计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (3): 200-205.DOI: 10.3778/j.issn.1002-8331.1610-0004

• 图形图像处理 • 上一篇    下一篇

改进的SIFT算法图像匹配研究

冯文斌1,刘宝华2   

  1. 1.燕山大学 河北省并联机器人与机电系统实验室,河北 秦皇岛 066004
    2.燕山大学 机械工程学院,河北 秦皇岛 066004
  • 出版日期:2018-02-01 发布日期:2018-02-07

Research on image matching based on improved SIFT algorithm

FENG Wenbin1, LIU Baohua2   

  1. 1.Hebei Provincial Key Lab of Parallel Robot and Mechatronic System, Yanshan University, Qinhuangdao, Hebei 066004, China
    2.School of Mechanical Engineering, Yanshan University, Qinhuangdao, Hebei 066004, China
  • Online:2018-02-01 Published:2018-02-07

摘要: 对SIFT(尺度不变特征变换)算法特征描述子维数过高,导致匹配速度过慢、匹配率低等问题,提出了一种分级放射状分区的方法来构建特征描述子,将特征点邻域划分为8个区域,统计各个区域内8个方向的梯度方向直方图,得到64维特征描述子,使特征描述子维数降低50%。同时因马氏距离考虑了特征描述子向量间的相关性,在匹配时用马氏距离双向匹配方法代替欧氏距离进行匹配,并用RANSAC(随机抽样一致性)方法消除误配点。实验结果表明,改进的SIFT算法保留了SIFT算法对模糊、压缩、旋转和缩放等不变性优势,并提高了匹配速度,正确匹配率平均增加10%~15%。

关键词: 尺度不变特征变换, 特征描述子, 马氏距离, 欧氏距离, 随机抽样一致性

Abstract: For the feature descriptors’ dimensions based on the SIFT algorithm are too high, resulting to low speed, low matching rate and other issues, a kind of hierarchical radial partition method is proposed to construct feature descriptor. The feature point neighborhood is divided into 8 regions, counting 8 directions’ gradient direction histogram in each region to get a 64 dimensions descriptor, which the dimensions of feature descriptors are reduced by 50%. At the same time, because the Mahalanobis distance considering the correlation between the feature descriptor vectors, using the two-direction matching method of Mahalanobis distance instead of Euclidean distance when matching, the RANSAC method is used to eliminate the mismatch points. The theoretical analysis and simulation results show that improved SIFT algorithm retains SIFT algorithm for fuzzy, compression, rotation and scaling invariance advantages, improves the matching speed, and the average rate of true-match increases of 10%~15%.

Key words: Scale-Invariant Feature Transform(SIFT), feature descriptor, Mahalanobis distance, Euclidean distance, Random Sample Consensus(RANSAC)