I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
4
,
A
ugus
t
2025
, pp.
3274
~
3286
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
4
.pp
3274
-
3286
3274
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
M
od
i
f
i
e
d
z
e
r
o
-
r
e
f
e
r
e
n
c
e
d
e
e
p
c
u
r
ve
e
st
i
m
at
i
on
f
or
c
on
t
r
ast
q
u
al
i
t
y e
n
h
a
n
c
e
m
e
n
t
i
n
f
ac
e
r
e
c
ogn
i
t
i
on
M
u
h
am
m
ad
K
ah
f
i
A
u
li
a
1
, D
yah
A
r
u
m
in
g T
ya
s
2
1
M
a
s
t
e
r
P
r
ogr
a
m
i
n C
om
put
e
r
S
c
i
e
nc
e
s
, F
a
c
ul
t
y of
M
a
t
he
m
a
t
i
c
s
a
nd N
a
t
ur
a
l
S
c
i
e
nc
e
s
,
U
ni
ve
r
s
i
t
a
s
G
a
dj
a
h
M
a
da
, Y
ogya
k
a
r
t
a
, I
ndone
s
i
a
2
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd E
l
e
c
t
r
oni
c
s
, F
a
c
ul
t
y of
M
a
t
he
m
a
t
i
c
s
a
nd
N
a
t
ur
a
l
S
c
i
e
nc
e
s
,
U
ni
ve
r
s
i
t
a
s
G
a
dj
a
h
M
a
da
,
Y
ogya
ka
r
t
a
, I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
J
ul
11, 2024
R
e
vi
s
e
d
A
pr
11, 2025
A
c
c
e
pt
e
d
J
un 8, 2025
Face
recognition
systems
remain
challenged
by
variable
lighting
cond
itions.
While
zero
-
reference
deep
curve
estimati
on
(
Zero
-
DCE
)
effe
ctively
enhances
low
-
light
images,
it
frequently
induces
overexposure
in
n
ormal
-
and
high
-
brightness
scenarios.
This
study
introduces
modified
Zero
-
DCE
combined
with t
hree establ
ished enh
ancement t
echniques
: contras
t stre
tching
(CS),
contrast
limited
adaptive
histogram
equalization
(CLAHE
),
and
brightness
preserving
dynamic
histogram
equalization
(BP
DHE).
Evaluations
employed
the
extended
Yale
face
database
B
an
d
face
recognition
technology
(FERET)
datasets,
with
10
representative
s
amples
assessed
usi
ng
the
blind/referenceless
image
spatial
quality
ev
aluator
(BRISQUE)
metric.
Modified
Zero
-
DCE
with
BPDHE
produced
o
ptimal
enhancement
quality
,
achieving
a
mean
BRISQUE
score
of
16.018.
On
the
extended
Yale
face
database
B,
visual
geometry
group
16
(
V
GG16
)
integrated
with
modified
Zero
-
DCE
and
CLAHE
attained
8
3.65%
recognition
accuracy,
representing
a
6.08
-
percentage
-
point
impro
vement
over
conventional
Zero
-
DCE.
For
the
200
-
subject
FERET
subset,
r
esidual
network
50
(
ResNet50
)
with
modified
Zero
-
DCE
and
CLAHE
ac
hieved
67.41%
accuracy.
Notably,
standard
Zero
-
DCE
with
CLAHE
demon
strated
superior
robustness
in
extremely
low
-
light
conditions,
highlighti
ng
the
illumination
-
dependent
performance
characteristics
of
these
enhan
cement
approaches.
K
e
y
w
o
r
d
s
:
B
r
ig
ht
ne
s
s
D
e
e
p l
e
a
r
ni
ng
F
a
c
e
r
e
c
ogni
ti
on
I
m
a
ge
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
Z
e
r
o
-
r
e
f
e
r
e
nc
e
de
e
p c
ur
ve
e
s
ti
m
a
ti
on
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
D
ya
h A
r
um
in
g T
ya
s
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
a
nd E
le
c
tr
oni
c
s
, F
a
c
ul
ty
of
M
a
th
e
m
a
ti
c
s
a
nd N
a
tu
r
a
l
S
c
ie
nc
e
s
U
ni
ve
r
s
it
a
s
G
a
dj
a
h M
a
da
B
ul
a
ks
um
ur
, C
a
tu
r
tu
ngga
l,
D
e
pok, S
le
m
a
n, Y
ogya
k
a
r
ta
, I
ndone
s
ia
E
m
a
il
:
dya
h.a
r
um
in
g.t
@
ugm
.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
F
a
c
e
r
e
c
ogni
ti
on
is
c
r
uc
ia
l
f
or
id
e
nt
if
yi
ng
a
nd
ve
r
if
yi
ng
f
a
c
ia
l
c
la
im
s
[
1]
.
H
ow
e
ve
r
,
f
a
c
e
r
e
c
ogni
ti
on
s
ys
te
m
s
f
a
c
e
c
ha
ll
e
ng
e
s
due
to
dyn
a
m
ic
e
nvi
r
onm
e
nt
s
,
pa
r
ti
c
ul
a
r
ly
dyna
m
ic
li
ght
in
g
a
nd
lo
w
-
li
ght
s
c
e
na
r
io
s
.
T
he
s
e
a
f
f
e
c
t
th
e
f
a
c
ia
l
f
e
a
tu
r
e
s
,
r
e
duc
in
g
a
c
c
ur
a
c
y
a
nd
in
c
r
e
a
s
in
g
m
is
id
e
nt
if
ic
a
ti
on
[
2]
,
[
3
]
.
A
ddr
e
s
s
in
g
th
e
s
e
c
ha
ll
e
nge
s
r
e
qui
r
e
s
in
nova
ti
ve
a
ppr
oa
c
he
s
to
e
nha
n
c
e
im
a
ge
qua
li
ty
be
f
or
e
r
e
c
ogni
ti
on,
e
m
pha
s
iz
in
g
th
e
ne
e
d f
or
e
f
f
e
c
ti
ve
pr
e
pr
oc
e
s
s
in
g m
e
th
ods
.
O
ne
pr
om
is
in
g
a
ppr
oa
c
h,
z
e
r
o
-
r
e
f
e
r
e
nc
e
de
e
p
c
ur
ve
e
s
ti
m
a
ti
on
(
Z
e
r
o
-
D
C
E
)
[
4]
,
ha
s
pr
ove
n
e
f
f
e
c
ti
ve
in
e
nha
n
c
in
g
th
e
c
ont
r
a
s
t
qu
a
li
ty
of
lo
w
-
li
ght
im
a
ge
s
.
Z
e
r
o
-
D
C
E
is
li
ght
w
e
ig
ht
a
nd c
om
put
a
ti
ona
ll
y
e
f
f
ic
ie
nt
,
m
a
ki
ng
it
s
ui
ta
bl
e
f
or
r
e
a
l
-
ti
m
e
a
ppl
ic
a
ti
ons
[
5]
.
H
o
w
e
ve
r
,
it
s
li
m
it
a
ti
ons
s
uc
h
a
s
ove
r
e
xpos
ur
e
in
nor
m
a
l
or
hi
gh
-
br
ig
ht
ne
s
s
im
a
ge
s
hi
ghl
ig
ht
th
e
ne
e
d
f
or
m
odi
f
ic
a
ti
ons
to
im
pr
ove
a
da
pt
a
bi
li
ty
a
c
r
o
s
s
br
oa
d
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
odi
fi
e
d z
e
r
o
-
r
e
fe
r
e
nc
e
d
e
e
p c
u
r
v
e
e
s
ti
m
at
io
n f
or
c
ont
r
a
s
t
qua
li
ty
…
(
M
uhamm
ad K
ahf
i
A
ul
ia
)
3275
li
ght
in
g
c
ondi
ti
ons
[
6]
.
O
ve
r
e
xpos
ur
e
c
a
n
le
a
d
to
lo
s
s
of
f
a
c
ia
l
f
e
a
tu
r
e
s
,
de
c
r
e
a
s
in
g
f
a
c
e
r
e
c
ogni
ti
on
a
c
c
ur
a
c
y
or
c
a
us
in
g
m
is
id
e
nt
if
ic
a
ti
on
[
7]
,
[
8]
.
P
r
e
vi
ous
s
tu
di
e
s
ha
v
e
p
r
opos
e
d
va
r
io
us
e
nha
n
c
e
m
e
nt
s
to
Z
e
r
o
-
D
C
E
,
s
uc
h
a
s
Z
e
r
o
-
D
C
E
+
+
[
5]
a
nd
Z
e
r
o
-
D
C
E
T
in
y
[
9]
,
bu
t
th
e
s
e
m
odi
f
ic
a
ti
ons
of
te
n
f
a
ll
s
hor
t
of
s
ig
ni
f
ic
a
nt
ly
im
pr
ovi
ng
im
a
ge
qua
li
ty
.
T
he
s
e
li
m
it
a
ti
ons
unde
r
s
c
or
e
th
e
ne
e
d
f
or
a
lt
e
r
na
ti
ve
a
ppr
oa
c
he
s
to
ba
la
nc
e
e
f
f
ic
ie
nc
y a
nd pe
r
f
or
m
a
nc
e
.
T
hi
s
s
tu
dy
pr
opos
e
s
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
,
de
s
ig
ne
d
to
ov
e
r
c
om
e
th
e
s
e
s
hor
tc
om
in
gs
th
r
ough
a
r
c
hi
te
c
tu
r
a
l
c
ha
nge
s
a
nd
in
te
gr
a
ti
on
w
it
h
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
:
c
ont
r
a
s
t
s
tr
e
tc
hi
ng
(
C
S
)
[
10]
,
c
ont
r
a
s
t
li
m
it
e
d
a
da
pt
iv
e
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
(
C
L
A
H
E
)
[
11]
,
a
nd
br
ig
ht
ne
s
s
pr
e
s
e
r
vi
ng
dyna
m
ic
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
(
B
P
D
H
E
)
[
12]
.
T
he
s
e
tr
a
di
ti
ona
l
m
e
t
hods
ha
ve
be
e
n
w
id
e
ly
us
e
d
in
dom
a
in
s
li
ke
m
e
di
c
a
l
im
a
gi
ng
a
nd
f
a
c
ia
l
r
e
c
ogni
ti
on
due
to
th
e
ir
a
da
pt
a
bi
li
t
y
to
va
r
yi
ng
br
ig
ht
ne
s
s
le
ve
ls
.
B
y
c
om
bi
ni
ng
th
e
s
tr
e
ngt
hs
of
bot
h
tr
a
di
ti
ona
l
a
nd
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
m
e
th
ods
,
w
e
a
im
to
a
ddr
e
s
s
dyna
m
ic
li
ght
in
g
c
ha
ll
e
nge
s
m
or
e
e
f
f
e
c
ti
ve
ly
.
