基于K近邻一致性的特征匹配内点选择算法

宽基线图像特征匹配是计算机视觉应用中一项极具挑战性的工作。由于图像之间存在较大的差异,宽基线图像初始特征匹配的结果中不可避免地包含大量的外点。提出了K近邻一致性算法来实现从宽基线图像初始匹配结果中快速选出高可靠性的点对。该算法采用仿射不变的结构相似度来衡量两组K近邻特征点的结构相似性。K近邻一致性算法采取由粗到精的策略,通过K近邻对应一致性检测和K近邻结构一致性检测两个步骤来选择内点。实验结果表明,提出的算法在查准率、查全率和运行速度等方面接近或优于当前几种最新的内点选择算法,可适用于存在大范围的视点、尺度和旋转变化的宽基线图像。

Inlier Selection Algorithm for Feature Matching Based on K Nearest Neighbor Consistency

Feature matching for wide baseline images is an extremely challenging task in computer vision applications.A large number of outliers are inevitably included in the initial matching results due to significant changes between views of wide baseline images.An inlier selection algorithm called K nearest neighbor consistency (KNNC) was proposed to efficiently select matches with high reliability from initial feature matching results of wide baseline images.An affine-invariant structure similarity is utilized to measure the degree of structure similarity between two groups of K nearest neighboring features.Adopting the coarse-to-fine strategy,KNNC algorithm selects inliers by the processes of K nearest neighbor correspondence consistency checking and K nearest neighbor structure consistency checking.Experimental results show that the proposed algorithm approximates or surpasses several state-of-the-art inlier selection algorithms in performance on precision,recall and computational time,and is applicable to wide baseline images with large differences in viewpoint,scale and rotation.