Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258

International Journal of Advance Research and Innovation

DT-CWT Based Block Nonlinear De-noising

Sushil Kumar
Department of Mathematics, Rajdhani College, University of Delhi, New Delhi, India
Abstract
For image denoising, Block thresholding is considered to be a better
Article Info
strategy than the term-by-term thresholding. There are number of Wavelet
Article history: based thresholding techniques for image denoising such as VisuShrink,
Received 02 April 2016 BayesShrink, SureShrink, NeighShrink, BiShrink, ProbShrink, Sure-LET
Received in revised form and BlockShrink. Selesnick has extended the BiShrink method to DT-
20 May 2016 CWT and he has shown that the DT-CWT achieves better results than
Accepted 28 May 2016 DWT for image denoising. Dengwen and Xiaoliu have shown that
Available online 15 June 2016 BlockShrink enjoys a number of advantages over the other conventional
Keywords image denoising methods. Their experimental results show that
DT-CWT, Wavelet Thresholding, BlockShrink outperforms significantly classic SureShrink method and
NeighShrink method. In this paper we extend BlockShrink approach to
BiShrink, BlockShrink, PSNR
DT-CWT and compare it with the BiShrink approach proposed by
Selesnick. We have analysed these methods of noise removal from
degraded images with Gaussian noise and compared the results in term of
PSNR.

1. Introduction
Image denoising is used to recover the image X from its
noisy image, Y given by Y= X Θ є, where є is the that the first method usually involves more computation not
contaminating noise and Θ is any composition of additive necessarily resulting in better performance. Balster et al. [1]
noise and/or multiplicative noise. Wavelet thresholding (or have shown that the selective wavelet shrinkage method
"shrinkage") is one of the techniques used for image which either selects or rejects a wavelet coefficient is
deonoising. We first apply the wavelet transform, T, to the statistically better than the probabilistic method because the
noisy image Y, and then apply a nonlinear estimation former can identify a narrow interval for the estimated
operator, D to the wavelet coefficients of high frequency parameter, which is used to adjust the wavelet coefficient,
sub-bands, either individually or in a group of coefficients. with a higher confidence level than the latter.
It has been shown that the reduction of absolute value in Some popular wavelet shrinkage methods are VisuShrink
wavelet coefficients is successful in signal restoration [19]. [13], BayesShrink [5], SureShrink[17], NeighShrink [6],
Finally, compute the inverse transform, T-1 to get an BivariateShrink[25], ProbShrink [19], Sure-LET [3] and
estimated image, X. In other words, X = T-1 D (TY). BlockShrink[9].
Two well-known shrinkage methods are hard In the VisuShrink method, the wavelet coefficients are
thresholding and Soft thresholding [12]. Hard thresholding shrinked according to the soft-shrinkage rule using the
consists of setting to zero all wavelet coefficients whose universal Threshold, VisuShrink shows better denoising
magnitude is less than a threshold value whereas in Soft performance than the Unversal threshol, but it yield an
thresholding, the wavelet coefficients above the threshold overly smoothed images because the universal threshold T
are shrunk toward the origin. In practice, hard thresholding is high if the number of pixels in the image are high. Just
is preferred to soft thresholding, since for soft thresholding like VisuShrink, SureShrink also applies the soft shrinkage
even large coefficients lying out of noise can shrunk and rule, but it uses independently chosen thresholds for each
hence creates undesirable bias [27]. subband through the minimization of the Stein’s unbiased
In fact, there are two basic approaches to modifying the risk estimate (SURE) [26]. SureShrink performs better than
coefficient, namely, probabilistic wavelet shrinkage and VisuShrink, producing more detailed images.[10]. In Bayes
selective wavelet shrinkage. In the first method, the shrink, thresholding is done at each sub band in the wavelet
decomposition which improves outcome and also
magnitude of the wavelet coefficient is reduced by the
completely denoise the flat regions of the image. But it is
probability of its contribution to the overall quality of the less sensitive to the noise around edges [11].
image. The second method uses a binary method where the In the NeighShrink approach, the wavelet coefficients are
reduction of coefficient magnitude is either 0 or 1, i.e. shrunk in overlapping blocks rather than individually or
coefficients are either selected or removed [1]. It is noted term by term as VisuShrink or SureShrinK. It is observed
that NeighShrink outperforms VisuShrink and SureShrink.
Corresponding Author, BiShrink [24] uses a bivariate shrinkage function taking into
E-mail address: [email protected] account the intrascale variability of wavelet coefficients by
Phone No--+91-971-123-4705 capturing the dependence between a wavelet coefficient and
All rights reserved: http://www.ijari.org its parent. ProbShrink [20] estimate the probability that a

