The Spectral Matching Algorithm Based on
Hyperbolic Mahalanobis Metric
Wen-xia Bao
1,2
, Guo-fen Yu
1
, Gen-sheng Hu
1
, Dong Liang
1
, Shao-mei Yan
1
1 Key Lab of Intelligent Computing & Signal Processing, Ministry of Education
Anhui University
Hefei 230039, China
2 Key Laboratory of Polarization Imaging Detection Technology Anhui Province
Hefei 230031, China
Abstract—In order to solve the problem of the traditional image
matching algorithm based on the spectral feature, the spectral
matching algorithm based on hyperbolic mahalanobis metric is
proposed in this paper. The algorithm first introduces a
hyperbolic metric that has better adaptability to the sample data.
And the hyperbolic mahalanobis metric is defined according to
the statistical properties of the data. For a point in a given point-
set, the sub point-set is selected according to the hyperbolic
mahalanobis metric and the weighted graph of the sub point-set
is constructed. The eigenvalue vector and the spectral gap vector
are obtained by the singular value decomposition (SVD) of the
adjacency matrix of the weighted graph, which construct the
hyperbolic mahalanobis metric spectral feature. Finally, the
matching matrix is constructed based on the similarity between
the hyperbolic mahalanobis metric spectral feature and
geometric relations between feature points. Thereby establish the
matching mathematical model and introduce the greedy
algorithm to obtain the matching results. A large number of
experimental results show that the proposed algorithm improves
the matching accuracy and the robustness. And the algorithm
extends the application range of the spectral matching algorithm.
Keywords-image matching; hyperbolic mahalanobis metric;
spectral feature
I. INTRODUCTION
Image feature matching is an essential problem in the field
of computer vision and pattern recognition. It has a wide
application in object detection and recognition, image
registration, stereo correspondence, 3D reconstruction and
image mosaic, etc. In recent years, the spectral graph theory [1,
2, 3, 4, 5, 6] is applied to the image matching as an efficient
mathematical description tool. For example, a matching
algorithm based on pairwise constraints can deal with different
size point-set [7]. The algorithm constructed assignments graph
and calculated its adjacency matrix. The principal eigenvector
corresponding to the largest eigenvalue of the matrix is taken as
the indicator vector to obtain the matching results. Based on the
work of [7], Yuan et al. [8] employed a weighted voting
strategy to speed up the computation of the spectral matching.
A spectral matching algorithm based on nonsubsampled
contourlet transform (NSCT) and scale-invariant feature
transform (SIFT) was proposed in [9]. The image is
decomposed into a low frequency and high frequency parts
based on the NSCT. And the SIFT is employed to extract
feature points of the low frequency image. And then constructs
the proximity matrix based on these feature points. Finally,
performs the spectral decomposition on the proximity matrix to
get the matching result. Reference [10] presents a fast and
efficient spectral matching algorithm, the proximity matrix is
approximated by a linear combination of the Kronecker
products of the highly compressed bases and index matrices.
And then calculate the eigenvector of the proximity matrix to
obtain the matching result. The algorithm improves the
computational speed and therefore can be adapted to the large-
scale point-set matching. A monotonic invariant intensity
descriptor (MIID) [11] was proposed based on spectral
embedding and nonsubsampled contourlet transform. The
MIID is constructed by the spectral vector and the spectral gap
vector of the signless Laplacian matrix. And the MIID is robust
to monotonic illumination changes. In [12], the weighted graph
is constructed based on the neighborhood of the feature points
and calculated the normalized Laplacian matrix of the graph.
Then the distribution of the matrix is summarized as a
histogram to obtain the spectral descriptor. And the MIID
combined the approximate distance order to obtain the
correspondence between the points.
These spectral feature matching algorithms adopt Euclidean
distance when constructing spectral features and calculating
similarity. The Euclidean distance considers the input sample
space as isotropic. However, this isotropic assumption can not
hold in many practical applications. Because this isotropic can
not fairly reflect the underlying relationships between the
sample data. Therefore, the mahalanobis distance which
considered the correlation among different data dimensions is
used to replace the Euclidean distance in the image matching
[13, 14, 15]. The mahalanobis distance is applied to the speed
up robust features (SURF) algorithm to eliminate the initial
mismatches [13]. And the least-square method and the
RANSAC algorithm are combined to obtain the Homography
matrix and the location information to complete the matching.
The algorithm can improve the matching accuracy. In order to
overcome the problems in the traditional graph transformation
matching (GTM), [14] proposed a weighted graph
transformation matching algorithm (WGTM) based on
mahalanobis distance. The algorithm first constructed K
neighbor graph for the points, and calculated the weight matrix
based on the median and angle distance of the points and its
adjacent K points. Finally, the iterations are used to eliminate
2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics( CISP-BMEI 2017)
978-1-5386-1936-0/17/$31 ©2017 IEEE