SlideShare a Scribd company logo
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 273
A Hybrid SVD Method Using Interpolation Algorithms for
Image Compression
J. Izadian , A. Hosaini Jalal_Izadian@yahoo.com
Department of Mathematics, Faculty of Sciences,
Mashhad Branch, Islamic Azad University,
Mashhad, Iran
M. Jalili Jalili.maryam@yahoo.com
Department of Mathematics,
Neyshabur Branch, Islamic Azad University,
Neyshabur, Iran
Abstract
In this paper the standard SVD method is used for image processing and is combined with some
interpolation methods as linear and quadratic interpolation for reconstruction of compressed
image.The main idea of the proposed method is to select a particular submatrix of main image
matrix and compress it with SVD method, then reconstruct an approximation of original image by
interpolation method. The numerical experiments illustrate the performance and efficiency of
proposed methods.
Keywords: SVD Method, Interpolation Method, Image Reconstruction, Compressing Image,
Lossless Compression, Lossy Compression.
1. INTRODUCTION
Image compression techniques plays an important role in transmision and storage of image
information in computer science and related domains.The goal of image compression is to obtain
a representation that minimizes the bits volume while still maintaining the important meaning and
the intrinsic structure of the original image. Image compression techniques can be classified into
two groups. Lossless compression and lossy compression.
In lossless compression the reconstructed image is identical to the original one and deduces a
low compression ratio while lossy compression methods allow a loss in the actual image data. So
the original image cannot be created excactly from the compressed image. But these methods
deduce high compression ratio. There is many ways to compress depending on the application
field. One popular method for compressing an image is the wavelet teqnique [1]. Wavelet
functions form an orthonormal basis on which it is possible to project each data set. Another
important method is the Singular Value Decomposition (SVD). This technique is based on the
factoriziation of the real matrix of image in three matrices that can be used to reconstruct the
main image or an approximation of it, for more information see [2,3,4].
In this paper the SVD method is combined with two interpolation processes, that are linear 2D
interpolation (triangular interpolation), and bilinear interpolation, to decrease the volume of
transfered image . This paper is orgonized as follows. Section 2 is specified to describe the basic
concepts of proposed method. Section 3 presents the numerical experiments . Finally section 4
concludes the paper with discussion.
2. THE BASIC CONCEPTS OF METHOD
Suppose that a nm× pixel gray-scale image is given. Each pixel having some level of black and
white given by some integers that can be selected between 0 and 255 or a real number
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 274
between zero and one. In the case of integer, each integer requires approximately one byte to
store. Then the resulting image has approximately nm× bytes volume. If the image is coloured
in (RGB) system the image contains three nm× matrices for red, green, and blue colours, then
nm××3 pixels image. The Singular Value Decomposition (SVD) for a gray-scale image for
compressing nm× pixels image matrix is based on the following decomposition of A ,
.= T
USVA
Actually, recalling from advanced linear algebra one has the following theorem, that is
demonstrated in advanced linear algebra (for example see [5,6]).
Theorem 2.1 Any nonzero real nm× matrix A with rank 0>r , there are an orthogonal mm×
matrix U , and an orthogonal nn× matrix V such that SAVUT
= is an nm× "diagonal
matrix" of the following form
,
00
0
= 




D
S
0,>...),,...,(= 211 rrdiagD σσσσσ ≥≥
where rσσσ ,...,, 21 are non-zero sigular values of A . This factorization is called the Singular
Value Decomposition (SVD) of A .
By using above theorem and assuming that
,1,2,...,=],......[= 121 miuuuuuU mrr +
where miui 1,...,=, are the column of U , by considering the orthogonality of U , one has



≠ .0
=1
==
ji
ji
uu ijj
T
i δ
Similarly for orthogonal nn× matrix ,V by considering.
,= 2
1














T
n
T
T
T
v
v
v
V
K
where iv are for ni 1,...,= are columns of ,V it yields



≠ .0
=1
==
ji
ji
vv ijj
T
i δ
Here, S is nm× diagonal matrix with singular values of A on the diagonal of D in rectangular
matrix S . The matrix S can be presented by the following matrix:
1
2
0 0 0
0 0
0 0 0
0 0 0
r
S
σ
σ
σ
 
 
 
 =
 
 
  
K
M M O M
K
K K
For =1,2,...,i n , iσ are singular values of matrix A. It is proved in Theorem(2.1) that
1 2 ... 0rσ σ σ≥ ≥ ≥ > and 1 2 ... 0r r nσ σ σ+ += = = = . For =1,2,...,i m , iσ are called singular values of
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 275
matrix
T
A .The vector iv for =1,2,...,i n are called the right singular vector of A , and iu
for =1,2,...,i m . are the left singular vectors of A (see[4,5]). then
1
2
1 2 1 1 1 1 2 2 2
.
[ ... ] .
.
.
T T T T
m m r r r
n
v
v
A USV u u u u S u v u v u v
v
σ σ σ−
 
 
 
 
= = = + + + 
 
 
 
