NEIGHBORHOOD REGRESSION FOR EDGE-PRESERVING IMAGE SUPER-RESOLUTION
Yanghao Li, Jiaying Liu
∗
, Wenhan Yang, Zongming Guo
Institute of Computer Science and Technology, Peking University, Beijing, P.R.China, 100871
ABSTRACT
There have been many proposed works on image super-
resolution via employing different priors or external databas-
es to enhance HR results. However, most of them do not
work well on the reconstruction of high-frequency detail-
s of images, which are more sensitive for human vision
system. Rather than reconstructing the whole components
in the image directly, we propose a novel edge-preserving
super-resolution algorithm, which reconstructs low- and high-
frequency components separately. In this paper, a Neighbor-
hood Regression method is proposed to reconstruct high-
frequency details on edge maps, and low-frequency part is
reconstructed by the traditional bicubic method. Then, we
perform an iterative combination method to obtain the esti-
mated high resolution result, based on an energy minimiza-
tion function which contains both low-frequency consistency
and high-frequency adaptation. Extensive experiments eval-
uate the effectiveness and performance of our algorithm. It
shows that our method is competitive or even better than the
state-of-art methods.
Index Terms— Image Super-Resolution (SR), Edge-
Preserving, Neighborhood Regression, High-frequency De-
tails
1. INTRODUCTION
Image Super-Resolution (SR) reconstruction is currently a
popular research area in signal processing. In many digital
imaging applications, high-resolution images are often de-
sired for later image processing. The SR task exactly focuses
on the enhancement of image resolution. In general, given
one or more low resolution (LR) images, it is responsible for
mapping them to high resolution (HR) images. However, s-
ince SR image reconstruction is generally a severely ill-posed
problem, many methods for SR reconstruction have been pro-
posed these years. They can be roughly divided into three cat-
egories. Interpolation-based methods use linear or non-linear
interpolation algorithms to restore a single image, such as
New Directed Interpolation (NEDI) [1]; multi-frames-based
methods [2, 3] aim to utilize information from a set of LR im-
ages to compose an HR image; and learning-based methods
∗Corresponding author
This work was supported by National Natural Science Foundation of China
under contract No. 61472011, National High-tech Technology R&D Pro-
gram (863 Program) of China under Grant 2013AA013504 and Beijing Nat-
ural Science Foundation under contract No.4142021.
use machine learning techniques in SR reconstruction, which
are very popular in recent years.
The key part of learning-based methods is to learn the
mapping between LR and HR images, and different methods
are proposed to model the relationship. Following the max-
imum a posteriori (MAP), some proposed methods formed
Markov Random Field (MRF) to connect the LR and HR im-
ages and then restored the HR images. Sparse representation
based methods [4, 5] learned their own coupled LR and HR
dictionaries to represent the relationship instead, based up-
on sparse signal representation. Moreover, assuming that two
manifolds of the LR and HR image patches are locally in sim-
ilar geometries, Neighbor Embedding (NE) methods [6, 7] es-
timated HR images by linearly combining the HR neighbors.
Because NE approaches do not need to learn dictionaries or
solve MRF, they dramatically reduce the execution complex-
ity.
However, all of the above methods directly restore the w-
hole components in the image. Since low-frequency compo-
nent contains most of energy in images, restoring the whole
parts together causes that the methods mainly apply on the
low-frequency component. So it results in ignoring the high-
frequency details of images, which are more sensitive for hu-
man vision system. In SR reconstruction, most work can do
well in the low-frequency part since it is more coherent and
less complex. Nevertheless, because the high-frequency part
has more variations and represents details, restoration for this
part is much harder and remains as a challenge of SR recon-
struction. Thus, we pay more attention to the reconstruction
of high-frequency details of the images in this paper.
In our work, we also consider that the SR reconstruction
includes two stages: reconstruction in the low-frequency part
and high-frequency details. We propose a new Neighborhood
Regression method for edge-preserving Super-Resolution
(NRSR), which mainly focuses on the reconstruction of high-
frequency details and the combination with reconstruction of
low-frequency part. Fig.1 illustrates the framework of our
approach. We consider low- and high-frequency components
of images separately. Then, we propose a Neighborhood
Regression method for reconstruction on edge maps, which
represent high-frequency details. Finally, we develop an in-
corporation method to combine low-frequency consistency
and high-frequency adaptation to obtain the final result.
The remainder of the paper is organized as follows: we
1201978-1-4673-6997-8/15/$31.00 ©2015 IEEE ICASSP 2015