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
.
15
, N
o.
1
,
F
e
br
ua
r
y
2026
, pp.
213
~
228
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
15
.i
1
.pp
213
-
228
213
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
C
om
p
ar
i
son
of
i
m
age
e
n
h
an
c
e
m
e
n
t
m
e
t
h
od
s f
or
p
r
at
i
m
a t
h
e
f
t
d
e
t
e
c
t
i
on
u
si
n
g ar
t
i
f
i
c
i
al
i
n
t
e
l
l
i
ge
n
c
e
M
ad
e
S
u
d
ar
m
a
1
,
N
i
Wayan
S
r
i
A
r
iy
an
i
2
,
I
P
u
t
u
A
gu
s
E
k
a D
ar
m
a U
d
ayan
a
3
,
I
d
a B
agu
s
G
d
e
P
r
an
at
ayan
a
4
, L
ie
Jas
a
1
1
E
l
e
c
t
r
i
c
a
l
E
ngi
ne
e
r
i
ng S
t
udy
P
r
ogr
a
m
, F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng, U
da
ya
na
U
ni
ve
r
s
i
t
y
, D
e
npa
s
a
r
, I
ndone
s
i
a
2
I
ndus
t
r
i
a
l
E
ngi
ne
e
r
i
ng
S
t
udy P
r
ogr
a
m
, F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng, U
da
ya
na
U
ni
ve
r
s
i
t
y, D
e
npa
s
a
r
, I
ndone
s
i
a
3
I
nf
or
m
a
t
i
c
s
S
t
udy P
r
og
r
a
m
, F
a
c
ul
t
y of
T
e
c
hnol
ogy a
nd I
n
f
or
m
a
t
i
c
s
, I
ns
t
i
t
ut
e
B
us
i
ne
s
s
a
nd T
e
c
hnol
ogy I
ndone
s
i
a
, D
e
npa
s
a
r
, I
ndone
s
i
a
4
A
gr
oe
c
ot
e
c
hnol
ogy
S
t
udy P
r
ogr
a
m
, F
a
c
ul
t
y of
A
gr
i
c
ul
t
u
r
e
, U
da
ya
na
U
ni
ve
r
s
i
t
y, D
e
npa
s
a
r
, 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
N
ov
30
,
2024
R
e
vi
s
e
d
N
ov
17
,
2025
A
c
c
e
pt
e
d
J
a
n
10
,
2026
The
theft
of
pratima
in
Balinese
temples
threatens
the
spiritual
and
cultural
balance
of
the
community.
These
sacred
objects,
regarded
as
manifes
tations
of
God
in
Hinduism,
hold
profound
religious
significance,
and
the
ir
loss
represents
both
material
and
spiritual
desecration.
To
address
this
iss
ue,
this
study
investigates
a
security
system
that
leverages
image
enhancem
ent
for
low
-
light
detection.
Four
techniques
—
contrast
limit
ed
adaptive
hist
ogram
equalizati
on
(CLAHE),
adaptive
histog
ram
equalizati
on
(AHE),
histogram
equaliza
tion
(HE),
and
gamma
correction
—
were
evaluated
to
i
mprove
image
quality.
CLAHE
yielded
the
lowest
mean
squared
error
(M
SE)
of
21.16
and
the
highest
peak
signal
-
to
-
noise
ratio
(PSNR)
of
38.13
d
B.
For
object
detection,
VGG
-
19
and
AlexNet
were
assessed.
The
best
configurat
ion,
VGG
-
19
with
HE,
reached
83.33%
accuracy
and
9
3.75%
recall,
and
achieved
a
receiver
operating
characteristic
area
under
the
curve
(
ROC
AUC
)
of
0.90±
0.02
across
five
runs.
Thresholds
derived
from
the
ROC
analysis
were
selected
using
the
Youden
J
statistic
to
b
alance
sensitivity
and
specificity.
The
approach
outperformed
lightweig
ht
and
classical
baselines
in
AUC,
indicati
ng
superior
discrim
ination
und
er
low
illumination.
These
findings
show
that
superior
image
quality
do
es
not
always
align
with
higher
detection
accuracy,
and
they
highlig
ht
the
importance
of
pairing
effec
tive
enhance
ment
with
robust
detector
s
for
temple
security.
The
study
contributes
practi
cal
insights
for
pres
erving
Balinese
cultura
l
and
spiritual
heritag
e
by
strengthe
ning
effor
ts
to
protec
t
pratima
against
theft.
K
e
y
w
o
r
d
s
:
A
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
I
m
a
ge
pr
oc
e
s
s
in
g
I
m
a
ge
qua
li
ty
e
nha
nc
e
m
e
nt
P
r
a
ti
m
a
s
e
c
ur
it
y s
ys
te
m
S
ur
ve
il
la
nc
e
s
ys
te
m
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
:
I
P
ut
u A
gus
E
ka
D
a
r
m
a
U
da
ya
na
I
nf
or
m
a
ti
c
s
S
tu
dy P
r
og
r
a
m
, F
a
c
ul
ty
of
T
e
c
hnol
ogy a
nd I
nf
or
m
a
ti
c
s
I
ns
ti
tu
te
B
us
in
e
s
s
a
nd T
e
c
hnol
ogy I
ndone
s
ia
T
uka
d P
a
ke
r
is
a
n S
tr
e
e
t,
D
e
np
a
s
a
r
C
it
y, B
a
li
P
r
ovi
nc
e
, I
ndone
s
ia
E
m
a
il
:
a
gus
.e
ka
da
r
m
a
@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
he
te
r
m
"
P
ur
a
"
or
ig
in
a
te
s
f
r
om
S
a
ns
kr
it
s
uf
f
ix
e
s
s
uc
h
a
s
(
pur
,
pur
i,
pur
a
,
pur
a
m
,
a
nd
por
e
)
,
w
hi
c
h
m
e
a
n
c
it
y,
f
or
ti
f
ie
d
c
it
y,
or
c
it
y
w
it
h
to
w
e
r
s
a
nd
pa
la
c
e
s
.
I
n
B
a
li
,
th
is
te
r
m
ha
s
e
vol
ve
d
in
to
a
s
p
e
c
if
ic
de
s
ig
na
ti
on
f
or
pl
a
c
e
s
of
w
or
s
hi
p,
w
hi
le
"
P
u
r
i"
r
e
f
e
r
s
to
th
e
r
e
s
id
e
nc
e
s
of
ki
ngs
a
nd
nobl
e
s
.
T
he
f
unc
ti
ons
of
a
P
ur
a
c
a
n
be
c
a
te
gor
iz
e
d
ba
s
e
d
on
c
e
r
ta
in
c
ha
r
a
c
te
r
is
ti
c
s
th
a
t
r
e
f
le
c
t
s
oc
ia
l,
pol
it
ic
a
l,
e
c
onomi
c
,
or
ge
ne
a
lo
gi
c
a
l
bond
s
w
it
hi
n
th
e
c
om
m
uni
ty
.
F
or
e
xa
m
pl
e
, s
oc
ia
l
ti
e
s
m
a
y
r
e
la
te
to
r
e
s
id
e
nt
ia
l
a
r
e
a
s
(
te
r
r
it
or
ia
l)
or
th
e
ve
ne
r
a
ti
on
of
a
s
a
c
r
e
d
te
a
c
he
r
[
1]
.
O
ne
of
th
e
e
s
s
e
nt
ia
l
e
le
m
e
nt
s
of
a
P
ur
a
a
s
a
pl
a
c
e
of
w
or
s
hi
p
i
s
th
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
.
15
, N
o.
1
,
F
e
br
ua
r
y
20
26
:
213
-
228
214
pr
a
ti
m
a
.
A
pr
a
ti
m
a
is
a
s
a
c
r
e
d
obj
e
c
t
in
th
e
f
or
m
of
a
s
ta
tu
e
,
b
e
li
e
ve
d
to
c
ont
a
in
m
y
s
ti
c
a
l
or
s
pi
r
it
ua
l
pow
e
r
.
T
hi
s
s
ta
tu
e
s
e
r
ve
s
a
s
a
s
ym
bol
f
or
c
om
m
uni
c
a
ti
ng
w
it
h
G
od.
A
lt
hough
it
m
a
y
r
e
s
e
m
bl
e
a
n
or
di
na
r
y
obj
e
c
t,
a
pr
a
ti
m
a
ha
s
unde
r
gone
a
pur
if
ic
a
ti
on
pr
oc
e
s
s
pe
r
f
or
m
e
d
a
c
c
or
di
ng
to
th
e
be
li
e
f
s
of
th
e
H
in
du
c
om
m
uni
ty
in
B
a
li
.
I
n
th
is
c
ont
e
xt
,
th
e
pr
a
ti
m
a
is
c
ons
id
e
r
e
d
th
e
dw
e
ll
in
g
pl
a
c
e
of
G
od,
or
I
da
S
a
ng
H
ya
ng
W
id
hi
W
a
s
a
,
a
nd
is
us
e
d
a
s
a
m
e
a
ns
of
w
or
s
hi
p
by
H
in
dus
[
2]
.
H
ow
e
ve
r
,
th
e
r
e
c
e
nt
th
e
f
t
of
p
r
a
ti
m
a
ha
s
be
c
om
e
a
tr
oubl
in
g
is
s
ue
f
or
th
e
B
a
li
ne
s
e
c
om
m
uni
ty
.
T
hi
s
a
c
t
not
onl
y
c
a
us
e
s
m
a
te
r
ia
l
lo
s
s
e
s
but
a
ls
o
br
in
gs
non
-
m
a
te
r
ia
l
ha
r
m
, a
f
f
e
c
ti
ng t
he
s
pi
r
it
ua
l
a
nd mys
ti
c
a
l
ba
la
nc
e
of
s
oc
ie
ty
. F
or
t
he
B
a
li
ne
s
e
H
in
du c
om
m
uni
ty
, t
he
th
e
f
t
of
pr
a
ti
m
a
is
s
e
e
n
a
s
a
de
s
e
c
r
a
ti
on
of
r
e
li
gi
on,
a
s
pr
a
ti
m
a
is
r
e
ga
r
de
d
a
s
a
hi
ghl
y
s
a
c
r
e
d
obj
e
c
t.
A
ddi
ti
ona
ll
y,
m
a
ny
B
a
li
ne
s
e
a
r
e
unw
il
li
ng
to
a
c
c
e
pt
ba
c
k
a
s
to
le
n
pr
a
ti
m
a
,
a
s
it
is
c
ons
id
e
r
e
d
to
ha
ve
lo
s
t
i
ts
s
a
nc
ti
ty
.
A
s
a
r
e
s
ul
t,
s
to
le
n
pr
a
ti
m
a
a
r
e
ty
pi
c
a
ll
y
s
to
r
e
d
in
th
e
B
a
li
R
e
gi
ona
l
M
us
e
um
a
nd
a
r
e
no
lo
nge
r
us
e
d
in
r
e
li
gi
ous
c
e
r
e
m
oni
e
s
.
G
iv
e
n
th
is
s
it
ua
ti
on,
a
n
e
f
f
e
c
ti
ve
s
e
c
ur
it
y
s
ys
te
m
is
ne
e
d
e
d
to
a
ddr
e
s
s
th
e
is
s
ue
of
pr
a
ti
m
a
th
e
f
t.
I
n
th
is
r
e
ga
r
d,
th
e
a
ppl
ic
a
ti
on
of
im
a
ge
e
nha
nc
e
m
e
nt
m
e
th
ods
is
c
r
uc
ia
l
to
im
pr
ovi
ng
im
a
ge
qua
li
ty
in
s
ubopti
m
a
l
li
ght
in
g
c
ondi
ti
ons
.