333
IJARI
 Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258
International Journal of Advance Research and Innovation
given coefficient contains a significant noise-free image. We analyse these methods of noise removal from
component. SURE-LET [18] directly parameterizes the degraded images with Gaussian noise and compare the
deonoising, process as a sum of elementary nonlinear results in term of PSNR.
In the subsequent section 2, we summarize the features
processes with unknown weights. It need not hypothesize a
of Dual Tree Complex Wavelet Transform. In Section 3, we
statistical model for the noiseless image while it minimizes give the experimental results by implementing the proposed
an estimate of the mean squared error between the noiseless image denoising algorithms using Matlab 7.0 on different
image and the denoised one by the SURE. Consequently, it grey scale image formats of size 512 x 512.
computes the unknown weights by solving a linear system
2. Dual Tree Complex Wavelet Transform
of equations. Bi Shrink, Prob Shrink and SURE-LET
methods have all been devised for both redundant and non Dual-Tree complex wavelet transform ([21],[14]) uses two
redundant wavelet transforms real DWT trees to implement its real part and imaginary
part, separately that result in decomposition with a much
higher degree of directionality than that possessed by the
traditional DWT. There are two versions of the 2D dual-tree
wavelet transform: the real 2-D dual-tree DWT is 2-times
Probabilistic
expansive, while the complex 2-D dual-tree DWT is 4-times
expansive. Both types have wavelets oriented in six distinct
Wavelet Shrinkage
directions. The real 2-D dual-tree DWT of an image is
implemented using two critically-sampled separable 2-D
Selective
DWTs in parallel. Then for each pair of sub-bands we take
the sum and difference. The six wavelets associated with the
real 2D dual-tree DWT are strongly oriented in {+15o, +45o,
+75o, -15o, -450, -75o} direction (see Fig. 2). The complex
Fig. 1 Block diagram of Wavelet Shrinkage methods 2-D dual-tree DWT also gives rise to wavelets in six distinct
directions, however, in this case there are two wavelets in
Block Shrink utilizes the pertinence of the neighbour each direction ( as shown in Fig. 3). In each direction, one
wavelet coefficients by using the block thresholding
of the two wavelets can be interpreted as the real part of a
scheme. It can decide the optimal block size and threshold
for every wavelet subband by minimizing Stein’s unbiased complex-valued 2D wavelet, while the other wavelet can be
risk estimate (SURE). The block thresholding interpreted as the imaginary part of a complex-valued 2D
simultaneously keeps or kills all the coefficients in groups wavelet. Because the complex version has twice as many
rather than individually, enjoys a number of advantages wavelets as the real version of the transform, the complex
over the conventional term-by-term thresholding. The block version is 4-times expansive.
thresholding increases the estimation precision by utilizing
the information about the neighbor wavelet coefficients. It
outperforms the classic SureShrink and NeighShrink [9],
[11].
Dixit and Sharma [11] have observed that recently
proposed wavelet methods like ProbShrink, BlockShrink
and NeighShrink Sure produce better visual images.
Lal et al. [16] have compared performances of Separable
DT-DWT, Real DT-DWT, Complex DT-DWT, RealDDT-
DWT and Complex DDDT-DWT and observed that
Complex DDDT-DWT outperforms the other
wavelettransforms and is effective for the very highly
corrupted images.
Chinnarao and Madhavilatha [7] have proposed a
conatextual information based thresholding method in DT-
CWT and observed that their method is highly suitable at
high noise levels as compared to low noise levels.
In this paper, we extend the BlockShrink approach to
DT-CWT and compare it with the BiShrink approach
proposed by Selesnick. Since it is noted that many of the Fig. 2 Frequency domain partition in DT-CWT resulting
wavelet based denoising algorithms are suffering from shift from two level decomposition [14, 21]
variance and lack of directionality, Dual Tree Complex
Wavelet Transform (DT-CWT) and complex Double
Density Dual Tree Discrete Wavelet Transform (DDDT-
DWT) have been proposed to decompose the image and
shrinkage operation to eliminate the noise from the noisy

