remote sensing

Article
Speckle Noise Reduction Technique for SAR Images
Using Statistical Characteristics of Speckle Noise
and Discrete Wavelet Transform
Hyunho Choi and Jechang Jeong *
Department of Electronics and Computer Engineering, Hanyang University, 222, Wangsimni-ro, Seongdong-gu,
Seoul 04763, Korea; [email protected]
* Correspondence: [email protected]; Tel.: +82-2-2220-0369




Received: 2 May 2019; Accepted: 14 May 2019; Published: 18 May 2019


Abstract: Synthetic aperture radar (SAR) images map Earth’s surface at high resolution, regardless of
the weather conditions or sunshine phenomena. Therefore, SAR images have applications in various
fields. Speckle noise, which has the characteristic of multiplicative noise, degrades the image quality of
SAR images, which causes information loss. This study proposes a speckle noise reduction algorithm
while using the speckle reducing anisotropic diffusion (SRAD) filter, discrete wavelet transform
(DWT), soft threshold, improved guided filter (IGF), and guided filter (GF), with the aim of removing
speckle noise. First, the SRAD filter is applied to the SAR images, and a logarithmic transform is
used to convert multiplicative noise in the resulting SRAD image into additive noise. A two-level
DWT is used to divide the resulting SRAD image into one low-frequency and six high-frequency
sub-band images. To remove the additive noise and preserve edge information, horizontal and
vertical sub-band images employ the soft threshold; the diagonal sub-band images employ the IGF;
while, the low- frequency sub-band image removes additive noise using the GF. The experiments
used both standard and real SAR images. The experimental results reveal that the proposed method,
in comparison to state-of-the art methods, obtains excellent speckle noise removal, while preserving
the edges and maintaining low computational complexity.

Keywords: synthetic aperture radar images; speckle reducing anisotropic diffusion; speckle noise;
discrete wavelet transform; improved guided filter

1. Introduction
Synthetic Aperture Radar (SAR) images employ active sensors that detect microwave radiation,
which has longer wavelength than visible light that is detected in passive sensors, such as the optical
sensor. Therefore, the surface of the Earth can be observed at high resolution, regardless of weather
conditions and sun phenomena [1]. The active sensor of SAR images is also used with satellites
or unmanned aerial vehicles (UAVs), as the development of the active sensor technology applied
to SAR images enabled high-resolution target detection and identification. SAR images are widely
used in applications of a variety of fields, such as the military, agricultural, weather forecasting, and
environmental analysis, etc. [2]. Due to the advantages of SAR images and their various applications,
research on the technology behind SAR images is being actively conducted around the world (image
enhancement [3–5], image classification [6,7], image segmentation [8,9], etc.).
In contrast with the optical sensor, the active sensor of the SAR is accompanied by speckle noise
that arises from the coherent imaging mechanism. Speckle noise in SAR images is generated by the
random interference of many elementary reflectors within one resolution cell [10]. This noise has
different features from the noise observed in images that were obtained by passive sensors, such as the

Remote Sens. 2019, 11, 1184; doi:10.3390/rs11101184 www.mdpi.com/journal/remotesensing

Remote Sens. 2019, 11, 1184 2 of 27

optical sensor. Speckle noise appears as a form of multiplicative noise in SAR images and it has the
characteristics of a Rayleigh distribution [11]. SAR images are used by observers to extract information
and identify targets. Speckle noise degrades SAR images and thus interferes with the transfer of image
information to the observer. Therefore, the development of effective filtering methods in the reduction
of speckle noise is critical for the analysis of information that is contained in various SAR images.
Numerous studies have been conducted with the aim of extracting image information from SAR
images by removing speckle noise. Five main categories of methods were applied in these studies:
linear filtering, nonlinear filtering, partial differential equation (PDE) filtering, hybrid methods, and
filtering methods that are based on the discrete wavelet transform (DWT).
The linear filter convolutes an image with a symmetric mask and then reconstructs each pixel
value as a weighted average value of the neighboring pixel values. The mean filter and the Gaussian
filter are typical linear filtering techniques that are effective in simple and smoothing speckle noise
reduction. A mean value of several pixel values around the target pixel substitute the mean filter.
The target pixel is located in the center. The mean filter exhibits a low-edge preservation performance,
because it does not consider the flat and homogeneous areas in the image [12]. The Gaussian filter uses
a Gaussian function of two dimensions (2D) as a convolution mask. This filtering technique uses the
Gaussian function as the mask weight value. The Gaussian function gives a larger weight to the center
of the mask. Moreover, the Gaussian filtering technique shows excellent performance in terms of noise
removal with a small variance; however, a blurring phenomenon appears in the edge areas.
The nonlinear filter extracts the edge regions in the image while using the statistical values (e.g.,
mean, median, standard deviation (SD), etc.) in the mask. As a median filter, a bilateral filter (BF) [13],
and a non-local mean (NLM) filter [14], these filtering techniques preserve the edge areas and remove
the noise in the image. The median filter removes noise by replacing the pixel value in the mask with a
median value (order statistics). Therefore, this median filtering scheme exhibits excellent noise removal
performance in the homogenous regions; however, it has low edge preservation performance at the
edges. The BF performs filtering by employing a Gaussian filter coefficient, when considering the
distance between the center and the neighboring pixel, as well as the pixel value difference. At the
edges, where the difference between the center pixel and the neighboring pixel is large, a small filter
coefficient value is employed. In the homogeneous regions, Gaussian filtering is applied to remove
noise, which has a large filter coefficient. The BF can preserve edge information while reducing
noise by the use of these two filter coefficients. Hence, the noise is removed following the same
principles of Gaussian filtering, whereas speckle noise has a Rayleigh distribution. Consequently, this
method exhibits a low speckle noise reduction performance. The NLM filter is a filtering method that
improves the performance based on the BF [15]. The BF assigns weights by determining the similarity
between the pixel units, whereas the NLM filter was extended to allocate weights that are based on the
similarity of the mask. This attribute represents the biggest difference between the BF and the NLM
filter, and it contributes to the noise reduction performance. However, it is suitable to be reduced by
the additive white Gaussian noise (AWGN) [16]. Therefore, the NLM filter has a low speckle noise
reduction performance.
The core idea of the PDE technique is to treat image processing as a discrete processing and
not a continuous process. The PDE method converts a noisy image into a form of PDEs to obtain
a noise-free image while using PDEs [17]. Filtering methods that are based on the PDE, such as
the anisotropic diffusion (AD) filter [18] and the adaptive window diffusion (AWAD) method [19],
have been proposed as other noise removal techniques. The AD filtering method employs a gradient
operator to identify the gradient changes in the image that are caused by noise and the edge effect.
Nearest-neighbor weighted averaging removed small gradient changes caused by noise, while large
gradient changes that are caused by edges are preserved [3]. This technique of the AD filter obtains
satisfactory results with respect to smoothing additive noise in the image. However, it exhibits a low
noise reduction performance the speckle noise (multiplicative noise), because the gradient operator
of the AD filter cannot classify the noise and edges in the speckle imagery. The size and direction of
Remote Sens. 2019, 11, 1184 3 of 27

the mask are adjusted according to the structure of the image, and an adaptive window anisotropic
diffusion (AWAD) is proposed. The AWAD method can control the mask size and direction, which
thereby shows an excellent edge preservation performance. However, the AWAD algorithm that
is based on a new diffusion function has low speckle noise removal performance, because the new
diffusion function does not perform speckle noise removal in homogeneous regions.
The abovementioned methods can be used in various combinations (hybrid filter) to enhance
the speckle noise removal and the edge preservation performance of each filtering method.
Deledalle et al. [20] proposed a probabilistic patch-based (PPB) algorithm based weight in the NLM
filter. The PPB algorithm uses a statistically grounded similarity criterion, which depends on the
noise distribution model, to remove the speckle noise in the SAR image. In the SAR images, the
PPB method obtains an excellent performance with respect to speckle removal and preservation of
edges; however, the PPB algorithm that was employed by the NLM filter has very high computational
complexity [21]. A 2S-PPB [22] is the extended method that is based on the PPB algorithm. The 2S-PPB
method for removing speckle noise in SAR images exploits a two-step strategy: (1) a non-local weighted
estimation analyzes the redundancy in time; and, (2) the non-local estimation in space is used in the
second step. The 2S-PPB algorithm can efficiently remove the speckle noise, but it exhibits widespread
artifacts, because it has a limitation of spatiotemporal similarity, such as watercolor strokes around
the edge regions [23,24]. The SAR-BM3D [24] algorithm, which maintains edge information while
smoothing homogeneous regions, uses a non-local filtering method and wavelet shrinkage in the
three-dimensional (3D) domain. The undecimated wavelet transform combined with the local linear
least minimum mean square error (LLMMSE), which uses the estimation standard to determine
shrinkage wavelet coefficients, is employed to evaluate the sparse coefficients. The SAR-BM3D method
has an excellent speckle noise removal performance, but it exhibits heavy computational complexity
because of the NLM filter. Therefore, this algorithm cannot be employed in a real-time system.
The DWT method can analyze the signal localization using both time and frequency, unlike
spatial filters that use only the size and orientation of the local mask. Since the 1990s, DWT has been
widely used in various image processing fields, including blocking-artifact reduction [25], image
fusion [26–28], and object detection [29–31], because of these advantages. In particular, the introduction
of DWT offered a new method to reduce speckle noise in the transform domain. The DWT becomes
one of the most researched methods for speckle noise reduction in SAR images, because of valuable
attributes, like time-frequency localization and multiresolution. Most of the methods applying the
DWT for general speckle noise removal proceed as follows. First, a decomposition of the image using
the DWT is performed. Subsequently, the wavelet function is applied to reduce the unnecessary
wavelet coefficients in the wavelet domain, such as noise. Classical threshold methods, such as hard
threshold and soft threshold [32], universal threshold [33], Stein’s unbiased risk estimate (SURE)
threshold [34], and Bayes shrink threshold [35], were developed and modified for each image to remove
the unnecessary wavelet coefficients in the wavelet domain. Finally, the processed wavelet coefficients
by the inverse DWT synthesize the noise-free image. A larger number of despeckling methods based
on the transform domain [36–39] have been studied. With the aim of reducing speckle noise in SAR
images, Yang et al. [35] proposed an adaptive speckle noise algorithm based on an improved wavelet
threshold. The improved wavelet threshold method is applied in high-frequency sub-band images
of the wavelet domain. After the speckle noise removal process, a forward and backward (FAB)
method is used to remove the residual noise from the image; however, the algorithm shows low
speckle noise suppression ability due to the low ability of choosing the optimal parameters for FAB.
Amini et al. [36] proposed a de-speckling method that is based on the expectation maximization (EM)
algorithm and the DWT. The EM method estimates the noise coefficients in each sub-band image using
the hidden Markov model (HMM) parameters. The de-speckling algorithm represents the artifacts
around the edge areas when the HMM parameter estimation shows incorrect values. Li et al. [37]
proposed a Bayesian multiscale method for de-speckling the SAR images in a non-homomorphic
framework. The linear decomposition method was used to treat the speckle noise (multiplicative noise)
Remote Sens. 2019, 11, 1184 4 of 27

in the non-homomorphic framework. Subsequently, a two-sided generalized Gamma distribution was

used in an earlier step to process the heavy-tailed nature of the wavelet coefficients of the noise-free
reflectivity. Based on this, the maximum a-posteriori (MAP) method was used in an analytical wavelet
shrinkage function. The MAP method employed a heterogeneity-adaptive threshold to select the best
estimates of the noise-free wavelet coefficients. Controlling a tunable parameter is difficult; hence,
this parameter affects the optimal heterogeneity-adaptive weight function. As a result, the algorithm
exhibits a blurring phenomenon around the edge areas in the actual SAR images. Rajesh et al. [38]
presented a combination of a spatial filter as a preprocessing step and adaptive threshold in the
frequency domain. The algorithm employs a Wiener filter, among other spatial filters, and an adaptive
soft threshold of wavelet coefficients in the wavelet domain. The Wiener filter shows excellent additive
noise removal performance in the image [40]. The algorithm did not consider this ability of the Wiener
filter, which results in a low speckle suppression performance. Despite extensive efforts, as mentioned
above, conventional algorithms exhibit low performance in terms of speckle noise removal, edge
information preservation, and computational complexity.
In this study, we employ a speckle reducing anisotropic diffusion (SRAD) filtering method as
a preprocessing filter to reduce the speckle noise and preserve the edge information. The SRAD
filtering result image is applied to a logarithmic transform for the conversion of multiplicative noise
to an additive noise. The soft threshold, guided filter (GF), and improved guided filter (IGF) in the
wavelet domain are employed to further reduce the additive noise in the SRAD filtering result image.
A diagonal sub-band image in the wavelet domain has a lower energy than the vertical and horizontal
sub-band images. Hence, the IGF applied as a new edge-aware weighting method is employed in the
diagonal sub-band image to preserve weak edges and remove the noise. To the same end, the soft
threshold is applied to the vertical and horizontal sub-band images. Moreover, the GF removed the
noise in the approximate sub-band image. Finally, a noise-free image is obtained by an exponential
transform and wavelet reconstruction. The proposed algorithm is implemented to remove the speckle
noise, preserve the edges, and reduce computational complexity.
This paper is organized, as follows. Section 2 describes the evaluation metrics and the proposed
algorithm in detail. In Section 3, simulated and real SAR images are used to analyze the experimental
results of qualitative, quantitative, and computational complexity. Section 4 presents a discussion.
Section 5 concludes the paper.