T
o
e
va
lu
a
te
th
e
pr
opos
e
d
m
e
th
od,
w
e
e
m
pl
oy
two
c
om
pl
e
m
e
nt
a
r
y
m
e
tr
ic
s
:
th
e
bl
in
d/
r
e
f
e
r
e
nc
e
le
s
s
im
a
ge
s
pa
ti
a
l
qua
li
ty
e
va
lu
a
to
r
(
B
R
I
S
Q
U
E
)
[
13
]
a
nd
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
B
R
I
S
Q
U
E
m
e
a
s
ur
e
s
pe
r
c
e
pt
ua
l
qua
li
ty
w
it
hout
r
e
qui
r
in
g
a
r
e
f
e
r
e
nc
e
im
a
ge
,
m
a
ki
ng
it
id
e
a
l
f
or
a
s
s
e
s
s
in
g
c
ont
r
a
s
t
e
nha
n
c
e
m
e
nt
e
f
f
e
c
ts
.
C
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y di
r
e
c
tl
y e
va
lu
a
te
s
t
he
pe
r
f
or
m
a
nc
e
of
f
a
c
e
r
e
c
ogni
ti
on mode
ls
(
vi
s
ua
l
ge
om
e
tr
y gr
oup
16
(
V
G
G
16
)
[
14]
a
nd
r
e
s
id
ua
l
ne
twor
k
50
(
R
e
s
N
e
t5
0
)
[
15]
)
p
r
e
pr
oc
e
s
s
e
d
w
it
h
our
m
e
th
ods
.
T
he
s
e
m
e
tr
ic
s
c
om
pr
e
he
ns
iv
e
ly
unde
r
s
ta
nd t
he
e
nha
nc
e
m
e
nt
s
'
i
m
pa
c
t
on i
m
a
g
e
qua
li
ty
a
nd r
e
c
ogni
ti
on a
c
c
ur
a
c
y.
T
he
da
ta
s
e
ts
u
s
e
d
in
th
is
s
tu
dy
a
r
e
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
d
a
ta
ba
s
e
B
[
16]
a
nd
f
a
c
e
r
e
c
ogni
ti
on
te
c
hnol
ogy
(
F
E
R
E
T
)
[
17]
,
[
18
]
,
c
hos
e
n
f
or
th
e
ir
di
ve
r
s
e
li
ght
in
g
c
ondi
ti
ons
a
nd
s
ubj
e
c
t
va
r
ia
bi
li
ty
.
T
he
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
t
a
ba
s
e
B
in
c
lu
d
e
s
im
a
g
e
s
c
a
pt
ur
e
d
unde
r
c
ont
r
ol
le
d
li
ght
in
g
va
r
ia
ti
ons
,
w
hi
le
F
E
R
E
T
of
f
e
r
s
a
w
id
e
r
a
nge
of
pos
e
s
,
e
xpr
e
s
s
io
n
s
,
a
nd
s
ubj
e
c
t
de
m
ogr
a
phi
c
s
.
T
he
in
c
lu
s
io
n
of
th
e
lo
w
-
li
ght
da
ta
s
e
t
(
L
O
L
)
[
19]
or
tr
a
in
in
g
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
e
ns
ur
e
s
th
e
m
ode
l
is
opt
im
iz
e
d
f
or
lo
w
-
li
ght
c
ondi
ti
ons
,
a
li
gn
in
g
w
it
h t
he
s
tu
dy'
s
obj
e
c
ti
ve
t
o a
ddr
e
s
s
l
ig
ht
in
g c
ha
ll
e
nge
s
c
om
pr
e
he
ns
iv
e
ly
.
T
hi
s
s
tu
dy
br
id
ge
s
Z
e
r
o
-
D
C
E
'
s
a
da
pt
a
bi
li
ty
ga
p
in
dyn
a
m
ic
li
ght
in
g
by
c
om
bi
ni
ng
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
it
h
tr
a
di
ti
ona
l
e
nha
nc
e
m
e
nt
m
e
th
ods
,
im
pr
ovi
ng
b
ot
h
c
ont
r
a
s
t
qua
li
ty
a
nd
r
e
c
ogni
ti
on
a
c
c
ur
a
c
y.
T
he
a
ppr
oa
c
h
e
nh
a
nc
e
s
r
e
li
a
bi
li
ty
f
or
r
e
a
l
-
w
or
ld
a
ppl
ic
a
ti
on
s
li
ke
s
ur
ve
il
la
n
c
e
a
nd
m
obi
le
a
ut
he
nt
ic
a
ti
on.
R
e
s
ul
ts
d
e
m
ons
tr
a
te
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
'
s
pot
e
nt
ia
l
to
a
dva
nc
e
f
a
c
e
r
e
c
ogni
ti
on
a
dopt
io
n,
und
e
r
s
c
or
in
g
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
'
s
c
r
it
ic
a
l
r
ol
e
i
n opti
m
iz
in
g de
e
p l
e
a
r
ni
ng mode
ls
.
2.
L
I
T
E
R
A
T
U
R
E
R
E
V
I
E
W
R
e
s
e
a
r
c
h
on
f
a
c
ia
l
r
e
c
ogni
ti
on
e
xpl
or
e
s
va
r
io
us
m
e
th
odol
ogi
e
s
.
A
c
om
m
on
a
ppr
oa
c
h
in
vol
ve
s
c
onvolut
io
na
l
ne
ur
a
l
ne
tw
or
ks
(
C
N
N
)
w
it
h
pr
e
tr
a
in
e
d
m
ode
ls
li
ke
V
G
G
16
a
nd
R
e
s
N
e
t5
0
[
20]
,
[
21]
.
W
he
n
te
s
te
d
on
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
,
V
G
G
16
a
nd
R
e
s
N
e
t5
0
a
c
hi
e
v
e
d
46.64
a
nd
46.73%
a
c
c
ur
a
c
y,
r
e
s
pe
c
ti
ve
ly
[
20]
.
A
not
he
r
s
tu
dy
r
e
por
te
d
hi
ghe
r
a
c
c
ur
a
c
ie
s
of
70.78
a
nd
64.87%
[
21]
.
D
e
s
pi
te
im
pr
ove
m
e
nt
s
, t
he
s
e
m
ode
ls
s
tr
uggl
e
w
it
h br
ig
ht
ne
s
s
va
r
ia
ti
ons
, r
e
qui
r
in
g a
ddi
ti
ona
l
e
nha
nc
e
m
e
nt
s
.
Z
e
r
o
-
D
C
E
im
p
r
ove
s
c
ont
r
a
s
t
in
lo
w
-
li
ght
i
m
a
ge
s
[
4]
.
Z
e
r
o
-
D
C
E
is
li
ght
w
e
ig
ht
,
f
a
s
t,
a
nd
s
upe
r
io
r
i
n
non
-
uni
f
or
m
li
ght
in
g
c
ondi
ti
ons
a
nd
lo
w
-
li
ght
in
g
c
a
s
e
s
[
5]
.
L
i
e
t
al
.
[
22]
us
e
d
Z
e
r
o
-
D
C
E
f
or
d
r
ow
s
in
e
s
s
de
te
c
ti
on,
in
c
r
e
a
s
in
g
a
c
c
ur
a
c
y
f
r
om
73.56
to
86.75%
.
Z
hou
[
23]
a
ppl
ie
d
it
f
or
bl
in
k
de
te
c
ti
on,
im
pr
ovi
ng
a
c
c
ur
a
c
y
f
r
om
58.80
to
76.50
%
(
r
ig
ht
e
ye
)
a
nd
70.60
to
94.1
0%
(
le
f
t
e
ye
)
.
G
uo
e
t
al
.
[
4]
f
ound
Z
e
r
o
-
D
C
E
ne
a
r
ly
m
a
tc
he
d
R
e
ti
ne
xN
e
t
in
pr
e
c
is
io
n
-
r
e
c
a
ll
but
ou
tp
e
r
f
or
m
e
d
ot
he
r
m
e
th
ods
l
ik
e
lo
w
-
li
ght
im
a
ge
e
nha
nc
e
m
e
nt
(
L
I
M
E
)
a
nd E
nl
ig
ht
e
n
G
A
N
. W
e
i
e
t
al
.
[
6]
m
od
if
i
e
d Z
e
r
o
-
D
C
E
t
o ope
r
a
te
i
n t
he
hue
, s
a
tu
r
a
ti
on,
a
nd va
lu
e
(
H
S
V
)
c
ol
or
s
pa
c
e
, a
c
hi
e
vi
ng t
he
hi
ghe
s
t
pe
a
k
s
ig
na
l
-
to
-
noi
s
e
r
a
ti
o
(
P
S
N
R
)
o
f
16.75 dB
a
nd l
ow
e
s
t
m
e
a
n
a
bs
ol
ut
e
e
r
r
or
(
M
A
E
)
of
98.78,
out
pe
r
f
o
r
m
in
g
E
n
li
ght
e
n
G
A
N
a
nd
R
e
ti
ne
xN
e
t.
H
ow
e
ve
r
,
va
r
ia
nt
s
li
ke
Z
e
r
o
-
D
C
E
+
+
[
5]
a
nd
Z
e
r
o
-
D
C
E
T
in
y
[
9]
s
how
e
d
m
in
im
a
l
i
m
pr
ove
m
e
nt
.
M
or
e
a
dva
nc
e
d
m
odi
f
ic
a
ti
ons
,
in
c
lu
di
ng
z
e
r
o
-
r
e
f
e
r
e
nc
e
r
e
s
id
ua
l
a
tt
e
nt
io
n
de
e
p
c
ur
ve
e
s
ti
m
a
t
io
n
(
Z
e
r
o
-
R
A
D
C
E
)
[
24]
,
z
e
r
o
-
r
e
f
e
r
e
nc
e
lo
w
-
li
ght
im
a
ge
e
nha
n
c
e
m
e
nt
w
it
h
in
tr
in
s
ic
noi
s
e
r
e
duc
ti
on
(
Z
e
r
o
-
L
E
I
N
R
)
[
25]
,
a
nd
B
é
z
i
e
r
c
ur
ve
e
s
ti
m
a
ti
on
(
B
é
z
ie
r
CE
)
[
26]
, e
nha
nc
e
c
ont
r
a
s
t
a
nd r
e
duc
e
noi
s
e
.
T
r
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
t
e
c
hni
que
s
, s
uc
h a
s
C
L
A
H
E
,
ha
ve
a
ls
o i
m
pr
ove
d f
a
c
e
r
e
c
ogni
ti
on.
