多相似内容图像的特征匹配

Feature Matching between Images with Multiple Similar Contents

针对当场景中包含多处相似内容时利用灰度或梯度信息进行特征匹配误差较大的问题,提出采用拓扑相似性度量以去除灰度或梯度分布相似内容导致的误匹配的方法.首先提出了K近邻算法,根据相似内容间的匹配相比非相似内容间的误匹配有着明显更小的最短最近邻距离的性质,得到相似内容间多对多的特征匹配作为待匹配关系;然后提出了平面投影的5条拓扑相似性约束条件,利用多单应矩阵将待匹配对划分至多处平面,对各平面上特征点进行分级三角剖分,根据拓扑相似性约束条件去除误匹配,并将多对多的匹配降为一对一匹配.实验结果表明,文中方法可以去除区域边缘以外的误匹配.

Intensity-or gradient-based similarity measurement leads to outliers caused by similar objects. To address this problem, a novel feature matching method using topology similarity constraints is proposed. Using Euclidean distance to measure similarity, similar contents are significantly closer than dissimilar contents. K nearest neighbor algorithm is proposed to find matched candidates which have significantly smaller distances. With plane-to-plane homography, matched features are divided into several feature-sets by planes. Then all feature-sets are hierarchically triangulated. With five topology similarity constraints, outliers are removed, and m ： n-matches are reduced to 1 ： 1-matches. The experimental results show that the proposed method successfully removed the outliers except the stray points.