一种基于SIFT特征的快速图像匹配算法  

A QUICK IMAGE MATCHING ALGORITHM BASED ON SIFT FEATURE

作　　者：杨松[1,2] 邵龙潭[1] 宋维波[2] 刘威[2] 

机构地区：[1]大连理工大学工业装备结构分析国家重点实验室,辽宁大连116024 [2]大连海洋大学信息工程学院,辽宁大连116023

出　　处：《计算机应用与软件》2016年第7期186-189,256,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(50905022;51309047);国家高技术研究发展计划项目(2010CB7315022);工业装备结构分析国家重点实验室专项基金项目(S09104)

摘　　要：经典的SIFT算法具有良好的尺度、旋转、光强不变特性而广泛应用于图像匹配。图像特征点较少时,匹配过程使用穷举法查找最近邻匹配点;当图像特征点较多时采用KD-Tree结构,而其检索过程存在"回溯"现象,这两种方法的匹配效率都不高。为了提高特征点的匹配速度,提出改进的SP-Tree结构解决"回溯"问题。在结点集分割时设置参数合理确定左右超平面位置,引入平衡因子作为结点分割方法选择的依据,采用近似最近邻搜索算法加快特征点匹配速度。给出算法的详细实现过程,并应用两幅图像进行验证。实验结果表明:SIFT特征向量采用改进SP-Tree结构在损失少部分匹配点的同时,提高了SIFT特征点的整体匹配速度,适合于图像特征的实时匹配过程。Due to its good invariant characteristics in scaling,rotation and light intensity,the classic SIFT algorithm has been widely used in image matching. If there are fewer image feature points,the exhaustion method is used to find the nearest matching point. If there are more image feature points,KD-tree will then be used,but the backtracking phenomenon exists in its retrieval process,so the matching efficiency of both methods are low. In order to improve feature points matching speed,we propose an improved SP-Tree structure to solve the backtracking problem. The parameter α is set to determine a reasonable location about hyper-plane in node set segmentation,and a balancing factors ρ is introduced as the choice basis for different node segmentation method,and the approximate nearest searching algorithm is adopted,which can accelerate the speed of feature points matching. In the paper we give the detailed implementation process of the algorithm and the validations with two standard images. Experimental results show that the SIFT feature vector,by using a modified SP-Tree structure,at the expense of few matching points,greatly improves the overall speed of SIFT feature points matching. It is suitable for image features matching in real time.

关 键 词：图像匹配 SIFT特征 KD-TREE SP-Tree 最近邻搜索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]