Digital Signal Processing Reference
In-Depth Information
ered a nonparametric method, i.e., it makes no a priori assumption. It is distinct from
parametric methods in which parameters must be estimated for a particular model
that is assumed a priori. For example, the most popular parametric method is that of
using least squares estimation.
13.3.1 Additive Gaussian White Noise Model
The following is a measurement model:
Y(k)
=
X(k)
+
B(k)
(13.10)
where
Y(k)
is the measurement signal of size
N
,
X(k)
is the a priori unknown
useful signal, and
B(k)
is a random noise perturbation (usually assumed to be white
Gaussian with variance
σ
2
).
The elimination or reduction of this additive noise can be achieved nonlinearly by
using multiresolution analysis under the assumption that the appropriate choice of a
decomposition basis allows discrimination of the useful signal from noise. The un-
derlying idea is that the useful signal can be described by a small number of coeffi-
cients of high-amplitude wavelets, and that the noise is spread across all coefficients.
This hypothesis justifies, in part, the traditional use of denoising by thresholding.
If
d
j
(k)
represent the wavelet coefficients of the measured signal, the estimation
of the wavelet coefficients of the useful signal, denoted
d
j
(k)
, is generated by two
types of thresholding:
•
Hard thresholding
d
j
(k)
if
d
j
(k)
|
|
>S
,
d
j
(k)
=
(13.11)
0
otherwise;
•
Soft thresholding
⎧
⎨
⎩
d
j
(k)
S
if
d
j
(k)>S
,
d
j
(k)
+
S
if
d
j
(k)<
−
S
,
0
−
d
j
(k)
=
(13.12)
otherwise
σ
√
2ln
N
where
N
is the size of the measured signal, and
σ
represents the
noise standard deviation.
A robust estimator of
σ
is given by:
=
with
S
Med
d
1
(k)
σ
=
1
.
4826
×
(13.13)
d
1
(k)
where Med
|
|
designates the median value of the wavelet coefficients, for
d
1
(k)
j
=
1, in increasing order
{|
|
,
0
≤
k
≤
N/
2
}
.
Search WWH ::
Custom Search