T
E
L
KO
MNIK
A
, V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9,
p
p.
3
01
0
~
3018
IS
S
N: 1
69
3
-
6
93
0
,
accr
ed
ited
F
irst
Gr
ad
e b
y K
em
en
r
istekdikti,
Decr
ee
No: 2
1/E/
K
P
T
/20
18
DOI:
10.12928/TE
LK
OM
N
IK
A
.v
1
7
i
6
.
12701
◼
30
10
Rec
ei
v
ed
Ma
r
c
h
18
,
20
1
9
;
Rev
i
s
ed
J
u
l
y
2
,
20
1
9
;
A
c
c
ep
ted
J
u
l
y
18
, 2
0
1
9
Batik
i
mag
e
retr
ieval usi
ng
c
on
volu
tio
n
al
neu
ral
net
w
ork
Her
i
P
r
as
etyo
*
1
,
B
er
t
o
n
A
r
ie
P
u
t
r
a A
ka
r
d
ihas
2
Dep
a
rtm
e
n
t
o
f
In
f
o
rm
a
ti
c
s
,
U
n
i
v
e
rs
i
ta
s
Se
b
e
l
a
s
M
a
re
t
(UN
S)
,
Su
ra
k
a
rta
,
I
n
d
o
n
e
s
i
a
*C
o
rre
s
p
o
n
d
i
n
g
a
u
th
o
r,
e
-
m
a
i
l
:
h
e
ri
.
p
ra
s
e
ty
o
@s
ta
ff
.u
n
s
.a
c
.i
d
1
,
b
e
rto
n
a
ri
e
1
2
3
@
s
tu
d
e
n
t.
u
n
s
.
a
c
.i
d
2
Ab
strac
t
Th
i
s
p
a
p
e
r
p
re
s
e
n
t
s
a
s
i
m
p
l
e
te
c
h
n
i
q
u
e
fo
r
p
e
rfo
r
m
i
n
g
Ba
ti
k
i
m
a
g
e
re
tri
e
v
a
l
u
s
i
n
g
th
e
Con
v
o
l
u
ti
o
n
a
l
Neu
r
a
l
Ne
t
work
(CN
N)
a
p
p
ro
a
c
h
.
Two
C
NN
m
o
d
e
l
s
,
i
.e
.
s
u
p
e
r
v
i
s
e
d
a
n
d
u
n
s
u
p
e
rv
i
s
e
d
l
e
a
rn
i
n
g
a
p
p
ro
a
c
h
,
a
r
e
c
o
n
s
i
d
e
re
d
t
o
p
e
r
fo
rm
e
n
d
-
to
-
e
n
d
f
e
a
tu
re
e
x
tr
a
c
t
i
o
n
i
n
o
rd
e
r t
o
d
e
s
c
ri
b
e
t
h
e
c
o
n
t
e
n
t
o
f
Ba
t
i
k
i
m
a
g
e
.
Th
e
d
i
s
ta
n
c
e
m
e
tri
c
s
m
e
a
s
u
r
e
th
e
s
i
m
i
l
a
ri
ty
b
e
twe
e
n
t
h
e
q
u
e
ry
a
n
d
t
a
rg
e
t
i
m
a
g
e
s
i
n
d
a
ta
b
a
s
e
b
a
s
e
d
o
n
th
e
fe
a
tu
r
e
g
e
n
e
ra
t
e
d
fr
o
m
CN
N
a
rc
h
i
te
c
tu
re
.
As
re
p
o
r
te
d
i
n
th
e
e
x
p
e
r
i
m
e
n
ta
l
s
e
c
ti
o
n
,
th
e
p
r
o
p
o
s
e
d
s
u
p
e
r
v
i
s
e
d
CN
N
m
o
d
e
l
a
c
h
i
e
v
e
s
b
e
t
te
r
p
e
rf
o
rm
a
n
c
e
c
o
m
p
a
r
e
d
to
u
n
s
u
p
e
rv
i
s
e
d
CN
N
i
n
th
e
B
a
ti
k
i
m
a
g
e
re
tri
e
v
a
l
s
y
s
t
e
m
.
In
a
d
d
i
ti
o
n
,
i
m
a
g
e
fe
a
tu
r
e
c
o
m
p
o
s
e
d
fro
m
t
h
e
p
ro
p
o
s
e
d
CN
N
m
o
d
e
l
y
i
e
l
d
s
b
e
tt
e
r
p
e
rf
o
rm
a
n
c
e
c
o
m
p
a
re
d
t
o
t
h
a
t
o
f
th
e
h
a
n
d
c
r
a
ft
e
d
fe
a
tu
re
d
e
s
c
r
i
p
to
r
.
Y
e
t,
i
t
d
e
m
o
n
s
tr
a
te
s
th
e
s
u
p
e
r
i
o
ri
t
y
p
e
rfo
rm
a
n
c
e
o
f
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
e
d
a
p
p
ro
a
c
h
i
n
t
h
e
Ba
t
i
k
i
m
a
g
e
re
tri
e
v
a
l
s
y
s
te
m
.
Key
w
ords
:
a
u
to
e
n
c
o
d
e
r,
CN
N
,
d
e
e
p
l
e
a
rn
i
n
g
,
f
e
a
tu
r
e
e
x
tra
c
t
i
o
n
,
i
m
a
g
e
re
tri
e
v
a
l
Copy
righ
t
©
2
0
1
9
Uni
v
e
rsi
t
a
s
Ahm
a
d
D
a
hl
a
n.
All
righ
t
s
r
e
s
e
rve
d
.
1.
Int
r
o
d
u
ctio
n
T
he
bra
i
n
i
s
an
a
ma
z
i
ng
or
ga
n
i
n
t
he
h
um
a
n
bo
dy
.
W
i
th
ou
r
bra
i
ns
,
w
e
c
an
un
d
e
r
s
tan
d
what
w
e
s
ee
,
s
m
el
l
,
t
as
te,
he
ar
a
nd
t
ou
c
h.
T
h
e
i
nfa
nt
brai
n
wei
gh
t
i
s
on
l
y
ab
ou
t
ha
l
f
a
k
i
l
og
r
a
m
bu
t
c
a
n
s
o
l
v
e
a
bi
g
pr
ob
l
e
m,
a
nd
ev
en
s
up
erc
om
p
ut
ers
c
an
no
t.
A
fte
r
s
ev
eral
mo
nt
hs
of
b
i
r
th,
the
b
ab
y
c
an
r
ec
og
n
i
z
e
t
h
e
fac
e
of
h
i
s
pa
r
en
ts
,
d
i
s
c
ern
d
i
s
c
r
ete
ob
j
ec
ts
fr
o
m
t
he
b
ac
k
ground
,
an
d
b
eg
i
n
to
s
pe
ak
.
W
i
th
i
n
on
e
y
ea
r
th
e
ba
by
ha
s
an
i
ntu
i
t
i
on
a
bo
ut
na
t
ural
ob
j
e
c
ts
,
c
an
fol
l
ow
ob
j
ec
ts
an
d
u
nd
ers
ta
nd
th
e
me
an
i
n
g
of
a
s
ou
nd
.
W
h
e
n
the
y
are
c
h
i
l
dre
n,
th
ey
c
an
un
de
r
s
tan
d
gramm
ar an
d h
av
e
th
o
us
an
ds
of
words
i
n t
he
i
r
v
oc
ab
u
l
ary
.
B
ui
l
d
i
n
g
ma
c
h
i
ne
s
th
at
ha
v
e
i
nt
el
l
i
g
en
c
e
l
i
k
e
ou
r
bra
i
n
s
are
no
t
ea
s
y
,
to
ma
k
e
ma
c
hi
ne
s
wi
th
art
i
fi
c
i
al
i
nte
l
l
i
g
en
c
e
w
e
h
av
e
t
o
s
o
l
v
e
v
ery
c
om
p
l
ex
c
om
p
uti
ng
pro
bl
e
ms
t
ha
t
we
h
av
e
ev
en
s
tr
ug
gl
e
d
w
i
th
,
pro
bl
em
s
t
ha
t
ou
r
bra
i
ns
c
a
n
s
o
l
v
e
i
n
a
ma
tt
er
of
s
ec
on
ds
.
T
o
ov
erc
om
e
th
i
s
probl
em
,
we
ha
v
e
t
o
d
ev
e
l
op
oth
er
way
s
to
progr
a
m
c
om
pu
ters
tha
t
h
av
e
be
e
n
us
ed
i
n
th
i
s
de
c
ad
e
.
T
h
erefore
the
r
e
aris
es
an
ac
ti
v
e
fi
el
d
of
arti
fi
c
i
al
c
om
pu
ter
i
n
tel
l
i
ge
nc
e
an
d
a
l
s
o
c
om
mo
nl
y
c
al
l
e
d d
ee
p
l
e
arni
n
g
[1]
.
Nowad
ay
s
A
r
ti
fi
c
i
al
i
nt
el
l
i
g
en
c
e
h
as
un
de
r
go
ne
v
ery
r
ap
i
d
d
ev
el
o
pm
en
t.
A
i
ha
s
be
en
us
ed
i
n
ma
ny
fi
e
l
ds
of
r
es
e
arc
h,
i
n
th
e
fi
el
d
of
c
om
pu
t
er
v
i
s
i
on
C
on
te
nt
-
B
as
ed
Im
ag
e
R
etri
ev
a
l
(
CB
IR)
ha
s
be
e
n
de
v
e
l
o
pe
d
i
n
mu
l
ti
-
l
ev
e
l
s
c
he
me
s
w
i
t
h
l
ow
-
l
ev
e
l
fe
atu
r
es
to
hi
gh
-
l
ev
el
fea
tures
.
Conv
ol
uti
on
a
l
N
eu
r
a
l
Netw
ork
(
CNN)
ha
s
be
e
n
s
uc
c
es
s
ful
l
y
us
ed
to
b
e
an
eff
ec
ti
v
e
de
s
c
r
i
p
tor
fea
ture
an
d
g
ai
n
ac
c
urate
r
es
ul
ts
.
In
g
en
era
l
,
th
e
fea
tu
r
es
ga
i
n
by
th
e
de
e
p
l
e
arni
ng
me
t
ho
d
are
tr
ai
ne
d
by
m
i
m
i
c
h
um
an
pe
r
c
ep
ti
on
s
throu
gh
v
ario
us
op
erati
on
s
s
uc
h
as
c
on
v
ol
uti
on
an
d
po
o
l
i
n
g.
D
ee
p
l
ea
r
n
i
ng
ha
s
be
c
om
e
a
de
s
c
r
i
pt
or
fe
at
ure
th
at
i
s
be
tt
er
th
an
l
ow
-
l
ev
el
fe
atu
r
es
.