2. Proposed Algorithm
In this study, we propose an algorithm for the reduction of speckle noise and the preservation
of the edges in the SAR image (Figure 1). The proposed algorithm employs the SRAD filtering
method as a preprocessing filter instead of directly applying the wavelet domain, as the SRAD can
be directly applied to the SAR image, because it uses the image without log-compressed data [39].
However, the SRAD filtering result image still includes the speckle noise, which represents a form
of multiplicative noise. Since most of the filtering methods are developed for reducing the AGWN,
the logarithmic transform is applied to the resulting SRAD image to convert the multiplicative noise
into additive noise [41], after which the resulting SRAD image contains additive noise. Subsequently,
the two-dimensional (2D) DWT transforms the SRAD filtering result image, which represents the
logarithmic transform, into four sub-band images (vertical sub-band image (LH), horizontal sub-band
image (HL), diagonal sub-band image (HH), and approximate sub-band image (LL)). We employed the
DWT performed until two-level decomposition. An effect of algorithm and analysis of results is tested
at one to two decomposition level. The two-level decomposition of the DWT shows the best results [42].
Most of the speckle noise occurs in high-frequency sub-band images [43]. Therefore, the soft threshold
of the wavelet coefficients is only applied to the horizontal and vertical sub-band images, which have
similar energy, to preserve the original signal and remove the noise signal. However, the diagonal
sub-band image has a low energy when compared to the vertical and horizontal sub-band images. For
the diagonal sub-band image, we employ an IGF that is based on a new edge-aware weighting method
Remote Sens. 2019, 11, 1184 5 of 27

to preserve
Remote a low
Sens. 2019, 11, xoriginal
FOR PEER signal and suppress the noise signal using this new edge-aware weighting
REVIEW 5 of 26
method. The approximate sub-band image contains significant components of the image and is less
of the image
affected by the and
noiseis less
[44];affected
however,by the noise exists
[44]; however, the noise exists
in the approximate in theimage.
sub-band approximate
The GF sub-
[45]
band
is image. to
employed The GF [45]
reduce is employed
the speckle noise to
andreduce thethe
preserve speckle
edgesnoise
in theand preserve the
approximate edges image.
sub-band in the
approximate
Each sub-band sub-band
image, onceimage. theEach sub-band
noise image,
is removed, is once the noise by
reconstructed is removed, is reconstructedand
wavelet reconstruction, by
wavelet
the reconstruction,
exponential transformand the exponential
is performed transform
to reverse is performed
the logarithmic to reverse
transform. Finally,the
welogarithmic
obtain the
transform. Finally,
despeckled image. we obtain the despeckled image.

Figure 1. Block diagram of the proposed algorithm.

2.1. Speckle Reducing

2.1. Speckle Reducing Anisotropic
Anisotropic Diffusion
Diffusion
As
As mentioned
mentioned above,
above, the the ADAD [18]
[18] performs
performs poorly
poorly inin terms
terms ofof edge
edge preservation
preservation inin the
the presence
presence
of
of speckle
speckle noise. It removes
noise. It removes the the additive
additive noise
noise from
from the
the image, thereby causing
image, thereby causing aa loss
loss of
of detailed
detailed
information.
information. The SRAD method modifies the AD filter to improve the edge detection accuracy in
The SRAD method modifies the AD filter to improve the edge detection accuracy in
speckled
speckled images. The instantaneous
images. The instantaneous coefficient
coefficient of
of variation
variation (ICOV)
(ICOV) isis merged
merged into
into the
the edge
edge detector.
detector.
The
The method
method for
for removing
removing specklespeckle noise
noise using
using ICOV
ICOV is is described
described below.
below.
The output image I ( x, y; t ) is obtained by a
The output image 𝐼(𝑥, 𝑦; 𝑡) is obtained by a PDE model PDE model of SRAD when an
of SRAD intensity
when image I (x,
an intensity y; 0)
image
has
𝐼(𝑥, finite
𝑦; 0) power andpower
has finite no zero and values
no zeroover the image
values of the
over the 2D coordinate
image grid Ω (Equation
of the 2D coordinate grid 𝛺 (1)).
(Equation
(1)). 

 ∂I (x, y; t)/∂t = div[c(q)∇I (x, y; t)]
𝜕𝐼(𝑥, 𝑦; 𝑡)/𝜕𝑡 = 𝑑𝑖𝑣[𝑐(𝑞)∇𝐼(𝑥, 𝑦; 𝑡)]

(1)
 
∂I (x,y;t)

I ( x, y; 0 ) = I ( x, y ) , =0

∂Ω
 0
𝜕𝐼(𝑥, ∂ n𝑦; 𝑡)
→
(1)


𝐼(𝑥, 𝑦; 0) = 𝐼 (𝑥, 𝑦), ( )| = 0
𝜕𝑛⃗
where I0 is the initial noisy image, t is time, and div is the divergence operator. ∇ is the gradient
where 𝐼 is the initial noisy image, →𝑡 is time, and 𝑑𝑖𝑣 is the divergence operator. ∇ is the gradient
operator, ∂Ω is the boundary of Ω, n is the unit vector, and I (x, y; t) is the output image. c(q) is the
operator, 𝜕𝛺 is the boundary of 𝛺, 𝑛⃗ is the unit vector, and 𝐼(𝑥, 𝑦; 𝑡) is the output image. 𝑐(𝑞) is
diffusion coefficient.
the diffusion coefficient.
The diffusion coefficient c(q) plays a crucial role in SRAD by determining the diffusion scale.
The diffusion coefficient 𝑐(𝑞) plays a crucial role in SRAD by determining the diffusion scale.
It encourages diffusion in the homogeneous regions and restriction near the edges of the image.
It encourages diffusion in the homogeneous regions and restriction near the edges of the image. The
The formula of the diffusion coefficient can be alternatively expressed as:
formula of the diffusion coefficient can be alternatively expressed as:
11
(q) =
c𝑐(𝑞) = 1 + [ f 2 (x, y; t) − f 2 (t)] /T ] (2)
(2)
1 + [𝑓 (𝑥, 𝑦; 𝑡) − 0𝑓 (𝑡)]/𝑇]
or
or h i
f 2 (x, y; t) − f02 (t)
c(q) = exp {− [𝑓 (𝑥, 𝑦; 𝑡) − 𝑓 (𝑡)]} (3)
𝑐(𝑞) = exp{− T } (3)
𝑇
Here, the ICOV 𝑓(𝑥, 𝑦; 𝑡) can detect edges of the images and speckle noise. The ICOV exhibits high
values at edge regions and low values in the homogeneous regions. It can be estimated using the
following equation:
Remote Sens. 2019, 11, 1184 6 of 27

Here, the ICOV f (x, y; t) can detect edges of the images and speckle noise. The ICOV exhibits high
values at edge regions and low values in the homogeneous regions. It can be estimated using the
following equation:
v
u
t 1 ∇I 2 − 1 ∇2 I 2
u
u
u     
2 I 42 I
f (x, y; t) = h   2 i2 (4)
1 ∇ I
1+ 4 I

where ∇2 represents the Laplace operator, f0 is the coefficient of variation at the time, and T is the
threshold of the diffusion coefficient (Equations (2) and (3)). The value of c(q) tends to zero when
f 2 (x, y; t) − f02 (t) is greater than T, as the diffusion stops. In the opposite case, the value of c(q)
approaches 1 when f 2 (x, y; t) − f02 (t) is less than T, thus the diffusion is applied as the filter in the
homogeneous regions. The threshold value of the diffusion coefficient has an effect on the reduction of
speckle noise and in the preservation of edge information.
 
T = f02 (t) 1 + f02 (t) (5)

where p
var[z(t)]
f0 (t) = (6)
z(t)

Here, z(t) and var[z(t)] are the intensity mean and variance, respectively, over a homogeneous region
at t. The f0 (t) of an automatic determination can be estimated, as follows:

f0 (t) ≈ f0 exp(−ρt) (7)

where ρ is a constant and f0 is the coefficient of variation in the observed image. The SRAD
filtering method can directly process the data and preserve important information in the image
without performing log-compression [39]. Therefore, the SRAD filtering technique can be used as a
preprocessing filter.

2.2. Logarithmic Transformation

Equation (8) shows the model degraded by the speckle noise in the SAR images [46]. The resultant
image is the product of the speckle noise and original image.

R(x, y) = O(x, y) × M(x, y) + A(x, y) (8)

where R(x, y) is degraded image of the SAR image. O(x, y) is the original image, M(x, y) is the speckle
noise, and A(x, y) is the additive noise. Since the additive noise affects the SAR images less than
multiplicative noise, it is ignored, and Equation (9) is obtained.

R(x, y) = O(x, y) × M(x, y) (9)

If a model of the multiplicative noise represents the speckle noise, as in Equation (9), it is difficult
to separate the original image and the noise component. When a logarithmic transform is applied to
SAR images containing the multiplicative noise (speckle noise), the speckle noise appears in the form
of additive noise, as follows:
 
logR(x, y) = log[O(x, y) × M(x, y)] = log O(x, y) + log M(x, y) = L(x, y) + S(x, y) (10)

F(x, y) = L(x, y) + S(x, y) (11)

where F(x, y), L(x, y) and S(x, y) are the logarithms of R(x, y), O(x, y), and M(x, y), respectively.
F(x, y) represents the characteristics of an AWGN with an average of 0 and a variance of σ2 . In this
 images.
For the 2D image, the basic idea of the DWT is described, as follows. One-level DWT transforms
the SAR images with speckle noise into four sub-band images: the approximate sub-band image (𝐿𝐿 )
and three detailed sub-band images (vertical coefficients (𝐿𝐻 ), horizontal coefficients (𝐻𝐿 ), and
diagonal coefficients ( 𝐻𝐻 )) (Figure 2b). Figure 2c shows the results of the two-level wavelet
Remote Sens. 2019, 11, 1184 7 of 27
decomposition. The two-level DWT decomposes the 𝐿𝐿 sub-band image that was obtained from
the one-level wavelet decomposition in the same manner to obtain four sub-band images (𝐿𝐿 , 𝐿𝐻 ,
𝐻𝐿 , we
study, anduse𝐻𝐻a ). The approximate
logarithmic sub-band
transform image
to convert the(𝐿𝐿multiplicative
) contains thenoise
low-frequency
into AWGN coefficients,
and attemptandto
detailed sub-band images ( 𝐿𝐻 , 𝐻𝐿
additionally remove the noise in the wavelet domain., 𝐻𝐻 𝐿𝐻 , 𝐻𝐿 , and 𝐻𝐻 ) depict the high-frequency
coefficients. The detailed sub-band images present most information regarding the image, including
theDiscrete
2.3. noise and edgeTransform
Wavelet information. The approximate sub-band image includes important information
about the SAR images, such as the texture.
The DWT is employed to remove noise in the various high- and low-frequency coefficients of
SAR2.4.images. It analyzes multiresolution sub-band images by adjusting the scaling and translation
Soft Threshold
parameters; hence, as the scaling parameter increases, extending the signal lowers the spatial resolution.
Various threshold methods exist ([32–34]). The most commonly used wavelet functions are the
The extended scaling parameter can obtain a sub-band image representing low-frequency coefficients.
soft and hard threshold. These threshold methods are used to reduce the speckle noise in the SAR
In the opposite case, a high-frequency sub-band image can be obtained. The translation parameter
images. Although both thresholds set to zero when the coefficients are smaller than the threshold,
moves along the time axis. As this parameter value increases, it moves to the right. With these two
these thresholds have the main difference. The former function suppresses the coefficients that are
parameters, the DWT can obtain an approximate sub-band image and detailed sub-band images.
larger than the threshold, while the latter function leaves them unchanged [43].
For the 2D image, the basic idea of the DWT is described, as follows. One-level DWT transforms
The hard threshold removes the coefficients that are below the threshold value 𝑇, as determined
the SAR images with speckle noise into four sub-band images: the approximate sub-band image
by the noise variance. The hard threshold is depicted, as follows:
(LL1 ) and three detailed sub-band images (vertical coefficients (LH1 ), horizontal coefficients (HL1 ),
𝑤, |𝑤| > 𝑇
and diagonal coefficients (HH1 )) (Figure 𝑊 2b).= Figure 2c shows the results of the two-level wavelet (12)
0, |𝑤| ≤ 𝑇
decomposition. The two-level DWT decomposes the LL1 sub-band image that was obtained from
thewhere 𝑤 iswavelet
one-level the wavelet coefficient in
decomposition the𝑇same
and is themanner
threshold value.four
to obtain 𝑊 sub-band
represents the (LL
images wavelet
2 , LH2 ,
coefficient after the hard threshold is applied. The hard threshold is known to
HL2 , and HH2 ). The approximate sub-band image (LL2 ) contains the low-frequency coefficients, and have discontinuity in
the noise-free image, since the wavelet coefficient at the threshold is suddenly zeroed.
detailed sub-band images (LH1 , HL1 , HH1 LH2 , HL2 , and HH2 ) depict the high-frequency coefficients. In the hard
Thethreshold
detailedmethod,
sub-band theimages
wavelet coefficients
present that do not exceed
most information the given
regarding threshold
the image, valuethe
including are noise
zeroed. and
The other wavelet coefficients remain unchanged. Therefore, the hard
edge information. The approximate sub-band image includes important information about thethreshold yields artifacts in the
SAR
despeckled image [47]. The soft threshold applies the signum function in its model to overcome these
images, such as the texture.
issues of the hard threshold (Equation (13)).

