Basrah Journal of Science (A) Vol.

31(1),111-120, 2013

Comparing study using DWT / CT transforms in image

denoising process

Iman Qays Abduljaleel Amal Hameed Khaleel Suhad Muhajer Kareem

[email protected] [email protected] [email protected]

Basra University\ Science Collage\ Computer Science Dept.

Abstract
Images may contain different types of noises; removing noise from image is often the first step in
image processing, and remains a challenging problem in spite of the sophistication of recent research.
This paper presents an efficient image denoising scheme based on two types of multi-resolution
transforms, namely, the Discrete Wavelet Transform (DWT) and the Curvelet Transform(CT). each
subbands components of an uses transform are denoising using two steps: 1) passed through principal
component analysis (PCA) denoising procedure. 2) The image data obtain from PCA procedure can be
denoising either by hard or soft thresholding techniques. The effectiveness of the methods was
compared using parameters like MSE and PSNR. We find that using CT more efficient then using
DWT and the qulity of denoising increase when we using PCA denoising procedure.
Use the matlab version of 8th in the different treatment stages.

Keywords: image denoising , wavelet transform, curvelet transform, PCA, soft thresholding, hard
thresholding.

I. Introduction acquisition or transmission, while retaining its

Image denoising is a procedure in digital quality [S.Satheesh et.al 2011]. Although many
image processing aiming at the removal of related studies have been done, developing a
noise, which may corrupt an image during its method that is capable to suppress the additive

111
 Iman Qays Abduljaleel et.al Comparing study using DWT / CT ….

noise totally from the noisy image without multi-resolution nature and ability to produce
corrupting the image details is still a high level of noise reduction. Wavelet fails to
challenging task. Hence, the required essential give sparse representation along c² curve.
criteria in designing a denoising technique are: Wavelet effectively represents discontinuities
1) the additive noise in smooth regions should for one dimensional signal. Curvelet transform
be completely removed, 2) the edges should not (CT) overcomes limitations of wavelet
be blurred or sharpened, 3) the texture details transform (WT). CT is a multiscale pyramid
should be maintained, 4) the overall contrast with many direction and position at each length
should be preserved and 5) the additional scale and needle shaped element at fine scale
artifacts should not be appeared in the restored [Anil et.al 2010].
image. However, it is very difficult to develop a The organization of the paper is as follows:
denoising method that matches all the section 2 contented the basic principles used in
mentioned criteria [S. Suhaila et.al 2010]. this paper, section 3 satisfied the proposed
over the last decade, there has been abundant algorithm, section 4 presents experiment results
interest in wavelet methods for noise removal and section 5 concludes the paper.
in signals and images. The new ridgelet and
curvelet transforms were developed over
II. Basic Principles
several years in an attempt to break an inherent
1. Discrete Wavelet Transform(DWT)
limit plaguing wavelet denoising of images.
The DWT is identical to a hierarchical
This limit arises from the well-known and
subband system where the subbands are
frequently depicted fact that the two-
logarithmically spaced in frequency and
dimensional (2-D) wavelet transform of images
represent octave-band decomposition. Due to
exhibits large. Wavelet coefficients even at fine
the decomposition of an image using the DWT
scales, all along the important edges in the
the original image is transformed into four
image, so that in a map of the large wavelet
pieces which is normally labeled as LL, LH,
coefficients one sees the edges of the images
HL and HH as in the schematic depicted in Fig.
repeated at scale after scale. While this effect is
1 a. The LL subband can be further
visually interesting, it means that many wavelet
decomposed into four subbands labeled as LL2,
coefficients are required in order to reconstruct
LH2, HL2 and HH2 [H. Dhillon et.al 2011].
the edges in an image properly [J. Starck et.al
As the discrete wavelet transform (DWT)
2002]. In non-linear techniques, wavelet based
corresponds to basis decomposition, it provides
image denoising methods have attracted due to
112
 Basrah Journal of Science (A) Vol.31(1),111-120, 2013

