 Open Access
 Total Downloads : 289
 Authors : Poojaverma, Gurdeep Kaur, Dr. Naveen Dhillon
 Paper ID : IJERTV2IS4680
 Volume & Issue : Volume 02, Issue 04 (April 2013)
 Published (First Online): 23042013
 ISSN (Online) : 22780181
 Publisher Name : IJERT
 License: This work is licensed under a Creative Commons Attribution 4.0 International License
Performance Investigation Of Different Wavelet Families To Optimize Mse Of Digital Image
*PoojaVerma ** GurdeepKaur***Dr. Naveen Dhillon
*M.Tech. (ECE), R.I.E.T., Phagwara**M.Tech.(CSE), D.A.V.I.E.T., Jalandhar
***HOD (ECE), R.I.E.T.Phagwara.
Abstract
Denoising is based on incorporating neighbouring wavelet coefficients, with different threshold value for different subband. The choice of the threshold estimation is carried out by analyzing the statistical parameters of the wavelet subband coefficients. To prove the efficiency of this method in image denoising, this method is compared with various conventional wavelet denoising approaches like VisuShrink, and NeighShrink algorithm which is based on neighbouring wavelet coefficients with universal threshold,which gives significant improvement of Mean Square Error (MSE) in both the cases. The observed result shows that, the proposed method yields superior image quality and better MSE.
Keywords:Imagedenoising, wavelet transform, VS, NS, RNS, MSE.

Introduction
Image noise is undesired variation in pixel intensity values in a captured or transmitted image. Various waveletbased methods have been proposed for the purpose of image enhancement and restoration. Basic wavelet image restoration methods are based on thresholding in the sense that each wavelet coefficient of the image is compared to a given threshold; if the coefficient is smaller than the threshold, then it is set to zero, otherwise it is kept or slightly reduced in magnitude. The intuition behind such an approach follows from the fact that the wavelet transform is efficient at energy compaction, thus small wavelet coefficient are more likely due to noise, and large coefficient are generally due to important image features, such as edges. The main idea behind it is that, if the wavelet coefficients estimates are bigger in absolute value of a certain specified threshold then the
same value is either retained as such or is diminished by the amount corresponding to the threshold. The smaller coefficients are instead eliminated, hence sparsifying the wavelet expansion. There are two basic approaches to image denoising, spatial domain methods and transform domain methods. The main difference between these two categories is that a transform domain method decomposes the image by a chosen basis before further processing [3] while a spatial domain method processes the observed image data directly.
Transform domain methods have developed rapidly since Donoho's soft thresholding technique [4], which was introduced in 1995. The noise is considered as a highfrequency component in the transform domain for both fast Fourier transform (FFT) and discrete wavelet transform (DWT) and hence thresholding or truncating eliminates noise. The advantage of transform domain methods is that images often have sparse representations in transform domain. Thus dealing with the transform domain is very efficient.
Problem of image denoising can be summarized as follows:Let A(i,j) be the noisefree image and B(i,j) the image corrupted with independent gaussian noise Z(i,j),
, = , + , (1.1)
Z(I,j) has normal distribution N(0,1).In the wavelet domain the problem can be formulated as
Y(I,j)=W(I,j)+N(I,j) (1.2)
Where Y(I,j) is noisy wavelet coefficient;W(I,j) is true coefficient and N(I,j) is independent Gaussian noise.In this paper the performance of various image denoising algorithms is evaluated in terms of MSE.

Discrtete Wavelet Transform

Wavelet Thresholding Process
Wavelet thresholding for image denoising attempts to remove the noise present in the signal while preserving most of the signal characteristics, regardless of its frequency content. The complete process of image denoising is shown in figure 2.1 [5] and it involves the following steps:

Acquire the noisy digital signal.

Compute a linear forward discrete wavelet transform of the noisy signal.

Perform a nonlinear thresholding operation on the wavelet coefficients of the noisy signal.