(a) (b) (c)

Figure
Figure 2. 2.Two-dimensional
Two-dimensional(2D)
(2D)image
image decomposition
decomposition result
result by
bydifferent
differentdiscrete
discretewavelet
wavelettransform
transform
(DWT)levels.
(DWT) levels. (a)
(a)Image
Imageof of
Napoli; (b) One-level
Napoli; wavelet
(b) One-level decomposition;
wavelet and, (c) Two-level
decomposition; wavelet
and, (c) Two-level
decomposition.
wavelet decomposition.

2.4. Soft Threshold

Various threshold methods exist ([32–34]). The most commonly used wavelet functions are the
soft and hard threshold. These threshold methods are used to reduce the speckle noise in the SAR
images. Although both thresholds set to zero when the coefficients are smaller than the threshold,
these thresholds have the main difference. The former function suppresses the coefficients that are
larger than the threshold, while the latter function leaves them unchanged [43].
The hard threshold removes the coefficients that are below the threshold value T, as determined
by the noise variance. The hard threshold is depicted, as follows:

w, |w| > T
(
Whard = (12)
0, |w| ≤ T
Remote Sens. 2019, 11, 1184 8 of 27

where w is the wavelet coefficient and T is the threshold value. Whard represents the wavelet coefficient
after the hard threshold is applied. The hard threshold is known to have discontinuity in the noise-free
image, since the wavelet coefficient at the threshold is suddenly zeroed. In the hard threshold
method, the wavelet coefficients that do not exceed the given threshold value are zeroed. The other
wavelet coefficients remain unchanged. Therefore, the hard threshold yields artifacts in the despeckled
image [47]. The soft threshold applies the signum function in its model to overcome these issues of the
hard threshold (Equation (13)).

sgn(w)(|w| − T ), |w| > T

(
Wso f t = (13)
0, |w| ≤ T

Here, sgn depicts the signum function. Wso f t is the wavelet coefficient after the shrinkage of the
soft threshold.
In the soft threshold method, the wavelet coefficients are zero if they are below the threshold.
The wavelet coefficients above the threshold are shrunk by the threshold value. Hence, the soft
threshold provides smooth results without artifacts. When compared to the hard threshold, the
soft threshold generally exhibits excellent preservation of detail at the expense of computational
complexity [43]. We apply the soft threshold to these sub-band images, since the horizontal and vertical
sub-band images have a similar energy [48].

2.5. Guided Filter

Zhang and Gunturk [47] mentioned that noise may exist in the approximate sub-band image
and detailed sub-band images in the wavelet domain. Gao et al. [43] have divided the 2D SAR
images into low, medium, and high-frequency sub-band images using a 2D fast Fourier transform
(FFT). The authors applied the approximate sub-band image through low-pass filtering to reduce the
noise in the approximate sub-band image. This method shows low speckle noise removal and edge
preservation ability. In [47,49], the authors applied the BF [13] to the approximate sub-band image to
suppress speckle noise. BF is capable of preserving the edges and shows an excellent noise removal
performance since it is difficult to distinguish between original signal and noise in the approximate
sub-band image; however, it exhibits gradient distortion and high complexity [47]. We applied the
GF [45] in the approximate sub-band image to overcome these problems. The process of removing
the noise using the GF is as follows. GF models the output image qi for the guidance image Ik of the
window ωk region with the center pixel k in the image, as follows:

qi = ak Ik + bk , ∀i ∈ ωk (14)

where ak and bk are linear coefficients estimated form the window ωk . Equation (15) removes unwanted
texture or noise to determine the linear coefficients.

qi = pi − ni (15)

Here, pi and ni denote the input image and noise component, respectively. The linear coefficients
are obtained by Equation (16) to minimize the difference between the input image pi and the output
image qi . X
(ak Ik + bk − pi )2 + εa2k

E(ak , bk ) = (16)
i∈ωk

where ε is a normalization parameter that serves to prevent ak from becoming infinitely large.
The minimization method of the liner coefficient in Equation (16) is as follows:

1
− µk pk
P
|ω| i∈ωk Ii pi
ak = (17)
σ2k +ε
Remote Sens. 2019, 11, 1184 9 of 27

bk = pk − ak µk (18)

Here, µk and σ2k are the mean and variance of the guidance image in the window ωk . |ω| represents
the number of pixels in the mask ωk , and pk = |ω|
1
pi . As the window size ωk and ε adjust, the noise is
removed and the edge areas are preserved. Therefore, these parameters are adjusted according to the
characteristics of the approximate sub-band image to remove the additive noise and to preserve edge
information.

2.6. Improved Guided Filter

2.6.1. A New Edge-Aware Weighting

The horizontal and vertical sub-band images of DWT have the same energy, while the diagonal
sub-band image has lower energy at the same scale [48]. We propose new edge-aware weighting for
effectively detecting and preserving weak edge information (Equation (19)). The gradient operator is
an effective operator for detecting sharp edge regions in the image and protecting against unnecessary
blurring phenomena around the edges. However, the gradient operator produces wide and blurred
edges when the edge regions are not sharp. In contrast, since a Laplacian operator uses a second-order
derivative operator that has a zero crossing level, it can detect the weak information of the edges while
using the zero crossing level. Therefore, we can detect weak edges in the diagonal sub-band image in
the wavelet domain.
1 + ||∆L|| 2
 
h= (19)
1 + ||∇L||
Here, ∆ and ∇ are the Laplacian and Gradient operators, respectively. The value of h is larger than 1
when h is located at weak edges; however, it is smaller than 1 if h is in the homogeneous regions.

2.6.2. The Proposed Filter

The new edge-aware weighting h of Equation (19) is incorporated into the cost function E(ak , bk )
of Equation (20). As mentioned above, an IGF obtains a solution that minimizes the input image pi and
the output image qi , while maintaining the linear model of Equation (14). A cost function with applied
new edge-ware weighting is expressed. as follows:
X ε

E(ak0 , bk0 ) = (ak0 Ik0 + bk0 − pi )2 + a2k0 (20)
h
i∈ωk

The optimal values of ak0 and bk0 are computed as:

1
− µk0 pk0
P
|ω| i∈ωk0 Ii pi
ak 0 = ε
(21)
σ2k0 + h

bk0 = pk0 − ak0 µk0 (22)

The final value of q̂i is given as follows:

q̂i = ak Ik + bk (23)

Here, ak and bk are the mean values of ak0 and bk0 within the window, respectively.

2.7. Evaluation Metrics

We used peak signal-to-noise (PSNR), structural similarity (SSIM), and equivalent number of
looks (ENL) to compare the performance of speckle noise reduction in the SAR images [41]. The PSNR
Remote Sens. 2019, 11, 1184 10 of 27

depicts the maximum signal-to-noise ratio. The PSNR is an objective measurement method that is
used to evaluate image quality, and it is defined as follows:
!
255
PSNR = 20log10 √ (24)
MSE

where the mean square error (MSE) is given by:

M− 1 N – 1
1 X X 2
MSE = Y(x, y) − Z(x, y) (25)
MN
x=0y=0

where M and N represent the number of pixels in the vertical and horizontal directions of the image,
respectively. Y(x, y) is the pixel value at the position of the original image (x, y) and Z(x, y) is the
pixel value at the coordinates of (x, y) in the filtered image. The filtered image Z(x, y) has a smaller
MSE as the image approaches the original image Y(x, y). Larger PSNR values imply better noise
reduction performances. The SSIM is an index that indicates the similarity between the original image
Y(x, y) and the filtered image Z(x, y). The SSIM is given, as follows:
  
2µx µ y + c1 2covxy + c2
SSIM(x, y) =    (26)
µ2x + µ2y + c1 σ2x + σ2y + c2

Here, µx and µx are the mean values of x and y, respectively. σ2x and σ2y present the variance of x and
y, respectively. covxy is the covariance of x and y. c1 and c2 are the two variables used to stabilize
the division that can occur with a weak denominator. When the value of SSIM is closer to 1, there is
no difference between the original and the filtered image. The equivalent number of look (ENL) is
used to evaluate the speckle noise reduction performance of the homogeneous regions in the image.
It is a standard metric in the absence of reference images that is widely used to evaluate despeckling
performance. The ENL is defined as:
µ2z
ENL = 2 (27)
σz
where µz and σz are the estimated mean and standard deviation of the filtered SAR image. Larger ENL
values indicate excellent speckle noise removal ability.

3. Experimental Results

3.1. Experiments on Standard Images

In this study, we selected 8-bit gray standard images (Airplane, Baboon, Barbara, Boat, Cameraman,
Fruits, Hill, House, Lena, Man, Monarch, Napoli, Peppers, and Zelda) with 256 × 256, 512 × 512, and 748
× 512 pixels in order to evaluate the performance of speckle noise removal and edge preservation
(Figure 3). Speckle noise (σ = 0.04) was added to each image. The existing methods (NLM [14],
Guided [45], Frost [50], Lee [51], Bitonic [52], weighted-least-squares (WLS) [53], non-local low-rank
(NLLR) [54], anisotropic diffusion filter with memory based on speckle statistics (ADMSS) [55],
SRAD [39], SRAD-Guided [56], SAR-BM3D [24]) and the proposed algorithm were used to compare
the speckle noise suppression performance. Tables 1 and 2 present the simulation conditions for the
standard images. The optimal parameters of the SRAD-guided method are the same as those of the
proposed algorithm (Table 2). The MATLAB 2018b software was used in a computer environment
[Intel (R) Core (TM) i5-8500 CPU @ 3.0 GHz with 16 GB RAM].
 Time step = 0.01 Mask size = 3 × 3
Mask size = 3 × 3
Peppers Exponential decay rate = 1 Regularization parameter =
Regularization parameter = 0.001
Number of iterations = 120 1e−10
Time step = 0.01 Mask size = 3 × 3
Mask size = 3 × 3
Zelda Exponential decay rate = 1 Regularization parameter =
Regularization parameter = 0.001
Remote Sens. 2019, 11,Number
1184 of iterations = 140 1e−10 11 of 27

Remote Sens. 2019, 11, x FOR PEER REVIEW 12 of 26

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n)

