Color Video Denoising Using 3D Framelet Transform

International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com

Volume 2, Issue 3, May June 2013 ISSN 2278-6856

Volume 2, Issue 3 May June 2013 Page 130

Abstract: Digital video denoising is highly desirable process
in order to improve the video quality. Many transformations
are used for denoising process among of them is framelet
transform (Double-Density Wavelet Transform) which is a
perfect in denoising process by avoiding the problems in the
other transformations. In this paper we propose a mixing
subbands and thresholds methods to denoise a noisy video
with Gaussian white noise type. This method named Mixing
Subbands and Double Wiener filter Thresholding methods
(MSDWTM) , it is applied first with double local wiener filter
then with one local and one global wiener filter . This method
is performed first with two dimensional fast discrete framelet
transform (2-D DDWT) and then with three dimensional fast
discrete framelet transform (3-D DDWT) .The results show
that our (MSDWTM) gives a better denoising results
comparative with the original methods.

Keywords: Framelet, two dimensional Double-Density
Wavelet Transform (2-D DDWT) , three dimensional
Double-Density Wavelet Transform (3-D DDWT), hard
threshod, soft threshold, semisoft threshold .

1. INTRODUCTION
Video has become a part of our everyday life; think of
television broadcast for example. Many other video
applications include: videophone, teleconferencing,
satellite observations, autonomous navigation,
astronomical and medical imaging[1] .On the other hand,
video is the technology of electronically capturing,
recording,processing, storing, transmitting, and
reconstructing a sequence of still images representing
scenes in motion [2]. These image sequences are often
corrupted by noise, for example, due to bad reception of
television pictures. Some noise sources are located in a
camera and become active during image acquisition
under bad Lightning conditions. Other noise sources are
due to transmission over analogue channels [1].So, Video
denoising plays an important role in video processing. It
is critical, since the presence of noise in a video sequence
degrades both its visual quality and the effectiveness of
subsequent processing tasks [3]. Also it is highly
desirable in order to enhance the overall perceptual
quality, increase compression effectiveness, facilitate
transmission bandwidth reduction, and facilitate accuracy
in processes like feature extraction and pattern
recognition that might be involved [2]. So, image
processing in some of its application needs a
transformation process to solve problems. The
transformation is a process that translates an object from
a given domain to another in order to have some
important implicit information, which can be used for its
recognition. One of the conventional transformation is the
Fourier Transform (FT) which usually transforms the
signal from its time domain to the frequency domain.
Another form of transformation was a Cosine Transform
(CT), like the (FT), CT uses sinusoidal basis functions.
The difference is that the CT basis functions are not
complex; they use only cosine functions and not sine
functions [4]. The Discrete wavelet transform(DWT) is a
powerful tool to analyze the local information of a signal
and obtain a space-frequency description of it [5], but it
suffers from three major disadvantages, (1) Shift-
sensitivity, (2) Poor directionality, and (3) Lack of phase
information. These disadvantages severely restrict its
scope for certain signal and image processing
applications (e.g. edge detection , image registration /
segmentation , motion estimation ) [6]. Other extensions
of standard DWT such as Wavelet Packet Transform
(WPT) and Stationary Wavelet Transform (SWT) reduce
only the first disadvantage of shift- sensitivity but with
the cost of very high redundancy and involved
computation. Recent research suggests the possibility of
reducing two or more of these disadvantages [7]
.Introducing the Framelet Transform (FrT) which is
called in some references Double-Density Wavelet
Transform (DDWT) as the tight-frame equivalent of
Daubechies orthonormal wavelet transform; the wavelet
filters are of minimal length and satisfy certain important
polynomial properties in an oversampled framework.
Because the DDWT, at each scale, has twice as many
wavelets as the DWT, it achieves lower shift sensitivity
than the DWT [7] .
2. RELATED WORKS
Selesnick I., and Wagner C., (2004) worked on the
development and implementation of double-density DWT
(Framelet Transform) for 1-D signals and 2-D images
with applications of image denoising[8]. In (2008) Al-
Taai H. N., propose an algorithm for computing Fast 1-
Color Video Denoising using 3D Framelet
Transform
Assist.Prof.Dr. Matheel E.