 
K
When compressing the image, the sum is not performed to the very last . The singular values
with small enough values are dropped. A matrix of rank k is obtained by truncating these sums
after the first k terms, denoted by
,= 221111
T
kkk
TT
k vuvuvuA σσσ +++ K (1)
is chosen as approximation of A . The total storage volume for kA will be 1)( ++ nmk . This
matrix is an approximation of A , that can be used as an approximation of compressed image A .
For having a criterion for storage volume decrease, the compression ratio is defined as follows:
.
1)(
=
++ nmk
mn
CR
We also use the Schur norm to measure the quality of obtained image kA , called Mean Square
Error (MSE), which is introduced by following expression:
( ) 2
2
=1 =1
1 1
= = ( ) .
m n
k
K ij ij
i j
MSE A A a a
mn mn
− −∑∑
where nmijaA ×)(= and .)(= )(
nm
k
ijk aA × An error value can be defined related to sum of iσ by
following relation:
,1=
1=
1=
i
r
i
i
k
i
ke
σ
σ
∑
∑
−
A typical choice of k is so that the storage space required for kA will be less than .
5
1
For having
an superior compression ratio, one can utilize a submatrix of A for small k and use an
interpolation process to reconstruct kA .
Consider m and n are even, and A
~
is a submatrix of order
22
nm
× of A , by omitting the
element of A which are situated on rows or column of odd numbers. This matrix is approximated
by SVD method with a suitable value of k . Then an interpolation method is used to reconstruct
an approximation of kA , that can be chosen as approximation of kA as obtained image. In linear
case, consider the chosen submatrix of A as:
,)~(=
~
22
)(
nm
k
ijk aA
×
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 276
where m and n are even numbers and the elements of k can be determined from
nm
k
ijk aA ×)(= )(
by the following relations:
,=)~(,=)~( )(
31)1,2(1)2(
)(
1,
)(
11)1,2(1)2(
)(
,
k
ji
k
ji
k
ji
k
ji aaaa +−+−++−+−
,=)~(,=)~( )(
31)3,2(1)2(
)(
11,
)(
11)3,2(1)2(
)(
1,
k
ji
k
ji
k
ji
k
ji aaaa +−+−+++−+−+
.
2
1,2,...,=,
2
1,2,...,=
n
j
m
i
The matrix kA
~
is of order ,
22
nm
× therefore it has
4
3
less element than kA . This matrix is
chosen as the compression of A that results a compression ratio given by
,
2)(
2
=
1)
22
(
=
++++ nmk
mn
nm
k
mn
CR
Therefore, in this method the compression ratio is doubled in comparison with kA . The matrix kA
~
is supposed as a base for interpolating method to reconstruct an approximation of kA . In the
linear case, this matrix is calculated with the following equations:
,)(=, nmijkkk aAAA ×
∗∗∗
;
,~=,~= )(
1,11)3,2(1)2(
)(
11)1,2(1)2(
k
jiji
k
ijji aaaa +
∗
+−+−
∗
+−+−
,~=,~= )(
11,31)3,2(1)2(
)(
1,31)1,2(1)2(
k
jiji
k
jiji aaaa ++
∗
+−+−+
∗
+−+−
)/2,~~(=)/2,~~(= )(
1,
)(
,21)1,2(1)2(
)(
,
)(
1,11)2,2(1)2(
k
ji
k
jiji
k
ji
k
jiji aaaaaa +
∗
+−+−+
∗
+−+− ++
)/2,~~(=)/2,~~(= )(
11,
)(
1,31)2,2(1)2(
)(
11,
)(
1,21)3,2(1)2(
k
ji
k
jiji
k
ji
k
jiji aaaaaa +++
∗
+−+−+++
∗
+−+− ++
]/4.~~~~[= )(
11,
)(
,
)(
1,
)(
1,21)2,2(1)2(
k
ji
k
ji
k
ji
k
jiji aaaaa ++++
∗
+−+− +++
In order to describe the hybrid SVD using bilinear interpolation, the following submatrix of kA is
considerd
,)~(=
~
33
)(
nm
k
ijk aA
×
where kA is considered a nm× matrix which m and n are the integers that are divisible by 3.
Then kA
~
can be defined as follows
,)(= )(
nm
k
ijk aA ×
,=)~(,=)~( )(
41)1,3(1)3(
)(
1,
)(
11)1,3(1)3(
)(
,
k
ji
k
ji
k
ji
k
ji aaaa +−+−++−+−
,=)~(,=)~( )(
41)4,3(1)3(
)(
11,
)(
11)4,3(1)3(
)(
1,
k
ji
k
ji
k
ji
k
ji aaaa +−+−+++−+−+
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 277
.
3
1,2,...,=,
3
1,2,...,=
n
j
m
i
Since the matrix kA
~
is a matrix of order ,
33
nm
× the compression ratio is obtained by
,
3)(
3
=
1)
33
(
=
++++ nmk
mn
nm
n
mn
CR
Now for each 44× block of matrix kA which is reduced to 22× block matrix kA
~
, the following
bilinear interpolation is applied on ,[0,1]2
11),(,
2
1=
2
1=
),( ~),(=),(
~
−+−+∑∑ sjrisr
sr
ji
ayxLyxP
where srL , is defined by the following equations:
,
0)(1
0)(
1)(0
1)(
=),(,
1)(0
1)(
1)(0
1)(
=),( 1211
−
−
−
−
−
−
−
− yx
yxL
yx
yxL
.
0)(1
0)(
0)(1
0)(
=),(,
1)(0
1)(
0)(1
0)(
=),( 2221
−
−
−
−
−
−
−
− yx
yxL
yx
yxL
or
),(1=),(1),1)((=),( 1211 xyyxLyxyxL −−−
.=),(),(1=),( 2221 xyyxLyxyxL −
The matrix
∗
kA of order nm× which is an approximation of kA is given as follows:
,)()(= )(
nm
k
ijnmijk aaA ××
∗∗
≅
where
∗
ija can be determined by the following formulae
,~=,~= )(
1,11)4,3(1)3(
)(
,11)1,3(1)3(
k
jiji
k
jiji aaaa +
∗
+−+−
∗
+−+−
,~=,~= )(
11,41)4,3(1)3(
)(
1,41)1,3(1)3(
k
jiji
k
jiji aaaa ++
∗
+−+−+
∗
+−+−
,~
3
1~
3
2
=,0)
3
1
(
~
= )(
1,
)(),(
11)2,3(1)3(
k
ji
k
ij
ji
ji aaPa +
∗
+−+− +
,~
3
2~
3
1
=,0)
3
2
(
~
= )(
1,
)(),(
11)3,3(1)3(
k
ji
k
ij
ji
ji aaPa +
∗
+−+− +
,~
3
1~
3
2
=)
3
1
(0,
~
= )(
1,
)(),(
21)1,3(1)3(
k
ji
k
ij
ji
ji aaPa +
∗
+−+− +
,~
3
2~
3
1
=)
3
2
(0,
~
= )(
1,
)(),(
31)1,3(1)3(
k
ji
k
ij
ji
ji aaPa +
∗
+−+− +
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 278
,~
3
1~
3
2
=)
3
1
(1,
~
= )(
11,
)(
1,
),(
21)4,3(1)3(
k
ji
k
ji
ji
ji aaPa +++
∗
+−+− ++
,~
3
2~
3
1
=)
3
2
(1,
~
= )(
11,
)(
1,
),(
31)4,3(1)3(
k
ji
k
ji
ji
ji aaPa +++
∗
+−+− ++
,~
3
1~
3
2
=,1)
3
1
(
~
= )(
11,
)(
1,
),(
41)2,3(1)3(
k
ji
k
ji
ji
ji aaPa +++
∗
+−+− ++
,~
3
2~
3
1
=,1)
3
2
(
~
= )(
11,
)(
1,
),(
41)3,3(1)3(
k
ji
k
ji
ji
ji aaPa +++
∗
+−+− ++
,~
9
1~
9
2~
9
2~
9
4
=)
3
1
,
3
1
(
~
= )(
11,
)(
1,
)(
1,
)(),(
21)2,3(1)3(
k
ji
k
ji
k
ji
k
ij
ji
ji aaaaPa ++++
∗
+−+− +++
,~
9
2~
9
1~
9
4~
9
2
=)
3
2
,
3
1
(
~
= )(
11,
)(
1,
)(
1,
)(),(
31)2,3(1)3(
k
ji
k
ji
k
ji
k
ij
ji
ji aaaaPa ++++
∗
+−+− +++
,~
9
2~
9
4~
9
1~
9
2
=)
3
1
,
3
2
(
~
= )(
11,
)(
1,
)(
1,
)(),(
21)3,3(1)3(
k
ji
k
ji
k
ji
k
ij
ji
ji aaaaPa ++++
∗
+−+− +++
,~
