基于聚类一致性的多单应性矩阵计算与图像配准  

Multi-Homography Calculation and Image Registration based on Cluster Consensus

作　　者：丁浩宇[1] 段嘉旭 郭建强[1] 高晓蓉[1] 

机构地区：[1]西南交通大学物理科学与技术学院光电工程研究所,四川成都610031

出　　处：《通信技术》2017年第8期1674-1682,共9页Communications Technology

摘　　要：图像配准在超声无损检测领域是一项关键性技术,可以有效消除图像中存在的偏差,提高检测准确度。为了有效提高图像配准的精度和效率,提出了一种基于模糊聚类和最大相关的图像配准方法。该算法通过对图像聚类形成多个点集,然后计算出对于每个点集对应的单应性矩阵,从而形成多种点坐标变换模式,并通过点在图像平面上按一定规律的移动实现图像配准。具体地,实验部分分别采集了多幅具有随机偏差的超声TOFD图像,和多幅利用计算机软件模拟得到的合成孔径聚焦成像算法形成的医学囊肿图像。在图像中含有噪声的情况下,所提的配准算法与传统的RANSAC算法配准结果相比,配准图像的信噪比有了较大幅度提升,均方根误差有了一定减小,表明所提方法具有较好的稳定性和图像配准效果。Image registration, as a key technique in nondestructive testing field, can be used to eliminate the deviation in images and raise the test accuracy and reliablity. To improve the accuracy and efficiency of image registration, a new registration method based on the combination of fuzzy clustering and maximum correlation is proposed. This method generates multiple pixel-sets via clustering, and the homography matrix is calculated by some operations of the pixel-sets, then a multi-transform model is thus acquired. Image registration is realized by the movement of the pixels. The experiment part illustrates the sampling of ultrasonic TOFD images with random deviation and the generation of the software-simulated medical cyst images via synthetic aperture focus imaging technique. Compared with the registration result by the algorithm based on RANSAC, and by taking the noise into consideration, the signal-to-noise ratio（SNR） is improved and the root mean square error（RMSE） reduced by the proposed algorithm, and all this means that this method is of fairly good stability and could achieve optimal image-registration result.

关 键 词：图像配准 单应性矩阵 RANSAC 模糊聚类 点最大相关 信噪比 均方根误差 

分 类 号：O426[理学—声学]