Figure
Figure 3. Images
3. Images usedused in the
in the experiments.(a)
experiments. (a) Airplane
Airplane (512
(512××512);
512);(b)(b)
Baboon
Baboon(512(512
× 512); (c) Barbara
× 512); (c) Barbara
(512 × 512); (d) Boat (512 × 512); (e) Cameraman (256 × 256); (f) Fruits (512 × 512);
(512 × 512); (d) Boat (512 × 512); (e) Cameraman (256 × 256); (f) Fruits (512 × 512); (g) Hill(g) Hill (512 × 512);
(512(h)
× 512);
House (256 × 256); (i) Lena (512 × 512); (j) Man (512 × 512); (k) Monarch (748 × 512); (l) Napoli (512 × 512);
(h) House (256 × 256); (i) Lena (512 × 512); (j) Man (512 × 512); (k) Monarch (748 × 512); (l) Napoli (512 ×
(m) Peppers (256 × 256); and, (n) Zelda (512 × 512).
512); (m) Peppers (256 × 256); and, (n) Zelda (512 × 512).
Tables 3Table
and 41.exhibit
Optimaltheparameters
PSNR (dB)ofand
theSSIM values
existing of the in
methods despeckled
the standard standard images while
images.
using the existing filtering methods and the proposed algorithm. The best and second-best values
among all of theMethods
despeckling methods are denoted in Optimal
red andParameters
blue color, respectively. Table 3 reports
the PSNR valuesNLM Mask size =
of different filtering methods on the fourteen standard
3×3 images. The SRAD filtering
method exhibits Frost
the best speckle noise removal performanceMask size in= 3the
× 3 Baboon (PSNR = 23.52 dB) and
Napoli (PSNR = 26.41Lee dB) images. The PSNR values of MaskSAR-BM3D
size = 3 × 3 are quite similar to that of the
proposed algorithm;
Bitonichowever, SAR-BM3D in the Airplane,
Mask sizeBarbara,
= 3 × 3 Fruits, Hill, and House images
exhibit slightly larger values of PSNR when compared with the proposed method. The PSNR values
WLS Mask size = 3 × 3, λ = 3
of the proposed algorithm show the best performance of speckle noise removal in the images (Boat =
NLLR β = 10, H = 10
27.55 dB; Cameraman = 26.87 dB; Lena = 30.13 dB; Man = 28.55 dB; Monarch = 29.64 dB; Peppers = 28.44
ADMSS ∆t = 0.5, σ = ρ = 0.1, niter = 15
dB; and, Zelda = 32.77 dB) (Table 3).
In Table 4, the performances of the existing Number of rows/cols
filtering methods of block
and the= 9, proposed algorithm are
Maximum size of the 3rd dimension of a stack = 16,
compared in terms of SSIM. As aforementioned, the SRADoffiltering
Diameter technique
search area = 39, provides the best edge
SAR-BM3D
preservation performance in two images (Baboon = Dimension
0.65 and Napoli
of step == 3,
0.77). In the Airplane, Barbara,
Cameraman, Fruits, House, Lena, Monarch, andParameter of the 2D Kaiser
Zelda, SAR-BM3D window
provides = 2,
the best edge preservation
Transform UDWT = daub4
performance while the proposed algorithm exhibits the second-best performance (Airplane, Barbara,
House, Lena, Monarch, and Zelda). The SAR-BM3D and the proposed method show the same edge
preservation performance in Cameraman, Fruits, and Hill. The proposed algorithm shows the highest
edge preservation performance in the following images: Boat = 0.73; Man = 0.77; and Peppers = 0.84.
We confirm that the results obtained by the proposed algorithm, which is displayed in Tables 3 and
4, demonstrate an excellent performance and rank at least second among the existing filtering
techniques.
Remote Sens. 2019, 11, 1184 12 of 27

Table 2. Optimal parameters of the proposed method in the standard images.

SRAD Filter IGF GF

Time step = 0.01
Mask size = 33 × 33 Mask size = 3 × 3
Airplane Exponential decay rate = 1
Regularization parameter = 0.0001 Regularization parameter = 0.001
Number of iterations = 115
Time step = 0.01
Mask size = 5 × 5 Mask size = 3 × 3
Baboon Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 50
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Barbara Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 70
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Boat Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 100
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Cameraman Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 200
Time step = 0.01
Mask size = 17 × 17 Mask size = 3 × 3
Fruits Exponential decay rate = 1
Regularization parameter = 0.01 Regularization parameter = 0.001
Number of iterations = 150
Time step = 0.01
Mask size = 17 × 17 Mask size = 3 × 3
Hill Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 100
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
House Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 190
Time step = 0.01
Mask size = 17 × 17 Mask size = 3 × 3
Lena Exponential decay rate = 1
Regularization parameter = 0.01 Regularization parameter = 0.001
Number of iterations = 150
Time step = 0.01
Mask size = 17 × 17 Mask size = 3 × 3
Man Exponential decay rate = 1
Regularization parameter = 0.01 Regularization parameter = 0.001
Number of iterations = 100
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Monarch Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 100
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Napoli Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 80
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Peppers Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 120
Time step = 0.01
Mask size = 3 × 3 Mask size = 3 × 3
Zelda Exponential decay rate = 1
Regularization parameter = 1e−10 Regularization parameter = 0.001
Number of iterations = 140

Tables 3 and 4 exhibit the PSNR (dB) and SSIM values of the despeckled standard images while
using the existing filtering methods and the proposed algorithm. The best and second-best values
among all of the despeckling methods are denoted in red and blue color, respectively. Table 3 reports
the PSNR values of different filtering methods on the fourteen standard images. The SRAD filtering
method exhibits the best speckle noise removal performance in the Baboon (PSNR = 23.52 dB) and
Napoli (PSNR = 26.41 dB) images. The PSNR values of SAR-BM3D are quite similar to that of the
proposed algorithm; however, SAR-BM3D in the Airplane, Barbara, Fruits, Hill, and House images exhibit
slightly larger values of PSNR when compared with the proposed method. The PSNR values of the
proposed algorithm show the best performance of speckle noise removal in the images (Boat = 27.55 dB;
Cameraman = 26.87 dB; Lena = 30.13 dB; Man = 28.55 dB; Monarch = 29.64 dB; Peppers = 28.44 dB; and,
Zelda = 32.77 dB) (Table 3).
Remote Sens. 2019, 11, 1184 13 of 27

Table 3. Peak signal-to-noise (PSNR) (in dB) results for each standard image.

Noisy NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
Airplane 16.53 19.12 19.14 22.06 23.78 26.18 24.97 17.39 23.43 26.97 26.53 28.10 27.45
Baboon 18.49 21.13 21.09 21.08 21.91 21.97 22.12 19.53 18.28 23.52 22.07 22.51 22.92
Barbara 19.16 22.40 22.05 22.34 23.26 23.68 23.78 20.39 20.50 24.99 23.75 28.32 24.59
Boat 18.46 21.81 21.68 23.36 19.41 26.39 25.50 19.65 20.14 27.37 26.59 27.20 27.55
Camera-man 18.66 21.65 21.59 22.41 22.85 24.43 25.03 19.75 17.59 26.73 24.71 26.35 26.87
Fruits 17.08 19.96 19.98 22.30 24.08 26.33 26.31 18.04 22.07 27.45 26.93 27.68 27.45
Hill 19.79 23.54 23.38 24.64 25.48 27.58 26.75 21.26 24.92 28.25 27.82 28.30 28.27
House 17.93 21.16 21.02 23.26 25.06 27.38 25.93 19.09 22.46 27.58 27.81 29.83 28.58
Lena 18.84 22.45 22.31 24.29 25.88 28.54 27.39 20.11 21.88 29.69 28.99 29.91 30.13
Man 19.51 23.07 22.94 24.41 26.15 27.46 26.46 20.83 20.82 28.31 27.68 27.71 28.55
Monarch 20.19 24.55 24.10 25.11 26.76 27.70 25.87 21.99 24.00 29.50 28.03 29.54 29.64
Napoli 21.00 24.62 24.27 24.06 24.48 24.34 23.69 22.71 22.90 26.41 24.34 25.14 25.70
Peppers 18.74 22.05 21.79 23.50 22.92 26.62 25.77 19.96 18.13 28.29 27.22 27.13 28.44
Zelda 21.18 26.23 25.94 26.71 28.62 31.40 30.66 23.19 29.28 32.67 32.20 32.38 32.77

Table 4. Structural similarity (SSIM) results for each standard image.

Noisy NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
Airplane 0.21 0.29 0.28 0.37 0.50 0.66 0.70 0.25 0.73 0.72 0.76 0.84 0.82
Baboon 0.49 0.56 0.56 0.47 0.54 0.52 0.53 0.53 0.39 0.65 0.53 0.56 0.61
Barbara 0.44 0.61 0.57 0.50 0.60 0.64 0.67 0.55 0.52 0.68 0.65 0.84 0.69
Boat 0.33 0.46 0.44 0.47 0.60 0.68 0.67 0.40 0.39 0.71 0.70 0.72 0.73
Camera-man 0.42 0.49 0.48 0.48 0.57 0.67 0.73 0.45 0.36 0.76 0.74 0.80 0.80
Fruits 0.18 0.28 0.27 0.33 0.48 0.64 0.70 0.23 0.43 0.76 0.76 0.78 0.78
Hill 0.38 0.56 0.54 0.53 0.64 0.69 0.68 0.49 0.58 0.73 0.71 0.73 0.73
House 0.25 0.41 0.38 0.41 0.53 0.67 0.71 0.33 0.53 0.78 0.76 0.84 0.78
Lena 0.29 0.45 0.43 0.45 0.60 0.73 0.75 0.38 0.47 0.81 0.75 0.84 0.83
Man 0.37 0.56 0.54 0.54 0.66 0.72 0.71 0.50 0.50 0.76 0.74 0.76 0.77
Monarch 0.31 0.60 0.55 0.53 0.69 0.81 0.83 0.47 0.80 0.86 0.88 0.90 0.89
Napoli 0.49 0.72 0.69 0.61 0.69 0.70 0.68 0.67 0.66 0.77 0.70 0.73 0.75
Peppers 0.36 0.54 0.52 0.54 0.65 0.77 0.77 0.46 0.36 0.82 0.82 0.83 0.84
Zelda 0.35 0.61 0.58 0.55 0.70 0.80 0.82 0.51 0.77 0.86 0.85 0.87 0.86
Remote Sens. 2019, 11, 1184 14 of 27

In Table 4, the performances of the existing filtering methods and the proposed algorithm are
compared in terms of SSIM. As aforementioned, the SRAD filtering technique provides the best edge
preservation performance in two images (Baboon = 0.65 and Napoli = 0.77). In the Airplane, Barbara,
Cameraman, Fruits, House, Lena, Monarch, and Zelda, SAR-BM3D provides the best edge preservation
performance while the proposed algorithm exhibits the second-best performance (Airplane, Barbara,
House, Lena, Monarch, and Zelda). The SAR-BM3D and the proposed method show the same edge
preservation performance in Cameraman, Fruits, and Hill. The proposed algorithm shows the highest
edge preservation performance in the following images: Boat = 0.73; Man = 0.77; and Peppers = 0.84.
We confirm that the results obtained by the proposed algorithm, which is displayed in Tables 3 and 4,
demonstrate an excellent performance and rank at least second among the existing filtering techniques.
From Tables 3 and 4, in the standard images, we analyze the performances of each method to
evaluate the soft threshold, IGF, and GF in the wavelet domain (Table 5). In Baboon (PSNR = 22.69
(−0.83), SSIM = 0.60 (−0.05), Barbara (PSNR = 24.32 (−0.67), SSIM = 0.66 (−0.02)), Boat (PSNR = 27.23
(−0.14), SSIM = 0.70 (−0.01), Hill (PSNR = 28.06 (−0.19), SSIM = 0.72 (−0.01)), Man (PSNR = 28.14
(−0.17), SSIM = 0.75 (−0.01), and Napoli (PSNR = 25.72 (−0.69), SSIM = 0.75 (−0.02) images, it is
confirmed that the noise reduction and the edge preservation performances are low when the soft
threshold method is compared with the SRAD result image. In comparison with the SRAD result,
only the edge preservation performance of the soft threshold has been improved in Airplane (PSNR
= 26.94 (−0.03), SSIM = 0.73 (+0.01)) image. Cameraman (PSNR = 26.69 (−0.04), SSIM = 0.76 (0.00)),
Fruits (PSNR = 27.44 (−0.01), SSIM = 0.76 (0.00)), Monarch (PSNR = 29.42 (−0.08), SSIM = 0.86 (0.86
(0.00), and Peppers (PNSR = 28.21 (−0.08), SSIM = 0.82 (0.00)) images have a low speckle noise removal
and the same edge preservation performances. In the Lena (PSNR = 29,69 (0.00), SSIM = 0.81 (0.00))
and Zelda (PSNR = 32.67 (0.00), SSIM = 0.86 (0.00)) images, the soft threshold technique shows the
same noise removal and edge preservation abilities as SRAD; the House (PSNR = 27.92 (+0.34), SSIM
= 0.70 (−0.08)) image has only enhanced the noise suppression ability; however, it shows low edge
preservation performance.
From Table 5, in the Man image, the IGF only exhibits reduced noise removal and the same edge
preservation abilities (PSNR = 28.30 (−0.01), SSIM = 0.76). The Napoli (PSNR = 26.41, SSIM = 0.77)
image has the same noise suppression and edge preservation performances. In the House (PSNR =
27.98 (+0.40), SSIM = 0.70 (−0.08)), enhanced noise reduction, and decreased edge preservation abilities
are showed when the IGF method is applied. The IGF compared to the SRAD result image shows
better noise suppression performance and the same edge preservation ability (Airplane (PSNR = 26.98
(+0.01), SSIM = 0.72), Baboon (PSNR = 23.53 (+0.01), SSIM = 0.65), Barbara (PSNR = 25.00 (+0.01), SSIM
= 0.68), Boat (PSNR = 27.39 (+0.02), SSIM = 0.71), Cameraman (PSNR = 26.74 (+0.01), SSIM = 0.76),
Hill (PSNR = 28.27 (+0.02), SSIM = 0.73), Lena (PSNR = 29.70 (+0.01), SSIM = 0.81), Monarch (PSNR
= 29.51 (+0.01), SSIM = 0.86), Peppers (PSNR = 28.31 (+0.02), SSIM = 0.82), and Zelda (PSNR = 32.68
(+0.01), SSIM = 0.86) images). Fruits (PSNR = 27.46 (+0.01), SSIM = 0.78) image show enhanced noise
reduction and edge preservation performances.
The result of comparing PSNR and SSIM values of GF and SRAD filtered images are as follows
(Table 5): In the Napoli (PSNR = 26.39 (−0.02), SSIM = 0.78 (+0.01)) image, the noise reduction
performance is reduced and the edge preservation ability is improved. The noise suppression
performance is improved; however, the edge preservation ability is maintained in the House (PSNR
= 28.64 (+1.06), SSIM = 0.78) and Zelda (PSNR = 32.78 (+0.11), SSIM = 0.86) images. In contrast, the
Fruits image exhibits the same noise removal ability and enhanced edge preservation performance
(PSNR = 27.45, SSIM = 0.78 (+0.02)). The Airplane, Baboon, Barbara, Boat, Cameraman, Hill, Lena, Man,
Monarch, and Peppers images have an enhanced noise suppression and edge preservation performances
(Airplane (PSNR = 27.48 (+0.51), SSIM = 0.82 (+0.10)), Baboon (PSNR = 23.78 (+0.26), SSIM = 0.66
(+0.01)), Barbara (PSNR = 25.30 (+0.31), SSIM = 0.71 (+0.03)), Boat (PSNR = 27.67 (+0.30), SSIM = 0.73
(+0.02)), Cameraman (PSNR = 26.91 (+0.17), SSIM = 0.80 (+0.04)), Hill (PSNR = 28.41 (+0.16), SSIM =
0.74 (+0.01)), Lena (PSNR = 29.72 (+0.03), SSIM = 0.82 (+0.01)), Man (PSNR = 28.52 (+0.21), SSIM = 0.77
Remote Sens. 2019, 11, 1184 15 of 27

