ISSN : 0976-8491(Online) | ISSN : 2229-4333(Print) IJCST Vol.

2, SP 1, December 2011

Wavelet Based Image De-noising of Non Logarithmic

Transformed Data.
1
T Sreekanth Rao, 1P Gangamohan, 2P Nagarjuna Reddy, 3B Prathyusha
Dept. of Signal Processing, School of Engineering,
1

Blekinge Institute of Technology, SE, Karlskrona, Sweden

2
Dept. of Electrical Engineering, Kakatiya Institute of Technology & Science, Warangal, AP, India
3
Dept. of ECE, Malla Reddy College of Engineering & Technology, Hyderabad, AP, India

Abstract Our research question is, could we design a technique to

Enhancement of speckled image based on orthogonal wavelet
transform can keep the image details intact. Although, de-noising
techniques such as a soft-thresholding and hard-thresholding
are simple in implementation, they require multilevel wavelet
decomposition. However, the Lee filtering technique requires
only a single level decomposition, logarithmically transformed
noisy image data is used for speckle reduction. In this paper,
we propose an efficient algorithm for speckle reduction without
using logarithmic transform of noisy image data and also which
requires only single level decomposition.
enhance the speckled image in wavelet domain without using
logarithmic transform of input data and also which requires
only a single level decomposition? We put forward a hypothesis
that the problem can be solved by dividing the sub-bands into
grid of blocks and modifying them, the goal is to improve Signal
to Noise Ratio (SNR) and obtain minimum MSE by keeping ET
intact. The main contribution of this paper is to model and
implement the technique in both Discrete Wavelet Transform
(DWT) and DD-DWT domain. The results are validated in MAT-
LAB.

Keywords IV. Problem Solution

Discrete Wavelet Transform, Double Density Discrete Wavelet
Transform, Threshold Techniques, soft-thresholding, hard- A. Discrete Wavelet Transform (DWT)
thresholding, orthogonal wavelet transform. DWT is a set of orthogonal iterative filter banks called wavelets.
The procedure of wavelet decomposition consists of consecutive
I. Introduction operation of row wise and column wise of input image data
Synthetic Aperture Radar (SAR) can collect information about followed by down sampling to obtain wavelet coefficients and
the parts of the world that are perpetually cloud covered. Due also called sub bands [2]. At one level decomposition of input
to its coherent nature, a speckle is originated. The property data DWT comprises of four sub bands LL (low/low), LH ,HL,
of speckle noise is multiplicative in amplitude and additive in HH(high/high) and DD-DWT comprises of LL, LH1, LH2, HL1,
phase. Reduction of speckle is a challenge because a noise HL2, HL3, HH1, HH2, HH illustrated in fig. 1.The LL sub bands
removal introduces blurring of image [3]. Double Density are used for multilevel decomposition [1].
Discrete Wavelet Transform (DD-DWT) is designed from scaling
and wavelet filter banks having property of shift invariant over
decimated Wavelet Transform [1]. Presently, the research is
focused on the de-noising techniques which require a single
level decomposition, as they take less Execution Time (ET)
and preserve image information. The aim of this paper is to
develop an efficient technique which requires only a single
level decomposition.

II. Review of the State of Art

The soft-threshold and hard-threshold de-noising techniques
are implemented on multilevel decomposed wavelet sub-bands
to obtain smoothed image [4]. These de-noising techniques
employ straightforward nonlinear thresholding of the noisy
wavelet coefficient [2]. The basic idea of Lee filter technique Fig. 1: Block diagram of DD-DWT and DWT.
is to transform the multiplicative model of speckle to additive
Gaussian model by applying logarithmic transform and B. Soft Thresholding and Hard Thresholding
then the speckle reduction is to be applied at the first level Donoho proposed Soft-thresholding and Hard-thresholding on
decomposition [3]. sub band coefficients. These techniques are used for multilevel
decomposed sub bands [3]. Choose Threshold T where
III. Problem Statement and Main Contribution σ is standard deviation of HH sub band.
In threshold techniques the threshold point is to be known for Soft-thresholding is given as:
minimum Mean Square Error (MSE) and requires multilevel
decomposition, which takes more ET and returns the smoothed
image where the edges of the image are lost. Logarithmic (1)
transform of input data in Lee filtering causes the loss of image
information for small back scattered values of SAR image [3].