W
it
hout
e
nha
nc
e
m
e
nt
,
C
N
N
a
c
hi
e
ve
d
97.2%
a
c
c
ur
a
c
y,
but
w
it
h
e
nha
nc
e
m
e
nt
s
,
a
c
c
ur
a
c
y
r
os
e
to
99.8%
[
27]
.
S
im
il
a
r
im
pr
ove
m
e
nt
s
w
e
r
e
not
e
d
w
it
h
C
S
a
nd
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
on
f
a
c
ia
l
e
m
ot
io
n
r
e
c
ogni
ti
on
(
F
E
R
)
da
ta
s
e
ts
[
28]
.
O
th
e
r
s
tu
di
e
s
a
ppl
ie
d
C
S
to
m
a
gne
ti
c
r
e
s
ona
nc
e
im
a
gi
ng
(
M
R
I
)
im
a
ge
s
[
29
]
,
C
L
A
H
E
to
f
a
c
ia
l
s
ki
n
im
a
ge
s
[
30]
,
a
nd
B
P
D
H
E
to
X
-
r
a
y
im
a
ge
s
[
31]
.
W
he
n
a
ppl
ie
d
to
pa
r
ti
c
ul
a
r
r
e
s
e
a
r
c
h
obj
e
c
ts
,
th
e
C
S
,
C
L
A
H
E
,
a
nd
B
P
D
H
E
m
e
th
ods
a
r
e
th
e
s
up
e
r
io
r
c
ont
r
a
s
t
qua
li
ty
e
nha
nc
e
m
e
nt
m
e
th
od
s
.
H
ow
e
ve
r
,
f
ur
th
e
r
r
e
s
e
a
r
c
h
is
ne
e
de
d
to
de
te
r
m
in
e
w
hi
c
h
m
e
th
od
is
s
ui
ta
bl
e
f
or
ove
r
c
om
in
g
th
e
w
e
a
kne
s
s
e
s
of
th
e
Z
e
r
o
-
D
C
E
m
e
th
od on f
a
c
ia
l
im
a
ge
s
w
it
h br
ig
ht
ne
s
s
pr
obl
e
m
s
.
T
hi
s
s
tu
dy
m
odi
f
ie
s
Z
e
r
o
-
D
C
E
to
c
r
e
a
te
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
a
nd
e
va
lu
a
te
s
C
S
,
C
L
A
H
E
,
a
nd
B
P
D
H
E
c
om
bi
ne
d
w
it
h
Z
e
r
o
-
D
C
E
f
or
f
a
c
e
r
e
c
ogni
ti
on
us
in
g
V
G
G
16
a
nd
R
e
s
N
e
t5
0.
M
e
th
ods
a
r
e
c
om
pa
r
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 4, A
ugus
t
2025
:
3274
-
3286
3276
us
in
g
B
R
I
S
Q
U
E
[
13]
f
or
c
ont
r
a
s
t
qua
li
ty
a
nd
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
T
he
be
s
t
e
nha
n
c
e
m
e
nt
m
e
th
od
ha
s
th
e
lo
w
e
s
t
B
R
I
S
Q
U
E
, a
nd t
he
be
s
t
m
od
e
l
a
c
hi
e
ve
s
t
he
hi
ghe
s
t
a
c
c
u
r
a
c
y.
3.
M
E
T
H
O
D
T
hi
s
s
e
c
ti
on
d
e
s
c
r
ib
e
s
th
e
d
a
ta
s
e
t,
m
e
th
odol
ogy,
a
nd
ke
y
m
odi
f
ic
a
ti
ons
to
Z
e
r
o
-
D
C
E
f
or
L
I
M
E
.
I
t
in
tr
oduc
e
s
th
e
or
ig
in
a
l
Z
e
r
o
-
D
C
E
,
th
e
pr
opos
e
d
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
,
a
nd
tr
a
di
t
io
na
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
te
c
hni
que
s
.
F
in
a
ll
y, i
t
di
s
c
us
s
e
s
t
h
e
f
a
c
e
de
te
c
ti
on a
nd de
e
p l
e
a
r
ni
ng mode
ls
us
e
d f
or
e
va
lu
a
ti
on.
3.1.
D
at
as
e
t
s
T
hi
s
s
tu
dy
ut
il
iz
e
s
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
t
a
ba
s
e
B
[
1
6]
a
nd
F
E
R
E
T
[
17]
,
[
18]
f
or
f
a
c
e
r
e
c
ogni
ti
on
a
nd
th
e
L
O
L
da
ta
s
e
t
[
19]
to
tr
a
in
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
.
T
he
e
xt
e
nde
d
Y
a
le
f
a
c
e
d
a
ta
ba
s
e
B
,
w
it
h
2,414
im
a
ge
s
of
38
s
ubj
e
c
ts
,
w
a
s
c
hos
e
n
f
or
it
s
c
ont
r
ol
le
d
li
ght
in
g
va
r
ia
ti
ons
,
w
hi
le
th
e
F
E
R
E
T
da
ta
s
e
t,
c
ont
a
in
in
g
14,051
im
a
ge
s
of
1,204
s
ubj
e
c
t
s
,
pr
ovi
de
s
di
ve
r
s
e
pos
e
s
a
nd
e
xpr
e
s
s
io
ns
.
T
he
L
O
L
da
ta
s
e
t
w
a
s
s
e
le
c
te
d t
o m
a
in
ta
in
c
on
s
is
te
nc
y
w
it
h t
he
or
ig
in
a
l
Z
e
r
o
-
D
C
E
t
r
a
in
in
g s
e
tu
p, e
ns
ur
in
g a
f
a
ir
c
om
pa
r
is
on.
S
om
e
F
E
R
E
T
c
la
s
s
e
s
c
ont
a
in
onl
y t
w
o i
m
a
ge
s
, m
a
ki
ng da
ta
s
e
t
s
pl
it
ti
ng i
m
pr
a
c
ti
c
a
l.
T
o a
ddr
e
s
s
t
hi
s
,
im
a
ge
s
w
e
r
e
dupl
ic
a
te
d
to
c
r
e
a
te
te
n
pe
r
c
la
s
s
,
in
c
r
e
a
s
in
g
th
e
da
ta
s
e
t
s
iz
e
to
14,291
im
a
ge
s
.
T
hi
s
dupl
ic
a
ti
on
di
f
f
e
r
s
f
r
om
a
ugm
e
nt
a
ti
on, a
s
dupli
c
a
te
d i
m
a
ge
s
r
e
m
a
in
unc
ha
n
ge
d but c
a
n s
ti
ll
unde
r
go a
ugm
e
nt
a
ti
on dur
in
g
tr
a
in
in
g.
F
a
c
e
de
te
c
ti
on
us
in
g
m
ul
ti
-
ta
s
k
c
a
s
c
a
de
d
c
onvolut
io
na
l
ne
twor
ks
(
M
T
C
N
N
)
[
32]
w
a
s
a
ppl
ie
d
to
F
E
R
E
T
i
m
a
ge
s
be
f
or
e
t
r
a
in
in
g t
he
f
a
c
e
r
e
c
ogni
ti
on mode
l.
S
a
m
pl
e
s
of
t
he
e
xt
e
nde
d
Y
a
le
f
a
c
e
d
a
ta
ba
s
e
B
a
nd
F
E
R
E
T
da
ta
s
e
ts
a
r
e
s
how
n i
n F
ig
ur
e
s
1 a
nd 2.
F
ig
ur
e
1. T
he
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
b
a
s
e
B
s
a
m
pl
e
s
[
16]
F
ig
ur
e
2. F
E
R
E
T
s
a
m
pl
e
s
[
17]
, [
18]
T
he
L
O
L
da
t
a
s
e
t
c
on
s
i
s
ts
of
500
lo
w
-
br
ig
ht
a
n
d
5
00
nor
m
a
l
-
br
ig
ht
im
a
g
e
s
.
S
in
c
e
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
r
e
ly
on
de
e
p
c
ur
ve
e
s
ti
m
a
ti
on
n
e
twor
k
(
D
C
E
-
N
e
t)
,
a
n
un
s
up
e
r
vi
s
e
d
l
e
a
r
n
in
g
m
e
th
od
,
tr
a
in
in
g
do
e
s
n
ot
r
e
qui
r
e
s
t
a
nda
r
d
br
i
ght
n
e
s
s
im
a
g
e
s
.
M
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
a
s
tr
a
in
e
d
u
s
in
g
485
lo
w
-
br
ig
ht
n
e
s
s
i
m
a
g
e
s
f
or
t
r
a
in
in
g a
nd v
a
li
da
t
io
n. F
i
gur
e
3 s
ho
w
s
a
s
a
m
pl
e
of
t
h
e
l
o
w
-
br
ig
ht
n
e
s
s
L
O
L
da
t
a
s
e
t.
F
ig
ur
e
3. L
O
L
s
a
m
pl
e
s
[
19]
3.2.
M
e
t
h
od
d
e
s
ig
n
T
hi
s
r
e
s
e
a
r
c
h
m
e
th
od
c
ons
i
s
ts
of
th
r
e
e
s
ta
ge
s
:
m
odi
f
yi
ng
Z
e
r
o
-
D
C
E
,
pr
e
pr
oc
e
s
s
in
g,
a
nd
f
a
c
e
r
e
c
ogni
ti
on
m
ode
li
ng.
I
n
m
odi
f
yi
ng
Z
e
r
o
-
D
C
E
,
w
e
r
e
m
ove
a
lo
s
s
f
unc
ti
on
w
it
h
m
in
im
a
l
im
pa
c
t
a
nd
a
dj
us
t
th
e
D
C
E
-
N
e
t
a
r
c
hi
te
c
tu
r
e
by
m
odi
f
yi
ng
hype
r
pa
r
a
m
e
te
r
s
or
l
a
ye
r
s
,
in
s
pi
r
e
d
by
pr
io
r
s
tu
di
e
s
[
33]
s
how
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
odi
fi
e
d z
e
r
o
-
r
e
fe
r
e
nc
e
d
e
e
p c
u
r
v
e
e
s
ti
m
at
io
n f
or
c
ont
r
a
s
t
qua
li
ty
…
(
M
uhamm
ad K
ahf
i
A
ul
ia
)
3277
th
a
t
c
e
r
ta
in
lo
s
s
f
unc
ti
ons
c
ont
r
ib
ut
e
li
tt
le
to
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
.
