Accelerated Joint Image Despeckling Algorithm in the Wavelet and Spatial Domains

Adib Akl & Charles Yaacoub
International Journal of Image Processing (IJIP), Volume (9) : Issue (3) : 2015 102
Accelerated Joint Image Despeckling Algorithm in the Wavelet
and Spatial Domains
Adib Akl adibakl@usek.edu.lb
Faculty of Engineering, Department of Telecommunications
Holy Spirit University of Kaslik (USEK)
Jounieh, P.O. Box 446, Lebanon
Charles Yaacoub charlesyaacoub@usek.edu.lb
Faculty of Engineering, Department of Telecommunications
Holy Spirit University of Kaslik (USEK)
Jounieh, P.O. Box 446, Lebanon
Abstract
Noise is one of the most widespread problems present in nearly all imaging applications. In spite
of the sophistication of the recently proposed methods, most denoising algorithms have not yet
attained a desirable level of applicability. This paper proposes a two-stage algorithm for speckle
noise reduction jointly in the wavelet and spatial domains. At the first stage, the optimal
parameter value of the spatial speckle reduction filter is estimated, based on edge pixel statistics
and noise variance. Then the optimized filter is used at the second stage to additionally smooth
the approximation image of the wavelet sub-band. A complexity reduction algorithm for wavelet
decomposition is also proposed. The obtained results are highly encouraging in terms of image
quality which paves the way towards the reinforcement of the proposed algorithm for the
performance enhancement of the Block Matching and 3D Filtering algorithm tackling multiplicative
speckle noise.
Keywords: Denoising, Spatial Filter, Speckle, Wavelet.
1. INTRODUCTION
Multiplicative in nature, speckle noise is a common problem found in different imaging
applications such as ultrasound, sonar and radar imaging [1, 2]. Originating from the
superposition of acoustical echoes coming with random phases and amplitudes during acquisition
or transmission, it tends to reduce the image resolution and contrast and blur important details
[1].
Despeckling can address the multiplicative nature of the noise, or transform the noisy image to
the logarithmic domain where multiplicative noise becomes additive, and apply additive noise
reduction techniques [2].
In the past decades, several algorithms have been proposed for image denoising. Hybrid order
statistics filters for speckle reduction are proposed in [2]. A preprocessing technique is used in
order to transform the noise in the logarithmic domain to a Gaussian-like noise, which allows for
better filtering results using known denoising techniques [3-5]. Lee [6-8], Kuan [3, 9], and Frost
[10] filters are still widely used in many applications. In general, they succeed to reduce speckle
in homogenous areas. However, in heterogeneous areas, speckle is retained. Therefore, they are
not able to perform a full removal of speckle without blurring any edges because they rely on local
statistical data related to the filtered pixel and this data depends on the occurrence of the filter
window over an area. Wavelet-based denoising techniques [4, 10, 11, 12] represent the image
jointly in the spatial and frequency domains. They rely on the sparse representation of the
wavelet sub-bands coefficients in order to be distinctly thresholded. The Block Matching and 3D

Filtering (BM3D) algorithm [13, 14] is a non-local denoising technique in a non-spatial domain. It
combines non-local image modeling and the sparse representation of the wavelet domain.
In the following, a two-stage despeckling algorithm is proposed. It consists of jointly denoising the
speckled image in the wavelet and spatial domains. At the first stage, the traditional speckle
reduction filter of Kuan [3] is adapted to the specificities of the filtered image by estimating its
optimal parameter value automatically, based on edge pixels statistics and noise variance. Then
the resulting sub-optimal spatial filter is used at the second stage to smooth the approximation
sub-band coefficients in the wavelet domain. Since wavelet decomposition is time-consuming
when dealing with large sized images, a complexity reduction algorithm is also proposed. Note
that the Kuan filter is used as the basis for the spatial filtering due to its versatility. However, other
algorithms could be used.
The remainder of this paper is organized as follows. Existing noise reduction methods are
reviewed in Section 2. The proposed algorithm is then described in Section 3. In Section 4,
results are shown and discussed. Finally, conclusions are drawn and prospects are provided in
Section 5.
2. EXISTING NOISE REDUCTION ALGORITHMS
In the past decades, several image denoising algorithms have been proposed. For instance, the
spatial means filters (arithmetic mean, geometric mean, harmonic mean and contraharmonic
mean filters [15]) smooth the local variations in an image by blurring the noise. The spatial order-
statistic filters (median, max, min, midpoint and alpha-trimmed mean filters [15]) are based on
ordering (ranking) the values of the pixels contained in the image area encompassed by the filter.
The spatial adaptive filters (adaptive local noise reduction and adaptive median filters [15])
change their smoothing behavior based on statistical characteristics of the image inside the filter
region. The spatial filters listed above have shown to successfully deal with a large panel of noisy
images in situations when only additive noise is present. However, they are not able to reduce
speckle noise which is non-additively combined with the underlying image.
Various nonlinear filtering techniques have also been proposed [16]. They seek to reduce the
effect of speckle noise while preserving the informative structure of the underlying image. Some
of the best known standard speckle noise reduction filters are the methods of Lee [6-8], Gamma
[24], Kuan [3, 9] and Frost [24]. These filters use the second-order sample statistics within a
minimum mean squared error estimation approach.
2.1 Lee Filter
The locally adaptive Lee multiplicative filter [7, 8] is based on a multiplicative noise image model
as follows:
( , ) ( , ) ( , ) ,g x y f x y n x y= × (1)
where g denotes the observed image, f the original image and n the multiplicative speckle noise
[6].
Based on an assumption that the noise is white, with unity mean, and uncorrelated with the image
f, the multiplicative Lee filter gives the best mean-squared estimate of f at each pixel ( , )g x y by:
( , )( , ) ( , ) ,x yz x y f k g x y f = + − 
(2)
where ( , )z x y is the filtered pixel, f is the mean of the original image pixels under the filtering
window (i.e. the local mean) and ( , )x yk is the gain factor given by:
2
( , ) 2 2 2
,
f
x y
n f
k
f
σ
σ σ
=
+
(3)

where
2
fσ is the variance of the original image pixels under the filtering window (local variance),
and
2
nσ is the noise variance. The local adaptation of the filter is based on the calculation of the
local statistics f and
2
fσ from the data sample estimates g and
2
gσ determined over a local
neighborhood window.
This requires the knowledge of the Coefficient of Variation (CoV), which is the ratio of the
standard deviation to the mean in homogeneous areas. If the original image is not available, the
noise variance can be determined from the number of looks or the Equivalent Number of Looks
(ENL). This parameter effectively controls the amount of smoothing applied to the image. The
filter size greatly affects the quality of the processed image. If the filter is too small, the noise
filtering algorithm is not efficient. If the filter is too large, some details of the image will be lost [17-
19].
An improved version of the basic Lee filter also exists; the modified Lee filter which relies on a
ratio-based edge detection algorithm used to estimate the edge strength at each pixel in the
image [20].
2.2 The Gamma Maximum A Posteriori Filter
The Gamma Maximum A Posteriori (MAP) filter [21] is used primarily to filter speckled radar data
while preserving high frequency features. It’s based on a Bayesian analysis of the image
statistics. It performs spatial filtering on each individual pixel in an image using the grey level
values in a square window surrounding each pixel. The Number of Looks (NLOOK) parameter
effectively controls the amount of smoothing applied to the image. It affects the speckle
coefficient of variation ( )uC and the upper speckle coefficient of variation max( )C . A small
NLOOK value leads to more smoothing and a larger NLOOK value preserves more
image features.
If Ci≤Cu (Ci is the image variation coefficient), the filtering window is over a
homogenous area and smoothing is applied. If Ci≥Cmax, the filtering window is over
an area of high local standard deviation (crossing edge-pixels), and therefore edge-pixel is
replicated [24].
If Cu<Ci<Cmax, the filtered pixel value based on the Gamma estimation of the
contrast ratios within the appropriate filter window is given by:
( 1)
( , ) ,
2
W NLOOK g D
z x y
W
− − × +
=
×
(4)
where:
2
2 2
2 2
1
, ( 1) 4 ( , ) and is the mean of
the window pixels.
u
i u
C
W D g W NLOOK W NLOOK g g x y g
C C
+
= = × − − + × × × ×
−
2.3 Kuan Filter
Similarly to the Lee filter, Kuan filter [3, 9, 22] is designed to smooth out speckle noise while
retaining shape features in the image by applying the Minimum Mean Square Error
(MMSE) criterion and it is applied to the logarithmic transformation of the noisy image. Kuan
filter is used primarily to filter speckled radar data. It performs spatial filtering on each individual

