图像特征点选取方法

A Method for Image Feature Points Selection

特征点是图像中常用的局部特征结构,现有的检测算子旨在提取图像中所有符合某一准则的特征点,往往会在同一位置得到大量的特征点,形成对图像局部区域的过描述,增加了高层处理任务的运算量。为了选取出一定数量较好的特征点来描述图像,常用方法是根据特征点响应函数来选取响应值大于某一阈值的特征点,但是这种选取方法并没有考虑特征点的分布均匀性和独特性,而特征点的分布均匀程度和独特性对于图像匹配的结果同样具有重要的影响。为此提出了一种新的特征点选取方法,首先基于检测算子得到的特征点构造最小生成树,然后定义特征点分布均匀性和独特性等选取准则,再通过删剪最小生成树,选取符合(准)最优准则的一组特征点。实验表明,该方法选取的特征点在图像描述和图像匹配等任务中优于传统的特征点选取方法。

Image points are important local features for image description and matching. However, popular point detectors usually extract thousands of points, which causes over-description and increases computational complexity for subsequent processing. To select a set of useful points to describe images, most methods simply choose points according to its response function. But these methods mostly overlook the distribution uniformity and distinctiveness of the points. In this paper, a no-vel method for selecting a set of points based on pruning of the MST（ Minimum Spanning Tree） is proposed. Firstly, a MST is constructed using the points extracted by some certain point detector. Then indicators such as ditribution uniformity and dis-tinctiveness based on MST are defined and objective function for selecting points is constructed. Finally, the MST is pruned to select a set of points which are optimal or suboptimal on stability, uniformity and distinctiveness in describing the structure of the image. Experimental results demonstrate that the proposed method outperforms traditional point selection methods.