A New Feature Matching Algorithm Based on the
RMAI for Aerial Image Registration
Zhaoxia Liu, Yaxuan Wang and Yu Jing
School of Software
Dalian University of Foreign Languages
Dalian, China
liu.zhaoxia0524@gmail.com
Abstract—To solve the problem of mismatches in image
registration, in this paper, a new feature matching algorithm is
proposed. Firstly, considering the intensity distribution and
affine invariant properties of triangle, the relative mass affine
invariants (RMAI) are defined to describe triangle regions. And
then a global feature matching strategy is designed based on
RMAI and graph structure. In the matching process, Genetic
Algorithm is applied to search two matched graphs.
Experimental results by registering aerial images show that the
proposed algorithm has higher accuracy and stability than
traditional Iterative Closest Point (ICP) and Coherent Point Drift
(CPD), the proposed algorithm can deal with images with similar
features, monotonous backgrounds and low overlapping areas.
Keywords—Global feature matching; affine invariants; image
registration
I. INTRODUCTION
Image registration has been widely applied in remote
sensing, medical image analysis, cartography, computer vision,
pattern recognition and so on [1]. Feature matching is to find
corresponding feature points from reference image and sensed
image. There are local feature matching and global feature
matching strategies.
Local matching only uses the feature similarity to establish
the correspondence, so the descriptor is important. There are
Scale-Invariant Feature Transform (SIFT) descriptor, moment-
based combined invariants, cross-correlation coefficient,
entropy and chain codes, Fourier descriptor, shape matrices[2]-
[4] and so on. Global matching uses spatial structure and
neighborhood relation to find corresponding feature points. The
popular approaches are ICP, CPD and TPS-RPM [5]-[7]. These
algorithms work well in some application, but they may fail in
registering images having the characteristics of similar feature,
monotonous background and low overlapping areas, such as
the images captured on the sea. The characteristic of which is
described in [8]. On the other hand, registration algorithms just
relying on either local or global matching do not match feature
points accurately. It is necessary to combine both of them. So,
in this paper, integrating local and global matching strategies, a
new feature matching algorithm is proposed for registering
aerial images with similar features, monotonous backgrounds
and low overlapping areas.
Firstly, relative mass affine invariants (RMAI) are
presented to reduce the interdependence between the points.
These invariants are described by considering the intensity
distribution of image and affine invariant properties of triangle.
The invariants are used for evaluating every triangle pair. Then,
global feature matching strategy is designed based on graph
structure constructed by triangulating point sets. Feature
matching is to find two matched graphs that include as many
corresponding points as possible. In the matching process,
Genetic Algorithm (GA) is used for finding the matched graphs.
Experimental results show that the proposed algorithm
performs well even under too many noises and is suitable for
images with similar feature, monotonous background and low
overlapping areas.
II. R
ELATIVE MASS AFFINE INVARIANTS
Given
m
V and
d
V are the feature point sets detected from
reference image and sensed image respectively. These points
are local extreme points. Therefore, the intensity distribution of
the regions around the points is uneven and contains plenty of
information. And regions in the triangles constituted by those
points are uneven too. Here, the difference between these
regions is compared to determine the correspondence of the
triangles. Affine transformation has some properties such as the
collinearity relation between points and preserving the ratio of
distances along a line. Collinearity relation means that three
points in a line will continue to be collinear after affine
transformation. Preserving ratios of distances means that the
ratio
21 32
/pppp
of collinear points
1
p ,
2
p ,
3
p is
preserved after affine transformation. Based on the two
properties above, another property of an area relative
invariance can be deduced, which is applied in several times in
[9]. And for two images the distribution of intensity of
corresponding triangle is identical, which derive that affine
maps have the properties of mass relative invariance. In this
paper, the RMAI is defined as follows:
For two-dimensional image, the formula of mass is given
by:
mS
(1)
where
is the intensity distribution of local image region, S
is the area of the region. Let
(,,)
iii
MMM
mS
and
2013 3rd International Conference on Computer Science and Network Technology
978-1-4799-0559-1/13/$31.00 ©2013 IEEE