201O International Conference on Mechanical and Electrical Technology (ICMET 2010)

Image Denoising Using Wavelet Transform

Mr. Sachin Ruikar Dr. DDDoye

Assistant Professor, Professor
Electronics & Telecommunication Department, Electronics & Telecommunication Department
STES's Sinhgad Acdemy of Engg, PUNE, INDIA SGGS Institute of Engineering and Technology,
ruikarsachin@gmail.com Nanded, India
dddoye@yahoo.com

Abstract-An image is often corrupted by noise in its transforming function and it is called the mother wavelet
acquisition and transmission. Removing noise from the original transform. The term mother wavelets mean a small wave.
image is still a challenging problem for researchers. In this
The smallness refers to the condition that this (window)
work new approach of threshold function developed for image
function is of finite length (compactly supported). The term
denoising algorithms. It uses wavelet transform in connection
translation is related to the location of the window, as the
with threshold functions for removing noise. Universal, Visu
Shrink, Sure Shrink and Bayes Shrink, normal shrink are
window is shifted through the signal[12]. This term
compared with our threshold function, it improves the SNR obviously corresponds to time information in the
efficiently. transformdomain. To construct the mother wavelet \II(x) ,
first determine a scaling function <I>(x), which satisfies the
Keywords-universashrink, baysshrink, sureshrink, two scale difference equation
normalshrink, vishushrink, etc.
<I> (x) = -Y2 Lk h(k) <I> (2x-k) ............................... (4)
Then , the wavelet kernel \II (x) is related to the scaling
I. INTRODUCTION function via
Noise reduction plays a fundamental role in image \II (x) = -Y2 Lk g(k) <I> (2x-k) ................................(5)
processing, and wavelet analysis has been demonstrated to Where g(k) = (-1/ h (1-k)
be a powerful method for performing image noise reduction The coefficients is h(k) in have to meet several
[1]. The procedure for noise reduction is applied on the conditions for the set of basis wavelet function in to be
wavelet coefficients achieved using the wavelet unique, orthonormal, and have degree of regularity . The
decomposition and representing the image at different scales. coefficient h(k) and g(k) play a very crucial role in a given
After noise reduction, the image is reconstructed using the discrete wavelet transform. To perform the wavelet
inverse wavelet transform. Decomposition and transform does not require the explicit form of the <l>(X) and
reconstruction are accomplished using two banks of filters \II (X) but only depends on h(k) and g(k) . Consider a J -
constrained by a perfect reconstruction condition. The level wavelet decomposition which can be written as
k
threshold selection for this denoising technique is Fo(x)= L Co,k <l>o,k (X)=Lk (Cj+l,k<l>j+l,k(x)+Lj�Jd j+l , k \II j+l ,k
application dependent. Section II summarizes review of (x) ........................(6)
wavelet transform. Section III deals with the types of the where coefficients Co,k are the given and coefficients Cj+1
threshold functions used. Sections IV explain Image and dj+l,k at scale j+1 are related to the coefficients Cj,k at
denoising algorithm. Section V Consists the performance scalej via
evaluation and results. Section VI describes the conclusions. Cj+l,n = LkCj,k h (k-2n)
II. REVIEW OF WAVELET TRANSFORM A+l,n=LkCj,kg(k-2n) .... ................................(7)
Where 0 � j � J .Thus provides recursive algorithm for
By wavelet transform [10] the decomposition of a signal wavelet decomposition through h(k) and g(k) and the final
with a family of real orthonormal bases \IIm,n (x) obtained outputs include a set of J-level wavelet coefficients dj,n , 1 �
through translation and dilation of a kernel function \II (x) j � J and the coefficients Cj,n for low resolution component
known as the mother wavelet, i.e.
m m <l>j,k(X) . By using a similar approach, we can derive recursive
\IIm,n(x)=2 <- 12)'¥(T x_n) .....................................(1) algorithm for function synthesis based on its wavelet
Where m and n are integers. Due to the orthonormal
coefficients dj,n , 1� j � J and Cj,n
property , the wavelet coefficients of a signal f(x) can be
Cj,k=LnCj+l,nh(k-2n)+Lndj+l,�(k-2n) ................(8)
easily computed via
+ It is convenient to view the decomposition as passing a
Cm,n= -J � f(x) \II m ,n (x) dx ................................ (2)
signal Cj,k through a pair of filters Hand G with impulse
And the synthesis formula
response h(n) and g(n) and down sampling the filtered
F(X)=Lm,nCm,n\llm,n(x) ...............................(3) signals by two, where h(n) and g(n) are defined as
The transformed signal is a function of two variables, . h(n) = h(-n) , g(n) = g(-n) ...............(9
the translation and scale parameter. The ,¥(t) is the

