一种基于聚类分析的红外图像配准算法  

An Infrared Image Registration Algorithm Based on Clustering Analysis

作　　者：尹丽华[1,2,3,4] 李范鸣[1,2,4] 刘士建[1,2,4] 王霄[1,2,3] 

机构地区：[1]中国科学院上海技术物理研究所,上海200083 [2]中国科学院大学,北京100049 [3]上海科技大学,上海200031 [4]中国科学院红外探测与成像技术重点实验室,上海200083

出　　处：《半导体光电》2017年第4期571-576,共6页Semiconductor Optoelectronics

基　　金：国家"863"计划项目(2011AA7031002G);中国科学院红外探测与成像技术重点实验室项目;中科院上海技术物理研究所创新基金项目(CX-60)

摘　　要：为了提高红外图像匹配的精度和效率,提出了一种将Harris-Laplace关键点提取和旋转不变LBP特征描述算子相结合的局部特征检测新算法,该算法不仅在图像的尺度、光照和角度发生变化时,仍然能够得到很好的检测效果,而且能很好地描述图像的局部纹理特征。特征向量描述完成后,为了进一步提高红外图像特征点匹配的正确率,提出了一种基于K-means聚类分析的图像匹配策略。先利用Cosine余弦相关匹配策略实现特征点的初步粗匹配,接着采用K-means聚类分析匹配策略剔除图像中大部分的错误匹配。实验表明:提出的算法表现出良好的鲁棒性,关键点提取的重复率(Repeatability)提高了9.2%。与传统的匹配算法相比,采用基于K-means聚类分析的匹配策略匹配精度可以提高5.05%,匹配时间可以缩短0.068s。该特征描述算法和基于K-means聚类分析的匹配算法满足了红外图像配准的高精度性和高实时性的要求。In order to improve the accuracy and efficiency of infrared image matching, a new local feature detection algorithm based on combination of Harris-Laplace and the rotation invariant LBP was proposed. The algorithm could not only get good detection effect when image scale, the light, the angle changed, but also a good description of the local texture features of images. After completion of feature vectors described, in order to further improve the accuracy of infrared image feature points matching, this paper presented an image matching strategy based on K-means clustering analysis. Firstly, Cosine correlation matching strategy was used to achieve initial coarse matching feature points~ Secondly, using K-means clustering analysis excluded most of the matching strategy image mismatching points. Experimental results show that the characterization algorithm maintained a good robustness and repetition rate （Repeatability） improved by 9.2%. Compared with the traditional matching algorithm, matching precision matching strategy based K-means clustering analysis can be increased by 5.05 %, matching time can be reduced by 0. 068 s. In this paper, the feature description algorithm and matching algorithm based on K-means clustering analysis could meet the high precision and high real-time requirements of image registration.

关 键 词：红外图像匹配 K-means聚类分析 LBP HARRIS-LAPLACE 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]