334
IJARI
 Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258
International Journal of Advance Research and Innovation

The package provided by Dengwen [8] contains the Matlab

codes for denoisinig grey scale images using BlockShrink
implemented with a decimated wavelet transform. The main
matlab files provided are:

1. denoisefun.m: This function is the denoising main

function using BlockShrink
2. SubbandThresholding.m: This function is used to
threshold the noisy subband according to
BlockShrink rule
3. parameters.m: This function select the optimal
block size and corresponding threshold
4. Parameters.mexw32: This file is significantly
faster than Parameters.m
5. Calc_MSE_PSNR: for computing MSE and
PSNR
6. Test: This script is for denoising demonstration.
Fig. 3 2-D dual- tree complex wavelet transform[14, 21]
4. Proposed Algorithms
We present the image denoisinig algorithms using Bivariate
Shrink thresholding and BlockShrink thresholding
3. Review of Thresholding Functions
implemented with a dual tree complex wavelet transform.
The matlab codes for Bivariate Shrinkage thresholding
In this section we present the basic steps of two implemented on DT-DWT are provided by Selesnick [23],
thresholding functions used in this paper. These are and for Blockshrink thresholding implemented on DWT are
Bivariate Shrinkage function and Block Shrinkage funciton. provided by Dengwen [8] as discussed in above section.
The basic steps of wavelet based image denoising are:
1. Decompose corrupted image by noise using wavelet The basic steps of these algorithms are given below.
transform.
2. Compute threshold in wavelet domain and apply to Algorithm 1 (BiShrink [23])
noisy coefficients. ………………………………………………….
3. Apply inverse wavelet transform to reconstruct image.
Sender and Selesnick [23] provide the Matlab 1. Read the image, s and resize it to 512 x 512.
implementation of both wavelet-based denoising using the 2. Obtain the noisy image x, given as x = s + g,
separable DWT and wavelet-based denoising using the dual- where g is the additive white Gaussian noise
tree DWT with bivariate shrinkage function. This shrinkage 3. Set the window size (=7).
function requires the prior knowledge of the noise variance 4. Extend the noisy image using symmetric
and the signal variance for each wavelet coefficient. The extension in order to improve the boundary
main matlab files provided for the Bivaraiate Shrinkage usig problem with the function symextend.m.
DT-CWT are: 5. Perform the 2D Dual tree DWT to level J = 6.
1. main_dtdwt.m: It loads the noisy image, calls the using cplxdual2D.m.
denoising routine and calculates the PSNR value 6. Estimate the noise variance. The noise variance
of denoised image. will be calculated using the robust median
2. denoising_dtdwt.m: It call number of subfunctions estimator.
for the calculations of the local adaptive image 7. Process each subband seperately in a loop. First
denoising. the real and imaginary parts of the coefficents and
2a. symextend.m: This function is used to the corresponding parent matrices are prepared for
extend the noisy image using symmetric extension each subband, and the matrices corresponding to
in order to reduce the boundary problem the real and imaginary parts of the parent matrix
2b. cplxdual2D.m: It calculates the forward are expanded using a function expand.m in order
dual-tree DWT to make the matrix size the same as the coefficient
2c. expand.m: The parent matrix is expanded matrix.
using this function in order to make the matrix 8. Estimate the signal variance and the threshold
size the same as the coefficient matrix. value : The signal variance for each coefficent is
2d. bishrink.m: It estimates the magnitude of estimated using the window size , and the
the complex coefficients. The coefficients are threshold value for each coefficient will be
estimated using the magnitudes of the complex calculated and stored in a matrix with the same
coefficient, its parent and the threshold value with size as the coefficent matrix.
this function. 9. Estimate the magnitude of the complex
2e. icplxdual2D.m: It calculates the inverse coefficients. The coefficients will be estimated
wavelet transform using the magnitudes of the complex coefficient,