A
l
t
ho
u
gh
no
w
the
CNN
m
od
u
l
e
ha
s
b
ec
om
e
s
tat
e
of
the
art
i
n
c
om
pu
t
er
v
i
s
i
o
n
thi
s
do
es
no
t
gu
ara
nte
e
th
e
fe
a
tures
o
bt
ai
n
ed
fr
o
m t
h
e h
i
gh
es
t l
ev
el
al
w
ay
s
ge
t t
he
b
es
t p
erf
ormanc
e
[
2]
.
In
th
e
C
on
t
en
t
-
B
as
ed
Ima
ge
R
etri
ev
a
l
s
y
s
tem
a
i
ms
to
prov
i
d
e
t
he
r
i
gh
t
way
to
do
the
br
ows
i
n
g,
r
etr
i
ev
i
n
g
a
n
d
s
ea
r
c
hi
ng
s
o
me
de
s
i
r
e
d
i
ma
ge
s
tha
t
ha
v
e
be
en
s
t
ored
i
n
t
he
i
m
ag
e
da
ta
ba
s
e.
T
he
i
ma
g
e
d
ata
ba
s
e
c
o
nta
i
ns
m
an
y
i
m
ag
e
s
tha
t
ha
v
e
be
en
s
tored
a
nd
arr
an
ge
d
i
n
a
s
torag
e
de
v
i
c
e.
Us
ua
l
l
y
,
the
s
i
z
e
of
th
e
i
ma
ge
da
t
a
ba
s
e
i
s
v
ery
l
arge
s
o
th
at
t
he
pr
oc
es
s
of
s
ea
r
c
hi
ng
for
s
pe
c
i
f
i
c
i
ma
g
es
ma
nu
a
l
l
y
r
eq
u
i
r
es
a
l
ot
of
ti
me
,
an
d
c
a
us
es
c
on
d
i
ti
on
s
t
ha
t
are
un
c
om
f
ortab
l
e
f
or
th
e
us
er.
F
or
ex
am
pl
e,
B
at
i
k
i
s
a
c
ul
t
ural
he
r
i
tag
e
of
th
e
arc
hi
p
el
ag
o
Ind
o
ne
s
i
a
tha
t
ha
s
a
h
i
gh
v
al
ue
a
nd
b
l
en
d
of
art,
l
a
de
n
w
i
th
ph
i
l
os
op
h
i
c
al
m
ea
n
i
n
gs
an
d
me
an
i
ng
f
ul
s
y
m
bo
l
s
th
at
s
ho
w
th
e
way
of
thi
nk
i
n
g
of
the
pe
op
l
e
ma
k
i
ng
i
t.
B
at
i
k
i
s
a
c
r
aft
tha
t
ha
s
be
en
a
pa
r
t
of
I
nd
on
es
i
a
n
c
u
l
ture
es
p
ec
i
a
l
l
y
J
av
an
es
e
f
or
a
l
on
g
t
i
me
,
ba
t
i
k
h
av
e
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
us
i
ng
c
on
v
ol
ut
i
o
na
l
ne
ural
ne
t
wor
k
(
Her
i
P
r
as
ety
o
)
3011
a
l
ot
of
m
oti
v
es
,
pa
tte
r
n
a
nd
c
o
l
o
r
s
o
t
o
tak
e
s
pe
c
i
f
i
c
b
ati
k
pi
c
t
ure
fr
om
the
d
ata
b
as
e
v
ery
c
ha
l
l
en
g
i
n
g
[3]
.
T
hi
s
pa
pe
r
off
ers
a
s
ol
u
ti
on
to
us
e
c
o
nv
ol
u
ti
o
na
l
ne
ura
l
ne
tw
ork
s
to
c
arr
y
ou
t
CB
IR
t
as
k
s
to
s
ol
v
e
prob
l
em
s
t
ha
t
oc
c
ur
i
n
t
ak
i
n
g
ba
ti
k
i
m
ag
es
.
T
h
e
m
eth
od
i
nte
nd
e
d
i
s
to
produc
e
e
ffe
c
ti
v
e
i
ma
ge
de
s
c
r
i
pto
r
s
fr
om
the
CNN
arc
hi
tec
ture.
Des
c
r
i
p
tors
of
thi
s
fea
ture
are
v
ery
i
mp
ortant
for
c
on
t
e
nt
-
ba
s
e
d
s
h
oo
t
i
ng
s
y
s
tem
s
.
T
he
Ima
ge
fea
ture
i
s
us
ed
t
o
i
m
prov
e
the
p
erfor
ma
nc
e a
nd
t
o s
ol
v
e p
r
ob
l
em
s
i
n e
x
i
s
ti
n
g b
ati
k
s
ho
oti
ng
s
y
s
tem
s
.
2.
Co
n
t
ent
-
b
as
ed Im
age R
etrie
v
al
S
ys
t
e
m
Ima
g
e
r
etri
ev
a
l
i
s
a
c
om
p
ute
r
s
y
s
tem
for
s
ea
r
c
hi
ng
an
d
r
etri
ev
i
ng
a
s
pe
c
i
fi
c
i
ma
ge
i
n
l
arge
or
b
i
g
s
i
z
e
of
i
m
ag
e
da
ta
ba
s
es
.
T
h
e
c
l
as
s
i
c
al
ap
proac
h
ap
p
en
ds
on
the
me
ta
da
t
a
s
uc
h
as
t
ex
ts
,
k
ey
wor
ds
,
or
de
s
c
r
i
pt
i
o
ns
e
mb
e
dd
e
d
i
n
a
n
i
ma
ge
.
T
hu
s
,
the
i
ma
ge
r
etri
ev
al
c
an
be
p
erfor
m
ed
w
i
th
t
he
s
ea
r
c
h
k
ey
as
afo
r
em
en
t
i
on
ed
tex
t,
k
ey
wor
ds
,
e
tc
.
T
hi
s
tec
hn
i
q
ue
i
s
i
ne
ffi
c
i
en
t
s
i
nc
e
t
he
ma
nu
al
i
m
ag
e
an
no
ta
ti
o
n
s
are
ti
m
e
-
c
on
s
um
i
n
g
a
nd
ex
h
au
s
ti
ng
proc
es
s
.
E
v
en
t
ho
u
gh
,
l
arg
e
am
ou
nts
o
f
au
tom
ati
c
i
m
ag
es
an
no
t
ati
on
s
ha
v
e
be
e
n
prop
os
ed
i
n
l
i
t
erature
[4
]
,
an
i
ma
g
e
r
etri
ev
a
l
s
y
s
tem
w
i
th
c
on
ten
t
a
nn
o
tat
i
on
s
ti
l
l
c
an
no
t
d
el
i
v
er
s
ati
s
fac
tory
r
es
ul
t
.
CB
IR
i
s
c
om
pu
t
er
ap
p
l
i
c
a
ti
on
de
a
l
i
ng
w
i
th
th
e
s
ea
r
c
hi
ng
pr
ob
l
em
s
ov
er
l
ar
ge
-
s
c
al
e
i
ma
ge
d
ata
b
as
e.
C
B
IR,
al
s
o
r
ec
og
ni
z
ed
as
Q
ue
r
y
-
B
as
ed
I
ma
ge
C
o
nte
nt
(
Q
B
IC)
an
d
Cont
e
nt
-
B
as
ed
V
i
s
ua
l
Inf
o
r
ma
ti
on
S
ea
r
c
h
(
CB
V
IR)
,
di
ff
ers
wi
th
th
e
c
on
te
nt
-
ba
s
ed
ap
pro
ac
h.
T
he
C
B
IR
an
a
l
y
z
es
th
e
i
m
ag
e
c
o
nte
nt
r
at
he
r
t
ha
n
me
tad
at
a
i
nfo
r
ma
t
i
on
s
uc
h
i
m
ag
e
k
ey
w
ords
,
tag
s
, o
r
i
ma
ge
de
s
c
r
i
pt
i
on
s
[5]
.
In
t
hi
s
pa
pe
r
,
the
us
a
bi
l
i
ty
o
f CNN
m
od
e
l
i
s
ex
ten
de
d
to
th
e
C
B
IR
ta
s
k
. T
he
m
ai
n
r
ea
s
on
i
s
the
s
up
erio
r
i
ty
pe
r
for
ma
nc
e
off
ere
d
by
CNN
mo
d
e
l
c
om
p
ared
t
o
the
ha
n
dc
r
a
fte
d
f
ea
ture
i
n
the
c
om
pu
t
er
v
i
s
i
o
n
an
d
r
ec
og
ni
ti
o
n
tas
k
s
.
T
he
CNN
or
De
ep
Le
arn
ne
t
wor
k
ac
hi
ev
es
the
ou
ts
tan
d
i
n
g
r
etri
ev
a
l
p
erf
ormanc
e
i
n
th
e
Im
ag
eN
et
c
ha
l
l
en
ge
[6
]
.
T
he
CN
N
mo
de
l
i
ns
p
i
r
es
the
ot
he
r
de
ep
l
ea
r
n
i
n
g
-
ba
s
ed
ap
pro
ac
he
s
,
s
uc
h
as
A
l
ex
N
et
[7]
,
V
G
G
Net
[
8]
,
G
oo
g
l
eL
eNet
[
9]
,
Mi
c
r
os
oft
R
es
Net
[1
0]
,
e
t
c
.,
to
tac
k
l
e
t
he
o
bs
ol
et
e
of
h
an
dc
r
aft
ed
f
ea
tur
e
i
n
th
e
i
ma
g
e
r
etri
ev
al
do
ma
i
n.
T
he
CNN
mo
de
l
r
ec
ei
v
es
a
three
-
di
m
en
s
i
on
a
l
i
ma
ge
of
s
i
z
e
ℎ
×
×
,
where
ℎ
an
d
are
s
pa
t
i
al
di
me
ns
i
o
ns
an
d
i
s
th
e
n
um
b
er
of
c
h
an
ne
l
s
.
T
h
i
s
i
ma
g
e
i
s
f
urthe
r
proc
es
s
ed
tho
r
ou
gh
th
e
CNN
arc
hi
te
c
ture
c
on
s
i
s
ti
n
g
s
ev
eral
c
on
v
ol
uti
on
s
,
ma
x
-
po
ol
i
n
gs
,
an
d
ac
ti
v
a
ti
o
n
fun
c
ti
on
s
to
pe
r
form
en
d
-
to
-
en
d
i
m
ag
e
f
ea
t
ure
g
en
er
ati
o
n.
