A Robust Method for Aligning Large-Photometric-
Variation and Noisy Images
Shiqian WU
#+1
, Wangming XU
*2
,Jun JIANG
#3
, Yimin QIU
*4
, Liangcai ZENG
#+5
#
College of Machinery and Automation, Wuhan University of Science and Technology
Hubei,430081, China
{
1
shiqian.wu,
3
jiangjun85,
5
zengliangcai}@wust.edu.cn
*
College of Information Science and Engineering, Wuhan University of Science and Technology
Hubei,430081, China
{
2
xuwangming,
4
qiuyimin}@wust.edu.cn
+
Hubei Collaborative Innovation Center for Advanced Steels
Hubei,430081, China
Abstract—We propose a novel robust method to accurately
align large-photometric-variation and noisy images emerged
from high dynamic range (HDR) imaging. First, the keypoints
are detected from optimal multi-binary images to eliminate the
noise and photometric effects. The feature descriptor, which is
translation-, rotation- and scale-invariant and robust to noise
and photometric changes, is then developed, and feature
matching is implemented efficiently by use of the structure
information of the keypoints. Finally, mutual information (MI) is
employed in the RANSAC method for homography estimation
and performance assessment, which makes the results accurate
and stable. Experiments carried out on synthesized images
demonstrate that the proposed method is much more robust to
both photometric changes and noise than the SIFT method, the
state-of-the-art alignment method.
I. INTRODUCTION
Image registration or alignment has been a fundamental
problem in image processing and computer vision. There are a
large number of techniques proposed for a variety of
applications, such as, different viewpoints, different durations,
and/or different sensors. Generally speaking, the registration
methods can be classified into two categories [1]: pixel-based
(area-based) algorithms and feature-based algorithms. Pixel-
based methods work by directly minimizing pixel-to-pixel
dissimilarities to find motion parameters between two images,
which are dependent of image intensities. Feature-based
methods, on the other hand, first extract distinctive features
from each image, then match and warp the features to derive
parametric transformations. As feature-based methods do not
work directly with image intensities, it is frequently used
when photometric changes. It is demonstrated in [2] that SIFT
(scale-invariant feature transform) [3] provides the best
performance among the feature descriptors.
Unlike conventional cases of varying illumination, image
sequences from video surveillance which operate day and
night, and multi-exposed images for high dynamic range
(HDR) imaging, always have significant photometric
variations such as shadows, highlights, and photometric
changes. Such image pairs pose great challenges for
registration using aforementioned methods: 1) One feature
detected in one image (e.g., indoor feature in Fig. 1(a)) may
not occur in another one (Fig. 1(b)), which results in difficulty
to use feature-based method; 2)The intensities are not linearly
related due to non-linear imaging system, which yields
difficulty to normalize these images for area-based
registration; 3) Each image contains severely shadow
/highlight regions, which implies information loss and few
features are detected; 4) A specific intensity in one image
may map to multiple intensities in the other images, and vice
versa, for example, two saturated pixels Z
1
(u) = 255, Z
1
(v) =
255 in Fig. 1 (a) may become Z
2
(u) = 255, Z
2
(v) = 240 in Fig.
1(b); 5) The images, especially from satellites and remote
sensing are always very noisy due to traveling through turbid
medium, which breaks the intensity relationship.
(a) (b)
Fig. 1 Photometric-variation images
To alleviate the effect of photometric variation on
extracting feature points, Tomaszewska and Mantiuk [4] and
Gevrekci and Gunturk [5] proposed to detect feature points in
contrast domain. It is shown in [5] that the repeatability rate of
the feature detector can on average be improved by about 25%.
On the other hand, photometric registration was proposed via
camera response function (CRF) [6]. Another technique in
consideration of photometric changes is to perform joint
geometric and photometric registration [7-8]. Recently, a
hybrid scheme employing an area-based method first, then a
feature-based method [9], and a two-stage method comprising
image normalization and local-binary-pattern (LBP)
representation [10], were developed respectively.
This work is extension of our previous alignment paradigm
[10] in the following aspects: 1) The proposed method works
not only for images which move in plane, but also for images
from different viewpoints. Therefore, feature-based method is
978-1-4673-7478-1/15/$31.00 ©2015 IEEE