w w w. i j c s t. c o m   International Journal of Computer Science & Technology  213

IJCST Vol. 2, SP 1, December 2011 ISSN : 0976-8491(Online) | ISSN : 2229-4333(Print)

Hard-thresholding is given as: ac tual S pec kled image

(2)

100 100

C. Lee Filtering
The multiplicative model is firstly converted to Additive Gaussian 200 200
model by applying logarithmic transform of input data and then
the speckle reduction at one level of decomposition [3]. 300 300

D. Proposed Algorithm 400 400

The procedure of Wavelet decomposition consists of consecutive 100 200 300 400 100 200 300 400
operation of row wise and column wise of input image data Fig. 3 : Actual image (left), speckle infected image (right).
followed by down sampling to obtain Wavelet coefficients and
also called sub bands [2].When the noisy image is single level The images after applying de-noising techniques in DWT
decomposed, DWT comprises of four sub-bands: LL (low/ domain are shown in fig. 3 and images after applying de-noising
low), LH, HL, HH (high/high) and DD-DWT comprises of eight techniques in DD-DWT domain are shown in fig. 4.
sub-bands: LL, LH1, LH2, HL1, HL2, HL3, HH1, HH2, HH [2]. S oft Thres hold DW T Hard Thres hold DW T
Although, the LL and HH sub-bands for DWT and DD-DWT are
not similar, we use the same abbreviations which make the
description of this proposed algorithm easy. All sub-bands 100 100
except LL are divided into grid of blocks shown in fig. 1.
200 200

300 300

400 400
100 200 300 400 100 200 300 400
Lee F ilter DW T P ropos ed DW T
Fig. 2: Block diagram of the comparison of block of sub-
bands for DWT and DD-DWT.
100 100
We set the variation factor for each block of HH sub-band. The
variation factor of block is given as:
200 200

where is a value greater than 1 to be chosen, and 300 300

are the standard deviation and variation factor of block
of HH sub-band respectively. The standard deviation of each 400 400
block of all sub-bands is compared to the variation factor of 100 200 300 400 100 200 300 400
corresponding block in HH sub-band. If the standard deviation Fig. 4 : I m a g e s o f d i f fe r e n t d e - n o i s i n g
of a block is less than variation factor, each pixel in that block techniques applied on DWT sub-bands.
is replaced by its pixel mean value. Otherwise, the block is kept S oft Thres hold DD-DW T Hard Thres hold DD-DW T
unchanged as illustrated in Eq. (2). The algorithm for block
is given as:
100 100

200 200
where represents the function on block , is a
standard deviation of block of sub-band which is compared to 300 300
that of HH sub-band. This procedure is done to all sub-bands at
one level of decomposition. The obtained sub-bands are further 400 400
100 200 300 400 100 200 300 400
smoothed by replacing a pixel by the mean of its adjacent pixels. Lee F ilter DD-DW T P ropos ed DD-DW T
The inverse Wavelet Transform is then applied on sub-bands
50
to obtain de-noised image. The setup of this experiment is
100 100
implemented in MAT-LAB, with an image of dimension 400 x
150
400 pixels, 8-bit pixel intensity image with 256 levels of gray.
200 200
The speckle image is generated using function: imnoise(cricket.
250
jpg,’s’,v), the value of v, which represents variance of speckle 300 300
noise is set at 0.8. The image is enhanced by using proposed
350
algorithm in both DD-DWT and DWT domains by taking 400 400
and each block in sub-bands of dimension 8 x 8 pixels. The 100 200 300 400 100 200 300 400

values of SNR, MSE and ET are computed and compared with Fig. 5 : Images of different de-noising techniques applied on
that of existing de-noising techniques formulated in Table I. DD-DWT sub-bands.

