1046 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 3, MARCH 2012
Image Registration Under Illumination Varia
Using Region-Based Confidence Weighted
M -Estimators
Mohamed M. Fouad, Member, IEEE,Richard M. Dansereau, Senior Member, IEEE,and
Anthony D. Whitehead, Member, IEEE
Abstract—We present an image registration model for image sets
with arbitrarily shaped local illumination variations between im-
ages. Any nongeometric variations tend to degrade the geometric
registration precision and impact subsequent processing. Tradi-
tional image registration approaches do not typically account for
changes and movement of light sources, which result in interimage
illumination differences with arbitrary shape. In addition, these
approaches typically use a least-square estimator that is sensitive to
outliers, where interimage illumination variations are often large
enough to act as outliers. In this paper, we propose an image reg-
istration approach that compensates for arbitrarily shaped inter-
image illumination variations, which are processed using robust
-estimators tuned to that region. Each -estimator for each il-
lumination region has a distinct cost function by which small and
large interimage residuals are unevenly penalized. Since the seg-
mentation of the interimage illumination variations may not be
perfect, a segmentation confidence weighting is also imposed to re-
duce the negative effect of mis-segmentation around illumination
region boundaries. The proposed approach is cast in an iterative
coarse-to-fine framework, which allows a convergence rate similar
to competing intensity-based image registration approaches. The
overall proposed approach is presented in a general framework,
but experimental results use the bisquare -estimator with region
segmentation confidence weighting. A nearly tenfold improvement
in subpixel registration precision is seen with the proposed tech-
nique when convergence is attained, as compared with competing
techniques using both simulated and real data sets with interimage
illumination variations.
Index Terms—Image registration, local illumination changes,
-estimation, segmentation confidence weighting.
I. I NTRODUCTION
G
EOMETRIC image registration (GIR) is the process of
determining the point-by-point correspondence between
two views of a scene and the transformation to align these cor-
respondences. The GIR of a set of images is a common pre-
processing step in many applications [1], such as generating
Manuscript received March 09, 2011; revised June 30, 2011; accepted Au-
gust 17, 2011. Date of publication September 08, 2011; date of current version
February 17, 2012. The associate editor coordinating the review of this manu-
script and approving it for publication was Dr. Chares Creusere.
M. M. Fouad is with the Department of Company Engineering, Military Tech-
nical College, Cairo, Egypt (e-mail: mmafoad@ieee.org).
R. M. Dansereau is with the Department of Systems and Company En-
gineering, CarletonUniversity, Ottawa,ON K1S 5B6, Canada(e-mail:
rdanse@sce.carleton.ca).
A. D. Whitehead is with the School of Information Technology, Carleton Uni-
versity, Ottawa, ON K1S 5B6, Canada (e-mail: awhitehe@connect.carleton.ca).
Digital Object Identifier 10.1109/TIP.2011.2167344
panoramic image mosaics [2], performing super-resolution
hancement [3], image stitching [4], change detection [5], m
ical imaging [6], and remote-sensing applications [7]. For
and similar applications, subpixel registration accuracy is
essary for satisfactory postprocessing results.
A primary problem in many image registration scenarios
that, often, there also exists nongeometrical transform var
within the image set. These variations manifest from many
tors, including changes in lighting, motion/deformations o
tures in the scene, changes in occluded and occluding obje
and changes in exposure, aperture, focus, lens distortion o
photometric factors. In addition, multimodal data sets cou
clude larger variations in terms of imaging modality, resol
and information content. Literature exists on handling ma
these confounding factors for GIR, although arguably few
proaches have been proposed to flexibly handle local chan
in illumination, with the multimodal techniques being a no
exception that can have good registration performance bu
the level of fine subpixel precision that we seek. The focus
this paper is hence to look at how interimage illumination
tions in the scene can be compensated within an intensity-
GIR framework, with particular attention to the arbitrary s
typical for illumination variations resulting from shadows
from one or more light sources. Note that the light source
assumed to be far field from the scene.
Interimage variations in illumination do indeed impact the
GIR, as discussed in the following literature. Refs. [8] and
present similar global-illumination models (GIMs) with sca
gain and offset modeling in their GIR to relate the intensity
levels of an image pair to improve the registration perform
In [8],an additional smoothness constraint, with predefined
neighborhoods, is also imposed. In [10], an affine illumina
model(AIM) is proposed with an extended gain and offset
model that also used triangular or quadrilateral region sup
(not arbitrarily shaped) to allow an improved modeling of
illumination variations. In [11],a dualinverse compositional
algorithm is presented based on an assumption that the geo-
metric and photometric transformations can be replaced, thus
impeding the use of explicit local photometric transformat
Although Bartoli [11] models the photometric changes ins
of illumination variations, he does so in a fashion using sca
gain and offset, when dealing with gray images, quite the
as in the GIM. In this paper, we will refer to the GIM and t
AIM as our comparator group.
The approaches in [8]–[11] do not take into account mul-
tiple shading levels with arbitrariness in the shape ofthe
1057-7149/$26.00 © 2011 IEEE