基于CA-SIFT的图像特征两级匹配算法  

Two-level matching algorithm for image features based on CA-SIFT

作　　者：焦卫东[1] 焦一哲 JIAO Weidong;JIAO Yizhe(Tianjin Key Laboratory for Advanced Signal Processing,CAUC,Tianjin 300300,China)

机构地区：[1]中国民航大学天津市智能信号与图像处理重点实验室,天津300300

出　　处：《中国民航大学学报》2025年第2期73-82,共10页Journal of Civil Aviation University of China

基　　金：国家重点基础研究发展计划项目(2020YFB1600101)。

摘　　要：针对欧氏空间尺度不变特征变换(SIFT,scale invariant feature transform)算法在彩色图像匹配过程中丢失图像光谱信息、匹配精度低、计算量大等问题,利用Clifford代数(CA,Clifford algebra)对多维空间的表达能力提出一种基于CA-SIFT的图像匹配算法。首先,将图像转换到CA空间表示,同时保留图像空间和光谱信息,通过共形几何代数内积运算构造度量函数,提高特征点搜索效率,在CA空间中检测特征点;其次,采用图像特征两级匹配策略,即将CA-SIFT特征描述向量转换为哈希编码,由暴力匹配得到粗匹配结果;最后,采用网格运动统计(GMS,grid-based motion statistics)方法完成精匹配。实验结果表明:本文算法性能优于SIFT算法,提取的特征点对数量最多提升近54%;图像匹配方面,平均匹配精度达到98%以上,实现了高精度、适用于多数场景的图像匹配方法。To address the problems of loss of image spectral information,low matching accuracy and large computational effort in the process of color image matching by the scale invariant feature transform(SIFT)algorithm in Euclidean space,an image matching algorithm based on CA-SIFT is proposed using the expressiveness of Clifford algebra(CA)for multidimensional space.Firstly,the image is transformed to CA space representation,while retaining the image space and spectral information,and the metric function is constructed by the inner product operation of the conformal geometric algebra to improve the efficiency of feature point search and detect feature points in CA space.Secondly,a two-stage image feature matching strategy is adopted,the CA-SIFT feature description vector is converted into a hash code,and the coarse matching results are obtained by brute force matching.Finally,a grid-based motion statistics(GMS)method is used to complete the fine matching.The experimental results show that the proposed algorithm outperforms the SIFT algorithm,and the number of extracted feature point pairs is improved nearly 54%.In terms of image matching,the average matching accuracy reaches over 98%,achieving a highly accurate and applicable image matching method for most scenes.

关 键 词：特征匹配 CLIFFORD代数 特征检测 内积 网格运动统计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]