IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 11, NOVEMBER 2013 4161
Image Noise Reduction via Geometric Multiscale
Ridgelet Support Vector Transform and
Dictionary Learning
Shuyuan Yang, Member, IEEE, Wang Min, Linfang Zhao, and Zhiyi Wang
Abstract—Advances in machine learning technology have
made efficient image denoising possible. In this paper, we propose
a new ridgelet support vector machine (RSVM) for image noise
reduction. Multiscale ridgelet support vector filter (MRSVF) is
first deduced from RSVM, to produce a multiscale, multidirec-
tion, undecimated, dyadic, aliasing, and shift-invariant geometric
multiscale ridgelet support vector transform (GMRSVT). Then,
multiscale dictionaries are learned from examples to reduce
noises existed in GMRSVT coefficients. Compared with the avail-
able approaches, the proposed method has the following charac-
teristics. The proposed MRSVF can extract the salient features
associated with the linear singularities of images. Consequently,
GMRSVT can well approximate edges, contours and textures in
images, and avoid ringing effects suffered from sampling in the
multiscale decomposition of images. Sparse coding is explored
for noise reduction via the learned multiscale and overcomplete
dictionaries. Some experiments are taken on natural images, and
the results show the efficiency of the proposed method.
Index Terms—Ridgelet support vector machine, ridgelet
support vector filter, multidirection, noise reduction, dictionary
learning.
I. INTRODUCTION
M
AKING a close inspection of the recent process made
in image processing we will find that much advance-
ment can be attributed to better modeling and representa-
tions of images. Recently kernel technologies have made
more efficient image processing possible. For example, Kernel
Regression (KR) [1]–[4] and Support Vector Machine (SVM)
[5]–[7] have been used for nonlinear representations of images
to present better modeling of two-dimensional (2D) images.
When modeling images, the geometric regularity of images
Manuscript received June 21, 2012; revised February 2, 2013, April 26,
2013, and June 1, 2013; accepted June 3, 2013. Date of publication June
26, 2013; date of current version September 11, 2013. This work was
supported in part by the National Basic Research Program of China (973
Program) under Grant 2013CB329402, the National Science Foundation of
China under Grants 61072108, 61271290, 61272282, NCET-10-0668, and
9140A24070412DZ0101, and the National Research Foundation for the Doc-
toral Program of Higher Education of China under Grant 20120203110005.
The associate editor coordinating the review of this manuscript and approving
it for publication was Prof. Hsueh-Ming Hang.
S. Yang, L. Zhao, and Z. Wang are with the Key Lab of Intelligent
Perception and Image Understanding of Ministry of Education, Xidian
University, Xi’an 710071, China (e-mail: syyang2009@gmail.com;
xiyouzhaolf@163.com; w5z2y0@163.com).
M. Wang is with National Key Lab of Radar Signal Processing, Xidian
University, Xi’an 710071, China (e-mail: wangmin@xidian.edu.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.2013.2271114
is important in inciting the human visual attention, so the
geometric regularities along singularity of edges or contours
should be emphasized for more accurate representation. Sev-
eral works have explored the adaptive approximation of images
[8]–[10] that allow to adaptively choose steering kernels. How-
ever, they only explore the variable weight factors of isotropic
kernels to formulate a data-adaptive regression, to preserve
fine details and local characteristics of images. Moreover, the
steering kernels need to be iteratively chosen from a gradient
or similarity estimation, which makes the location of the
geometric regularity very complex.
When the image is partitioned into sufficiently small
patches, the geometric regularity can be reduced to the linear
singularity along edges. Therefore, in order to exploit the
anisotropic regularity along edges, an appropriate regression
should include elongated functions that are nearly parallel
to the edges. Recently there has been a growing interest
in Geometric Multiscale Analysis (GMA) of images, and
its fast growth provides a group of new basis such as
Ridgelet [11], Curvelet [12], Contourlet [13], which can
effectively detect and capture different types of geometric
information of images. Among them, Ridgelets have proven
to be very efficient in estimating multidimensional surfaces
exhibiting specific sorts of spatial inhomogeneities, such as
linear and curvilinear singularities [11]. In this paper, based on
the previous work on Ridgelet kernel regression [15], [16], we
combine Ridgelets with Least Square-Support Vector Machine
(LS-SVM), to propose a new Ridgelet Support Vector Machine
(RSVM) for noise reduction of images. Images are regarded as
a 2D function whose inputs are the coordinate of pixels, and
outputs are the pixel values. RSVM is used to approximate
the function by formulating an optimization problem with a
set of linear equality constraints. Firstly a Multiscale Ridgelet
Support Vector Filter (MRSVF) is deduced from RSVM, to
produce a multiscale, multidirection, undecimated, dyadic,
aliasing and shift-invariant transform, Geometric Multiscale
Ridgelet Support Vector Transform (GMRSVT). Then several
multiscale and overcomplete dictionaries are learned from
the GMRSVT coefficients, and then used to reduce noises
via a sparse assumption of clean images [17]–[29]. Using
overcomplete representations and sparse assumption to denoise
images has drawn a lot of research attention in very recent
years [17]. Its basic idea is that clean image patches can
be sparsely coded by a dictionary while noisy image patches
cannot. Learning a dictionary that yields sparse coding of a
1057-7149 © 2013 IEEE