pixel in an image using the grey level values in a square window surrounding each pixel [24]. The
resulting grey-level value for the smoothed pixel is given by:
( , ) ( , ) (1 ) ,z x y g x y W g W= × + − (5)
where g(x,y) is the center pixel of the filtering window, g is the mean value of intensity within the
filtering window, and W is a weighting function which depends on the Number of Looks (NLOOK)
parameter [23] and given by:
2 2
2
1 /
,
1
u i
u
C C
W
C
−
=
+
(6)
where Cu and Ci are the estimated noise variation coefficient and the image
variation coefficient, respectively:
1
, ,
NLOOK
g
u iC C
g
σ
= = (7)
where gσ is the standard deviation of intensity within the filtering window.
Same as the Lee filter case, a small NLOOK value corresponds to a high noise variation
and a low W and therefore leads to more smoothing, while a larger NLOOK value
corresponds to a low noise variation and a high W which preserves more image
features.
Theoretically, the correct value for NLOOK should be the Equivalent Number of Looks (ENL) of
the image. Depending on the application, the user may experimentally adjust the NLOOK value
so as to control the effect of the filter [22].
2.4 Frost Filter
Frost method consists in adjusting the filter's parameters according to local area statistics about
the target pixel [24]. When uniform regions are filtered, the filter acts as a mean filter and when
high contrast regions are filtered, the filter acts as a high-pass filter with rapid decay of elements
away from the filter center. Thus, large uniform areas will tend to be smoothed out and speckle
removed, while high contrast edges and other objects will retain their values [24]. The frost filter
can be considered as an adaptive-weighted-mean filter since it uses an adaptive filtering
algorithm, which is an exponential damped convolution kernel that adapts itself to features by
computing a set of weighting factors for each pixel within the filtering window as follows:
2
,
g
DAMP T
g
nM e
σ     − × × 
     = (8)
where DAMP is a factor that determines the extent of the exponential damping for the image, gσ
is the standard deviation of the filter window, g is the mean value within the window and T is the
absolute value of the grey level distance between the center pixel and its surrounding pixels in
the filter window [24].
The Enhanced Frost filter is an extension to the basic filter that further divides the image into
homogeneous, heterogeneous and isolated point target areas [19]. It can significantly improve the
ability of speckle mitigation in the vicinity of edges and small features, thus retaining more image
details.
2.5 Denoising Using Wavelets
Wavelet transform [25] has been extensively studied in recent years and used for many
application domains, mainly in image compression and noise reduction. It consists of a set basis
functions used to analyze signals in both spatial and frequency domains simultaneously. The

basic functions of the wavelet transform help to isolate signal discontinuities since high pass
filtering is used to obtain detail information and low pass filtering is used to retrieve a smoother
approximation of the signal, making it possible to analyze the signal at different scales [22].
Multi-resolution processing of the Discrete Wavelet Transform (DWT)
Multi-resolution processing consists in constructing a set of child wavelets from a mother wavelet
using scaling and wavelet functions [22]. These two functions form a filter bank consisting of a
low pass and high pass filters. The idea of multi-resolution processing through wavelet
decomposition is to pass the signal through the filter bank; the signal is decomposed to detail
coefficients (output of the high pass filter) and approximation coefficients (output of the low pass
filter). Then, the filter outputs are down-sampled by 2 in the purpose of discarding half the
samples. The decomposition is repeated to further increase the frequency resolution [25-27].
Application to Image Denoising
Considering an image corrupted by an additive noise and modeled as ,g f n= + where f
denotes the unknown, noise-free image and n the noise [28, 29], wavelet-based denoising
(Figure 1) consists in:
1) Applying the DWT to g.
2) Thresholding the detail coefficients (Wavelet shrinkage).
3) Inverse transforming (IDWT) the result to obtain an estimation z of the original image f.
FIGURE 1: DWT Denoising Block Diagram.
When dealing with speckled images, a logarithmic transformation is applied to the noisy image
before wavelet decomposition, to transform the multiplicative noise, into additive noise and after
wavelet reconstruction, an exponential transformation is applied to reverse the logarithmic
operation.
In wavelet domain, the output of the low pass filter consists of the high magnitude and low
frequency components (approximation coefficients) and the output of the high pass filter consists
of the low magnitude and high frequency components (detail coefficients). Figure 2 shows a one
level, 2D wavelet decomposition scheme [30] where L(.) represents the low pass filtering
operator, and the subscript (L) represents a low pass filtered output. Similarly, H(.) and the
subscript (H) represent high pass filtering and a high-pass filtered output, respectively. The
symbol (↓2) represents a down-sampling operator by a factor of 2.
As a result of filtering and down sampling (Figure 2), four sub-bands are obtained: scaling
(approximation) coefficients (gLL), horizontal detail coefficients (gHL), vertical detail coefficients
(gLH) and diagonal detail coefficients (gHH). The (HL), (LH) and (HH) sub-bands, represent the
high frequency (and low magnitude) components and the (LL) sub-band represents the low
frequency (and high magnitude) components. At the next level of decomposition, only the (LL)
component is passed to the decomposition process.
The reconstruction process or the Inverse Discrete Wavelet Transform (IDWT) consists of
assembling back the wavelet coefficients to the original image. After each inverse low and high
pass filtering, an up-sampling process (zeros insertion) is required to reverse the decomposition
process.

The main idea of wavelet denoising is to threshold only the high frequency components while
preserving most of the features in the image by retaining the approximation sub-band.
There are two thresholding methods frequently used; hard [26, 31] and soft [32] thresholding. In
hard thresholding, the input is kept if its amplitude is greater than a threshold T, otherwise it is
forced to zero. In soft thresholding, if the absolute value of the input is less than or equal to a
threshold T, then the output is forced to zero, otherwise, the output is a scaled version of the
input.
FIGURE 2: One level, 2D Wavelet Decomposition Block Diagram.
Finding a suitable threshold is an important task in wavelet shrinkage. In fact, choosing a very
large threshold will shrink to zero almost all coefficients, which results in over smoothing the
image, while choosing a very small threshold will yield a noisy result which is close to the input
[29]. VisuShrink [25] is an approach that uses a Universal (global) threshold applied to all sub-
bands and scales after decomposition [33]. However, this threshold can be unwarrantedly large
because it depends on the number of pixels yielding overly smoothed images. In addition, it
ignores the difference between sub-bands at different scales [34]. BayesShrink [25, 31] is a sub-
band adaptive threshold selection technique that determines a specific threshold for each sub-
band assuming a Generalized Gaussian Distribution (GGD) [35]. This method depends on the
standard deviation of the noise-free image, and the GGD shape parameter [34]. The method
consists of finding, for each sub-band, a threshold that minimizes the expected value of the mean
square error (Bayesian Risk) [32, 36].
2.6 The Block Matching and 3D Filtering (BM3D) Algorithm
The BM3D algorithm is a non-local denoising technique [13, 14] in a non-spatial domain. It
combines non-local modeling and the sparse representation of the wavelet domain. The algorithm
is divided into three processing stages: Block matching, 3D filtering, and aggregation.
In the block matching stage, the noisy image g is divided into blocks x XG ∈ where x represents
the position of each block in the whole image X. For each block, the patches with most
resemblance are grouped in a 3D array. Therefore, groups whose elements have a high degree
of similarity are constructed separately. Two patches are similar if the Euclidean distance
between them is less than a given threshold [14, 37]. The block matching advantages are the
induction of a high correlation in the third dimension of the 3D array and the improvement of the
dispersion of all possible configurations of details present in the image [13].