9
4~
9
2~
9
2~
9
1
=)
3
2
,
3
2
(
~
= )(
11,
)(
1,
)(
1,
)(),(
31)3,3(1)3(
k
ji
k
ji
k
ji
k
ij
ji
ji aaaaPa ++++
∗
+−+− +++
In fact, one has
0,1,2,3=,)
3
,
3
(
~
= ),(
)(11)),3((11)3( sr
sr
Pa ji
sjri
∗
++−++−
Then,
∗
A is obtained as suitable approximation of .kA
3. NUMERICAL EXPERIMENTS
In this section some test examples are presented. In these examples we compare visually and
numerically the compressing. The reconstructed image and the compression ratio and, the mean
square error ( MSE) in each case are presented. In the tables, IL , and BI , are used for linear
and bilinear interpolation, respectively.
3.1 Example
In this example we consider a 300144× pixels black and white image from MATLAB gallery.
The initial and final images are given in Figure 1.(a), and (b), the results for simple SVD linear
interpolation and bilinear interpolation are presented in Figures 2. and 3, respectively. In Table
3.1. numerical results for three methods are given.
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 279
method k e CR MSE cpu time
SVD 9 0.4304 10.78658 1.2458×10^-4 8.7622
SVD 16 0.3384 6.0674 8.0855× 10^-5 8.6436
SVD 25 0.258 3.8831 5.5209× 10^-5 8.0967
SVD 36 0.1889 2.6966 3.9410 ×10^-5 9.4412
SVD,LI 9 0.3873 10.7143 2.5198 ×10^-4 8.1953
SVD,LI 16 0.2860 6.00268 1.6349 ×10^-4 8.4375
SVD,LI 25 0.1985 3.8571 1.1551 ×10^-4 8.7658
SVD,LI 36 0.1228 2.6786 8.1363 ×10^-5 11.4910
SVD,BI 9 0.3335 95.3642 4.001 ×10^-4 7.3140
SVD,BI 16 0.2308 53.6424 2.5966 ×10^-4 7.9940
SVD,BI 25 0.1579 34.3311 1.6774 ×10^-4 11.3992
SVD,BI 36 0.0454 23.8411 1.0217 ×10^-4 12.6754
TABLE 3.1: The results for black and white image.
[original] [final]
FIGURE 1: The results of SVD method for k=36.
[original] [final]
FIGURE 2: The results of SVD and linear interpolation method for k=36.
[original] [final]
FIGURE 3: The results of SVD and bilinear interpolation method for k=36.
3.2 Example
In this example a 5123843 ×× pixel colored image from MATLAB gallery is selected. The initial
and final images are given in Figure 4. and the results for SVD linear interpolation , and bilinear
interpolation are shown in Figure 4-5, respectively. The numerical results are presented in Table
3.2.
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 280
method M e CR MSE time
SVD 9 0.4913 24.3538 5.9974×10^-5 25.9324
SVD 16 0.3969 13.6990 3.3733 ×10^-5 26.5473
SVD 25 0.32242 8.7674 1.9898 ×10^-5 26.3382
SVD 36 0.2685 6.0884 1.3075 ×10^-5 28.4017
SVD,LI 9 0.7508 24.2726 2.2499 ×10^-4 11.3439
SVD,LI 16 0.7158 13.6533 2.0673 ×10^-4 11.8093
SVD,LI 25 0.7084 8.7381 2.0631 ×10^-4 12.4191
SVD,LI 36 0.6945 6.0681 2.0611×10^-4 14.2736
SVD,BI 9 0.7370 217.0066 7.0313 ×10^-4 8.7288
SVD,BI 16 0.7001 122.0662 6.9817 ×10^-4 10.8634
SVD,BI 25 0.6922 78.1224 6.9801 ×10^-4 12.1687
SVD,BI 36 0.6535 54.2517 6.978 ×10^-4 13.1736
TABLE 3.2: The results for black and white.
[original] [final]
FIGURE 4: The results for SVD method for k=36.
[original] [final]
FIGURE 5: The results for SVD and linear interpolation method.
[original] [final]
FIGURE 6: The results for SVD and bilinear interpolation method.
3.3 Example
In this example an other colored image is chosen from MATLAB gallery which has 3183183 ××
pixels. The initial and final images are given in Figures 7-9. and the results for simple SVD linear
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 281
interpolation, and bilinear interpolation are shown in the same Figures, respectively. The
numerical results are presented in Table 3.3.
method M im in e CR MSE time
SVD 9 3 3 0.4304 17.6389 8.8662× 10^-5 17.8987
SVD 16 4 4 0.3384 9.9212 4.4780 ×10^-5 18.5721
SVD 25 5 5 0.2580 6.3500 2.9478× 10^-5 25.36016
SVD 36 6 5 0.1889 4.4097 1.968 ×10^-5 32.3716
SVD,LI 9 3 3 0.7493 17.5562 3.5891× 10^-4 10.3739
SVD,LI 16 4 4 0.7192 9.8754 3.5681 ×10^-4 11.7131
SVD,LI 25 5 5 0.6943 6.3202 3.5630× 10^-4 13.90725
SVD,LI 36 6 5 0.6738 4.3891 3.5615 ×10^-4 17.5310
SVD,BI 9 3 3 0.7330 155.7969 0.0016 9.1011
SVD,BI 16 4 4 0.6995 87.6358 0.0016 10.1484
SVD,BI 25 5 5 0.6720 56.0869 0.0016 12.1943
SVD,BI 36 6 5 0.6535 38.9491 0.0016 13.17828
Table 3.3: The results for colour image .
[original] [final]
FIGURE 7: The results for SVD method for k=36.
[original] [final]
FIGURE 8: The results for SVD and linear interpolation method for k=36
J. Izadian , A. Hosaini & M. Jalili
International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 282
[original] [final]
FIGURE 9: The results of SVD and bilinear interpolation method for k=36
4. CONCLUSION
In this paper two hybrid SVD methods, using linear interpolation, and bilinear interpolation are
presented. The results show the preference of bilinear interpolation combined with SVD. The
linear method is also economically acceptable.
5. REFERENCES
[1] J. Stoer, R. Bulirsch. Intoduction to Numerical Analysis. translated by R.Bartels ,
W.Gautschi. C.Witzgall. Springer-Verlag, second edition, (1991).
[2] H. Abdi. “Singular Value Decomposition (SVD) and Generalized Singular Value
Decomposition (GSVD).” In N.J. Salkind (Ed.): Encyclopedia of Measurement and
Statistics. Thousand Oaks (CA): Sage. pp. 907-912, 2007.
[3] R. Ashino, A. Morimoto, M. Nagase, R.Vaillancourt. “Image Compression with
Multiresulotion Singular Value Decomposition and Other Methods.” Mathematical and
Computer Modelling. Vol. 41(6-7). pp. 773-790, 2005.
[4] P.Dostert. “An application of linear algebra to image compression.” The university of
arizona. July 2009.
[5] M.Sing Song. “Wavelet Image Compression.” Mathematical subject classification, 1991.
[6] M.Sing Song. “Wavelet Image Compression”, Southern Illinois University Edwardsville,
Edwardsville, IL62026, USA, 2004.

