Rigid Image Registration by using Corner and Edge Contents
with Application to Super-Resolution
Duangnapa Wangrattanapranee
Department of Communications and Integrated Systems
Tokyo Institute of Technology, Meguro-ku, Tokyo, 152-8552, Japan
duangnapa@nh.cradle.titech.ac.jp
Akinori Nishihara
Center for Research and Development of Educational Technology
Tokyo Institute of Technology, Meguro-ku, Tokyo, 152-8552, Japan
aki@cradle.titech.ac.jp
Abstract
Image registration method by using corner and edge con-
tents is proposed. In this method, the algorithm runs as a
coarse-to-fine strategy to find a precise alignment between
images in a rigid planar motion. In a coarse stage, land-
marks in both images are detected by Harris corner detec-
tion. Rough alignment is geometrically found by matching
a group of corners, instead of one-on-one, across two im-
ages. Next in a refining stage, Radon transform is applied to
compute the precision of registration in a subpixel accuracy.
Our algorithm is compared with the FFT-based method on
both simulated and real world images. With the same ac-
curacy level with FFT-based method, the results show su-
periorities in both degree of freedom and computational
complexity. Finally, super-resolution imaging is introduced
to demonstrate the capability of the proposed registration
method on real world application.
1 Introduction
Image registration is the process of aligning two sets of
data or images which differ in some details into the same co-
ordinate system. The differences can be caused by different
time taken, different perspective viewpoints, and/or differ-
ent image acquisition devices. Image registration is a cru-
cial step in many applications e.g. super-resolution imag-
ing, medical imaging, image mosaicing, satellite imaging,
etc.
Comprehensive surveys of image registration methods
by [1, 2] show that the important factor in image regis-
tration problem is to assume the transformation model and
correctly find its parameters. This makes image registration
method as an application-dependent solution. One way to
categorize the existing methods is by using a feature cri-
terion: feature-based methods estimate the transformation
parameters by extracting feature points of two images, and
feature-less methods estimate the transformation parame-
ters by either determining the intensity information of the
entire images or applying some transformations e.g. FFT-
based methods, projection-based methods.
To our knowledge, the FFT-based methods are one of the
most widely used methods due to its simplicity and ability
in subpixel motion estimation. However, it requires very
high computational cost which limits existing methods to a
certain degree of motion in order to speed up the algorithm,
such as in [5]. In this paper, we propose a hybrid regis-
tration method that overcomes those limitations: no limita-
tions in global planar motion estimation with effective com-
putation time, by combining the use of motion flexibility
from feature-based methods with the high accuracy align-
ment from feature-less methods. We first extract the land-
mark features across the two images. Rather than matching
the one-by-one feature points, we group them into clusters
and then perform a group matching instead. The correspon-
dences between the best matched portions and its runner up
give out rough geometric estimates of the transformation
parameters. Next, the precise value is determined by apply-
ing the Radon transform on edge-images. Details of theo-
ries behind each stage are described in Section 2. In Sec-
tion 3, the experimental results on both simulated and real
world images are shown. In Section 4, we provide a possi-
ble application of our method in super-resolution imaging.
Finally, the conclusion of this article is given in Section 5.
Digital Image Computing: Techniques and Applications
978-0-7695-3456-5/08 $25.00 © 2008 IEEE
DOI 10.1109/DICTA.2008.24
32