a non redundant and unique representation of isolate different scales. While ridgelets have
the signal. Several properties of the wavelet global length and variable widths, curvelets in
transform, which make this representation addition to a variable width have a variable
attractive for denoising, are [H. Dhillon et.al length and so a variable anisotropy [J. Starck
2011] : et.al 2002].
 Multiresolution - image details of The decomposition is the sequence of
different sizes are analyzed at the following steps (see Fig. 1) [J. Starck et.al
appropriate resolution scales. 2002]:
 Sparsity - the majority of the wavelet 1.Subband decomposition: The image f is
coefficients are small in magnitude. decomposed into subbands such that :
 Edge detection - large wavelet f →(P0 f ,Δ1 f ,Δ2 f ,...) .
coefficients coincide with image edges. 2.Smooth partitioning: Each subband is then

 Edge clustering - the edge coefficients smoothly windowed into squares.

within each sub band tend to form 3.Renormalization: Each resulting square is

spatially connected clusters renormalized to unit scale

4.Ridgelet analysis: Each square is analyzed via

2. Curvelet Transform(CT) the discrete ridgelet transform.

CT is one of the latest developments of non- Both curvelets and ridgelets take the form of

adaptive transforms today. Compared to basis elements which exhibit very high

wavelet, curvelet provides a more sparse directional sensitivity and are highly

representation of the image, with improved anisotropic. Both occur at all scales, locations,

directional elements and better ability to and orientations. However, whereas ridgelets

represent edges and other singularities along have unit length, curvelets occur at all dyadic

curves[B. Zhang et.al 2011]. Curvelets have lengths and exhibit an anisotropy increasing

useful geometric features that set them apart with decreasing scale like a power law.

from wavelets. It uses the ridgelet transform as Curvelet obeys a scaling relation where by the

a component step and implements curvelet width of a curvelet element is about the square

subbands using a filter bank of wavelet filters. of its length; i.e. width ~ length² [A. Majumdar

This transform combines multiscale ridgelets et.al 2007].

with a spatial bandpass filtering operation to

113
 Iman Qays Abduljaleel et.al Comparing study using DWT / CT ….

Fig.1 : Overview of Curvelet Transform Scheme.

3. Principal Component Analysis (PCA) The auto-covariance matrix M for the ensemble
PCA is one of the most well known methods X is defined by
for data analysis. The objective of this method …(3)

is dual. Primarily it is to massage input data in a

Where M is d²*d² matrix, with elements
form where the multiple dimensions of the
…(4)
input data are mutually decorrelated. As a side
effect the significance of each decorrelated We extract the eigenvectors of M as columns in
dimension, in terms of energy contribution in a matrix V and its eigenvalues as diagonal
the original input, is also estimated. Upon entries in a matrix A [Pierre 2005, Masoud et.al
concluding the computations the user is 2008].
presented with a ranking on which dimensions
are most important and based on this ranking 4. Salt and pepper noise
we can throw away non-important dimensions Salt and pepper noise is a form of noise
and reduce the dimensionality of the data with typically seen on images. It represents itself as
minimal side effects[Pierre 2005]. randomly occurring white and black pixels. A
Let us consider a set of N sample images “spike” or impulse noise drives the intensity
{X1,X2,…,XN} of d*d dimensions be an values of random pixels to either their
ensemble of vectors and Let be the average maximum or minimum values. The resulting
vector in the ensemble. black and white flecks in the image resemble
salt and pepper[G. Iango et.al 2011].
…(1)