(+0.01)), Monarch (PSNR = 29.71 (+0.21), SSIM = 0.89 (+0.03)), Peppers (PSNR = 28.38 (+0.09), SSIM =
0.84 (+0.02))).

Table 5. PSNR and SSIM results of each method in the proposed algorithm for each standard image.

SRAD Filter Soft Threshold IGF GF Proposed

PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
26.94 0.73 26.98 0.72 27.48 0.82
Airplane 26.97 0.72 27.45 0.82
(−0.03) (+0.01) (+0.01) (0.00) (+0.51) (+0.10)
22.69 0.60 23.53 0.65 23.78 0.66
Baboon 23.52 0.65 22.92 0.61
(−0.83) (−0.05) (+0.01) (0.00) (+0.26) (+0.01)
24.32 0.66 25.00 0.68 25.30 0.71
Barbara 24.99 0.68 24.59 0.69
(−0.67) (−0.02) (+0.01) (0.00) (+0.31) (+0.03)
27.23 0.70 27.39 0.71 27.67 0.73
Boat 27.37 0.71 27.55 0.73
(−0.14) (−0.01) (+0.02) (0.00) (+0.30) (+0.02)
26.69 0.76 26.74 0.76 26.90 0.80
Cameraman 26.73 0.76 26.87 0.80
(−0.04) (0.00) (+0.01) (0.00) (+0.17) (+0.04)
27.44 0.76 27.46 0.78 27.45 0.78
Fruits 27.45 0.76 27.45 0.78
(−0.01) (0.00) (+0.01) (+0.02) (0.00) (+0.02)
28.06 0.72 28.27 0.73 28.41 0.74
Hill 28.25 0.73 28.27 0.73
(−0.19) (−0.01) (+0.02) (0.00) (+0.16) (+0.01)
27.92 0.70 27.98 0.70 28.64 0.78
House 27.58 0.78 28.58 0.78
(+0.34) (−0.08) (+0.40) (−0.08) (+1.06) (0.00)
29.69 0.81 29.70 0.81 29.72 0.82
Lena 29.69 0.81 30.13 0.83
(0.00) (0.00) (+0.01) (0.00) (+0.03) (+0.01)
28.14 0.75 28.30 0.76 28.52 0.77
Man 28.31 0.76 28.55 0.77
(−0.17) (−0.01) (−0.01) (0.00) (+0.21) (+0.01)
29.42 0.86 29.51 0.86 29.71 0.89
Monarch 29.50 0.86 29.64 0.89
(−0.08) (0.00) (+0.01) (0.00) (+0.21) (+0.03)
25.72 0.75 26.41 0.77 26.39 0.78
Napoli 26.41 0.77 25.70 0.75
(−0.69) (−0.02) (0.00) (0.00) (−0.02) (+0.01)
28.21 0.82 28.31 0.82 28.38 0.84
Peppers 28.29 0.82 28.44 0.84
(−0.08) (0.00) (+0.02) (0.00) (+0.09) (+0.02)
32.67 0.86 32.68 0.86 32.78 0.86
Zelda 32.67 0.86 32.77 0.86
(0.00) (0.00) (+0.01) (0.00) (+0.11) (0.00)
Avg. −0.19 −0.01 +0.04 0.00 +0.24 +0.02

Figures 4–6 show the despeckled images that are provided by existing filtering methods and the
proposed algorithm for the standard images. Figure 4b–l, Figures 5b–l and 6b–l exhibit the filtering
result images of GF, Frost filter, Lee filter, Bitonic filter, WLS filter, NLLR method, ADMSS method,
SRAD filter, SRAD-Guided algorithm, SAR-BM3D, and the proposed algorithm, respectively. In the
Cameraman image, the issue of speckle noise residue in homogeneous regions appears in Figure 4b–e,g,h.
Figure 4e, which is compared to Figure 4b–d,g,h exhibits reduced speckle noise but not quite. As shown
in Figure 4f,i,j, some edges are lost in the edge regions, whereas the homogeneous regions remain with
the speckle noise. The speckle noise reduction and edge preservation performance is noticeable in
SAR-BM3D and the proposed algorithm (Figure 4k,l). The SAR-BM3D and the proposed algorithm
exhibit similar performance with respect to the edge preservation, and show the strongest speckle
removal ability in the homogeneous regions. However, the proposed algorithm has the best speckle
noise removal performance in the homogeneous areas, as SAR-BM3D exhibits artifacts in these regions.
Remote Sens. 2019, 11, 1184 16 of 27
Remote Sens. 2019, 11, x FOR PEER REVIEW 15 of 26

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure
Figure 4.4. Performance
Performance comparison
comparison of different techniques
of different in Cameraman
techniques image. (a)
in Cameraman Noisy;(a)
image. (b) Noisy;
Guided;(b)
(c) Frost; (d) Lee; (e) Bitonic; (f) weighted-least-squares (WLS); (g) non-local low-rank (NLLR);
Guided; (c) Frost; (d) Lee; (e) Bitonic; (f) weighted-least-squares (WLS); (g) non-local low-rank (NLLR); (h)
(h)anisotropic
anisotropicdiffusion
diffusionfilter
filterwith
withmemory
memorybased
basedononspeckle
specklestatistics (ADMSS);
statistics (ADMSS); (i)(i)
speckle
specklereducing
reducing
anisotropic
anisotropic diffusion
diffusion (SRAD);(j)(j)SRAD-Guided;
(SRAD); SRAD-Guided; (k) (k) SAR-BM3D;
SAR-BM3D; and,and,(l)
(l)Proposed
Proposedalgorithm.
algorithm.

Figures 4–6 show the despeckled images that are provided by existing filtering methods and the
proposed algorithm for the standard images. Figure 4b–l, Figure 5b–l and Figure 6b–l exhibit the
homogeneous regions remain with the speckle noise. The speckle noise reduction and edge
preservation performance is noticeable in SAR-BM3D and the proposed algorithm (Figure 4k,l). The
SAR-BM3D and the proposed algorithm exhibit similar performance with respect to the edge
preservation, and show the strongest speckle removal ability in the homogeneous regions. However,
the proposed
Remote Sens. 2019,algorithm
11, 1184 has the best speckle noise removal performance in the homogeneous 17
areas,
of 27
as SAR-BM3D exhibits artifacts in these regions.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure 5. Performance
Performance comparison
comparison of of different techniques
techniques in Monarch image. (a)
(a) Noisy;
Noisy; (b)
(b) Guided;
Guided;
(c) Frost;
Frost;(d)
(d)Lee;
Lee;(e)
(e)Bitonic;
Bitonic; (f) WLS; (g) NLLR; (h) ADMSS; (i) SRAD; (j) SRAD-Guided; (k)
(f) WLS; (g) NLLR; (h) ADMSS; (i) SRAD; (j) SRAD-Guided; (k) SAR-BM3D; SAR-
BM3D;
and, (l) and, (l) Proposed
Proposed algorithm.algorithm.

GF, Frost
The GF, Frost filter,
filter, Lee
Leefilter,
filter,Bitonic
Bitonicfilter,
filter,NLLR
NLLRmethod,
method, ADMSS
ADMSS method,
method, andand
SRADSRAD filter
filter do
do not
not perform
perform well well
forfor speckle
speckle noise
noise removalininthe
removal thehomogeneous
homogeneousregions,
regions,and
and the
the speckle
speckle noise
persists
persists ininthe
thefiltering
filteringresult images
result (Figure
images 5b–e,g–i).
(Figure The WLS
5b–e,g–i). Thefilter
WLS and the SRAD-Guided
filter algorithm
and the SRAD-Guided
perform
algorithmbetter thanbetter
perform the above
thanfiltering
the above methods
filteringwith regardwith
methods to speckle
regardnoise reduction
to speckle noise(Figure 5f,j).
reduction
However,5f,j).
(Figure the WLS filter and
However, thetheWLS
SRAD-Guided
filter andalgorithm exhibit a blurring
the SRAD-Guided phenomenon
algorithm exhibit ina the image.
blurring
The filtering result
phenomenon in theimage
image. that was
The obtained
filtering by image
result the proposed algorithm
that was obtainedhad by similar visual algorithm
the proposed quality as
SAR-BM3D.
had similar The
visual SAR-BM3D
quality asachieves
SAR-BM3D. excellent
Theedge preservation
SAR-BM3D performance,
achieves excellent however it exhibits
edge preservation
artifacts in the homogeneous regions (Figure 5k). In Figure 5l, the proposed algorithm exhibits strong
speckle noise removal ability while maintaining the edges. The qualitative result of Figure 5 represents
the same result as in Figure 6.
 Remote Sens. 2019, 11, x FOR PEER REVIEW 17 of 26

performance, however it exhibits artifacts in the homogeneous regions (Figure 5k). In Figure 5l, the
Remote Sens. 2019,
proposed 11, 1184 exhibits strong speckle noise removal ability while maintaining the edges. The
algorithm 18 of 27
qualitative result of Figure 5 represents the same result as in Figure 6.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure 6. Performance comparison of different techniques in Peppers image. (a) Noisy; (b) Guided;
(c) Frost; (d) Lee; (e) Bitonic; (f) WLS; (g) NLLR; (h) ADMSS; (i) SRAD; (j) SRAD-Guided; (k) SAR-BM3D;
and, (l) Proposed algorithm.
 SRAD- SAR-
NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD Proposed
Guided BM3D
ROI1
50.80 17.89 47.59 64.94 91.46 165.71 21.61 18.59 114.10 125.44 135.16 141.78
(61 × 71)
ROI2
40.87 16.25 37.85 49.15 64.98 118.11 19.41 16.78 81.01 88.88 85.09 99.92
(51 ×Sens.
Remote 71) 2019, 11, 1184 19 of 27

Table 8. ENL results for SAR image2.

3.2. Experiments on Real SAR Images
SRAD- SAR-
NLM Guided
In this section, two realFrost Lee Bitonic
SAR images showingWLS NLLR
different ADMSS
scenes areSRAD
used for Proposed
the evaluation of the
Guided BM3D
conventional
ROI1 filtering methods and the proposed algorithm on real SAR images (Figure 7). The real
29.53 16.17 48.47 62.14 99.30 207.56 21.20 201.56 146.91 174.02 186.54 205.89
SAR × 71) shows a scene from a photojournal [256 × 256, 8 bit, X-band] [57]. The real SAR image2
(61image1
ROI2a rural scene in Bedfordshire [512 × 512, 8 bit, X-band] [2,58]. As mentioned in Section 3.1,
depicts 28.66 13.13 39.43 50.79 80.55 180.37 20.56 124.83 117.17 141.30 129.35 160.67
the(81 × 51) parameters of the conventional methods maintained the same values as those in the
optimal
standard image (Table 1). Table 6 shows the optimal parameters of the proposed algorithm.