3D filtering is a procedure that jointly filters a group of similar blocks by exploiting the similarities
between the grouped images and inside each image in the group. It consists of three different
sections: 3D transform, shrinkage and inverse 3D transform. Those three procedures are jointly
applied [13] on the block. The 3D transform consists of applying a 2D wavelet transform on each
patch of the block then a 1D wavelet or Discrete Cosine Transform (DCT) on the resulting
patches. As a result of the sparse representation given by the 3D transform, the shrinkage
process can effectively attenuate the noise by eliminating the coefficients relying under a certain
chosen threshold. The inverse 3D transform consists of assembling back the thresholded
coefficients to reconstruct the 3D block.
Finally, the aggregation stage consists of combining the patches in the 3D filtered groups
including the reference patch. A trivial solution is to compute the weighted mean of all the
estimated patches overlapping at a pixel position [13].
3. THE PROPOSED WAVELET/SPATIAL DESPECKLING ALGORITHM
In this section, a two-stage despeckling algorithm which consists in jointly denoising the corrupted
image in the wavelet and spatial domains is proposed. At the first stage, an automatic estimation
of the optimal Kuan filter parameter value based on edge pixels statistics and noise variance is
developed. Then the resulting adaptive filter is used at the second stage to spatially smooth the
approximation sub-band coefficients in the wavelet domain. Since a large number of wavelet
levels makes the despeckling algorithm computationally expensive, a complexity reduction
algorithm for wavelet decomposition is also proposed. Note that the proposed enhancement
targets the Kuan filter due to its versatility and adaptability to speckle noise strength, while a
similar study could have been performed with another spatial despeckling method to be used in
the hybrid filter that will be discussed in Section 3.2; we therefore focus on the Kuan filter without
loss of generality.
3.1 The Adaptive Spatial Filter
As explained earlier, Kuan filtering relies on the Number of Looks (NLOOK) parameter which
significantly affects the filtering performance and is usually taken equal to the Equivalent Number
of Looks (ENL) of the image. This parameter is manually adjusted, in order to control the strength
of the smoothing applied to the image.
We start by implementing the additive model of the Kuan filter using a set of test images (e.g.
Lena, Mandrill, LivingRoom, … [38]) corrupted by speckle noise with different distributions. By
setting NLOOK = ENL, we notice that filtered images are over-blurred, that’s why came our idea
of analyzing the Kuan filter performance with respect to the NLOOK parameter and proposing a
novel technique for selecting a suitable NLOOK value that yields near-optimal performance
compared to manual parameter selection, therefore eliminating the need for several runs of the
filter to reach the optimal performance. In this purpose, we define the adjusted NLOOK value as:
_ (1 ) ,NLK AD A ENL= + (9)
where A is the adjustment coefficient.
Figure 3 shows an example of the Lena image filtered with NLOOK=ENL (middle) and NLK_AD
as defined in (9), for A = 0.65 (right). It can be clearly seen that the image with the adjusted
NLOOK parameter (NLK_AD) is much sharper and its details are better preserved.
In the first part of this study, we aim at finding a closed form expression for the adjustment
coefficient that yields near optimal performance, without the need for several runs of the filter with
different NLOOK values. Therefore, we consider the effect of the filter window size, the image
detail, the noise distribution and the noise variance on the optimal adjustment coefficient A and
consequently, the NLK_AD parameter. Thus, all the listed parameters are fixed and the effect of
varying each one of them is studied.

FIGURE 3: Kuan filter results with different NLOOK values. Gaussian and Uniform noise distributions are
used in the top and bottom rows respectively. Each row shows (from left to right): the noisy image, the Kuan
filtered image using NLOOK=ENL and the corresponding Kuan filtered image using NLK_AD with A=0.65.
An 11×11 window with images of size 256×256 are used.
Effect of the Algorithm Parameters on the Optimal Adjustment Coefficient
A Gaussian speckle noise having a variance of 2600 is added to the Cameraman image [38] in
order to find the adjustment coefficient A that yields the best result using different window size.
The Peak Signal-to-Noise Ratio (PSNR) and the Signal to Mean Squared Error (S/MSE)
variations with respect to the NLK_AD parameter are plot and ensure that, for different odd-sized
filter windows, the optimal values of A are nearly constant.
On the other hand, test images are distorted with multiplicative speckle noise using the Uniform,
Gaussian, Rayleigh, and Erlang distributions while fixing all the other parameters and the optimal
value of A remains the same regardless of the noise distribution, for each test image.
In the purpose of studying the influence of the image itself on the filtering process, the filter is
applied on different images, having different features, from [38] and [39] with all the other
parameters being fixed. It can be noticed that the optimal adjustment coefficient varies with image
details. For example, the Pirate image [38] has a high percentage of edges, therefore, the
obtained optimal adjustment coefficient is greater than with the Lena image, which is expected
since more image features have to be preserved.
To find the relationship between the optimal adjustment coefficient A and the amount of image
detail, edge detection is performed on the set of images and the curves presenting the optimal A
with respect to the noise variance and the number of edge pixels in each image are drawn.
The Logarithmic and Polynomial Approximations
Since the noise variance affects the filtering process, the variation of A with the noise variance
and image detail is modeled as:
,d vA A A= + (10)
where Ad and Av are the detail and the noise variance components in A, respectively.

We observe that a higher noise variance corresponds to a lower value of A, which is expected
since a lower A value leads to more smoothing. Additionally, a higher noise variance corresponds
to higher edge percentage (Pe) values because the edge detection algorithm cannot completely
differentiate between noise and small edges. Curves for different noise variances have similar
behavior with respect to the percentage of edges in the image, therefore, we take as reference
one of the curves (for a given variance).
The curve with the solid dark line in Figure 4 shows the measured Ad with respect to Pe. It can
be seen that, for Pe exceeding 2%, Ad varies in a logarithmic or second order polynomial fashion.
The former can be expressed as:
ln( ) ,d eA Pα β= + (11)
and the latter as:
2
,d e eA P Pλ δ µ= + + (12)
where α = 0.4015, β = 0.054, λ = –0.01215, δ = 0.20512 and μ= –0.01394 determined by fitting
the measured results with (11) and (12) using the Least Squares method.
It can be observed in Figure 4 that, as the image detail (Pe) increases, Ad increases and
consequently, NLK_AD increases. Therefore, less smoothing is performed, as expected. The
logarithmic and polynomial approximations mainly overlap with measured data, except for
Pe < 2%, which proves high accuracy of the models in (11) and (12).
A similar study with respect to the noise variance shows that Av (Figure 5) can be accurately
modeled by a second order polynomial as:
2
( 2100) ( 2100) ,vA a v b v= − + − (13)
where a = –2x10
-8
and b = –42x10
-6
determined by curve fitting.
Note that for v = 2100, Av = 0. In fact, Av reflects the variation of the adjustment coefficient with
respect to the noise variance compared to the reference v = 2100 which has been used to derive
Ad in (11) and (12). On the other hand, Av decreases as v increases, which is also expected
since more smoothing is required (lower value of NLK_AD) when the noise increases.
FIGURE 4: Measured and approximated values for Ad for a noise variance of 2100.

FIGURE 5: Measured (solid line) and approximated (dashed line) values for Av.
3.2 The Hybrid Wavelet-Spatial Filter
In wavelet denoising, only detail coefficients are thresholded and approximation coefficients
remain stagnant, as explained earlier. This is intuitive since the approximation component is a
low-frequency image, usually assumed to be noise-free. In fact, this is only a theorectical aspect
since practically, approximation coefficients also contain speckle noise as it can be clearly
observed in the example shown in Figure 6. Wherefore, the proposed algorithm consists of
smoothing the approximation coefficients in addition to detail coefficients thresholding as follows:
1) Applying the Discrete Wavelet Transform to the speckled image.
2) Thresholding the detail coefficients (Wavelet shrinkage).
3) Spatial Smoothing of the approximation sub-band coefficients.
4) Applying the Inverse Discrete Wavelet Transform to the result.
FIGURE 6: Cameraman image of size 512×512 (up-left) [38], corrupted by speckle noise of unit mean and a
variance of 2400 (up-right), decomposed structure (down left) and the denoised image (down right). Note the
existence of noise in the approximation image of the decomposed structure and the remaining noise in the
denoised image (using sym1 wavelets with BayesShrink threshold selection).