Compute the linear inverse wavelet transform of the thresholded wavelet coefficients.
This simple fourstep process is known as wavelet thresholding or shrinkage.
Noisy Signal
Forward wavelet transform
Forward wavelet transform
Thresholding
Thresholding
Inverse wavelet transform
Inverse wavelet transform
Denoised Signal
Figure 2.1 Wavelet based image denoising [5]


Threshold Selection:
As one may observe, threshold determination is an important question when applying the wavelet thresholding scheme. A small threshold may yield a
result close to the input, but the result may be still be noisy. A large threshold produces a signal with a large number of zero coefficients. This leads to an overly smooth signal and smoothness generally suppresses the details and edges of the original signal and causes blurring and ringing artifact.

VisuShrink Algorithm
For VisuShrink algorithm [5], the wavelet coefficientsd of the noisy signal are obtained first.
Then with the universal threshold = 2 log n2 , ( is the noise level and n is the length of the noisy signal) the coefficients d= {di}, where i = 1, 2. . . n are shrinked according to the softshrinkage rule or soft thresholding methodgiven
(d) = sign di . di , di
0, di <

NeighShrink Algorithm
NeighShrink algorithm [9] threshold the wavelet coefficients according to the magnitude of the square sum of all the wavelet coefficients within the neighbourhood window.
It is based on the incorporating neighbouring wavelet coefficients with universal threshold. The NeighShrink algorithm is described as follows.

Incorporating Neighbouring Wavelet Coefficients
The wavelet transform can be accomplished by applying the lowpass and highpass filters on the same set of low frequency coefficients recursively. That means wavelet coefficients are correlated in a small neighbourhood. A large wavelet coefficient will probably have large coefficients at its neighbour locations. Therefore, Cai et al. [23] proposed the following wavelet denoising scheme for 1D signal by incorporating neighbouring coefficients into the thresholding process.
Let dj,k is the set of wavelet coefficients of the noisy 1D signal than in equation 2.1
, = (,) + (,) + (,+)(2.1)
If s2 j,k is less than or equal to2, then set the wavelet coefficient dj,k to zero. Otherwise, these coefficients shrink according to equation 2.2
, = ,(1/ (,)) (2.2)
Where = 2 log n. and n is the length of the signal. Note that the first (last) term in s2(j,k) is omit if dj,k is at the left (right) boundary of level j wavelet coefficients. For image denoising, the wavelet coefficients are arranged as a square matrix. For every
level of wavelet decomposition, first produce four frequency subbands, namely, LL, LH, HL, and HH. Since the Gaussian noise will be averaged out in the low frequency wavelet coefficients, so keep the small coefficients in these frequencies, only wavelet coefficients in the high frequency levels need to be threshold. That means only the high frequency subbands LH, HL and HH need to be thresholded. For every wavelet coefficient dj,k of our interest, so consider a neighbourhood window Qj,k around it [24] and choose the window by having the same number of pixels above, below, and on the left or right of the pixel to b threshold. That means the neighbourhood window size should be 3 Ã— 3, 5 Ã— 5, 7 Ã— 7, 9 Ã— 9, etc.
figure 2.2 illustrates a 3 Ã— 3 neighbourhood window centered at the wavelet coefficient to be thresholded. It should be mentioned in this algorithm that different wavelet coefficient subbands are threshold independently. This means when the small window surrounding the wavelet coefficient to be thresholded touches the coefficients in other subbands, we do not include those coefficients in the calculation. For 2D the square of summation around the window of wavelet coefficients is given by equation 2.3.
, = ,, (2.4)
Where the shrinkage factor can be defined as equation 2.5
, = ( / (,))+(2.5)
Here, the + sign in the formula means it takes nonnegative value, and = 2 log n2 is the threshold for the image. This thresholding formula is a
modification to the classical soft thresholding scheme
developed by Donoho and his coworkers [4]. The neighbourhood window size around the wavelet coefficient to be thresholded has influence on the denoising ability of this algorithm. The larger the window size, the relatively smaller the threshold, If the size of the window around the pixel is too large, a lot of noise will be kept, so an intermediate window size of 3 Ã— 3 or 5 Ã— 5 should be used. The neighbour wavelet image denoising algorithm can be described as follows:

Perform forward 2D wavelet decomposition on the noisy image.
,
,
(,) = ,
(,)(2.3)

