1

SPECKLE NOISE REDUCTION OF SAR IMAGE USING WAVELET PACKET

TRANSFORM
S. Suganya
1
, T.Veerakumar
2

1, 2 Department of Electronics and Communication Engineering, PSG College of Technology
Peelamedu, Coimbatore-641 004, TamilNadu, India
[email protected], [email protected]

AbstractRemoving noise from the original signal is still a
challenging problem for researchers. Even though so many algorithms
are proposed by different researchers, each algorithm has its
assumptions, advantages and limitations. We propose a new method for
removing speckle noise from SAR images, based on wavelet packet
transform with BayesShrink. Wavelet Packet Transform has rich set of
bases than wavelet. In this paper, the thresholding techniques like
VishuShrink and BayesShrink are applied to the high frequency
subbands of wavelet packets. The performance of this method
substantially surpasses that of previously published methods, both
visually and in terms of PSNR and IEF.

Keywords: Image Denoising, Wavelet Transform, Wavelet Packet
Transform, BayesShrink, VisuShrink.

1. Introduction
Synthetic Aperture Radar (SAR) is playing an increasingly
important role in gathering information from the earths surface.
SAR images are corrupted by speckle noise due to random
interference of electromagnetic waves. Speckle noise [1] is a
common phenomenon in all coherent imaging systems like laser,
acoustic and SAR imagery. Reduction of speckle noise is very
important in real time applications. Numerous filters have been
proposed to reduce the speckle noise, such as local statistical
methods [2], sigma filter [3], median filter [4], homomorphic filter
[5], Lee filter [6], Kuan filter [7] and Frost filter [8], which change
their filtering coefficients adaptively depending on the local
statistics within a fixed sliding window. Although these filters are
known to be efficient in speckle reduction, more or less useful
information in SAR images are damaged. Hence, we have to go for
the frequency domain approach such as wavelet based algorithm
[9].
Wavelet representations are convenient in image denoising
because natural image samples have a varying statistical behavior.
Wavelet Packet Transform (WPT) is more flexible compared to
Wavelet transform (WT). WPT has special abilities to achieve
higher discrimination by analyzing the higher frequency domain of
a signal.
The wavelet packet decomposition is an expansion of the wavelet
decomposition.
But in wavelet packet decomposition each detail coefficient as well
as the approximation coefficient will be further decomposed. The
wavelet packet decomposition will generate a large number of
bases compared to wavelet decomposition. Since the wavelet
packets have a frequential support which decreases exponentially
at high frequencies, WPT can provide more detailed
decomposition for the high frequency signals.
This paper is organized as follows: Section 2 deals with Wavelet
packet transform, the thresholding techniques is described in
Section 3, proposed algorithm is discussed in section 4,
experimental results are listed in Section 5 and Section 6
describes the conclusion.

2. Wavelet Packet Transform
Wavelet Packet Transform was introduced by Coifman,
Meyer and Wickerhauser [10] by generalizing the link between
multiresolution approximations and wavelet functions. Wavelet
packet transforms are more adaptive than the wavelet transform
because they offer a rich library of bases from which the best
one can be chosen for a certain class of images. A space Aj of a
multiresolution approximation is decomposed in a lower
resolution space Aj+1 plus a detail space Dj+1. This is done by
dividing the orthogonal basis
Z n )} n 2 t ( {
j
j
e o
of Aj into
two new orthogonal bases
Z n )} n 2 t ( {
1 j
1 j
e o
+
+
of Aj+1
and
Z n )} n 2 t ( {
1 j
1 j
e v
+
+
of Dj+1.
Here, both the approximation and detail spaces are divided
further. The recursive splitting of vector spaces is represented
in a binary tree [11]. The binary tree of wavelet packet
decomposition is shown in figure 1.