T
he
im
a
ge
e
nha
nc
e
m
e
nt
m
e
th
ods
to
be
a
ppl
ie
d
in
c
lu
de
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
)
,
a
da
pt
iv
e
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
(
A
H
E
)
,
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
(
H
E
)
,
a
nd
ga
m
m
a
c
or
r
e
c
ti
on.
T
he
s
e
f
our
m
e
th
ods
w
il
l
be
te
s
te
d
to
de
te
r
m
in
e
w
hi
c
h
is
th
e
m
os
t
e
f
f
e
c
ti
ve
i
n e
nha
nc
in
g i
m
a
ge
qua
li
ty
i
n da
r
k a
r
e
a
s
.
F
ur
th
e
r
m
or
e
,
to
de
te
c
t
th
e
f
t
pe
r
pe
tr
a
to
r
s
,
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
-
ba
s
e
d
m
e
th
ods
us
in
g
VGG
-
19
a
nd
A
le
xN
e
t
a
r
c
hi
te
c
tu
r
e
s
w
il
l
be
e
m
pl
oye
d.
V
G
G
-
19
a
nd
A
le
xN
e
t
a
r
e
two
C
N
N
a
r
c
hi
te
c
tu
r
e
s
known
f
or
th
e
ir
e
f
f
e
c
ti
ve
ne
s
s
in
obj
e
c
t
de
te
c
ti
on
ta
s
ks
.
I
n
th
is
s
tu
dy,
bot
h
C
N
N
a
r
c
hi
te
c
tu
r
e
s
w
il
l
be
a
ppl
ie
d
to
a
na
ly
z
e
im
a
ge
s
a
nd
id
e
nt
if
y
s
us
pi
c
io
u
s
a
c
ti
vi
ti
e
s
.
T
hi
s
r
e
s
e
a
r
c
h
w
il
l
c
om
pa
r
e
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
th
e
f
our
im
a
ge
e
nha
nc
e
m
e
nt
m
e
th
ods
in
im
pr
ovi
ng
im
a
ge
qua
li
ty
a
nd
e
va
lu
a
te
th
e
obj
e
c
t
de
te
c
ti
on
pe
r
f
or
m
a
nc
e
o
f
th
e
two
C
N
N
a
r
c
hi
te
c
tu
r
e
s
m
e
nt
io
ne
d.
T
h
e
goa
l
is
to
de
te
r
m
in
e
w
hi
c
h
im
a
ge
e
nh
a
nc
e
m
e
nt
m
e
th
od
yi
e
ld
s
th
e
be
s
t
r
e
s
ul
ts
in
lo
w
-
li
ght
c
ondi
ti
ons
a
nd
w
hi
c
h
C
N
N
a
r
c
hi
te
c
tu
r
e
of
f
e
r
s
th
e
hi
ghe
s
t
a
c
c
ur
a
c
y
in
de
te
c
ti
ng
pr
a
ti
m
a
th
e
f
t
pe
r
pe
tr
a
to
r
s
.
W
it
h
th
is
a
ppr
oa
c
h,
it
is
hope
d
t
ha
t
th
e
s
e
c
ur
it
y
s
ys
te
m
s
in
pr
a
ti
m
a
s
to
r
a
ge
lo
c
a
ti
ons
c
a
n be
i
m
pr
ove
d, pr
e
ve
nt
in
g t
he
f
t
in
t
he
f
ut
ur
e
.
T
he
nove
lt
y
of
th
i
s
s
tu
dy
li
e
s
in
it
s
c
om
pr
e
he
n
s
iv
e
e
va
lu
a
ti
on
of
im
a
ge
e
nha
n
c
e
m
e
nt
m
e
th
od
s
c
om
bi
ne
d
w
it
h
a
dva
nc
e
d
C
N
N
a
r
c
hi
te
c
tu
r
e
s
f
or
a
ddr
e
s
s
in
g
th
e
uni
que
c
ha
ll
e
nge
s
of
th
e
f
t
de
te
c
ti
on
in
B
a
li
ne
s
e
te
m
pl
e
s
.
U
nl
ik
e
pr
e
vi
ous
s
tu
di
e
s
th
a
t
of
te
n
f
oc
us
on
g
e
ne
r
a
l
obj
e
c
t
de
te
c
ti
on
or
im
a
ge
e
nha
nc
e
m
e
nt
in
w
e
ll
-
li
t
e
nvi
r
onm
e
nt
s
,
th
is
r
e
s
e
a
r
c
h
s
pe
c
if
ic
a
ll
y
t
a
r
ge
ts
lo
w
-
li
ght
c
ondi
ti
ons
c
om
m
on
in
te
m
pl
e
s
e
tt
in
gs
,
ut
il
iz
in
g
C
L
A
H
E
,
A
H
E
,
H
E
,
a
nd
ga
m
m
a
c
or
r
e
c
ti
on
to
im
p
r
ove
im
a
ge
c
la
r
it
y.
F
ur
th
e
r
m
o
r
e
,
th
e
us
e
of
VGG
-
19
a
nd
A
le
xN
e
t
a
r
c
hi
te
c
tu
r
e
s
pr
ovi
de
s
va
lu
a
bl
e
in
s
ig
ht
s
in
to
th
e
tr
a
de
-
of
f
s
be
twe
e
n
im
a
ge
qua
li
ty
a
nd
de
te
c
ti
on
a
c
c
ur
a
c
y,
w
hi
c
h
ha
s
not
be
e
n
e
xt
e
ns
iv
e
ly
e
xpl
or
e
d
i
n
th
is
c
ont
e
xt
.
B
y
ta
il
or
in
g
s
ta
t
e
-
of
-
th
e
-
a
r
t
A
I
te
c
hnol
ogi
e
s
to
pr
ot
e
c
t
pr
a
ti
m
a
,
th
is
s
tu
dy
not
onl
y
c
ont
r
ib
ut
e
s
to
th
e
f
ie
ld
of
im
a
ge
pr
oc
e
s
s
in
g
a
nd
s
e
c
ur
it
y
s
ys
te
m
s
but
a
ls
o
s
e
r
ve
s
to
pr
e
s
e
r
ve
th
e
c
ul
tu
r
a
l
a
nd
s
pi
r
it
ua
l
he
r
it
a
ge
of
B
a
li
.
T
h
e
f
in
di
ngs
a
r
e
e
xp
e
c
te
d
to
gui
de
f
ut
ur
e
de
ve
lo
pm
e
nt
s
i
n s
e
c
ur
it
y s
y
s
te
m
s
f
or
c
ul
tu
r
a
l
a
nd r
e
li
gi
ous
s
it
e
s
f
a
c
in
g
s
im
il
a
r
c
ha
ll
e
nge
s
.
2.
S
T
A
T
E
O
F
T
H
E
A
R
T
T
hi
s
s
e
c
ti
on
r
e
vi
e
w
s
tw
o
c
om
pl
e
m
e
nt
a
r
y
a
r
e
a
s
t
o
s
uppor
t
pr
a
ti
m
a
th
e
f
t
de
t
e
c
ti
on
u
nde
r
lo
w
-
li
g
ht
c
ondi
ti
ons
in
te
m
pl
e
e
nvi
r
o
nm
e
nt
s
.
T
h
e
f
ir
s
t
is
l
ow
-
li
ght
im
a
g
e
e
nh
a
nc
e
m
e
nt
(
L
L
I
E
)
,
w
hi
c
h
im
pr
ov
e
s
vi
s
u
a
l
qua
li
ty
a
s
pr
e
-
pr
o
c
e
s
s
i
ng s
t
e
p
s
o t
h
a
t
r
e
l
e
va
n
t
obj
e
c
t
s
be
c
om
e
m
or
e
de
t
e
c
t
a
bl
e
. T
he
s
e
c
ond i
s
s
ur
ve
il
la
n
c
e
t
h
e
f
t
or
a
nom
a
l
y
de
t
e
c
ti
o
n,
w
hi
c
h
e
v
a
lu
a
te
s
m
ode
l
pe
r
f
or
m
a
nc
e
in
r
e
a
l
-
w
or
ld
m
o
ni
to
r
in
g
s
c
e
na
r
io
s
.
T
h
e
r
e
vi
e
w
hi
ghl
ig
ht
s
a
dom
a
in
g
a
p
b
e
tw
e
e
n
ge
n
e
r
a
l
s
tu
di
e
s
a
n
d
te
m
p
le
c
ont
e
xt
s
th
a
t
e
xhi
bi
t
non
uni
f
or
m
il
l
um
in
a
ti
on,
in
tr
ic
a
t
e
or
n
a
m
e
nt
s
,
a
nd
pr
a
c
ti
c
a
l
c
on
s
tr
a
in
ts
f
or
e
dge
-
de
vi
c
e
de
pl
oym
e
nt
.
T
a
bl
e
1
pr
e
s
e
n
ts
a
c
om
pa
r
a
t
iv
e
s
um
m
a
r
y of
L
L
I
E
s
tu
di
e
s
,
c
ove
r
i
ng me
th
od
s
,
e
va
l
ua
ti
o
n da
ta
s
e
t
s
, r
e
por
te
d
m
e
tr
i
c
s
,
a
nd
ke
y r
e
s
ul
t
s
.
T
a
bl
e
1.
C
om
pa
r
a
ti
ve
s
um
m
a
r
y of
L
L
I
E
m
e
th
ods
R
e
s
e
a
r
c
h
M
e
t
hod
D
a
t
a
s
e
t
s
M
e
t
r
i
c
s
R
e
s
ul
t
R
e
t
i
ne
x
-
N
e
t
[
3]
,
2018
D
e
e
p
r
e
t
i
ne
x
de
c
om
pos
i
t
i
on of
i
l
l
um
i
na
t
i
on a
nd
r
e
f
l
e
c
t
a
nc
e
L
ow
-
l
i
ght
da
t
a
s
e
t
(
L
O
L
)
;
c
om
m
onl
y c
om
pa
r
e
d on
L
I
M
E
, M
E
F
, N
P
E
, D
I
C
M
,
VV
P
e
a
k s
i
gna
l
-
to
-
noi
s
e
r
a
t
i
o (
P
S
N
R
)
,
s
t
r
uc
t
ur
a
l
s
i
m
i
l
a
r
i
t
y i
nde
x (
S
S
I
M
)
E
a
r
l
y be
nc
hm
a
r
k
a
nc
hor
e
d on L
O
L
.
Z
e
r
o
-
D
C
E
[
4]
,
2020
Z
e
r
o
-
r
e
f
e
r
e
nc
e
de
e
p
c
ur
ve
e
s
t
i
m
a
t
i
on
D
I
C
M
, L
I
M
E
, M
E
F
, N
P
E
,
V
V
, S
I
C
E
;
s
om
e
t
i
m
e
s
L
O
L
N
I
Q
E
, L
O
E
;
P
S
N
R
/
S
S
I
M
i
f
G
T
e
xi
s
t
s
L
i
ght
w
e
i
ght
a
nd r
obus
t
t
o nonuni
f
or
m
l
i
ght
i
ng.
E
n
l
i
gh
t
e
nG
A
N
[
5]
,
202
1
U
npa
i
r
e
d G
A
N
f
or
l
ow
-
l
i
ght
e
nha
nc
e
m
e
nt
L
I
M
E
, M
E
F
, N
P
E
, D
I
C
M
,
r
e
a
l
-
w
or
l
d i
m
a
ge
s
N
I
Q
E
, s
ubj
e
c
t
i
ve
t
e
s
t
s
;
l
i
m
i
t
e
d S
S
I
M
/
P
S
N
R
S
t
r
ong pe
r
c
e
pt
ua
l
qua
l
i
t
y
w
i
t
h hi
ghe
r
i
nf
e
r
e
nc
e
c
os
t
.