Abdulmunim
1
, Rabab F.

Abass
2

1
University of Technology, computer sciences dept, Head of Multimedia part,
Al-Sina'a Street , Tel Muhammed, Iraq ,Capital Baghdad

2
University of Technology, computer sciences dept. master student,



D, 2-D (FrT) [9]. Finally in (2012) S.Sulochana and
R.Vidhya proposed the denoising method which uses
(FrT) to decompose the image and performed shrinkage
operation to eliminate the noise using different thresholds
methods [10] .
3. WIENER FILTER
The Wiener filter, also called a minimum mean square
estimator (developed by Norbet Wiener in 1942),
alleviates some of the difficulties inherent in the another
filters by attempting to model the error in the restored
image through the use of statistical methods. After the
error is modeled, the average error is mathematically
minimized, thus the term minimum mean square
estimator was proposed. The resulting equation is the
Wiener filter [4]:

) , (
) , (
) , (
) , (
) , (
2
(
+
=
-
v u S
v u S
v u H
v u H
v u
W
R
I
n

where:
) , ( v u H
-
is the complex conjugate of
) , ( v u H
.
2
) , ( ) , ( v u I v u S
I
=
is the power spectrum of the original
image.

2
) , ( ) , ( v u N v u S
n
=
is the power spectrum of the noise.
This equation assumes a square image of size N N . The
complex conjugate can be found by negating the
imaginary part of a complex number. If we assume that
the noise term ) , ( v u Sn is zero, this equation reduces to an
inverse filter since
) , ( ) , ( ) , (
2
v u H v u H v u H
-
=
, as the
contribution of the noise increases, the filter gain
decreases.
This seems reasonable, in portions of the spectrum
uncontaminated by noise we have an inverse filter,
whereas in portions of the spectrum heavily corrupted by
noise, the filter attenuates the signal, with the amount of
attenuation being determined by the ratio of the noise
spectrum to the uncorrupted image spectrum.
The Wiener filter is applied by multiplying it by the (FT)
of the degraded image, and the restored image is obtained
by taking the inverse (FT) of the result, as follows[4]:
)] , ( D ) , ( R [ F )] , (
[ F ) , (
W
-1 -1
v u v u v u I c r I = =

4. FRAMELET TRANSFORM(FRT)
Framelet are very similar to wavelets but have some
important differences. In particular, whereas wavelets
have an associated scaling function () and wavelet
function () , framelets have one scaling function ()
and two wavelet functions 1() and 2() .
The scaling function () and the wavelets 1() and
2() are defined through these equations by the low-pass
(scaling) filter h
0
(n) and the two high-pass (wavelet)
filters h
1
(n) and h
2
(n). equation (3) and (4) represents the
two functions [9].
) 2 ( 2 ) ( ) ( k t k h t
k
=
) 2 ( 2 ) ( ) ( k t k g t
k
=
4.1 Computation Method of Framelet Transform for 2-
D Signal
A 2-D separable transform is equivalent to two 1-D
transforms in series. It is implemented as 1-D row
transform followed by a 1-D column transform on the
data obtained from the row transform. To compute a
single level discrete framelet transform for 2-D signal
using separable method the next step should be followed
[9] :
1. Checking input dimensions: Input matrix should be of
length NxN, where N must be even and N >=length
(analysis filters).
2. Construct a transformation matrix using
transformation matrices .
3. Transformation of input rows by apply matrix
multiplication to the 3N/2xN constructed transformation
matrix by the NxN input matrix as in equation (5).

| | | |
N N N
N
X W Y

= .
2
3

4. Transformation of input columns: can be done as
follows:
a. Transpose the row transformed (3N/2)xN
matrix resulting from step 3.
b. Apply matrix multiplication to the
(3N/2)xN constructed transformation matrix by
the NxN column matrix.

| | | |
2
3
2
3
.
N
N N
N
Y W YY

' =

The final discrete framelet transformed matrix is equal
to:

| |
2
3
2
3 N N
o
Y Y Y

' =

4.2 Computation Method of Framelet Transform for 3-
D Signal
The structures are defined in 3-D and the transformation
algorithm is applied in x-, y- and z-direction:
Lets take a general 3-D signal, for example any NxNxM
matrix, and apply the following steps[7]:
1. Let X be the NxNxM input 3-D signal.