(a) (b)

Figure 7. Real
Figure 7. Real SAR
SAR images
images used
used in
in the
the experiments.
experiments. (a)
(a) SAR
SAR image1
image1 [58];
[58]; (b)
(b) SAR
SAR image2
image2 [59].
[59].

Table 6. Optimal parameters of the proposed algorithm in the Synthetic Aperture Radar (SAR) image.
Figure 8 shows the results of the simulated SAR images. Figure 8 shows that some filters, such
as Guided, Frost, Lee, Bitonic, NLLR, and SRAD, do not exhibit
SRAD Filter IGF strong speckle noise GF removal ability
(Figure 8b–e,g,i). Tables 6 and 7Timeillustrate that
step = 0.01 the WLS filter represents
Mask size = 33 × 33 the best
Mask sizevalue;
ENL = 3 × 3 however,
SAR image1 Exponential decay rate = 1 Regularization Regularization
it exhibits a blurring phenomenon in the image (Figure 8f). The SAR-guided method when compared
Number of iterations = 140 parameter = 0.0001 parameter = 0.001
with the SAR-BM3D and the proposed methods show inferior speckle noise removal and edge
Time step = 0.01 Mask size = 33 × 33 Mask size = 3 × 3
preservation performances.
SAR image2
The SAR-BM3D method has
Exponential decay rate = 1
an excellent speckle
Regularization
noise reduction and
Regularization
edge preservation abilities;Number
however, artifacts
of iterations in the homogeneous
= 145 regions areparameter
parameter = 0.0001 observable (Figure 8k).
= 0.001
The proposed algorithm compared with the SAR-BM3D method has a strong speckle noise removal
ability; however,
Tables 7 and 8itlist
does
theexhibit limited
ENL values thatlow edge
were preservation
computed in twoperformance in some
regions of interest edgeofareas
(ROIs) the
(Figure 8l).
real SAR images. The GF in the ROI shows a low ENL value (SAR image1 (ENL1 = 17.89, ENL2 =
16.25), SAR image2 (ENL1 = 16.17, ENL2 = 13.13)), while the conventional filtering methods have
Table 9. ENL results of each method in the proposed algorithm for each real SAR images.
better results, as shown in Tables 7 and 8. According to Table 7, the NLM filter (ROI1 = 50.80, ROI2 =
40.87), Frost filterSRAD = 47.59, Soft
(ROI1Filter ROI2 = 37.85), Lee filter
Threshold (ROI1 = 64.94, ROI2
IGF GF = 49.15), Proposed
Bitonic filter
(ROI1 = 91.46, ROI2 = 64.98), NLLR method (ROI1 = 21.61, ROI2 = 19.41), ADMSS methodROI-2
ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 (ROI1
=SAR
18.59, ROI2 = 16.78),
image1 114.10 SRAD
81.01
114.62
filter (ROI1 =
81.59 118.84
114.10, ROI2 =
84.09
81.01),136.52
and the 97.05
SRAD-Guided
141.78 method
99.92
(ROI1 = 125.44, ROI2 = 88.88) do(+0.52) (+0.58)
not exhibit (+4.74)
excellent (+2.29)
speckle noise(+22.42)
removal(+16.04)
performance in terms of
the
SARENL. The SAR-BM3D
image2 146.91 (ROI1147.76
117.17 = 135.16,118.50
ROI2 =148.93 119.27
85.09), WLS 203.02 157.24
(ROI1 = 165.71, ROI2205.89
= 118.11), and
160.67
(+0.85) (+1.33) (+2.02) (+2.10) (+56.11) (+40.07)
proposed (ENL1 = 141.78, ENL2 = 99.92) methods show similar values of excellent speckle noise
Avg. +0.69 +0.96 +3.38 +2.20 +39.27 +68.56
suppression ability. The SAR-BM3D, WLS, and proposed methods outperform the existing filtering
methods, as implied by the higher ENL values were obtained by these methods in comparison to
those of the existing filtering methods. Table 8 shows that the NLM filter (ROI1 = 29.53, ROI2 = 28.66),
Frost filter (ROI1 = 48.47, ROI2 = 39.43), Lee filter (ROI1 = 62.14, ROI2 = 50.79), Bitonic filter (ROI1 =
99.30, ROI2 = 80.55), and NLLR method (ROI1 = 21.20, ROI2 = 20.56) have low speckle noise removal
performances. The WLS (ROI1 = 207.56, ROI2 = 180.37), ADMSS (ROI1 = 201.56, ROI2 = 124.83), SRAD
filter (ROI1 = 146.91, ROI2 = 117.17), SRAD-Guided (ROI1 = 174.02, ROI2 = 141.30), SAR- BM3D (ROI1
= 186.54, ROI2 = 129.35), and proposed (ROI1 = 205.89, ROI2 = 160.67) algorithms exhibit similar ENL
values. Among these techniques, the WLS, ADMSS, and proposed methods represent a better ENL
result when compared to the conventional filtering methods in speckle noise removal. Tables 7 and 8
Remote Sens. 2019, 11, 1184 20 of 27

depict that the WLS filter outperforms all of the filtering methods in terms of the ENL, while the
proposed method ranks second in the speckle noise suppression performance.

Table 7. Equivalent number of looks (ENL) results for SAR image1.

NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
ROI1
50.80 17.89 47.59 64.94 91.46 165.71 21.61 18.59 114.10 125.44 135.16 141.78
(61 × 71)
ROI2
40.87 16.25 37.85 49.15 64.98 118.11 19.41 16.78 81.01 88.88 85.09 99.92
(51 × 71)

Table 8. ENL results for SAR image2.

NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
ROI1
29.53 16.17 48.47 62.14 99.30 207.56 21.20 201.56 146.91 174.02 186.54 205.89
(61 × 71)
ROI2
28.66 13.13 39.43 50.79 80.55 180.37 20.56 124.83 117.17 141.30 129.35 160.67
(81 × 51)

From Tables 7 and 8, the data from Table 9 are analyzed for evaluating the performance of
each method in the proposed method for the real SAR images. In the real SAR image1 and image2,
the soft threshold, the IGF, and the GF methods, as compared to the SRAD filtering result image,
shows enhanced noise suppression ability. In the SAR image1, the soft threshold, the IGF, and the GF
techniques have enhanced noise removal ability (ROI-1 (ENL = 114.62 (+0.52)) and -2 (ENL = 81.59
(+0.58), ROI-1 (ENL = 118.84 (+4.74)) and -2 (ENL = 84.09 (+2.29), ROI-1 (ENL = 136.52 (+21.90)) and
-2 (ENL = 97.05 (+16.04)). The noise reduction performance of the soft threshold (ROI-1 (ENL = 147.76
(+0.85)) and ROI-2 (ENL = 118.50 (+1.33)), IGF (ROI-1 (ENL = 148.93 (+2.02)) and ROI-2 (ENL = 119.27
(+2.10)), and the GF (ROI-1 (ENL = 203.02 (+56.11)) and ROI-2 (ENL = 157.24 (+40.07)) methods in the
SAR image2 exhibits improved ability (Table 9).

Table 9. ENL results of each method in the proposed algorithm for each real SAR images.

SRAD Filter Soft Threshold IGF GF Proposed

ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 ROI-2 ROI-1 ROI-2
SAR 114.62 81.59 118.84 84.09 136.52 97.05
114.10 81.01 141.78 99.92
image1 (+0.52) (+0.58) (+4.74) (+2.29) (+22.42) (+16.04)
SAR 147.76 118.50 148.93 119.27 203.02 157.24
146.91 117.17 205.89 160.67
image2 (+0.85) (+1.33) (+2.02) (+2.10) (+56.11) (+40.07)
Avg. +0.69 +0.96 +3.38 +2.20 +39.27 +68.56

Figure 8 shows the results of the simulated SAR images. Figure 8 shows that some filters,
such as Guided, Frost, Lee, Bitonic, NLLR, and SRAD, do not exhibit strong speckle noise removal
ability (Figure 8b–e,g,i). Tables 6 and 7 illustrate that the WLS filter represents the best ENL value;
however, it exhibits a blurring phenomenon in the image (Figure 8f). The SAR-guided method when
compared with the SAR-BM3D and the proposed methods show inferior speckle noise removal and
edge preservation performances. The SAR-BM3D method has an excellent speckle noise reduction and
edge preservation abilities; however, artifacts in the homogeneous regions are observable (Figure 8k).
The proposed algorithm compared with the SAR-BM3D method has a strong speckle noise removal
ability; however, it does exhibit limited low edge preservation performance in some edge areas
(Figure 8l).
Remote Sens. 2019, 11, 1184 21 of 27
Remote Sens. 2019, 11, x FOR PEER REVIEW 20 of 26

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure8.8. Performance
Figure Performance comparison
comparisonofofdifferent
differenttechniques
techniquesinin
SAR image2.
SAR (a) Noisy;
image2. (b) Guided;
(a) Noisy; (c)
(b) Guided;
Frost; (d) Lee; (e) Bitonic; (f) WLS; (g) NLLR; (h) ADMSS; (i) SRAD; (j) SRAD-Guided; (k) SAR-BM3D;
(c) Frost; (d) Lee; (e) Bitonic; (f) WLS; (g) NLLR; (h) ADMSS; (i) SRAD; (j) SRAD-Guided; (k) SAR-BM3D;
and,(l)
and, (l)Proposed
Proposedalgorithm.
algorithm.

3.3. Computational Complexity

Remote Sens. 2019, 11, 1184 22 of 27

3.3. Computational Complexity

Tables 10 and 11 in this section present the time costs of the existing methods and the proposed
algorithm on 14 standard images and two real SAR images. The experimental environments are those
that are referred to in Section 3.1. Table 10 shows that the proposed method is much faster than
Lee, NLLR, ADMSS, and SAR-BM3D. The running time of the proposed algorithm with 14 standard
images is approximately 5.06 s on average. Table 11 denotes that the proposed algorithm has less
computational complexity than the NLLR, ADMSS, and SAR-BM3D methods. The average execution
time in the real SAR images is approximately 4.50 s.

Table 10. Computational complexity results (in seconds) of the de-speckling methods for each
standard image.
NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
Airplane 0.48 0.16 1.86 6.41 0.09 3.51 1052.12 196.87 5.51 5.92 61.50 5.70
Baboon 0.48 0.11 2.00 7.29 0.10 0.48 1030.23 173.14 2.45 2.61 59.84 2.76
Barbara 0.50 0.12 2.05 7.28 0.08 1.00 1003.88 162.76 3.44 3.74 59.36 3.74
Boat 0.48 0.11 2.01 7.28 0.09 0.98 1007.25 174.22 5.06 5.48 61.07 5.36
Cameraman 0.12 0.08 0.52 1.88 0.03 0.46 211.28 21.64 1.55 1.21 14.45 1.91
Fruits 0.48 0.11 2.03 7.31 0.09 0.97 1012.13 181.41 7.40 7.85 62.17 7.84
Hill 0.48 0.11 1.98 7.25 0.09 0.99 1061.75 162.39 4.98 5.62 61.38 5.28
House 0.12 0.09 0.53 1.92 0.03 0.49 231.46 28.01 1.54 1.05 14.34 1.83
Lena 0.48 0.16 1.86 6.47 0.10 1.00 1081.19 170.44 7.53 8.03 60.16 7.71
Man 0.48 0.11 1.99 7.32 0.09 1.09 1057.03 165.61 5.23 5.70 60.15 5.40
Monarch 0.73 0.13 2.85 9.77 0.12 1.51 1661.38 277.26 8.48 5.96 87.94 8.93
Napoli 0.50 0.12 1.90 6.64 0.08 1.07 1060.22 168.11 4.02 4.10 59.55 4.31
Peppers 0.12 0.09 0.50 1.71 0.03 0.50 218.14 26.96 1.04 1.16 14.49 1.32
Zelda 0.48 0.12 1.88 6.88 0.09 0.99 1001.87 164.50 7.10 7.45 59.57 7.32
Avg. 0.42 0.12 1.71 6.10 0.08 1.07 906.42 148.09 4.67 4.71 52.57 5.06

Table 11. Computational complexity results of the de-speckling methods for the real SAR images.
NLM Guided Frost Lee Bitonic WLS NLLR ADMSS SRAD SRAD-Guided SAR-BM3D Proposed
SAR image1 0.16 0.08 0.46 0.45 0.10 0.17 222.16 29.41 1.14 1.08 14.40 1.56
SAR image2 0.47 0.19 1.73 6.23 0.12 0.71 1071.83 191.19 7.09 7.48 62.95 7.45
Avg. 0.32 0.14 1.10 3.34 0.11 0.44 647.04 110.30 4.12 4.28 38.68 4.50

Tables 12 and 13 present the computing time of each step of the proposed method for the standard
and real SAR images. The SRAD method has a high computing time, because the SRAD filter uses an
iterative method to remove the speckle noise (standard images = 4.76 s (91.56%); SAR images = 4.12 s
(91.56%)). In finding an optimal threshold value for classifying an original signal and a noise signal,
the computing time of the soft threshold method is low, because, when compared to the computing
time of the SRAD filter, it takes approximately 0.11 s (standard images) and 0.10 s (real SAR images).
The IGF and the GF work very fast, because they only take approximately 6% (standard images) and
9% (real SAR images), respectively, of the total time. The main reason for this low time consumption is
the use of a box filter in the GF [45]. The box filter can efficiently use a computational complexity in
O(N ) time by employing the integral image method [59]. The IGF is a method developed based on the
GF; hence, it has a low computation time.