An important issue in the approximation image spatial smoothing would be the preservation of the
image characteristics. In other words, the algorithm must efficiently smooth out the noise in the
approximation image without blurring its details or reducing the contrast and quality. Otherwise,
the image structure and dynamics would be lost. Therefore, the approximation sub-band
coefficients in the wavelet domain are smoothed using the previously developed adaptive Kuan
filter which leads to a better speckle removal without image distortion, as will be seen in Section
4.
3.3 The Proposed Wavelet Acceleration Algorithm
This section deals with accelerating the wavelet filtering process used in the proposed method.
The acceleration algorithm consists in finding the number of decomposition levels that yields near
optimal performance, without the need for several runs of the wavelet filter with different number
of decomposition levels, regardless of the image size, its edge percentage, the noise variance,
the wavelet type and base, the wavelet threshold selection technique and the wavelet
thresholding technique.
Relative Difference Threshold Selection
We start by applying 100 times, wavelet denoising on an image corrupted by Gaussian speckle
noise using different number of decomposition levels (from 1 level to the maximum number of
decomposition levels calculated as in [40]). For every run, the simulated speckle noise is
regenerated and the PSNR, the Coefficient of Correlation (CoC) and the ENL metrics of the
denoised image are calculated and saved.
The second step is the calculation of the relative difference between the ENL of each two
consecutive decomposition levels, for the purpose of finding a reasonable relative difference
threshold that allows the algorithm to stop before reaching the maximum number of
decomposition levels, without loss in performance. The same is performed for different images
from [38] and [39], different noise variances (2500, 2600, 3000, 3100, 3500, 4100, 4600 and
6000), different image sizes (128× 128, 256× 256, 512× 512 and 1024× 1024), different wavelet
types and bases (db2, db4, db7, sym1, sym2 and sym6) and using bayesShrink threshold
selection.
After analyzing the results, we noticed that a relative difference smaller than 0.2 leads to an
approximately steady PSNR value. As a result, 0.2 is taken as a threshold. Therefore, our
proposed algorithm consists of applying a level of wavelet decomposition and denoising,
computing the relative difference between the corresponding ENL and the ENL of the previous
decomposition level, and stop when the relative difference becomes less than 0.2 as shown in
Figure 7.
By limiting the number of decomposition levels, a computational gain is expected at the expense
of some loss in image quality, as will be shown in Section 4.
FIGURE 7: Flowchart of the Proposed Acceleration Algorithm.
YES
NO
Level of wavelet decomposition
BayesShrink Thresholding
Wavelet reconstruction
Relative Difference
R = (ENL(i) - ENL(i-1))/(ENL(i-1)) R<T END

4. PRACTICAL RESULTS
In this section, the proposed algorithms are put in practice. Besides the subjective qualitative
evaluation based on human perception, a quantitative analysis is essential. Thus, an objective
benchmark is used to study the quality-related outcomes of the filtering process. Namely, the
Peak Signal-to-Noise Ratio (PSNR), the Signal-to-Mean-Squared-Error (S/MSE), the Coefficient
of Correlation (CoC) and the Equivalent Number of Looks (ENL).
Peak Signal-to-Noise Ratio (PSNR)
PSNR is considered to be the least complex metric, as it defines the image quality degradation as
a plain pixel by pixel error power estimate. PSNR is an engineering term for the ratio between the
maximum possible power of a signal and the power of corrupting noise that affects the fidelity of
its representation [17]. In image processing, the signal is the original image, and the noise is the
error (blur) introduced by the denoising procedure.
Signal-to-Mean Squared Error (S/MSE)
In order to quantify the achieved performance improvement, the Signal-to-Mean Squared Error
can also be computed, based on the original and the noisy/denoised images as follows:
[ ]
[ ]
1 1
2
0 0
10 1 1
2
0 0
( , )
/ 10log ,
( , ) ( , )
m n
i j
m n
i j
I i j
S MSE
K i j I i j
− −
= =
− −
= =
 
 
 
=  
 
− 
 
 
∑ ∑
∑ ∑
(14)
where I is the original image and K the noisy or denoised image. This measure corresponds to
the classical SNR in the case of additive noise [41].
Coefficient of Correlation (CoC)
In ultrasound imaging, it is important to suppress speckle noise while at the same time preserving
the edges of the original image that often constitute features of interest for diagnosis. For this
reason, we also considered a qualitative measure for edge preservation, the Coefficient of
Correlation (CoC) metric:
( )
( ) ( )
,
,
, . ,
I I K K
I I I I K K K K
CoC
 ∆ − ∆ ∆ − ∆
=
 ∆ − ∆ ∆ − ∆  ∆ − ∆ ∆ − ∆
(15)
where I∆ and K∆ are the high-pass filtered versions of I and K respectively, obtained with a
3×3 pixel standard approximation of the Laplacian operator [25], I∆ is the mean value
of I∆ and K∆ is the mean value of K∆ and the operator  is given by:
1 1
0 0
( , ) ( , ). ( . ) ,
m n
i j
I K I i j K i j
− −
= =
 = ∑ ∑ (16)
where m×n is the size of the corresponding images I and K.
The CoC cannot exceed 1 in absolute value. It is 1 in the case of an increasing linear relationship
and - 1 in the case of a decreasing linear relationship. Its value lies in between in all other cases,
indicating the degree of linear dependence between the images. The absolute value of the
correlation measure should be close to unity to an optimal effect of edge preservation [41].
Equivalent Numbers of Looks (ENL)
The Equivalent Number of Looks is also a good approach for either estimating the speckle noise
variance in a noisy image (in general a SAR image), or evaluating the performance obtained after
filtering. ENL is often calculated over a uniform region, but due to difficulties in identifying uniform
areas, the image is divided into small areas of 25 x 25 pixels, the ENL is computed for each area

(17), then the average of these ENL values is taken as explained in [12]. Note that this latter
method is used throughout this paper. A large ENL value usually corresponds to a better
performance. The formula for the ENL calculation is given by:
2
mean
standard deviation
ENL
 