(1)
(3)
(4)
) 5 (
) 6 (
) 7 (
) 2 (


2. Apply 2D fast discrete framelet transform algorithm
to each NxN input matrix, which result in a
(3N/2)x(3N/2) x M matrix.

3. Apply 1-D DDWT algorithm to each of the (3N/2) x
(3N/2) elements in all M matrices in z-direction, which
can be done as follows:

a. For each i,j construct the Mx1 input vector
Y(i,j)=[ ai,j bi,j ci,j di,j ]t 1xM where i,j =0,1,2,..,3N/2
b. Construct an (3N/2)xN transformation matrix;
using transformation matrices.
c. Apply matrix multiplication to the (3N/2)xN
constructed transformation matrix by the Mx1
input vector.
4. Repeat step 3 for all i, j to get YY matrix (3N/2) x
(3N/2) x (3M/2) matrix.
4.3 Computation of Inverse Framelet Transform for
2-D Signal
To compute a single level 2-D inverse discrete framelet
transform the next steps should be followed[6] :
1. Let Yo be the 3N/2 3N/2 framelet transformed
matrix.
2. Construct N 3N/2 reconstruction matrix, T=WT,
using transformation matrices.
3. Reconstruction columns: By applies matrix
multiplication to the Nx3N/2 reconstruction matrix by
the 3N/2
3N/2 framelet transformed matrix.

| | | |
2
3
2
3
2
3
.
N N
o
N
N
Y T YY

=
1. Reconstruction rows: can be done as follows:
a. Transpose the column reconstructed matrix
resulting from step 3.
| |
N
N
Y Y Y
' =
2
3

b. Apply matrix multiplication by multiplying the
reconstruction matrix with the resultant transpose
matrix.
| | | |
N
N N
N
Y T X

=
2
3
2
3
.
4.4 Computation of Inverse Framelet Transform for
3-D Signal
To compute a single level IDDWT for 3-D signal the next
steps should be followed [7]:
1. Let Y be the (3N/2) (3N/2) (3M/2) framelet
transformed matrix
2. Construct M (3M/2) reconstruction matrix. .
3. Apply 1-D IDDWT algorithm to each of the (3N/2)
(3N/2) elements in all (3M/2) matrices in z-direction.
4. Construct (Nx3N/2) reconstruction matrix, T=W
T

using transformation matrices .
5. Apply 2-D IDDWT algorithm to each (3N/2) (3N/2)
result matrix from step 3.
5. TYPE OF THRESHOLDING
Thresholding is one of the most commonly used
processing tools in wavelet signal processing. It is widely
used in noise reduction, signal and image compression or
recognition [11].

5.1 Hard Threshold
Hard Thresholding is also called "kill / keep" strategy
[12] or "gating" [13]. If the signal or a coefficient value is
below a present value it is set to zero, that is [11]:

s
>
= =
. | | , 0
; | | ,
) , (
Thv G
Thv G G
Thv G T X
j
k
j
k
j
k j
k h
j
k

where Thv is the threshold value or the gate value.
Hard thresholding can be described as the usual process
of setting to zero the wavelet coefficients whose absolute
values are less than or equal to the threshold value Thv.