More Related Content

PDF
Blind Source Separation Using Hessian Evaluation
PPTX
PosterPresentations.com-3 6x48-Template-V5 - 副本
PDF
International Journal of Computational Engineering Research(IJCER)
PDF
GREY LEVEL CO-OCCURRENCE MATRICES: GENERALISATION AND SOME NEW FEATURES
PDF
Auto Regressive Process (1) with Change Point: Bayesian Approch
PDF
Estimation of mean and its function using asymmetric loss function
PDF
Estimation of mean and its function using asymmetric loss function
PDF
2012 mdsp pr11 ica part 2 face recognition
Blind Source Separation Using Hessian Evaluation
PosterPresentations.com-3 6x48-Template-V5 - 副本
International Journal of Computational Engineering Research(IJCER)
GREY LEVEL CO-OCCURRENCE MATRICES: GENERALISATION AND SOME NEW FEATURES
Auto Regressive Process (1) with Change Point: Bayesian Approch
Estimation of mean and its function using asymmetric loss function
Estimation of mean and its function using asymmetric loss function
2012 mdsp pr11 ica part 2 face recognition

What's hot (20)

PDF
Cy33602608
PDF
International journal of engineering and mathematical modelling vol2 no1_2015_1
PDF
2012 mdsp pr13 support vector machine
PDF
Human Face Detection Based on Combination of Logistic Regression, Distance of...
PPTX
Computer Graphics Unit 1
PPTX
Relaxation method
PDF
2012 mdsp pr10 ica
PPTX
Output primitives computer graphics c version
PDF
2012 mdsp pr09 pca lda
PDF
Linear Regression Ordinary Least Squares Distributed Calculation Example
PDF
Applied mathematics for complex engineering
PPTX
A mid point ellipse drawing algorithm on a hexagonal grid
PPT
Line drawing algo.
PDF
Image Restitution Using Non-Locally Centralized Sparse Representation Model
PDF
Conjugate Gradient method for Brain Magnetic Resonance Images Segmentation
PDF
Applied mathematics for CBE assignment
PDF
Hierarchical Vertebral Body Segmentation Using Graph Cuts and Statistical Sha...
PDF
Sixth order hybrid block method for the numerical solution of first order ini...
Cy33602608
International journal of engineering and mathematical modelling vol2 no1_2015_1
2012 mdsp pr13 support vector machine
Human Face Detection Based on Combination of Logistic Regression, Distance of...
Computer Graphics Unit 1
Relaxation method
2012 mdsp pr10 ica
Output primitives computer graphics c version
2012 mdsp pr09 pca lda
Linear Regression Ordinary Least Squares Distributed Calculation Example
Applied mathematics for complex engineering
A mid point ellipse drawing algorithm on a hexagonal grid
Line drawing algo.
Image Restitution Using Non-Locally Centralized Sparse Representation Model
Conjugate Gradient method for Brain Magnetic Resonance Images Segmentation
Applied mathematics for CBE assignment
Hierarchical Vertebral Body Segmentation Using Graph Cuts and Statistical Sha...
Sixth order hybrid block method for the numerical solution of first order ini...
Ad

Similar to A Hybrid SVD Method Using Interpolation Algorithms for Image Compression (20)

