面向近重复图像匹配的SIFT特征裁减算法

A SIFT Pruning Algorithm for Efficient Near-Duplicate Image Matching

针对图像提取出的SIFT特征数目通常很大、无法精确控制,导致系统效率不高且不稳定的问题,提出一种SIFT特征裁减算法.通过对SIFT关键点的对比度和主曲率比加权来衡量其匹配能力;在定位关键点和计算关键点方向2个步骤增加对关键点的二次筛选,提取出设定阈值数目内对匹配最有效的关键点.实验结果表明,该算法能有效地控制SIFT特征数量,比已有裁减算法具有更高的匹配准确度;与原始未裁减算法相比,在保证匹配准确度的同时,大大提高了系统的效率和稳定性.

The number of SIFT features extracted from an image is usually large and cannot be adequately controlled,which usually results in poor system performance of low efficiency and instability.A SIFT pruning algorithm is proposed to address the above issues in this work.The algorithm measured discriminative power of keypoints by combining the weighted contrast and ratio of the principal curvature,then extracted the proper number of most significant keypoints within a given range through a two-phase filter process in the steps of keypoint localization and orientation assignment.The experiments show that the proposed algorithm can effectively control the number of features and provide higher accuracy than the previous pruning algorithm.The experiments also indicate that the proposed pruning algorithm achieves much higher efficiency and stability with a comparable matching accuracy in comparison to the original non-pruning SIFT algorithm.