Abstract: A new image matching algorithm based on manifold regularization is proposed to efficiently match image pairs with large amount of data. The feature points of the two images are characterized the same underlying manifold, with the similarity constrain of the feature points coming from the different images calculated by the Manhattan distance between SIFT descriptors and spatial constrain calculated by distance between feature points from the same image which is the regularization of the objective function. The unified embedding of feature points can be obtained directly by solving the eigen-value problem. Test results indicate that the proposed method has a higher performance than SVD-SIFT and LE-SIFT methods, and this method has linear complexity, which is suitable for dealing with large number of feature points.

Key words: spatial constraints, manifold, regularization, Manhattan distance

摘要： 针对大数据量的影像匹配问题，提出了一种基于正则化流形学习的影像匹配方法。该方法利用曼哈顿距离方法计算特征点SIFT描述符的相似性，并在此基础上增加同一幅影像中特征点之间的空间结构关系作为正则化项，采用流形学习方法将两幅影像中的特征点共同映射到同一流形空间；根据最小距离方法进行特征点的匹配。通过选取四种不同来源的影像对进行实验，与SVD-SIFT方法、LE-SIFT方法进行综合分析，结果表明该方法在匹配性能上优于现有方法，而且该方法具有线性复杂度，适用于处理特征点数量较大的影像匹配。

关键词: 空间约束, 流形学习, 正则化, 曼哈顿距离