5.2 Soft Threshold
Soft Thresholding is an alternative scheme of hard
thresholding and can be stated as[12]:
s
>
= =
. | | , 0
; | | ), | (| * ) (
) , (
Thv G
Thv G Thv G G sign
Thv G T X
j
k
j
k
j
k
j
k j
k s
j
k

Where :

0. 1
; 0 G if , 0
; 0 , 1
) (
j
k
<
=
> +
=
j
k
j
k
j
k
G if
G if
G sign

Soft thresholding is an extension of hard thresholding,
firstly setting to zero the wavelet coefficients whose
absolute value are less than or equal to Thv, then
shrinking the non zero coefficients towards zero by a
threshold value Thv[12].

5.3 Semisoft Threshold
Bruce and Gao showed that hard thresholding would
cause a bigger variance, while soft thresholding will tend
to have a bigger bias because all larger coefficients are
reduced by Thv. To prevent the drawback of hard and soft
thresholding, they proposed a semi-soft thresholding
approach as given in the following equation:

{ }
| |
if ) (

0
) , (
>
s <
s
= =
Thv G if G
Thv G Thv
Thv Thv
Thv G Thv
G sign
Thv G if
Thv G T X
j
k
j
k
j
k
j
k j
k
j
k
j
k soft semi
j
k

) 8 (
) 9 (
) 10 (
) 11 (
) 12 (
) 13 (


The aim of semi-soft threshold is to offer a compromise
between hard and soft thresholding by changing the
gradient of the slope. This scheme requires two
thresholds, a lower threshold Thv and an upper
threshold Thv , where Thv is estimated to be twice the
value of lower threshold Thv . There is no attenuation for
inputs beyond Thv . For inputs below or equal to Thv ,
the output is forced to zero. For inputs that lie between
Thv and Thv the output depends on the gradient
formula:
{ }
Thv Thv
Thv G Thv
G sign
j
k
j
k
) (
[13].
6.THE PROPOSED METHOD (MSDWTM)
The general block diagram of video denoising using 2-D
MSDWTM is explained in figure (6.1), where the digital
video will be converted to frames, adding Gaussian white
noise then perform the proposed 2-D MSDWTM (which
will be explained in detail in figure (6.2) ) , the final step
is to test the denoised frame according to RMSE,SNR and
PSNR measurement .

Figure 6.1: main block diagram of video denoising using
2-D MSDWTM
The framework of 2-D MSDWTM is illustrated in Figure
(6.2), the algorithm is stated bellow:
1. The same noisy frame is denoised using double local
wiener filter 33 and 99 kernel .
2. Apply 2-D DDWT decomposition on each of them to
introduce eight sub bands of each of them (LL, LH1,
LH2, H1L, H2L, H1H1, H1H2, H2H1 and H2H2) .
3. Perform mixing sub band (replace H1H1, H1H2, H2H1
and H2H2 sub band from D2 with H1H1,H1H2,H2H1 and
H2H2 from D1) .
4. Apply the threshold methods on the detail sub bands
LH1,LH2,H1L,H2L,H1H1,H1H2,H2H1 and H2H2) .
5. Finally perform inverse 2-D DDWT to produce the
denoised frame.

Figure 6.2: The framework of 2-D MSDWTM
The general block diagram of video denoising using 3-D
MSDWTM is explained in figure (6.3), where the digital
video will be converted to frames, adding Gaussian white
noise then perform the proposed 3-D MSDWTM (which
will be explained in detail in figure (6.4) ) , the final step
is to test the denoised frames according to RMSE,SNR
and PSNR measurement .

Figure 6.3: main block diagram of video denoising using
3-D MSDWTM

The framework of 3-D MSDWTM is illustrated in Figure


(6.4), the algorithm is stated bellow:
1. The same eight noisy frames is denoised using double
local wiener filter 33 and 99 kernel for each of them .
2. Apply 3-D DDWT decomposition to introduce two
groups of decomposed frames (D1, D2) each of them has
twelve frames .
3. Perform mixing sub bands (replace the blue sub bands
from D2 with the red sub bands from D1) .
4.Apply the threshold methods (LLL without
thresholding) .
5.Finally perform inverse 3-D DDWT to produce the
denoised eight frames.
Note that W1 is local wiener filter [3,3] and W2 is local
wiener filter [9,9].