978-1-4244-8102-6/10/$26.00 © 2010 IEEE 509

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

The pair of filters H and G corresponds to the half band The choice of a threshold is an important point of
low pass and high pass filters, respectively, and is called the interest. It plays a major role in the removal of noise in
quadrature mirror filters in the signal processing it is shown images because denoising most frequently produces
in Figure 1. The reconstruction procedure is implemented by smoothed images, reducing the sharpness of the image. Care
up sampling the subsignal Cj+1 and dj+1 (inserting zero should be taken so as to preserve the edges of the denoised
between the neighboring samples) and filtering with hen) image..
and g(n), respectively, and adding these two filtered signals
together[13] [14]. Usually the signal decomposition scheme
is performed recursively to the output of the low pass filter h. III. TYPES OF THRESHOLD
It leads to the conventional wavelet transform or the so Noise is present in an image either in an additive or
called pyramid structured wavelet decomposition. multiplicative form. An additive noise follows the rule, w (x,
Presenting the simplest form of wavelets, the Haar y) = s (x y) +n(x, y),
basis[6][7]. It covers one-dimensional wavelet transforms While the multiplicative noise satisfies w(x, y)=s(x, y)xn(x,
and basis functions. y), Where s(x,y) is the original signal, n(x,y) denotes the
noise introduced into the signal to produce the corrupted
image w(x,y), and (x,y) represents the pixel location[14].
Gaussian Noise is evenly distributed over the signal. This
means that each pixel in the noisy image is the sum of the
levelj+l
true pixel value and a random Gaussian distributed noise
value. Salt and Pepper Nois is an impulse type of noise,
which is also referred to as intensity spikes. This is caused
where
0 convolve with filter X and generally due to errors in data transmission. The corrupted
pixels are set alternatively to the minimum or to the
dOWllsampling maximum value, giving the image a "salt and pepper" like
appearance. Unaffected pixels remain unchanged. Speckle
Figure 1. One level downsampling
Noise is a multiplicative noise. Th is type of noise occurs
As mentioned earlier, the wavelet equation produces
different wavelet families like Daubechies, Haar, coiflets, in almost all coherent imaging systems such as laser,
etc. Wavelets are classified into a family by the number of acoustics and SAR (Synthetic Aperture Radar) imagery. The
vanishing moments N. Within each family of wavelets there source of this noise is attributed to random interference
are wavelet subclasses distinguished by the number of between the coherent returns[I], [2], [3]. Fully developed
coefficients and by the level of iterations. The wavelet speckle noise has the characteristic of multiplicative noise.
decomposition of an image is done as follows: In the first A. Universal Threshold
level of decomposition, the image is split into 4 subbands,
The universal threshold can be defined as
namely the HH, HL, LH and LL subbands as shown in
•

Figure 2. The HH subband gives the diagonal details of the T = aJ21og(N) ... ... ... ... ... ... ... (10)
image; the HL subband gives the horizontal features while N being the signal length, (J being the noise variance is well
the LH subband represents the vertical structures [4] [5]. known in wavelet literature as the Universal threshold. It is
The LL subband is the low resolution residual consisting of the optimal threshold in the asymptotic sense and minimizes
low frequency components and it is this subband which is the cost function of the difference between the function.
further split at higher levels of decomposition [13]. However, it is useful for obtain a starting value when
LL, HL, nothing is known of the signal condition. One can surmise
HL,
LH,
that the universal threshold may give a better estimate for
HH,
HL, the soft threshold if the number of samples is large[2][26].

LH, HH, B. Visu Shrink

Visu Shrink was introduced by Donoho[5]. It uses a
threshold value t that is proportional to the standard
deviation of the noise. It follows the hard thresholding rule.
An estimate of the noise level (J was defined based on the
LH, HH, median absolute deviation given by
Median(�g I j=I,k : k = 0,1,......2
j-I
-I})
a=
�

...... (11)
0.6745
Figure 2. subbands of image after decomposition

510

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

Where &-1, k corresponds to the detail coefficients in the From the definition of additive noise we have w(x, y)=s(x,
wavelet transform. VisuShrink does not deal with y)+n(x, y).
minimizing the mean squared error. It can be viewed as Since the noise and the signal are independent of each other,
general-purpose threshold selectors that exhibit near optimal it can be stated that
minimal error properties and ensures with high probability a2w =a2s +�
that the estimates are as smooth as the true underlying .....................................(15)
functions. However, VisuShrink is known to yield recovered a2w can be computed as shown below:
images that are overly smoothed. This is because
VisuShrink removes too many coefficients. Another
disadvantage is that it cannot remove speckle noise. It can
2 1 � 2
a w = -2 L. w
n x,y=i
(x,y) ...............................
(16)

only deal with an additive noise. VisuShrink follows the The variance of the signal, �s is computed as
global thresholding scheme where there is a single value of 2 2
threshold applied globally to all the wavelet coefficients[3]. as = Jmax(a w -0' ,0) .............(17)
With a2and a2so the Bayes threshold is computed from
C. Sure Shrink
Equation (14). Using this threshold, the wavelet coefficients
A threshold chooser based on Stein's Unbiased Risk are threshold at each band.
Estimator (SURE) was proposed by Donoho and Johnstone
and is called as Sure Shrink. It is a combination of the E. Normal Shrink
universal threshold and the SURE threshold [5]. This The threshold value which is adaptive to different
method specifies a threshold value tj for each resolution subband characteristics
2I
TN= pa
levelj in the wavelet transform which is referred to as level
dependent thresholding. The goal of Sure Shrink is to lay
minimize the mean squared error[16], defined as, Where the scale parameter � is computed once for each
MSE = -2
1 n 2
�::CZ(x,y) - S(x,y)) scale using the following equation.

�lO� L; )
................ (12)
n x,y=i
p= ..................... (18)
Where z(x,y) is the estimate of the signal while s(x,y) is
the original signal without noise and n is the size of the
signal. Sure Shrink suppresses noise by thresholding the Lk is the length of the subband at Kth scale
empirical wavelet coefficients. The Sure Shrink threshold t* a2 is the noise variance, which is estimated from the
is defined as subband HH using equation (11).

t* = min(t,a�2Iogn) .................. (13) F. Proposed Threshold

Where t denotes the value that minimizes Stein's This function is calculated by
Unbiased Risk Estimator, a is the noise variance computed newth = ,Jr-2m - -)
- -xI- o-g(M
from Equation, and n is the size of the image. Sure Shrink
Where, M is the total number of pixel of an image.
follows the soft thresholding rule. The thresholding
employed here is adaptive, i.e., a threshold level is assigned IV. IMAGE DENOISING ALGORITHM
to each dyadic resolution level by the principle of
The problem boils down to finding an optimal threshold
minimizing the Stein's Unbiased Risk Estimator for
such that the mean squared error between the signal and its
threshold estimates. It is smoothness adaptive, which means
estimate is minimized. The different methods for denoising
that if the unknown function contains abrupt changes or
we investigate differ only in the selection of the threshold.
boundaries in the image, the reconstructed image also does.
Soft thresholding is used for all the algorithms due to the
D. Bayes Shrink following reasons: Soft thresholding has been shown to
Bayes Shrink was proposed by Chang, Yu and Vetterli. achieve near minimax rate over a large number of Besov
The goal of this method is to minimize the Bayesian risk, spaces [3]. Moreover, it is also found to yield visually more
and hence its name, Bayes Shrink [22]. It uses soft pleasing images. Hard thresholding is found to introduce
thresholding and is subband-dependent, which means that artifacts in the recovered images. We now study four
thresholding is done at each band of resolution in the thresholding techniques Universal, VisuShrink, Sure Shrink
wavelet decomposition. Like the Sure Shrink procedure, it is and Bayes Shrink and investigate their performance for
smoothness adaptive. The Bayes threshold, tB, is defined as denoising on Lena image. Fig 3. Shows that denoising
tE =�/ as ...................................(14) technique in details such as
Where a2 is the noise variance and as is the signal variance x: the original noise-free digital one or two-dimensional
without noise. The noise variance a2 is estimated from the signal which has M samples.
subband HH by the median estimator shown in equation O. w: an additive white Gaussian noise with zero mean and
variance w, which is assumed to the same size M but

511

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

independent of the original signal x, y=x+w: the noisy TABLEr. RESULT ANALYSIS FOR LENA IMAGE

version of the original noise-free signal.

