve the consistent strategies of random samples to purify the matching and eliminate the mismatching through the calculation of their geometrical registration errors of transformation matrix. Simulation results showed that： this algorithm had stronger anti Gaussian noise performance and higher matching accuracy, as well as the shortest consumption of time compared with the current image registration techniques measured based on l_1 distance or l_2 distance similarity. The algorithm mentioned can accurately complete the image feature registration.