U
R
e
t
i
ne
x
-
N
e
t
[
6]
,
2022
D
e
e
p unf
ol
di
ng of
R
e
t
i
ne
x f
or
m
ul
a
t
i
on
L
O
L
pl
u
s
non
-
pa
i
r
e
d s
e
t
s
P
S
N
R
, S
S
I
M
, L
P
I
P
S
H
i
gh a
c
c
ur
a
c
y i
n e
xt
r
e
m
e
l
ow
l
i
ght
, he
a
vi
e
r
m
ode
l
.
GA
-
R
e
t
i
ne
x
[
7]
,
2023
R
e
t
i
ne
x w
i
t
h gl
oba
l
a
t
t
e
nt
i
on
C
om
m
onl
y L
I
M
E
, M
E
F
,
N
P
E
, D
I
C
M
P
S
N
R
, S
S
I
M
B
e
t
t
e
r
gl
oba
l
c
ons
i
s
t
e
nc
y
on br
oa
d s
c
e
ne
s
.
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
C
om
par
is
on of i
m
age
e
nhanc
e
m
e
nt
m
e
th
ods
f
or
p
r
at
ima the
ft
d
e
te
c
ti
on
…
(
M
ade
Sudar
m
a)
215
T
hi
s
ov
e
r
vi
e
w
in
di
c
a
te
s
th
a
t
li
ght
w
e
ig
ht
de
te
r
m
in
is
ti
c
a
ppr
o
a
c
he
s
r
e
m
a
in
r
e
le
va
nt
a
s
c
ont
r
ol
le
d
ba
s
e
li
ne
s
a
nd
a
r
e
s
ui
ta
bl
e
f
or
de
pl
oym
e
nt
in
lo
w
-
li
ght
te
m
pl
e
s
c
e
na
r
io
s
.
T
a
bl
e
2
c
om
pa
r
e
s
s
ur
ve
il
la
nc
e
th
e
f
t
de
te
c
ti
on s
tu
di
e
s
. I
nc
lu
di
ng pr
obl
e
m
de
f
in
it
io
ns
, da
ta
s
e
ts
, e
va
lu
a
ti
on me
tr
ic
s
, a
nd r
e
s
ul
ts
.
T
a
bl
e
2.
C
om
pa
r
a
ti
ve
s
um
m
a
r
y of
s
ur
ve
il
la
nc
e
t
he
f
t
de
te
c
ti
on
R
e
s
e
a
r
c
h
P
r
obl
e
m
D
a
t
a
s
e
t
s
M
e
t
r
i
c
s
R
e
s
ul
t
U
C
F
-
c
r
i
m
e
[
8]
, 2018
V
i
de
o a
nom
a
l
y de
t
e
c
t
i
on
i
n t
he
w
i
l
d (
bur
gl
a
r
y,
a
nd
r
obbe
r
y)
1.9k vi
de
os
, 128
hour
s
R
O
C
-
A
U
C
,
AP
S
t
a
nda
r
d l
a
r
ge
-
s
c
a
l
e
w
e
a
kl
y l
a
be
l
e
d C
C
T
V
be
nc
hm
a
r
k.
S
e
ns
i
t
i
ve
a
c
t
i
vi
t
y
[
9]
,
2021
S
hopl
i
f
t
i
ng pr
e
di
c
t
i
on
vi
a
s
oc
i
a
l
s
i
gna
l
c
u
e
s
M
a
nua
l
l
y a
nnot
a
t
e
d
r
e
t
a
i
l
vi
de
os
A
c
c
ur
a
c
y,
i
nt
e
r
pr
e
t
a
bi
l
i
t
y
H
i
ghl
i
ght
s
t
r
a
ns
pa
r
e
nt
de
c
i
s
i
on f
a
c
t
or
s
.
S
ho
pl
i
f
t
i
ng
d
e
t
e
c
t
i
on
[
10]
,
20
23
S
hopl
i
f
t
i
ng c
l
a
s
s
i
f
i
c
a
t
i
on
w
i
t
h 2D
or
3D
C
N
N
a
nd
hybr
i
ds
900 i
ns
t
a
nc
e
s
A
c
c
ur
a
c
y,
F1
-
s
c
or
e
S
t
r
uc
t
ur
e
d da
t
a
s
e
t
f
or
vi
de
o
-
ba
s
e
d s
hopl
i
f
t
i
ng
be
nc
hm
a
r
ki
ng.
3.
M
E
T
H
O
D
T
h
e
s
tu
dy
e
m
pl
o
ys
a
f
i
ve
-
s
ta
ge
e
x
pe
r
im
e
nt
a
l
pi
p
e
l
in
e
.
T
h
e
i
n
it
i
a
l
s
t
a
g
e
c
ov
e
r
s
da
ta
s
e
t
a
c
qu
is
it
i
on,
c
on
tr
a
s
t
e
nh
a
n
c
e
m
e
nt
th
r
ou
gh C
L
A
H
E
,
pr
e
pr
oc
e
s
s
in
g, a
n
d
m
od
e
l
im
p
le
m
e
n
ta
ti
o
n us
in
g
A
le
xN
e
t
a
n
d V
G
G
-
19.
T
he
s
e
c
ond
s
ta
ge
m
ir
r
or
s
th
is
pr
oc
e
dur
e
but
ut
il
iz
e
s
A
H
E
i
ns
te
a
d
of
C
L
A
H
E
.
T
he
th
ir
d
s
ta
ge
pe
r
f
or
m
s
e
nha
nc
e
m
e
nt
us
in
g
HE
.
I
n
th
e
f
our
th
s
ta
ge
,
e
nha
n
c
e
m
e
nt
is
e
xc
lu
de
d,
a
nd
onl
y
pr
e
pr
oc
e
s
s
in
g
a
nd
m
ode
l
a
ppl
ic
a
ti
on
a
r
e
c
ondu
c
te
d.
T
he
f
in
a
l
s
ta
ge
pe
r
f
or
m
s
a
c
c
u
r
a
c
y
a
s
s
e
s
s
m
e
nt
,
r
e
s
ul
t
in
te
r
pr
e
ta
ti
on,
a
nd
c
om
pa
r
a
ti
ve
a
na
ly
s
is
b
e
twe
e
n e
nh
a
nc
e
d a
nd non
-
e
nh
a
nc
e
d
s
c
e
na
r
io
s
. F
ig
ur
e
1 de
pi
c
ts
t
he
ove
r
a
ll
w
or
kf
lo
w
.
F
ig
ur
e
1. R
e
s
e
a
r
c
h
f
lo
w
3.1.
D
at
as
e
t
T
he
in
it
ia
l
pha
s
e
of
th
is
r
e
s
e
a
r
c
h
in
vol
ve
s
da
ta
s
e
t
a
c
qui
s
it
io
n
f
r
om
a
publ
ic
ly
a
va
il
a
bl
e
r
e
pos
it
or
y
on
th
e
K
a
ggl
e
pl
a
tf
or
m
.
A
s
s
how
n
in
T
a
bl
e
3,
th
e
da
ta
s
e
t
c
on
s
is
t
s
of
921
im
a
ge
s
c
a
te
gor
iz
e
d
in
to
two
c
la
s
s
e
s
:
559
im
a
ge
s
c
ont
a
in
in
g
in
di
vi
dua
ls
a
nd
362
im
a
ge
s
w
it
hout
in
d
iv
id
ua
ls
.
T
hi
s
c
om
pos
it
io
n
pr
ovi
de
s
a
de
qua
te
va
r
ia
ti
on
be
twe
e
n
c
la
s
s
e
s
,
e
n
a
bl
in
g
th
e
m
ode
l
to
c
a
pt
ur
e
di
s
ti
nc
ti
ve
vi
s
ua
l
pa
tt
e
r
ns
f
or
im
pr
ove
d
c
la
s
s
if
ic
a
ti
on pe
r
f
or
m
a
nc
e
.
F
r
om
T
a
bl
e
3,
it
c
a
n
be
obs
e
r
ve
d
th
a
t
th
e
da
ta
s
e
t
c
ont
a
in
s
a
gr
e
a
te
r
num
be
r
of
im
a
ge
s
w
it
h
in
di
vi
dua
ls
c
om
pa
r
e
d
to
th
os
e
w
it
hout
.
T
hi
s
va
r
ia
ti
on
e
ns
ur
e
s
th
a
t
th
e
m
ode
l
is
e
xpos
e
d
to
di
ve
r
s
e
vi
s
ua
l
c
ha
r
a
c
te
r
is
ti
c
s
,
w
hi
c
h
is
e
s
s
e
nt
ia
l
f
or
a
c
hi
e
vi
ng
a
c
c
ur
a
te
a
nd
g
e
ne
r
a
li
z
e
d
c
la
s
s
if
ic
a
ti
on
r
e
s
ul
ts
.
T
h
e
da
ta
s
e
t
in
th
is
s
tu
dy
is
r
e
la
ti
ve
ly
m
ode
s
t
in
s
iz
e
(
921
im
a
ge
s
)
a
nd
doe
s
not
ye
t
in
c
lu
de
ve
r
i
f
ie
d
pr
a
ti
m
a
th
e
f
t
e
ve
nt
s
,
in
tr
oduc
in
g
a
dom
a
in
ga
p
w
it
h
r
e
a
l
-
w
or
ld
de
pl
oym
e
nt
.
T
o
c
ur
b
ove
r
ly
opt
im
is
ti
c
e
s
ti
m
a
te
s
,
r
e
s
ul
ts
in
th
e
r
e
s
ul
ts
s
e
c
ti
on
a
r
e
s
um
m
a
r
iz
e
d
us
in
g
f
ol
d
-
ba
s
e
d
m
e
a
ns
w
it
h
c
onf
id
e
nc
e
in
te
r
va
ls
,
w
hi
le
a
ugm
e
nt
a
ti
ons
a
r
e
ta
il
or
e
d
to
lo
w
-
li
gh
t
te
m
pl
e
s
c
e
ne
s
to
be
tt
e
r
a
ppr
oxi
m
a
te
f
ie
ld
c
ondi
ti
ons
.
L
ooki
ng
a
he
a
d,
c
oor
di
na
ti
on
w
it
h
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
.
15
, N
o.
1
,
F
e
br
ua
r
y
20
26
:
213
-
228
216
te
m
pl
e
a
ut
hor
it
ie
s
is
pl
a
nn
e
d
to
e
th
ic
a
ll
y
r
e
-
e
n
a
c
t
ni
ght
s
c
e
na
r
io
s
,
br
oa
de
n
m
ul
ti
-
s
it
e
a
nd
m
ul
ti
-
de
vi
c
e
d
a
ta
c
ol
le
c
ti
on, a
nd pe
r
f
or
m
ha
r
d
-
ne
ga
ti
ve
m
in
in
g f
r
om
pi
lo
t
de
pl
oym
e
nt
s
.
T
a
bl
e
3. D
a
ta
s
e
t
c
ount
D
a
t
a
s
e
t
s
A
m
ount
of
da
t
a
T
he
phot
o c
ont
a
i
ns
s
om
e
one
559
T
he
phot
o doe
s
not
c
ont
a
i
n a
nyone
362
3.3.
H
is
t
ogr
am
e
q
u
al
iz
at
io
n
A
hi
s
to
gr
a
m
r
e
pr
e
s
e
nt
s
t
he
di
s
tr
ib
ut
io
n
of
pi
xe
l
in
t
e
n
s
it
ie
s
i
n
a
n
im
a
g
e
a
nd
he
l
ps
de
t
e
r
m
in
e
w
he
th
e
r
th
e
im
a
ge
a
p
pe
a
r
s
d
a
r
k or
br
ig
ht
[
11]
,
[
12]
.
