IEEE SIGNAL PROCESSING LETTERS, VOL. 18, NO. 2, FEBRUARY 2011 103
Mean Square Error Estimation in Thresholding
Soosan Beheshti, Senior Member, IEEE, Masoud Hashemi, Student Member, IEEE, Ervin Sejdic ´ , Member, IEEE,
and Tom Chau, Senior Member, IEEE
Abstract—We present a novel approach to estimating the mean
square error (MSE) associated with any given threshold level in
both hard and soft thresholding. The estimate is provided by using
only the data that is being thresholded. This adaptive approach
provides probabilistic confidence bounds on the MSE. The MSE
bounds can be used to evaluate the denoising method. Our sim-
ulation results confirm that not only does the method provide an
accurate estimate of the MSE for any given thresohlding method,
but the proposed method can also search and find an optimum
threshold for any noisy data with regard to MSE.
Index Terms—Confidence bounds, estimation evaluation, mean
square error (MSE), thresholding.
I. INTRODUCTION
E
STIMATING a set of unknown parameters corrupted by
additive noise is one of the most important problems in
science and engineering. A common measure for the quality of
the estimators is the mean-square error (MSE). In many areas
(e.g., estimation, denoising, modeling), the aim is to estimate
and/or minimize a form of the MSE. When a suboptimal esti-
mation algorithm in the sense of MSE is used, it is of our interest
to compare the resultant MSE to the minimum (optimum) MSE
(MMSE). However, computing the MMSE is usually cumber-
some as well. To overcome this problem various application-de-
pendent methods have been proposed to estimate bounds for
MMSE. For example, in [1] a lower bound was provided for
error estimation of diffusion filters which is based on covariance
inequality and Cramer–Rao bounds [2]. Similar developments
can be found in [3]–[6] which are geared towards estimating a
desired unknown parameter from an observed pulse-frequency
modulated signal. MSE estimators can also be found in appli-
cations dealing with linear estimation and regression models
(e.g., [7], [8]). Nevertheless, generalizations of these findings
are often limited as they often deal with specific applications,
classes of signals, and/or a range of signal to noise ratio. In par-
ticular, these previous findings are generally not applicable in
denoising applications, where estimation of MSE for various
threshold values is desired.
Manuscript received October 05, 2010; revised November 18, 2010; accepted
November 21, 2010. Date of publication December 10, 2010; date of current
version December 23, 2010. The associate editor coordinating the review of
this manuscript and approving it for publication was Prof. Markku Renfors.
S. Beheshti is with the Department of Electrical and Computer Engineering,
Ryerson University, Toronto, ON M5B 2K3 Canada (e-mail: soosan@ee.ry-
erson.ca).
M. Hashemi is with the Institute of Biomaterials and Biomedical Engineering
(IBBME), University of Toronto, Toronto, ON M5S 1A1 Canada (e-mail: sayed-
masoud.hashemiamroabadi@utoronto.ca).
E. Sejdic ´ is with the Division of Gerontology, Beth Israel Deaconess Medical
Center and Harvard Medical School, Harvard University, Boston, MA 02215
USA (e-mail: esejdic@bidmc.harvard.edu).
T. Chau is with Bloorview Research Institute and the IBBME, University of
Toronto, Toronto, ON M5S 1A1 Canada (e-mail: tom.chau@utoronto.ca).
Digital Object Identifier 10.1109/LSP.2010.2097590
In this letter, we propose a method for the MSE estimation
in denoising applications. Different denoising techniques exist
for removing noise from the desired data, ranging from linear
transformations such as Wiener filters [9] to modern approaches
that are mostly concentrated on using wavelet coefficients and
shrinkage for noise cancelation [10]–[14]. Our development is
in line with these modern approaches. Specifically, our method
is applicable for any thresholding technique and any type of
signal structure. To achieve this goal, we first study the structure
of the MSE in a thresholding scenario. Then, we provide prob-
abilistic bounds on the MSE by using only the available noisy
data. A similar method has been presented for MSE estimation
for subspace selection and linear modeling (e.g., [15] and [16]).
To develop an MSE estimator for the purpose of thresholding,
the proposed method is based on the same fundamental princi-
ples as this previous technique. Here we provide probabilistic
confidence bounds on the MSE as a function of the available
data.
The letter is arranged as follows. Section II summarizes the
background of thresholding and formulates the problem under
consideration. In Section III, the structure of MSE in thresh-
olding is provided. Section IV introduces the method for the
MSE estimation by using only the available noisy data, while
Section V covers the simulation results. Section VI draws the
concluding remarks.
II. PROBLEM STATEMENT AND MOTIVATION
The unknown noiseless data is corrupted by an ad-
ditive white Gaussian noise . The noisy data
of length :
(1)
is available for . The additive noise is a
sample of a zero mean random variable with variance .
The data is projected onto another orthogonal basis that presents
the noiseless data with possibly fewer nonzero coefficients than
the data length. The most popular example of such a basis is
the class of orthogonal wavelets. The associated coefficients of
the data in the new bases are the result of the following inner
product
(2)
where , are the complete orthonormal basis,
and are the coefficients of noisy and noiseless data, and is
the coefficient of the additive noise. Due to the orthonormality
of the transformation, the transformed additive noise is zero
mean with the same noise variance .
Thresholding is classified into the two types of hard and soft.
Assume that the value of threshold is , then we denote the hard
thresholded coefficients with and the soft thresholded
version with . We use the notation of where both hard
and soft thresholding satisfy conditions.
1070-9908/$26.00 © 2010 IEEE