PPTX
Face Recognition using PCA-Principal Component Analysis using MATLAB
PDF
Soạn thảo văn bản bằng LATEX
PDF
Neighbour Local Variability for Multi-Focus Images Fusion
PDF
Neighbour Local Variability for Multi-Focus Images Fusion
PDF
Neighbour Local Variability for Multi-Focus Images Fusion
PDF
MULTIFOCUS IMAGE FUSION USING MULTIRESOLUTION APPROACH WITH BILATERAL GRADIEN...
PDF
Investigation on the Pattern Synthesis of Subarray Weights for Low EMI Applic...
PDF
On image intensities, eigenfaces and LDA
PDF
Neuro -fuzzy-networks-for-identification-of-mathematical-model-parameters-of-...
PDF
Content Based Image Retrieval Using Gray Level Co-Occurance Matrix with SVD a...
PDF
ALEXANDER FRACTIONAL INTEGRAL FILTERING OF WAVELET COEFFICIENTS FOR IMAGE DEN...
PDF
Scaling Transform Methods For Compressing a 2D Graphical image
PDF
A New Method Based on MDA to Enhance the Face Recognition Performance
PDF
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
PDF
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
PDF
Detection of Seam Carving in Uncompressed Images using eXtreme Gradient Boosting
PDF
svd15_86.pdf for SVD study and revosited
PDF
A STUDY AND ANALYSIS OF DIFFERENT EDGE DETECTION TECHNIQUES
PDF
Blind Image Seperation Using Forward Difference Method (FDM)
PDF
A Quantitative Comparative Study of Analytical and Iterative Reconstruction T...
Face Recognition using PCA-Principal Component Analysis using MATLAB
Soạn thảo văn bản bằng LATEX
Neighbour Local Variability for Multi-Focus Images Fusion
Neighbour Local Variability for Multi-Focus Images Fusion
Neighbour Local Variability for Multi-Focus Images Fusion
MULTIFOCUS IMAGE FUSION USING MULTIRESOLUTION APPROACH WITH BILATERAL GRADIEN...
Investigation on the Pattern Synthesis of Subarray Weights for Low EMI Applic...
On image intensities, eigenfaces and LDA
Neuro -fuzzy-networks-for-identification-of-mathematical-model-parameters-of-...
Content Based Image Retrieval Using Gray Level Co-Occurance Matrix with SVD a...
ALEXANDER FRACTIONAL INTEGRAL FILTERING OF WAVELET COEFFICIENTS FOR IMAGE DEN...
Scaling Transform Methods For Compressing a 2D Graphical image
A New Method Based on MDA to Enhance the Face Recognition Performance
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
Detection of Seam Carving in Uncompressed Images using eXtreme Gradient Boosting
svd15_86.pdf for SVD study and revosited
A STUDY AND ANALYSIS OF DIFFERENT EDGE DETECTION TECHNIQUES
Blind Image Seperation Using Forward Difference Method (FDM)
A Quantitative Comparative Study of Analytical and Iterative Reconstruction T...
Ad

Recently uploaded (20)

PPTX
human mycosis Human fungal infections are called human mycosis..pptx
PDF
Anesthesia in Laparoscopic Surgery in India
PDF
ANTIBIOTICS.pptx.pdf………………… xxxxxxxxxxxxx
PPTX
Tissue processing ( HISTOPATHOLOGICAL TECHNIQUE
PDF
The Lost Whites of Pakistan by Jahanzaib Mughal.pdf
PPTX
school management -TNTEU- B.Ed., Semester II Unit 1.pptx
PDF
OBE - B.A.(HON'S) IN INTERIOR ARCHITECTURE -Ar.MOHIUDDIN.pdf
PDF
Chinmaya Tiranga quiz Grand Finale.pdf
PDF
Complications of Minimal Access Surgery at WLH
PPTX
Cell Types and Its function , kingdom of life
PDF
FourierSeries-QuestionsWithAnswers(Part-A).pdf
PPTX
Lesson notes of climatology university.
PDF
grade 11-chemistry_fetena_net_5883.pdf teacher guide for all student
PPTX
Pharmacology of Heart Failure /Pharmacotherapy of CHF
PDF
Abdominal Access Techniques with Prof. Dr. R K Mishra
PDF
O5-L3 Freight Transport Ops (International) V1.pdf
PPTX
Microbial diseases, their pathogenesis and prophylaxis
PDF
Supply Chain Operations Speaking Notes -ICLT Program
PPTX
Institutional Correction lecture only . . .
PPTX
Cell Structure & Organelles in detailed.
human mycosis Human fungal infections are called human mycosis..pptx
Anesthesia in Laparoscopic Surgery in India
ANTIBIOTICS.pptx.pdf………………… xxxxxxxxxxxxx
Tissue processing ( HISTOPATHOLOGICAL TECHNIQUE
The Lost Whites of Pakistan by Jahanzaib Mughal.pdf
school management -TNTEU- B.Ed., Semester II Unit 1.pptx
OBE - B.A.(HON'S) IN INTERIOR ARCHITECTURE -Ar.MOHIUDDIN.pdf
Chinmaya Tiranga quiz Grand Finale.pdf
Complications of Minimal Access Surgery at WLH
Cell Types and Its function , kingdom of life
FourierSeries-QuestionsWithAnswers(Part-A).pdf
Lesson notes of climatology university.
grade 11-chemistry_fetena_net_5883.pdf teacher guide for all student
Pharmacology of Heart Failure /Pharmacotherapy of CHF
Abdominal Access Techniques with Prof. Dr. R K Mishra
O5-L3 Freight Transport Ops (International) V1.pdf
Microbial diseases, their pathogenesis and prophylaxis
Supply Chain Operations Speaking Notes -ICLT Program
Institutional Correction lecture only . . .
Cell Structure & Organelles in detailed.