HE
e
nh
a
n
c
e
s
im
a
ge
c
ont
r
a
s
t
by
r
e
m
a
p
pi
ng pixe
l
in
t
e
n
s
it
ie
s
s
o
t
ha
t
gr
a
y
le
v
e
l
s
a
r
e
m
or
e
un
if
or
m
ly
di
s
tr
ib
ut
e
d.
T
h
e
r
e
m
a
ppi
ng
i
s
p
e
r
f
or
m
e
d
th
r
o
ugh
a
tr
a
n
s
f
or
m
a
ti
on
f
un
c
ti
on
T
,
w
r
it
te
n
a
s
s
=
T
(
r
)
,
w
hi
l
e
th
e
or
ig
in
a
l
v
a
lu
e
c
a
n
be
r
e
c
o
ve
r
e
d
w
i
th
r
=
T
⁻
¹(
s
)
.
B
ot
h
m
a
ppi
ngs
ope
r
a
te
w
it
hi
n
th
e
r
a
nge
0
≤ r
, s
≤1
. T
he
go
a
l
of
t
hi
s
t
e
c
hni
qu
e
i
s
t
o obta
in
a
m
or
e
e
ve
n hi
s
to
gr
a
m
w
he
r
e
e
a
c
h gr
a
y l
e
v
e
l
c
o
nt
a
in
s
a
s
im
il
a
r
numb
e
r
of
pi
x
e
l
s
,
a
s
r
e
pr
e
s
e
n
te
d
by t
h
e
pr
ob
a
bi
li
t
y di
s
tr
i
but
io
n f
un
c
ti
on
i
n (
1)
[
1
1]
, [
13]
, [
14]
.
P
(
)
=
ℎ
=
−
1
,
0
≤
≤
−
1
(
1)
W
he
r
e
t
he
gr
a
y l
e
ve
l
(
k)
i
s
nor
m
a
li
z
e
d a
ga
in
s
t
th
e
m
a
xi
m
um
gr
a
y va
lu
e
(
L
-
1
)
. I
n
t
he
s
pe
c
if
ie
d gr
a
y s
c
a
le
, t
he
va
lu
e
r
k=
0r
k=
0r
k=
0
in
di
c
a
te
s
th
e
c
ol
or
bl
a
c
k,
w
hi
l
e
r
k=
1r
k
=
1r
k=
1
in
di
c
a
te
s
th
e
c
ol
or
w
hi
te
[
15]
,
[
16]
.
A
not
he
r
f
or
m
ul
a
t
ha
t
c
a
n be
us
e
d t
o c
a
lc
ul
a
te
HE
f
or
a
n i
m
a
ge
w
it
h a
k
-
bi
t
g
r
a
y s
c
a
le
i
s
a
s
s
how
n i
n (
2)
.
T
hi
s
e
xpl
a
na
ti
on
hi
ghl
ig
ht
s
th
e
im
por
ta
nc
e
of
nor
m
a
li
z
a
ti
on
in
e
n
s
ur
in
g
a
c
c
ur
a
te
di
s
tr
ib
ut
io
n
of
pi
xe
l
in
te
ns
it
ie
s
a
c
r
os
s
t
he
i
m
a
ge
.
=
(
(
2
−
1
)
ℎ
)
(
2)
3.4.
A
d
ap
t
iv
e
h
is
t
ogr
am
e
q
u
al
iz
at
io
n
A
H
E
e
nha
nc
e
s
im
a
ge
c
ont
r
a
s
t
by
di
vi
di
ng
th
e
im
a
ge
in
to
s
e
ve
r
a
l
r
e
gi
ons
a
nd
pe
r
f
o
r
m
in
g
HE
on
e
a
c
h
one
in
di
vi
dua
ll
y.
B
y
ge
n
e
r
a
ti
ng
m
ul
ti
pl
e
lo
c
a
l
hi
s
to
gr
a
m
s
,
A
H
E
a
dj
us
t
s
in
te
ns
it
y
va
lu
e
s
di
f
f
e
r
e
nt
ly
in
e
a
c
h
s
e
gm
e
nt
,
w
hi
c
h
im
pr
ove
s
vi
s
ib
il
it
y
of
f
in
e
de
ta
il
s
[
17]
.
U
nl
ik
e
gl
oba
l
HE
,
A
H
E
a
ppl
ie
s
a
n
a
da
pt
iv
e
s
tr
a
te
gy
th
a
t
be
tt
e
r
ha
ndl
e
s
va
r
ia
ti
on
s
a
c
r
os
s
di
f
f
e
r
e
nt
im
a
ge
a
r
e
a
s
[
18]
.
A
s
a
r
e
s
ul
t,
it
is
pa
r
ti
c
ul
a
r
ly
e
f
f
e
c
ti
ve
in
boos
ti
ng
lo
c
a
l
c
ont
r
a
s
t
a
nd
s
ha
r
pe
ni
ng
e
dge
s
th
r
oughout
th
e
im
a
ge
.
B
e
c
a
u
s
e
of
it
s
lo
c
a
li
z
e
d
pr
oc
e
s
s
in
g,
A
H
E
i
s
a
ls
o c
om
m
onl
y r
e
f
e
r
r
e
d t
o a
s
lo
c
a
l
hi
s
to
gr
a
m
pr
oc
e
s
s
in
g
[
11]
.
3.5.
C
on
t
r
as
t
li
m
it
e
d
ad
ap
t
iv
e
h
is
t
ogr
am
e
q
u
al
iz
a
t
io
n
C
L
A
H
E
i
s
a
r
e
f
in
e
d
v
e
r
s
io
n
of
A
H
E
th
a
t
pr
ovi
d
e
s
a
s
im
p
le
r
a
n
d
m
or
e
e
f
f
ic
i
e
nt
c
ont
r
a
s
t
e
nh
a
nc
e
m
e
nt
a
ppr
oa
c
h
[
19]
. T
he
m
e
th
od
i
s
e
f
f
e
c
t
iv
e
in
s
e
p
a
r
a
ti
ng
f
or
e
gr
oun
d
obj
e
c
t
s
f
r
om
th
e
b
a
c
kgr
o
und,
r
e
duc
in
g
noi
s
e
,
a
nd
im
pr
ov
in
g
ov
e
r
a
ll
c
ont
r
a
s
t
in
a
n
im
a
ge
[
20]
.
B
e
c
a
us
e
o
f
it
s
pr
a
c
ti
c
a
l
im
pl
e
m
e
nt
a
ti
on
a
nd
a
bi
li
ty
t
o
e
nha
n
c
e
f
e
a
t
ur
e
s
w
it
ho
ut
e
x
c
e
s
s
iv
e
l
y
a
m
p
li
f
yi
ng
n
oi
s
e
, C
L
A
H
E
is
w
id
e
ly
us
e
d
in
g
e
n
e
r
a
l
im
a
ge
e
nh
a
nc
e
m
e
n
t
ta
s
k
s
.
I
n
C
L
A
H
E
,
th
e
im
a
g
e
is
f
ir
s
t
di
vi
de
d
in
to
s
e
ve
r
a
l
s
u
b
-
r
e
gi
ons
,
a
f
te
r
w
hi
c
h
a
hi
s
t
ogr
a
m
i
s
g
e
ne
r
a
te
d
f
or
e
a
c
h
r
e
gi
on.
T
he
s
e
hi
s
to
gr
a
m
s
a
r
e
th
e
n
c
li
pp
e
d
ba
s
e
d
on
a
pr
e
d
e
f
in
e
d
th
r
e
s
h
ol
d
to
pr
e
v
e
nt
ove
r
-
e
nha
nc
e
m
e
nt
.
T
he
c
li
pp
e
d
pi
xe
l
c
ount
s
a
r
e
r
e
di
s
tr
ib
ut
e
d
e
ve
nl
y
a
c
r
os
s
a
l
l
in
te
n
s
it
y
l
e
ve
l
s
,
e
ns
ur
in
g
c
ont
r
ol
l
e
d
c
o
nt
r
a
s
t
im
pr
ove
m
e
nt
.
T
h
e
a
ve
r
a
ge
n
um
be
r
of
pi
x
e
ls
a
s
s
i
gne
d
t
o
e
a
c
h i
nt
e
ns
it
y l
e
ve
l
i
s
d
e
f
in
e
d i
n (
3)
.
=
−
X
−
(
3
)
r
e
pr
e
s
e
nt
s
th
e
m
e
a
n
pi
xe
l
c
ount
in
a
s
ub
-
im
a
ge
,
w
he
r
e
is
th
e
t
ot
a
l
num
be
r
of
gr
a
y
le
ve
ls
,
−
de
not
e
s
th
e
num
be
r
of
pi
xe
ls
a
lo
ng
th
e
X
-
a
xi
s
,
a
nd
−
c
or
r
e
s
ponds
to
th
e
num
be
r
of
pi
xe
ls
a
lo
ng
th
e
Y
-
a
xi
s
. U
s
in
g t
hi
s
va
lu
e
, t
h
e
c
li
p l
im
it
f
or
t
he
hi
s
to
gr
a
m
c
a
n be
de
te
r
m
in
e
d t
hr
ough (
4)
.
=
(
4)
H
e
r
e
,
is
t
he
c
li
p l
im
it
, w
hi
le
de
f
in
e
s
t
he
m
a
xi
m
um
a
ll
ow
a
bl
e
a
ve
r
a
ge
pi
xe
ls
f
or
e
a
c
h i
nt
e
ns
it
y l
e
ve
l
in
a
s
ub
-
im
a
ge
.
A
ny
hi
s
to
gr
a
m
bi
n
th
a
t
e
xc
e
e
ds
th
is
li
m
it
w
il
l
be
c
li
ppe
d.
T
he
c
li
ppe
d
pi
xe
l
c
ount
s
a
r
e
th
e
n
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
C
om
par
is
on of i
m
age
e
nhanc
e
m
e
nt
m
e
th
ods
f
or
p
r
at
ima the
ft
d
e
te
c
ti
on
…
(
M
ade
Sudar
m
a)
217
r
e
di
s
tr
ib
ut
e
d
uni
f
or
m
ly
a
c
r
os
s
a
ll
gr
a
y
le
ve
ls
,
w
it
h
th
e
nu
m
be
r
of
r
e
di
s
tr
ib
ut
e
d
pi
xe
ls
pe
r
le
ve
l
c
a
lc
ul
a
te
d
us
in
g (
5)
.
=
(
5)
I
n
th
is
c
ont
e
xt
,
M
s
pe
c
if
ie
s
th
e
a
r
e
a
s
iz
e
,
N
r
e
pr
e
s
e
nt
s
th
e
in
te
n
s
it
y
le
ve
l,
a
nd
\
a
lp
ha
i
s
th
e
c
li
p
f
a
c
to
r
th
a
t
c
ont
r
ol
s
th
e
d
e
gr
e
e
of
e
nh
a
nc
e
m
e
nt
w
it
hi
n
th
e
r
a
nge
of
0
to
100
[
19]
.
T
he
s
e
p
a
r
a
m
e
te
r
s
c
ol
le
c
ti
ve
ly
de
te
r
m
in
e
how
th
e
hi
s
to
gr
a
m
is
a
dj
us
te
d
dur
in
g
th
e
C
L
A
H
E
pr
oc
e
s
s
.
P
r
ope
r
tu
ni
ng
of
th
e
s
e
va
lu
e
s
e
na
bl
e
s
C
L
A
H
E
t
o e
nha
nc
e
c
ont
r
a
s
t
e
f
f
e
c
ti
ve
ly
w
hi
le
m
in
im
iz
in
g nois
e
a
m
pl
if
ic
a
ti
on.
3.6.
