Академический Документы
Профессиональный Документы
Культура Документы
, (5)
in which c
is a predetermined threshold.
In the implementation of the image or image se-
quence coding, we usually use the mean square error
(MSE) as a distortion measure. For our application, the
MSE between two image sequence blocks U and V is
dened as
d
mse
(U, V) = E[(U(x, y, t ) V(x, y, t ))
2
], (6)
where E denotes the mean operator.
2.2. The Computerized Tomographic
Image Reconstruction
Figure 1 illustrates the basic projection concept, in
which a linear transformation is dened as
P
(z) P(z, ) =
_
ray AB
f (x, y) ds,
< z < , 0 < , (7)
where (z, ) represents the coordinates of the AB ray
[17] relative to the object, line AB is the path of the
ray, s is the distance along line AB, and f (x, y) is
a 2D spatial function (image). The image reconstruc-
tion problem is to determine f (x, y) from P
(z). For a
digital image, the integral value can be rewritten as the
Figure 1. The projection of a 2D spatial function (image).
summation of projection
p
(z) =
N
z
i =1
f
i,z
, z = 1, 2, . . . , Z, (8)
where f
i,z
and N
z
denote the gray-level from the i th
pixel and the total number of pixels of the zth ray,
respectively. Z is the total number of rays. The theory
of reconstructing an image from its projection values
can be found in [18].
3. The Adaptive Fractal Image Sequence
Coding Scheme
The major elements of our proposed adaptive frac-
tal image sequence coding technique are discussed in
this section. It includes solid image block classica-
tion scheme based on the algorithm of 3D projection
classication and adaptive fractal image sequence cod-
ing technique by the 3D variable shape decomposition
algorithm. The classication algorithm is derived and
extended from the computerized tomographic recon-
struction skill.
3.1. The 3D Projection Classication Algorithm
For a n n n (n >1) solid image block Q, we could
dene the projection directions Ray(, ), , =0
,
45
, 90
, 135
), = 0
, 45
, 90
, 135
, is of
equivalence. Therefore, the total number of the pro-
jection directions is 13.
A set of 2D projection array data can be obtained
by averaging gray-levels along the thirteen directions.
The 2D projection array data P
l
and the unidirectional
Figure 2. A n n n solid image block Q, where a, b, . . . , h are
corners of Q.
272 Chao and Chieu
reconstructed solid image block
Q
l
, which is made by
back projecting P
l
, along the l direction are dened as
P
l
= project
l
(Q) and
Q
l
= backproject
l
(P
l
), respec-
tively, where l belongs to one of the 13 projection di-
rections. The concept can be implemented by using the
following formulae. For the reason of space saving, we
only list three projection formulae. The formulae not
listed are similar to those formulae listed below.
