不同距离测度的SIFT特征描述符相似性度量比较

发布时间：2018-03-06 12:21

本文选题：影像匹配　切入点：SIFT算子　出处：《遥感信息》2017年01期 　论文类型：期刊论文

【摘要】：鉴于尺度不变特征转换(Scale-Invariant Feature Transform,SIFT)匹配算法特征点描述符的相似性度量大多采用欧氏距离,而目前诸多学者提出的距离测度中除欧氏距离还有其他应用较为广泛的如城市距离、棋盘距离、马氏距离、余弦距离、相关性距离等,该文采用这6种距离用于特征点描述符的相似性度量,在实验中分别统计这6种距离测度在最近邻/次近邻比值取值从0.3至0.9的匹配结果。实验结果表明,6种距离测度中综合性能最好的是城市距离,其次是棋盘距离,而欧氏距离与余弦距离由于计算耗时以及误匹配点相对较多其综合性能并不是很好,性能最差的是马氏距离。因此,在实际的SIFT匹配算法中,特征点描述符相似性度量应采用基于城市距离或棋盘距离的测度准则,而不是欧氏距离或余弦距离。
[Abstract]:Since the similarity measures of scale-Invariant Feature transform (sift) matching algorithm are mostly Euclidean distance, many scholars have put forward the Euclidean distance as the most widely used distance measure, such as urban distance. The chessboard distance, Markov distance, cosine distance, correlation distance and so on are used to measure the similarity of feature point descriptors. In the experiment, the matching results of the nearest neighbor / next nearest neighbor ratio between 0.3 and 0.9 are calculated respectively. The experimental results show that the best comprehensive performance of the six distance measures is the city distance, followed by the chessboard distance. However, the Euclidean distance and cosine distance are not very good due to the computation time and the relative number of mismatch points, and the worst performance is the Markov distance. Therefore, in the actual SIFT matching algorithm, The similarity measure of feature point descriptor should be based on city distance or chessboard distance instead of Euclidean distance or cosine distance.
【作者单位】：辽宁工程技术大学测绘与地理科学学院;国家测绘地理信息局卫星测绘应用中心;
【基金】：辽宁工程技术大学测绘与地理科学学院研究生优秀学位论文培育计划(YS201503) 辽宁省教育厅重点实验室项目(LJZS001)
【分类号】：P23;TP391.41

【相似文献】