Figure 1 Binary tree of wavelet packet spaces

3. Thresholding Techniques
Thresholding [12] is a simple non-linear technique, which
operates on one wavelet packet coefficient at a time. The choice
of a threshold is an important point of interest. It plays a major
role in the removal of noise in images because denoising most
0
L
D

0
1 L
D
+
D
1
1 L
D
+
D
0
2 L
D
+
D
1
2 L
D
+
D
3
2 L
D
+
D
2
2 L
D
+
D
0
L
A

0
1 + L
A
D
1
1 + L
A
D
0
2 + L
A
D
1
2 + L
A
D
3
2 + L
A
D
2
2 + L
A
D
X

2
frequently produces smoothed images, reducing the sharpness of
the image. Wavelet packet coefficient is compared to a given
threshold and is set to zero if its magnitude is less than the
threshold; otherwise, it is kept or modified (depending on the
thresholding rule). The threshold acts as an oracle which
distinguishes between the insignificant coefficients likely due to
noise, and the significant coefficients consisting of important
signal structures. Thresholding rules are especially effective for
signals with sparse or near-sparse representations where only a
small subset of the coefficients represents all or most of the signal
energy. Soft thresholding is a method used to modify the noisy
wavelet coefficients. It shrinks the coefficients above the threshold
in absolute value. Thresholding essentially creates a region around
zero where the coefficients are considered negligible. Outside of
this region, the thresholded coefficients are kept to full precision.
Hard thresholding may seem to be natural; the continuity of soft
thresholding has some advantages. It makes algorithms
mathematically more tractable. Their most well-known
thresholding are VisuShrink[13], and BayesShrink[14].

A. VisuShrink
For image denoising, VisuShrink is known to yield overly
smoothed images. This is because of its threshold choice. It can be
viewed as general purpose threshold selectors that exhibit near
optimal min max error properties and ensures with high probability
that the estimates are as smooth as the true underlying functions.
The VisuShrink threshold T is given in (1)
2log T M o = (1)
where is the noise variance and M is the number of pixels in the
image. VisuShrink follows the global thresholding scheme where
there is a single value of threshold applied globally to all the
wavelet coefficients.

B. BayesShrink
In this paper, we removed the speckle noise of the WPT subbands
with the help of the ideas and methods of BayesShrink soft-
thresholding [9].
Let us consider the image as follows
n x f + = (2)
where x is the original image, n is the noise and they are
independent of each other, hence
2
n
2
x
2
f
o o o + = (3)
where
2
n
o is the variance of the noise, if the image noise could
not know in advance, can be used to estimate the noise variance
with median estimator. The variance of noise is

6745 . 0
HH Median
~
n
= o

(4)
where HH is the wavelet packet coefficients of the high frequency
subband. The variance of the wavelet packet subband is calculated
by the equation is

=
=
mn
1 j , i
2
j , i
2
f
WP
mn
1
~
o (5)
where mn is the size of the subband under consideration. WPi,j
is the wavelet packet subband.The original image variance
(
2
x
~
o ) is estimated by the formula is ) 0 ,
~ ~
max(
~ 2
n
2
f
2
x
o o o =

(6)

We refer to our method as Bayesshrink which perform soft
thresholding, wavelet packet subbands dependent threshold (T),
which is given by
x
2
n
~
~
T
o
o
= .

4. Proposed Methodology
The flow chart of the proposed algorithm is shown in figure
2.



Figure 2 Flow chart of Proposed Algorithm

The noise variance is obtained from the high frequency
subband of first level wavelet packet decomposition. Using the
noise variance and the variance of each subband, calculate the
approximated variance original image subbands.
From the approximated variance and noise variance find the
threshold for each subbands. The obtained threshold value is
compared with each subband coefficients and applies the soft
thresholding algorithm.

Get the Noisy image
Apply the Wavelet Packet Transform
Find the Noise variance, (
2
n
o )
Find the Threshold value (T) using VishuShrink and
BayesShrink method
Apply Soft-thresholding for the high frequency WPT
subbands
Take Inverse Wavelet Packet Transform
Denoised image
3
Inverse WPT is then applied to the soft threshold coefficients.
Finally, the performance metrics like Peak Signal to Noise Ratio
(PSNR) and Image Enhancement Factor (IEF) are calculated.