P
Ray(90
,90
)
(i, j ) =
1
n
n1
k=0
Q(i, j, k),
for i, j = 0, . . . , n 1. (9.a)
Q
Ray(90
,90
)
(i, j, k) = P
Ray(90
,90
)
(i, j ),
for i, j, k =0, . . . , n 1. (9.b)
P
Ray(45
,45
)
(i, j )
=
_
_
1
j +1
j
k=0
Q(k, j k, i k),
for j = 0, . . . , n 1 and i = j, . . . , n 1
1
i +1
i
k=0
Q(k, j k, i k),
for i = 0, . . . , n 2
and j = i +1, . . . , n 1
1
i j +n
n1+i j
k=0
Q(i k, j i +k, k),
for i = 1, . . . , n 1
and j = n, . . . , n 1 +i
1
j +1
j
k=0
Q(i +k n +1, j k, n 1 k),
for i = n, . . . , 2n 2
and j = 0, . . . , 2n 2 i
1
2n 1 i
2n2i
k=0
Q(n 1 k, i + j +k 2n
+2, i +k n +1),
for n > 2, i = n +1, . . . , 2n 2,
j = 2n 1 i, . . . , n 2
1
3n 2 i j
3n3i j
k=0
Q(i + j +k 2n
+2, n 1 k, 2n 2 j k),
for i = n, . . . , 2n 2
and j = n 1, . . . , 3n 3 i
0, otherwise
(10.a)
Q
Ray(45
,45
)
(i, j, k)
=
_
_
P
Ray(45
,45
)
(i +k, i + j ),
for j, k = 0, . . . , n 1
and i = 0, . . . , n 1 k
P
Ray(45
,45
)
(i +k, j k +n 1),
for k = 1, . . . , n 1, i = n k, . . . , n 1
and j = 0, . . . , n 1
(10.b)
P
Ray(135
,90
)
(i, j )
=
_
_
1
i +1
i
k=0
Q(i k, j, n 1 k),
for i, j = 0, . . . , n 1
1
2n 1 i
2n2i
k=0
Q(n 1 k, j,
2n 2 i k),
for i = n, . . . , 2n 2, j = 0, . . . , n 1
(11.a)
Q
Ray(135
,90
)
(i, j, k) = P
Ray(135
,90
)
(n 1 +i k, j ),
for i, j, k = 0, . . . , n 1 (11.b)
As a remark, Q() and
Q
l
() denote the gray-levels of
Q and
Q
l
, respectively. P
l
() are obtained integers by
rounding. Both the total equations for projection and
backprojection are 13. Atypical example of projection
and backprojection operation is shown in Fig. 3 by
using (9.a) and (9.b), respectively.
With the above knowledge, we are able to dene a
3D projection classication algorithm as follows:
(1) A solid image block has the smooth class if all
the MSEs between Q and
Q
l
are smaller than a
threshold t
s
.
(2) A simple texture class solid image block has the
smallest MSE between Q and
Q
l
. Moreover, the
Figure 3. A typical example of projection and backprojection op-
eration.
Adaptive Fractal Image Sequence Coding 273
smallest MSE is smaller than a threshold t
e
as
well.
(3) Otherwise, the solid image block is classied to
the mixed texture class solid image block.
In practical applications, t
e
is larger than t
s
. By using
the above algorithm, we may have 15 classes (13 sim-
ple texture classes, one mixed texture class, and one
smooth class) to be classied. All these classied tex-
tures can be viewed as the attributions for the solid
image blocks.
3.2. The 3D Variable Shape
Decomposition Algorithm
In this subsection, we will describe how to incorpo-
rate the projection classication algorithminto variable
shape solid image block transformation.
For the reason of computation complexity, the test-
ing image sequences are arranged as several non-
overlappings regular 3D testing image sequence block
sequentially. For a testing image sequence block U,
we can divide U into non-overlapping solid range
block R
i
of size N N (N/2), i = 1, . . . , N
R
. N
R
denotes the total number of the solid range blocks in
U and N is even number. The solid range block is
rst split into four non-overlapping subblocks having
size (N/2) (N/2) (N/2), as shown in Fig. 4. We
employ the 3D projection classication algorithm to
estimate the attribution of the four subblocks. By com-
paring the attribution of the four subblocks, a conven-
tional merging rule is proposed to merge the subblocks
which hold the same attributions. As shown in Fig. 4,
we can get one of the 15 possible merging patterns. A
close observation on Fig. 4 concludes that the 10 shape
Figure 4. Four subblocks can be merged to 15 possible merging
patterns.
Figure 5. Ten shape components.
components in Fig. 5 can morphologically describe
all of the 15 forms of the merging patterns. In other
words, any solid range block can be dened by these
10 shape components as well as the 15 merging pat-
terns. The corresponding merging pattern, shape com-
ponents, and attribution of shape component are the
necessary information for searching algorithm design
when the fractal encoding is proceeded. The searching
algorithm will be presented in next subsection.