L
et
be
a
v
ec
tor
da
t
a
l
oc
at
ed
at
s
pa
ti
a
l
p
os
i
t
i
on
(
,
)
i
n
s
pe
c
i
f
i
c
l
ay
ers
. The CNN
c
om
p
ute
s
a n
ew
da
t
a
as
fo
l
l
ow:
Y
ij
=
f
ks
(
{
X
si
+
δ
i
,
sj
+
δ
j
}
0
≤
δ
i
,
δ
j
<
k
)
(
1)
where
an
d
de
n
ote
k
erne
l
s
i
z
e a
n
d s
tr
i
de
,
r
es
pe
c
t
i
v
el
y
. The
fu
nc
t
i
on
i
s
t
he
l
ay
er
ty
pe
us
e
d
s
uc
h
as
m
atri
x
do
t
mu
l
ti
pl
i
c
ati
on
for
c
on
v
ol
uti
on
a
l
l
ay
ers
,
ma
x
s
p
ati
al
for
ma
x
p
oo
l
i
ng
l
ay
ers
,
no
n
l
i
n
ea
r
f
un
c
ti
on
s
f
or ac
ti
v
ati
o
n f
un
c
ti
on
s
, a
nd
oth
er t
y
pe
s
of
l
ay
ers
. Th
i
s
fo
r
m
of
fu
nc
ti
on
al
i
ty
i
s
ma
i
nta
i
ne
d u
s
i
ng
k
ern
el
s
i
z
e a
n
d s
tep
c
o
mp
os
i
ti
on
w
hi
l
e s
t
i
l
l
us
i
n
g t
h
e t
r
a
ns
format
i
on
r
u
l
es
.
f
ks
°
g
k
′
s
′
=
(
f°g
)
k
′
+
(
k
−
1
)
s
′
,
s
s
′
(
2)
W
h
i
l
e
a
g
en
eral
n
etwo
r
k
c
om
pu
t
es
ge
ne
r
a
l
no
nl
i
ne
ar
fun
c
ti
o
ns
,
a
ne
tw
ork
wi
th
on
l
y
l
ay
ers
of
t
hi
s
f
orm
c
o
mp
ut
es
a
n
on
l
i
n
ea
r
fi
l
t
er,
wh
i
c
h
we
c
al
l
a
de
ep
fi
l
ter
or
f
ul
l
y
c
on
v
ol
ut
i
o
na
l
ne
twork
.
F
CN
na
tura
l
l
y
o
pe
r
ate
s
at
a
ny
s
i
z
e
i
np
ut
an
d
prod
uc
es
th
e
ap
pro
pria
te
s
pa
t
i
a
l
di
m
en
s
i
on
s
.
T
h
e
l
os
s
fu
n
c
ti
on
i
s
v
a
l
ue
d
c
o
mp
os
e
d
wi
t
h
t
he
F
CN
de
f
i
ne
s
ta
s
k
.
If
th
e
l
os
s
fun
c
ti
on
is
a
s
u
m
ov
er
the
s
p
ati
al
d
i
m
en
s
i
o
ns
o
f
th
e
fi
n
al
l
ay
er
(
;
)
=
∑
′
(
;
)
,
the
p
aram
ete
r
gra
di
en
t
w
i
l
l
be
a
s
um
ov
er
t
he
pa
r
am
et
er
grad
i
e
nts
of
e
ac
h
of
i
ts
s
p
ati
al
c
om
po
n
en
ts
.
T
h
us
s
toc
ha
s
ti
c
gradi
e
nt
on
c
om
pu
t
e
d
on
who
l
e
i
ma
ge
s
wi
l
l
b
e
the
s
am
e
as
the
s
toc
h
as
ti
c
grad
i
e
nt
o
n
′
,
tak
i
n
g
a
l
l
t
he
f
i
n
al
r
ec
e
pti
v
e
fi
e
l
ds
as
mi
ni
b
atc
h.
W
h
e
n
c
al
c
u
l
at
i
ng
thi
s
r
ec
ep
t
i
v
e
f
i
el
d
i
s
d
on
e
r
ep
ea
ted
l
y
wi
t
h
forw
ard
an
d
b
ac
k
war
d
pro
pa
g
ati
on
op
erat
i
on
s
fee
d
ba
c
k
wi
l
l
be
mo
r
e
ef
f
ec
ti
v
e i
f th
e c
a
l
c
ul
a
ti
o
n
i
s
do
n
e l
ay
er by
l
ay
er i
n a
l
l
i
m
ag
e
s
c
om
pa
r
e
d t
o
c
om
pu
t
i
n
g
p
atc
h
by
p
atc
h
to
t
he
pa
r
t
of
the
i
ma
ge
.
A
n
i
l
l
us
tr
ati
on
of
a
CNN
op
e
r
ati
on
c
an
b
e
s
ee
n i
n Fi
gu
r
e
1.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
b
er 20
19
:
30
1
0
-
3018
3012
T
he
prop
os
ed
CNN
mo
d
e
l
c
on
s
tr
uc
ts
the
fea
ture
d
es
c
r
i
pto
r
fr
om
B
ati
k
i
ma
g
e.
T
hi
s
fea
ture
d
es
c
r
i
pto
r
i
s
t
o
m
e
as
ure
th
e
s
i
mi
l
ar
i
ty
b
etwe
en
qu
ery
a
nd
targ
et
i
m
ag
es
i
n
d
ata
b
as
e
un
de
r
th
e
K
-
Neares
t
Ne
i
g
hb
ors
(
K
NN)
[
11
]
s
tr
ate
gy
.
T
hi
s
K
NN
t
ec
hn
i
qu
e
p
erf
orms
s
i
m
i
l
arit
y
ma
tc
hi
ng
wi
th
the
d
i
s
tan
c
e
s
c
ore
c
r
i
ter
i
on
.
T
h
i
s
pa
pe
r
i
nv
es
ti
g
ate
s
two
C
NN
m
od
e
l
s
i
n
the
tr
a
i
n
i
ng
s
ta
ge
,
i
.e
.
wi
th
s
up
erv
i
s
e
d
an
d
un
s
u
p
erv
i
s
ed
l
ea
r
n
i
ng
ap
pro
ac
h
es
.
F
i
gu
r
e
1
i
l
l
us
tr
ate
s
an
ex
am
pl
e
of
propos
e
d
s
up
erv
i
s
ed
C
N
N
arc
hi
t
ec
ture
f
or
B
at
i
k
i
ma
ge
r
etri
ev
al
.
T
he
s
up
erv
i
s
ed
t
ermi
no
l
o
gy
r
efe
r
s
to
the
ut
i
l
i
z
ati
o
n
of
c
l
as
s
l
a
be
l
,
w
he
r
ea
s
un
s
up
erv
i
s
ed
di
s
ob
ey
s
the
i
m
ag
e
l
a
be
l
i
n
th
e
tr
a
i
n
i
ng
proc
es
s
.
A
ut
oe
nc
od
er
i
s
s
i
m
pl
e
ex
am
p
l
e
of
un
s
up
erv
i
s
e
d
CN
N
m
eth
od
whi
c
h
c
om
pres
s
es
the
da
t
a
fe
atu
r
es
i
nt
o
s
ma
l
l
er
s
i
z
e
an
d
r
ec
ov
ers
ba
c
k
to
th
e o
r
i
gi
na
l
da
ta
[
1
2]
.
F
i
gu
r
e
1.
I
l
us
tr
at
i
on
op
erati
on
us
i
ng
C
NN
3.
Me
t
h
o
d
T
hi
s
s
ec
ti
on
pres
en
ts
two
me
t
ho
ds
for
ge
n
erati
ng
t
he
f
ea
t
ure
d
es
c
r
i
pto
r
i
n
th
e
B
ati
k
i
ma
ge
r
etri
ev
a
l
s
y
s
tem
.
W
e
fi
r
s
t
l
y
ex
p
l
ai
n
t
he
s
up
erv
i
s
ed
CNN
m
od
e
l
.
T
he
n
,
th
e
un
s
u
pe
r
v
i
s
ed
CA
E
mo
de
l
[1
3]
i
s
s
u
bs
eq
u
en
tl
y
de
s
c
r
i
be
d i
n t
h
i
s
s
ec
ti
on
.
3.1
.
S
u
p
er
vis
ed L
ea
r
n
ing
T
he
CNN
m
od
e
l
i
s
the
s
u
pe
r
v
i
s
ed
de
e
p
l
ea
r
n
i
n
g
-
ba
s
ed
ap
proac
h
c
om
m
on
l
y
u
s
ed
i
n
the
i
ma
ge
c
l
as
s
i
fi
c
a
ti
o
n
[
1
4]
,
pre
di
c
t
i
on
[15
]
,
s
eg
me
n
tat
i
on
,
an
a
l
y
s
i
s
[1
6]
,
e
tc
.
T
he
s
up
erv
i
s
ed
CNN
mo
de
l
c
o
ns
i
s
ts
of
s
ev
eral
l
ay
ers
s
uc
h
as
c
on
v
ol
ut
i
on
al
l
ay
er,
m
ax
po
o
l
i
ng
l
ay
er,
et
c
.
T
he
s
e
l
ay
ers
are
r
ep
ea
t
ed
ov
er s
ev
eral
t
i
me
s
an
d f
ed
i
nto
the
fu
l
l
y
c
on
n
ec
ted
l
ay
e
r
at
t
he
en
d
of
CNN l
ay
er
[17
]
.
O
ur
propos
ed
i
m
ag
e
r
e
tr
i
e
v
al
s
y
s
tem
em
pl
oy
s
th
e
C
NN
arc
hi
tec
t
ure
wi
th
s
i
x
c
on
v
ol
ut
i
o
na
l
l
ay
ers
a
nd
tw
o
ful
l
y
c
on
n
ec
ted
l
ay
ers
to
ge
ne
r
ate
B
at
i
k
fea
t
ure
d
es
c
r
i
pto
r
.
T
ab
l
e
1
s
um
ma
r
i
z
es
th
e
CNN ar
c
h
i
t
ec
ture us
ed
i
n o
ur
propos
e
d m
et
ho
d
.
T
ab
l
e
1.