335
IJARI
 Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258
International Journal of Advance Research and Innovation
its parent and the threshold value with a Matlab new7.tif 37.9214 37.8906
function bishrink.m.
10. Calculate the inverse wavelet transform using new11.tif 35.8946 35.8222
icplxdual2D.m.
11. Extract the image. The neccessary part of the final new12,tif 38.1366 38.2194
image is extracted in order to reverse the
symetrical extension.
shagun.jpg 36.4878 36.1837
………………………………………….
Algorithm 2 (BlockShrink [8]) zoneplate.png 28.0138 32.7194
………………………………………………
1. Read the image s, and resize it to 512 x 512. peppers.jpg 36.9105 37.0584
2. Corrupt the image by additive White Gaussian
Noise. It is given as x = s + g where x is noisy tooth1.jpg 35.1693 35.5138
image corrupted by additive white Gaussian noise
g of standard deviation . Both s and x are of same cameraman.tif 36.8094 36.7104
sizes.
3. Perform the 2D Dual tree DWT to x to level J = 5 New8.tif 34.5610 34.5968
using cplxdual2D.m.
4. Estimate the noise variance, sigma, using robust
median estimator, if it is provided by the user
5. Normalise the noise level of the noisy
Table2. Comparison of BiShrink and BlockShrink image
coefficients.
6. Extract all six detail subbands in each scale. denoising approaches using wavelet thresholding and DT-
7. Apply different threshold values with soft DWT thresholding methods (I1: barbara.png; I2:lena.png;
Thresholding for each detail subband coefficients I3:parliament.bmp; I4:lena.jpg)
with the function ‘SubbandThresholding’.
8. Reconstruct the denoised image by taking the BiSh_ BlSh_ BlSh_
inverse DT-DWT. DWT BiSh_DT DWT DTDWT
…………………………………………………….. DWT
5. Experimental Results
In this section, we give the results obtained on I1 35.5457 33.3797 35.727 33.6762
implementing the proposed method of image denoising, i.e.,
Bivariate Shrink and Block Shrink based on DT-CWT I2 37.7429 35.8798 37.8393 35.1921
thresholding methods on Matlab 7.0. The algorithm is
I3 34.8633 36.3445 34.6432 36.0398
implemented on different grey scale image formats of size
512 x 512, and some of the results are summarized in Table
I4 32.1841 38.4773 32.3865 38.3097
1 and Table 2, respectively. The Table 3 summarizes the
results of proposed algorithms for the brain image: mri_jpe
wtth standard deviation, sigma of additive Gaussian noise
taken from 10 to 100, respectively. The results of the tables Table 3: Results of Proposed algorithms for the brain
image: mri_jpe
1 to 3 are shown through fig. 4 to fig. 6, respectively.
Sigma Bi_ Block
Table 1. PSNR values of denoised images resulting from Shrink
the proposed method based on DT-CWT Shrink