5. Simulation Results and Discussion
The algorithm is tested with the SAR images of size 256 256.
Peak signal-to-noise ratio (PSNR) and IEF are used to estimate the
performance of the denoised image. The PSNR and IEF values are
calculated from the equation (7) and (8) respectively.
|
|
.
|

\
|
=
MSE
255
log 20 dB in PSNR
10

(7)

= =
= =

=
N
1 i
N
1 j
2
N
1 i
N
1 j
2
)) j , i ( x ) j , i ( z (
)) j , i ( x ) j , i ( f (
IEF

(8)
where

= =
=
N
1 i
N
1 j
2
2
)) j , i ( z ) j , i ( x (
N
1
MSE

(9)

where MSE is the Mean Square Error, N2 is the size of the image,
x is the original image and z is the denoised image and f is the
noisy image.

The following table shows the performance analysis by comparing
the PSNR values for wavelet and wavelet packet for the
thresholding techniques namely BayesShrink and VisuShrink by
using Haar filter for different variance. The level of decomposition
is three. From the table 1, it is obvious that the performance of
wavelet packet is better than the wavelet.

The wavelet packet based methods have drawbacks that the run
time is longer than the wavelet based method because the wavelet
packet decomposition is more time-consuming. The wavelet packet
based denoising method can achieve the good performance
especially for the SAR images with much detailed information.

Table 1 gives the comparison of the proposed scheme against the
existing scheme for GeoEye1 satellite image of Kutztown
University and the corresponding plot is shown in figure 3. From
the figures 3 and 4, we observe that the proposed WPT based
BayesShrink algorithm gives better result than the other spatial and
frequency domain approach.









Table 1 PSNR and IEF values for WPT and various filters
of GeoEye1 Image
Filters Variance
0.01 0.08 0.16
PSNR
in dB
IEF PSNR
in dB
IEF PSNR
in dB
IEF
Lee 23.03 0.30 22.47 2.00 21.80 3.38
Stick 17.79 0.12 16.81 0.76 16.06 1.25
Kuwahara 24.22 0.40 21.69 1.75 20.09 2.36
Kuwahara
Fast
21.31 0.18 20.00 1.13 19.07 1.79
Median 23.68 0.35 22.00 1.83 20.76 2.68
Frost 25.70 0.53 22.36 2.03 20.31 2.52
WT Vishu 25.05 1.15 21.86 2.32 20.79 3.33
WT Bayes 29.10 0.44 23.16 1.73 21.47 2.69
WPT
Vishu
28.79 0.50 22.10 1.91 20.90 2.88
WPT
Bayes
29.34 1.28 23.32 2.47 21.91 3.41

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
16
18
20
22
24
26
28
30
Variance-->
P
S
N
R

i
n

d
B
-
-
>
PSNR Vs Variance


Lee
Stick
WTVishu
WTBayes
WPTVishu
WPTBayes

Figure 3 Plot for comparing of various thresholding techniques

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
0
0.5
1
1.5
2
2.5
3
3.5
Variance-->
I
E
F
-
-
>
IEF Vs Variance


Lee
Stick
WTVishu
WTBayes
WPTVishu
WPTBayes

Figure 4 IEF values of various thresholding techniques
4
The results for comparison of various filters with the proposed
methodology for GeoEye1 Image are shown in the figure 5.


(a) (b)

(c) (d)

(e) (f)

(g) (h)

(i) (j)


(k) (l)
Figure 5 Original and denoised images
(a)Original image (b)Noisy image ((c) Lee filter(d)Stick filter(e)Kuwahara
filter (f)Kuwahara fast filter(g)Median filter (h)Frost filter(i)WT
BayesShrink(j)WT VisuShrink (k)WPT BayesShrink (l)WPT VisuShrink

Table 2 gives the result of the comparison of various WPT
filters for west concordoorthophoto image and the
corresponding plot is shown in figure 6. From the table, it is
possible to observe that the choice of the wavelet filters affects
the performance of the proposed algorithm.

In this case, the performance of La8 filter is better than other
wavelet filters. La8 wavelet coefficients is proportional to
change between avarage over scale and its surroundings, the
change measured by length of filter by 2 first differences and
the average is localiized weighted average.

But in Daubhechies filter wavelet coeffients are proportional to
difference between localized weighted average and its
surroundings. Hence, the performance of La8 better than the
other fiters.

