June 10, 2010 / Vol. 8, No. 6 / CHINESE OPTICS LETTERS 573
Image registration method based on improved Harris
corner detector
Qi Zeng (QQQ lll)
∗
, Liu Liu (444 èèè), and Jianxun Li ( oooïïïÊÊÊ)
Automation Department, School of Electronic & Information and Electrical Engineering, Shanghai Jiao Tong University,
Shanghai 200240, China
∗
E-mail: zqmark@gmail.com
Received November 2, 2009
Harris corner detector is a classic tool to extract feature. It is stable to illumination change and rotation
but unstable to more complicated transform. In order to register images with different viewpoints, we
extend Harris corner detector to scale-space to gain invariance to scale change, then we apply affine
shap e adaptation to the scale invariant point until convergence is reached, giving it invariance to affine
transform. With these local features, we use general feature descriptor and matching algorithm to generate
matches and then use the matches to calculate the geometric transform matrix, which enables the final
registration. Result shows that our algorithm can get more accurate matches than scale invariant feature
transform SIFT, and less difference exists between registered images.
OCIS co des: 100.2000, 100.5010.
doi: 10.3788/COL20100806.0573.
Image registration is a process of finding a proper ge-
ometric transform between two images that can align
corresponding points in them. It is the foundation of
applications, such as image fusion, medical image pro-
cessing, and three-dimensional (3D) image reconstruction
and is widely used in medical imaging and remote sens-
ing. Recently, image registration has been a topic widely
discussed, and methods with high efficiency and accu-
racy have been developed. For example, Guizar-Sicairos
et al. proposed to use nonlinear optimization and matrix-
multiply discrete Fourier transforms to register two-
dimensional (2D) images with sub-pixel accuracy
[1]
. In
medical image registration, Modersitzki proposed to inte-
grate the concept of local rigidity to the Flexible Image
Registration Toolbox (FLIRT), giving extra constraint
for rigid object in non-rigid optimization process
[2]
.
In medical image registration, many methods, like the
maximization of mutual information (MMI) method
[3]
,
treat the registration process as an optimization prob-
lem and use this measure as an object function to find
the best transform. However, because of the diversity
of image registration problem, one framework for image
registration is to assume that the transform between im-
ages is of a certain kind (rotation, scaling, etc.), then
to use local feature matches to calculate the transform
matrix. For example, when translation, rotation, and
scaling are present, a registration method was proposed
to deal with that situation
[4]
. In this framework, local
feature matching is of utmost importance. A good lo-
cal feature should generally have a clear mathematically
well-founded definition, also, the local image structure
around the local feature is rich in terms of local informa-
tion contents, such as derivative information
[5]
, curvature
information, etc. Most imp ortantly, a good local feature
should be tolerant to image noise, changes in illumina-
tion, scaling, rotation, as well as changes in viewpoint.
The most classic local feature is Harris corner
[6]
detec-
tor, it is stable to illumination change and rotation but
unstable to more complicated transform. Scale invariant
feature transform (SIFT)
[7]
detector, which takes advan-
tage of the scale invariant nature of scale-space represen-
tation, is stable to scale change but not very good at lo-
cating corners which often correspond to significant local
structures. In this letter, we propose a method that can
register images with affine and scale transform. Begin-
ning with traditional Harris corner detector, we extend
it to scale-space, giving the detector invariance to scale
transform, and then we apply affine shape adaptation,
giving the detector invariance to affine transform. We
use default descriptor and matching algorithm to gen-
erate matches, and use those matches to calculate ge-
ometrical transform parameters. Finally, we transform
one image using the geometrical transformation matrix
to align with the other image.
Geometric transform between images is the founda-
tion in our image registration. To describe the transform,
we introduce projective geometry
[8]
. Unlike traditional
Cartesian coordinate, a point in projective coordinate is
defined as a vector of three elements: the first two are
x and y, and the third coordinate is introduced to deal
with the situation of infinite point. The third element
is 1 when the point is not infinite and 0 when the point
is infinite. Because when divided by 0, any small value
becomes infinite, using projective coordinate makes an
infinite point homogeneous with any other point in the
space. A geometric transform can be written in matrix
form as
x
0
1
x
0
2
x
0
3
=
"
h
11
h
12
h
13
h
21
h
22
h
23
h
31
h
32
h
33
#Ã
x
1
x
2
x
3
!
.
(1)
A geometric transform can be further divided to
isometies, similarity transforms, and affine transforms.
Here we focus on affine transforms because of its ubiq-
uity:
Ã
x
0
y
0
1
!
=
"
a
11
a
12
t
x
a
21
a
22
t
y
0 0 1
#Ã
x
y
1
!
. (2)
1671-7694/2010/060573-04
c
° 2010 Chinese Optics Letters