=  
 
. (17)
Note that PSNR, S/MSE, and CoC are quantitative metrics, they are based on a pixel by pixel
calculation and the original image is needed for computation, whereas ENL is based on the mean
and standard deviation, and it can be computed without the need of the original image [22].
4.1 The Performance of the Proposed Adaptive Spatial Filter
It is trivial that a good performing spatial filtering is necessary to get a denoised image of good
quality using the proposed approach. Otherwise, the image characteristics will be lost. In this
section, the NLOOK parameter estimation previously developed is exploited and compared to the
results obtained with the optimal NLOOK value determined manually. Simulations were
performed using different standard images from [38] and [39], corrupted by speckle noise of
variance ranging from 2100 to 4100. Since many of the samples lead to an approximately same
percentage (Pe) of edge pixels, and therefore to a similar algorithm behavior, table 1 summarizes
sample results obtained for 8 different images having different Pe values.
It can be observed that, except for Pe < 2 (e.g. Lady image), both the logarithmic and polynomial
estimations result in near optimal values for the adjustment parameter, and the PSNR loss does
not exceed 0.33 dB and 0.16 dB with logarithmic and polynomial estimations, respectively. In fact,
the behavior of the filter when dealing with few-detailed images (Lady image in table 1) is
expected, since both proposed approximations fail to be accurate in cases of very low Pe values
as shown in Figure 4. However, images with a percentage of edge pixels less than 2 rarely exist
(in our test datasets from [38] and [39], only 11% of the images had Pe < 2). Furthermore, the
approximation error can be largely reduced by setting Ad to a constant value (approximately 0.43)
for such low values of Pe as implies Figure 4.
It is important to mention that experimental results with such high noise variances are shown to
highlight the decent behavior of the proposed method in critical conditions. Moreover, when
dealing with slightly speckled images, no filtering is performed (W ≈ 1 in (5)), as will be verified in
section 4.2.
Figure 8-a shows the Cumulative Density Function (CDF) of the PSNR loss obtained with both
estimations, using the whole set of test images. It can be noticed that in both methods, 50% of
the filtered images undergo a loss in PSNR not exceeding 0.05 dB, whereas 85% and 89%
undergo a PSNR loss that does not exceed 0.11 dB in logarithmic and polynomial
approximations, respectively.
Even though values of Pe are more likely to occur between 2% and 8%, let us analyze the
proposed models’ behavior for values of Pe outside this interval. It can be deduced from Figure 4
that for low values of Pe, the measured Ad seems to saturate at a value close to 0.43, while the
logarithmic approximation reaches 0 for values of Pe close to 0.9% as shown in Figure 8-b, and
becomes negative for lower values.
Similarly, the polynomial approximation also fails to model Ad at low values of Pe, but results in
slightly better performance compared to the logarithmic model. This problem can be solved by
using any of the two models only when Pe exceeds a fixed threshold (e.g. for Pe ≥ 2%), or by
setting Ad to a constant value (e.g. Ad = 0.43) when Pe is below that threshold. On the other
hand, the polynomial approximation reaches a peak for Pe ≈ 8.5%, and then decreases as Pe
increases. Therefore, the polynomial model becomes inaccurate in this case. The same does not
occur with the logarithmic approximation since it keeps increasing as Pe increases.

Noise
variance
Image Pe (%)
Speckled
Image
PSNR
Log approximation Polynomial approximation Optimal values
100xA NLK_AD PSNR 100xA NLK_AD PSNR 100xA NLOOK PSNR
2100
Lady 0.9719 26.1012 4.2606 12.1041 24.276 17.393 13.6457 25.695 42 16.5228 28.246
WomanDarkHair 2.6634 20.6058 44.7361 30.1964 29.115 44.617 30.1987 29.152 44.617 30.1987 29.152
I06 3.1488 20.5841 51.2811 28.3623 27.569 52.98 28.2943 27.963 53.321 28.5143 27.988
Lena 4.4518 20.7322 65.3613 28.6966 27.766 65.837 28.8355 27.765 65.3613 28.6966 27.766
Board 6.0764 19.6519 77.8522 26.12 24.224 78.376 26.349 24.237 78.376 26.349 24.237
I13 6.2301 21.5436 79.2398 24.9994 26.599 73.987 24.9655 26.541 78.595 24.4561 26.632
LivingRoom 7.0168 20.9285 83 28.1954 25.992 82.703 28.0957 26.012 83.63 28.4462 26.028
Mandrill 11.597 20.5485 103.801 30.417 24.865 73.053 25.8463 24.713 103.801 30.417 24.865
2600
Lady 0.9887 25.2349 2.3487 10.611 24.451 15.098 11.9622 25.912 43 14.8931 28.838
WomanDarkHair 3.0743 20.8759 47.8966 26.4071 28.423 47.58 26.3707 28.435 44 25.7209 28.493
I06 3.3248 19.6843 52.0432 23.9424 27.107 50.524 23.5761 27.132 50.819 23.5999 27.281
Lena 4.5025 19.855 63.216 24.2587 27.195 65 24.5917 27.199 63.726 24.4509 27.2
Board 6.0455 18.7322 73 22.0966 23.658 75.598 22.5739 23.668 75.048 22.4301 23.698
I13 6.2467 20.5233 74.9848 22.1354 26.260 75.935 22.2013 26.191 75.213 22.5847 26.261
LivingRoom 6.5968 20.041 78.5514 23.8295 25.51 82 24.4238 25.512 78.436 23.787 25.57
Mandrill 11.382 19.6209 100.451 26.1791 24.308 72.045 22.5894 24.322 83 23.8904 24.365
3100
Lady 0.9901 24.5231 -1.1945 9.3747 24.321 11.523 10.5359 25.819 40 13.2466 29.024
WomanDarkHair 3.027 20.177 43.674 22.3196 27.926 43.361 22.2835 27.985 41 21.8922 28.119
I06 3.5269 18.9928 45.997 19.767 26.501 45.624 19.5671 26.663 51.407 20.223 26.614
Lena 4.4182 19.1324 58.8572 20.8946 26.767 62 21.2741 26.724 59.31 20.8339 26.785
Board 6.339 18.0201 73.3508 19.8109 23.235 69 19.2277 23.228 73.601 19.7984 23.266
I13 6.356 19.8108 75.473 19.4415 25.622 74.423 19.4445 25.992 74.337 19.5426 25.874
LivingRoom 7.2327 19.2983 78.6463 21.3184 25.078 77.193 21.1876 25.087 76 20.8615 25.121
Mandrill 12.095 18.9053 99.2916 23.3454 23.803 62.725 19.2675 23.891 79 21.0759 23.956
3600
Lady 0.9953 23.9122 -5.5841 8.2428 24.029 7.0173 9.2931 25.582 36 11.8257 28.94
WomanDarkHair 3.1719 19.5572 40.9514 19.5081 27.608 40.641 19.6611 27.589 37 18.9953 27.672
I06 3.6442 18.3526 43.3436 17.2142 26.201 43.221 16.8099 26.204 49.012 17.5012 26.254
Lena 4.5715 18.5216 55.6266 18.2715 26.398 56.18 18.363 26.464 56.18 18.363 26.464
Board 6.3707 17.4144 68.9511 17.4682 22.829 69.161 17.6717 22.837 65 17.0354 22.845
I13 6.4014 19.0212 70.3316 17.9968 25.596 70.44 17.2634 25.526 72.256 17.1994 25.598
LivingRoom 7.1453 18.7005 73.5582 18.6148 24.753 72.328 18.5073 24.736 72 18.5868 24.797
Mandrill 11.821 18.2575 93.7692 20.7329 23.4 60.474 17.1004 23.621 73.55 18.5456 23.6257
4100
Lady 0.9971 23.4004 11.1116 7.1373 23.456 1.4499 8.1364 25.193 35 10.8747 29.183
WomanDarkHair 3.2227 19.0674 35.9893 16.9639 27.374 35.689 16.9408 27.299 35.9893 16.9639 27.374
I06 3.618 17.7216 38.2526 14.7722 25.921 38.221 14.7677 24.999 44.098 15.2344 25.923
Lena 4.7283 17.9916 51.3807 16.1424 26.11 52.024 16.1452 26.118 52.024 16.1452 26.118
Board 6.3188 16.8767 63.0227 15.4386 22.472 63.297 15.4994 22.455 57 14.8499 22.531
I13 6.3904 18.7049 64.2214 15.9697 25.455 64.031 15.0996 25.402 60.215 15.0027 25.491
LivingRoom 7.9094 18.1322 72.0373 16.8162 24.427 68.422 16.5037 24.457 66 16.2858 24.465
Mandrill 12.867 17.733 91.5761 18.8708 23.045 44.941 14.2858 23.313 67 16.3813 23.372
TABLE 1: Summary of results obtained with different 512×512 speckled images using the proposed
adaptive spatial filter.