A Hybrid SVD Method Using Interpolation Algorithms for Image Compression

  • 1. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 273 A Hybrid SVD Method Using Interpolation Algorithms for Image Compression J. Izadian , A. Hosaini Jalal_Izadian@yahoo.com Department of Mathematics, Faculty of Sciences, Mashhad Branch, Islamic Azad University, Mashhad, Iran M. Jalili Jalili.maryam@yahoo.com Department of Mathematics, Neyshabur Branch, Islamic Azad University, Neyshabur, Iran Abstract In this paper the standard SVD method is used for image processing and is combined with some interpolation methods as linear and quadratic interpolation for reconstruction of compressed image.The main idea of the proposed method is to select a particular submatrix of main image matrix and compress it with SVD method, then reconstruct an approximation of original image by interpolation method. The numerical experiments illustrate the performance and efficiency of proposed methods. Keywords: SVD Method, Interpolation Method, Image Reconstruction, Compressing Image, Lossless Compression, Lossy Compression. 1. INTRODUCTION Image compression techniques plays an important role in transmision and storage of image information in computer science and related domains.The goal of image compression is to obtain a representation that minimizes the bits volume while still maintaining the important meaning and the intrinsic structure of the original image. Image compression techniques can be classified into two groups. Lossless compression and lossy compression. In lossless compression the reconstructed image is identical to the original one and deduces a low compression ratio while lossy compression methods allow a loss in the actual image data. So the original image cannot be created excactly from the compressed image. But these methods deduce high compression ratio. There is many ways to compress depending on the application field. One popular method for compressing an image is the wavelet teqnique [1]. Wavelet functions form an orthonormal basis on which it is possible to project each data set. Another important method is the Singular Value Decomposition (SVD). This technique is based on the factoriziation of the real matrix of image in three matrices that can be used to reconstruct the main image or an approximation of it, for more information see [2,3,4]. In this paper the SVD method is combined with two interpolation processes, that are linear 2D interpolation (triangular interpolation), and bilinear interpolation, to decrease the volume of transfered image . This paper is orgonized as follows. Section 2 is specified to describe the basic concepts of proposed method. Section 3 presents the numerical experiments . Finally section 4 concludes the paper with discussion. 2. THE BASIC CONCEPTS OF METHOD Suppose that a nm× pixel gray-scale image is given. Each pixel having some level of black and white given by some integers that can be selected between 0 and 255 or a real number
  • 2. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 274 between zero and one. In the case of integer, each integer requires approximately one byte to store. Then the resulting image has approximately nm× bytes volume. If the image is coloured in (RGB) system the image contains three nm× matrices for red, green, and blue colours, then nm××3 pixels image. The Singular Value Decomposition (SVD) for a gray-scale image for compressing nm× pixels image matrix is based on the following decomposition of A , .= T USVA Actually, recalling from advanced linear algebra one has the following theorem, that is demonstrated in advanced linear algebra (for example see [5,6]). Theorem 2.1 Any nonzero real nm× matrix A with rank 0>r , there are an orthogonal mm× matrix U , and an orthogonal nn× matrix V such that SAVUT = is an nm× "diagonal matrix" of the following form , 00 0 =      D S 0,>...),,...,(= 211 rrdiagD σσσσσ ≥≥ where rσσσ ,...,, 21 are non-zero sigular values of A . This factorization is called the Singular Value Decomposition (SVD) of A . By using above theorem and assuming that ,1,2,...,=],......[= 121 miuuuuuU mrr + where miui 1,...,=, are the column of U , by considering the orthogonality of U , one has    ≠ .0 =1 == ji ji uu ijj T i δ Similarly for orthogonal nn× matrix ,V by considering. ,= 2 1               T n T T T v v v V K where iv are for ni 1,...,= are columns of ,V it yields    ≠ .0 =1 == ji ji vv ijj T i δ Here, S is nm× diagonal matrix with singular values of A on the diagonal of D in rectangular matrix S . The matrix S can be presented by the following matrix: 1 2 0 0 0 0 0 0 0 0 0 0 0 r S σ σ σ        =        K M M O M K K K For =1,2,...,i n , iσ are singular values of matrix A. It is proved in Theorem(2.1) that 1 2 ... 0rσ σ σ≥ ≥ ≥ > and 1 2 ... 0r r nσ σ σ+ += = = = . For =1,2,...,i m , iσ are called singular values of
