IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 14, NO. 10, OCTOBER 2005 1647
Image Up-Sampling Using Total-Variation
Regularization With a New Observation Model
Hussein A. Aly, Member, IEEE, and Eric Dubois, Fellow, IEEE
Abstract—This paper presents a new formulation of the regu-
larized image up-sampling problem that incorporates models of
the image acquisition and display processes. We give a new ana-
lytic perspective that justifies the use of total-variation regulariza-
tion from a signal processing perspective, based on an analysis that
specifies the requirements of edge-directed filtering. This approach
leads to a new data fidelity term that has been coupled with a
total-variation regularizer to yield our objective function. This ob-
jective function is minimized using a level-sets motion that is based
on the level-set method, with two types of motion that interact si-
multaneously. A new choice of these motions leads to a stable solu-
tion scheme that has a unique minimum. One aspect of the human
visual system, perceptual uniformity, is treated in accordance with
the linear nature of the data fidelity term. The method was imple-
mented and has been verified to provide improved results, yielding
crisp edges without introducing ringing or other artifacts.
Index Terms—Data fidelity, gamma correction, image up-sam-
pling, interpolation, level-sets motion (LSM), observation model,
regularization, total variation.
I. INTRODUCTION
D
IGITAL-image magnification with higher perceived res-
olution is of great interest for many applications, such
as law enforcement and surveillance, standards conversions for
broadcasting, printing, aerial- and satellite-image zooming, and
texture mapping in computer graphics. In such applications, a
continuous real-world scene is projected by an ideal (pin-hole)
optical system onto an image plane and cropped to a rectangle
. The resulting continuous image is acquired by a physical
camera to produce a digital lower resolution (LR) image (i.e.,
lower than desired) defined on a lattice (following the notation
of [1], [2]). This camera, including the actual optical compo-
nent, is modeled as shown in Fig. 1 as a continuous-space filter
followed by ideal sampling on . The problem dealt with in
this paper is, given the still LR image , obtain the best per-
ceived higher resolution (HR) image defined on a denser sam-
pling lattice . Here, we hypothesize that an ideal HR image
defined on a denser lattice can be obtained in principle di-
rectly from by a virtual camera, which can similarly be mod-
Manuscript received March 13, 2004; revised September 14, 2004. This work
was supported in part by the Ministry of Defence, Egypt, and in part by the Nat-
ural Sciences and Engineering Research Council of Canada. The associate editor
coordinating the review of this manuscript and approving it for publication was
Dr. Thierry Blu.
H. A. Aly is with the Ministry of Defence, Cairo, Egypt (e-mail:
haly@ieee.org).
E. Dubois is with the School of Information Technology and Engineering,
University of Ottawa, Ottawa, ON K1N 6N5 Canada (e-mail: edubois@uot-
tawa.ca).
Digital Object Identifier 10.1109/TIP.2005.851684
eled by filtering with a continuous-space filter followed by
ideal sampling on . Our goal is then to obtain an estimate of
denoted by with the highest perceptual quality.
Many solution methods for the image magnification problem
exist in the literature, with a broad quality range. We can
categorize these solution methods into model-based and
non-model-based ones. Non-model-based methods use linear
or nonlinear (adaptive) interpolation. Linear interpolators
range from straightforward pixel repeat [which is also called
zero-order hold (ZOH)], bilinear, or bicubic [3] interpolation
to embedding in spline kernel spaces [4]–[7]. Simple linear
methods suffer from staircasing (blocking) of oblique edges,
blurring of the object boundaries and texture, and ringing in
smooth regions that are adjacent to edges. Splines produce
better quality up-sampled images than those obtained by
straightforward linear interpolators, but are known to produce
oscillatory edges with significant ringing near them. Analysis
of this effect based on image isophotes (iso-intensity contours)
can be found in [8]. These drawbacks of the linear methods
have led to research in adaptive methods, whose goal is to
preserve the sharpness of strong edges in the up-sampled image
. They adapt the interpolation method used according to
the edges given in the LR image and, hence, are generally
called edge-directed interpolation [9]–[12]. Another nonlinear
approach exploits local correlation of the samples without
explicitly extracting edges by defining a local metric that de-
termines the local participation weight of each sample of in
interpolating a sample of [13]–[17]. Adaptive methods can
produce clearly visible edges as compared to those produced
by the linear class, enhancing the overall perceived quality of
the resulting images. However, this class has the drawbacks of
relying on good edge estimation or local correlation and every
implementation is sensitive to the orientations of the edges. De-
spite the fact that the sharpness of the edges is being enhanced
by adaptive methods, the crispness of long edges is not well
handled and they are usually wavy, and blotching occurs on the
boundaries of edges. Furthermore, there is no solid theoretical
base that unifies the realization of the approaches of this class
and every approach stands on its own.
Model-based image up-sampling methods rely on modeling
the imaging processes and using sophisticated regularization
methods describing a priori constraints. According to the for-
mulation of Fig. 1, it can be shown that can be related to
by down-sampling, as shown in Fig. 2 [18], or, more generally,
by arbitrary rate conversion in the case of . Without loss
of generality, we only consider the case of ; the more
general case can be handled by an up-sampling step to an inter-
1057-7149/$20.00 © 2005 IEEE