基于SIFT的图像匹配实时性改进  

Improvement of Real-Time Performance of Image Matching Based on SIFT

作　　者：董锦涛 陈水忠[2] 徐恺[2] 揭斐然[2] DONG Jintao;CHEN Shuizhong;XU Kai;JIE Feiran(Science and Technology on Electro-Optical Control Laboratory,Luoyang 471000 China;Luoyang Institute of Electro-Optical Equipment AVIC,Luoyang 471000 China)

机构地区：[1]光电控制技术重点实验室,河南洛阳471000 [2]中国航空工业集团公司洛阳电光设备研究所,河南洛阳471000

出　　处：《电光与控制》2020年第3期80-83,88,共5页Electronics Optics & Control

摘　　要：在图像匹配中,图像分辨率越高,可提取的特征越多,匹配精度也越高,但相应的匹配耗时也会越长。因此,对基于SIFT的图像匹配算法进行改进。首先,使用第二组高斯金字塔上的特征点进行粗匹配,由基于GMS改进的RANSAC剔除误匹配点并初步得到仿射矩阵,大致确定匹配区域;然后,筛选剩余特征点中位于该区域的部分进行细匹配,并用初步得到的仿射矩阵筛选匹配对;最后,基于最小二乘算法再次去误匹配并快速求解最终的仿射变换矩阵。经实验验证,所提算法在保证亚像素精度的条件下,很大程度上减少了算法的匹配耗时,提高了实时性。In image matching higher resolution of an image means higher matching accuracy and there are more features that can be extracted but the matching time is longer.Thus the image matching algorithm based on SIFT is improved in this paper.Rough matching is performed on the feature points of the second group of Gaussian pyramids.The improved RANSAC based on GMS is used to eliminate the mismatched points and obtain the affine matrix and the matching area is roughly determined.During the process of accurate matching the remaining feature points in the same matching area are filtered and the matching pairs are filtered by using the previous affine matrix and the least squares algorithm is used to eliminate the mismatched points again and quickly solve the final affine matrix.The experimental results show that this method greatly shortens the matching time and improves real-time performance under the condition of guaranteeing the sub-pixel accuracy.

关 键 词：图像匹配 SIFT 实时性 GMS 最小二乘法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]