A Fast Fractal Image Coding Scheme
A Fast Fractal Image Coding Scheme
new fast fractal image coding (FFIC) scheme is presented, in which the fractal parameters are directly obtained from the range and domain blocks. Compared with the conventional fractal image coding (FIC) scheme, FFIC is 6 times faster in compression speed at a little loss of PSNR of about 0.5dB.
Abstract-A
For firthcr details on fractal image compression scheme, please see [ 11. FIC uses fill-search algorithm to find the best a b , T and g and is very slow in compression speed. This paper proposes a fast fractal image coding (FFIC) scheme It utilizes DCT to classify image blocks, firther to fix the geometric transform T, and then obtains n and b directly from f and T(@ in terms with solving an optimum problem.
3
>J
.=
( 2iJJ
U
+U
(1)
where
We classify image blocks as 10 types as illustrated in Fig. 1, where B denotes pixels with bigger grey values and S represents pixels with smaller grey values. For examples, type 0 indicates image blocks whose pixel values vary slowly and type 1 describes image blocks whose pixels in the right are almost bigger than in the left. Type 9 is composed of what cant be categorized by the forgoing 9 types.
t,j=N-l
,,,=O
(f . - a . T ( g ) . . - b ) 1,J
3 J
where a ~(0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8)
9
T E (10 > 4 > 4 4 > 4 4 4 > 4 ) ; b = f - a - T(g)= - a . g ; g E {all possible domain blocks ( N x N )E G };
3
9
wPe9,
U-7803-2912-0
1047
type0 if C(O,O)/E2 99.99 % , else type 1 if IC(O,l)l>>IC(1,O)I and C(0,l) < 0 , else
type if IC(l,O)i >> IC(0,l)I and C(1,O) > 0 , else type 5 iflC(0,l)i = iC(1,O)I and C(0,l) > Oand C(1,O) > 0,else type 6 iflC(0,l)i =lC(l,O)l and C(0,l) < Oand C(1,O) < 0,else
Fig. 2 three DCT coefficients Having defined 10 lunds of image blocks and chosen 3 DCT coefficients, we present the following classification algorithm. Class!fication Alnori thm Mark E as the image energy, an image block is categorized as:
type 9
Once the types of range blocks(R) and domain blocks(D) are known, it is easy to decide the optimum geometric transform T* by looking up table 1.
Table 1 geometric transform T between different kinds of images (transform domain block D only)
AD type0
type 1 type 2 type 3 type 4 type 5 type 6 type 7 type 8 type 9
lo lo lo lo lo lo lo lo lo
type1 lo
1 1
type2
1 1
type3
17
type4
type5
type6 lo
1 1
type7
I1
type8 lo
11
type9
Is
17
12
Is
lo
12
lo
17
Is
lo
12
Is
17
I1
lo
17
Is
lo ll lo
I1
lo
I1
1 3 . lo lo
17
lo lo
I2
12
lo lo
l4
lo lo 1 5 .
15 . lo
17
k
lo
1s
17
Is
17
15
13
lo lo
lo
lo
lo lo
17 1s
lo
13
lo
lo
lo
lo lo
lo lo lo lo lo lo lo lo lo
1048
the same way as conventional FIC. The software platform is Borlrrnd C++ 3.1 and the hardware 1s Pentium 586/90MHz PC. No optimum in programming is performed for speed. The standard 2 image Lena (512 ~ 5 1 with 8 bbp) is partitioned into four parts and each part is encoded independently.
a. =
I , J d
c [(A,,-3h,-41 c [h,-h)2]
I , ] = N -1
(3)
1.1=0
b'= f -o'.h
where h = T' ( g ).
(4)
4 Implementing FFIC
size takes and In Our experiment,the range S. Domain lS posltloned at the multlplesOf possible domain blocks To encode an image, its
are using the above method and the classification results are stored in the memory. A range block is also classified. If it belongs to type 0, the mean pixel value is encoded as its private pixels; otherwise, table 1 and (3) and (4) are used to fix the T, a and b respectively. We allocate 3 bits for T(8 transforms), 3 bits for n ( limiting it in the interval
The experiment results are shown in table 2. The functions of PSNR and Time with respect to CR are plotted in Fig. 3 and Fig. 4, respectively The solutions of n and b in three schemes are the sameby ( 3 ) and (4). FIC-1 does not utilize geometric transform and FIC-2 tries 8 geometric transforms and chooses the best one. FFIC is the fast fractal image coding scheme. From Fig. 3 and Fig. 4, we can see that FIC-2 achieves the best compression quality at the same CR among above three coding schemes. However, it takes terrible long time in compression, FFIC sLstimes faster than FIC-2 at a little degradation of PSNR of about 0.5dB. Compared with FIC-1, FFIC obtains about 0.6dB gains of PSNR with near,y the SaMe compression speed,
6 Conclusions
from o*l to o.8 With quantization and bits for b (limiting it in the range from -255 to 255 with quantization step 2). Other parameters are encoded in
Table 2 test results using different coding schemes CR time 7.88 8.83 8.67 test 1 PSNR Time dB minute 3194 81 33.01 419 32.56 72
This paper studies the relationship between transform coefficients a d geometric $eatures of an image, and thus utilizes Discre Cosine Transform
CR
time 955 12.00 10.57
test 2 PSNR
test 3 CR PSNR Time time dB minute 11.64 31.62 65 13.83 31.91 373 58 12.30 31.92
R 14.72 14.31
Time
55
1049
33.5
33
32.5
A
m
-U
( I L
3231.5 -
3 1 30.5
30
to fix the geometric transform parameter T quickly. It also obtains parameters a and b in terms with solving m problem. Experiment results show that this 1s a PromisW way to speed U the fractal image P compression progress.
References
1 A E Jacquin, "Fractal image coding. A review,"
Proc. IEEE, vol. 81, no. 10, pp. 1451-1465, Oct 1993. 2 B.E. Wohlberg and G.de Jager, "Fast image domain fractal compression by DCT domain block matching," Electron Lett , vol 3 1, pp
869-870, May 1995. 3 Y Zhao and B. Yuan, "Image compression using fractals and discrete cosine transform," Electron Lett, vol 30, pp 474-475, March 1994 4 R N.Bracewell, K.-Y. Chang, A K. Jha and Y -H Wang, "Affine theorem for two-dimensional fourier transform," Electron Lett, vol. 29, pp 304, Feb. 1993. 5 A K. Jain, Fundamentals of digital image processing, Prentice Hall, Englewood Cliffs, NJ 07632, 1989.
1050