Image Bishrink BlockShrink 10 38.996 39.3364

barbara.png 33.3797 33.6762 40 31.2812 31.5986

lena.png 35.8798 35.1921 80 27.5422 27.9882

c2.bmp 36.3445 36.0390 100 26.4751 26.8314

mri.jpe 38.9919 39.3364

lena.jpg 38.4773 38.3097

336
IJARI
 Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258
International Journal of Advance Research and Innovation
fail to show better results in terms of PSNR for .png image
Bivariate v/s Block Shrinkage using DT-CWT formats. Further, from the Fig. 4, one can observe that DT-
CWT based denoising using Block Shrink function gives
39 better result in terms of PSNR for the image
37 “Zoneplate.png” that contains much more contours than the
35 other images, as compared to DT_CWT based denoising
PSNR

33 Bishrink using Bi-Shrink function. Some of the noised and denoised

31 images obtained from the implementation of proposed
29 algorithms are shown in Fig. 7 below.
27 BlockShri
25 nk
lena.png
c2.bmp

tooth1.jpg
lena.jpg
barbara.png

mri.jpe

New8.tif
new11.tif
new12,tif
shagun.jpg

peppers.jpg

cameraman.tif
new7.tif

zoneplate.png
Bivariate v/s Block Shrinkage
Image: mri.jpe
40 BiShrink
38
36
BlockSh
34
Fig.4 The results of PSNR of proposed method based on rink

PSNR
Bishrink and BlockShrink thresholding methods 32
30

39 28

38 26
37 24
10 40 80 100
36
35 Noise density
34 I1
Fig.6 The results of proposed method based on Bishrink and
33 I2 Block Shrink thresholding methods for the barin image:
32 mri_jpe
31 I3 Noised Image
Denoised Image

30 I4
29

Fig.5 PSNR values obtained of proposed algorithms compared with

corresponding Wavelet based thresholding techniques for four
images Fig.7 (a) The noised and denoised images new11.tif
Denoised Image
Noised Image

From Fig. 4, we observe that overall Block shrinkage

method using DT-DWT performs better than Bivariate
Shrinkage method using DT-CWT for image denoising in
terms of PSNR. In Fig. 6, we observe that even when noise
density is increasing, Block shrink based proposed
algorithm performs better than Bi-Shrink based algorithm in
terms of PSNR. Finally, we compared the proposed
algorithms based on Bi-Shrink function and Block Shrink
function using DT-CWT with the corresponding Wavelet
based thresholding, and summarized the results for some of
the images in Fig. 5. We observe that the proposed Fig.7(b) The noised and denoised images mri.jpe
algorithms perform better for the .bmp and .jpg formats but