(a) (b)
FIGURE 8: (a) CDF of the loss in PSNR due to logarithmic (dotted line) and polynomial (solid line)
approximations. (b) Logarithmic (dashed line) and polynomial (solid line) approximations as defined in (11)
and (12).
Table 2, 3, 4 and 5 show a comparison of the Lee filter, standard (NLOOK=ENL) Kuan filter,
Gamma filter, Frost filter, Enhanced Frost filter, and the proposed enhanced Kuan filter using both
the logarithmic and polynomial approximations with a speckle noise of Gaussian, Uniform,
Rayleigh and Exponential distributions respectively.
The same is performed for different images of different edge percentage and the results were
approximately similar. It can be clearly seen that the performance of the filters is affected by the
noise distribution. For the examples shown in Tables 2-5, the Lee multiplicative filter performs
better than the Enhanced Frost filter with Uniformly distributed noise, and worse with Gaussian
noise, the Kuan filter gives better results than the Lee multiplicative filter except for the case with
NLOOK=ENL (where this latter parameter is not adjusted). Depending on the noise distribution,
either of the approximations (logarithmic or polynomial) can outperform the other in Kuan filtering
with adjusted NLOOK parameter.
In tables 2-5, it can also be observed that the best PSNR is obtained with either the Gamma MAP
filter or the proposed (modified) Kuan filter, whereas the Frost and enhanced Frost filters
preserve image features more than other filters as it can be noticed from CoC values.
PSNR S/MSE CoC ENL
Gaussian noise
Noisy image 19.2482 13.5918 0.1781 13.3615
Lee multiplicative filtered image 26.9749 21.3185 0.3028 98.0283
Kuan filtered image(NLOOK=ENL) 22.8675 17.2111 0.1155 31.6124
K.filtered image (Log. approximation) 27.2248 21.5685 0.3106 106.6198
K.filtered image
(Poly. approximation)
27.2389 21.5826 0.3154 110.6581
Gamma filtered image 28.1836 22.5272 0.3892 233.4135
Frost filtered image 26.3968 20.7404 0.4225 233.3297
Enhanced Frost filtered image 27.0823 21.4259 0.4394 240.8940
TABLE 2: Comparison of speckle noise filtering techniques on a 512×512 Lena image corrupted by a
Gaussian speckle noise using an 11×11 window size.

PSNR S/MSE CoC ENL
Uniform noise
Noisy image 19.1606 13.5042 0.1760 13.1826
K.filtered image
27.7968 22.1405 0.3403 145.0305
Uniform speckle noise using an 11×11 window size.
PSNR S/MSE CoC ENL
Rayleigh noise
Noisy image 19.4281 13.7718 0.1760 13.7621
K.filtered image
27.3523 21.6960 0.3236 116.7191
Rayleigh speckle noise using an 11×11 window size.
PSNR S/MSE CoC ENL
Exponential noise
Noisy image 20.0551 14.3987 0.1746 15.3187
K.filtered image
27.1044 21.4480 0.3684 169.5444
TABLE 5: Comparison of speckle noise filtering techniques on a 512×512 Lena image corrupted by an
Exponential speckle using an 11×11 window size.
Table 6 shows a comparison between the proposed adaptive Kuan filter and the basic and
enhanced versions of Lee, Kuan, and Frost filters using five different images. It can be noticed
that the optimal Kuan filter yields the highest PSNR, with the drawback of performing several filter
runs in order to tune the filter to yield the best possible output quality. Furthermore, by setting the
NLOOK parameter to ENL, Lee and Frost filters outperform the Kuan filter in their basic and
enhanced versions. Moreover, our proposed solution approaches the performance of the optimal
Kuan filter, with a negligible PSNR loss, and outperforms the basic and enhanced Lee and Frost
filters, for all the speckled test images used in the filtering process.

PSNR WomanD.H. Pirate Lena CameraMan LivingRoom
Noisy image 19.55 20.02 19.42 20.1 19.9
Lee 25.39 25.91 27.14 26.01 26.2
Frost 23.69 25.08 26.38 25.9 25.89
Enhanced Lee 25.61 26.01 27.30 26.78 26.88
Enhanced Frost 25.47 25.60 27.08 26.12 26.18
Kuan 23.34 24.21 24.02 24.92 25.1
Kuan (poly. approximation) 25.78 26.57 27.35 26.95 27.21
Kuan (log. approximation) 25.88 26.59 27.33 26.89 27.1
Kuan (optimal NLOOK) 25.92 26.70 27.44 27.11 27.35
TABLE 6: PSNR obtained with different despeckling filters using images of 512×512 size and an 11×11
filtering window.
It is important to note that in the literature [13, 14, 37], the Block Matching and 3D filtering (BM3D)
algorithm is tested on images corrupted by additive noise. Moreover, the method's different
parameters should be adequately chosen (depending on the noise strength, image size and
percentage of edge pixels etc) in order to perform well dealing with speckle noise. Those
parameters are the reference block size, the searching window size, the number of similar
patches in the 3D block and the shrinking threshold. Therefore, finding the optimal set of
parameters dealing with speckle noise is a very difficult task and choosing a non-optimal set of
parameters usually fails to outperform traditional despeckling filters. On the contrary, the
proposed adaptive filer can automatically estimate the suitable ENL parameter which will be used
to spatially smooth the approximation image in the wavelet domain without need for many filter
runs with different parameters.
Figure 9 shows a real spine lumbar MRI image obtained from [42], filtered with Enhanced Frost,
Enhanced Lee, and the proposed adaptive Kuan filter using the polynomial approximation. A
visual inspection of the results shows that all the filters succeed to reduce speckle noise and
result in a pleasant visual appearance, preserving edges and contours. However, compared to
our proposed solution, the enhanced Lee and Frost filters show over-blurred images.
FIGURE 9: Spine lumbar MRI image [42]. Top left: original (noisy) image. Top right: Enhanced Frost filtered
image. Bottom left: Enhanced Lee filtered image. Bottom right: result of filtering with our proposed solution
using the polynomial approximation. A 13×13 window size is used for all the results.

Figure 10 shows a real cardiac catheterization speckled image obtained from [43], filtered with
the enhanced versions of Frost and Lee and with the proposed Kuan filter (polynomial
approximation). It can be observed that all the filters succeed to reduce speckle noise. However,
the proposed adaptive Kuan filter outperforms the enhanced Frost and enhanced Lee filters in
edge preservation. In other words, the Kuan filter with the proposed polynomial approximation
yields visually better performance with the resulting image efficiently smoothed, while significantly
preserving image details.
FIGURE 10: Cardiac catheterization image [43]. Top left: original (noisy) image. Top right: Enhanced Frost
filtered image. Bottom left: Enhanced Lee filtered image. Bottom right: result of filtering with our proposed
solution using the polynomial approximation. A 13×13 window size is used for all the results.
4.2 Performance of the Hybrid Wavelet-Spatial Despeckling Filter
This section deals with implementing the proposed hybrid wavelet-spatial domain filer. First,
wavelet denoising with one decomposition level is applied on the 25 images of TID2013 database
[39], each corrupted by 5 different levels of Gaussian speckle noise, then the results are
compared to the same algorithm but using our enhanced Kuan filter to smooth the approximation
image before wavelet reconstruction. We also apply wavelet denoising with the number of
decomposition levels determined dynamically as proposed earlier, and compare the results to the
same algorithm but using our enhanced Kuan filter to smooth the approximation image before
wavelet reconstruction. Finally, the proposed enhanced Kuan filter is applied to the approximation
image in each reconstruction level. Table 7 presents the gains (average, minimum, maximum,
and standard deviation) with respect to the noisy image using different evaluation metrics (PSNR,
CoC and ENL). It can be noticed that a Kuan smoothing on the approximation image with one
decomposition level leads to an increase of 6.12 dB in average PSNR compared to 4.88 dB
without the approximation image smoothing.
It is important to note that Kuan smoothing on the last decomposition level does not result in any
significant enhancement. In fact, the smoothing strength of our adaptive Kuan filter depends on
the NLK_AD parameter determined dynamically, which is very high in the approximation image of
the last level, due to repetitive smoothing at every decomposition. Therefore, W ≈ 1 (in (5)) and
no filtering is performed in this case. Moreover, a Kuan smoothing filter on the approximation
image of each level before reconstruction, leads to an increase of 0.2 dB in average PSNR
compared to the case where approximation smoothing is performed only at the last
decomposition level, without blurring the denoised image, which is obviously seen from the CoC
metric.