Name of Name of Original Noisy After After
Y=DWT (y): the DWT of the noisy signal y. the noise the ImageSNR Image denoising denoising
T(Y,A): the thresholding transformation with threshold ,. Technique In DB SNR in (Recovered using
XA: the thresholded wavelet coe±cients obtained after DB image) Block
applying the thresholding operator x: the denoised version SNR in DB only
Gaussian Visushrink 13.2505 13.5400 14.2110 13.5583
of the noisy image y, which represents an approximation of Image Universal 9.2979 13.4046
the original image x. threshold
Sureshrink 16.8720 13.5442
Normal 15.4684 13.5418
Shrink
Bays 15.3220 13.5452
T(Y,'\) shrink
NEW 15.0061
A.J!I!!"oach
Figure 3: Denoising using DWT & IDWT Speckle Visushrink 13.2505 12.6249 12.2023 12.6241
The goal behind using wavelets for denoising purposes is Universal 9.8297 12.3477
to smooth out some areas where noise is present but to threshold
Sureshrink 14.7922 12.6302
leave other areas unaffected. The addition of noise to a
Normal 15.0069 12.6260
signal will contribute noisy coefficients. Unlike noisy Shrink
coefficients that are contributing to all coefficients, signal is Bays 14.9716 12.6735
contributed to only a few high amplitude coefficients. shrink
Therefore, since the signal is concentrated in a few NEW 14.5171
�roach
coefficients, it is possible to threshold out the noise only Salt & Visushrink 13.2505 14.0302 11.8957 13.3760
leaving the signal. In order to pick a threshold by which to Pepper Universal 9.6062 15.6247
filter out noise requires complete knowledge of the signal threshold
and an estimate of the noise process. One global and non­ Sureshrink 17.2951 13.3792
linear method involves knowing the standard deviation of Normal 17.2498 13.3849
Shrink
the noise and the number of wavelet coefficients or signal Bays 16.9775 13.3718
samples shrink
NEW 15.5274
A. Proposed algorithm: Approach
We have considered the block of the approximation Poisson Visushrink 13.2505 13.7428 11.6988 13.7780
Universal 9.5227 12.7115
coefficient. Then calculate the threshold function using the
threshold
Bays threshold, normal threshold, Sure Shrink, VisuShrink Sureshrink 14.3082 13.7788
& universal threshold. In denoising each coefficient is Normal 14.5519 13.7481
threshold by comparing against a threshold; if the Shrink
coefficient is smaller than the threshold it is set to zero, Bays 14.6518 13.7785
shrink
otherwise, it is kept. Then we will recover the denoised
NEW 13.9969
image. A....I!.I!.roach
V. RESULTS

We compared various denoising method on several test

images widely used in image processing community. Here, VI. CONCLUSION AND FUTURE SCOPE
we report the result only for the Lena image Result shown in
tabular format. By applying the new threshold function our In the case where an image is corrupted with Gaussian
result is improved drastically. From the experimental and noise, the wavelet shrinkage denoising has proved to be
mathematical results it can be concluded that for Gaussian nearly optimal. The output from Bayes Shrink method is
noise image Normal threshold & Sure Shrink gives better much closer to the high quality image and there is no
result. In case of Poisson noise Bays threshold, Normal blurring in the output image unlike the other two
threshold & Sure Shrink gives better quality result. For salt methods.VisuShrink cannot denoise multiplicative noise
and pepper noise, Bays threshold & Normal threshold gives unlike BayesShrink. Denoising salt and pepper noise using
us better result. For speckle noise Bays threshold, Normal VisuShrink and Universal has proved to be inefficient.
threshold & Sure Shrink gives better quality result. Result Since selection of the right denoising procedure plays a
analysis for input image Lena size (512 x 512) is shown in major role, it is important to experiment and compare the
table 1. methods. As future research, we would like to work further
on the comparison of the denoising techniques. Besides, the