T
he
S
up
erv
i
s
ed
C
NN A
r
c
hi
tec
ture f
or B
ati
k
I
ma
ge
Retr
i
ev
al
L
a
y
e
r
T
y
p
e
S
iz
e
Ou
t
p
u
t
S
h
a
p
e
I
n
p
u
t
(
1
2
8
,
1
2
8
,
3
)
-
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
8
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
1
2
8
,
1
2
8
,
8
)
Max
P
o
o
li
n
g
8
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
6
4
,
6
4
,
8
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
1
6
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
6
4
,
6
4
,
1
6
)
Max
P
o
o
li
n
g
1
6
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
3
2
,
3
2
,
1
6
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
3
2
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
3
2
,
3
2
,
3
2
)
Max
P
o
o
li
n
g
3
2
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
1
6
,
1
6
,
3
2
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
6
4
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
1
6
,
1
6
,
6
4
)
Max
P
o
o
li
n
g
64
(2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
8
,
8
,
6
4
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
1
2
8
(
3
x
3
)
f
i
lt
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
in
g
(
8
,
8
,
1
2
8
)
Max
P
o
o
li
n
g
1
2
8
(
2
x
2
)
f
i
lt
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
in
g
(
4
,
4
,
1
2
8
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
2
5
6
(
3
x
3
)
f
i
lt
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
in
g
(
4
,
4
,
2
5
6
)
Max
P
o
o
li
n
g
2
5
6
(
2
x
2
)
f
i
lt
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
in
g
(
2
,
2
,
2
5
6
)
Flat
e
r
n
+
D
r
o
p
o
u
t
(
3
0
%
)
(
1
,
1
,
1
0
2
4
)
1
0
2
4
n
e
u
r
o
n
s
256
D
e
n
s
e
2
5
6
n
e
u
r
o
n
s
97
S
o
f
t
max
9
7
w
a
y
97
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
us
i
ng
c
on
v
ol
ut
i
o
na
l
ne
ural
ne
t
wor
k
(
Her
i
P
r
as
ety
o
)
3013
A
fte
r
p
erfor
m
i
ng
s
i
x
c
on
v
o
l
ut
i
on
an
d
ma
x
-
po
ol
i
ng
op
erati
o
ns
,
an
i
n
pu
t
i
ma
ge
o
f
s
i
z
e
128
×
128
×
3
i
s
c
on
v
erted
i
nto
ne
w
r
e
p
r
es
en
tat
i
on
w
i
t
h
di
me
ns
i
on
al
i
ty
2
×
2
×
256
.
T
hi
s
ne
w
da
ta
r
ep
r
es
e
nta
t
i
on
i
s
t
he
n
fl
a
tte
n
to
be
c
o
me
on
e
d
i
me
ns
i
on
a
l
da
ta
of
s
i
z
e
1
×
1
×
1024
.
T
h
is
fl
att
en
d
ata
i
s
s
ub
s
e
qu
e
nt
l
y
proc
es
s
ed
an
d
tr
a
i
n
ed
wi
th
t
he
Mu
l
t
i
-
La
y
er
P
erc
ep
tr
on
(
M
LP
)
.
Her
ei
n,
t
he
ML
P
r
ec
ei
v
es
1
02
4
i
n
pu
t
fe
atu
r
e
a
nd
fee
ds
i
n
to
10
24
i
n
pu
t
ne
urons
.
The
hi
d
de
n a
nd
ou
tp
ut
l
ay
ers
ar
e
s
et
as
25
6
a
nd
97
,
r
es
pe
c
t
i
v
el
y
.
T
he
v
al
u
e
of
97
i
n
o
utp
u
t
l
ay
er
s
i
s
e
qu
i
v
a
l
e
nt
to
tha
t
of
th
e
de
s
i
r
e
d
c
l
as
s
target,
i
.e.
t
he
n
um
b
er
of
B
at
i
k
i
ma
ge
c
l
as
s
es
us
ed
i
n
the
pro
po
s
e
d
i
ma
ge
r
etr
i
ev
a
l
s
y
s
tem
.
3.2
. C
AE
Un
sup
er
vis
ed L
ea
r
n
ing
T
hi
s
pa
pe
r
a
l
s
o
c
on
s
i
d
ers
the
ot
he
r
CNN
mo
de
l
,
na
me
l
y
C
on
v
ol
uti
on
a
l
A
uto
-
E
nc
od
er
(
CA
E
)
,
for
ge
n
erati
ng
i
m
ag
e
fe
atu
r
e.
T
he
C
A
E
i
s
an
u
ns
up
erv
i
s
ed
d
ee
p
l
ea
r
n
i
ng
-
ba
s
ed
me
t
ho
d,
i
.e.
th
e
i
ma
g
e
l
a
be
l
i
s
no
t
r
eq
u
i
r
ed
i
n
the
tr
a
i
n
i
ng
proc
es
s
.
In
order
to
ge
ne
r
at
e
i
ma
ge
f
ea
t
ure,
thi
s
te
c
hn
i
qu
e
l
ea
r
ns
a
nd
c
ap
tures
t
he
i
nf
ormat
i
o
n f
r
o
m i
np
ut
da
ta
di
r
ec
t
l
y
wi
t
ho
u
t th
e
av
ai
l
a
bi
l
i
ty
of
c
l
as
s
l
ab
e
l
.
T
he
C
A
E
i
nv
ol
v
es
t
wo
pa
r
ts
,
i
.e.
en
c
o
de
r
an
d
de
c
od
er
bl
oc
k
s
.
T
h
e
e
nc
od
er
bl
oc
k
proc
es
s
es
the
s
am
p
l
e
da
t
a
c
on
s
i
s
ti
ng
s
am
pl
es
an
d
fea
tures
to
y
i
el
d
t
he
o
u
tpu
t
.
In
the
o
pp
os
i
te
s
i
de
,
t
he
d
ec
o
de
r
ai
ms
to
r
ec
o
ns
tr
uc
t
the
orig
i
na
l
s
am
pl
e
da
ta
fr
om
the
.
Le
t
′
be
the
r
ec
on
s
tr
uc
ted
da
ta
produc
e
d
at
th
e
de
c
o
de
r
s
i
de
.
T
he
ma
i
n
g
oa
l
of
C
A
E
i
s
to
mi
ni
mi
z
e
the
d
i
ff
erenc
e
be
tw
ee
n
t
he
or
i
gi
na
l
da
ta
an
d
r
ec
on
s
tr
uc
ted
v
ers
i
on
′
.
S
pe
c
i
f
i
c
al
l
y
,
the
en
c
o
de
r
s
i
mp
l
y
m
ap
s
t
he
i
n
pu
t
i
nt
o
ne
w
r
ep
r
es
e
n
tat
i
on
w
i
th
th
e
he
l
p
of
fun
c
ti
on
.
T
hi
s
proc
es
s
c
an
be
f
ormu
l
at
ed
as
fo
l
l
ow:
Y
=
f
(
X
)
=
s
f
(
WX
+
b
X
)
(
3)
where
d
en
ot
es
th
e
no
n
l
i
ne
ar
ac
ti
v
at
i
on
fu
nc
ti
o
n
i
n
en
c
o
de
r
s
i
de
.
CA
E
s
i
mp
l
y
pe
r
f
orms
a
l
i
n
ea
r
op
erati
on
i
f
o
ne
s
i
mp
l
y
us
es
i
d
en
t
i
ty
fun
c
ti
on
for
.
T
he
and
∈
are
en
c
od
er
pa
r
am
ete
r
s
,
r
es
pe
c
t
i
v
el
y
,
r
efe
r
r
i
n
g
as
w
ei
gh
t
ma
tr
i
x
an
d
bi
as
v
ec
tor.
In
c
on
tr
a
s
t,
the
de
c
od
er
r
ec
on
s
tr
uc
ts
′
fr
om
r
ep
r
es
en
tat
i
o
n
by
m
ea
ns
of
f
un
c
ti
on
.
T
hi
s
proc
es
s
c
an
be
s
i
mp
l
y
i
l
l
us
tr
ate
d a
s
:
X
′
=
g
(
Y
)
=
s
g
(
W
′
Y
+
b
Y
)
(
4)
where
r
ep
r
es
en
ts
the
ac
t
i
v
ati
on
fu
nc
ti
o
n
i
n
de
c
o
de
r
s
i
de
.
T
he
an
d
are
th
e
bi
a
s
v
ec
tor
an
d w
ei
gh
t
ma
tr
i
x
, r
es
pe
c
ti
v
el
y
, d
e
no
t
i
ng
as
de
c
o
de
r
p
aramet
er.
S
tr
i
c
tl
y
s
p
ea
k
i
ng
,
t
he
CA
E
mo
de
l
s
ea
r
c
he
s
the
g
l
o
ba
l
or
n
ea
r
o
pti
m
u
m
pa
r
am
ete
r
=
(
,
,
)
i
n
the
tr
a
i
n
i
ng
proc
es
s
.
T
h
i
s
t
as
k
i
s
eq
ui
v
al
en
t
to
th
e
mi
ni
mi
z
at
i
o
n
proc
es
s
of
l
os
s
fu
nc
ti
on
ov
er al
l
da
tas
et
un
de
r
th
e f
o
l
l
owi
n
g o
bj
ec
ti
v
e f
u
nc
ti
on
:
θ
=
min
θ
L
(
X
,
X
′
)
=
min
θ
L
(
X
,
g
(
f
(
X
)
)
)
(
5)
where
(
∙
,
∙
)
de
no
tes
t
he
au
t
o
-
en
c
od
er
l
os
s
fun
c
t
i
o
n.
In
thi
s
pa
p
er,
we
s
i
mp
l
y
u
s
e
l
i
ne
a
r
r
ec
on
s
tr
uc
ti
on
2
for
l
os
s
fu
nc
ti
on
,
or
c
om
mo
n
l
y
r
efe
r
r
ed
as
M
ea
n
S
q
ua
r
e
d
E
r
r
o
r
(
MS
E
)
[
18
]
.
T
hi
s
l
os
s
fu
nc
ti
o
n
i
s
fo
r
ma
l
l
y
de
fi
ne
d
as
:
L
2
(
θ
)
=
∑
‖
x
i
−
x
i
′
‖
2
n
i
=
1
=
∑
‖
x
i
−
g
(
f
(
x
i
)
)
‖
2
n
i
=
1
(
6)
where
∈
,
′
∈
′
an
d
∈
,
r
es
pe
c
ti
v
el
y
d
en
ot
e
th
e
ori
gi
na
l
i
np
ut
da
t
a,
r
ec
on
s
tr
uc
te
d
d
ata
,
an
d
ne
w c
o
mp
ac
t re
pres
en
tat
i
on
o
f i
np
ut
da
t
a.
In
th
i
s
p
ap
er,
th
e
C
A
E
a
r
c
hi
tec
ture
was
b
ui
l
t
wi
t
h
fou
r
en
c
od
i
n
g
b
l
oc
k
s
an
d
fo
ur
de
c
od
i
ng
s
tag
es
.
T
h
i
s
arc
hi
tec
ture
i
nc
l
u
de
s
a
s
tac
k
ed
Conv
o
l
ut
i
o
na
l
A
uto
-
E
nc
od
er.
T
he
s
um
ma
r
y
o
f
C
A
E
arc
hi
te
c
ture
us
ed
i
n
th
i
s
pa
pe
r
c
an
b
e
s
ee
n
i
n
T
ab
l
e
2.
S
u
pp
os
e
tha
t
a
n
i
np
ut
i
m
ag
e
i
s
of
s
i
z
e
128
×
128
×
3
.