F
ig
ur
e
4
il
lu
s
tr
a
te
s
th
is
pr
oc
e
s
s
,
w
hi
c
h
in
vol
ve
s
s
e
le
c
ti
ng
a
n
a
ppr
oa
c
h,
r
e
m
ovi
ng
a
lo
s
s
f
unc
ti
on,
m
odi
f
yi
ng
th
e
D
C
E
-
N
e
t,
a
nd
e
va
lu
a
ti
ng
r
e
s
ul
ts
to
f
in
a
li
z
e
t
he
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
.
F
ig
ur
e
4. M
odi
f
yi
ng Z
e
r
o
-
D
C
E
s
ta
ge
T
he
pr
e
pr
oc
e
s
s
in
g
s
ta
g
e
in
vol
ve
s
a
ppl
yi
ng
Z
e
r
o
-
D
C
E
or
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
,
c
onve
r
ti
ng
im
a
ge
s
to
gr
a
ys
c
a
le
,
a
nd
opt
io
na
ll
y
e
nha
nc
in
g
c
ont
r
a
s
t
u
s
in
g
C
S
,
C
L
A
H
E
,
or
B
P
D
H
E
,
r
e
s
ul
ti
ng
in
s
ix
da
ta
s
e
t
va
r
ia
ti
ons
.
B
R
I
S
Q
U
E
s
c
or
e
s
a
r
e
c
a
lc
ul
a
te
d
on
a
s
ub
s
e
t
of
10
i
m
a
ge
s
f
r
om
th
e
e
xt
e
nde
d
Y
a
l
e
f
a
c
e
da
ta
b
a
s
e
B
a
nd
F
E
R
E
T
da
ta
s
e
ts
to
a
s
s
e
s
s
e
nha
nc
e
m
e
nt
e
f
f
e
c
ti
ve
ne
s
s
,
f
ol
lo
w
e
d
by
e
va
lu
a
ti
ons
on
th
e
f
ul
l
da
ta
s
e
ts
f
or
va
li
da
ti
on. F
ig
ur
e
5 outl
in
e
s
t
hi
s
pr
e
pr
oc
e
s
s
in
g s
t
a
ge
be
f
or
e
m
o
vi
ng t
o f
a
c
e
r
e
c
ogni
ti
on mode
li
ng.
I
n
th
e
f
a
c
e
r
e
c
ogni
ti
on
m
ode
li
ng
s
ta
ge
,
th
e
da
t
a
s
e
t
s
a
r
e
s
pl
it
in
to
60%
tr
a
in
in
g,
20%
va
li
da
ti
on,
a
nd
20%
te
s
t
da
ta
,
w
it
h
a
ugm
e
nt
a
ti
on
te
c
hni
que
s
a
ppl
ie
d.
V
G
G
16
a
nd
R
e
s
N
e
t5
0
m
ode
ls
a
r
e
tr
a
in
e
d
f
or
20
e
poc
hs
us
in
g
A
da
m
opt
im
iz
a
ti
on
w
it
h
a
0.001
le
a
r
ni
ng
r
a
t
e
a
nd
de
c
a
y
r
a
te
s
of
0.9
a
nd
0.999.
F
ig
u
r
e
6
de
ta
il
s
th
is
s
ta
ge
,
w
he
r
e
m
ode
ls
a
r
e
te
s
te
d
f
or
a
c
c
ur
a
c
y
to
de
te
r
m
in
e
th
e
be
s
t
a
ppr
oa
c
h
f
o
r
ha
ndl
in
g
br
ig
ht
ne
s
s
va
r
ia
ti
ons
i
n f
a
c
e
r
e
c
ogni
ti
on.
F
ig
ur
e
5. P
r
e
pr
oc
e
s
s
in
g s
ta
ge
F
ig
ur
e
6. F
a
c
e
r
e
c
ogni
ti
on mode
li
ng s
ta
ge
T
o
e
va
lu
a
te
th
e
im
pa
c
t
of
our
m
odi
f
ic
a
ti
ons
,
w
e
us
e
B
R
I
S
Q
U
E
f
or
im
a
ge
qua
li
ty
a
s
s
e
s
s
m
e
nt
a
nd
c
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y f
or
m
ode
l
pe
r
f
or
m
a
nc
e
. B
R
I
S
Q
U
E
, a
no
-
r
e
f
e
r
e
nc
e
m
e
tr
ic
, i
s
e
s
s
e
nt
ia
l
f
or
da
ta
s
e
ts
l
ik
e
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
a
nd
F
E
R
E
T
,
w
hi
c
h
la
c
k
n
or
m
a
l
il
lu
m
in
a
ti
on
r
e
f
e
r
e
nc
e
s
.
B
y
c
om
bi
ni
ng
pe
r
c
e
pt
ua
l
qua
li
ty
a
s
s
e
s
s
m
e
nt
w
it
h
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y,
w
e
e
ns
ur
e
th
a
t
th
e
pr
opos
e
d
m
odi
f
ic
a
ti
ons
e
nha
nc
e
bot
h i
m
a
ge
qua
li
ty
a
nd f
a
c
e
r
e
c
ogni
ti
on pe
r
f
or
m
a
nc
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 4, A
ugus
t
2025
:
3274
-
3286
3278
3.3. Z
e
r
o
-
r
e
f
e
r
e
n
c
e
d
e
e
p
c
u
r
ve
e
s
t
im
at
io
n
Z
e
r
o
-
D
C
E
[
4]
e
nha
nc
e
s
lo
w
-
li
ght
im
a
ge
c
ont
r
a
s
t
by
e
s
ti
m
a
ti
ng
hi
gh
-
or
de
r
c
ur
ve
s
th
a
t
a
dj
us
t
pi
xe
l
-
w
is
e
dyna
m
ic
r
a
nge
,
im
pr
ovi
ng
br
ig
ht
ne
s
s
w
hi
le
pr
e
s
e
r
vi
ng
c
ont
r
a
s
t.
U
nl
ik
e
C
N
N
-
or
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
G
A
N
)
-
ba
s
e
d
m
e
th
ods
,
it
ope
r
a
te
s
w
it
hout
pa
ir
e
d
or
unpa
ir
e
d
tr
a
in
in
g
da
ta
.
T
h
e
e
nha
nc
e
m
e
nt
i
s
dr
iv
e
n
by
th
e
li
ght
-
e
nha
nc
e
m
e
nt
c
ur
ve
(
L
E
-
c
ur
ve
)
,
ba
s
e
d
on
a
qua
dr
a
ti
c
f
unc
ti
on,
w
it
hi
n
th
e
D
C
E
-
N
e
t,
a
s
s
how
n i
n F
ig
ur
e
7.
F
ig
ur
e
7. D
C
E
-
N
e
t
a
r
c
hi
te
c
tu
r
e
[
4]
D
C
E
-
N
e
t
c
ons
is
ts
of
s
e
v
e
n
c
onvolut
io
n
la
ye
r
s
w
it
h
s
ym
m
e
t
r
ic
a
l
s
ki
p
c
onne
c
ti
ons
.
T
h
e
f
ir
s
t
s
ix
la
ye
r
s
us
e
32
c
onvolut
io
n
ke
r
ne
ls
(
3×
3,
s
tr
id
e
1)
w
it
h
R
e
L
U
a
c
ti
va
ti
on,
w
hi
le
th
e
f
in
a
l
la
ye
r
ha
s
24
ke
r
ne
l
s
(
3×
3,
s
tr
id
e
1)
w
it
h
T
a
nh
a
c
ti
va
ti
on,
ge
ne
r
a
ti
ng
24
c
ur
ve
p
a
r
a
m
e
te
r
m
a
ps
f
or
e
ig
ht
it
e
r
a
ti
ons
(
th
r
e
e
pe
r
R
G
B
c
ha
nne
l)
.
A
s
a
n
uns
up
e
r
vi
s
e
d
m
e
th
od,
Z
e
r
o
-
D
C
E
r
e
li
e
s
on
f
our
lo
s
s
f
unc
ti
ons
:
s
pa
ti
a
l
c
ons
i
s
te
nc
y,
e
xpo
s
ur
e
c
ont
r
ol
, c
ol
or
c
ons
ta
nc
y, a
nd i
ll
um
in
a
ti
on s
m
oot
hne
s
s
.
S
pa
ti
a
l
c
ons
is
te
nc
y l
o
s
s
i
s
s
how
n i
n (
1)
.
=
1
∑
∑
(
|
−
|
−
|
−
|
)
2
∈
(
)
=
1
(
1)
W
he
r
e
K
is
th
e
num
b
e
r
of
lo
c
a
l
r
e
gi
ons
,
Ω
(
i)
is
th
e
pi
xe
l
a
dj
a
c
e
nt
to
th
e
c
e
nt
e
r
pi
xe
l,
Y
a
nd
I
a
r
e
th
e
a
ve
r
a
g
e
in
te
ns
it
y
va
lu
e
s
of
th
e
lo
c
a
l
r
e
gi
ons
in
th
e
ups
c
a
le
d
a
nd
or
ig
in
a
l
im
a
ge
,
r
e
s
pe
c
ti
ve
ly
.
E
xpos
ur
e
c
ont
r
ol
lo
s
s
i
s
s
how
n i
n (
2)
.
=
1
∑
|
−
|
=
1
(
2)
W
he
r
e
is
th
e
num
be
r
of
lo
c
a
l
non
-
ove
r
la
ppi
ng
r
e
gi
ons
,
a
nd
is
th
e
good
e
xpos
ur
e
le
v
e
l
(
de
f
a
ul
t
va
lu
e
is
0.6)
.
C
ol
or
c
ons
ta
nc
y l
os
s
i
s
s
how
n i
n (
3)
.
=
∑
(
−
)
2
,
=
{
(
,
)
,
(
,
)
,
(
,
)
}
∀
(
,
)
∈
(
3)
W
he
r
e
is
th
e
a
ve
r
a
ge
in
te
n
s
it
y
va
lu
e
of
c
ha
nne
l
of
th
e
e
nha
nc
e
d
im
a
ge
,
is
th
e
a
ve
r
a
ge
in
te
n
s
it
y
va
lu
e
of
c
ha
nne
l
,
(
,
)
is
th
e
c
ha
nne
l
pa
ir
,
is
r
e
d
c
ha
nne
l,
is
gr
e
e
n
c
ha
nne
l,
a
nd
is
bl
u
e
c
h
a
nne
l.
I
ll
um
in
a
ti
on s
m
oot
hne
s
s
l
os
s
i
s
s
how
n i
n (
4)
.
=
1
∑
∑
(
|
+
|
)
2
,
=
{
,
,
}
∈
=
1
(
4)
W
he
r
e
is
th
e
pi
xe
l
-
w
is
e
a
lp
ha
m
a
p
f
or
one
of
,
or
c
ha
nne
l
a
t
th
e
-
th
it
e
r
a
ti
on,
is
th
e
num
be
r
of
it
e
r
a
ti
ons
,
∇
a
nd
∇
a
r
e
th
e
hor
iz
ont
a
l
a
nd
ve
r
ti
c
a
l
gr
a
di
e
nt
ope
r
a
t
io
ns
.