For any i th solid range block R
i
, we may have at
most four associated individual shape components. By
indexing the associated individual shape components
clockwise with an integer j , we are able to dene the
shape components of solid range block as R
m
i, j
which
will be used in the searching algorithm design. Here
m denotes the mth shape component, m = 0, 1, . . . , 9,
and j 0, 1, 2, 3. The above procedure of producing
R
m
i, j
from R
i
is named as the 3Dvariable shape decom-
position algorithm.
For the searching space, the testing image sequence
block U is also divided into N
D
m
overlapable solid
domain blocks that can be indexed as D
m
k
m
, k
m
=1,
. . . , N
D
m
, according to which mth shape component of
the shape component of solid range block is referring.
Here referring means that the solid domain blocks are
selected to have the same shape as the referring shape
component but the size of selected blocks is larger.
The projection classication algorithm is followed to
classify each solid domain block into its correspond-
ing attribution. Using this procedure, we could create
a class pool C
m
A
in accordance with the referred mth
shape component and the attribution A of the solid do-
main block, where A belongs to one of the 15 classes.
Apparently, the total number of class pool is 150. All
the class pools are served to create the total class
pool, C, of solid domain block within our proposed
searching algorithm except those 10 smooth class
pools.
274 Chao and Chieu
3.3. The Adaptive Fractal Image Sequence
Encoding and Decoding
The variable shape contractive transformation w
m
i, j
for
R
m
i, j
is , where , , and are spatial contrac-
tion, gray-level permutation, and massic transforma-
tion, respectively. , , and are proposed as follows.
Spatial Contraction (). The spatial contraction ()
shrinks the solid domain block D
m
k
m
to the size of
the R
m
i, j
and translates the solid domain block D
m
k
m
to
the position of the R
m
i, j
, where m
k
m
)
x+p,y+q,t +r
, is the average gray-level value of
adjacent
x
y
t
pixels in D
m
k
m
. We formulate it
as follow:
_
D
m
k
m
_
x+p,y+q,t +r
=
1
x
1
i =0
y
1
j =0
t
1
k=0
f (x + p
x
+i, y +q
y
+ j, t +r
t
+k), (12)
Figure 6. Thirty-three gray-level permutations for shape-9, where a, b, . . . , and h are positions of eight corners of the original solid image
block.
where (x, y, t ) and f () denote the starting position
and gray-levels of pixel of the D
m
k
m
. In general,
x
,
y
,
and
t
are set to be natural numbers.
Gray-Level Permutation (). () principally per-
forms rotations and reections on the output of the
(D
m
k
m
k
m
k
m
))
in order that the contrast and brightness of
v
((D
m
k
m
))
are similar to those of the R
m
i, j
. Let be the scal-
ing factor and denotes the brightness offset, the
Adaptive Fractal Image Sequence Coding 275
transformation () can be dened as
v
_
_
D
m
k
m
___
=
v
_
_
D
m
k
m
__
+. (13)
Given two blocks H R
m
i, j
and G
v
((D
m
k
m
)), we
can nd
=
E[H G] E[H] E[G]
E[G
2
] (E[G])
2
=
Cov(H, G)
Var(G)
(14)
and
= E[H] E[G] (15)
by using (6), (13), and the least square rule. Var() and
Cov() are the variance operator and the covariance
operator, respectively. If G H, (14) can be given by
=
Cov(H, G H)
Var(G)
=
Var(H)
Var(G)
. (16)
If the attribution of R
m
i, j
belongs to the smooth
class, then Var(H)
=0,
=0, and
= E[H]. On the
contrary, if the attribution of R
m
i, j
belongs to ano-
ther texture classes, we choose {0.2, 0.3, 0.4, 0.5,
0.6, 0.7, 0.8, 0.9} which is the same as that in [3].
We nowpropose a searching algorithmto nd a con-
stractive transformation w
m
i, j
for the encoding of R
m
i, j
.
The w
m
i, j
searching algorithm is given as follows.