A
s
i
t
c
an
be
i
nfe
r
r
e
d
fr
om
T
ab
l
e
2,
t
hi
s
i
m
ag
e
i
s
c
on
v
o
l
v
e
d
fou
r
ti
m
es
to
ob
ta
i
n
n
ew
s
i
mp
l
er
a
nd
c
om
p
ac
t
r
ep
r
es
en
tat
i
o
n.
T
hi
s
proc
es
s
c
an
be
al
s
o
c
on
s
i
de
r
e
d
as
r
e
pe
t
i
ti
v
e
en
c
od
i
n
g.
Her
ei
n
,
th
e
ne
w
r
e
pres
en
tat
i
o
n
i
s
r
e
ga
r
de
d
a
s
ne
ura
l
c
od
e
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
b
er 20
19
:
30
1
0
-
3018
3014
wi
th
d
i
m
en
s
i
on
a
l
i
ty
4
×
4
×
128
.
B
y
u
s
i
ng
th
e
ba
c
k
war
d
a
pp
r
o
a
c
h
an
d
d
ec
od
i
ng
proc
es
s
,
thi
s
ne
ura
l
c
od
e
c
an
b
e
r
ec
ov
ered
ba
c
k
to
y
i
el
d
t
h
e
r
ec
on
s
tr
uc
te
d
i
ma
ge
o
f
orig
i
n
al
s
i
z
e
128
×
128
×
3
.
T
hi
s
r
ev
ers
e
proc
es
s
p
erfor
ms
th
e
d
ec
on
v
ol
uti
on
an
d
un
po
o
l
i
ng
op
era
ti
o
n
s
.
T
he
C
A
E
n
eu
r
a
l
c
o
de
c
an
be
furt
he
r
uti
l
i
z
e
d
as
th
e
fea
ture
d
es
c
r
i
pto
r
i
n
th
e
p
r
op
os
ed
B
ati
k
i
ma
ge
r
etr
i
ev
a
l
s
y
s
tem
.
T
ab
l
e
2.
T
he
C
A
E
A
r
c
hi
tec
t
ure for
B
at
i
k
Im
a
ge
R
etri
ev
al
S
y
s
tem
L
a
y
e
r
T
y
p
e
S
iz
e
Ou
t
p
u
t
S
h
a
p
e
I
n
p
u
t
(
1
2
8
,
1
2
8
,
3
)
-
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
3
2
(
3
x
3)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
1
2
8
,
1
2
8
,
3
2
)
Max
P
o
o
li
n
g
+
D
r
o
p
o
u
t
3
2
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
6
4
,
6
4
,
3
2
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
6
4
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
6
4
,
6
4
,
6
4
)
Max
P
o
o
li
n
g
+
D
r
o
p
o
u
t
6
4
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
3
2
,
3
2
,
6
4
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
6
4
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
3
2
,
3
2
,
6
4
)
Max
P
o
o
li
n
g
+
D
r
o
p
o
u
t
6
4
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
1
6
,
1
6
,
6
4
)
C
o
n
v
o
lut
ion
a
l
+
R
e
lu
1
2
8
(
3
x
3
)
f
i
lt
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
in
g
(
1
6
,
1
6
,
1
2
8
)
Max
P
o
o
li
n
g
+
D
r
o
p
o
u
t
1
2
8
(
2
x
2
)
f
i
lt
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
in
g
(
8
,
8
,
1
2
8
)
Max
P
o
o
li
n
g
(
N
e
u
r
a
l
C
o
d
e
)
1
2
8
(
2
x
2
)
f
i
lt
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
in
g
(
4
,
4
,
1
2
8
)
U
n
p
o
li
n
g
1
2
8
(
2
x
2
)
f
i
lt
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
in
g
(
8
,
8
,
1
2
8
)
De
c
o
n
v
o
v
u
t
ion
+
R
e
lu
1
2
8
(
3
x
3
)
f
i
lt
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
in
g
(
8
,
8
,
1
2
8
)
U
n
p
o
o
li
n
g
+
D
r
o
p
o
u
t
1
2
8
(
2
x
2
)
f
i
lt
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
in
g
(
1
6
,
1
6
,
1
2
8
)
D
e
c
o
n
v
o
v
u
t
ion
+
R
e
lu
6
4
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
1
6
,
1
6
,
6
4
)
U
n
p
o
o
li
n
g
+
D
r
o
p
o
u
t
6
4
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
3
2
,
3
2
,
6
4
)
D
e
c
o
n
v
o
v
u
t
ion
+
R
e
lu
6
4
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
3
2
,
3
2
,
6
4
)
U
n
p
o
o
li
n
g
+
D
r
o
p
o
u
t
6
4
(
2
x
2
)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
6
4
,
6
4
,
6
4
)
D
e
c
o
n
v
o
v
u
t
ion
+
R
e
lu
3
2
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
6
4
,
6
4
,
3
2
)
U
n
p
o
o
li
n
g
+
D
r
o
p
o
u
t
3
2
(
2
x
2)
f
il
t
e
r
s
,
2
s
t
r
ide
,
0
p
a
d
d
ing
(
1
2
8
,
1
2
8
,
3
2
)
D
e
c
o
n
v
o
v
u
t
ion
+
S
igmoid
3
2
(
3
x
3
)
f
il
t
e
r
s
,
1
s
t
r
ide
,
2
p
a
d
d
ing
(
1
2
8
,
1
2
8
,
3
)
3.3
. Le
ar
n
ing
p
r
o
ce
ss
an
d
Hyper
p
ar
em
eter
T
u
n
ing
T
he
CNN
mo
de
l
i
s
v
ery
s
e
ns
i
ti
v
e
to
hy
p
erpar
a
me
t
er
c
ha
ng
es
i
n
the
l
ea
r
n
i
ng
pr
oc
es
s
,
s
i
nc
e
i
t
ut
i
l
i
z
es
the
Res
tr
uc
tured
L
i
n
ea
r
U
ni
t
(
R
eL
u)
(
)
=
(
0
,
)
f
or
i
ts
ac
ti
v
ati
on
f
un
c
ti
on
.
T
hi
s
f
un
c
ti
on
i
s
wi
t
h
t
he
gr
ad
i
en
t
de
s
c
en
t
ma
k
i
ng
i
t
v
ery
un
s
ta
bl
e
i
n
c
om
p
aris
on
wi
t
h
the
ta
nh
an
d
s
i
g
mo
i
d
ac
t
i
v
ati
on
f
un
c
ti
on
s
.
Co
mp
ar
ed
to
t
h
e
a
foreme
nti
on
e
d
ac
ti
v
a
ti
o
n
f
un
c
ti
o
ns
,
ReL
u
y
i
el
ds
an
i
de
nti
c
a
l
err
or w
i
t
h 2
5%
l
es
s
i
t
erat
i
on
i
n
l
e
arni
n
g s
tag
e
[7]
.
In
the
tr
ai
ni
n
g
proc
es
s
o
f
ou
r
pro
po
s
ed
i
m
ag
e
r
e
tr
i
ev
al
s
y
s
tem
,
we
s
i
mp
l
y
s
pl
i
t
the
i
ma
g
e
da
t
as
et
as
two
f
ol
ds
,
i
.
e.
75
%
an
d
2
5%
for
tr
ai
n
i
ng
a
nd
tes
t
i
ng
p
urpos
e,
r
es
pe
c
ti
v
e
l
y
.
T
he
A
da
pt
i
v
e
M
om
en
t
E
s
ti
ma
ti
on
(
A
da
m)
[19]
i
s
ex
p
l
oi
te
d
for
CN
N
op
t
i
m
i
z
er
wi
t
h
l
e
arni
ng
r
at
e
0.0
0
01
.
W
e
s
i
mp
l
y
e
mp
l
oy
the
M
ea
n
S
qu
are
E
r
r
or
(
M
S
E
)
[20
]
f
or
c
a
l
c
ul
a
ti
n
g
the
l
os
s
fun
c
t
i
on
.
F
or
av
oi
di
n
g
t
he
ov
erfi
tt
i
ng
probl
em
an
d
de
al
i
ng
wi
th
s
ma
l
l
s
i
z
e
of
da
t
as
et,
t
he
propos
ed
s
y
s
tem
us
es
da
t
a
a
ug
me
nt
ati
on
t
ec
hn
i
qu
e
to
i
mp
r
ov
e
the
da
ta
v
aria
t
i
o
n.
T
he
tr
ai
n
i
n
g
an
d
t
es
ti
n
g
proc
es
s
es
are
c
on
du
c
te
d
un
d
er
th
e
I
nte
l
Co
r
e
i
5
20
1
0
pr
oc
es
s
or.
F
r
om
o
ur
ex
pe
r
i
me
nt,
the
s
up
erv
i
s
e
d
CNN
a
nd
CA
E
mo
d
el
s
r
eq
ui
r
e
aro
un
d
10
ho
urs
an
d
3
da
y
s
,
r
es
pe
c
ti
v
el
y
,
for
the
tr
ai
n
i
n
g
pr
oc
es
s
.
A
t
t
h
e
en
d
of
tr
a
i
n
i
n
g
proc
es
s
,
t
wo
d
ee
p
l
ea
r
n
i
ng
ba
s
e
d
m
od
e
l
s
prod
uc
e
a
s
et
o
f
i
ma
g
e
f
ea
tur
es
wh
i
c
h
c
an
b
e
us
e
d
for
the
de
s
c
r
i
pto
r
i
n
th
e
B
at
i
k
i
ma
g
e
r
etri
ev
a
l
.
T
h
es
e
i
ma
ge
f
ea
t
ures
are
s
i
mp
l
y
ob
ta
i
ne
d
fr
o
m
th
e
l
as
t
l
ay
er
an
d
ne
ural
c
o
de
l
ay
er
o
f
s
up
erv
i
s
ed
CNN
an
d C
A
E
mo
de
l
s
, res
pe
c
ti
v
el
y
.
4.
E
xper
i
men
t
al
S
t
u
d
y
E
x
ten
s
i
v
e
ex
pe
r
i
me
nts
we
r
e
c
arr
i
ed
ou
t
t
o
i
nv
es
t
i
g
ate
an
d
ex
am
i
n
e
the
pr
o
po
s
ed
me
th
od
p
erfor
ma
nc
e
i
n
th
e
B
at
i
k
i
ma
g
e
r
etri
ev
al
s
y
s
tem
.
F
i
r
s
tl
y
,
w
e
g
i
v
e
a
b
r
i
ef
de
s
c
r
i
p
ti
o
n
ab
ou
t
th
e
i
ma
g
e
d
ata
s
et
u
s
ed
i
n
th
e
ex
pe
r
i
me
nt.
T
he
eff
ec
t
i
v
en
es
s
of
the
prop
o
s
ed
me
tho
d
i
s
s
ub
s
eq
ue
ntl
y
ob
s
erv
ed
un
de
r
v
i
s
ua
l
i
nv
es
t
i
ga
ti
o
n.