Table 12. Time consumption (in seconds) of each step (proposed algorithm) in the standard image.

Time for Time for Soft Time for

Image Size Time for GF Total Time
SRAD Threshold IGF
Airplane 512 × 512 5.51 0.11 0.05 0.03 5.70
Baboon 512 × 512 2.45 0.11 0.17 0.03 2.76
Barbara 512 × 512 3.44 0.11 0.16 0.03 3.74
Boat 512 × 512 5.06 0.11 0.16 0.03 5.36
Cameraman 256 × 256 1.55 0.10 0.23 0.03 1.91
Fruits 512 × 512 7.40 0.11 0.30 0.03 7.84
Remote Sens. 2019, 11, 1184 23 of 27

Table 12. Cont.

Time for Time for Soft Time for

Image Size Time for GF Total Time
SRAD Threshold IGF
Hill 512 × 512 4.98 0.11 0.16 0.03 5.28
House 512 × 512 1.54 0.10 0.16 0.03 1.83
Lena 512 × 512 7.53 0.11 0.04 0.03 7.71
Man 512 × 512 5.23 0.12 0.05 0.03 5.40
Monarch 748 × 512 8.48 0.12 0.30 0.03 8.93
Napoli 512 × 512 4.02 0.12 0.13 0.04 4.31
Peppers 256 × 256 1.04 0.11 0.14 0.03 1.32
Zelda 512 × 512 7.10 0.12 0.10 0.03 7.32
Avg. 4.67 0.11 0.15 0.03 4.96

Table 13. Time consumption (in seconds) of each step (proposed algorithm) for the real SAR images.

Time for Time for Soft Time for

Image Size Time for GF Total Time
SRAD Threshold IGF
SAR Image1 256 × 256 1.14 0.10 0.29 0.03 1.56
SAR Image2 512 × 512 7.09 0.11 0.22 0.03 7.45
Avg. 4.12 0.10 0.26 0.03 4.50

4. Discussion
This study used the statistical characteristics of speckle noise and the DWT to remove the speckle
noise in SAR images. The proposed algorithm applies the SRAD filter, soft threshold, GF, and the IGF.
The speckle noise in SAR images is modelled as multiplicative noise. However, most of the filtering
methods were developed for AWGN, as additive noise in imaging and sensing systems is most common.
Therefore, conventional filtering methods are unable to remove speckle noise. The SRAD filtering
method, in contrast, uses the ICOV to directly apply a diffusion process in all areas, except for the
edge regions, by separating the edge areas and noise from SAR images with speckle noise. The SRAD
filter exhibits excellent speckle noise removal and edge preservation. From the experimental results,
the SRAD filtering scheme demonstrates the best speckle noise suppression and edge preservation
performance among the single filtering methods. Based on this finding, the SRAD filtering technique
was used as a preprocessing filter. In order to further remove the speckle noise remaining in the SRAD
filtering result image, the logarithmic transform is used to convert the multiplicative noise (speckle
noise) to additive noise. The SRAD filtering result image with the additive noise is decomposed
into one low-frequency sub-band image (LL2 ) and six high-frequency sub-band images (LH1 , HL1 ,
HH1 LH2 , HL2 , and HH2 ) while using a two-level DWT. In the wavelet domain, horizontal (HL1 , HL2 )
and vertical (LH1 , LH2 ) sub-band images have the same energy, while, in comparison, the diagonal
(HH1 , HH2 ) sub-band images have lower energies on the same scale. The former sub-band images
were applied to soft threshold to remove the additive noise. The IGF method with new edge-aware
weighting based on the Gradient and the Laplacian operators is applied to the latter sub-band images
in order to remove the additive noise and preserve low edge information. We applied the guide filter
to remove the additive noise that is present in the approximate (LL2 ) sub-band image. In some of
the standard images, the proposed algorithm does not represent the best speckle noise removal and
edge preservation performance in terms of PSNR and SSIM (Tables 3 and 4). When the soft threshold,
IGF, and GF are employed in the wavelet domain after the SRAD filter application in the proposed
algorithm, when compared with the SRAD filter, the proposed algorithm shows enhanced speckle
noise and edge preservation performance in the Airplane (PSNR = 0.48 dB; SSIM = 0.10), Boat (PSNR
= 0.18 dB; SSIM = 0.02), Cameraman (PSNR = 0.14 dB; SSIM = 0.04), Lena (PSNR = 0.44 dB; SSIM =
0.02), Man (PSNR = 0.24 dB; SSIM = 0.01), Monarch (PSNR = 0.14 dB; SSIM = 0.03), Peppers (PSNR =
0.15 dB; SSIM = 0.02), and Zelda (PSNR = 0.15 dB; SSIM = 0.02) images. We analyzed the performance
of each technique (soft threshold, IGF and GF) in the wavelet domain to achieve these results from
Remote Sens. 2019, 11, 1184 24 of 27

standard images (Table 5). Among these methods, the soft threshold shows low speckle noise and
edge preservation abilities in most standard images (average: PSNR = −0.19 dB; SSIM = −0.01). The
IGF method exhibits a limited improvement in the speckle noise suppression performance (PSNR =
+0.04 dB (avg.)). The GF technique contributes most of the speckle noise removal and edge preservation
performances among each method in the wavelet domain (PSNR = +0.24 dB, SSIM = +0.02 (average)).
When the same method that is applied in the standard images is applied to real SAR images, the
proposed algorithm shows an improved speckle noise reduction performance in the ROI of the two
real SAR images (SAR image1 (ROI-1: ENL = 27.68; ROI-2: ENL = 18.91) and SAR image2 (ROI-1:
ENL = 58.98; ROI-2: ENL = 43.50)). From the ENL results in the real SAR images, we analyzed the
contributions of the soft threshold, IGF, and GF to the noise suppression performance (Table 9). The soft
threshold shows improved speckle noise rejection performance for ROI-1 (ENL = +0.69) and -2 (ENL =
+0.85) in SAR image1 and 2. The IGF method exhibits enhanced speckle noise removal ability over the
soft threshold (ROI-1 (ENL = +3.38) and ROI-2 (ENL = +2.20)). The GF technique was confirmed to
have improved speckle noise reduction performance in terms of ENL = +39.27 at ROI-1 and ENL =
+68.56 at ROI-2. The GF method has been found to have the greatest contribution to speckle noise
removal in the wavelet domain.
The proposed algorithm shows the performance within the second-best value in all of the standard
images and real SAR images in Table 3, Table 4, Table 7, and Table 8. The proposed method performs
better than any nonlinear filter and hybrid method in the different images that have characteristics
that include low-frequency components. Although the SAR-BM3D method, when compared to the
conventional algorithms, exhibits excellent speckle noise reduction and edge preservation abilities,
it employs noise reduction based on the NLM filter. The computational complexity of the SAR-BM3D
algorithm is high, since the NLM filter needs to search regions. Therefore, it is difficult to obtain
real time observations using the SAR-BM3D. The proposed algorithm exhibits a 10–30 times lower
computational complexity when compared with SAR-BM3D (Table 10). In the proposed algorithm,
the SRAD filter has a high computational complexity, because it uses an iterative method to remove
the speckle noise (standard images = 4.76 s (91.56%); SAR images = 4.12 s (91.56%)); however, the time
that is consumed by the soft threshold in finding an optimal value to classify an original signal and a
noise signal is low (0.11 s (standard images) and 0.10 s (SAR images)). Moreover, the IGF and the GF
have a low computational complexity (standard image ≈ 6%; real SAR images ≈ 9%), because the box
filter in the IGF and the GF can efficiently employ computing time (O(N )). Moreover, the proposed
method exhibits speckle noise suppression and edge preservation performance similarly to SAR-BM3D
(Tables 3 and 4). In the real SAR images, the WLS filter exhibits the best speckle noise removal
performance in terms of ENL (Tables 7 and 8). However, the resulting WLS image exhibits a blurring
phenomenon (Figure 8f). As mentioned in [60], high ENL values do not always imply the best speckle
noise suppression performance; further, blurring is observed in the image. Tables 7 and 8 indicate
that SAR-BM3D and the proposed method provide satisfactory speckle noise removal (Figure 8k,l).
As mentioned above, the proposed algorithm has low computational complexity, which is about
8–13 times that of the SAR-BM3D method (Table 11). The experimental results demonstrate that the
proposed method exhibits excellent speckle noise reduction, while preserving edge information and
maintaining low computational complexity.

5. Conclusions
In summary, we proposed a novel algorithm that is based on statistical characteristics of speckle
noise and the DWT to remove speckle noise from the SAR images. For this purpose, the SRAD
filtering method, which can be directly applied to the SAR image, is used as a preprocessing filter.
The logarithmic transform is employed to convert the multiplicative noise in the resulting SRAD image
to additive noise. In order to further remove the additive noise from the SRAD filter result image,
the two-level DWT converts the SRAD filter result image into one approximate sub-band image and
six detailed sub-band images. The IGF is applied to diagonal sub-band images, which have lower
Remote Sens. 2019, 11, 1184 25 of 27

energy within the same scale, to remove the additive noise and preserve edge information. Meanwhile,
the horizontal and vertical sub-band images, which exhibit higher energy than the diagonal sub-band
images, are treated with the soft threshold. The GF is applied to remove the additive noise that is present
in the approximate sub-band image. The experiments in this study used both standard images and real
SAR images. The experimental results demonstrate that the proposed method is able to obtain excellent
speckle noise removal and edge preservation at low computational complexity when compared with
the state-of-the art methods. In future research, we aim to study a novel filtering technique that can
remove noise while preserving edge information in the approximate sub-band image.

