3736 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 15, NO. 12, DECEMBER 2006
Image Denoising Via Sparse and Redundant
Representations Over Learned Dictionaries
Michael Elad and Michal Aharon
Abstract—We address the image denoising problem, where
zero-mean white and homogeneous Gaussian additive noise is to
be removed from a given image. The approach taken is based
on sparse and redundant representations over trained dictio-
naries. Using the K-SVD algorithm, we obtain a dictionary that
describes the image content effectively. Two training options are
considered: using the corrupted image itself, or training on a
corpus of high-quality image database. Since the K-SVD is limited
in handling small image patches, we extend its deployment to
arbitrary image sizes by defining a global image prior that forces
sparsity over patches in every location in the image. We show how
such Bayesian treatment leads to a simple and effective denoising
algorithm. This leads to a state-of-the-art denoising performance,
equivalent and sometimes surpassing recently published leading
alternative denoising methods.
Index Terms—Bayesian reconstruction, dictionary learning, dis-
crete cosine transform (DCT), image denoising, K-SVD, matching
pursuit, maximum a posteriori (MAP) estimation, redundancy,
sparse representations.
I. INTRODUCTION
I
N THIS paper, we address the classic image denoising
problem: An ideal image
is measured in the presence of
an additive zero-mean white and homogeneous Gaussian noise,
, with standard deviation . The measured image is, thus
(1)
We desire to design an algorithm that can remove the noise from
, getting as close as possible to the original image, .
The image denoising problem is important, not only because
of the evident applications it serves. Being the simplest possible
inverse problem, it provides a convenient platform over which
image processing ideas and techniques can be assessed. Indeed,
numerous contributions in the past 50 years or so addressed
this problem from many and diverse points of view. Statistical
estimators of all sorts, spatial adaptive filters, stochastic anal-
ysis, partial differential equations, transform-domain methods,
splines and other approximation theory methods, morphological
analysis, order statistics, and more, are some of the many direc-
tions explored in studying this problem. In this paper, we have
no intention to provide a survey of this vast activity. Instead,
Manuscript received December 18, 2005; revised May 3, 2006. The associate
editor coordinating the review of this manuscript and approving it for publica-
tion was Dr. Tamas Sziranyi.
The authors are with the Department of Computer Science, The Technion—
Israel Institute of Technology, Haifa 32000, Israel (e-mail: elad@cs.technion.
ac.il; michalo@cs.technion.ac.il).
Digital Object Identifier 10.1109/TIP.2006.881969
we intend to concentrate on one specific approach towards the
image denoising problem that we find to be highly effective and
promising: the use of
sparse and redundant representations over
trained dictionaries.
Using redundant representations and sparsity as driving
forces for denoising of signals has drawn a lot of research
attention in the past decade or so. At first, sparsity of the unitary
wavelet coefficients was considered, leading to the celebrated
shrinkage algorithm [1]–[9]. One reason to turn to redundant
representations was the desire to have the shift invariance
property [10]. Also, with the growing realization that regular
separable 1-D wavelets are inappropriate for handling images,
several new tailored multiscale and directional redundant
transforms were introduced, including the curvelet [11], [12],
contourlet [13], [14], wedgelet [15], bandlet [16], [17], and the
steerable wavelet [18], [19]. In parallel, the introduction of the
matching pursuit [20], [21] and the basis pursuit denoising [22]
gave rise to the ability to address the image denoising problem
as a direct sparse decomposition technique over redundant
dictionaries. All these lead to what is considered today as some
of the best available image denoising methods (see [23]–[26]
for few representative works).
While the work reported here is also built on the very same
sparsity and redundancy concepts, it is adopting a different
point of view, drawing from yet another recent line of work
that studies example-based restoration. In addressing general
inverse problems in image processing using the Bayesian
approach, an image prior is necessary. Traditionally, this has
been handled by choosing a prior based on some simplifying
assumptions, such as spatial smoothness, low/max-entropy,
or sparsity in some transform domain. While these common
approaches lean on a guess of a mathematical expression for the
image prior, the example-based techniques suggest to learn the
prior from images somehow. For example, assuming a spatial
smoothness-based Markov random field prior of a specific
structure, one can still question (and, thus, train) the derivative
filters to apply on the image, and the robust function to use in
weighting these filters’ outcome [27]–[29].
When this prior-learning idea is merged with sparsity and re-
dundancy, it is the dictionary to be used that we target as the
learned set of parameters. Instead of the deployment of a pre-
chosen set of basis functions as the curvelet or contourlet would
do, we propose to learn the dictionary from examples. In this
work we consider two training options: 1) training the dictio-
nary using patches from the corrupted image itself or 2) training
on a corpus of patches taken from a high-quality set of images.
This idea of learning a dictionary that yields sparse represen-
tations for a set of training image-patches has been studied in
a sequence of works [30]–[37]. In this paper, we propose the
1057-7149/$20.00 © 2006 IEEE