  • 3. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 275 matrix T A .The vector iv for =1,2,...,i n are called the right singular vector of A , and iu for =1,2,...,i m . are the left singular vectors of A (see[4,5]). then 1 2 1 2 1 1 1 1 2 2 2 . [ ... ] . . . T T T T m m r r r n v v A USV u u u u S u v u v u v v σ σ σ−         = = = + + +          K When compressing the image, the sum is not performed to the very last . The singular values with small enough values are dropped. A matrix of rank k is obtained by truncating these sums after the first k terms, denoted by ,= 221111 T kkk TT k vuvuvuA σσσ +++ K (1) is chosen as approximation of A . The total storage volume for kA will be 1)( ++ nmk . This matrix is an approximation of A , that can be used as an approximation of compressed image A . For having a criterion for storage volume decrease, the compression ratio is defined as follows: . 1)( = ++ nmk mn CR We also use the Schur norm to measure the quality of obtained image kA , called Mean Square Error (MSE), which is introduced by following expression: ( ) 2 2 =1 =1 1 1 = = ( ) . m n k K ij ij i j MSE A A a a mn mn − −∑∑ where nmijaA ×)(= and .)(= )( nm k ijk aA × An error value can be defined related to sum of iσ by following relation: ,1= 1= 1= i r i i k i ke σ σ ∑ ∑ − A typical choice of k is so that the storage space required for kA will be less than . 5 1 For having an superior compression ratio, one can utilize a submatrix of A for small k and use an interpolation process to reconstruct kA . Consider m and n are even, and A ~ is a submatrix of order 22 nm × of A , by omitting the element of A which are situated on rows or column of odd numbers. This matrix is approximated by SVD method with a suitable value of k . Then an interpolation method is used to reconstruct an approximation of kA , that can be chosen as approximation of kA as obtained image. In linear case, consider the chosen submatrix of A as: ,)~(= ~ 22 )( nm k ijk aA ×
  • 4. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 276 where m and n are even numbers and the elements of k can be determined from nm k ijk aA ×)(= )( by the following relations: ,=)~(,=)~( )( 31)1,2(1)2( )( 1, )( 11)1,2(1)2( )( , k ji k ji k ji k ji aaaa +−+−++−+− ,=)~(,=)~( )( 31)3,2(1)2( )( 11, )( 11)3,2(1)2( )( 1, k ji k ji k ji k ji aaaa +−+−+++−+−+ . 2 1,2,...,=, 2 1,2,...,= n j m i The matrix kA ~ is of order , 22 nm × therefore it has 4 3 less element than kA . This matrix is chosen as the compression of A that results a compression ratio given by , 2)( 2 = 1) 22 ( = ++++ nmk mn nm k mn CR Therefore, in this method the compression ratio is doubled in comparison with kA . The matrix kA ~ is supposed as a base for interpolating method to reconstruct an approximation of kA . In the linear case, this matrix is calculated with the following equations: ,)(=, nmijkkk aAAA × ∗∗∗ ; ,~=,~= )( 1,11)3,2(1)2( )( 11)1,2(1)2( k jiji k ijji aaaa + ∗ +−+− ∗ +−+− ,~=,~= )( 11,31)3,2(1)2( )( 1,31)1,2(1)2( k jiji k jiji aaaa ++ ∗ +−+−+ ∗ +−+− )/2,~~(=)/2,~~(= )( 1, )( ,21)1,2(1)2( )( , )( 1,11)2,2(1)2( k ji k jiji k ji k jiji aaaaaa + ∗ +−+−+ ∗ +−+− ++ )/2,~~(=)/2,~~(= )( 11, )( 1,31)2,2(1)2( )( 11, )( 1,21)3,2(1)2( k ji k jiji k ji k jiji aaaaaa +++ ∗ +−+−+++ ∗ +−+− ++ ]/4.~~~~[= )( 11, )( , )( 1, )( 1,21)2,2(1)2( k ji k ji k ji k jiji aaaaa ++++ ∗ +−+− +++ In order to describe the hybrid SVD using bilinear interpolation, the following submatrix of kA is considerd ,)~(= ~ 33 )( nm k ijk aA × where kA is considered a nm× matrix which m and n are the integers that are divisible by 3. Then kA ~ can be defined as follows ,)(= )( nm k ijk aA × ,=)~(,=)~( )( 41)1,3(1)3( )( 1, )( 11)1,3(1)3( )( , k ji k ji k ji k ji aaaa +−+−++−+− ,=)~(,=)~( )( 41)4,3(1)3( )( 11, )( 11)4,3(1)3( )( 1, k ji k ji k ji k ji aaaa +−+−+++−+−+
  • 5. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 277 . 3 1,2,...,=, 3 1,2,...,= n j m i Since the matrix kA ~ is a matrix of order , 33 nm × the compression ratio is obtained by , 3)( 3 = 1) 33 ( = ++++ nmk mn nm n mn CR Now for each 44× block of matrix kA which is reduced to 22× block matrix kA ~ , the following bilinear interpolation is applied on ,[0,1]2 11),(, 2 1= 2 1= ),( ~),(=),( ~ −+−+∑∑ sjrisr sr ji ayxLyxP where srL , is defined by the following equations: , 0)(1 0)( 1)(0 1)( =),(, 1)(0 1)( 1)(0 1)( =),( 1211 − − − − − − − − yx yxL yx yxL . 0)(1 0)( 0)(1 0)( =),(, 1)(0 1)( 0)(1 0)( =),( 2221 − − − − − − − − yx yxL yx yxL or ),(1=),(1),1)((=),( 1211 xyyxLyxyxL −−− .=),(),(1=),( 2221 xyyxLyxyxL − The matrix ∗ kA of order nm× which is an approximation of kA is given as follows: ,)()(= )( nm k ijnmijk aaA ×× ∗∗ ≅ where ∗ ija can be determined by the following formulae ,~=,~= )( 1,11)4,3(1)3( )( ,11)1,3(1)3( k jiji k jiji aaaa + ∗ +−+− ∗ +−+− ,~=,~= )( 11,41)4,3(1)3( )( 1,41)1,3(1)3( k jiji k jiji aaaa ++ ∗ +−+−+ ∗ +−+− ,~ 3 1~ 3 2 =,0) 3 1 ( ~ = )( 1, )(),( 11)2,3(1)3( k ji k ij ji ji aaPa + ∗ +−+− + ,~ 3 2~ 3 1 =,0) 3 2 ( ~ = )( 1, )(),( 11)3,3(1)3( k ji k ij ji ji aaPa + ∗ +−+− + ,~ 3 1~ 3 2 =) 3 1 (0, ~ = )( 1, )(),( 21)1,3(1)3( k ji k ij ji ji aaPa + ∗ +−+− + ,~ 3 2~ 3 1 =) 3 2 (0, ~ = )( 1, )(),( 31)1,3(1)3( k ji k ij ji ji aaPa + ∗ +−+− +
  • 6. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 278 ,~ 3 1~ 3 2 =) 3 1 (1, ~ = )( 11, )( 1, ),( 21)4,3(1)3( k ji k ji ji ji aaPa +++ ∗ +−+− ++ ,~ 3 2~ 3 1 =) 3 2 (1, ~ = )( 11, )( 1, ),( 31)4,3(1)3( k ji k ji ji ji aaPa +++ ∗ +−+− ++ ,~ 3 1~ 3 2 =,1) 3 1 ( ~ = )( 11, )( 1, ),( 41)2,3(1)3( k ji k ji ji ji aaPa +++ ∗ +−+− ++ ,~ 3 2~ 3 1 =,1) 3 2 ( ~ = )( 11, )( 1, ),( 41)3,3(1)3( k ji k ji ji ji aaPa +++ ∗ +−+− ++ ,~ 9 1~ 9 2~ 9 2~ 9 4 =) 3 1 , 3 1 ( ~ = )( 11, )( 1, )( 1, )(),( 21)2,3(1)3( k ji k ji k ji k ij ji ji aaaaPa ++++ ∗ +−+− +++ ,~ 9 2~ 9 1~ 9 4~ 9 2 =) 3 2 , 3 1 ( ~ = )( 11, )( 1, )( 1, )(),( 31)2,3(1)3( k ji k ji k ji k ij ji ji aaaaPa ++++ ∗ +−+− +++ ,~ 9 2~ 9 4~ 9 1~ 9 2 =) 3 1 , 3 2 ( ~ = )( 11, )( 1, )( 1, )(),( 21)3,3(1)3( k ji k ji k ji k ij ji ji aaaaPa ++++ ∗ +−+− +++ ,~ 9 4~ 9 2~ 9 2~ 9 1 =) 3 2 , 3 2 ( ~ = )( 11, )( 1, )( 1, )(),( 31)3,3(1)3( k ji k ji k ji k ij ji ji aaaaPa ++++ ∗ +−+− +++ In fact, one has 0,1,2,3=,) 3 , 3 ( ~ = ),( )(11)),3((11)3( sr sr Pa ji sjri ∗ ++−++− Then, ∗ A is obtained as suitable approximation of .kA 3. NUMERICAL EXPERIMENTS In this section some test examples are presented. In these examples we compare visually and numerically the compressing. The reconstructed image and the compression ratio and, the mean square error ( MSE) in each case are presented. In the tables, IL , and BI , are used for linear and bilinear interpolation, respectively. 3.1 Example In this example we consider a 300144× pixels black and white image from MATLAB gallery. The initial and final images are given in Figure 1.(a), and (b), the results for simple SVD linear interpolation and bilinear interpolation are presented in Figures 2. and 3, respectively. In Table 3.1. numerical results for three methods are given.