Table 2 PSNR values for WPT filters of West
concordoorthophoto Image

Filters
PSNR in dB
Noise Variance
0.01 0.02 0.04 0.08 0.1 0.12 0.16
Haar 26.25 24.24 22.40 20.70 20.16 19.50 19.02
D4 26.80 24.81 22.91 21.13 20.74 20.28 19.70
Bi9 27.00 24.75 22.67 20.73 20.17 19.43 18.83
Bi5 24.48 23.01 21.21 19.24 18.59 18.08 18.67
La8 27.08 25.13 23.33 23.33 21.49 20.94 19.77


5

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
18
19
20
21
22
23
24
25
26
27
28
Wavelet Filters-->
P
S
N
R

i
n

d
B
-
-
>
PSNR Vs Wavelet Filters


Haar
Db2
Bi9/7
Bi5/3
La8

Figure 6 Plot for comparing of WPT filters for West concordoorthophoto

The denoised image obtained using the proposed method and the
existing methods are given in figure 7. From the figure, it shows
that the visual quality of the proposed method better than the
existing methods.


(a) (b)

(c) (d)

(e) (f)

(g) (h)

(i) (j)

(k) (l)
Figure 7 Original and denoised West concordorthophoto images (a) Original
image (b) Noisy image (c) Lee filter (d) Stick filter (e) Kuwahara filter (f)
Kuwahara fast filter (g) Median filter (h) Frost filter (i) WT BayesShrink (j)
WT VisuShrink (k) WPT BayesShrink (l) WPT VisuShrink


6. Conclusion
In this paper, removal of speckle noise in SAR image
using wavelet packet transform by incorporating BayesShrink
and VisuShrink thresholding techniques has been presented.
BayesShrink performs better than VisuShrink in terms of
PSNR. The resultant SAR image using BayesShrink gives
smoother results and more visually appealing than one obtained
by VisuShrink. The denoising of the image using wavelet
packet achieves better performance than wavelet denoising.

References
[1] Goodman J W, Some fundamental properties of speckle, J. Opr.Soc. Am.,
66 (1976) 1145).
[2] Lee J S, Speckle analysis and smoothing of synthetic aperture radar
images, Comput. Graphics Image Graphics, 17 (1981) 32).
[3] Lee J S, A simple speckle smoothing algorithm for synthetic aperture radar
images, IEEE Trans. Systems Man Cybernet., 13 (1983) 89).
[4] Lim J S, Two-Dimensional Signal and Image Processing, (Prentice-Hall,
Englewood Cliffs, New Jersey), 1990 ,476.
[5] Franceschetti, G & Pascazio V, Iterative homomorphic technique for
speckle reduction in synthetic-aperture radar imaging, J. Opt. Soc. Amer., 12
(1995) 686).
6
[6] Lee S, Digital image enhancement and noise filtering by use of local statistics,
IEEE Trans. Pattern Anal. Machine Intell., 2 (1980) 165).
[7] Kuan D T, Sawchuk A A, Strand T C & Chavel P, Adaptive noise smoothing
filter for images with signal dependent noise,IEEE Trans. Pattern Anal. Machine
Intell., 2 (1985) 165).
[8] Frost V S, Stiles I A, et al, A model for radar images and its application to
adaptive digital filtering of multiplicative noise, IEEE Trans. Pattern Anal.
Machine Intell., 4 (1982) 157).
[9] Chi M, Lu G, & Huang S Z, De-Noising the Speckle Noises of the SAR Image
via the Wavelet Soft-Thresholding Method, Journal of Xiamen University (Nature
Science), 41 (2002) 756).
[10] Xiong Z, Ramchandran K & Orchard M T, Wavelet packets coding using
space-frequency quantization, IEEE Trans. Image Processing, 7 (1998) 892).
[11] Mallat S, A Wavelet tour of signal processing, {Academic press, New York),
1999}.
[12] Donoho D L, De-Noising by Soft Thresholding, IEEE Trans. Info. Theory
43 (1993) 933).
[13] Donoho D L & Johnstone I M , Ideal spatial adaptation via wavelet shrinkage,
Biometrika, 81 (1994) 425).
[14] Chang S, Yu B & Vetterli M, Adaptive Wavelet Thresholding for Image
denoising and Compression, IEEE Trans. Image Processing, 9 (2000) 1532).