T
he
s
e
gr
a
di
e
nt
s
m
e
a
s
ur
e
how
th
e
in
te
ns
it
y i
n
c
ha
nge
s
a
c
r
os
s
a
dj
a
c
e
nt
pi
xe
l
s
.
T
he
t
ot
a
l
lo
s
s
of
t
h
e
f
our
l
os
s
f
unc
ti
ons
i
s
de
s
c
r
ib
e
d i
n (
5)
.
=
+
+
+
(
5)
W
he
r
e
a
nd
a
r
e
l
os
s
w
e
ig
ht
s
.
3.
4
.
M
od
if
ie
d
Z
e
r
o
-
r
e
f
e
r
e
n
c
e
d
e
e
p
c
u
r
v
e
e
s
t
i
m
at
io
n
A
m
odi
f
ic
a
ti
on
e
xpe
r
im
e
nt
w
a
s
c
onduc
t
e
d
by
r
e
m
ovi
ng
on
e
lo
s
s
f
unc
ti
on
f
r
om
th
e
f
our
lo
s
s
f
unc
ti
ons
i
n Z
e
r
o
-
D
C
E
.
T
he
r
e
s
ul
ts
of
r
e
m
ovi
ng one
l
os
s
f
unc
ti
on w
e
r
e
t
e
s
te
d on a
n i
m
a
ge
f
r
om
t
he
e
xt
e
nde
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
odi
fi
e
d z
e
r
o
-
r
e
fe
r
e
nc
e
d
e
e
p c
u
r
v
e
e
s
ti
m
at
io
n f
or
c
ont
r
a
s
t
qua
li
ty
…
(
M
uhamm
ad K
ahf
i
A
ul
ia
)
3279
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
.
T
he
r
e
s
ul
ts
of
th
e
lo
s
s
f
unc
ti
on
r
e
m
ova
l
e
xpe
r
im
e
nt
s
a
r
e
s
how
n
in
F
ig
ur
e
8.
F
ig
ur
e
8
il
lu
s
tr
a
te
s
th
e
im
pa
c
t
of
r
e
m
ovi
ng
e
a
c
h
lo
s
s
f
unc
ti
on.
F
ig
ur
e
8(
a
)
is
th
e
in
put
im
a
ge
,
w
hi
le
F
ig
ur
e
8(
b)
s
how
s
Z
e
r
o
-
D
C
E
w
it
h
a
ll
lo
s
s
e
s
.
R
e
m
ovi
ng
s
pa
ti
a
l
c
ons
is
te
n
c
y
lo
s
s
(
)
(
F
ig
ur
e
8(
c
)
)
le
a
ds
to
ove
r
e
xpos
ur
e
,
w
hi
le
r
e
m
ovi
ng e
xpos
ur
e
c
ont
r
ol
lo
s
s
(
)
(
F
ig
ur
e
8(
d
)
)
,
a
f
f
e
c
ts
l
um
in
a
nc
e
c
ont
r
ol
. F
ig
ur
e
8(
e
)
s
how
s
c
ol
or
c
ons
ta
nc
y
lo
s
s
(
)
r
e
m
ova
l,
r
e
duc
in
g
c
ol
or
f
id
e
li
ty
,
a
nd
F
ig
u
r
e
8(
f
)
i
ll
us
tr
a
te
s
how
r
e
m
ovi
ng
i
ll
um
in
a
ti
on
s
m
oot
hne
s
s
lo
s
s
(
)
,
in
tr
oduc
e
s
a
r
ti
f
a
c
ts
.
W
hi
le
a
ll
lo
s
s
e
s
a
r
e
im
por
ta
nt
,
ha
d
m
in
im
a
l
im
pa
c
t
on
br
ig
ht
ne
s
s
a
nd
f
a
c
ia
l
de
ta
il
s
.
R
e
m
ovi
ng
it
in
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
im
pr
ove
d
e
f
f
ic
ie
nc
y
w
it
hout
s
a
c
r
if
ic
in
g
c
ont
r
a
s
t,
us
e
f
ul
f
or
r
e
a
l
-
ti
m
e
f
a
c
e
r
e
c
ogni
ti
on
. T
he
f
or
m
ul
a
f
or
t
ot
a
l
lo
s
s
us
e
d i
n
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
, i
ni
ti
a
ll
y
s
how
n i
n (
5)
, now c
ha
nge
s
t
o (
6)
.
=
+
+
(
6)
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
F
ig
ur
e
8.
T
h
e
r
e
s
ul
ts
of
t
h
e
lo
s
s
f
un
c
ti
on
r
e
m
ov
a
l
e
xp
e
r
im
e
n
ts
of
(
a
)
i
np
ut
im
a
g
e
,
(
b)
Z
e
r
o
-
D
C
E
,
(
c
)
w
it
hou
t
,
(
d)
w
it
hout
e
x
p
,
(
e
)
w
it
hout
co
l
, a
nd
(
f
)
w
it
hout
tv
A
T
he
a
r
c
hi
te
c
tu
r
e
of
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
(
m
odi
f
ie
d
D
C
E
-
N
e
t)
is
not
to
o
di
f
f
e
r
e
nt
f
r
om
th
a
t
of
Z
e
r
o
-
D
C
E
(
D
C
E
-
N
e
t)
.
T
he
m
ode
l
us
e
s
f
e
w
e
r
f
il
te
r
s
(
8
in
s
t
e
a
d
of
32)
in
th
e
c
onvolut
io
n
la
ye
r
.
I
t
a
ls
o
in
c
lu
de
s
a
n
e
xt
r
a
pool
in
g
la
ye
r
(
M
a
xP
ool
2D
)
a
f
te
r
e
a
c
h
c
onc
a
t
e
na
ti
on
a
nd
dr
opout
la
ye
r
.
T
hi
s
pool
in
g
la
ye
r
is
in
s
pi
r
e
d
by
th
e
s
pa
ti
a
l
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
in
Z
e
r
o
-
R
A
D
C
E
[
24]
.
F
ig
ur
e
9
il
lu
s
tr
a
te
s
th
e
a
r
c
hi
te
c
tu
r
e
of
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
, i
.e
.,
m
odi
f
ie
d
D
C
E
-
N
e
t.
F
ig
ur
e
9. M
odi
f
ie
d Z
e
r
o
-
D
C
E
a
r
c
hi
te
c
tu
r
e
I
n
F
ig
ur
e
9,
th
e
gr
a
y
bl
oc
k
is
th
e
c
onvolut
io
n
la
ye
r
,
th
e
r
e
d
bl
oc
k
is
th
e
dr
opout
la
ye
r
,
a
nd
th
e
bl
ue
bl
oc
k
is
th
e
pool
in
g
la
ye
r
.
M
odi
f
ie
d
D
C
E
-
N
e
t
ha
s
7,776
pa
r
a
m
e
te
r
s
,
s
ig
ni
f
ic
a
nt
ly
f
e
w
e
r
th
a
n
D
C
E
-
N
e
t’
s
79,416,
due
to
r
e
duc
e
d
f
il
te
r
s
pe
r
la
y
e
r
.
T
hi
s
m
a
ke
s
it
li
ght
e
r
a
nd
f
a
s
te
r
f
or
im
a
ge
e
nha
nc
e
m
e
nt
.
S
in
c
e
doe
s
not
ha
ve
a
s
ig
ni
f
ic
a
nt
e
f
f
e
c
t
on
th
e
Z
e
r
o
-
D
C
E
ba
s
e
d
on
F
ig
ur
e
10,
a
n
a
tt
e
m
pt
w
a
s
m
a
de
to
e
li
m
in
a
te
in
t
he
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
.
T
he
r
e
s
ul
ts
of
t
he
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
c
om
pa
r
is
on s
ta
ge
w
it
h
a
nd w
it
hout
a
r
e
s
how
n i
n F
ig
ur
e
10.
F
ig
ur
e
10
c
om
pa
r
e
s
Z
e
r
o
-
D
C
E
a
nd
m
od
if
ie
d
Z
e
r
o
-
D
C
E
w
i
t
h
a
nd
w
it
hout
.
A
s
a
ba
s
e
li
ne
,
F
ig
ur
e
10(
a
)
s
how
s
t
he
i
nput
i
m
a
ge
us
e
d f
or
e
nha
nc
e
m
e
nt
.
F
ig
ur
e
s
10(
b)
a
nd 10(
d)
s
how
m
in
im
a
l
d
if
f
e
r
e
nc
e
s
w
he
n
is
pr
e
s
e
nt
.
H
ow
e
ve
r
,
F
ig
ur
e
s
10(
c
)
a
nd
10(
e
)
r
e
ve
a
l
th
a
t
r
e
m
ovi
ng
in
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
ove
r
e
xpos
ur
e
,
c
ont
r
a
di
c
ti
ng
’
s
in
te
nde
d
f
unc
ti
on
in
Z
e
r
o
-
D
C
E
.
T
hi
s
hi
ghl
ig
ht
s
th
e
ne
e
d
f
or
f
ur
th
e
r
a
na
ly
s
is
of
in
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
.
S
in
c
e
m
od
if
ie
d
Z
e
r
o
-
D
C
E
ha
s
be
e
n
f
ound,
it
c
a
n
be
us
e
d
f
or
f
a
c
e
im
a
ge
pr
e
pr
oc
e
s
s
in
g
a
lo
ng
w
it
h
Z
e
r
o
-
D
C
E
,
C
S
,
C
L
A
H
E
,
a
nd
B
P
D
H
E
,
th
e
n
c
om
bi
ne
d
a
nd
e
va
lu
a
te
d us
in
g B
R
I
S
Q
U
E
a
nd e
nde
d w
it
h a
c
om
pa
r
a
ti
ve
s
tu
dy.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 4, A
ugus
t
2025
:
3274
-
3286
3280
(
a
)
(
b)
(
c
)
(
d)
(
e
)
F
ig
ur
e
10.
C
om
pa
r
is
on of
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
it
h a
nd w
it
hout
of
(
a
)
i
nput
i
m
a
ge
,
(
b)
Z
e
r
o
-
D
C
E
,
(
c
)
Z
e
r
o
-
D
C
E
w
it
hout
,
(
d)
M
odi
f
ie
d Z
e
r
o
-
D
C
E
w
it
h
s
p
a
, a
nd
(
e
)
M
odi
f
ie
d Z
e
r
o
-
D
C
E
w
it
hout
s
p
a
3.5. T
r
ad
it
io
n
al
c
on
t
r
as
t
q
u
al
i
t
y e
n
h
an
c
e
m
e
n
t
m
e
t
h
od
s
T
hi
s
s
tu
dy
us
e
d
th
r
e
e
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
:
C
S
,
C
L
A
H
E
,
a
nd
B
P
D
H
E
.
