局部密度差异引导的图像特征匹配算法

Image feature matching guided by local density difference

摘要:

为解决预定义参数化模型的特征匹配方法通用性较低且鲁棒性较差的问题,根据正确匹配和误匹配的空间分布具有显著差异的现象,提出一种局部密度差异引导的特征匹配(RFM-LoDD)算法。首先,将假定特征匹配转换为能够表征特征匹配性质的空间样本点,并引入概率距离计算样本点的局部密度。其次,在随机选取的40幅涉及不同变换模型的图像对上测试算法的最优参数设置,确定了具有全局最优的密度阈值和其他参数。最后,将样本点的局部密度与密度阈值进行比较,当样本点的局部密度大于密度阈值,则认为该样本点代表的假定匹配为正确匹配,否则,就认为其代表的假定匹配为误匹配。在代表图像对和公开数据集上进行的实验表明,RFM-LoDD算法在各种匹配场景下都能够保持良好的鲁棒性,特别是在内点率较低的Retina数据集和AdelaideRMF数据集上相比于先进的算法均取得了领先的F分数。此外,RFM-LoDD算法具有准线性的时间复杂度,在4个公开数据集上的平均运行时间约为40 ms,时间成本相比于经典的随机抽样一致性（RANSAC）算法降低了两个数量级。

Abstract:

To address the low generalization and poor robustness in feature matching methods that rely on pre-defined parameterized models, based on the observation that the spatial distribution of correct match and mismatch has significant differences, a feature matching algorithm guided by local density difference (RFM-LoDD) is proposed. Firstly, the putative feature matches are converted into spatial sample points that can characterize the nature of the feature matches, and the probabilistic distance is introduced to calculate the local density of the sample points. Secondly, the optimal parameter settings of the algorithm are tested on 40 randomly selected image pairs involving different transformation models,determining the globally optimal density threshold and other parameters. Finally, the local density of the sample points is compared with the density threshold. When the local density of a sample point is greater than the density threshold, the putative feature match represented by the sample point is considered to be a correct match, otherwise, it is a mismatch. Experiments conducted on representative image pairs and public datasets demonstrate that the RFM-LoDD algorithm maintains good robustness in various matching scenarios. Notably, it achieves leading F-scores on the Retina dataset and AdelaideRMF dataset with low inlier rates compared to advanced algorithms. Additionally, the RFM-LoDD algorithm has quasi-linear time complexity, with an average run time of about 40 ms on the four public datasets, significantly reducing time cost by two orders of magnitude compared to the classical random sample consensus (RANSAC) algorithm.