G
am
m
a
c
o
r
r
e
c
t
io
n
G
a
m
m
a
c
or
r
e
c
ti
on
a
dj
us
ts
th
e
r
e
la
ti
ons
hi
p
be
twe
e
n
in
put
a
nd
out
put
in
te
ns
it
y
va
lu
e
s
,
in
tr
oduc
in
g
a
non
-
li
ne
a
r
m
a
ppi
ng
be
twe
e
n
th
e
m
.
W
he
n
th
e
ga
m
m
a
va
lu
e
is
e
qua
l
to
1,
th
e
m
a
ppi
ng
be
ha
ve
s
li
ne
a
r
ly
.
A
ga
m
m
a
va
lu
e
be
lo
w
1
ge
ne
r
a
te
s
br
ig
ht
e
r
out
put
s
,
w
hi
le
va
lu
e
s
a
bove
1
r
e
s
ul
t
in
da
r
ke
r
out
put
s
[
21]
.
T
he
br
ig
ht
ne
s
s
of
e
a
c
h
pi
xe
l
is
th
e
r
e
f
o
r
e
de
te
r
m
in
e
d
by
ga
m
m
a
pa
r
a
m
e
te
r
(
γ
)
,
a
pos
it
iv
e
c
ons
ta
nt
.
I
f
γ
e
xc
e
e
ds
1,
th
e
im
a
ge
be
c
om
e
s
da
r
ke
r
;
if
γ
is
be
lo
w
1,
th
e
im
a
ge
a
ppe
a
r
s
br
ig
ht
e
r
.
T
hi
s
a
dj
us
tm
e
nt
is
of
te
n
us
e
d
to
c
ont
r
ol
im
a
ge
c
ont
r
a
s
t
[
22]
.
F
ig
ur
e
2
il
lu
s
tr
a
te
s
th
e
e
f
f
e
c
t
o
f
di
f
f
e
r
e
nt
ga
m
m
a
va
lu
e
s
,
s
how
in
g
γ
<
1
f
or
in
c
r
e
a
s
in
g
br
ig
ht
ne
s
s
,
γ
=
1
f
or
a
li
ne
a
r
r
e
s
pons
e
,
a
nd
γ
>
1
f
or
de
c
r
e
a
s
in
g
br
ig
ht
ne
s
s
.
T
hi
s
vi
s
ua
li
z
a
ti
on
he
lp
s
c
la
r
if
y t
he
non
-
li
ne
a
r
be
ha
vi
or
i
nt
r
oduc
e
d by ga
m
m
a
c
or
r
e
c
ti
on.
F
ig
ur
e
2. G
a
m
m
a
c
or
r
e
c
ti
on
3.7.
A
u
gm
e
n
t
at
io
n
T
he
a
ugm
e
nt
a
ti
on
pha
s
e
in
th
is
s
tu
dy
a
ppl
ie
s
va
r
io
us
te
c
hni
que
s
to
in
c
r
e
a
s
e
th
e
di
v
e
r
s
it
y
of
th
e
im
a
ge
da
ta
s
e
t
[
23]
.
T
he
r
e
s
c
a
li
ng
ope
r
a
ti
on
nor
m
a
li
z
e
s
pi
xe
l
va
lu
e
s
to
th
e
0
–
1
r
a
nge
,
e
na
bl
in
g
m
or
e
s
ta
bl
e
m
ode
l
tr
a
in
in
g.
R
ot
a
ti
on
a
dj
u
s
tm
e
nt
e
xpo
s
e
s
th
e
m
od
e
l
to
di
f
f
e
r
e
nt
im
a
ge
or
ie
nt
a
ti
ons
,
w
hi
le
z
oom
in
g
in
or
out
a
ll
ow
s
th
e
m
ode
l
to
le
a
r
n
obj
e
c
ts
of
va
r
yi
ng
s
iz
e
s
.
H
or
iz
ont
a
l
f
li
ppi
ng
a
dds
a
ddi
ti
ona
l
va
r
ia
ti
on
by
m
ir
r
or
in
g
th
e
im
a
ge
,
he
lp
in
g
r
e
duc
e
ove
r
f
it
t
in
g
r
is
ks
[
24]
.
T
he
da
ta
s
e
t
is
s
e
pa
r
a
te
d
in
to
tr
a
in
in
g,
va
li
da
ti
on,
a
nd
te
s
ti
ng
por
ti
ons
to
e
ns
ur
e
r
e
li
a
bl
e
m
ode
l
e
va
lu
a
ti
on.
T
o
f
ur
th
e
r
m
in
im
iz
e
ove
r
f
it
ti
ng,
tr
a
in
in
g
a
nd
te
s
ti
ng
f
ol
lo
w
a
s
tr
a
ti
f
ie
d
f
iv
e
-
f
o
ld
s
c
he
m
e
in
w
hi
c
h
s
pl
it
s
a
r
e
m
a
d
e
w
it
h
s
pa
ti
a
l
a
w
a
r
e
ne
s
s
to
pr
e
ve
nt
le
a
ka
g
e
be
twe
e
n
vi
s
ua
ll
y
s
im
il
a
r
s
c
e
n
e
s
.
H
ype
r
pa
r
a
m
e
te
r
s
a
r
e
tu
ne
d
e
xc
lu
s
iv
e
ly
on
th
e
tr
a
in
in
g
f
ol
ds
w
it
hi
n
th
is
ne
s
te
d
pr
oc
e
s
s
.
C
onf
id
e
nc
e
in
te
r
va
ls
a
t
th
e
95
pe
r
c
e
nt
le
ve
l
a
r
e
c
om
put
e
d
us
in
g
nonpa
r
a
m
e
tr
ic
boot
s
tr
a
ppi
ng
on
pe
r
-
im
a
ge
pr
e
di
c
ti
ons
to
a
s
s
e
s
s
e
s
ti
m
a
to
r
s
ta
bi
li
ty
.
O
ve
r
a
l
l,
th
e
s
e
a
ugm
e
nt
a
ti
on
s
tr
a
te
gi
e
s
in
c
r
e
a
s
e
th
e
va
r
ia
bi
li
ty
of
t
he
t
r
a
in
in
g da
ta
a
nd c
ont
r
ib
ut
e
t
o hi
ghe
r
m
ode
l
a
c
c
ur
a
c
y
[
25]
.
3.8.
C
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
A
C
N
N
is
a
ty
pe
of
m
ul
ti
-
la
ye
r
pe
r
c
e
pt
r
on
(
M
L
P
)
de
s
ig
ne
d
s
pe
c
if
ic
a
ll
y
f
or
pr
oc
e
s
s
in
g
two
-
di
m
e
ns
io
na
l
im
a
ge
d
a
ta
[
26]
–
[
28]
.
C
N
N
s
im
it
a
te
th
e
vi
s
ua
l
m
e
c
ha
ni
s
m
of
th
e
hum
a
n
br
a
in
,
e
na
bl
in
g
c
om
put
e
r
s
to
id
e
nt
if
y
a
nd
di
s
ti
ngui
s
h
obj
e
c
ts
th
r
ough
a
pr
oc
e
s
s
known
a
s
im
a
ge
r
e
c
ogni
ti
on.
A
s
a
de
e
p
le
a
r
ni
ng
c
la
s
s
if
ic
a
ti
on
m
ode
l,
C
N
N
us
e
s
c
onvolut
io
na
l
la
ye
r
s
to
a
ppl
y
f
i
lt
e
r
s
to
in
pu
t
da
ta
.
S
im
il
a
r
to
ot
he
r
ne
ur
a
l
ne
twor
k
a
r
c
hi
te
c
tu
r
e
s
,
C
N
N
s
c
on
s
is
t
of
ne
ur
ons
e
q
ui
ppe
d
w
it
h
w
e
ig
ht
s
,
bi
a
s
e
s
,
a
nd
a
c
ti
va
ti
on
f
unc
ti
ons
.
T
he
ir
tr
a
in
in
g
r
e
li
e
s
on
ba
c
kpr
opa
ga
ti
on,
w
hi
le
f
or
w
a
r
d
pr
opa
ga
ti
on
is
us
e
d
dur
in
g
c
la
s
s
if
ic
a
ti
on
[
29]
.
A
ty
pi
c
a
l
C
N
N
is
c
om
pos
e
d
of
th
r
e
e
pr
im
a
r
y
la
ye
r
s
:
c
onvolut
io
na
l
la
ye
r
s
,
pool
in
g
la
ye
r
s
,
a
nd
f
ul
ly
c
onne
c
te
d
la
ye
r
s
.
T
he
c
onvolut
io
na
l
la
ye
r
s
e
r
ve
s
a
s
th
e
c
or
e
of
C
N
N
a
r
c
hi
te
c
tu
r
e
.
I
n
th
is
la
ye
r
,
f
il
te
r
s
—
c
om
m
onl
y
3×
3
in
s
iz
e
—
s
li
de
ove
r
th
e
in
put
to
e
xt
r
a
c
t
e
s
s
e
nt
i
a
l
im
a
ge
f
e
a
tu
r
e
s
.
T
h
e
s
e
f
il
te
r
s
c
a
pt
ur
e
s
pa
ti
a
l
r
e
la
ti
ons
hi
ps
be
twe
e
n
ne
ig
hbor
in
g
pi
xe
ls
a
nd
c
a
n
ge
ne
r
a
te
e
f
f
e
c
ts
s
uc
h
a
s
e
dge
de
te
c
ti
on,
bl
ur
r
in
g,
a
n
d
s
ha
r
pe
ni
ng.
T
he
s
tr
id
e
pa
r
a
m
e
te
r
de
te
r
m
in
e
s
how
m
a
ny
pi
xe
ls
th
e
f
il
te
r
m
ove
s
a
t
e
a
c
h
s
te
p;
f
or
e
xa
m
pl
e
,
a
s
tr
id
e
of
1 s
hi
f
ts
th
e
f
il
te
r
by one
p
ix
e
l
[
30]
.
P
a
ddi
ng i
s
a
ppl
ie
d
w
he
n t
he
f
il
te
r
d
im
e
ns
io
ns
do not pe
r
f
e
c
tl
y
f
it
th
e
i
nput
. A
n i
ll
us
tr
a
ti
on of
a
c
onvolut
io
na
l
la
ye
r
i
s
pr
ovi
de
d i
n F
ig
ur
e
3.
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
.
15
, N
o.
1
,
F
e
br
ua
r
y
20
26
:
213
-
228
218
F
ig
ur
e
3. C
onvolut
io
na
l
la
ye
r
T
he
f
or
m
ul
a
us
e
d
to
de
te
r
m
in
e
th
e
out
put
s
iz
e
of
th
e
f
e
a
tu
r
e
m
a
p
in
a
c
onvolut
io
na
l
la
ye
r
is
s
how
n
in
(
6)
.
I
t
c
a
lc
ul
a
te
s
th
e
r
e
s
ul
ti
ng
di
m
e
n
s
io
ns
a
f
te
r
c
onvolut
io
n
by
c
ons
id
e
r
in
g
th
e
in
put
s
iz
e
,
ke
r
ne
l
s
iz
e
,
s
tr
id
e
, a
nd pa
ddi
ng
.
(
,
ℎ
)
=
[
+
2
−
s
]
+
1
(
6)
W
he
r
e
(
,
ℎ
)
is
r
e
s
ul
ti
ng
out
put
s
iz
e
;
is
K
e
r
ne
l
s
iz
e
;
is
s
iz
e
of
th
e
s
tr
id
e
;
is
s
iz
e
of
th
e
pa
ddi
ng
;
a
nd
is
v
a
lu
e
of
t
he
i
nput
i
m
a
ge
s
iz
e
.