Apply the proposed shrinkage scheme to threshold the wavelet coefficients using a neighbourhood
Where dj,k is the wavelet coefficient after 2D discrete wavelet transform and Qj,k is the window size centered at the wavelet coefficients to be thresholded as shown in figure 2.2.
3Ã—3 window Qj,k
Wavelet coefficient to be thresholded
Figure 2.2 An illustration of the neighbourhood window centered at the wavelet coefficient to be thresholded [9].
When the above summation has pixel indices out of the wavelet subband range, the corresponding terms in the summation is omitted.
For the wavelet coefficient to be thresholded [25], it is shrinked according to the following equation 2.4
window Qj,kand the universal threshold 2 log n2

Perform inverse 2D wavelet transform on the thresholded wavelet coefficients.
This algorithm is known as NeighShrink algorithm. Because VisuShrink algorithm kills too many small wavelet coefficients, so this shrinkage schemes gives the better result.


Limitation of NeighShrink Algorithm:
In the above mention that this algorithm is based on soft thresholding technique that is based on kill or shrink rule according to the wavelet coefficients and threshold value but it is use the universal threshold for every subbands. Normally in wavelet subbands, as the level increases the coefficients of the subband becomes smoother [1]. For example the subband HL2 is smoother than the corresponding subband in the first level (HL1) and so the threshold value of HL2 should be smaller than that for HL1. This is the limitation of this method which is use universal threshold for every subbands. This limitation is overcome in our proposed method. In propose proposed method we take the NeighShrink algorithm with different threshold value for different subbands which is based on Generalized Gaussian Distribution (GGD) modeling of subband coefficients.


Revised NeighdShrinkAlgorithm(proposed method)
In the NeighShrink algorithm different wavelet coefficient subbands are shrinked independently, but the threshold keep unchanged in all subbands. The shortcoming of this method is that the threshold in all subbands is suboptimal. The optimal of every subband should be datadriven and maximize the peak signal to noise ratio (PSNR). We will improve NeighShrink by determining an optimal threshold for every wavelet subbandwhich is based on Generalized Gaussian Distribution (GGD) [1] modeling of subband coefficients. In this proposed method, the choice of the threshold () estimation is carried out by analyzing the statistical parameters of the wavelet subband coefficients like standard deviation, arithmetic mean and geometrical mean as shown in equation 2.6
=C((AMGM) (2.6)
Here is the noise variance of the corrupted image [21],[22] .
The term C is depend on number of decomposition level and the level where the subband is available at that time which is given in equation 2.7.
C= ( ) (2.7)
Where, L is the no. of wavelet decomposition level, k is the level at which the subband is available.
The Arithmetic Mean and Geometric Mean of the subband matrix d(j,k) are given in equation 2.8 and 2.9.
m . m d(j,k)
=
1 1
1 1
decomposed in second step. HL1LL2, HL1LH2, HL1HL2, HL1HH2, be the subbands when HL1 is decomposed in second step and HH1LL2, HH1LH2, HH1HL2, HH1HH2 are the subbands when HH1 is decomposed. The total no. of subbands after second decomposition level is 16. After L decompositions, a total of D(L) = subbands are obtained. Where L is the no. of decomposition level.
LL1LL2
LL1HL2
HL1LL2
HL1HL2
LL1LH2
LL1HH2
HL1LH2
HL1HH2
LH1LL2
LH1HL2
HH1LL2
HH1HL2
LH1LH2
LH1HH2
HH1LH2
HH1HH2
Fig 2.3 Subband structure after two level packet decomposition.
(2) Compute the threshold value for each subband, except the approximate coefficients band using equation (2.5) after finding out its following
terms.Obtain the noise variance from equation (2.10)
j= k=
M2
(2.8)
Find the term C for each subband using equation [1] (2.7).Calculate the term AMGM for each subband
= =
= =
= [ . (,)](2.9)
Steps of Revised NeighShrink algorithm:
The Complete algorithm of proposed wavelet based image denoising technique is explained in the following steps.
(1) Perform the DWT of the noisy image using Mallat algorithm [18] upto L levels to obtain (3L+1) subbands, for L=2 levels subbands are named as HH1, LH1, HL1, HH2, LH2, HL2 and LL2.In figure 2.3 the LL1, LH1, HL1 and HH1 be the four subbands of image after first decomposition step and LL1LL2, LL1LH2, LL1HL2, LL1HH2 are the four subbands of image when LL1 subband is decomposed in second decomposition step. Similarly LH1LL2, LH1LH2, LH1HL2, LH1HH2 be the subbands when LH1 is
(except approximate coefficients subband) using equations (2.8) and (2.9).