Searching Algorithm
For the encoding of R
m
i, j
.
If the attribution of the R
m
i, j
is smooth,
Calculate the mean value E[R
m
i, j
] and let the pa-
rameters of w
m
i, j
include both E[R
m
i, j
] and the attri-
bution of the R
m
i, j
.
else /
k
m
k
m
k
m
)).
Moreover, the d
mse
(R
m
i, j
, (
v
((D
m
k
m
)))) must
have a minimal distortion. The parameters of w
m
i, j
should include the starting position of D
m
k
m
,
v
(),
, , and the attribution of the R
m
i, j
.
End of Searching Algorithm
The encoding algorithm of a testing image sequence
block is summarized as follows.
Encoding Algorithm
1. A testing image sequence block U is applied as an
input.
2. Partition the testing image sequence block U to the
corresponding solid domain blocks D
m
k
m
. For each
D
m
k
m
, apply the projection classication algorithmto
create the total class pool C.
3. The testing image sequence block U is divided to
non-overlapping N N (N/2) solid range blocks
R
i
, i = 1, . . . , N
R
.
4. For each solid range blocks R
i
, i = 1, . . . , N
R
,
(a) Apply the 3D projection classication algo-
rithm and the 3D variable shape decomposition
algorithm to create one of the fteen possible
merging patterns.
For the encoding of R
m
i, j
,
Do the w
m
i, j
searching algorithm.
(b) Store the form of the merging pattern of the R
i
and all parameters of the w
m
i, j
.
End of Encoding Algorithm
The decoding algorithm of a testing image seq-
uence block can be described by the following proce-
dure.
Decoding Algorithm
1. Initialize an arbitrary image sequence block V.
2. Assign the iteration number, k, to be 0.
3. Do
(a) Calculate a new fractal image sequence block
V by
all of R
m
i, j
w
m
i, j
().
(b) k = k +1.
(c) Replace V by the new fractal image sequence
block
V, until d
mse
(V,
V) c
. c
is an error
bound.
End of Decoding Algorithm
4. Simulation Result
By applying the above algorithms to each testing im-
age sequence block sequentially, we performcomputer
simulation of the original image sequences of sales-
man. There are 64 (#0 to #63) frames in this testing
image sequence and each frame has spatial resolution
of 360288 pixels with 8 bits gray-level for each pixel.
276 Chao and Chieu
Figure 7. The original eight frames (#8 to #15) of salesman.
Each testing image sequence block U is assigned to
have eight frames, and the original eight frames (#8 to
#15) of salesman are displayed in Fig. 7. We employ
the 3D projection classication algorithm and the 3D
variable shape decomposition algorithm to the adap-
tive fractal image sequence coding for each testing im-
age sequence block. Let the size of R
i
is 8 8 4
and
x
=
y
=
t
=2. The values for t
s
and t
e
are 20
and 36, respectively. Three bits are required for scal-
ing factor of each texture block, and each brightness
offset requires 8 bits. Because the shift along x and y
axial direction for the overlapping solid domain blocks
are 4-pixel per step, () requires 13 bits to represent
the place of the solid domain blocks of the testing im-
age sequence block U. We need one bit to represent the
attribution (smooth or texture class) of the R
m
i, j
.
Let n
, n
, n
v
, n
, n
E
, and n
A
, denote the number
of bits for , ,
v
, , the mean value of the pixels in
the smooth block, and the attribution of shape compo-
nent, respectively. From previous discussion, we know
that the required bits for encoding w
m
i, j
, n
w
m
i, j
, can be
summarized as follows:
Adaptive Fractal Image Sequence Coding 277
Figure 8. The decoded frames (#8 to #15) from salesman at the tenth reconstructed iteration.
(1) n
w
m
i, j
= n
A
+n
E
= 1 +8 = 9 (bits), if the attribu-
tion of R
m
i, j
is smooth.