Gaussian noise
Gains with respect to the noisy image
PSNR CoC ENL
AVG MIN MAX STD AVG MIN MAX STD AVG MIN MAX STD
1L without Kuan
smoothing
4.8808 3.0946 5.996 0.6681 0.0863 0.0143 0.0999 0.0121 21.2313 18.224 24.956 1.6711
1L with Kuan
smoothing
6.1168 4.8226 6.9996 0.5514 0.0836 0.014 0.1089 0.013715 41.3668 29.2124 49.224 2.8868
Dynamic without
Kuan smoothing
7.1225 5.826 7.9986 0.5538 0.1142 0.0543 0.971 0.079 60.0037 30.4054 65.2006 4.4478
Dynamic with Kuan
smoothing on the
last level
7.1228 5.8206 7.9991 0.5534 0.1153 0.0599 0.9174 0.0744 60.0042 30.4023 65.2505 4.4484
Dynamic with Kuan
smoothing on all
reconstructed levels
7.3229 5.9146 8.2906 0.556 0.1132 0.0443 0.9121 0.0742 64.7011 40.4054 72.3504 4.5485
TABLE 7: Evaluation metrics for wavelet-based denoising with and without Kuan smoothing. Sym1 wavelets
and BayesShrink threshold selection technique are used.
4.3 Performance Evaluation for the Proposed Complexity Reduction Algorithm
In this section, the proposed accelerated algorithm is tested with different scenarios. Figure 11
shows an example of the denoising results obtained using the Lena image corrupted by a
Gaussian speckle of variance 3100, Haar, db4, sym4, and bior6.8 wavelets, Soft and Hard
VisuShrink and BayesShrink.
The algorithm results in two wavelet decomposition levels. The first and second columns of
Figure 11 show the images obtained with one and two decomposition levels respectively. It can
be seen that two decomposition levels give better performance than one decomposition level as
expected. In addition, Universal Soft and Hard thresholding give more pleasant results than
BayesShrink.
The same work is done with different standard test images, different noise variances and
distributions, different thresholding methods, different wavelet types and families and using
different decomposition levels, with images from [38, 39]. From our intensive test scenarios, we
observed that a “best” denoising setup using wavelets cannot be generalized. In other words,
there is no one optimal threshold selection technique and wavelet type that always give the best
performance; performance depends on the wavelet thresholding algorithm, the wavelet family, the
noise variance, the number of decomposition levels and the noisy image itself.
In the purpose of studying the complexity gain and PSNR loss when we stop at the estimated
number of decomposition levels denoted L_opt_algo, we apply 10
4
times wavelet denoising on
the images from TID2013 database [39] corrupted by five different levels of multiplicative
Gaussian speckle noise, using BayesShrink threshold selection, ‘sym1’ wavelet type and different
number of decomposition levels.
Figure 12 shows the histogram of the difference between the number of decomposition levels that
yields the highest PSNR (i.e. to which we refer as “optimal”) and the number L_opt_algo
obtained with our proposed method. It can be observed that 49% of the time, our algorithm
successfully determines optimal number of levels whereas it performs 1 additional level for 11%
of the time, and one and two levels less than the optimal number for 23% and 17% of the time,
respectively. It is important to mention that these numbers could vary with a different dataset or
simulation setup, but they give a general overview of the behavior of the proposed algorithm. As a
result of erroneously estimating the optimal number of decomposition levels, some PSNR loss is
incurred.
Figures 13 and 14 respectively show the histogram and the cumulative density function (CDF) of
the PSNR loss due to the proposed algorithm. It can be observed that the loss does not exceed

0.04 dB and thus, can be considered as negligible. Therefore, the proposed acceleration
algorithm has a good performance in terms of robustness against PSNR loss.
Figure 13 shows that there exists only 2% of no loss (loss of 0 dB), whereas Figure 12 indicates
that 49% of the images should have no loss, which could seem contradictory; in fact, referring to
Figure 14, we notice that the PSNR loss is less than 0.01dB for 49% of the images, which can be
explained to be due to numerical rounding errors in computer simulations and therefore the
results in Figures 13 and 14 are not contradictory but rather consistent with those in Figure 12.
FIGURE 11: Denoising using wavelets. 1
st
row: original image (left), speckled image (right). 2
nd
, 3
rd
and 4
th
row: db4 wavelet denoised images using the Universal Threshold with soft thresholding, the Universal
Threshold with hard thresholding and with BayesShrink. Results obtained with one decomposition level are
on the left and those obtained using two decomposition levels are on the right.

FIGURE 12: Histogram of the difference between L_opt_algo and the optimal number of decomposition
levels.
FIGURE 13: Histogram of the PSNR loss when using the acceleration algorithm.
FIGURE 14: CDF of the PSNR loss when using the acceleration algorithm.
The advantage of the acceleration algorithm is the gain in complexity while keeping the PSNR
loss negligible. In the purpose of analyzing this gain, the CPU time usage is computed for a
number of ‘sym1’ wavelet decomposition levels followed by BayesShrink thresholding and the
corresponding wavelet reconstruction. Figure 15 shows the processor time usage (average of
100 runs) on a 2 GHz Intel Core 2 Duo CPU, for images from the TID2013 database [39],
corrupted by 5 levels of multiplicative Gaussian speckle noise. The figure shows that one could
save 0.12 microseconds per pixel by performing one decomposition level instead of two, and 0.10
microseconds per pixel for two levels instead of three. The incremental gain exponentially
decreases as the number of decomposition levels increases, which is expected due to the size of
the approximation image that is exponentially reduced (by a power of 2 in each dimension) for
each additional level.

FIGURE 15: CPU time usage according to the number of decomposition levels.
5. CONCLUSION
We have proposed a hybrid wavelet-spatial denoising algorithm based on two-stage processing.
An automatic parameter selection technique for the spatial Kuan filter is proposed in the first
stage, which is later used to spatially smooth the approximation image of the wavelet coefficients
in the second stage. An acceleration algorithm for wavelets computation is also proposed. It
consists in selecting a suitable number of wavelet decomposition levels yielding near-optimal
performance and eliminating the need for additional and unnecessary decomposition levels.
The procedures have been materialized on a set of different images under different conditions. In
order to evaluate the quality of the results, objective metrics were used besides subjective ones
identifying the method's capacity in reducing speckle noise. The obtained results quality was
highly encouraging, in terms of speckle reduction and image characteristics preservation, and
proved that the proposed acceleration algorithm is advantageous in complexity gain which
suggests that further research in this direction could be promising, in particular in the optimal
choice of the decomposition level where spatial smoothing can be performed on the
approximation component.
The evaluation of repetitive runs of the spatial optimized filter is of our interest, to verify whether it
allows for better noise suppression or reduction while critically preserving edges and texture.
Moreover, it is known that smoothing is usually performed in a pre-processing phase before edge
detection. In the proposed algorithm, edge detection was performed for the purpose of
smoothing. Therefore, we propose the use of our enhanced Kuan filter recursively for the sake of
edge detection. Finally, we aim at reinforcing the use of the proposed algorithm to enhance the
performance of the BM3D denoising algorithm, leaving this perspective an open discussion for
future considerations.
6. ACKNOWLEDGMENT
This work has been partly supported by a research grant from the Higher Center for Research at
the Holy Spirit University of Kaslik (USEK), Lebanon.
7. REFERENCES
[1] A. Isar, S. Moga and D. Isar, "A new denoising system for SONAR images," ACM Journal
on Image and Video Processing, January 2009.
[2] I. Shanthi and M.L. Valarmathi, "Speckle Noise Suppression of SAR Image using Hybrid
Order Statistics Filters," International Journal of Advanced Engineering Sciences and
Technologies, Vol. 5, No. 2, pp. 229-235, 2011.