After subtracting the average from each 5. Thresholding Techniques

element of X, we get a modified ensemble of Thresholding is a procedure which takes place

vectors X'= {X'1,X'2,…,X'N} with after decomposing a signal at a certain

decomposition level. A threshold is applied to
…(2)
coefficients for each level from 1 to N (last
114
 Basrah Journal of Science (A) Vol.31(1),111-120, 2013

decomposition level). By applying a hard

threshold the coefficients below this threshold III. Proposed algorithm
level are zeroed, and the output after a hard This section describes the image denoising
threshold is defined by this equation[J. Karam algorithm. The algorithm is very simple to
2008] :- implement and computationally more efficient
(see Fig. 2).
…(5)
We can see in Fig. 2 that we converted an input
where x(t) is the input speech signal and δ is the image to noised image by using salt and pepper
threshold. An alternative is soft thresholding at noise because this type of noise is caused by
level δ which defined by this equation [Pierre errors in data transmission and we can
2005] :- measured the performance of proposed
algorithm by using this type of noise.
…(6)

Input noisy image

DWT CT

PCA algorithm to all DWT coefficient PCA algorithm to all CT coefficient

hm
Estimate soft Estimate hard Estimate soft Estimate hard
thresholding thresholding thresholding thresholding

IDWT IDWT ICT ICT

denoised image1 denoised image2 denoised image3 denoised image4

Fig. 2: proposed algorithms scheme.

The wavelet denoising method consists of the  Calculate the DWT of the input image at
following steps: scale 2 and mother wavelet db1.

115
 Iman Qays Abduljaleel et.al Comparing study using DWT / CT ….

 For each subband coefficients calculate the we can see that thresholding step performed
PCA algorithm using Kaiser's rule to zeros the first action of image denoising by
all value less than Kaiser’s value. removing unaccepted value less than
 Using thresholding ( hard or soft) to threshold value.
denoising each DWT subband coefficients.  Reconstruct the denoised image from the
 Reconstruct the denoised image from the modified CT coefficients by using inverse
modified DWT coefficients by using CT (ICT).
inverse DWT (IDWT). We use statistical tool to measure the
enhancement of images. The Root Mean Square
Another algorithm was proposed by using Error (RMSE) and Peak Signal-to-Noise Ratio
another multi-resolution transform which (PSNR) are used to evaluate the enhancement
named Curvelet transform (CT) and we can of images [L.R. Litwin 1998].
defined as below:
 Calculate the CT of the input image at scale
…(7)
2 and angle 8. We get a set of subbands wj,
each subband wj contains Nj coefficients
and corresponds to a given resolution level.
…(8)
 For each subband coefficients calculate the
PCA algorithm using Kaiser's rule to zeros
all value less than Kaiser’s value.
 Using thresholding : If a CT subband IV. Experiment Result
coefficients is smaller than a predefined The results of the proposed method are tested
threshold it will be set to zero; otherwise it on different Gray-Scale images such as
will shrunk in the absolute value by an cameraman.jpg and lena.jpg of size ( 512*512
amount of the threshold, this function pixels ). The collected testing images are
named soft thresholding. Same as soft corrupted by salt and pepper noise and then
threshold If a CT subband coefficients is these are used for proposed denoising
smaller than a predefined threshold it will algorithm. Fig. 4 and Fig. 5 illustrates the
be set to zero; otherwise it kept unchanged, denoising algorithms results ( PSNR and
this function named hard thresholding. So RMSE ) .

116
Basrah Journal of Science (A) Vol.31(1),111-120, 2013

(a) (b) (c)

(d) (e) (f)

Fig.4 : (a) original image1, (b) image1 with salt and pepper noise, (c) denoised
image1 by CT_PCA_HARD algorithm (PSNR=16.8429 , RMSE=181.28 ), (d)
denoised image1 by CT_PCA_SOFT algorithm (PSNR=15.6577 , RMSE=
211.51), (e) denoised image1 by DWT_PCA_HARD algorithm (PSNR=20.3849,
RMSE=121.01 ), (f) denoised image1 by DWT_PCA_SOFT algorithm(PSNR=
20.3246, RMSE=127.39 )