T
he
n,
th
e
o
bj
ec
ti
v
e
pe
r
form
an
c
e
c
om
pa
r
i
s
o
ns
are
furt
he
r
e
v
al
ua
t
ed
to
ov
erl
oo
k
t
he
eff
ec
t
of
di
ff
erent
d
i
s
tan
c
e
m
etri
c
s
a
nd
s
up
erio
r
i
ty
of
t
he
propos
e
d
me
th
od
i
n c
o
mp
ar
i
s
on
w
i
th
the
f
ormer
c
om
pe
t
i
ng
s
c
he
me
s
.
4.1
.
D
ataset
T
hi
s
e
x
p
erim
en
t
uti
l
i
z
es
a
s
et
of
B
ati
k
i
m
ag
es
,
r
efe
r
ed
as
B
a
ti
k
i
ma
ge
da
t
as
et
,
ov
er
v
ario
us
pa
tt
erns
,
c
ol
ors
,
an
d
mo
ti
fs
.
T
h
i
s
i
ma
ge
da
ta
ba
s
e
c
on
s
i
s
ts
of
15
5
2
i
ma
ge
.
T
h
i
s
da
ta
ba
s
e
i
s
furt
he
r
d
i
v
i
de
d
i
nt
o
9
7
i
m
ag
e
c
l
as
s
es
.
E
a
c
h
c
l
as
s
c
on
t
ai
ns
a
s
et
of
s
i
mi
l
ar
i
m
a
ge
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
us
i
ng
c
on
v
ol
ut
i
o
na
l
ne
ural
ne
t
wor
k
(
Her
i
P
r
as
ety
o
)
3015
r
eg
ardi
ng
to
th
ei
r
mo
t
i
fs
an
d
c
on
t
en
t
a
pp
e
aranc
e.
E
ac
h
i
ma
ge
c
l
as
s
o
wns
16
s
i
m
i
l
ar
i
ma
g
es
,
i
n
whi
c
h
a
l
l
i
ma
g
es
b
el
on
g
i
ng
to
th
e
s
a
me
c
l
as
s
ar
e
c
o
n
s
i
de
r
ed
as
s
i
m
i
l
ar
i
ma
ge
s
.
F
i
gu
r
e
2
g
i
v
es
s
ev
eral
ex
a
mp
l
es
of
B
at
i
k
i
ma
ge
s
fr
om
t
he
d
ata
s
e
t.
4.2
.
P
r
ac
t
ica
l App
lic
atio
n
o
n
Batik Im
age Ret
r
ie
va
l
T
hi
s
s
ub
-
s
ec
ti
o
n
ev
al
ua
t
e
s
the
pe
r
form
an
c
e
o
f
the
propos
ed
me
t
ho
d
un
d
er
v
i
s
ua
l
i
nv
es
ti
ga
t
i
on
.
T
he
pro
po
s
e
d
me
t
ho
d
uti
l
i
z
es
th
e
i
ma
ge
fe
atu
r
e
o
bta
i
n
ed
fr
o
m
CNN
an
d
CA
E
ap
pro
ac
h
for
pe
r
for
mi
n
g
B
ati
k
i
m
ag
e
r
e
tr
i
ev
a
l
s
y
s
tem
.
T
he
c
orr
ec
tne
s
s
of
the
pro
po
s
ed
m
eth
o
d
i
s
de
t
ermi
ne
d
whet
he
r
t
he
s
y
s
tem
r
etu
r
ns
a
s
et
of
r
etr
i
ev
ed
i
ma
ge
s
c
orr
ec
tl
y
or n
ot.
F
i
gu
r
e
3
d
i
s
pl
ay
s
the
r
etri
e
v
ed
i
ma
ge
s
r
etu
r
n
ed
by
the
propos
ed
i
ma
ge
r
etri
ev
al
s
y
s
tem
us
i
ng
t
he
CNN
an
d
CA
E
i
ma
ge
fea
t
ures
.
W
e
on
l
y
s
ho
w
s
i
x
-
tee
n
r
etr
i
ev
ed
i
m
ag
es
arr
an
ge
d
i
n
as
c
en
di
ng
ma
n
ne
r
b
as
ed
o
n
the
i
r
s
i
m
i
l
arit
y
s
c
ore.
T
h
e
s
i
mi
l
ari
ty
c
r
i
teri
o
n
i
s
me
as
ured
us
i
ng
the
di
s
tan
c
e
s
c
ore
an
d
g
i
v
en
at
th
e
t
op
of
ea
c
h
i
ma
g
e.
S
m
al
l
er
di
s
ta
nc
e
v
a
l
u
e
i
n
di
c
ate
s
mo
r
e
s
i
mi
l
ar
be
tw
ee
n
th
e
qu
ery
an
d
targ
et
i
ma
ge
i
n
d
ata
ba
s
e.
A
s
s
h
own
i
n
t
hi
s
fi
g
ure,
the
pro
po
s
ed
me
th
od
w
i
th
CNN
fea
t
ure
r
etu
r
ns
a
l
l
r
e
tr
i
ev
e
d
i
ma
ge
s
c
orr
ec
tl
y
.
It
i
s
l
i
tt
l
e
r
eg
r
ett
a
bl
e
th
at
the
pro
po
s
e
d m
e
tho
d w
i
th
CA
E
fe
a
ture o
nl
y
produc
es
s
i
x
r
etri
ev
ed
i
m
ag
es
c
orr
ec
t
l
y
.
F
i
gu
r
e
2.
S
om
e i
ma
ge
s
am
pl
es
i
n t
he
B
at
i
k
da
t
as
et
(
a)
(
b)
F
i
gu
r
e
3.
P
erfor
m
an
c
e
ev
al
ua
ti
on
i
n t
erms
of
v
i
s
ua
l
i
nv
es
ti
ga
ti
o
n
for the
prop
os
ed
me
th
od
w
i
th:
(
a) CN
N
, a
n
d (b)
C
A
E
i
ma
ge
fe
a
ture
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
b
er 20
19
:
30
1
0
-
3018
3016
4.3
.
Co
mp
ar
i
son
of
P
o
r
p
o
se
d
M
eth
o
d
s wit
h
Dir
eff
e
r
ent
Distan
c
e M
et
r
ic
s
T
hi
s
s
ub
-
s
ec
ti
on
r
ep
orts
t
h
e e
ff
ec
t o
f
d
i
ffe
r
en
t
di
s
ta
nc
e m
etri
c
s
on
th
e p
r
o
po
s
e
d
me
th
od
.
In
t
hi
s
ex
pe
r
i
me
nt,
thre
e
d
i
s
tan
c
e
m
etri
c
s
,
na
me
l
y
E
uc
l
i
de
a
n
[
21
]
,
M
an
ha
tta
n
[
22
]
,
an
d
B
r
ay
-
Cur
t
i
s
di
s
t
an
c
e
[23]
,
are
ex
te
ns
i
v
el
y
ex
am
i
ne
d
ov
er
t
wo
pe
r
form
an
c
e
c
r
i
t
erio
n,
i
.e.
prec
i
s
i
on
a
nd
r
ec
a
l
l
r
ate
. Thes
e t
w
o s
c
ores
are f
ormal
l
y
de
f
i
n
ed
as
:
p
i
(
n
)
=
R
V
n
(
7)
r
i
(
n
)
=
R
V
M
(
8)
where
(
)
an
d
(
)
de
n
ote
s
th
e
pr
ec
i
s
i
on
an
d
r
ec
a
l
l
r
ate
,
r
es
pe
c
ti
v
e
l
y
,
i
f
i
ma
g
e
i
s
turn
e
d
as
qu
ery
i
ma
ge
.
T
he
s
y
mb
o
l
s
an
d
r
ep
r
es
e
nt
the
nu
mb
er
of
r
etri
ev
ed
i
m
ag
es
an
d
t
o
tal
i
ma
g
es
i
n
d
ata
ba
s
e
w
hi
c
h
i
s
r
el
ev
an
t
t
o
i
ma
g
e
,
r
es
pe
c
ti
v
e
l
y
.
i
s
th
e
n
um
b
er
of
i
m
ag
es
whi
c
h
are
r
el
ev
an
t
to
qu
ery
i
ma
ge
ob
tai
n
ed
at
r
etri
ev
ed
i
m
ag
es
.
F
i
gu
r
e
4
s
h
ows
th
e
pe
r
for
ma
nc
e
c
om
p
aris
on
ov
er
v
a
r
i
ou
s
d
i
s
tan
c
e
m
etri
c
s
i
n
te
r
ms
of
P
r
ec
i
s
i
o
n
a
nd
R
ec
al
l
s
c
ore
s
.
A
l
l
i
ma
g
es
i
n
da
tab
as
e
a
r
e
c
ho
s
en
as
qu
ery
i
ma
ge
.
T
he
n
um
b
er
of
r
etri
ev
ed
i
ma
g
es
are
s
et
as
=
{
1
,
2
,
…
,
16
}
.
I
n
m
os
t
c
as
es
,
B
r
ay
-
Cur
t
i
s
di
s
t
an
c
e
y
i
e
l
ds
th
e
be
s
t
r
etri
ev
al
p
erfor
ma
nc
e
c
o
m
pa
r
ed
to
th
at
of
the
ot
he
r
di
s
tan
c
e
m
etri
c
s
f
or
b
oth
CNN
a
nd
CA
E
i
ma
ge
fe
atu
r
e
.
In
t
he
B
at
i
k
i
ma
ge
r
etr
i
ev
al
s
y
s
tem
,
th
e
B
r
ay
-
Cur
ti
s
d
i
s
tan
c
e
be
c
om
es
a
g
oo
d
c
an
di
da
te
fo
r
m
ea
s
uri
ng
t
h
e s
i
m
i
l
ar
i
ty
be
twee
n t
h
e q
u
ery
an
d t
arget
i
m
ag
es
i
n
da
tab
as
e.
T
ab
l
e
3
ta
bu
l
at
es
mo
r
e
c
o
mp
l
ete
c
o
mp
ars
i
on
s
for
t
he
propos
e
d
i
ma
g
e
r
etri
ev
al
s
y
s
tem
us
i
ng
CNN
an
d
CA
E
f
ea
t
u
r
es
ov
er
v
ario
us
di
s
ta
nc
e.
T
hi
s
c
om
p
aris
on
i
s
ev
al
u
at
ed
i
n
t
erms
of
av
erage
r
ec
a
l
l
r
ate
wi
t
h
t
he
n
um
b
er
of
r
etr
i
ev
ed
i
ma
ge
s
as
=
16
.
Her
e
i
n,
al
l
i
ma
ge
s
i
n
da
ta
ba
s
e
are
turn
ed
as
q
ue
r
y
i
m
ag
e.