[3] D. T. Kuan, "Adaptive Noise Smoothing Filter for Images with Signal Dependent Noise,"
IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. PAMI-7, No. 2, pp. 165-
177, 1985
[4] D.H.H. Santosh, V.N.L. Kumar, D.R. Ramesh and P.S.Prasad, ‘’Efficient Techniques for
Denoising of Speckle and Highly Corrupted Impulse Noise Images,’’ IEEE Trans.
Electronics Computer Technology (ICECT), Vol. 3, pp. 253-257, 2011.
[5] R. Ahirwar and A. Choubey, "A Novel Wavelet-based Denoising Method of SAR Image
Using Interscale Dependency," 2011 International Conference on Computational
Intelligence and Communication Systems, Gwalior, India, October 2011.
[6] J. S. Lee, "Digital image enhancement and noise filtering by using local statistics," IEEE
Trans. on Pattern Analysis and Machine Intelligence, vol. PAM1-2, 1980.
[7] J.S. Lee and I. Jurkevich, “Segmentation of SAR images,” IEEE Trans. on Geoscience and
Remote Sensing, Vol. 27, No. 6, pp. 674-680, November 1989.
[8] J.S. Lee, “Speckle analysis and smoothing of synthetic aperture radar images,” Computer
Graphics and Image Processing, Vol. 17, pp. 24-32, 1981.
[9] D. Kuan, A. Sawchuk, T. Strand and P. Chavel, “Adaptive restoration of images with
speckle,” IEEE Trans., Acoustic Speech, Signal Processing, Vol. ASSP-35, No. 3, pp. 373-
383, March 1987.
[10] A. Isar, S. Moga and D. Isar, "A new denoising system for SONAR images," ACM Journal
on Image and Video Processing, January 2009.
[11] O.V. Michailovich and A. Tannenbaum, "Despeckling of medical ultrasound images," IEEE
Trans. on Ultrasonics, Ferroelectrics and Frequency Control, pp. 64-78, January 2006.
[12] R. Ahirwar and A. Choubey, "A Novel Wavelet-based Denoising Method of SAR Image
Using Interscale Dependency," 2011 International Conference on Computational
Intelligence and Communication Systems, Gwalior, India, October 2011.
[13] DabovKostadin, “Image and video Restoration with nonlocal transform-domain filtering,”
Ph.D. dissertation, Image/video restoration and enhancement. Finland: Tampere University
of technology, 2010.
[14] A. Buades, B. Coll, and J. M. Morel, “A non-local algorithm for image denoising,” IEEE
Trans. Comp. Vision. Pattern Reco., Vol.2, pp. 60-65, June 2005.
[15] Gonzales. Rafael C. and Woods. Richard E., “Digital Image Processing,” Third Edition,
Pearson Prentice Hall, pp. 311-334, 346-356, 463-476, 488-509, 2008.
[16] C. Ju and C. Moloney, “An Edge-Enhanced Modified Lee filter for the Smoothing of SAR
Image Speckle Noise,” IEEE Workshop on Nonlinear Signal Image Processing, Mackinac
Island, MI, 1997.
[17] Himani Goyal, “Qualitative and Quantitative Evaluation of Image Denoising Techniques,”
International Journal of Computer Applications, Vol. 8, No. 14, pp. 31-34, October
2010.
[18] D.M. Smith, “Speckle reduction and segmentation of Synthetic Aperture Radar images,”
International Journal of Remote Sensing, Vol. 17, No. 11, pp. 2043-2057, 1996.

[19] L. V. Patil, K. K. Rane, A. P. Salvi, and S. S. Shitole, “Evaluation of Statistical Digital Image
Filters,” European Journal of Scientific Research, June 2011.
[20] S.S. Ganugapati and C.R. Moloney, “A ratio based edge detector for speckled images
based on maximum strength edge pruning,” Proceedings of ICIP- 95, Washington, D.C.,
Vol. 2, pp. 165-168, October 1995.
[21] A. Lopes, E. Nezry, R. Touzi and H. Laur., “Maximum a posteriori speckle filtering and first
order texture models in SAR images,” Proceeding of Geoscience and Remote Sensing
Symposium ‘90, College Park, MD, Vol. 3, pp. 2409-2412, May 1990.
[22] Hong Sern Tan, “Denoising of Noise Speckle in Radar Image,” B.E. Thesis, School of
Information Technology and Electrical Engineering, The University of Queensland, October
2001.
[23] L. Klaine, B. Vozel, and K. Chehdi. “Adaptive filtering of multiplicative noise by a new
differential method,” Proceeding of the Sixth International Symposium on Signal Processing
and Its Applications’ 2003, July 2003.
[24] S. Aybike and S. B. Hasan, “Techniques of speckle noise reduction in ultrasound images,”
Proceeding of the 19th Signal Processing and Communications Applications
Conference, pp. 375-378, 2011.
[25] R. Rangarajan, R. Venkataramanan and S. Shah, “Image Denoising Using Wavelets:
Wavelets & time frequency,” Report, December 16, 2002.
Available at: http://www-personal.engin.umich.edu/~rvenkata/551_code/Report.pdf
[26] Sani M. Isa, Ary Noviyanto and Aniati Murni Arymurthy, “Optimal Selection of Wavelet
Thresholding Algorithm for ECG Signal Denoising,” IEEE Trans. on Advanced Computer
Science and Information System, pp. 365-370, December 2011.
[27] I. Daubechies, “Ten lectures on wavelets,” The Conference Board of the Mathematical
Sciences, Series in Applied Mathematics, Vol. 61, pp. 194-202, 1994.
Sciences, Series in Applied Mathematics, Vol. 61, pp. 198-202 and 254-256, 1994.
Sciences, Series in Applied Mathematics, Vol. 61, pp. 271-280.
[30] S. Mallat, “A theory for multiresolution signal decomposition: The wavelet representation,”
IEEE Trans., Pattern Anal. Machine Intell., Vol. 11, pp. 674–693, July 1989.
[31] S. Grace Chang, Bin Yu and Martin Vetterli, “Adaptive Wavelet Thresholding for Image
Denoising and Compression,” IEEE Trans on image processing, Vol. 9, pp. 1532-1546,
September 2000.
[32] David L. Donoho, “De-Noising by Soft-Thresholding,” IEEE Trans. on information theory,
Vol. 41, No. 3, pp. 613-627, May 1995.
[33] Donoho, D.L., Johnstone, “Spatial adaptation by wavelet shrinkage,” Biometrika, 81, pp.
425-455, 1994.
[34] Shi Guangmin and Li Fudong, “Image Denoising with Optimized Subband Threshold,”
Proceeding of the Fifth International Conference on Computational Intelligence and
Multimedia Applications, 2003.

[35] Yunyi Yan, Baolong Guo, Zhanlong Yang and Xiang Fu, “Image Noise Removal via
Wavelet Transform and r/K-PSO,” Fourth International Conference on Natural
Computation, pp. 544-547, 2008.
[36] Aleksandra Pizurica and Wilfried Philips, “Estimating the Probability of the Presence of a
Signal of Interest in Multiresolution Single- and Multiband Image Denoising,” IEEE Trans.
on image processing, Vol. 15, No. 3, March 2006.
[37] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by sparse 3D
transform-domain collaborative filtering,” IEEE Trans. Image Process., Vol. 16(8), pp. 2080-
2095, August 2007.
[38] ImageProcessingPlace.com, Image Databases.
[39] TAMPERE IMAGE DATABASE 2013 TID2013, version 1.0.
[40] M. Misiti, Y. Misiti, G. Oppenheim, and J.M. Poggi, Matlab Wavelet Toolbox User’s guide
version 3, 2005.
[41] A. Achim, A. Bezerianos and P. Tsakalides, “Ultrasound Image Denoising ViaMaximum A
Posteriori Estimation Of Wavelet Coefficients,” Proceedings of the 23
rd
annual international
conference of the IEEE, Vol. 3, pp. 2553-2556, 2001.
[42] M.R 12369 R_11_1200, Bellevue Medical Center, Mansourieh, El Metn, Lebanon.
[43] Interventional Cardiology Service, C07-372, N.D. Maritime hospital, Byblos, Lebanon.

Accelerated Joint Image Despeckling Algorithm in the Wavelet and Spatial Domains

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to Accelerated Joint Image Despeckling Algorithm in the Wavelet and Spatial Domains (20)

Recently uploaded (20)

Accelerated Joint Image Despeckling Algorithm in the Wavelet and Spatial Domains