基于权重阈值的图像匹配方法  

Image Registration Method Based on Weight Threshold

作　　者：丁苏楠 张秋菊 DING Sunan;ZHANG Qiuju(School of Mechanical Engineering,Jiangnan University,Wuxi,Jiangsu 214122,China)

机构地区：[1]江南大学机械工程学院,江苏无锡214122

出　　处：《轻工机械》2020年第5期67-73,共7页Light Industry Machinery

基　　金：国家自然科学基金:新型仿生柔性腕手构型与抓取操控性能研究(51575236)。

摘　　要：针对尺度不变特征变换(SIFT)算法在低照度情况下存在匹配特征点数目少,会丢失部分重要信息的问题,课题组提出先对图像进行预处理,即增强对比度,提高亮度,使特征信息更明显,特征点更充足,再使用SIFT算法提取特征点;针对SIFT算法因特征描述符维数过多导致耗时长,实时性差的问题,课题组提出一种新的特征点描述方式,降低了描述符维数,提高了算法运行速度;最后,针对SIFT算法匹配过程中对描述符所有维度设置统一阈值易造成误匹配的问题,课题组提出一种权重阈值的方法,对距离特征点不同位置的种子点设置不同的阈值,提高匹配正确率。实验表明:与SIFT算法以及PCA-SIFT算法相比,改进的算法匹配精度提高了10%~20%;同时,匹配时间也有所提高。该算法既有效提高了匹配正确率,又缩短了算法运行时间。Focusing on the issue of the fewer number of feature points and partial important information loss under low illumination of the scale-invariant feature transform(SIFT)algorithm,a method of image pre-process was proposed,which is to enhance the contrast and improve the brightness for the more obvious feature information and the more abundant feature points before extracting the feature points by SIFT algorithm.Secondly,aiming at the time-consuming and poor real-time problems of the SIFT algorithm caused by the excessive feature descriptor dimension,a new feature point description method was proposed,which reduces the descriptor dimension and improves the running speed of the algorithm.Finally,for the problem of mismatching easily caused by setting uniform threshold for all dimensions of the descriptors in the SIFT algorithm registration process,a weight threshold method was proposed to improve the registration rate by setting different thresholds for the seed points at different positions of the feature points.Experimental results show that,compared with the SIFT algorithm and the PCA-SIFT algorithm,the algorithm registration accuracy is improved by 10%~20%,and the matching time is also improved.The algorithm can effectively improve the registration accuracy and shorten the running time of the algorithm.

关 键 词：图像处理 尺度不变特征变换算法 分层阈值 MATLAB 图像匹配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]