  • 7. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 279 method k e CR MSE cpu time SVD 9 0.4304 10.78658 1.2458×10^-4 8.7622 SVD 16 0.3384 6.0674 8.0855× 10^-5 8.6436 SVD 25 0.258 3.8831 5.5209× 10^-5 8.0967 SVD 36 0.1889 2.6966 3.9410 ×10^-5 9.4412 SVD,LI 9 0.3873 10.7143 2.5198 ×10^-4 8.1953 SVD,LI 16 0.2860 6.00268 1.6349 ×10^-4 8.4375 SVD,LI 25 0.1985 3.8571 1.1551 ×10^-4 8.7658 SVD,LI 36 0.1228 2.6786 8.1363 ×10^-5 11.4910 SVD,BI 9 0.3335 95.3642 4.001 ×10^-4 7.3140 SVD,BI 16 0.2308 53.6424 2.5966 ×10^-4 7.9940 SVD,BI 25 0.1579 34.3311 1.6774 ×10^-4 11.3992 SVD,BI 36 0.0454 23.8411 1.0217 ×10^-4 12.6754 TABLE 3.1: The results for black and white image. [original] [final] FIGURE 1: The results of SVD method for k=36. [original] [final] FIGURE 2: The results of SVD and linear interpolation method for k=36. [original] [final] FIGURE 3: The results of SVD and bilinear interpolation method for k=36. 3.2 Example In this example a 5123843 ×× pixel colored image from MATLAB gallery is selected. The initial and final images are given in Figure 4. and the results for SVD linear interpolation , and bilinear interpolation are shown in Figure 4-5, respectively. The numerical results are presented in Table 3.2.
  • 8. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 280 method M e CR MSE time SVD 9 0.4913 24.3538 5.9974×10^-5 25.9324 SVD 16 0.3969 13.6990 3.3733 ×10^-5 26.5473 SVD 25 0.32242 8.7674 1.9898 ×10^-5 26.3382 SVD 36 0.2685 6.0884 1.3075 ×10^-5 28.4017 SVD,LI 9 0.7508 24.2726 2.2499 ×10^-4 11.3439 SVD,LI 16 0.7158 13.6533 2.0673 ×10^-4 11.8093 SVD,LI 25 0.7084 8.7381 2.0631 ×10^-4 12.4191 SVD,LI 36 0.6945 6.0681 2.0611×10^-4 14.2736 SVD,BI 9 0.7370 217.0066 7.0313 ×10^-4 8.7288 SVD,BI 16 0.7001 122.0662 6.9817 ×10^-4 10.8634 SVD,BI 25 0.6922 78.1224 6.9801 ×10^-4 12.1687 SVD,BI 36 0.6535 54.2517 6.978 ×10^-4 13.1736 TABLE 3.2: The results for black and white. [original] [final] FIGURE 4: The results for SVD method for k=36. [original] [final] FIGURE 5: The results for SVD and linear interpolation method. [original] [final] FIGURE 6: The results for SVD and bilinear interpolation method. 3.3 Example In this example an other colored image is chosen from MATLAB gallery which has 3183183 ×× pixels. The initial and final images are given in Figures 7-9. and the results for simple SVD linear
  • 9. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 281 interpolation, and bilinear interpolation are shown in the same Figures, respectively. The numerical results are presented in Table 3.3. method M im in e CR MSE time SVD 9 3 3 0.4304 17.6389 8.8662× 10^-5 17.8987 SVD 16 4 4 0.3384 9.9212 4.4780 ×10^-5 18.5721 SVD 25 5 5 0.2580 6.3500 2.9478× 10^-5 25.36016 SVD 36 6 5 0.1889 4.4097 1.968 ×10^-5 32.3716 SVD,LI 9 3 3 0.7493 17.5562 3.5891× 10^-4 10.3739 SVD,LI 16 4 4 0.7192 9.8754 3.5681 ×10^-4 11.7131 SVD,LI 25 5 5 0.6943 6.3202 3.5630× 10^-4 13.90725 SVD,LI 36 6 5 0.6738 4.3891 3.5615 ×10^-4 17.5310 SVD,BI 9 3 3 0.7330 155.7969 0.0016 9.1011 SVD,BI 16 4 4 0.6995 87.6358 0.0016 10.1484 SVD,BI 25 5 5 0.6720 56.0869 0.0016 12.1943 SVD,BI 36 6 5 0.6535 38.9491 0.0016 13.17828 Table 3.3: The results for colour image . [original] [final] FIGURE 7: The results for SVD method for k=36. [original] [final] FIGURE 8: The results for SVD and linear interpolation method for k=36
  • 10. J. Izadian , A. Hosaini & M. Jalili International Journal of Image Processing (IJIP), Volume (6) : Issue (5) : 2012 282 [original] [final] FIGURE 9: The results of SVD and bilinear interpolation method for k=36 4. CONCLUSION In this paper two hybrid SVD methods, using linear interpolation, and bilinear interpolation are presented. The results show the preference of bilinear interpolation combined with SVD. The linear method is also economically acceptable. 5. REFERENCES [1] J. Stoer, R. Bulirsch. Intoduction to Numerical Analysis. translated by R.Bartels , W.Gautschi. C.Witzgall. Springer-Verlag, second edition, (1991). [2] H. Abdi. “Singular Value Decomposition (SVD) and Generalized Singular Value Decomposition (GSVD).” In N.J. Salkind (Ed.): Encyclopedia of Measurement and Statistics. Thousand Oaks (CA): Sage. pp. 907-912, 2007. [3] R. Ashino, A. Morimoto, M. Nagase, R.Vaillancourt. “Image Compression with Multiresulotion Singular Value Decomposition and Other Methods.” Mathematical and Computer Modelling. Vol. 41(6-7). pp. 773-790, 2005. [4] P.Dostert. “An application of linear algebra to image compression.” The university of arizona. July 2009. [5] M.Sing Song. “Wavelet Image Compression.” Mathematical subject classification, 1991. [6] M.Sing Song. “Wavelet Image Compression”, Southern Illinois University Edwardsville, Edwardsville, IL62026, USA, 2004.