Evolution-enhanced multiscale overcomplete dictionaries learning
for image denoising
Shuyuan Yang
a,
n
, Min Wang
b
, Meirong Wei
a
, Licheng Jiao
a
a
Key Lab of Intelligent Perception and Image Understanding of Ministry of Education, Xidian University, Xi’an 710071, China
b
National Key Lab of Radar Signal Processing, Department of Electrical Engineering, Xidian University, Xi’an 710071, China
article info
Article history:
Received 7 May 2011
Received in revised form
29 November 2011
Accepted 24 January 2012
Available online 20 February 2012
Keywords:
Multiscale overcomplete dictionaries
learning
Evolution-enhanced
Image denoising
abstract
In this paper, a multiscale overcomplete dictionary learning approach is proposed for image denoising
by exploiting the multisca le property and sparse re presentation of images. The images are firstly
sparsely represented by a translation invariant dictionary and then the coefficients are denoised using
some learned multiscale dictionar ies. Dictionaries learning can be reduced to a non-convex l
0
-norm
minimization problem with multiple variables, so an evolution-enh anced algorithm is proposed to
alternately optimize the variables. Some experiments are taken on comparing the performance of our
proposed method with its counterparts on some benchmark natural images, and the superiorities of our
proposed method to its counterparts can be observed in both the visual result and some numerical
guidelines.
& 2012 Elsevier Ltd. All rights reserved.
1. Introduction
Overcomplete (or redundancy) is important in transformation-
based image denoising methods to have the shift invariance
property (Coifman and Donoho, 1994; Bui and Chen, 1998). For
example, with the growing realization of deficiencies of orthogonal
wavelets in denoising images, some redundant multiscale trans-
forms have been introduced, including undecimated Wavelets (Lang
et al., 1996), Curvel et (Starck et al., 2002), Contourle t (Do and
Vetterli, 2002), Wedgelet (Demaretetal.,2005), Bandelet (Zhang
et al., 2010), Shearlet (Blu and Luisier, 2007) and so on. In the past
decade, using spatial overcomplete representation and sparsity for
images denoising has drawn much attention of researches (Elad
et al., 2006; Elad and Aharon, 2006; Elad and Aharon, 2006). Its basic
idea is that the sparse representation (SR) of images will help in
automatically selecting the primary components in images while
reducing the noise components, as long as the dictionary can well
describe the characteristics of images. In more recent works (Aharon
et al., 2006; Protter and Elad, 2009; Chatterjee and Milanfar, 2009;
Turek et al., 2010; Dong et al., 2011), image patches prove to well
represent the statistical properties of the whole image, so a large
number of image patches are taken as the examples from which a
dictionary can be learned. The patches are taken from the noisy
image and then sparsely represented and restored, which lead to
state-of-the-art denoising result.
Although the SR-based denoising methods have proved to work
well on the natural images (Aharon et al., 2006; Protter and Elad,
2009; Chatterjee and Mila nfar, 2009 ; Turek et al., 2010; Dong et al.,
2011), the SR is all executed in the spatial domain. On the other hand,
it has been aware that making avail of the multiscale properties of
images will obtain better denoising result (Bui and Chen, 1998; Lang
et al., 1996; Starck et al., 2002; Do and Vetterli, 2002; Demaretetal.,
2005; Zhang et al., 2010; Blu and Luisier, 2007). Therefore, in this
study we take both the overcomplete representations of images in
spatial domain and transformation domain into account, and propose
a multiscale dictionaries learning approach for image denoising.
Some multiscale overcomplete dictionaries are learned from example
patches, and then used to reduce the noise distributed in the different
scales of images. We reduce the image denoising to an l
0
-norm
minimization problem with multiple variables. The available optimi-
zation schemes for this NP-hard problem can be mainly divided into
two catalogs: approximation method and relaxation method. The
approximation method includes greedy algorithms and shrinkage
algorithms. A greedy strategy abandons exhaustive search in favor
of a series of locally optimal single-term updates. Its basic idea is
to represent a signal as a weighted sum of atoms taken from a
dictionary, such as matching pursuit (Mallat and Zhang, 1993),
orthogonal matching pursuit (Tropp and Gilbert, 2007)andtheir
variants (Donoho et al., 2006; Needell, 2009). The approximation
method can correctly pick up atoms in the case of existing sparse
solution and the selection rule is simple to understand. However, it is
characteristics of heavy computation, slow convergence, and can only
work well in the noiseless case. The shrinkage strategy iterates
between shrinkage/thresholding operation and projection onto
Contents lists available at SciVerse ScienceDirect
journal homepage: www.elsevier.com/locate/engappai
Engineering Applications of Artificial Intelligence
0952-1976/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2012.01.021
n
Corresponding author.
E-mail address: syyang@xidian.edu.cn (S. Yang).
Engineering Applications of Artificial Intelligence 25 (2012) 1259–1264