基于马氏度量的图像谱特征描述

Image Spectral Feature Description Based on Mahalanobis Metric

传统的谱特征描述过程中采用的是不能反映样本间潜在关系的欧式距离进行度量的.为更好地区分数据之间的联系,提出基于马氏度量的图像谱特征描述算法.首先,对特征点及其周围特征点按照马氏距离进行分层,并在每层上面构造相应的结构图及计算其关联矩阵;接着,对关联矩阵进行谱分解得到其特征值向量和谱隙向量;然后分别用两者的最大值、平均值和方差统计量得到最终的马氏度量谱特征;最后,根据马氏度量谱特征之间的相似性和特征点之间距离关系来构建匹配数学模型,并用贪心算法求解得到特征点之间的匹配关系.实验结果表明,该算法提高匹配精度;同时将其应用于偏振图像的匹配问题上,并取得较好的匹配结果.

The traditional spectral features in the process of description use the European distance metric thatcan not reflect the potential relationship between the samples.In order to better distinguish the relation be-tween data,the image spectral feature description algorithm based on mahalanobis metric is proposed in thispaper.Firstly,the feature points and their surrounding points are layered according to the Mahalanobis dis-tance,and the corresponding structure graph is constructed on each layer and the correlation matrix is calcu-lated.Secondly,the eigenvalue vector and the spectral gap vector are obtained by spectral decomposition ofthe correlation matrix.And then,the maximum,mean and variance of the two vectors are calculated respective-ly to obtain the final mahalanobis metric spectral features.Finally,a matching mathematical model is con-structed based on the similarity between the mahalanobis metric spectral and the distance between featurepoints,and the matching relation between feature points is obtained by greedy algorithm.A large number ofexperimental results show that the proposed algorithm improves the matching accuracy.At the same time,it is applied to the matching of polarimetric images and a good matching result is obtained.