A
c
ti
va
ti
on
f
unc
ti
ons
a
r
e
a
ppl
ie
d
im
m
e
di
a
te
ly
a
f
te
r
th
e
c
onvolut
io
n
ope
r
a
ti
on
to
in
tr
oduc
e
non
-
li
ne
a
r
it
y.
A
m
ong
th
e
a
va
il
a
bl
e
a
c
ti
va
ti
on
f
unc
ti
ons
,
th
e
r
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
is
th
e
m
os
t
f
r
e
que
nt
ly
us
e
d
in
C
N
N
m
ode
ls
due
to
it
s
a
bi
li
ty
to
m
in
im
iz
e
e
r
r
or
s
a
nd
a
voi
d
s
a
tu
r
a
ti
on.
R
e
L
U
is
w
id
e
ly
im
pl
e
m
e
nt
e
d a
c
r
os
s
hi
dde
n l
a
y
e
r
s
be
c
a
us
e
of
i
ts
e
f
f
ic
ie
nc
y.
T
he
R
e
L
U
f
unc
ti
on i
s
de
f
in
e
d i
n (
7)
.
(
)
=
{
,
>
0
0
,
≤
0
(
7)
R
e
L
U
out
put
s
z
e
r
o
f
or
ne
ga
ti
ve
in
put
va
lu
e
s
a
nd
r
e
tu
r
ns
th
e
in
put
it
s
e
lf
w
he
n
th
e
va
lu
e
is
pos
it
iv
e
.
F
ol
lo
w
in
g
th
e
a
c
ti
va
ti
on
s
ta
ge
,
th
e
pool
in
g
la
ye
r
is
us
e
d
to
r
e
d
uc
e
th
e
s
pa
ti
a
l
di
m
e
ns
io
ns
of
th
e
f
e
a
tu
r
e
m
a
p
pr
oduc
e
d
by
c
onvolut
io
n
[
31]
.
T
h
e
two
c
om
m
on
pool
in
g
te
c
h
ni
que
s
a
r
e
m
a
x
pool
in
g
a
nd
a
ve
r
a
ge
pool
in
g.
I
n
m
a
x
pool
in
g,
th
e
f
e
a
tu
r
e
m
a
p
is
pa
r
ti
ti
one
d
in
to
s
m
a
ll
r
e
g
io
ns
,
a
nd
th
e
hi
ghe
s
t
va
lu
e
f
r
om
e
a
c
h
r
e
gi
on
is
s
e
le
c
te
d
to
f
or
m
a
dow
ns
a
m
pl
e
d
out
put
[
24]
,
[
32]
.
T
hi
s
pr
oc
e
s
s
he
lp
s
lo
w
e
r
di
m
e
ns
io
na
li
ty
a
nd
r
e
m
ove
le
s
s
r
e
le
va
nt
de
ta
il
s
w
hi
le
pr
e
s
e
r
vi
ng
c
r
it
ic
a
l
f
e
a
tu
r
e
s
[
28]
.
A
n
il
lu
s
tr
a
ti
on
o
f
th
e
pool
in
g
la
ye
r
is
pr
ovi
de
d
in
F
ig
ur
e
4.
F
ig
ur
e
4. M
a
x
pool
in
g
la
ye
r
T
he
f
ol
lo
w
in
g i
s
t
he
f
or
m
ul
a
us
e
d i
n
m
a
x pooli
ng
, w
hi
c
h c
a
n b
e
s
e
e
n i
n (
8)
.
(
,
ℎ
)
=
(
(
,
ℎ
)
−
1
−
)
+
1
(
8)
W
he
r
e
(
,
ℎ
)
is
r
e
s
ul
ti
ng
s
iz
e
of
he
ig
ht
a
nd
w
id
th
;
(
,
ℎ
)
is
pr
e
vi
ous
w
e
ig
ht
a
nd
he
ig
ht
s
iz
e
;
is
s
iz
e
of
th
e
s
tr
id
e
;
a
nd
is
s
iz
e
of
t
he
ke
r
ne
l
.
3.9.
A
le
xN
e
t
I
n
th
is
s
tu
dy,
A
le
xN
e
t
is
e
m
pl
oye
d
dur
in
g
th
e
m
ode
l
de
s
ig
n
s
t
a
ge
to
pr
oc
e
s
s
im
a
ge
s
th
a
t
ha
ve
be
e
n
pr
e
-
pr
oc
e
s
s
e
d
be
f
or
e
ha
nd
[
33]
–
[
35]
.
T
he
A
le
xN
e
t
a
r
c
hi
te
c
tu
r
e
in
c
lu
de
s
f
iv
e
c
onvolut
io
na
l
la
ye
r
s
a
nd
th
r
e
e
f
ul
ly
c
onne
c
te
d
la
ye
r
s
.
T
o
h
e
lp
r
e
duc
e
ov
e
r
f
it
ti
ng,
f
iv
e
dr
opout
la
ye
r
s
a
r
e
in
c
or
por
a
te
d
a
f
te
r
s
e
ve
r
a
l
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
C
om
par
is
on of i
m
age
e
nhanc
e
m
e
nt
m
e
th
ods
f
or
p
r
at
ima the
ft
d
e
te
c
ti
on
…
(
M
ade
Sudar
m
a)
219
c
onvolut
io
na
l
a
nd
f
ul
ly
c
onne
c
te
d
s
ta
ge
s
,
w
he
r
e
uni
t
s
a
r
e
r
a
ndoml
y
de
a
c
ti
va
te
d.
P
ool
in
g
la
ye
r
s
a
r
e
a
ls
o
a
ppl
ie
d
f
ol
lo
w
in
g
c
e
r
ta
in
c
onvolut
io
na
l
la
y
e
r
s
to
d
e
c
r
e
a
s
e
s
pa
ti
a
l
di
m
e
ns
io
ns
a
nd
e
xt
r
a
c
t
e
s
s
e
nt
ia
l
f
e
a
tu
r
e
s
.
T
he
f
in
a
l
pr
e
di
c
ti
ons
a
r
e
pr
oduc
e
d
th
r
ough
de
ns
e
la
ye
r
s
a
t
th
e
out
put
s
ta
ge
.
T
hi
s
a
r
c
hi
te
c
tu
r
e
e
na
bl
e
s
A
le
xN
e
t
to
e
f
f
e
c
ti
ve
ly
c
a
pt
ur
e
im
por
ta
nt
vi
s
ua
l
p
a
tt
e
r
ns
w
hi
l
e
im
pr
ovi
ng
ge
ne
r
a
li
z
a
ti
on
th
r
ough
dr
opout,
le
a
di
ng t
o m
or
e
r
e
li
a
bl
e
pr
e
di
c
ti
on outc
om
e
s
.
3.10.
V
G
G
-
19
I
n
th
is
s
tu
dy,
V
G
G
-
19
is
a
ppl
ie
d
in
th
e
m
ode
l
de
s
ig
n
pha
s
e
to
pr
oc
e
s
s
im
a
ge
s
th
a
t
ha
ve
unde
r
gone
pr
e
-
pr
oc
e
s
s
in
g.
T
he
V
G
G
-
19
a
r
c
hi
te
c
tu
r
e
is
c
om
pos
e
d
of
1
6
c
onvolut
io
na
l
la
ye
r
s
f
ol
lo
w
e
d
by
3
f
ul
ly
c
onne
c
te
d
la
ye
r
s
.
T
o
m
it
ig
a
te
ove
r
f
it
ti
ng,
dr
opout
la
ye
r
s
a
r
e
i
ns
e
r
te
d
a
f
te
r
s
e
ve
r
a
l
c
onvolut
io
na
l
a
nd
de
ns
e
s
ta
ge
s
,
w
h
e
r
e
s
e
le
c
te
d
uni
ts
a
r
e
te
m
por
a
r
il
y
di
s
a
bl
e
d
dur
in
g
tr
a
in
in
g.
T
he
ne
twor
k
e
m
pl
oys
s
m
a
ll
3
×
3
c
onvolut
io
na
l
f
il
te
r
s
w
it
h
a
s
tr
id
e
of
1
to
e
xt
r
a
c
t
de
ta
il
e
d
f
e
a
tu
r
e
s
f
r
om
in
put
im
a
ge
s
.
P
ool
in
g
ope
r
a
ti
ons
a
r
e
pl
a
c
e
d
a
f
te
r
s
pe
c
if
ic
c
onvolut
io
na
l
bl
oc
ks
to
pr
ogr
e
s
s
iv
e
ly
r
e
duc
e
s
pa
ti
a
l
di
m
e
ns
io
ns
w
hi
le
pr
e
s
e
r
vi
ng
e
s
s
e
nt
ia
l
p
a
tt
e
r
ns
[
33]
–
[
35]
.
F
in
a
l
pr
e
di
c
ti
ons
a
r
e
ge
ne
r
a
te
d
us
i
ng
de
ns
e
la
ye
r
s
a
t
th
e
out
put
s
ta
ge
. T
hi
s
s
e
tu
p
e
na
bl
e
s
V
G
G
-
19
to
f
oc
us
on
r
e
le
va
nt
vi
s
ua
l
s
tr
uc
tu
r
e
s
w
hi
le
be
ne
f
it
in
g
f
r
om
dr
opout
to
e
nh
a
nc
e
ge
ne
r
a
li
z
a
ti
on
a
nd
im
pr
ove
pr
e
di
c
ti
v
e
pe
r
f
or
m
a
nc
e
.
V
G
G
-
19
i
s
w
id
e
ly
u
s
e
d
a
s
a
r
e
f
e
r
e
nc
e
b
a
c
kbone
due
to
it
s
s
ta
bl
e
pe
r
f
or
m
a
nc
e
a
nd
r
e
pr
oduc
ib
il
it
y
a
c
r
os
s
di
f
f
e
r
e
nt
m
a
c
hi
ne
le
a
r
ni
ng
f
r
a
m
e
w
or
ks
.
I
ts
w
e
ll
-
e
s
ta
bl
is
he
d
a
r
c
hi
te
c
tu
r
e
a
ll
ow
s
f
or
c
ons
is
te
nt
c
om
pa
r
is
on
in
e
nha
nc
e
m
e
nt
a
nd
r
e
c
ogni
ti
on
ta
s
ks
,
pa
r
ti
c
ul
a
r
ly
unde
r
lo
w
-
li
ght
c
ondi
ti
ons
.
D
e
s
pi
te
th
e
e
xi
s
te
nc
e
of
m
or
e
r
e
c
e
nt
m
ode
ls
w
it
h
hi
ghe
r
pe
a
k
a
c
c
ur
a
c
y,
V
G
G
-
19
r
e
m
a
in
s
s
ui
ta
bl
e
f
or
s
c
e
n
a
r
io
s
r
e
qui
r
in
g
pr
e
di
c
ta
bl
e
la
te
nc
y
a
n
d
e
f
f
ic
ie
nt
r
e
s
our
c
e
u
s
a
ge
,
m
a
ki
ng
it
pr
a
c
ti
c
a
l
f
or
de
pl
oym
e
nt
s
w
it
h l
im
it
e
d c
om
put
a
ti
ona
l
c
a
pa
bi
li
ti
e
s
.
3.11.
I
m
age
q
u
al
it
y as
s
e
s
s
m
e
n
t
I
m
a
ge
qua
li
ty
a
s
s
e
s
s
m
e
nt
(
I
Q
A
)
in
th
is
s
tu
dy
ut
il
iz
e
s
m
e
a
n
s
qua
r
e
d
e
r
r
or
(
M
S
E
)
a
nd
P
S
N
R
to
m
e
a
s
ur
e
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
m
e
th
od
s
[
36]
.
