图像特征匹配中一种快速关键维过滤搜索算法

Key dimension filtering based search algorithm of B+Tree for image feature matching

为了解决宽基线多图匹配中匹配效率低和匹配精度不高的问题,使用经典的SIFT特征作为描述子,提出一种新的高维特征搜索算法.采用基于距离尺度的相似性度量准则,首先将图像高维特征集合分类,然后为每一个类建立B+Tree索引,最后在KNN(K Nearest Neighbor)搜索阶段应用基于关键维过滤的查找策略,实现高维特征的快速匹配.实验结果表明,与经典的BBF和LSH等KNN搜索算法相比较,关键维过滤搜索算法具有更高的搜索效率和搜索精度,有助于提升宽基线多图匹配性能.

In dealing with the issues of low efficiency and low accuracy in multiple wide-based-line image matching, this paper adopts the classical SIFT descriptor, and proposes a novel high dimensional feature search algorithm. This paper follows the distance-based similarity standard, and firstly partitions the image feature set into different classes, then establishes a B＋Tree for each class, and finally gives out a key dimension filtering strategy（KDF） in the KNN search step to speed up the high dimensional feature matching. Experimental results show that the proposed algorithm, which can obtain a higher accuracy with a lower time cost than the classical KNN search algorithm such as BBF, LSH and so on, would be a help to improve the capability of multiple wide-based-line image matching.