214 International Journal of Computer Science & Technology w w w. i j c s t. c o m

ISSN : 0976-8491(Online) | ISSN : 2229-4333(Print) IJCST Vol. 2, SP 1, December 2011

34
MS E vs Threshold P oint for S oft Threshold
34
MS E vs Threshold P oint for Hard Threshold
[3] Guozhong Chen, Xingzhao Liu, “An Improved Wavelet-
DD-DW T DD-DWT based Method for SAR Images Denoising Using Data
DWT
32 DW T 32
Fusion Technique,” IEEE Conference on Radar, pp. 4, Apr.
30 30 2006
[4] D. L. Donoho, “De-noising by Soft-thresholding,” IEEE
28
28
Trans. Information Theory, Vol. 41, No. 3, pp. 613-627,
May. 1995.

MSE
26 26
MSE

24
24
Sreekanth Rao Tangellapally born in
22 22 Warangal,India in 1987. He completed
his Bachelor degree in Electronics and
20
20
Instrumentation Engineering from Kakatiya
18 18
0 10 20 30 40 50 60 70 80
University, Warangal. He is currently undergoing
0 10 20 30 40 50 60
Threshold P oint Threshold P oint a Master’s programme in Electrical Engineering,
Fig. 6 : Graph of MSE vs. Threshold point for Soft-threshold with emphasis on Signal Processing at Blekinge
(left), MSE vs. Threshold point for Hard-threshold (right) in both Institute of Technology, Karlskrona, Sweden. His areas of
DWT and DD-DWT domain interests are Wireless Communications and Bio-medical Signal
Processing.
In Threshold techniques - Threshold point to be known for
minimum MSE plotted in Fig. 5 and requires multilevel
decomposition. Logarithmic transform of input data in Lee Paidi Ganga Mohan born in Nizamabad, India
Filtering. By considering these two points the proposed algorithm in 1986. He completed his Bachelor degree
is modeled and maintains good SNR and minimum MSE. in Electronics & Communication Engineering
By comparing the images in Fig. 3 and Fig. 4 the de-noised from Jawaharlal Technological University
image with the proposed algorithm in DD-DWT domain shows (JNTU) Hyderabad. He is currently pursuing a
good resolution. Master’s programme in Electrical Engineering,
with emphasis on Signal Processing at Blekinge
Table 1: Comparison of SNR, MSE and ET values of de-noising Institute of Technology, Karlskrona, Sweden.
techniques.
De-noising SNR in dB MSE ET
Techniques DD-DWT DWT DD-DWT DWT P Nagarjuna Reddy born in Kurnool, India
in 1983. He completed his Bachelor degree
Soft- 35.30 4.78 19.18 21.59 7.53 in Electrical & Electronics Engineering from
thresholding Jawaharlal Technological University (JNTU)
Hard- 35.15 4.76 19.83 21.68 7.45 Hyderabad and Master’s programme in
thresholding Electrical Engineering at JNTU, Kakinada.
Lee Filter 35.31 4.65 19.11 22.23 .31 His areas of interests are High Voltage
Proposed 35.73 17.36 20.85 Engineering and Signal Processing.
algorithm 34.93 2.69

B Prathyusha born in Warangal,India in

V. Conclusion 1987. She completed her Bachelor degree in
The proposed method for enhancement of SAR images Electronics and Communication Engineering
maintains good SNR and minimum MSE than Soft-thresholding, from Jawaharlal Technological University
Hard-thresholding and Lee Filter techniques shown in Table I. (JNTU). She is currently undergoing a Master’s
De-noising techniques applied in DD-DWT domain gives better programme in Signal Processing at MRCET, JNTU
results over DWT domain. Hyderabad. Her areas of interests are Speech
The proposed algorithm can be used as application Signal Processing and Bio-medical Signal Processing.
for fast image enhancement and by selecting the appropriate
block size or creating adaptive block size algorithm in image
sub-bands could be the future work.

References
[1] A. Jayawardena, “Design of Double Density Wavelet Filter
Banks,” Seventh International Symposium on Signal
Processing and Its Applications, Vol. 2, pp. 463-466, Jul.
2003.
[2] Yu-long, Chun-Yan Song, Chun-Hui Zhao, “Double-Density
Discrete Wavelet Transform Based Texture Classification”,
Third International Conference on Intelligent Information
Hiding and Multimedia Signal Processing, Vol. 2, pp. 91-
94, Nov. 2007.

w w w. i j c s t. c o m   International Journal of Computer Science & Technology  215