117
 Iman Qays Abduljaleel et.al Comparing study using DWT / CT ….

(a) (b) (c)

(d) (e) (f)

Fig.5 : (a) original image2, (b) image2 with salt and pepper noise, (c) denoised
image2 by CT_PCA_HARD algorithm (PSNR=16.1436 , RMSE=1580.2), (d)
denoised image2 by CT_PCA_SOFT algorithm (PSNR=14.9043, RMSE=
2103.01), (e) denoised image2 by DWT_PCA_HARD algorithm
(PSNR=19.2042, RMSE=781.02 ) , (f) denoised image2 by DWT_PCA_SOFT
algorithm (PSNR= 19.179, RMSE=785.55 )

V. Conclusion the CT is more efficient than DWT and it gives

In this work, we have introduced various the best results after using PCA algorithm in
techniques for removal of Salt and pepper noise CT coefficient . The proposed method is simple
from images. To find the most important value and easy to implement.
in each CT/DWT coefficients we used PCA
algorithm to removed non-important CT/DWT References
coefficients. At different scales the CT/DWT Anil A. patil, J. Singhai, (2010), “Image
coefficients vary, therefore we applied the Denoising Using Curvelet Transform : An
thresholding functions (soft and hard Approach For Edge Preservation”, Journal of
thresholding). The performance of proposed Scientific & Industrial Research, Vol. 69, pp.
algorithms is measured using quantitative 34-38.
performance measures such as RMSE and
PSNR. The experimental results indicate that
118
 Basrah Journal of Science (A) Vol.31(1),111-120, 2013

A. Majumdar, B.B. Chaudhuri,(2007),

"Curvelet-Based Multi SVM Recognizer for J. Starck, Emmanuel J. Candès, David L.
Offline Handwritten Bangla: A Major Indian Donoho,(2002), “The Curvelet Transform for
Script", Ninth International Conference on Image Denoising”. IEEE Transactions On
Document Analysis and Recognition (ICDAR Image Processing 11(6), pp. 670-684.
2007), Vol 1, pp.491-495.
L.R. Litwin, (1998),”Speech Coding with
B. Zhang, Y. Zhang, W. Lu and G. Wavelets”, IEEE Potentials, Vol.17, No.2, pp.
Han,(2011)," Phenotype Recognition by 38-41.
Curvelet Transform and Random Subspace
Masoud Mazloom, Saeed
Ensemble", Journal of Applied Mathematics &
Ayat,(2008),"Combinational Method for Face
Bioinformatics, vol.1, no.1,pp. 79-103.
Recognition: Wavelet, PCA and ANN", 8th
DICTA Conference, Canberra, Australia.
G. Iango and R. Marudhachalam,(2011),” New
Hybrid Filtering Techniques For Removal Of Pierre Divenyi,(2005)," Speech Separation by
Gaussian Noise From Medical Images”, ARPN Humans and Machines", Kluwer Academic
Journal of Engineering and Applied Sciences, Publishers, Boston, ISBN: 1-4020-8001-8.
VOL. 6, NO. 2, pp. 8-12.
S.Satheesh, K. Prasad, (2011),” Medical Image
H. Dhillon, G. D. Jindal, A. Girdhar, (2011), ” Denoising Using Adaptive Threshold Based On
A Novel Threshold Technique for Eliminating Contourlet Transform”, ACIJ, Vol.2, No.2, pp.
Speckle Noise In Ultrasound Images”, IPCSIT, 52-58.
IACSIT Press, Singapore, vol.10.
S. Suhaila, T. Shimamura, (2010),“A
Thresholding-Based Image Denoising Method”,
J. Karam, (2008),” End Point Detection for
World Academy of Science, Engineering and
Wavelet Based Speech Compression”, World
Technology, Vol. 65, pp. 892-897.
Academy of Science, Engineering and
Technology, Vol. 37, pp.327-330.

