基于光谱图像空间的F-SIFT特征提取与匹配  

Feature extraction and matching of F-SIFT based on spectral image space

作　　者：丁国绅 乔延利[1] 易维宁[1] 李俊 杜丽丽 DING Guo-shen;QIAO Yan-li;YI Wei-ning;LI Jun;DU Li-li(Key Laboratory of Optical Calibration and Characterization,Anhui Institute of Optics and Fine Mechanics,Chinese Academy of Sciences,Hefei 230031,China;University of Science and Technology of China,Hefei 230026,China;Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]中国科学院安徽光学精密机械研究所,通用光学定标与表征技术重点实验室,安徽合肥230031 [2]中国科学技术大学,安徽合肥230026 [3]西安科技大学安全科学与工程学院,陕西西安710054

出　　处：《光学精密工程》2021年第5期1180-1189,共10页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.41601379)。

摘　　要：针对传统SIFT算法提取到的图像特征点数量稀少的问题,可利用高光谱图像的特性,在光谱维度上取差重构了图像的尺度空间,使得提取到的特征点数量得到了极大地提高。但特征点数量的大幅增加导致算法的时间开销随之增大,并且有效特征点的占比较低。为了解决特征点的冗余问题,提高匹配效率,提出了一种基于光谱图像空间的F-SIFT算法。利用FAST算法能够在像素层面上快速判断的特性,构建了以当前像素为中心的八邻域准则,在提取图像的特征点之前对差分金字塔的对应位置的像素点做预筛选,使得特征点数量降低到了原来的10%以下;另外传统的匹配方法只统计了目标象元邻域内的像素信息,而忽略了象元的几何位置信息,因此本文扩展了特征描述符向量,首先利用最近邻与次近邻之比对特征描述符做一次粗匹配,记录可靠性程度并将其纳入描述符向量,接着按照相似性程度的高低从前20组匹配中迭代选取4组匹配点用于构造三角平面,利用象元的位置信息进行精确匹配。实验结果表明本文方法能够有效降低冗余特征点的数量,剔除误匹配。To solve the problem of the scarcity of the number of image feature points extracted by the traditional scale-invariant feature transform(SIFT)algorithms,in the previous work,we used the characteristics of a hyperspectral image and reconstructed the scale space of the image by considering the difference in the spectral dimension,so that the number of feature points extracted is greatly increased.However,the large increase in the number of feature points leads to the increase in the time cost of the algorithm,and the proportion of effective feature points is low.To solve the redundancy problem of feature points and im‐prove the matching efficiency,we propose a novel flip-SIFT(F-SIFT)algorithm based on the spectral image space.Based on the ability of the proposed algorithm to perform matching quickly at the pixel level,an eight-neighborhood criterion centered on the current pixel is constructed.Before the feature points of the image are extracted,the pixels in the corresponding position of the difference pyramid are pre-filtered,so that the number of feature points is reduced to less than one tenth of the original.In addition,the traditional matching method only counts the pixel information in the neighborhood of the target pixel,but ignores the geometric location information of the pixel.Therefore,in this study,we extended the feature descriptor vector.First,the feature descriptor is roughly matched by the comparison between the nearest neighbor and the next nearest neighbor,and the similarity degree is recorded and included in the descriptor vector.Then,according to the reliability degree,four groups of matching points are selected iteratively from the first 20 sets of matching points to construct the triangle plane and the pixel position information is used for accurate matching.Experimental results showed that the proposed method can effectively reduce the number of redundant feature points and eliminate mismatches.

关 键 词：尺度不变特征变换 光谱图像空间 八邻域准则 双重位置迭代匹配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]