A
s
r
ep
orted
i
n
thi
s
t
ab
l
e,
t
he
propos
e
d
me
th
od
w
i
th
s
up
erv
i
s
ed
CNN
de
l
i
v
ers
b
ett
er
pe
r
f
orm
an
c
e
c
om
p
ared
to
t
ha
t
of
CA
E
tec
hn
i
q
ue
.
T
he
i
m
ag
e
fea
ture
ob
ta
i
ne
d
fr
o
m
pro
po
s
ed
s
u
pe
r
v
i
s
ed
CNN
m
eth
o
d
i
s
mo
r
e
s
ui
t
ab
l
e
f
o
r
B
at
i
k
i
ma
g
e
r
etri
ev
al
ta
s
k
.
(
a)
(
b)
F
i
gu
r
e
4.
P
erfor
m
an
c
e c
o
m
pa
r
i
s
on
s
i
n t
erms
of
prec
i
s
i
on
a
nd
r
ec
a
l
l
r
a
tes
ov
er v
ario
us
d
i
s
tan
c
e
me
tr
i
c
s
wi
th
th
e i
ma
g
e f
ea
tures
f
r
om
: (a)
CNN
, a
n
d (b)
C
A
E
me
th
od
4.4
.
Co
mp
ar
i
son
again
st
F
o
r
mer
Me
t
h
o
d
s
T
hi
s
s
ub
-
s
ec
ti
o
n
s
um
ma
r
i
z
es
the
pe
r
f
orma
nc
e
c
o
mp
aris
o
n
b
etwe
en
th
e
pr
op
os
ed
s
up
erv
i
s
ed
CNN
me
t
ho
d
a
nd
form
er
ex
i
s
ti
n
g
s
c
he
me
s
on
B
at
i
k
i
m
ag
e
r
etr
i
ev
al
s
y
s
tem
.
T
hi
s
c
om
pa
r
i
s
o
n
i
s
c
on
du
c
t
ed
i
n
t
erms
of
A
v
erag
e
P
r
ec
i
s
i
on
Rec
a
l
l
(
A
P
R)
s
c
ore.
T
he
A
P
R
i
s
forma
l
l
y
d
efi
ne
d
as
:
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
us
i
ng
c
on
v
ol
ut
i
o
na
l
ne
ural
ne
t
wor
k
(
Her
i
P
r
as
ety
o
)
3017
=
1
∑
(
)
=
1
(
9
)
where
(
)
and
are
th
e
r
ec
al
l
r
ate
f
or
qu
ery
i
ma
ge
an
d
the
t
ota
l
n
um
b
er
of
i
ma
ge
s
i
n
da
ta
ba
s
e,
r
es
pe
c
ti
v
e
l
y
.
Her
ei
n,
a
l
l
i
m
ag
es
i
n
da
tab
as
e
are
turn
ed
as
q
ue
r
y
i
ma
g
e
i
nd
i
c
at
i
n
g
th
at
=
1552
.
T
hu
s
,
t
he
A
P
R
v
al
ue
i
s
av
eragi
ng
ov
er
a
l
l
qu
ery
i
ma
ge
s
.
T
he
nu
mb
er
of
r
et
r
i
ev
ed
i
ma
ge
s
i
s
s
et
as
16
y
i
el
di
n
g
=
16
.
T
o
ma
k
e
a
fa
i
r
c
om
pa
r
i
s
on
,
t
hi
s
ex
p
eri
me
nt
al
s
o
i
nv
es
ti
ga
tes
t
he
di
me
ns
i
on
al
i
ty
of
i
ma
ge
f
ea
t
ure.
T
ab
l
e
4
r
ep
orts
the
pe
r
f
orm
an
c
e c
o
mp
ar
i
s
on
i
n
te
r
ms
of
fe
atu
r
e
d
i
me
ns
i
o
na
l
i
ty
an
d A
P
R
v
al
ue
.
A
s
s
ho
wn
i
n
t
hi
s
t
a
bl
e,
the
prop
os
ed
s
u
pe
r
v
i
s
ed
CNN
y
i
e
l
ds
th
e
b
es
t
p
erfor
ma
nc
e
i
n
c
om
pa
r
i
s
o
n
wi
t
h
the
o
the
r
c
om
pe
t
i
ng
s
c
he
me
s
.
It
i
s
no
tew
orthy
tha
t
th
e
prop
os
ed
me
t
ho
d
r
eq
ui
r
es
l
owes
t
fea
t
ure
di
me
ns
i
on
a
l
i
ty
(
wi
t
h
ex
c
ep
ti
o
na
l
on
c
om
pa
r
i
s
o
n
t
o
LB
P
[20
]
s
c
he
me
)
.
T
hi
s
l
o
wer
d
i
me
ns
i
o
na
l
i
ty
i
nd
i
c
ate
s
t
he
fas
t
er
proc
e
s
s
on
K
NN
s
e
arc
hi
n
g
for
eff
ec
ti
v
e
B
a
ti
k
i
ma
ge
r
etr
i
ev
a
l
s
y
s
tem
.
T
hu
s
,
the
prop
os
ed
me
t
ho
d
c
an
b
e
c
on
s
i
de
r
ed
on
i
mp
l
e
me
n
ti
n
g
the
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
a
n
d c
l
as
s
i
f
i
c
ati
on
s
y
s
tem
.
T
ab
l
e
3.
A
P
R CNN
an
d CA
E
Met
h
o
d
E
u
c
li
d
e
a
n
Man
h
a
t
t
a
n
B
r
a
y
-
c
u
r
t
is
CNN
0
.
9
9
3
8
0
.
9
9
3
1
0
.
9
9
4
7
C
A
E
0
.
6
7
3
7
0
.
6
3
8
7
0
.
7
6
5
4
T
ab
l
e
4.
A
P
R Co
mp
ar
i
s
on
wi
th
F
ormer
M
eth
od
Met
h
o
d
Fea
t
u
r
e
S
i
z
e
APR
(
%
)
L
B
P
[
2
4
]
59
9
2
.
5
7
L
TP
[
2
5
]
118
9
5
.
6
5
C
L
B
P
[
2
6
]
118
9
5
.
1
7
L
D
P
[
2
7
]
236
9
3
.
5
2
Ga
b
o
r
Fil
t
e
r
[
2
8
]
144
9
6
.
5
5
OD
B
TC
+P
S
O
[
3
]
384
9
7
.
6
8
P
r
o
p
o
s
e
d
S
u
p
e
r
v
i
s
e
d
C
N
N
97
9
9
.
4
7
5.
Co
n
clus
ion
s
A
n
ew
c
on
ten
t
-
ba
s
ed
i
m
a
ge
r
etr
i
ev
al
s
y
s
tem
h
as
b
ee
n
pres
en
t
ed
i
n
thi
s
pa
p
er.
T
hi
s
s
y
s
tem
ac
hi
ev
es
t
he
r
etri
e
v
al
ac
c
urac
i
es
9
9.4
7%
an
d
76
.
54
%,
r
es
p
ec
ti
v
e
l
y
,
w
hi
l
e
th
e
i
ma
g
e
fea
ture
i
s
c
o
ns
tr
uc
ted
fr
o
m
CNN
an
d
C
A
E
d
ee
p
l
e
arni
n
g
-
ba
s
ed
arc
hi
tec
ture
on
B
at
i
k
i
m
ag
e
da
ta
ba
s
e.
T
h
e
CNN
o
utp
e
r
forms
the
form
er
ex
i
s
ti
ng
s
c
he
me
s
i
n
t
erms
of
r
etri
e
v
al
ac
c
urac
y
.
In
a
dd
i
ti
o
n,
i
t
r
eq
u
i
r
es
t
he
l
owes
t
i
ma
g
e
f
ea
tures
,
i
.
e.
97
fea
ture
di
me
ns
i
on
al
i
ty
,
c
om
p
ared
to
oth
er
me
t
ho
ds
.
F
or
fut
ure
wor
k
,
a
s
l
i
g
ht
mo
d
i
f
i
c
ati
o
n
c
an
b
e
c
arr
i
ed
ou
t
for
CA
E
m
od
e
l
by
ad
d
i
ng
f
ul
l
y
-
c
on
ne
c
te
d
l
ay
ers
be
fore
an
d
aft
er
th
e
ne
ura
l
c
od
e
s
ec
ti
on
.
T
h
i
s
s
c
en
ario
ma
y
r
ed
uc
e
t
he
di
m
en
s
i
on
a
l
i
ty
of
i
ma
g
e
fe
atu
r
e
,
at
the
s
a
me
t
i
me
,
i
t
i
mp
r
ov
es
t
he
p
erfor
ma
nc
e
for
B
at
i
k
i
m
ag
e
r
etri
ev
a
l
.
Ref
er
en
ce
s
[1
]
J
o
h
n
s
o
n
M
H.
Th
e
n
e
u
ra
l
b
a
s
i
s
o
f
c
o
g
n
i
ti
v
e
d
e
v
e
l
o
p
m
e
n
t.
In
:
Dam
o
n
W.
Ed
i
to
r
.
Ha
n
d
b
o
o
k
o
f
c
h
i
l
d
p
s
y
c
h
o
l
o
g
y
:
Cog
n
i
ti
o
n
,
p
e
rc
e
p
ti
o
n
,
a
n
d
l
a
n
g
u
a
g
e
.
H
o
b
o
k
e
n
:
J
o
h
n
W
i
l
e
y
&
So
n
s
In
c
.
1
9
9
8
:
1
-
4
9
.
[2
]
L
i
u
P,
e
t
a
l
.
Fu
s
i
o
n
o
f
d
e
e
p
l
e
a
rn
i
n
g
a
n
d
c
o
m
p
re
s
s
e
d
d
o
m
a
i
n
fe
a
tu
re
s
fo
r
c
o
n
te
n
t
-
b
a
s
e
d
i
m
a
g
e
re
tri
e
v
a
l
.
IEEE
Tra
n
s
a
c
ti
o
n
s
o
n
I
m
a
g
e
Pro
c
e
s
s
i
n
g
.
2
0
1
7
.
2
6
(
1
2
)
:
5
7
0
6
-
5
7
1
7
.
[3
]
Pra
s
e
t
y
o
,
H
,
e
t
a
l
.
Ba
t
i
k
I
m
a
g
e
Ret
ri
e
v
a
l
Us
i
n
g
O
DBTC
Fe
a
tu
re
a
n
d
Pa
rti
c
l
e
Swarm
O
p
ti
m
i
z
a
t
i
o
n
.
J
o
u
rn
a
l
o
f
Te
l
e
c
o
m
m
u
n
i
c
a
ti
o
n
,
El
e
c
tr
o
n
i
c
Co
m
p
u
t
e
r En
g
i
n
e
e
r
i
n
g
.