C
S
s
tr
e
tc
he
s
i
m
a
ge
c
ont
r
a
s
t
by e
xpa
ndi
ng i
nt
e
ns
it
y va
lu
e
s
be
twe
e
n
t
he
m
in
im
um
a
nd ma
xi
m
um
pi
xe
l
li
m
i
ts
[
28]
,
ty
pi
c
a
ll
y
f
or
8
-
bi
t
im
a
ge
s
(
0
-
255)
.
C
L
A
H
E
r
e
duc
e
s
noi
s
e
a
nd
a
r
ti
f
a
c
ts
by
di
vi
di
ng
th
e
im
a
ge
in
to
s
ub
-
bl
oc
ks
,
c
om
put
in
g
hi
s
to
gr
a
m
s
,
a
nd
a
ppl
yi
ng
a
tr
a
n
s
f
or
m
a
ti
on
w
it
h
a
c
li
p
bounda
r
y
pa
r
a
m
e
te
r
[
34]
.
T
he
m
odi
f
ie
d
hi
s
to
gr
a
m
s
a
r
e
in
te
r
pol
a
te
d
to
a
dj
us
t
pi
xe
l
in
te
n
s
it
y.
B
P
D
H
E
,
i
nt
r
oduc
e
d
by
I
br
a
hi
m
a
nd
K
ong
in
2007
[
12]
,
nor
m
a
li
z
e
s
c
ont
r
a
s
t
w
hi
le
pr
e
s
e
r
vi
ng
a
ve
r
a
ge
in
t
e
ns
it
y.
I
t
f
ol
lo
w
s
f
iv
e
s
te
p
s
:
hi
s
to
gr
a
m
s
m
oot
hi
ng,
de
te
c
ti
ng
lo
c
a
l
m
a
xi
m
a
,
m
a
ppi
ng
pa
r
ti
ti
ons
,
e
qua
li
z
in
g
pa
r
ti
ti
ons
,
a
n
d
nor
m
a
li
z
in
g
c
ont
r
a
s
t.
B
P
D
H
E
r
e
qui
r
e
s
no
pa
r
a
m
e
te
r
t
uni
ng, i
nt
r
oduc
e
s
m
in
im
a
l
a
r
ti
f
a
c
ts
, a
nd i
s
s
ui
ta
bl
e
f
or
r
e
a
l
-
ti
m
e
s
ys
te
m
s
.
3.
6
.
F
ac
e
d
e
t
e
c
t
io
n
T
he
F
E
R
E
T
da
ta
s
e
t
c
ont
a
in
s
non
-
f
a
c
e
c
ha
r
a
c
te
r
is
ti
c
s
,
r
e
qui
r
in
g
f
a
c
e
de
te
c
ti
on
be
f
or
e
tr
a
in
in
g
a
r
e
c
ogni
ti
on mode
l.
M
T
C
N
N
[
32]
w
a
s
us
e
d f
or
t
hi
s
t
a
s
k, a
s
i
t
a
c
c
ur
a
te
ly
de
te
c
ts
f
a
c
e
s
a
nd f
iv
e
ke
y l
a
ndm
a
r
ks
.
T
hi
s
s
te
p
e
ns
ur
e
s
hi
gh
-
qua
li
ty
f
a
c
e
lo
c
a
li
z
a
ti
on
b
e
f
or
e
a
ppl
yi
ng
V
G
G
16
a
nd
R
e
s
N
e
t5
0,
im
pr
ovi
ng
r
e
c
ogni
ti
on
a
c
c
ur
a
c
y
unde
r
va
r
ie
d
li
ght
in
g.
H
ow
e
ve
r
,
s
om
e
i
m
a
ge
s
w
e
r
e
not
de
te
c
te
d,
r
e
duc
in
g
th
e
da
ta
s
e
t
f
r
om
14,291 to 13,783 i
m
a
ge
s
. F
ig
ur
e
11 s
how
s
t
he
f
a
c
e
de
te
c
ti
on r
e
s
ul
t
f
or
t
he
f
ir
s
t
s
ubj
e
c
t.
F
ig
ur
e
11.
F
a
c
e
de
te
c
ti
on r
e
s
ul
ts
on t
he
f
ir
s
t
s
ubj
e
c
t
of
t
he
F
E
R
E
T
da
ta
s
e
t
3.
7
.
D
e
e
p
l
e
ar
n
in
g m
od
e
ls
T
he
de
e
p
le
a
r
ni
ng
m
ode
ls
us
e
d
in
th
is
s
tu
dy
a
r
e
V
G
G
16
a
nd
R
e
s
N
e
t5
0.
V
G
G
16
[
14
]
is
a
C
N
N
w
it
h
13
c
onvolut
io
n
la
ye
r
s
,
f
iv
e
pool
in
g
la
ye
r
s
,
a
nd
th
r
e
e
f
ul
ly
c
onne
c
te
d
la
ye
r
s
,
tr
a
in
e
d
on
I
m
a
ge
N
e
t
w
it
h
224×
224
R
G
B
im
a
ge
s
.
I
t
us
e
s
f
iv
e
m
a
x
-
pool
in
g
la
ye
r
s
a
n
d
e
nds
w
it
h
a
S
of
tM
a
x
a
c
ti
va
ti
on
f
unc
ti
on.
R
e
s
N
e
t5
0
[
15]
is
a
d
e
e
pe
r
C
N
N
w
it
h
50
c
onvolut
io
n
la
ye
r
s
,
of
f
e
r
in
g
lo
w
e
r
e
r
r
or
r
a
te
s
a
nd
f
a
s
te
r
c
la
s
s
if
ic
a
ti
on
th
a
n
V
G
G
16.
B
ot
h
m
ode
ls
a
r
e
w
id
e
ly
us
e
d
in
f
a
c
ia
l
r
e
c
ogni
ti
on,
w
it
h
V
G
G
16
e
xc
e
ll
in
g
i
n
a
c
c
ur
a
c
y
a
nd
R
e
s
N
e
t5
0
be
ne
f
it
in
g
f
r
om
r
e
s
id
ua
l
c
onne
c
ti
ons
f
or
e
f
f
ic
ie
nc
y.
T
hi
s
pr
ovi
de
s
a
ba
la
nc
e
d
c
om
pa
r
is
on of
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
e
f
f
e
c
ts
on dif
f
e
r
e
nt
ne
twor
k de
pt
hs
.
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
e
c
ti
on
e
va
lu
a
t
e
s
th
e
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
m
e
th
od,
a
na
ly
z
in
g
it
s
be
ha
vi
or
a
nd
c
ha
r
a
c
te
r
is
ti
c
s
.
I
t
di
s
c
us
s
e
s
th
e
r
e
s
ul
ts
of
a
ppl
yi
ng
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
a
nd
ot
he
r
m
e
th
ods
f
or
im
a
ge
e
nha
nc
e
m
e
nt
.
F
in
a
ll
y,
it
pr
e
s
e
nt
s
t
he
f
a
c
e
r
e
c
ogni
ti
on mode
li
ng r
e
s
ul
t
s
t
o a
s
s
e
s
s
t
he
m
e
th
od’
s
i
m
pa
c
t.
4
.1.
M
od
if
ie
d
Z
e
r
o
-
r
e
f
e
r
e
n
c
e
d
e
e
p
c
u
r
v
e
e
s
t
i
m
at
io
n
an
al
ys
is
T
e
n
im
a
g
e
s
w
e
r
e
r
a
ndoml
y
s
e
le
c
t
e
d,
f
iv
e
f
r
om
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
a
nd
f
iv
e
f
r
om
F
E
R
E
T
.
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
e
r
e
a
ppl
ie
d,
w
i
th
r
e
s
ul
ts
s
how
n
in
F
ig
ur
e
s
12
a
nd
13,
w
he
r
e
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
p
r
oduc
e
d
c
le
a
r
e
r
im
a
ge
s
.
T
he
a
na
ly
s
is
e
xa
m
in
e
s
’
s
c
ont
r
a
di
c
to
r
y
e
f
f
e
c
ts
,
c
ons
id
e
r
in
g
pool
in
g,
dr
opout,
r
e
duc
e
d
f
il
te
r
s
(
8
vs
.
32)
,
a
n
d
r
e
m
ova
l,
w
it
h
r
e
s
ul
ts
s
um
m
a
r
iz
e
d
in
T
a
bl
e
1
, w
he
r
e
"
W
/o
"
s
ta
nd
s
f
or
"
w
it
hout
."
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
odi
fi
e
d z
e
r
o
-
r
e
fe
r
e
nc
e
d
e
e
p c
u
r
v
e
e
s
ti
m
at
io
n f
or
c
ont
r
a
s
t
qua
li
ty
…
(
M
uhamm
ad K
ahf
i
A
ul
ia
)
3281
F
ig
ur
e
12. Z
e
r
o
-
D
C
E
r
e
s
ul
ts
F
ig
ur
e
13. M
odi
f
ie
d
Z
e
r
o
-
D
C
E
r
e
s
ul
ts
T
a
bl
e
1
s
how
s
th
a
t
w
hi
le
bot
h
th
e
pool
in
g
la
ye
r
a
nd
ut
il
iz
e
s
pa
ti
a
l
c
ha
r
a
c
te
r
is
ti
c
s
,
th
e
num
be
r
of
c
onvolut
io
n f
il
te
r
s
ha
s
t
he
m
os
t
s
ig
ni
f
ic
a
nt
i
m
pa
c
t
on Z
e
r
o
-
D
C
E
r
e
s
ul
ts
. I
n
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
, pooli
ng
i
s
m
or
e
e
f
f
e
c
ti
ve
th
a
n
in
r
e
duc
in
g
ove
r
e
xpos
ur
e
,
a
ll
ow
in
g
lu
m
in
a
nc
e
a
dj
us
tm
e
nt
w
it
hout
e
qua
li
z
in
g
lo
c
a
l
r
e
gi
ons
. T
hi
s
m
odi
f
ic
a
ti
on i
m
pr
ove
s
i
m
a
ge
c
l
a
r
it
y a
nd r
e
duc
e
s
c
om
put
a
ti
ona
l
lo
a
d.
