Documente Academic
Documente Profesional
Documente Cultură
,QWHUQDWLRQDO&RQIHUHQFHRQ6LJQDO3URFHVVLQJDQG,QWHJUDWHG1HWZRUNV63,1
International Conference on Signal Processing and Integrated Networks (SPIN)
B. S. P. Mishra
B. Majhi
S. Dehuri
Department of CSE
Seemanta Engg. College
ppsarangi@gmail.com
Department of CSE
KIIT, Bhubaneswar
mishra.bsp@gmail.com
Department of CSE
NIT, Rourkela
bmajhi@nitrkl.ac.in
Department of ICT
FM University,Balasore
satchi.lapa@gmail.com
KeywordsDifferential evolution, image enhancement, histogram equalization, contrast stretching, particle swarm optimization
I.
I NTRODUCTION
II.
95
(1)
In order to achieve an automatic image enhancement technique and no objective parameters adjustment by the human
intervention, a criterion to evaluate the quality of the enhanced
image should be chosen. This criterion is considered as tness
function of the DE. To describe the criterion, it is required to
mention that the enhanced image has more number of edgels
and also it has high intensity of edges [6]. But these two are not
sufcient to describe the tness function for a good enhanced
image. Another one measure has been added i.e. entropy value
of the image. Entropy value indicates the average information
content in the image. For enhanced image the histogram of the
gray-levels are uniformly distributed that increase the entropy
of the image. These three measures together describe the
following objective criterion and in this study it is used as
tness function to be maximized.
nedgels(IS )
(5)
.H(Ie )
F (Ie ) = log(log(E(IS )))
M xN
In the eq. (5), Ie is the enhanced image of the original image
produced by the transformation function mentioned in eq.
(4).The edges or edgels can be detected by using many efcient
edge detector algorithms. In this paper Sobel edge detector [15]
is used as automatic threshold detector [Rosin]. Other edge
detector algorithms such as Laplacian [15] and Canny [16]
can be used to nd edges or edgels of the enhanced image.
The sobel edge operator produce an output image Is from the
input enhanced image Ie as:
(6)
Is (i, j) = hIe2 + vIe2
where f(x, y) is the intensity value of the xth column and yth
row coordinate pixel of the input image and g(x, y) is the
modied intensity value of the same coordinate pixel of the
processed output image. T (.) is an operator on f, dened over
some neighborhood N of (x, y). For N we use mostly a square
sub image that is centered at (x, y). The histogram equalization
is a point processing technique, in which transformation of
one pixel depends on the histogram of the overall image.
Overall enhancement performance is good in this method but
sometimes local enhancement of the region is preferred. In
local enhancement, modication of a pixel based on graylevel distribution among its N sub images centered at that
pixel [11]. One of the examples of local enhancement method
is Adaptive histogram equalization (AHE) which works efciently on medical images [14]. However, AHE is a very
time consuming approach. Our approach used in this paper
for local enhancement is less time consuming method similar
to statistical scaling presented in [13].
A. Local enhancement method
Local enhancement method uses transformation function
that based on the gray-level distribution in the neighborhood
window of size (n x n) centered at each pixel in the input
image. The transformation function [11] applied to each pixel
at coordinates (x, y) using following equation
g(x, y) = A(x, y).[f (x, y) m(x, y)]
with
A(x, y) =
M
with0 < k < 1
(x, y)
(2)
(3)
In equation (2) and (3), m(x, y) the local mean and (x, y)
local standard deviation are computed in an n x n neighborhood
centered at (x, y) pixel of the input image. Thus, the values
255
ei
(7)
i=0
The population of DE consists of NP D dimensional parameters target vectorsXi,G = [X1,i,G , X2,i,G , ......XD,i,G ],i =
1, 2, ....., N P for each generation G. The parameters of each
vectors of the population are randomly select values, which
U
uniformly distributed on the intervals [xL
j , xj ].Then the tness
value of each individual target vector is calculated through
evolution. DE uses random parent selection regardless of their
tness values. In every generation, each individual xi,G gets a
chance to become parent and depends on the other randomly
select auxiliary parents. Three auxiliary parents xr1,G , xr2,G ,
and xr3,G are selected randomly such that r1, r2, r3
{1, 2, ..., N P } and i = r1 = r2 = r3. Then new vector xi,G
is called donor vector can be derived as follows:
vi,G = xr1,G + F (xr2,G xr3,G )
(8)
Size
couple
eight
pout
tire
zelda
TABLE II.
512
256
240
205
512
x
x
x
x
x
512
256
291
232
512
NP/G/W
Fitness
DV
BV
40/50/5
40/20/7
40/20/3
30/30/3
40/50/5
2.2492
0.8787
1.0922
1.9694
15.547
0.231
0.182
0.077
0.209
0.333
0.0070
0.0065
0.0087
0.0014
0.0043
Image
Measures
LCS
HE
PSO-IE
DE-IE
couple
Fitness
DV
BV
1.5600
0.2306
0.0039
4.0096
0.3194
0.0072
3.2222
0.3151
0.0044
3.3639
0.2828
0.0044
Eight
Fitness
DV
BV
1.3907
0.2381
0.0065
1.1286
1.1286
0.0072
2.4307
0.1139
0.0074
2.4756
0.1190
0.0072
Pout
Fitness
DV
BV
1.2516
0.1805
0.0077
1.2776
0.2912
0.0071
1.8746
0.0953
0.006
1.8842
0.0864
0.0087
Tire
Fitness
DV
BV
1.9514
0.2164
0.0014
1.8277
0.2643
0.0075
2.3345
0.1872
0.0013
2.3566
0.1871
0.0014
R EFERENCES
Fitness
DV
BV
1.8811
0.3739
0.0035
15.652
0.3610
0.0036
16.986
0.3284
0.0044
17.035
0.3348
0.0043
Zelda
TABLE III.
ACKNOWLEDGMENT
[1]
[2]
[3]
Image
Measures
LCS
HE
PSO-IE
DE-IE
couple
eight
pout
tire
zelda
4718
3115
3190
3198
3255
3357
5076
3599
3175
4011
8581
4121
3656
3101
3266
6960
6650
4562
3992
3268
7317
6772
5390
4040
3273
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
C ONCLUSION
[15]
99
(a)
(b)
(a)
(b)
(c)
(d)
(c)
(d)
(e)
(e)
Fig. 2.
Couple Image: (a) Original Image, (b) Linear Contrast Stretching(LCS), (c) Histogram Equalization (HE), (d) PSO based Method (PSO-IE),
and (e) DE based Method (DE-IE)
Fig. 4. Pout Image: (a) Original Image, (b) Linear Contrast Stretching (LCS),
(c) Histogram Equalization (HE), (d) PSO based Method (PSO-IE), and (e)
DE based Method (DE-IE)
(a)
(b)
(a)
(b)
(c)
(d)
(c)
(d)
(e)
(e)
Fig. 3. Eight Image: (a) Original Image, (b) Linear Contrast Stretching (LCS),
(c) Histogram Equalization (HE), (d) PSO based Method (PSO-IE), and (e)
DE based Method (DE-IE)
Fig. 5. Tire Image: (a) Original Image, (b) Linear Contrast Stretching (LCS),
(c) Histogram Equalization (HE), (d) PSO based Method (PSO-IE), and (e)
DE based Method (DE-IE)
100