Figure 6.4: The framework of 3-D MSDWTM
7.THE RESULTS
In this study, we test the result of 2-D DDWT and 3-D
DDWT with coefficients of filter listed in table (7.1) .

Table 7.1: The coefficients of a wavelet tight frame

We use two color AVI video types, the frames have been
resized to 128128 corrupted by Gaussian white noise
type, the three types of the thresholds (hard,soft and
semisoft) applied on the whole decomposed frame with
optimal threshold value. The threshold that producing the
minimum RMSE is the optimal one.
Table (7.2) shows the results of MSE and SNR using 2-D
DDWT from two AVI color video Vipmen and
Warnsigns respectively with noise level =25 .

Table 7.2: RMSE and SNR of 20
th
frame from two AVI
color video Vipmen and Warnsigns, denoised using 2-
D DDWT ,with noise level =25 .

Table (7.3) shows the results of RMSE and SNR using 3-
D DDWT from two AVI color video xylophone and
shuttle respectively for 20
th
frame (which is the first
frame of the noisy eight consecutive frames 20
th
,21
th
,22
th
, 23
th
,24
th
,25
th
,26
th
and 27
th
) .

th
frame from two AVI
color video Vipmenand Warnsigns, denoised using 3-D
DDWT ,with noise level =25 .



Table (7.4) shows the results of RMSE and SNR using the
proposed 2-D MSDWTM from two AVI color video
Vipmen and Warnsigns respectively.

th
frame from two AVI
color video Vipmen and Warnsigns, denoised using 2-D
MSDWTM ,with noise level =25 .

In the above 2-D MSDWTM ,the noisy frame denoised
with doubly local wiener filter ,when it denoised with one
local wiener filter and with another global wiener filter, it
gives a better result as can be seen in table (7.5) .

th
frame from two AVI
color video Vipmen and Warnsigns, denoised using 2-
D MSDWTM , with one local and one global wiener filter
with noise level =25

Figure (7.1) gives the denoised 20
th
frame of Vipmen and
Warnsigns respectively using 2-D MSDWTM with
three types of thresholds (Hard, Soft and Semisoft) .

(a) Original frames

(b) Noisy frames ( =25)

(c) denoised frames using Hard threshold

(d) denoised frames using soft threshold

(e) denoised frames using semisoft threshold
Figure 7.1 : The 20
th
frame from two AVI color video
Vipmen and Warnsignsrespectively, denoised using 2-
D MSDWTM,with noise level =25 .

Table (7.6) shows the results of RMSE and SNR using the
proposed 3-D MSDWTM from two AVI color video
Vipmen and Warnsigns respectively .

Table 7.6 : RMSE and SNR of 20
th
frame from two AVI
color video Vipmen and Warnsigns denoised using 3-D
MSDWTM ,with noise level =25 .

Noisy video name Denoising by 3-D DDWT
RMSE SNR dB
Threshold
type
RMSE
SNR dB

Vipmen
25.0963 14.1906
Hard 8.4700 23.4297
Soft 11.4730 20.3777
Semisoft 8.2410 23.6646
Warnsigns
25.0903 15.9045
Hard 10.6882 23.1562
Soft 12.9969 21.1432
Semisoft 10.3894 23.4047


In the above 3-D MSDWTM, the noisy frame denoised
with doubly local wiener filter ,when it denoised with one
local wiener filter (w1) and with another global wiener
filter (w2) for each of eight frames ,it gives a better result
as can be seen in table (7.7) .

th
frame from two AVI
color video Vipmen and Warnsigns, denoised using 3-D
MSDWTM , with one local and one global wiener filter
with noise level =25 .

Figure (7.2) gives the denoised 20
th
frame of Vipmen and
Warnsigns respectively using 3-D MSDWTM with
three types of thresholds (Hard, Soft and Semisoft) .