(2) n
w
m
i, j
= n
A
+n
+n
+n
v
+n
= 1 +13 +3 +
n
v
+8 = 25 +n
v
(bits), if the attribution of R
m
i, j
is not smooth, in which
n
m
v
_
_
6, for the case that R
m
i, j
is a shape-9
solid image block,
5, for the case that R
m
i, j
is a shape-5
or shape-6 solid image block,
4, otherwise.
(17)
The required average bit rate (BR) for each decoded
image sequence block can be calculated by using
BR =
all of R
m
i,j
n
w
m
i, j
+
N
R
i =1
n
R
i
360 288 8
(bits/pixel). (18)
n
R
i
requires four bits to accommodate the classication
of the merging patterns into 15 forms. The peak sig-
nal to noise ratio (PSNR
t
) of t th frame of the decoded
278 Chao and Chieu
Figure 9. Curves for salesman: (a) curve of PSNR and (b) curve of
BR.
image sequence is dened by
PSNR
t
= 10 log
_
255
2
d
mse
(K
t
, L
t
)
_
(dB), (19)
where K
t
and L
t
denote the t th frame of the origi-
nal image sequence and decoded image sequence, re-
spectively. The decoded image sequence (#8 to #15)
of salesman by the adaptive fractal image sequence
coding technique are displayed in Fig. 8. The curve of
PSNRs and BRs for the adaptive fractal image sequence
coding are shown in Figs. 9(a) and (b), respectively.
5. Conclusions
In this paper, we have described a new image sequence
classication scheme which is based on the 3D pro-
jection algorithm to estimate the texture attribution of
the image sequence block. We have also adopted the
3D variable shape decomposition skill, the projection
classication algorithm, and the fractal coding tech-
nique to form the new image sequence coding scheme.
It is a general phenomenon that the decoded image
sequence quality and the transmitting bit rate depend
on the choice of size of the solid block, block shape,
and block characteristics. An adaptive algorithm to al-
locate all these factors in a proper way is thus necessary.
In the fractal coding setup, we propose the projection
classication algorithm to encode variable shape. A
better trade-off between the image quality and trans-
mitting bit rate is obtained. From this study we have
the following conclusions:
(1) In our coding scheme, we use a solid range block
with xed size and we restrain the size to 884.
This choice is due to the trade-off between the im-
age quality and transmitting bit rate. For a larger
solid range block, the bit rate can be reduced
whereas the image quality is downgraded. On the
contrary, for a smaller solid image block, the im-
age quality can be improved however the bit rate
is increased.
(2) For the 9th shape components, we will have 33
gray-level permutations. We may neglect a gray-
level permutation that seldom occurs in order to
reduce the bit rate. Therefore, it only needs ve
bits to characterize the information for encoding.
(3) Further improvement of the proposed image se-
quence coding systembetween the transmitting bit
rate and the image sequence quality may be ex-
pected if the wavelet decomposition (or parallel
processing) is used. This is also our future research
topic.
Acknowledgments
The authors would like to express the most grateful
acknowledgments to the National Science Council of
the R.O.C. for the nancial support under the contract
number NSC-87-2213-E011-032.
References
1. M. Kunt, M. Benard, and R. Leonardi, Recent results in high-
compression image coding, IEEE Trans. Circuits and Systems,
Vol. 34, No. 11, pp. 13061336, 1987.
2. M.F. Barnsley and A.D. Sloan, A better way to compress im-
ages, Byte Mag., Vol. 13, No. 1, pp. 215224, 1988.
3. A.E. Jacquin, Image coding based on a fractal theory of it-
erated contractive image transformations, IEEE Trans. Image
Process., Vol. 1, No. 1, pp. 1830, 1992.
4. D.M. Monro and F. Dudbridge, Fractal block coding of im-
ages, Electronics Letters, Vol. 28, No. 11, pp. 10531055, 1992.