M
S
E
qua
nt
if
ie
s
th
e
a
ve
r
a
ge
s
qua
r
e
d
di
f
f
e
r
e
nc
e
be
twe
e
n
th
e
in
put
im
a
ge
a
nd
th
e
e
nha
n
c
e
d
r
e
s
ul
t
b
y
e
va
lu
a
ti
ng
pi
xe
l
in
te
ns
it
ie
s
a
t
c
or
r
e
s
ponding
pos
it
io
ns
[
37]
. T
he
M
S
E
va
lu
e
i
s
c
om
put
e
d u
s
in
g (
9)
.
M
S
E
=
1
∑
∑
(
(
−
)
−
(
−
)
)
2
=
1
=
0
−
1
=
0
(
9)
I
n
th
is
e
xpr
e
s
s
io
n,
a
nd
de
not
e
th
e
or
ig
in
a
l
a
nd
e
nha
nc
e
d
im
a
ge
s
,
w
hi
le
a
nd
r
e
pr
e
s
e
nt
pi
xe
l
c
oor
di
na
te
s
.
in
di
c
a
te
s
th
e
to
ta
l
num
be
r
of
pi
xe
ls
.
P
S
N
R
,
on
th
e
ot
he
r
ha
nd,
c
om
pa
r
e
s
th
e
m
a
xi
m
um
pos
s
ib
le
s
ig
na
l
s
tr
e
ngt
h
to
th
e
a
m
ount
of
noi
s
e
in
tr
oduc
e
d
dur
i
ng
pr
oc
e
s
s
in
g
[
38]
.
I
t
is
w
id
e
ly
u
s
e
d
to
a
s
s
e
s
s
th
e
f
id
e
li
ty
of
e
nha
nc
e
d
im
a
ge
s
;
hi
ghe
r
P
S
N
R
va
lu
e
s
c
or
r
e
s
pond
to
be
tt
e
r
vi
s
ua
l
qua
li
ty
[
39]
.
P
S
N
R
is
c
a
lc
ul
a
te
d us
in
g (
10)
.
=
10
10
(
(
−
1
)
2
)
(
10)
3.12.
M
od
e
l
e
val
u
at
io
n
T
he
s
y
s
te
m
’
s
pe
r
f
or
m
a
nc
e
is
a
s
s
e
s
s
e
d
us
in
g
a
c
onf
us
io
n
m
a
tr
ix
,
w
hi
c
h
s
e
r
ve
s
a
s
a
s
ta
nda
r
d
to
ol
f
or
e
va
lu
a
ti
ng
c
la
s
s
if
ic
a
ti
on
m
ode
ls
[
40]
.
I
t
s
um
m
a
r
iz
e
s
how
w
e
ll
th
e
m
ode
l
di
s
ti
ngui
s
he
s
be
twe
e
n
c
la
s
s
e
s
by
di
s
pl
a
yi
ng
th
e
c
ount
s
of
c
or
r
e
c
tl
y
a
nd
in
c
or
r
e
c
tl
y
pr
e
di
c
te
d
s
a
m
pl
e
s
.
T
hi
s
a
ll
ow
s
a
de
t
a
il
e
d
e
xa
m
in
a
ti
on
of
m
ode
l
be
ha
vi
or
a
c
r
os
s
di
f
f
e
r
e
nt
c
a
te
gor
ie
s
[
41]
.
T
he
c
onf
us
io
n
m
a
tr
ix
c
ont
a
in
s
in
f
or
m
a
ti
on
a
bout
bot
h
th
e
a
c
tu
a
l
la
be
ls
a
nd
th
e
m
ode
l’
s
pr
e
di
c
ti
ons
,
or
ga
ni
z
e
d
in
to
f
our
c
e
ll
s
r
e
pr
e
s
e
nt
in
g
tr
ue
a
nd
f
a
ls
e
out
c
om
e
s
f
or
e
a
c
h
c
la
s
s
.
F
r
om
th
is
m
a
tr
ix
,
s
e
ve
r
a
l
e
va
lu
a
ti
on
m
e
tr
ic
s
c
a
n
be
de
r
iv
e
d,
in
c
lu
di
ng
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd F
1
-
s
c
or
e
, pr
ovi
di
ng a
c
om
pr
e
he
ns
iv
e
vi
e
w
of
t
he
m
ode
l’
s
c
la
s
s
if
ic
a
ti
on pe
r
f
or
m
a
nc
e
.
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
4.1.
I
m
age
i
m
p
r
ove
m
e
n
t
r
e
s
u
lt
s
I
n
obj
e
c
t
de
te
c
ti
on,
pa
r
ti
c
ul
a
r
ly
in
hum
a
n
de
te
c
ti
on,
v
a
r
io
us
m
e
th
ods
a
r
e
u
s
e
d
to
e
nha
nc
e
a
c
c
ur
a
c
y
a
nd
c
la
r
it
y
in
de
te
c
ti
ng
hum
a
n
obj
e
c
t
s
in
di
gi
ta
l
im
a
g
e
s
.
T
h
e
m
e
th
ods
c
om
m
onl
y
e
m
pl
oye
d
in
c
lu
de
H
E
,
A
H
E
,
C
L
A
H
E
,
a
nd
ga
m
m
a
c
or
r
e
c
ti
on.
E
a
c
h
m
e
th
od
ha
s
it
s
ow
n
c
ha
r
a
c
te
r
is
ti
c
s
a
nd
a
dva
nt
a
g
e
s
in
im
pr
ovi
ng
im
a
ge
qua
li
ty
to
a
s
s
is
t
hum
a
n
de
te
c
ti
on
a
lg
or
it
hm
s
.
I
n
th
e
in
i
ti
a
l
s
ta
ge
of
th
e
s
tu
dy,
th
e
A
H
E
m
e
th
od
w
a
s
a
ppl
ie
d
to
e
nha
nc
e
im
a
g
e
qua
li
ty
.
F
ig
ur
e
5
pr
e
s
e
nt
s
th
e
e
nha
nc
e
m
e
nt
r
e
s
ul
ts
,
d
e
pi
c
te
d
th
r
ough
hi
s
to
gr
a
m
s
a
nd
a
c
om
pa
r
is
on
of
th
e
L
-
c
ha
nne
l
be
twe
e
n
th
e
or
ig
in
a
l
a
nd
A
H
E
-
pr
oc
e
s
s
e
d
im
a
ge
s
,
a
c
c
om
pa
ni
e
d
by
th
e
ir
pi
xe
l
in
te
ns
it
y di
s
tr
ib
ut
io
n.
T
he
r
e
s
ul
ts
of
a
d
a
pt
iv
e
A
H
E
a
da
pt
iv
e
ly
a
dj
us
t
th
e
c
ont
r
a
s
t
i
n
lo
c
a
l
a
r
e
a
s
.
I
ts
hi
s
to
gr
a
m
s
how
s
a
m
or
e
uni
f
or
m
in
te
ns
it
y
di
s
tr
ib
ut
io
n,
w
it
h
s
e
ve
r
a
l
s
ig
ni
f
ic
a
nt
pe
a
ks
,
a
s
s
how
n
in
F
ig
ur
e
5.
T
hi
s
in
di
c
a
te
s
th
a
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
.
15
, N
o.
1
,
F
e
br
ua
r
y
20
26
:
213
-
228
220
A
H
E
e
nha
nc
e
s
lo
c
a
l
c
ont
r
a
s
t
in
a
m
or
e
nua
nc
e
d
w
a
y
c
om
pa
r
e
d
to
tr
a
di
ti
ona
l
H
E
,
m
a
ki
ng
de
ta
il
s
in
va
r
io
us
pa
r
ts
of
th
e
im
a
ge
c
le
a
r
e
r
a
nd
m
or
e
vi
s
ib
le
.
S
im
il
a
r
ly
,
F
ig
ur
e
6
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
of
C
L
A
H
E
,
w
hi
c
h
li
m
it
s
th
e
de
gr
e
e
of
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
t
o pr
e
ve
nt
e
xc
e
s
s
iv
e
noi
s
e
a
m
pl
if
ic
a
ti
on.
A
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
6,
th
e
hi
s
to
gr
a
m
s
how
s
s
ig
ni
f
ic
a
nt
pe
a
ks
a
t
a
n
in
te
n
s
it
y
a
r
ound
200,
w
it
h
a
pi
xe
l
c
ount
of
a
ppr
oxi
m
a
te
ly
12,000.
T
hi
s
in
di
c
a
te
s
th
a
t
t
he
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
is
pe
r
f
or
m
e
d
in
a
c
ont
r
ol
le
d
m
a
nne
r
,
w
hi
c
h
he
lp
s
to
r
e
duc
e
noi
s
e
th
a
t
m
a
y
a
r
is
e
f
r
om
e
xc
e
s
s
iv
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
.
T
he
s
ubs
e
que
nt
s
t
a
ge
a
ppl
ie
s
th
e
ga
m
m
a
c
or
r
e
c
ti
on
m
e
th
od
to
e
nha
nc
e
im
a
ge
qua
li
ty
.
T
he
r
e
s
ul
t
s
of
th
is
im
pl
e
m
e
nt
a
ti
on a
r
e
pr
e
s
e
nt
e
d i
n F
ig
ur
e
7.
F
ig
ur
e
5.
A
da
pt
iv
e
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
F
ig
ur
e
6. 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
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
C
om
par
is
on of i
m
age
e
nhanc
e
m
e
nt
m
e
th
ods
f
or
p
r
at
ima the
ft
d
e
te
c
ti
on
…
(
M
ade
Sudar
m
a)
221
F
ig
ur
e
7
pr
e
s
e
nt
s
th
e
im
a
ge
r
e
s
ul
ti
ng
f
r
om
th
e
a
ppl
ic
a
ti
on
of
ga
m
m
a
c
or
r
e
c
ti
on
,
w
hi
c
h
pe
r
f
or
m
s
a
non
-
li
ne
a
r
a
dj
us
tm
e
nt
of
im
a
ge
br
ig
ht
ne
s
s
.
T
he
hi
s
to
gr
a
m
e
xhi
bi
ts
s
ig
ni
f
ic
a
nt
pe
a
ks
a
t
a
n
in
te
ns
it
y
of
a
ppr
oxi
m
a
te
ly
200,
w
it
h
a
pi
xe
l
c
ount
of
a
r
ound
17,500,
c
lo
s
e
ly
r
e
s
e
m
bl
in
g
th
e
or
ig
in
a
l
im
a
ge
but
w
it
h
a
s
li
ght
ly
m
odi
f
ie
d
in
te
ns
it
y
di
s
tr
ib
ut
io
n.
T
he
f
in
a
l
e
xpe
r
im
e
nt
e
m
pl
oys
th
e
HE
m
e
th
od
to
im
p
r
ove
im
a
ge
qua
li
ty
.
F
ig
ur
e
8
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
of
a
ppl
yi
ng
th
is
m
e
th
od,
il
lu
s
tr
a
te
d
th
r
ough
th
e
hi
s
to
gr
a
m
,
or
ig
in
a
l
L
-
c
ha
nne
l,
a
nd t
he
pr
oc
e
s
s
e
d L
-
c
ha
nne
l,
a
s
w
e
ll
a
s
t
he
pi
xe
l
in
t
e
ns
it
y di
s
tr
ib
ut
io
n.
T
he
r
e
s
ul
ts
of
H
E
,
a
s
s
how
n
in
F
ig
ur
e
8,
de
m
ons
tr
a
te
a
not
ic
e
a
bl
e
e
nh
a
nc
e
m
e
nt
in
c
ont
r
a
s
t
a
c
r
os
s
th
e
e
nt
ir
e
im
a
ge
.
T
he
hi
s
to
gr
a
m
di
s
pl
a
ys
a
m
or
e
uni
f
or
m
in
te
ns
it
y
di
s
tr
ib
ut
io
n
w
it
h
pe
r
io
di
c
pe
a
ks
r
a
ngi
ng
f
r
om
0
to
250.