T
a
bl
e
1. O
bs
e
r
va
ti
on r
e
s
ul
ts
ba
s
e
d on the
numbe
r
of
f
il
te
r
s
O
bs
e
r
va
t
i
on
P
ool
i
ng +
w
/
o
P
ool
i
ng +
w
/
o
W
/
o pool
i
ng +
W
/
o pool
i
ng +
w
/
o
U
s
i
ng 8 f
i
l
t
e
r
s
U
s
i
ng 32 f
i
l
t
e
r
s
T
h
e
s
e
c
o
nd
f
o
c
u
s
of
a
n
a
l
y
s
i
s
i
s
t
h
e
e
f
f
e
c
t
of
t
he
dr
o
pou
t
la
ye
r
i
n
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
,
w
h
e
r
e
t
h
e
m
od
e
l
us
e
s
8
f
il
te
r
s
,
poo
li
n
g,
a
nd
n
o
.
T
h
e
dr
o
po
ut
la
y
e
r
he
lp
s
r
e
d
uc
e
ov
e
r
f
i
tt
i
ng
b
y
r
a
n
dom
ly
s
e
tt
i
ng
s
om
e
e
l
e
m
e
nt
v
a
l
ue
s
t
o
z
e
r
o
in
f
e
a
tu
r
e
m
a
pp
in
g,
a
f
f
e
c
t
in
g
th
e
lo
s
s
f
unc
ti
o
n
c
a
l
c
u
la
ti
o
n.
T
a
bl
e
2
s
h
ow
s
t
ha
t
w
h
il
e
dr
op
out
s
lo
w
s
t
ot
a
l
lo
s
s
c
on
v
e
r
g
e
n
c
e
,
it
a
c
c
e
l
e
r
a
t
e
s
c
on
ve
r
ge
nc
e
in
s
p
e
c
if
i
c
l
os
s
c
o
m
p
on
e
nt
s
,
s
u
c
h a
s
.
T
he
r
e
s
ul
ts
of
t
h
e
m
od
if
i
e
d
Z
e
r
o
-
D
C
E
out
pu
t
im
a
ge
w
i
th
a
nd
w
it
ho
ut
t
h
e
dr
op
ou
t
l
a
y
e
r
c
a
n
b
e
s
e
e
n
i
n F
i
gur
e
14.
T
a
bl
e
2.
O
bs
e
r
va
ti
on
r
e
s
ul
ts
ba
s
e
d on the
numbe
r
of
f
il
te
r
s
D
r
opou
t
l
a
ye
r
s
E
xi
s
t
N
ot
E
xi
s
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 4, A
ugus
t
2025
:
3274
-
3286
3282
F
ig
ur
e
14
c
om
pa
r
e
s
th
e
out
put
of
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
it
h
a
nd
w
it
hout
th
e
dr
opout
la
ye
r
.
F
ig
ur
e
14(
a
)
s
how
s
t
he
r
e
s
ul
t
w
it
h d
r
opout, whil
e
F
ig
ur
e
14(
b
)
s
how
s
t
he
r
e
s
ul
t
w
it
hout
dr
opout. T
he
dr
opout
la
ye
r
s
ig
ni
f
ic
a
nt
ly
im
pa
c
ts
th
e
m
ode
l
by
s
e
tt
in
g
s
om
e
e
le
m
e
nt
va
lu
e
s
to
z
e
r
o
in
f
e
a
tu
r
e
m
a
ppi
ng,
pr
e
ve
nt
in
g
ove
r
-
br
ig
ht
e
ni
ng a
nd e
nha
nc
in
g i
m
a
ge
qua
li
ty
. T
he
pr
e
s
e
nc
e
of
dr
opout a
m
pl
if
ie
s
t
he
a
bs
e
nc
e
of
, a
f
f
e
c
ts
th
e
be
ha
vi
or
of
c
onvolut
io
n
f
il
te
r
s
,
a
nd
im
pr
ove
s
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
pool
in
g.
T
he
s
e
m
odi
f
ic
a
ti
ons
c
ol
le
c
ti
ve
ly
c
ont
r
ib
ut
e
to
r
e
duc
in
g
ove
r
e
xpos
ur
e
a
nd
e
nha
nc
in
g
lu
m
in
a
nc
e
a
dj
us
tm
e
nt
.
A
s
a
r
e
s
ul
t,
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
a
c
hi
e
ve
s
a
m
or
e
ba
la
nc
e
d
im
a
ge
e
nh
a
nc
e
m
e
nt
pr
oc
e
s
s
.
P
r
oc
e
s
s
in
g
s
pe
e
d
t
e
s
ts
w
e
r
e
c
ondu
c
te
d
on
2,414
im
a
ge
s
f
r
om
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
us
in
g
a
n
I
nt
e
l
C
or
e
i5
-
8250U
la
pt
op
w
it
hout
a
G
P
U
.
T
a
bl
e
3
s
how
s
th
a
t
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
p
r
oc
e
s
s
e
s
im
a
ge
s
f
our
ti
m
e
s
f
a
s
te
r
th
a
n
Z
e
r
o
-
D
C
E
.
T
hi
s
r
e
s
ul
t
hi
ghl
ig
ht
s
th
e
e
f
f
ic
ie
nc
y
of
th
e
pr
opos
e
d
m
e
th
od
in
r
e
a
l
-
w
or
ld
a
ppl
ic
a
ti
ons
.
T
h
e
ne
xt
s
t
e
p
is
to
c
om
bi
ne
th
i
s
a
ppr
oa
c
h
w
it
h
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
a
nd
pe
r
f
or
m
a
qua
nt
it
a
ti
ve
e
va
lu
a
ti
on
us
in
g
B
R
I
S
Q
U
E
. T
he
gr
e
e
n
-
c
ol
or
e
d c
e
ll
s
i
ndi
c
a
te
t
he
be
s
t.
(
a
)
(
b)
F
ig
ur
e
14. C
om
pa
r
is
on of
m
odi
f
ie
d Z
e
r
o
-
D
C
E
of
(
a
)
w
it
h d
r
op
out
l
a
ye
r
a
nd (
b)
w
it
hout
dr
opout l
a
ye
r
T
a
bl
e
3
.
T
im
e
c
om
pa
r
is
on r
e
s
ul
ts
of
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
M
e
t
hods
T
ot
a
l
t
i
m
e
(
s
)
A
ve
r
a
ge
t
i
m
e
(
s
)
Z
e
r
o
-
D
C
E
767.3902
0.3179
M
odi
f
i
e
d Z
e
r
o
-
D
C
E
206.8746
0.0857
4
.2.
P
r
e
p
r
oc
e
s
s
in
g s
t
age
r
e
s
u
lt
s
T
he
c
om
bi
na
ti
on
pr
oc
e
s
s
w
a
s
f
ir
s
t
te
s
te
d
on
te
n
s
e
le
c
te
d
im
a
ge
s
us
in
g
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
w
it
h
C
S
,
C
L
A
H
E
,
a
nd
B
P
D
H
E
.
E
va
lu
a
ti
on
w
a
s
c
o
nduc
te
d
w
it
h
a
nd
w
it
hout
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
,
a
nd
th
e
r
e
s
ul
t
s
f
or
th
e
f
ir
s
t
s
a
m
pl
e
a
r
e
s
how
n
in
T
a
bl
e
4,
w
it
h
th
e
be
s
t
r
e
s
ul
ts
hi
ghl
ig
ht
e
d
in
gr
e
e
n.
T
o
c
onf
ir
m
r
obus
tn
e
s
s
,
a
ll
im
a
ge
s
f
r
om
t
he
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
a
nd
F
E
R
E
T
da
ta
s
e
ts
w
e
r
e
t
e
s
te
d.
T
a
bl
e
4. B
R
I
S
Q
U
E
e
va
lu
a
ti
on r
e
s
ul
ts
of
f
ir
s
t
s
a
m
pl
e
I
nput
Z
e
r
o
-
D
C
E
Z
e
r
o
-
D
C
E
+
CS
Z
e
r
o
-
D
C
E
+
C
L
A
H
E
Z
e
r
o
-
D
C
E
+
B
P
D
H
E
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
+
CS
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
+
C
L
A
H
E
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
+
B
P
D
HE
S
c
or
e
57.8683
48.4495
43.6969
37.5207
37.5892
35.5319
32.8356
31.7629
T
he
a
ve
r
a
ge
B
R
I
S
Q
U
E
va
lu
e
w
a
s
c
a
lc
ul
a
te
d f
or
e
a
c
h i
m
a
ge
a
n
d c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
od, with
a
nd
w
it
hout
c
om
bi
na
ti
on.
T
a
bl
e
5
pr
e
s
e
nt
s
th
e
a
ve
r
a
g
e
B
R
I
S
Q
U
E
s
c
or
e
s
f
or
th
e
te
n
s
a
m
pl
e
im
a
ge
s
,
he
lp
in
g
to
de
te
r
m
in
e
th
e
m
os
t
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
f
or
m
in
im
iz
in
g
s
c
o
r
e
s
w
hi
le
pr
e
s
e
r
vi
ng
f
a
c
ia
l
de
ta
il
s
.
F
r
om
th
e
r
e
s
ul
ts
,
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
+
B
P
D
H
E
a
c
hi
e
ve
d
th
e
be
s
t
pe
r
f
o
r
m
a
nc
e
w
it
h
a
n
a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
of
16.0177, while B
P
D
H
E
out
pe
r
f
or
m
e
d C
S
a
nd C
L
A
H
E
w
he
n a
ppl
ie
d t
o Z
e
r
o
-
D
C
E
.
T
a
bl
e
5. A
ve
r
a
ge
B
R
I
S
Q
U
E
e
va
lu
a
ti
on r
e
s
ul
ts
of
10 s
a
m
pl
e
s
M
e
t
hods
W
/
o c
om
bi
na
t
i
on
CS
C
L
A
H
E
B
P
D
H
E
Z
e
r
o
-
D
C
E
43
.
6404
36
.
6863
27
.
9842
23
.
7520
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
29
.
3021
26
.
5361
19
.
7596
16
.
0177
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
odi
fi
e
d z
e
r
o
-
r
e
fe
r
e
nc
e
d
e
e
p c
u
r
v
e
e
s
ti
m
at
io
n f
or
c
ont
r
a
s
t
qua
li
ty
…
(
M
uhamm
ad K
ahf
i
A
ul
ia
)
3283
W
it
hout
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
,
bot
h
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
y
ie
ld
e
d
lo
w
e
r
s
c
or
e
s
,
de
m
on
s
tr
a
ti
ng
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
c
onve
nt
io
na
l
m
e
th
ods
.
T
h
e
c
om
bi
na
ti
on
of
th
e
s
e
m
e
th
ods
c
ons
is
te
nt
ly
im
pr
ove
d
B
R
I
S
Q
U
E
s
c
or
e
s
c
om
pa
r
e
d
to
us
in
g
t
he
m
a
lo
ne
.