119
 ‫‪Iman Qays Abduljaleel et.al‬‬ ‫‪Comparing study using DWT / CT ….‬‬

‫دراست هقارنت لتحويلي (‪ )DWT/CT‬الوستخذم اى في هعالجت‬

‫تنقيت الصور‬

‫سهاد ههجر كرين‬ ‫اهل حويذ خليل‬ ‫ايواى قيس عبذ الجليل‬
‫‪[email protected]‬‬ ‫‪[email protected]‬‬ ‫‪[email protected]‬‬

‫جاهعت البصرة\ كليت العلوم \ قسن علوم الحاسباث‬

‫الوستخلص‬
‫ححخٕ٘ انصٕر إَٔاػا يخخهفت يٍ انضٕضاء‪ٔ ,‬نٓذا حؼخبز يحأنت إسانت ْذِ انضٕضاء انخطٕة األٔنٗ ضًٍ سهسهت خطٕاث‬
‫يؼانجت حهك انصٕرة ٔانخ‪ ٙ‬السانج حؼذ ححذ‪ٚ‬ا نهكث‪ٛ‬ز يٍ انباحث‪ ٍٛ‬انؼايه‪ ٍٛ‬ف‪ْ ٙ‬ذا انًجال‪.‬‬
‫‪multi-‬‬ ‫ٔيٍ خالل ْذا انبحث َقذو آن‪ٛ‬ت يقخزحت إلسانت احذ إَاع انضٕضاء باالػخًاد ػهٗ َٕػ‪ ٍٛ‬يٍ انخحٕ‪ٚ‬الث انًخؼذدة انذقت‬
‫‪ ْٙٔ ,resolution‬انخحٕ‪ٚ‬م انًٕ‪ٚ‬ج‪ ٙ‬انًخقطغ )‪ٔ Discrete Wavelet Transform(DWT‬ححٕ‪ٚ‬م )‪ٔ . Curvelet (CT‬بحسب‬
‫انخحٕ‪ٚ‬م انًسخخذو َقٕو بخُق‪ٛ‬ت انؼُاصز انًكَٕت نكم حشيت جشئ‪ٛ‬ت ‪ subband‬يٍ انحشو انًكَٕت نهصٕرة األصه‪ٛ‬ت باسخخذاو انخطٕح‪ٍٛ‬‬
‫)‪ )2 . principal component analysis (PCA‬انب‪ٛ‬اَاث انًُقاة‬ ‫انخان‪ٛ‬خ‪ )1 : ٍٛ‬حًز‪ٚ‬ز انكخهت انًحذدة يٍ انؼُاصز إنٗ خٕارسي‪ٛ‬ت‬
‫‪soft‬‬ ‫‪ hard thresholding‬أ خٕارسي‪ٛ‬ت‬ ‫‪ PCA‬حُقٗ يزة أخزٖ باسخخذاو خٕارسي‪ٛ‬ت‬ ‫انخ‪َ ٙ‬حصم ػه‪ٓٛ‬ا يٍ خٕارسي‪ٛ‬ت‬
‫‪.thresholding‬‬
‫حقاص كفاءة األَظًت انًقخزحت باالػخًاد ػهٗ انًق‪ٛ‬اس‪ٔ .PSNR ٔ MSE ٍٛ‬قذ ٔجذَا اٌ اسخخذاو ححٕ‪ٚ‬م ‪ Curvelet‬كاٌ أكثز كفاءة‬
‫يقارَت بانخحٕ‪ٚ‬م انًٕ‪ٚ‬ج‪ ٙ‬انًخقطغ ‪ٔ , DWT‬اسدادث كفاءحّ ػُذ إنحاقّ بخٕارسي‪ٛ‬ت ‪ٔ .PCA‬قذ حى اسخخذاو بزَايج ‪matlab (R2011‬‬
‫)‪ a‬ف‪ ٙ‬كم يزاحم انًؼانجت‪.‬‬

‫‪120‬‬