IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 23, NO. 12, DECEMBER 2014 5123
Fast Image Upsampling via the Displacement Field
Lingfeng Wang, Huaiyu Wu, and Chunhong Pan
Abstract—In this paper, we present a fast image upsampling
method within a two-scale framework to ensure the sharp
construction of upsampled image for both large-scale edges
and small-scale structures. In our approach, the low-frequency
image is recovered via a novel sharpness preserving interpola-
tion technique based on a well-constructed displacement field,
which is estimated by a cross-resolution sharpness preserving
model. Within this model, the distances of pixels on edges are
preserved, which enables the recovery of sharp edges in the high-
resolution result. Likewise, local high-frequency structures are
reconstructed via a sharpness preserving reconstruction algo-
rithm. Extensive experiments show that our method outperforms
current state-of-the-art approaches, based on quantitative and
qualitative evaluations, as well as perceptual evaluation by a user
study. Moreover, our approach is very fast so as to be practical
for real applications.
Index Terms— Image upscaling, displacement field, super-
resolution.
I. INTRODUCTION
I
MAGE enlargement, or image upsampling, is a funda-
mental and challenging problem in the imaging research
area. The major difficulties come from two aspects. The first
aspect is how to recover sharp edges and textures while
suppressing visual artifacts, such as ringing, aliasing, blocking,
and blurring. Sharp edges are essentially helpful to make an
upsampled image visually clear. The second aspect is how
to ensure the upsampling operation fast enough to satisfy the
requirement of real applications.
To produce a high-resolution (HR) image, the simplest and
effective way is to apply analytical interpolation formula,
e.g., the bilinear and bicubic schemes. These methods are
very fast. However, they often produce undesired artifacts on
salient edges in the HR image, such as blurring, staircasing and
ringing. To preserve sharp edges, many edge-directed methods
have been proposed, where the sharp edges with minimal jaggy
or ringing artifacts can be produced by encoding edge based
knowledge. Recently, some example-based methods have been
proposed, where high-frequency information is learned from a
Manuscript received February 10, 2014; revised May 18, 2014 and July 27,
2014; accepted September 17, 2014. Date of publication September 25, 2014;
date of current version October 21, 2014. This work was supported in part
by the National Natural Science Foundation of China under Grant 91338202,
Grant 61175025, and Grant 61272049 and in part by the Beijing Natural
Science Foundation under Grant 4132075. The associate editor coordinat-
ing the review of this manuscript and approving it for publication was
Dr. Brendt Wohlberg.
The authors are with the Department of National Laboratory of
Pattern Recognition, Institute of Automation, Chinese Academy of Sciences,
Beijing 100190, China (e-mail: lfwang@nlpr.ia.ac.cn; hywu@nlpr.ia.ac.cn;
chpan@nlpr.ia.ac.cn).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIP.2014.2360459
set of training images. However, the high computational cost
makes these methods impractical for real applications.
To recover sharp edges in an effective and efficient manner,
we introduce a two-scale sharpness preserving upsampling
method. The flowchart of our method is shown in Fig. 1.
A key feature of our method is the separation of large-
scale sharp edges recovering from small-scale texture details
reconstructing. In large-scale interpolation, sharp edges in
the recovered HR result are ensured by a novel sharpness
preserving interpolation technology which adjusts the bicubic
upsampling image based on a well-constructed displacement
field. To recover small-scale textures, we present a sharpness
preserving reconstruction algorithm. Although our method
recovers large and small scale structures step-by-step, it does
not need to distinguish them, explicitly.
The advantages or details of our method are highlighted on
as follows (Codes are available at www.sigvc.org/lfwang.):
• A two-scale approach is proposed to recover HR images
via interpolation and reconstruction respectively. As a
result, both large-scale sharp edges and small-scale tex-
ture details can be faithfully recovered by our approach.
• A new interpolation algorithm based on the displacement
field is proposed to ensure the HR image preserve sharp-
ness from the LR image.
• A sharpness preserving reconstruction procedure is
applied to recover small-scale details while preserving
large-scale sharpness.
• Our approach is very fast compared with current state-
of-the-art approaches.
II. P
REVIOUS WORK
Single image upsampling has been extensively studied by
the communities in the field of computer graphics, computer
vision and image processing. In this section, we briefly intro-
duce the principles of the main approaches and the motivations
of our method.
The interpolation based methods [1]–[5] are widely used to
produce HR images because of their simplicity. The princi-
ple of these methods is to construct a rational interpolation
function. For example, in [3], they proposed a local adaptive
linear interpolation by the linear minimum mean square-
error estimation. Although complex interpolation functions are
constructed in these methods, they still tend to introduce visual
artifacts, such as ringing, aliasing, and blurring on edges.
To address this problem, in our interpolation based scheme, we
focus on initial placement of pixels other than the construction
of interpolation function. As a result, the sharp edges can be
well preserved in the HR result.
1057-7149 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.