A Multi-sensor Remote Sensing Image Matching Method Based on SIFT Operator
and CRA Similarity Measure
WU Yingdan
College of Science
Hubei University of Technology
Wuhan, China
e-mail: yd_wu2010@163.com
MING Yang
Geospatial Information and Digital Transportation Research Institute
CCCC Second Highway Consultants Co., Ltd
Wuhan, China
e-mail: ankelbreaker@163.com
Abstract—Aim to the difficulty of acquisition of conjugate
points from multi-source remote sensing imagery, a novel
matching method based on SIFT and CRA similarity measure
is proposed. Firstly, the SIFT operator is adopted to extract
feature points and coarse match is performed, the approximate
transformation relationship and the rotation angle between the
matched images are estimated by the above matching results.
Secondly, if there exists large rotation angle between the
matched images, rotation compensation on the image of
matching window should be carried out, and the CRA
similarity measure is used to search the corresponding points.
Finally, the mismatched points in each level of pyramid images
are eliminated by Quadratic polynomial with RANSAC
algorithm. Repeating above procedures until the original
image level, it realizes the automatic and reliable matching of
multi-source remote sensing imagery. The experiment of
matching of multiple SAR imagery and optical remote sensing
imagery is made, and the method is validated.
Keywords:
multi-source remote senseing imagery; image
matching; SIFT; CRA; RANSAC
I. INTRODUCTION
With the development of remote sensing technology,
platforms for remote sensing have been diversely increasing.
Using multi-source remote sensing image to obtain three-
dimensional space information has wide application
prospects. However, the acquisition of conjugate points’
coordinates mainly obtained by manual, and this is not only
time consuming, but also can not meet the requirements of
fast processing massive data. Therefore, research of
automatic matching for multi-source remote sensing imagery
has important significance.
Automatic matching for multi-source remote sensing
images has been greatly studied. Thepaut uses the edge
features for the registration of two images
[1]
, the drawback is
that because of different imaging mechanism of SAR and
optical sensor and the speckle noise in SAR image, the edge
matching is very difficult. Paul proposed a matching method
based on the regional characteristics of the optical image and
SAR image
[2]
, such as the area, perimeter and radius. But the
algorithm’s application scope is severely limited and feature
positioning accuracy is not high. Yong et al proposed the
algorithm for multi-source remote sensing image matching
based on information entropy which comprehensively
considers of the intensity information and the edge direction
information
[3]
. Li et al proposed a image matching algorithm
based on the improved maximum mutual information for
medical image, which has a good reference, but the
feasibility for remote sensing image matching needs further
study
[4]
. In recent years, Lowe proposed a new Scale-
invariant feature transform operator (SIFT for short)
[5]
,
which is not sensitive to changes of the image rotation,
scaling, and brightness, and also maintains the advantages of
relatively stableness on change of perspective, affine
transformation and noise. The drawback is low efficiency
and not high matching success rate. Inglada has taken the
comparison tests with several similarity measure, such as the
correlation coefficient, the minimum distance measure,
mutual information measure (MI) and cluster reward
algorithm (CRA), and the latter two measure obtained better
results
[6]
. Accordingly, this paper introduces a new multi-
sensor remote sensing image matching algorithm based on
SIFT operator and CRA similarity measure to solve
corresponding points automatic matching.
II. I
MAGE MATCHING STRATEGY
To match the multi-source remote sensing images, large
rotation angle, significant gray differences and high false
match rate are often encounter. In this paper, the pyramid
image matching strategy is adopted, and it ensures the
matching speed and reliability. The workflow of proposed
algorithm is shown in Fig. 1.
A. Pyramid Image Generation
Pyramid image can be generated by pixel average
method
[7]
, wavelet transform method
[8]
, Laplace method
[9]
and some other methods. In this paper, the pixel 3 × 3
averaging method is used to generate two levels of pyramid
image. Firstly, calculate the average gray value of the
element of 3 × 3 pixels in the original image, which is
assigned to the corresponding pixel in the first level of
pyramid image, and the same method is used to generate the
second level pyramid image from the first level pyramid
image.
B. SIFT Feature Matching
SIFT feature matching algorithm proposed by Lowe
includes four parts, such as scale space extreme exploration,
2011 International Conference on Intelligence Science and Information Engineering
978-0-7695-4480-9/11 $26.00 © 2011 IEEE
DOI 10.1109/ISIE.2011.78
115