337
IJARI
 Volume 4, Issue2 (2016) 333-338 ISSN 2347 - 3258
International Journal of Advance Research and Innovation
6. Conclusions [15] Kumar S. Image Denoising using Wavelet-like
Image denoising is an active and challenging topic of Transform, ICASET, 2016, 21-22
research. The major challenge lies in the fact that one does [16] Lal S., Chandra M., Upadhyay G.K., Gupta D.,
not know what the original signal is for a corrupted image. Removal of Addditive Gaussian noise by complex double
We have presented the DT-CWT based denoising using density dual tree discrete wavelet transform, MIT
BiShrink function and BlockShrink functiion and compared International Journal of Electronics and Communication
the results for different grey scale image formats in terms of Engineering, 1, 1, 2011, 8-16
PSNR. We observe that Block Shrink outperform BiShrink [17] Luisier F., Blu T., Unser M., “A new SURE approach
as well as earlier wavelet domain methods. to image denosing: interscale orthonormal wavelet
thresholding”, IEEE Transactions on Image Processing, 16
Acknowledgement (3), 2007, 593-606
We wish to thank I. W. Selesnick and Zhou Dengwen for [18] Luisier F., Blu T., Unser M. SURE-LET Matlab Codes.
2007b.<http://bigwww.epfl.ch/demo/suredenoising/matlab.h
providing their Matlab software for the image denoising for
tml>.
the purpose of research to us. [19] Piˇzurica A. and Philips W. Estimating the probability
of the presence of a signal of interest in multi resolution
References single and multiband image denoising. IEEE Trans. Image
Process. 15, 3, 2006, 654–665
[1] Balster Eric J., Zheng Y. F., and Ewing R. L., Feature-
[20] Pizzeria A., Philips W. ProbShrink Matlab Codes, 2006
Based Wavelet Shrinkage Algorithm for Image Denoising,
b. Located at the URL: <ttp:// telin.ugent.be/~sanja>.
IEEE Transactions on Image Processing, 14,(12), 2005
[21] Selesnick I. W., The design of approximate Hilbert
[2] Bhonsle D., Dewangan S., “Comparative Study of Dual-
transform pairs of wavelet bases, IEEE Trans. Signal
Tree Complex Wavelet Transform and Double Density
Processing, 5(50), 2002, 1144 - 1152
Complex Wavelet Transform for Image Denoising Using
[22] Sendur L. Selesnick I. W. Bivariate shrinkage functions
Wavelet-Domain”, International Journal of Scientific and
for wavelet-based denoising exploiting interscale
Research Publications, 2, 7, 2012
dependency,” IEEE Trans. Signal Processing, 50,11, 2002,
[3] Blu T., Luisier F., “The SURE-LET approach to image
2744–2756
denoising”, IEEE Trans. Image Process., 16, 2007,2778–
[23] Sendur L. Selesnick I.W. DT-CWT and BiShrink
2786
Matlab codes, 2002b.
[4] Cai T.T. Zhou H. H. A Data-Driven Block
<ttp://taco.poly.edu/WaveletSoftware/>.
Thresholding Approach To Wavelet Estimation.
[24] Selesnick I. W., Wavelet Software at Brooklyn Poly,
Ann.Statist., http://www.stat.yale.edu/~hz68/.
<eeweb.poly.edu/iselesni/Wavelet Software/denoise2.html>
[5] Chang S. G., Yu B., and Vetterli M. Adaptive wavelet
[25] Sendur L. and Selesnick I. W. Bivariate shrinkage with
thresholding for image denoising and compression. IEEE
local variance estimation, IEEE Signal Process. Lett. 9, 12,
Transaction Image Processing, 9, 2000, 1532-1546.
2002, 438–441
[6] Chen G.Y., Bui T.D., Krzyzak A. Image denoising using
[26] Stein C.Estimation of the mean of a multivariate
neighboring wavelet coefficients. Proceedings of IEEE
normal distribution, Ann. Statist. 9, 1981, 1135–1151
international conference on Acoustics, speech and signal
[27] Starck Jean-Luc, Murtagh Fionn, and Fadili Jaial M.
processing 04, 2004, 917-920
Sparse Image and Signal Processing. Cambridge University
[7] Chinnarao B. Madhavilatha M. Improved Image
Press, 2010
denoising algorithm using Dual Tree Complex Wavelet
Transform. International Journal of Computer Applications,
44, 20, 2012, 1-4
[8] Dengwen Zhou, 24430-BlockShrink-denoising,
<http://www.mathworks.com/matlabcentral/ fileexchange/>
[9] Dengwen Zhou and Xiaoliu Shen, Image denoising
using block thresholding,in Proc. 2008 congress on image
and signal processing,Sanya, Hainan, China, 2008, 335–338
[10] Dengwen Zhou, Wengang Cheng. Image denoising
with an optimal threshold and neighboring window.
Elsevier pattern Recognition, 29, 11, 2008, 1694–169
[11] Dixit A., Sharma P., A Comparative Study of Wavelet
Thresholding for Image Denoising, I.J. Image, Graphics and
Signal Processing, 12, 2014, 39-46
[12] Donoho D.L., Johnstone I.M. Ideal spatial adaptation
via wavelet shrinkage. Biometrika, 81, 1994, 425–455
[13] Donoho D.L., Johnstone I.M. Adapting to unknown
smoothness via wavelet shrinkage. Journal of the American
Statistical Association, 90(432), 1995, 1200-1224
[14] Kingsbury N.: Image processing with complex
wavelets. Phil. Trans. Royal Society London A, 9 (29),
1999, 2543 – 2560

338
IJARI