Put the threshold value in equation [9] (2.5) of all subband coefficients (except approximate coefficients subband) for calculating the shrinkage factor. And then find the noiseless coefficient using equation (2.4)

Perform the inverse DWT to reconstruct the denoised image.The information from the four sub images is upsampled and then filtered with the corresponding inverse filters along the columns The two results that belong together are added and then again upsampled and filtered with the corresponding inverse filters. The result of the last step is added together in order to get the original image again.Hence there is no loss of information when the image is composed at full precision.



RESULTS AND DISCUSSION
Table 3.1 MSE of the noisy images and denoised images of standard image testpat1 using db5 wavelet
S.No
Noise levels
MSE of noisy images
MSE of denoised images using different algorithms
VS
NGS
RNGS
1.
10
100
190.2
66.8179
53.7066
2.
15
225
288.7
132.3688
108.4250
3.
20
400
384.4
214.8365
176.5279
4.
25
625
478.8
307.0986
255.7125
5.
30
900
577.1
402.2768
342.6551
6.
35
1225
680.6
489.7501
432.3564
7.
40
1600
793
576.6547
524.8454
8.
45
2025
915.1
664.2541
619.3196
9.
50
2500
1046.8
747.9031
717.5827
MSE Vs Noise level
MSE Vs Noise level
1200
1200
1000
1000
800
800
600
VS
NGS
600
VS
NGS
400
400
200
200
0
0
1 2 3 4 5 6 7 8 9
Noise Level
1 2 3 4 5 6 7 8 9
Noise Level
RNGS
RNGS
MSE
MSE
Figure 3.1 MSE of the noisy images and denoised images of standard image testpat1 using db5 wavelet.
Table 3.2 MSE of the noisy images and denoised images of standard image testpat1 using sym5 wavelet
S.No
Noise levels
MSE of noisy images
MSE of denoised images using different algorithms
VS
NGS
RNGS
1.
10
100
182.6
65.1276
52.9235
2.
15
225
278.1
130.1315
106.2353
3.
20
400
368.4
212.2812
172.9380
4.
25
625
460.8
302.2990
251.1762
5.
30
900
556.9
394.2354
335.7390
6.
35
1225
658.5
487.5670
425.5164
7.
40
1600
766.8
577.4514
516.2479
8.
45
2025
884.0
667.1109
609.0710
9.
50
2500
1011.6
759.1345
703.8926
MSE Vs Noise level
MSE Vs Noise level
1200
1200
1000
1000
800
800
600
VS
NGS
600
VS
NGS
400
400
200
200
0
0
1 2 3 4 5 6 7 8 9
Noise Level
1 2 3 4 5 6 7 8 9
Noise Level
RNGS
RNGS
MSE
MSE
Figure 3.2 MSE of the noisy images and denoised images of standard image testpat1 using sym5 wavelet.

Conclusion:
In this paper work, firstly a comparative analysis between the two conventional denoising algorithms i.e. VisuShrink and NeighShrink has been made. Out of these two algorithms NeighShrink gives the better MSEthan the other two algorithms.The conventional NeighShrink algorithm is modified by considering the different threshold value for different subbands that is based on Generalized Gaussian Distribution (GGD) modeling.The results have shown that the denoising of images through the Revised NeighShrink algorithm achieved enhancement in MSE.

Future Scope
The field of images processing has been growing at a very fast pace. The day to day emerging technology requires more and more revolution and evolution in the images processing field.
The work proposed in this paper also portrays a small contribution in this regard. The proposed denoising technique can provide a good platform for further research work in this respect.
Future work may be done for improving the Mean Square Error by considering the adaptive window size for every sub band over Proposed algorithm.
REFERENCES

D. Gnanadurai, and V. Sadasivam, An Efficient Adaptive Thresholding Technique for Wavelet Based Images denoising, International Journal of Signal Processing 2, 2006.

A. K. Jain, Fundamentals of Digital Image Processing, PrenticeHall Inc, Englewood Cliffs, 1989.

Shensa, M. J., The Discrete Wavelet Transform: Wedding the A Trous and Mallat Algorithms Signal Processing, IEEE Transactions on vol.40(10), pp.2464 2482, 1992.