Adaptive Fractal Image Sequence Coding 279
5. E.W. Jacobs, Y. Fisher, and R.D. Boss, Image compression: A
study of the iterated transformmethod, Signal Process., Vol. 29,
No. 3, pp. 251263, 1992.
6. H. Li, M. Novak, and R. Forchheimer, Fractal-based image
sequence compression scheme, Optical Eng., Vol. 32, No. 7,
pp. 15881595, 1993.
7. G. Lu, Fractal image compression, Signal Process.: Image
Comm., Vol. 5, No. 4, pp. 327343, 1993.
8. M.F. Barnsley, Fractals Everywhere, Academic: San Diego,
CA, 1988.
9. B. Moghaddam, K.J. Hintz, and C.V. Stewart, Fractal image
compression and texture analysis, in Proc. SPIEConf. on Image
Understanding, 1990, Vol. 1406, pp. 4257.
10. A.E. Jacquin, A novel fractal block-coding technique of digital
images, in Proc. IEEE Int. Conf. on Acoust., Speech, Signal
Process., 1990, Vol. 4, pp. 22252228.
11. I.K. Kim and R.H. Park, Still image coding based on vec-
tor quantization and fractal approximation, IEEE Trans. Image
Process., Vol. 5, No. 4, pp. 589597, 1996.
12. D.M. Monro and F. Dudbridge, A hybrid fractal transform, in
Proc. IEEEInt. Conf. on Acoust., Speech, Signal Process., 1993,
Vol. 5, pp. 169172.
13. S. Lepsoy, G.E. Oien, and T.A. Ramstad, Attractor image com-
pression with a fast noniterative decoding algorithm, in Proc.
IEEE Int. Conf. on Acoust., Speech, Signal Process., 1993,
Vol. 5, pp. 337340.
14. G. Vines and M.H. Hayes, Adaptive IFS image coding with
proximity maps, in Proc. IEEE Int. Conf. on Acoust., Speech,
Signal Process., 1993, Vol. 5, pp. 349352.
15. L.C. Thomas and F. Deravi, Pruning of the transform space
in block-based fractal image compression, in Proc. IEEE Int.
Conf. on Acoust., Speech, Signal Process., 1993, Vol. 5, pp. 341
344.
16. C.J. Sze et al., Fractal image coding system based on an adap-
tive side-coupling quadtree structure, Image and Vision Com-
puting, Vol. 14, No. 6, pp. 401415, 1996.
17. R.A. Brooks and G. Dichiro, Principles of computer assisted
tomography (CAT) in radiographic and radioisotopic imaging,
Phys. Med. Biol., Vol. 21, pp. 689732, 1976.
18. G.T. Herman, Image Reconstruction fromProjections: The Fun-
damentals of Computerized Tomography, Academic Press: New
York, 1980, pp. 60.
19. A. Macovski, Medical Imaging Systems, Prentice-Hall:
Englewood Cliffs, NJ, 1983.
Her-ChangChaowas borninTaoyuan, TaiwanonOctober 25, 1964.
He received the S.B. and Ph.D. degrees, both in Electronic Engineer-
ing, from National Taiwan University of Science and Technology,
Taipei, Taiwan, ROC, in 1991 and 1998, respectively. Presently he
is an Assistant Professor at the Department of Electronic Engineer-
ing, VanNungInstitute of Technology, Chung-Li, Tao-Yuan, Taiwan,
ROC. His research interests are in the area of image processing, digi-
tal lter design, wavelet transforms, computer vision, and neural
networks.
Bin-Chang Chieu received the Ph.D. degree in Electrical Engineer-
ing from Rensselaer Polytechnic Institute, Troy, NY, USA, in 1989.
He is now Professor at the Department of Electronic Engineering,
National Taiwan University of Science and Technology, Taipei,
Taiwan, ROC. His current research interests are in image processing,
digital signal processing, neural networks, and computer vision.