(a) Original frames

(b) Noisy frames ( =25)

(c) denoised frames using Hard threshold

(d) denoised frames using soft threshold

(e) denoised frames using semisoft threshold
Figure 7.2 : The 20
th
frame from two AVI color video
Vipmen and Warnsignsrespectively , denoised using
3-D MSDWTM, with noise level =25 .

8. CONCLUSIONS AND SUGGESTIONS FOR
FUTURE WORKS
We concluded from the above previous results that 3-D
DDWT gives a better results than 2-D DDWT, because
the additional dimension (z) of transformation dealing
with temporal correlations between video frames. In the
proposed 2-D MSDWTM we concluded that by using the
double wiener filter with mixing the high sub band gives
a better results in subjective and objective than the
original 2-D DDWT. Also the proposed 3-D MSDWTM
achieve the better results than 3-D DDWT when one local
and one global wiener filter for the same frame was used.
Achieving good result than the double local wiener filter
was get.
From the earlier result and discussion, one can develop
the 3-D DDWT to 4-D DDWT, where the fourth
dimension is the local angle this improvement may
attenuate the noise .

References
[1] Abed AL-Ghanee F. Z., "Digital Video Enhancement
and Denoising ", M.Sc. Thesis, Univ. of Technology,
Computer Science and Information Sys. Dep.,
October 2006.

[2] Raiamani A., Krishnaveni, V., and Ferose, H. W., "
A new Denoising Approach for the Removal of
Impulse Noise from Color Image and Video
Sequences", Image Anal Stereol, Vol. 31, PP 185-
191, 2012.

[3] Xue G.," Temporal Denoising of High Resolution
Video", M.Sc. Thesis, Univ. of McMaster, Electrical
and Computer Eng. Dep. 2009.

[4] Umbangh S. E., "Computer Vision and Image
Processing", A practical approach using CVIP tools,
Prentice Hall, Inc., Upper Saddle River, NJ 07458
(U.S.A.), 1998.

[5] Bahich M., Bailich, M., Imloul, A., Afifi, M., and
Barj, E.," A Comparative Study of One and Two-
Dimensional Wavelet-Based Techniques for Noisy
Fringe Patterns Analysis", Optics Communications,
Vol. 290, PP 4348, 2013.

[6] Naher A. K.," A Framelet Based Approach for
Image Denoising", M.Sc. Thesis, Univ.of
Technology,Electrical and Electronic Eng.Dep.,July
2008 .

[7] Kadhim W. Q.," Optical Flow Estimation Using
Framelet Transform", M.Sc. Thesis, Univ. of
Technology, Computer and Information Iechnology
Eng.Dep., January 2010.


[8] Selesnick I., and Wagner, C., "Introduction of
Double-Density Wavelet Software"Electrical
Engineering, Polytechnic University, 2004 .

[9] Al-Taai H. N., A Novel Fast Computing Method
for Framelet Coefficients American Journal of
Applied Sciences Vol.5, No. 11, PP. 1522-
1527,2008.

[10] Sulochana S., and Vidhya, R. " Image Denoising
using Adaptive Thresholding in Framelet Transform
Domain" International Journal of Advanced
Computer Science and Applications,,Vol.3, NO. 9
PP. 192-196, 2012 .

[11] Goswami J. C., and Chan A. K., "Fundamentals of
Wavelets Theory, Algorithms, and Applications",
John Willy and Sons, 1999.

[12] Abdulmunim M. E. A., " Color Image Denoising
Using Discrete Multiwavelet Transform", Ph.D.
Thesis, Univ. of Technology, Computer Science and
Information Sys. Dep., 2004.

[13] Jakiwer M.J.,"Quantitative Analysis of Denoising
SAR Images ", M.Sc. Thesis, Univ. of Technology,
Computer Science,Dep., Dec.2003

Color Video Denoising Using 3D Framelet Transform

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Color Video Denoising Using 3D Framelet Transform

Diunggah oleh

Hak Cipta:

Format Tersedia

International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com

3N/2 framelet transformed matrix.

Anda mungkin juga menyukai