Author Contributions: H.C. designed the methodology, implemented the simulation, and wrote this paper.
J.J. wrote and edited this paper.
Funding: This research was not funded.
Acknowledgments: The authors would like to thank the reviewers’ for their good suggestions.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Yahya, N.; Karmel, N.S.; Malik, A.S. Subspace-Based Technique for Speckle Noise Reduction in SAR Images.
IEEE Trans. Geosci. Remote Sens. 2014, 52, 6257–6271. [CrossRef]
2. Liu, F.; Wu, J.; Li, L.; Jiao, L.; Hao, H.; Zhang, X. A Hybrid Method of SAR Speckle Reduction Based on
Geometric-Structural Block and Adaptive Neighborhood. IEEE Trans. Geosci. Remote Sens. 2018, 56, 730–747.
[CrossRef]
3. Guo, F.; Zhang, G.; Zhang, Q.; Zhao, R.; Deng, M.; Xu, K. Speckle Suppression by Weighted Euclidean
Distance Anisotropic Diffusion. Remote Sens. 2018, 10, 722. [CrossRef]
4. Yuan, Q.; Zhang, Q.; Li, J.; Shen, H.; Zhang, L. Hyperspectral image denoising employing a spatial-spectral
deep residual convolutional neural network. IEEE Trans. Geosci. Remote Sens. 2018, 57, 1205–1218. [CrossRef]
5. Zhang, Q.; Yuan, Q.; Zeng, C.; Wei, X.; Wei, Y. Missing data reconstruction in remote sensing image with a
unified spatial-temporal-spectral deep convolutional neural network. IEEE Trans. Geosci. Remote Sens. 2018,
56, 4274–4288. [CrossRef]
6. Yuan, Y.; Fang, J.; Lu, X.; Feng, Y. Remote sensing image scene classification using rearranged local features.
IEEE Trans. Geosci. Remote Sens. 2018, 57, 1779–1792. [CrossRef]
7. Tu, B.; Zhang, X.; Kang, X.; Zhang, G.; Li, S. Density peak-based noisy label detection for hyperspectral
image classification. IEEE Trans. Geosci. Remote Sens. 2019, 57, 1573–1584. [CrossRef]
8. Gemme, L.; Dellepiane, S.G. An automatic data-driven method for SAR image segmentation in sea surface
analysis. IEEE Trans. Geosci. Remote Sens. 2018, 56, 2633–2646. [CrossRef]
9. Duan, Y.; Liu, F.; Jiao, L.; Tao, X.; Wu, J.; Shi, C.; Wimmers, M.O. Adaptive hierarchical multinomial latent
model with hybrid kernel function for SAR image semantic segmentation. IEEE Trans. Geosci. Remote Sens.
2018, 56, 5997–6015. [CrossRef]
10. Liu, S.; Wu, G.; Zhang, X.; Zhang, K.; Wang, P.; Li, Y. SAR despeckling via classification-based nonlocal and
local sparse representation. Neurocomputing 2017, 219, 174–185. [CrossRef]
11. Xie, H.; Pierce, L.E.; Ulaby, F.T. Statistical properties of logarithmically transformed speckle. IEEE Trans.
Geosci. Remote Sens. 2002, 40, 721–727. [CrossRef]
12. Barash, D. A fundamental relationship between bilateral filtering, adaptive smoothing and the nonlinear
diffusion equation. IEEE Trans. Pattern Anal. Machine Intell. 2002, 24, 844–867. [CrossRef]
13. Tomasi, C.; Manduchi, R. Bilateral Filtering for Gray and Color Images. In Proceedings of the Sixth
International Conference on Computer Vision, Bombay, India, 4–7 January 1998; pp. 839–846.
14. Buades, A.; Coll, B.; Morel, J.-M. A non-local algorithm for image denoising. In Proceedings of the 2005
IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Diego, CA, USA,
20–25 June 2005; pp. 60–65.
15. Buades, A.; Coll, B.; Morel, J.M. A review of image denoising algorithms, with a new one. SIAM Multiscale
Model. Simul. 2005, 490–530. [CrossRef]
Remote Sens. 2019, 11, 1184 26 of 27

16. Torres, L.; Sant’Anna, S.J.S.; Freitas, C.D.C.; Frery, C. Speckle reduction in polarimetric SAR imagery with
stochastic distances and nonlocal means. Pattern Recognit. 2014, 141–157. [CrossRef]
17. Xu, W.; Tang, C.; Gu, F.; Cheng, J. Combination of oriented partial differential equation and shearlet transform
for denoising in electronic speckle pattern interferometry fringe patters. Appl. Opt. 2017, 56, 2843–2850.
[CrossRef] [PubMed]
18. Perona, P.; Malik, J. Scale-Space and Edge Detection Using Anisotropic Diffusion. IEEE Trans. Pattern Anal.
Mach. Intell. 1990, 12, 629–639. [CrossRef]
19. Li, J.C.; Ma, Z.H.; Peng, Y.X.; Huang, H. Speckle reduction by image entropy anisotropic diffusion. Acta Phys.
Sin. 2013, 62, 099501.
20. Deledalle, C.A.; Denis, L.; Tupin, F. Iterative weighted maximum likelihood denoising with probabilistic
patch-based weights. IEEE Trans. Image Process. 2009, 18, 2661–2672. [CrossRef]
21. Zhang, J.; Lin, G.; Wu, L.; Cheng, Y. Speckle filtering of medical ultrasonic images using wavelet and guided
filter. Ultrasonics 2016, 65, 177–193. [CrossRef] [PubMed]
22. Su, X.; Deledalle, C.; Tupin, F.; Sun, F. Two-step multitemporal nonlocal means for synthetic aperture radar
images. IEEE Trans. Geosci. Remote Sens. 2014, 52, 6181–6196.
23. Chierchia, C.; Mirelle, E.G.; Scarpa, G.; Verdoliva, L. Multitemporal SAR image despeckling based on
block-matching and collaborative filtering. IEEE Trans. Geosci. Remote Sens. 2017, 55, 5467–5480. [CrossRef]
24. Parrilli, S.; Poderico, M.; Angelino, C.V.; Verdoliva, L. A nonlocal SAR image denoising algorithm based on
LLMMSE wavelet shrinkage. IEEE Trans. Geosci. Remote Sens. 2012, 50, 606–616. [CrossRef]
25. Wu, M.-T. Wavelet transform based on Meyer algorithm for image edge and blocking artifact reduction.
Inf. Sci. 2019, 474, 125–135. [CrossRef]
26. Singh, R.; Khare, A. Fusion of multimodal medical images using Daubechies complex wavelet
transform—A multiresolution approach. Inform. Fusion 2014, 19, 49–60. [CrossRef]
27. Li, H.; Manjunath, B.S.; Mitra, S.K. Multisensor image fusion using the wavelet transform. Graph. Models
Image Process. 1995, 57, 235–245. [CrossRef]
28. Pajares, G.; de la Cruz, J.M. A wavelet-based image fusion tutorial. Pattern Recognit. 2004, 37, 1855–1872.
[CrossRef]
29. Huang, Z.-H.; Li, W.-J.; Wang, J.; Zhang, T. Face recognition based on pixel-level and feature-level fusion of
the top-level’s wavelet sub-bands. Inf. Fusion 2015, 22, 95–104. [CrossRef]
30. Hsia, C.-H.; Guo, J.-M. Efficient modified directional lifting-based discrete wavelet transform for moving
object detection. Signal Process. 2014, 96, 138–152. [CrossRef]
31. Liu, S.; Florencio, D.; Li, W.; Zhao, Y.; Cook, C. A Fusion Framework for Camouflaged Moving Foreground
Detection in the Wavelet Domain. IEEE Trans. Image Process. 2018, 27, 3918–3930. [CrossRef]
32. Donoho, D.L. De-noising by soft-thresholding. IEEE Trans. Inf. Theory 1995, 25, 613–627. [CrossRef]
33. Donoho, D.L.; Johnstone, I.M. Adapting to unknown smoothness via wavelet shrinkage. J. Am. Stat. Assoc.
1995, 90, 1200–1224. [CrossRef]
34. Chang, S.G.; Yu, B.; Vetterli, M. Adaptive wavelet thresholding for image denoising and compression.
IEEE Trans. Image Process. 2000, 9, 1532–1546. [CrossRef]
35. Yang, Y.; Ding, Z.; Liu, J.; Gao, Q.; Yuan, X.; Lu, X. An adaptive SAR image speckle noise algorithm based on
wavelet transform and diffusion equations for marine scenes. In Proceedings of the 2017 IEEE International
Geoscience and Remote Sensing Symposium, Fort Worth, TX, USA, 23–28 July 2017; pp. 1–4.
36. Amini, M.; Ahmad, M.O.; Swamy, M.N.S. SAR image despeckling using vector-based hidden markov model
in wavelet domain. In Proceedings of the 2016 IEEE Canadian Conference on Electrical and Computer
Engineering, Vancouver, BC, Canada, 15–18 May 2016; pp. 1–4.
37. Li, H.-C.; Hong, W.H.; Wu, Y.-R.; Fan, P.-Z. Bayesian Wavelet Shrinkage with Heterogeneity-Adaptive
Threshold for SAR images despeckling based on Generalized Gamma Distribution. IEEE Trans. Geosci.
Remote Sens. 2013, 51, 2388–2402. [CrossRef]
38. Rajesh, M.R.; Mridula, S.; Mohanan, P. Speckle Noise Reduction in Images using Wiener Filtering and
Adaptive Wavelet Thresholding. In Proceedings of the 2016 IEEE Region 10 Conference (TENCON),
Singapore, 22–25 November 2016; pp. 2860–2863.
39. Yu, Y.; Acton, S.T. Speckle reducing anisotropic diffusion. IEEE Trans. Image Process. 2002, 11, 1260–1270.
[PubMed]
Remote Sens. 2019, 11, 1184 27 of 27

40. Dass, R. Speckle noise reduction of ultrasound images using BFO cascaded with wiener filter and discrete
wavelet transform in homomorphic region. Procedia Comput. Sci. 2018, 132, 1543–1551. [CrossRef]
41. Singh, P.; Shree, R. A new SAR image despeckling using directional smoothing filter and method noise
thresholding. Eng. Sci. Technol. Int. J. 2019, 21, 589–610. [CrossRef]
42. Choi, H.H.; Lee, J.H.; Kim, S.M.; Park, S.Y. Speckle noise reduction in ultrasound images using a discrete
wavelet transform-based image fusion technique. Biomed. Mater. Eng. 2015, 26, 1587–1597. [CrossRef]
[PubMed]
43. Gao, F.; Xue, X.; Sun, J.; Wang, J.; Zhang, Y. A SAR Image Despeckling Method Based on Two-Dimensional S
Transform Shrinkage. IEEE Trans. Geosci. Remote Sens. 2016, 54, 3025–3034. [CrossRef]
44. Sivaranjania, R.; Roomi, S.M.M.; Senthilarasi, M. Speckle noise removal in SAR images using Multi-Objective
PSO (MOPSO) algorithm. Appl. Soft Comput. 2019, 76, 671–681. [CrossRef]
45. He, K.; Sun, J.; Tang, X. Guided image filtering. IEEE Trans. Pattern Anal. Mach. Intell. 2013, 35, 1397–1409.
[CrossRef]
46. Saevarsson, B.B.; Sveinsson, J.R.; Benediktsson, J.A. Combined Wavelet and Curvelet Denoising of SAR
Images. In Proceedings of the 2007 IEEE International Geoscience and Remote Sensing Symposium, Barcelona,
Spain, 23–28 July 2007; pp. 4235–4238.
47. Zhang, M.; Gunturk, B.K. Multiresolution Bilateral Filtering for Image Denoising. IEEE Trans. Image Process.
2008, 17, 2324–2333. [CrossRef]
48. Sheikh, H.R.; Bovik, A.C.; Cormack, L. No-reference Quality Assessment Using Natural Scene Statistics:
JPEG2000. IEEE Trans. Image Process. 2005, 14, 1918–1927. [CrossRef] [PubMed]
49. Wenxuan, S.; Jie, L.; Minyuan, W. An image denoising method based on multiscale wavelet thresholding and
bilateral filtering. Wuhan Univ. J. Nat. Sci. 2010, 15, 148–152.
50. Frost, V.S.; Stiles, J.A.; Shanmugan, K.S.; Holtzman, J.C. A model for radar images and its application to
adaptive digital filtering of multiplicative noise. IEEE Trans. Pattern Anal. Mach. Intell. 1982, PAMI-4, 66–157.
[CrossRef]
51. Lee, S.T. Digital image enhancement and noise filtering by use of local statistics. IEEE Trans. Pattern Anal.
Mach. Intell. 1980, PAMI-2, 165–168. [CrossRef]
52. Treece, G. The bitonic filter: Linear filtering in an edge-preserving morphological framework. IEEE Trans.
Image Process. 2016, 25, 5199–5211. [CrossRef] [PubMed]
53. Farbman, Z.; Fattal, R.; Lischinski, D.; Szeliski, R. Edge-preserving decomposition for multi-scale tone and
detail manipulation. ACM Trans. Graph. 2008, 27. [CrossRef]
54. Zhu, L.; Fu, C.-W.; Brown, M.S.; Heng, P.-A. A non-local low-rank framework for ultrasound speckle
reduction. In Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition,
Honolulu, HI, USA, 21–26 July 2017; pp. 5650–5658.
55. Ramos-Llordén, G.; Vegas-Sánchez-Ferrero, G.; Martin-Fernández, M.; Alberola-López, C.; Aja-Fernández, S.
Anisotropic diffusion filter with memory based on speckle statistics for ultrasound images. IEEE Trans.
Image Process. 2015, 24, 345–358. [CrossRef]
56. Hyunho, C.; Jechang, J. Speckle noise reduction in ultrasound images using SRAD and guided filter.
In Proceedings of the International Workshop on Advanced Image Technology, Chiang Mai, Thailand,
7–9 January 2018; pp. 1–4.
57. Jet Propulsion Laboratory. Available online: https://photojournal.jpl.nasa.gov/catalog/PIA01763 (accessed on
30 December 2018).
58. Dataset of Standard 512X512 Grayscale Test Images. Available online: http://decsai.ugr.es/cvg/CG/base.htm
(accessed on 30 December 2018).
59. Crow, F. Summed-area tables for texture mapping. In Proceedings of the 11th Annual Conference on
Computer Graphics and Interactive Techniques, New York, NY, USA, 1984; pp. 207–212.
60. Elad, M.; Ahalon, M. Image denoising via learned dictionaries and sparse representation. In Proceedings of
the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, New York, NY,
USA, 17–22 June 2006; pp. 895–900.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).