A
ddi
ti
ona
ll
y,
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
e
xhi
bi
te
d
s
upe
r
io
r
pr
oc
e
s
s
in
g
e
f
f
ic
ie
nc
y
a
nd
be
tt
e
r
c
ont
r
a
s
t
qua
li
ty
th
a
n
Z
e
r
o
-
D
C
E
.
T
he
ne
xt
s
te
p
is
c
a
lc
ul
a
ti
ng
th
e
a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
f
or
e
a
c
h
im
a
g
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
od
a
nd
da
ta
s
e
t.
F
ur
th
e
r
m
or
e
,
th
e
a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
f
or
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
da
ta
s
e
t
c
a
n
be
s
e
e
n
in
T
a
bl
e
6.
T
he
gr
e
e
n
-
c
ol
or
e
d c
e
ll
i
ndi
c
a
te
s
t
he
be
s
t.
F
or
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
ba
s
e
B
da
ta
s
e
t,
T
a
bl
e
6
s
h
ow
s
th
a
t
Z
e
r
o
-
D
C
E
c
om
bi
ne
d
w
it
h
C
L
A
H
E
a
c
hi
e
ve
d
th
e
be
s
t
B
R
I
S
Q
U
E
s
c
or
e
of
18.1781.
T
he
us
e
of
C
L
A
H
E
a
nd
B
P
D
H
E
s
ig
ni
f
ic
a
nt
ly
im
pr
ove
d
r
e
s
ul
ts
,
pa
r
ti
c
ul
a
r
ly
f
or
L
O
L
da
ta
s
e
ts
.
M
odi
f
ie
d
Z
e
r
o
-
D
C
E
a
ls
o
pe
r
f
or
m
e
d
w
e
ll
,
e
nha
nc
in
g
out
c
om
e
s
c
om
pa
r
e
d
to
Z
e
r
o
-
D
C
E
a
lo
ne
.
T
he
s
e
f
in
di
ngs
c
onf
ir
m
th
a
t
tr
a
di
ti
ona
l
c
ont
r
a
s
t
m
e
th
ods
a
r
e
e
s
s
e
nt
ia
l
f
or
im
pr
ovi
ng
B
R
I
S
Q
U
E
s
c
or
e
s
.
T
a
bl
e
7
pr
e
s
e
nt
s
t
he
a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
s
f
or
th
e
F
E
R
E
T
da
ta
s
e
t.
T
he
gr
e
e
n
-
c
ol
or
e
d c
e
ll
i
ndi
c
a
te
s
t
he
be
s
t.
T
a
bl
e
6. A
ve
r
a
ge
B
R
I
S
Q
U
E
e
va
lu
a
ti
on r
e
s
ul
ts
of
t
he
e
xt
e
nd
e
d
Y
a
le
f
a
c
e
da
ta
b
a
s
e
B
M
e
t
hods
W
/
o c
om
bi
na
t
i
on
CS
C
L
A
H
E
B
P
D
H
E
Z
e
r
o
-
D
C
E
30.8752
29.8683
18.1781
22.2609
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
28.1670
27.9141
22.2711
21.7172
T
a
bl
e
7. A
ve
r
a
ge
B
R
I
S
Q
U
E
e
va
lu
a
ti
on r
e
s
ul
ts
of
F
E
R
E
T
M
e
t
hods
W
/
o c
om
bi
na
t
i
on
CS
C
L
A
H
E
B
P
D
H
E
Z
e
r
o
-
D
C
E
43.2518
18.5525
12.1491
9.3183
M
odi
f
i
e
d
Z
e
r
o
-
D
C
E
22.0140
10.3947
10.5732
7.2470
T
a
bl
e
7
pr
e
s
e
nt
s
th
e
a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
s
f
or
th
e
F
E
R
E
T
da
ta
s
e
t,
w
h
e
r
e
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
+
B
P
D
H
E
a
c
hi
e
ve
d t
he
be
s
t
r
e
s
ul
t
w
it
h a
n a
ve
r
a
ge
B
R
I
S
Q
U
E
s
c
or
e
of
7.2470. T
he
e
f
f
e
c
ti
ve
ne
s
s
of
B
P
D
H
E
in
im
pr
ovi
ng
bot
h
Z
e
r
o
-
D
C
E
a
nd
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
is
e
vi
de
nt
,
a
s
it
c
ons
i
s
te
nt
ly
pr
oduc
e
d
s
c
or
e
s
unde
r
10.
M
odi
f
ie
d
Z
e
r
o
-
D
C
E
pe
r
f
or
m
e
d
w
e
ll
on
th
e
F
E
R
E
T
da
ta
s
e
t,
w
hi
c
h
h
a
s
nor
m
a
l
or
s
li
ght
ly
da
r
k
br
ig
ht
ne
s
s
le
ve
ls
,
w
he
r
e
a
s
Z
e
r
o
-
D
C
E
of
te
n
c
a
us
e
d
ove
r
e
xpo
s
ur
e
.
A
s
s
how
n
in
T
a
bl
e
s
5
a
nd
6,
c
om
bi
ni
ng
tr
a
di
ti
ona
l
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
yi
e
ld
s
be
tt
e
r
B
R
I
S
Q
U
E
s
c
or
e
s
t
ha
n w
it
hout
c
om
bi
na
ti
on.
F
r
om
T
a
bl
e
s
5
to
7,
Z
e
r
o
-
D
C
E
c
ons
i
s
te
nt
ly
ha
d
th
e
hi
ghe
s
t
B
R
I
S
Q
U
E
s
c
or
e
s
,
in
di
c
a
ti
ng
th
a
t
m
odi
f
ic
a
ti
ons
a
nd
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
s
ig
ni
f
ic
a
nt
ly
i
m
pr
ove
d
im
a
ge
qua
li
ty
.
M
odi
f
ie
d
Z
e
r
o
-
D
C
E
,
e
ve
n
w
it
hout
c
om
bi
na
ti
on,
out
pe
r
f
or
m
e
d
Z
e
r
o
-
D
C
E
.
G
iv
e
n
th
e
w
id
e
in
te
n
s
it
y
r
a
nge
of
da
ta
s
e
t
im
a
ge
s
,
C
L
A
H
E
a
nd
B
P
D
H
E
yi
e
ld
e
d
be
tt
e
r
im
pr
ove
m
e
nt
s
th
a
n
C
S
.
T
he
ne
xt
s
te
p
is
to
di
vi
de
th
e
da
ta
in
to
tr
a
in
in
g,
va
li
da
ti
on, a
nd t
e
s
t
s
e
t
s
f
or
t
he
f
a
c
e
r
e
c
ogni
ti
on mode
l.
4
.
3
.
F
ac
e
r
e
c
ogn
it
io
n
m
od
e
li
n
g r
e
s
u
lt
s
T
he
c
l
a
s
s
if
ie
r
a
nd
tr
a
ns
f
e
r
le
a
r
ni
ng
a
r
c
hi
te
c
tu
r
e
s
e
le
c
ti
on
a
r
e
il
l
us
tr
a
te
d
in
F
ig
ur
e
15,
s
how
c
a
s
in
g
th
e
be
s
t
-
pe
r
f
or
m
in
g
m
ode
ls
ba
s
e
d
on
th
e
e
xp
e
r
im
e
nt
a
l
r
e
s
ul
ts
.
I
n
th
is
c
onf
ig
ur
a
ti
on,
th
e
f
in
a
l
f
ul
ly
c
onne
c
te
d
la
ye
r
of
V
G
G
16
/
R
e
s
N
e
t5
0 i
s
r
e
m
ove
d. F
e
a
tu
r
e
e
xt
r
a
c
ti
on i
s
pe
r
f
or
m
e
d f
r
om
t
he
l
a
s
t
c
onvolut
io
na
l
bl
oc
k, a
nd
th
e
r
e
s
ul
ti
ng f
e
a
tu
r
e
s
a
r
e
t
he
n pa
s
s
e
d t
hr
ough a
ba
tc
h nor
m
a
li
z
a
ti
on l
a
ye
r
.
F
ig
ur
e
15. T
r
a
ns
f
e
r
l
e
a
r
ni
ng a
r
c
hi
te
c
tu
r
e
T
he
a
c
c
ur
a
c
y
r
e
s
ul
ts
of
th
e
f
a
c
e
r
e
c
ogni
ti
on
m
ode
l
f
or
th
e
e
xt
e
nde
d
Y
a
le
f
a
c
e
da
ta
b
a
s
e
B
a
r
e
s
how
n
in
T
a
bl
e
8.
T
he
gr
e
e
n
-
c
ol
or
e
d
c
e
ll
in
di
c
a
te
s
th
e
be
s
t
-
pe
r
f
or
m
in
g
m
ode
l
ba
s
e
d
on
a
c
c
ur
a
c
y
m
e
tr
ic
s
.
T
hi
s
c
om
pa
r
is
on
hi
ghl
ig
ht
s
th
e
im
pa
c
t
of
d
if
f
e
r
e
nt
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
ods
on
r
e
c
ogni
ti
on
a
c
c
ur
a
c
y,
de
m
ons
tr
a
ti
ng t
he
e
f
f
e
c
ti
ve
ne
s
s
of
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
i
n i
m
p
r
ovi
ng f
a
c
ia
l
f
e
a
tu
r
e
c
la
r
it
y.
T
a
bl
e
8
pr
e
s
e
nt
s
a
c
c
ur
a
c
y
r
e
s
ul
ts
.
T
he
be
s
t
pe
r
f
or
m
a
nc
e
(
83.65%
)
is
a
c
hi
e
ve
d
by
V
G
G
16+
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
+
C
L
A
H
E
,
s
how
in
g
a
6.08%
im
pr
ove
m
e
nt
ove
r
Z
e
r
o
-
D
C
E
a
lo
ne
.
W
hi
le
Z
e
r
o
-
D
C
E
s
li
ght
ly
out
pe
r
f
or
m
s
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
in
e
xt
r
e
m
e
lo
w
-
li
ght
c
a
s
e
s
,
t
he
la
tt
e
r
c
om
bi
ne
d
w
it
h
C
L
A
H
E
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
s
a
c
c
ur
a
c
y.
F
or
R
e
s
N
e
t5
0,
Z
e
r
o
-
D
C
E
r
e
duc
e
s
a
c
c
ur
a
c
y,
but
m
odi
f
ie
d
Z
e
r
o
-
D
C
E
a
lo
ne
r
e
a
c
he
s
Evaluation Warning : The document was created with Spire.PDF for Python.