T
hi
s
di
s
tr
ib
ut
io
n
in
di
c
a
te
s
a
s
ig
ni
f
ic
a
nt
im
pr
ov
e
m
e
nt
in
ove
r
a
ll
c
ont
r
a
s
t,
a
ll
ow
in
g
p
r
e
vi
ous
ly
in
di
s
ti
nc
t
de
ta
il
s
to
be
c
om
e
m
or
e
pr
om
in
e
nt
a
nd
vi
s
ua
ll
y
di
s
c
e
r
ni
bl
e
.
T
he
pr
oc
e
s
s
e
d
L
-
c
ha
nne
l
f
ur
th
e
r
il
lu
s
tr
a
te
s
th
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
,
w
hi
le
th
e
pi
xe
l
in
te
ns
it
y
di
s
tr
ib
ut
io
n
c
onf
ir
m
s
th
e
br
oa
d
e
r
s
pr
e
a
d
of
in
te
ns
it
y va
lu
e
s
c
om
pa
r
e
d t
o t
he
or
ig
in
a
l
im
a
ge
.
F
ig
ur
e
7.
G
a
m
m
a
c
or
r
e
c
ti
on
F
ig
ur
e
8. H
is
to
gr
a
m
e
qua
li
z
a
ti
on
4.2.
R
e
s
u
lt
s
of
i
m
age
e
n
h
an
c
e
m
e
n
t
e
val
u
at
io
n
I
n
th
is
s
tu
dy,
s
e
ve
r
a
l
im
a
ge
pr
oc
e
s
s
in
g
m
e
th
ods
w
e
r
e
a
ppl
ie
d
to
im
pr
ove
th
e
im
a
ge
qua
li
ty
of
a
s
im
ul
a
te
d
phot
o
th
e
f
t
e
xpe
r
im
e
nt
,
na
m
e
ly
H
E
,
A
H
E
,
C
L
A
H
E
,
a
nd
ga
m
m
a
c
or
r
e
c
ti
on.
H
E
is
us
e
d
to
e
nha
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
.
15
, N
o.
1
,
F
e
br
ua
r
y
20
26
:
213
-
228
222
th
e
ove
r
a
ll
c
ont
r
a
s
t
of
th
e
im
a
ge
by
di
s
tr
ib
ut
in
g
pi
xe
l
in
te
ns
it
ie
s
m
or
e
e
ve
nl
y.
A
H
E
,
on
th
e
ot
he
r
ha
nd,
a
dj
us
ts
t
he
c
ont
r
a
s
t
lo
c
a
ll
y, a
ll
ow
in
g f
or
de
ta
il
e
nha
nc
e
m
e
nt
i
n
s
pe
c
if
ic
a
r
e
a
s
of
t
he
i
m
a
ge
. C
L
A
H
E
is
s
im
il
a
r
to
A
H
E
but
li
m
it
s
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
to
r
e
du
c
e
pot
e
nt
i
a
l
noi
s
e
a
m
pl
if
ic
a
ti
on.
M
e
a
nw
hi
le
,
ga
m
m
a
c
or
r
e
c
ti
on
m
odi
f
ie
s
th
e
br
ig
ht
ne
s
s
of
th
e
im
a
ge
non
-
li
ne
a
r
ly
,
p
r
ovi
di
ng
a
m
or
e
na
tu
r
a
l
br
ig
ht
ne
s
s
a
dj
u
s
tm
e
nt
w
it
hout
dr
a
s
ti
c
a
ll
y
c
ha
ngi
ng
th
e
c
ont
r
a
s
t
r
a
ti
o
be
twe
e
n
pi
xe
ls
.
T
he
e
va
lu
a
ti
on
of
e
a
c
h
m
e
th
od
is
c
onduc
te
d
us
in
g
two
pr
im
a
r
y
m
e
tr
ic
s
:
M
S
E
a
nd
P
S
N
R
.
T
a
bl
e
4
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
of
th
e
im
a
ge
e
nha
nc
e
m
e
nt
e
va
lu
a
ti
on,
c
om
pa
r
in
g
th
e
pe
r
f
or
m
a
nc
e
of
e
a
c
h
m
e
th
od
ba
s
e
d
on
th
e
s
e
two
m
e
tr
ic
s
.
T
he
s
e
m
e
tr
ic
s
w
e
r
e
c
hos
e
n
due
to
th
e
ir
w
id
e
a
dopt
io
n
in
im
a
ge
pr
oc
e
s
s
in
g
r
e
s
e
a
r
c
h
a
nd
th
e
ir
a
bi
li
ty
to
pr
ovi
de
bot
h
qua
nt
it
a
ti
ve
a
nd
pe
r
c
e
pt
ua
l
in
s
ig
ht
s
in
to
im
a
ge
qua
li
ty
.
T
he
r
e
s
ul
ts
f
r
om
T
a
bl
e
4
f
or
m
th
e
ba
s
is
f
or
de
te
r
m
in
in
g
w
hi
c
h
e
nha
nc
e
m
e
nt
m
e
th
od of
f
e
r
s
t
he
m
os
t
e
f
f
e
c
ti
ve
ba
l
a
nc
e
be
tw
e
e
n
de
ta
il
pr
e
s
e
r
va
ti
on a
nd nois
e
r
e
duc
ti
on.
F
r
om
T
a
bl
e
4
,
H
E
s
h
ow
s
h
ig
h
e
r
r
or
a
nd
p
oor
qu
a
li
ty
(
M
S
E
1
2,2
49.
74;
P
S
N
R
7
.6
4
d
B
)
,
i
ndi
c
a
ti
ng
not
i
c
e
a
bl
e
n
oi
s
e
a
nd
a
r
ti
f
a
c
ts
.
A
H
E
pe
r
f
or
m
s
w
or
s
t
(
M
S
E
3
5,1
27.
59;
P
S
N
R
2.
79
d
B
)
,
r
e
f
le
c
ti
ng
s
e
v
e
r
e
di
s
to
r
ti
o
n
a
n
d
w
e
a
k
s
im
il
a
r
it
y
t
o
th
e
or
i
gi
n
a
l
.
I
n
c
o
nt
r
a
s
t,
C
L
A
H
E
d
e
li
ve
r
s
th
e
b
e
s
t
f
id
e
li
ty
a
nd
c
l
a
r
i
ty
w
i
th
th
e
lo
w
e
s
t
M
S
E
(
21
.16
)
a
n
d
th
e
hi
gh
e
s
t
P
S
N
R
(
3
8.1
32
d
B
)
,
pr
e
s
e
r
vi
n
g
de
ta
il
w
h
il
e
e
n
h
a
n
c
in
g
c
o
nt
r
a
s
t.
G
a
m
m
a
c
or
r
e
c
ti
o
n
i
s
a
c
c
e
pt
a
b
le
b
ut
c
le
a
r
l
y
be
hi
n
d
C
L
A
H
E
(
M
S
E
233
.
13;
P
S
N
R
2
6.9
7
dB
)
.
O
v
e
r
a
ll
,
C
L
A
H
E
i
s
th
e
m
o
s
t
s
ui
t
a
b
le
e
nh
a
n
c
e
m
e
nt
m
e
th
od
f
or
l
ow
-
li
g
ht
im
a
g
e
s
,
w
hi
l
e
H
E
a
n
d A
H
E
t
e
nd
t
o i
nt
r
od
uc
e
s
ub
s
t
a
n
ti
a
l
no
is
e
a
nd
d
is
to
r
t
io
n. F
ig
ur
e
9
vi
s
u
a
l
ly
c
o
r
r
ob
or
a
te
s
t
h
e
s
e
q
ua
nt
i
ta
ti
v
e
c
om
p
a
r
i
s
o
n
s
a
c
r
o
s
s
m
e
th
od
s
.
T
a
bl
e
4. R
e
s
ul
ts
of
im
a
ge
e
nh
a
nc
e
m
e
nt
e
v
a
lu
a
ti
on
I
m
a
ge
i
m
pr
ove
m
e
nt
m
e
t
hods
M
S
E
P
S
N
R
HE
12249.74
7.64
AHE
35127.59
2.79
C
L
A
H
E
21.16
38.132
G
a
m
m
a
c
or
r
e
c
t
i
on
233.13
26.97
F
ig
ur
e
9. R
e
s
ul
ts
of
im
a
ge
e
nha
nc
e
m
e
nt
gr
a
ph
B
a
s
e
d
on
T
a
bl
e
4
a
nd
F
ig
ur
e
9,
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
f
our
e
n
ha
nc
e
m
e
nt
m
e
th
od
s
unde
r
lo
w
-
li
ght
c
ondi
ti
ons
is
e
va
lu
a
te
d
u
s
in
g
M
S
E
a
nd
P
S
N
R
,
w
he
r
e
lo
w
e
r
M
S
E
a
nd
hi
ghe
r
P
S
N
R
in
di
c
a
te
be
tt
e
r
f
id
e
li
ty
a
nd
qua
li
ty
.
C
L
A
H
E
pe
r
f
or
m
s
be
s
t
(
M
S
E
21.16;
P
S
N
R
38.13
2
dB
)
,
s
how
in
g
s
tr
ong
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
w
it
hout
s
a
c
r
if
ic
in
g
im
po
r
ta
nt
de
ta
il
s
,
w
hi
c
h
is
m
os
t
s
uppor
ti
ve
f
or
obj
e
c
t
de
te
c
ti
on.
G
a
m
m
a
c
or
r
e
c
ti
on
r
a
nks
s
e
c
ond
(
M
S
E
233.13;
P
S
N
R
26.97
dB
)
,
s
ui
ta
bl
e
f
or
m
ode
r
a
te
e
nha
nc
e
m
e
nt
w
it
h
r
e
la
ti
ve
ly
c
ont
r
ol
le
d
a
r
ti
f
a
c
ts
.
I
n
c
ont
r
a
s
t,
H
E
(
M
S
E
12,249.74;
P
S
N
R
7.64
dB
)
a
nd
A
H
E
(
M
S
E
35,127.59;
P
S
N
R
2.79
dB
)
pr
oduc
e
la
r
ge
de
vi
a
ti
ons
f
r
om
th
e
r
e
f
e
r
e
nc
e
a
nd
lo
w
qua
li
ty
,
in
c
r
e
a
s
in
g
noi
s
e
a
nd
a
r
ti
f
a
c
ts
th
a
t
c
a
n
hi
nde
r
de
te
c
ti
on.
I
m
pl
ic
a
ti
on
f
or
th
e
p
ip
e
li
ne
:
pr
io
r
i
ti
z
e
C
L
A
H
E
a
s
th
e
de
f
a
ul
t
f
or
da
r
k
i
m
a
ge
s
,
w
it
h
G
a
m
m
a
c
or
r
e
c
ti
on
a
s
a
n
a
lt
e
r
na
ti
ve
w
he
n
na
tu
r
a
l
c
ont
r
a
s
t
is
s
uf
f
ic
ie
nt
o
r
w
he
n
pr
e
s
e
r
vi
ng
f
in
e
t
e
xt
ur
e
s
is
de
s
ir
e
d.
H
E
a
nd
A
H
E
a
r
e
not
r
e
c
om
m
e
nde
d
a
t
th
e
pr
e
pr
oc
e
s
s
in
g
s
ta
ge
be
c
a
us
e
th
e
y
de
gr
a
de
vi
s
ua
l
s
ig
na
l
qua
li
ty
a
nd
r
e
duc
e
t
he
l
ik
e
li
hood of
i
de
nt
if
yi
ng de
ta
il
s
i
n da
r
k r
e
gi
ons
.
Evaluation Warning : The document was created with Spire.PDF for Python.