512

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

complexity of the algorithms can be measured according to [14] Arthur Jr Weeks , Fundamental of Electronic Image Processing PHI
2005
the CPU computing time flops. This can produce a time
[15] Chang, S. G., Yu, B., and Vetterli, M. (2000). Adaptive wavelet
complexity standard for each algorithm. These two points
thresholding for image denoising and compression. IEEE Trans. on
would be considered as an extension to the present work Image Proc., 9, 1532-1546
done. [16] Chambolle, A, DeVore, R. A, Lee, N.-Y., and Lucier, BJ. (1998).
Nonlinear wavelet image processing: variational problems,
REFERENCES compression, and noise removal through wavelet shrinkage. IEEE
Trans. on Image Proc., 7,319-355
[1] F. Luisier, T. Blu, and M. Unser, "A new SURE approach to image
denoising: Inter-scale orthonormal wavelet thresholding," IEEE Trans. [17] H. Choi and R. G. Baraniuk, "Analysis of wavelet domain Wiener
Image Process., vol. 16, no. 3, pp. 593--606, Mar. 2007. filters," in IEEE Int. Symp. Time- Frequency and Time-Scale
Analysis, (Pittsburgh), Oct. 1998.
[2] x.-P. Zhang and M. D. Desai, "Adaptive denoising based on SURE
risk," IEEE Signal Process. Lett., vol. 5, no. 10, pp. 265-267, Oct. [18] S. Grace Chang, Bin Yu and M. Vattereli, Wavelet Thresholding for
1998. Multiple Noisy Image Copies, IEEE Trans. Image Processing, vol. 9,
pp.1631- 1635, Sept. 2000
[3] Thierry Blu, and Florian Luisier "The SURE-LET Approach to Image
Denoising". IEEE TRANSACTIONS ON IMAGE PROCESSING, [19] .H. Zhang, Aria Nosratinia, and R. O. Wells, Jr., "Image denoising
VOL. 16, NO. 11, pp 2778 - 2786 , NOV 2007 via wavelet-domain spatially adaptive FIR Wiener filtering", in IEEE
Proc. Int. Conf. Acoust., Speech, Signal Processing, Istanbul, Turkey,
[4] Bin Yu, Martin Vetterli, and Grace Chang, "Spatially Adaptive
June 2000
Wavelet Thresholding with Context Modeling for Image Denoising".
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 9, NO. 9, [20] H. A Chipman, E. D. Kolaczyk, and R. E. McCulloch: 'Adaptive
pp1522-1531 SEPTEMBER 2000 Bayesian wavelet shrinkage', J. Amer. Stat. Assoc., Vol. 92, No 440,
Dec. 1997, pp. 1413-1421
[5] D. L. Donoho and I. M. Johnstone, "Denoising by soft thresholding",
IEEE Trans. on Iriform. Theory, Vol. 41, pp. 613-627, 1995. [21] M. Lang, H. Guo and J.E. Odegard, Noise reduction Using
Undecimated Discrete wavelet transform, IEEE Signal Processing
[6] E. Stollnitz, T. DeRose, and D. Salesin, "Wav elets for Computer
Letters, 1995
Graphics: A Primer (part 1)", IEEE Computer Graphics and
Applications, vol. 15, no. 3, pp. 76-84, 1995 [22] T. D. Bui and G. Y. Chen, "Translation-invariant denoising using
multiwavelets", IEEE Transactions on Signal Processing, Vo1.46,
[7] C. Burrus, R. Gopinath, and H. Guo, Introduction to Wavelets and
No.l2, pp.3414-3420, 1998.
Wavelet Traniforms: A Primer, Prentice Hall, 1998
[23] J. Lu, J. B.Weaver, D.M. Healy, and Y. Xu, "Noise reduction with
[8] Donoho, D.L. and Johnstone, I.M. (1994) Ideal spatial adaptation via
multiscale edge representation and perceptual criteria," in Proc.
wavelet shrinkage. Biometrika, 81, 425-455.
IEEE-SP Int. Symp. Time- Frequency and Time-Scale Analysis,
[9] Donoho, D.L. and Johnstone, I.M. (1992) Adapting to Unknown Victoria, BC, Oct. 1992, pp. 555-558.
Smoothness via Wavelet Shrinkage. to appear JASA, 1995
[24] P. Moulin and J. Liu, "Analysis of multiresolution image denoising
[10] Tianhomg Chang & C.C Jay Kuo "Texture Analysis And schemes using generalized Gaussian and complexity priors", IEEE
Classification With Tree-structured Waelet Transform" IEEE Infor. Theory, Vol. 45, No 3, Apr. 1999, pp. 909-919.
Transaction On Image Processing Vol.2 No.4 Oct. 1993. pp429-440
[25] A Cohen and J. Kova·cevic', "Wavelets: The mathematical
[11] Ingrid Daubechies "The Wavelet Transform, Time-Frequency background," Proc. IEEE, vol. 84, pp. 514-522, 1996.
Localization and Signal Analysis" IEEE Transaction On Information
[26] S. Chang, B. Yu, and M. Vetterli, "Adaptive wavelet thresholding for
Theory vol.36, No.5,Sept-1990, pp961-1005
image denoising and compression," IEEE Trans. Image Processing,
[12] M. Sonka,V. Hlavac, R. Boyle Image Processing. Analysis. And vol. 9, pp. 1532-1546, Sept. 2000.
Machine Visison. Ppl0-21O & 646-670
[27] R. C. Gonzalez and R. E.Woods, Digital Image Processing. Reading,
[13] Raghuveer M. Rao., AS. Bopardikar Wavelet Transforms: MA: Addison-Wesley, 1993.
Introduction To Theory And Application Published By Addison­
[28] S. G. Mallat, A Wavelet Tour of Signal Processing. New York:
Wesly 2001 ppl-126
Academic, 1998.

513

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

514

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Mechanical and Electrical Technology (ICMET 2010)

515

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on April 01,2025 at 09:21:35 UTC from IEEE Xplore. Restrictions apply.