D. L. Donoho, denoising by softthresholding, IEEE Transactions on Information Theory, vol.41, pp.613 627, 1995.

S. KotherMohideen, Dr. S. ArumugaPerumal, Dr.M.MohamedSathik, Images Denoising using Discrete Wavelet transform, International Journal of Computer Science and Network Security, vol.8, January 2008.

ShuangboFei, Ruizhen Zhao, Adaptive Wavelet Shrinkage For Images denoising Based On SURE Rule Signal Processing, 8th International Conference vol.1, 2006.

S. Sudha, G. R. Suresh, R. Sukanesh, Wavelet Based Image Denoising using Adaptive Thresholding, International Conference on Computational Intelligence and Multimedia Applications. vol.3, pp.296 300, 2007.

Li Zhen, He Zhengjia, ZiYanyang, Wang Yanxue, Customized wavelet denoising using intra and interscale dependency for bearing fault detection, Journal of Sound`and Vibration.

Chen G Y, Bui T D, Krzyzak, A Images denoising Using Neighbouring Wavelet Coefficients, Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing ICASSP. Canada: pp.917920, 2004.

TinkuAcharya, Ajoy. K. Ray, Images processing principles and applications, Hoboken, New Jersey, A John Wiley and Sons, MC. Publication, 2005.

Detail information about the Multiresolution Analysis and The Continuous wavelet transform6,
<http://users.rowan.edu/~polikar/WAVELETS/WT part3. html>.

Detail information about the Fundamental Concepts of Transformations, http://users.rowan.edu/
~polikar/WAVELETS/WTpart1.html

Dongwook Cho, Tien D. Bui, and Guangyi Chen, Images denoising based on wavelet shrinkage using neighbor and level dependency, International Journal of Wavelets, Multiresolution and Information Processing vol. 7, no. 3, pp.299311, 2009.

Detail information about Multiresolution Analysis: The Discrete wavelet transform8,<http://users. rowan.edu/~polikar/WAVELETS/WTpart4.html>

Bruce W. S, Multirate and wavelet signal processing, Academic press, New York, 1996.

A. Graps, An Introduction to Wavelets, IEEE Journal of ComputationalScience and Engineering, vol.2(2), pp.50 61, 1995.

Dong Wei, Alan C. Bovik, and Brian L. Evans,Generalized Coiflets: A New Family of Orthonormal Wavelets.

S. G. Mallat, A theory for multiresolution signal decomposition: The wavelet representation, IEEE Trans. Pattern Anal. Mach. Intell. 11 (1989) pp.674693, 1989.

Abdulla Al Muhit, Md. Shabiul Islam and Masuri Othman, VLSI Implementation of Discrete Wavelet Transform (DWT) for Image Compression, 2nd International Conference on Autonomous Robots and Agents December 1315, 2004.

S. G. Chang, B. Yu and M. Vetterli, Adaptive wavelet thresholding for images denoising and compression, IEEE Trans. Images Processing. vol. 9, pp.15321546, 2000.

Maher A. SidAhmed., Images ProcessingTheory algorithm and aarchitecture, McGrawHill, pp.7880, 1995.

S. Grace Chang, Bin Yu and M. Vattereli, Spatially Adaptive Wavelet Thresholding with Context Modeling for Image Denoising, IEEETransaction – Images Processing, vol.9, pp.15221531, 2000.

T. T. Cai and B. W. Silverman, Incorporating information on neighbouring coefficients into wavelet estimation, The Indian Journal of Statistics, vol. 63, Series B, Pt. 2, pp.127148, 2001.

Jianxin Dai, images denoising based on combining neighbouring wavelet coefficients, Images and Signal Processing, CISP '09. 2nd International Congress, pp.1 3, 2009.

G. Y. Chen, T. D. Buiand A. Krzyzak, Images denoising using neighbouring wavelet coefficients, Integrated ComputerAided Engineering. 2005.

G. Y. Chen and T. D. Bui, Multiwaveletsdenoising using Neighbouring Coefficients, IEEE Signal ProcessingLetters, vol.10, no.7, pp.211214, 2003.