2
0
1
8
.
1
0
(
2
-
4
):
7
1
-
7
4
.
[4
]
Dat
ta
R,
L
i
J
,
Wa
n
g
JZ
.
Co
n
te
n
t
-
b
a
s
e
d
i
m
a
g
e
re
tr
i
e
v
a
l
:
a
p
p
ro
a
c
h
e
s
a
n
d
tr
e
n
d
s
o
f
th
e
n
e
w
a
g
e
.
Pro
c
e
e
d
i
n
g
s
o
f
th
e
7
th
AC
M
SI
G
M
M
i
n
te
rn
a
ti
o
n
a
l
wo
rk
s
h
o
p
o
n
M
u
l
ti
m
e
d
i
a
i
n
fo
r
m
a
ti
o
n
re
tri
e
v
a
l
.
2
0
0
5
.
[5
]
Ea
k
i
n
s
J
P,
G
ra
h
a
m
M
E
.
Con
t
e
n
t
b
a
s
e
d
i
m
a
g
e
re
tr
i
e
v
a
l
:
A
r
e
p
o
rt
to
t
h
e
J
ISC
te
c
h
n
o
l
o
g
y
a
p
p
l
i
c
a
t
i
o
n
s
p
ro
g
ra
m
m
e
.
1
9
9
9
.
[6
]
Rus
s
a
k
o
v
s
k
y
O
,
e
t
a
l
.
I
m
a
g
e
n
e
t
l
a
rg
e
s
c
a
l
e
v
i
s
u
a
l
re
c
o
g
n
i
ti
o
n
c
h
a
l
l
e
n
g
e
.
In
te
rn
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Com
p
u
te
r Vi
s
i
o
n
.
2
0
1
5
;
1
1
5
(3
)
:
211
-
2
5
2
.
[7
]
Kri
z
h
e
v
s
k
y
A
,
S
u
ts
k
e
v
e
r
I
,
Hin
to
n
GE
.
Im
a
g
e
n
e
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
wit
h
d
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
.
Ad
v
a
n
c
e
s
i
n
n
e
u
ra
l
i
n
fo
rm
a
ti
o
n
p
ro
c
e
s
s
i
n
g
s
y
s
te
m
s
.
2
0
1
2
:
1097
-
1
1
0
5
.
[8
]
Si
m
o
n
y
a
n
K
,
Z
i
s
s
e
r
m
a
n
A.
Ve
ry
d
e
e
p
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
two
r
k
s
fo
r
l
a
rg
e
-
s
c
a
l
e
i
m
a
g
e
re
c
o
g
n
i
ti
o
n
.
a
rXi
v
p
re
p
ri
n
t
a
rXi
v
.
2
0
1
4
.
[9
]
Sz
e
g
e
d
y
C
,
e
t
a
l
.
G
o
i
n
g
d
e
e
p
e
r
wit
h
c
o
n
v
o
l
u
ti
o
n
s
.
Pro
c
e
e
d
i
n
g
s
o
f
th
e
IEEE
c
o
n
f
e
re
n
c
e
o
n
c
o
m
p
u
te
r
v
i
s
i
o
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
i
t
i
o
n
.
2
0
1
5
:
1
-
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
b
er 20
19
:
30
1
0
-
3018
3018
[1
0
]
He
K
,
e
t
a
l
.
De
e
p
re
s
i
d
u
a
l
l
e
a
rn
i
n
g
fo
r
i
m
a
g
e
r
e
c
o
g
n
i
t
i
o
n
.
Pro
c
e
e
d
i
n
g
s
o
f
th
e
IEEE
c
o
n
f
e
re
n
c
e
o
n
c
o
m
p
u
te
r
v
i
s
i
o
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
i
ti
o
n
.
2
0
1
6
:
770
-
7
7
8
.
[1
1
]
Cov
e
r
T
,
Har
t
P.
Nea
r
e
s
t
n
e
i
g
h
b
o
r
p
a
tt
e
r
n
c
l
a
s
s
i
f
i
c
a
ti
o
n
.
I
EEE
tra
n
s
a
c
ti
o
n
s
o
n
i
n
fo
rm
a
t
i
o
n
th
e
o
ry
.
1967
;
1
3
(1
)
:
21
-
27.
[1
2
]
Pe
ts
c
h
a
rn
i
g
S,
L
u
x
M
,
Ch
a
tz
i
c
h
ri
s
to
f
i
s
S
.
D
i
m
e
n
s
i
o
n
a
l
i
t
y
r
e
d
u
c
ti
o
n
f
o
r
i
m
a
g
e
fe
a
t
u
re
s
u
s
i
n
g
d
e
e
p
l
e
a
rn
i
n
g
a
n
d
a
u
to
e
n
c
o
d
e
r
s
.
Pro
c
e
e
d
i
n
g
s
o
f
th
e
1
5
th
In
te
rn
a
ti
o
n
a
l
Wo
r
k
s
h
o
p
o
n
Con
t
e
n
t
-
Ba
s
e
d
M
u
l
t
i
m
e
d
i
a
I
n
d
e
x
i
n
g
,
AC
M
.
2
0
1
7
.
[1
3
]
M
a
s
c
i
J
,
e
t
a
l
.
Sta
c
k
e
d
c
o
n
v
o
l
u
ti
o
n
a
l
a
u
to
-
e
n
c
o
d
e
r
s
fo
r
h
i
e
r
a
rc
h
i
c
a
l
fe
a
t
u
re
e
x
tra
c
ti
o
n
.
In
t
e
rn
a
ti
o
n
a
l
Con
fe
re
n
c
e
o
n
A
rt
i
fi
c
i
a
l
Ne
u
ra
l
Ne
two
rk
s
.
2
0
1
1
:
52
-
59
.
[1
4
]
Wa
n
g
R
,
e
t
a
l
.
A
Cro
p
Pe
s
ts
Im
a
g
e
C
l
a
s
s
i
f
i
c
a
ti
o
n
Al
g
o
ri
th
m
Ba
s
e
d
o
n
Dee
p
Co
n
v
o
l
u
ti
o
n
a
l
Neu
ra
l
N
e
two
rk
.
TEL
KO
M
NI
KA
Te
l
e
c
o
m
m
u
n
i
c
a
ti
o
n
C
o
m
p
u
ti
n
g
El
e
c
tr
o
n
i
c
s
a
n
d
Con
t
ro
l
.
2
0
1
7
;
1
5
(3
)
:
1239
-
1
2
4
6
.
[1
5
]
Ba
h
a
ri
n
A
,
Ab
d
u
l
l
a
h
A
,
Yo
u
s
o
ff
SNM
.
Pre
d
i
c
ti
o
n
o
f
B
i
o
p
ro
c
e
s
s
Pr
o
d
u
c
ti
o
n
Us
i
n
g
D
e
e
p
Neu
r
a
l
Net
work
M
e
th
o
d
.
TE
L
KO
M
NIKA
Te
l
e
c
o
m
m
u
n
i
c
a
ti
o
n
Co
m
p
u
ti
n
g
E
l
e
c
tro
n
i
c
s
a
n
d
C
o
n
tro
l
.
2
0
1
7
;
1
5
(2
)
:
805
-
8
1
3
.
[1
6
]
Su
d
i
a
t
m
i
k
a
IBK
,
Ra
h
m
a
n
F
,
Tri
s
n
o
T,
Su
y
o
t
o
S
.
Im
a
g
e
fo
r
g
e
ry
d
e
t
e
c
t
i
o
n
u
s
i
n
g
e
rr
o
r
l
e
v
e
l
a
n
a
l
y
s
i
s
a
n
d
d
e
e
p
l
e
a
r
n
i
n
g
.
TEL
KO
M
NIKA
Te
l
e
c
o
m
m
u
n
i
c
a
ti
o
n
Co
m
p
u
ti
n
g
E
l
e
c
tro
n
i
c
s
a
n
d
Con
tro
l
.
2
0
1
9
;
1
7
(2
)
:
653
-
6
5
9
.
[1
7
]
Se
ti
a
wan
W
,
Uto
y
o
MI
,
Ru
l
a
n
i
n
g
t
y
a
s
R.
C
l
a
s
s
i
f
i
c
a
ti
o
n
o
f
n
e
o
v
a
s
c
u
l
a
ri
z
a
t
i
o
n
u
s
i
n
g
c
o
n
v
o
l
u
ti
o
n
a
l
neur
a
l
n
e
two
r
k
m
o
d
e
l
.
TEL
KO
M
NIKA
Te
l
e
c
o
m
m
u
n
i
c
a
t
i
o
n
C
o
m
p
u
ti
n
g
El
e
c
tro
n
i
c
s
a
n
d
Con
tro
l
.
2
0
1
9
;
1
7
(1
)
:
463
-
4
7
2
.
[1
8
]
M
e
n
g
Q
,
e
t
a
l
.
Re
l
a
ti
o
n
a
l
a
u
to
e
n
c
o
d
e
r
fo
r
f
e
a
tu
r
e
e
x
tra
c
ti
o
n
.
2
0
1
7
In
te
rn
a
ti
o
n
a
l
J
o
i
n
t
Con
f
e
re
n
c
e
o
n
Neu
ra
l
Ne
tw
o
rk
s
(I
J
CN
N).
2
0
1
7
:
364
-
3
7
1
.
[1
9
]
Ki
n
g
m
a
DP,
Ba
J
.
Ad
a
m
:
A
m
e
th
o
d
fo
r
s
to
c
h
a
s
ti
c
o
p
t
i
m
i
z
a
ti
o
n
.
a
rX
i
v
p
r
e
p
ri
n
t
a
rX
i
v
.
2
0
1
4
.
[2
0
]
Hag
a
n
M
T
,
M
e
n
h
a
j
MB
.
Tr
a
i
n
i
n
g
fe
e
d
fo
rward
n
e
two
r
k
s
wit
h
th
e
M
a
rq
u
a
rd
t
a
l
g
o
ri
t
h
m
.
IEE
E
tra
n
s
a
c
t
i
o
n
s
o
n
N
e
u
ra
l
Ne
two
r
k
s
.
1
9
9
4
;
5
(6
)
:
9
8
9
-
9
9
3
.
[2
1
]
Dan
i
e
l
s
s
o
n
PE
.
Eu
c
l
i
d
e
a
n
d
i
s
ta
n
c
e
m
a
p
p
i
n
g
.
Com
p
u
te
r
G
ra
p
h
i
c
s
i
m
a
g
e
p
r
o
c
e
s
s
i
n
g
.
1980
;
1
4
(3
)
:
2
2
7
-
2
4
8
.
[2
2
]
Craw
S
.
M
a
n
h
a
tt
a
n
d
i
s
t
a
n
c
e
.
In
:
Sa
m
m
u
t
C,
We
b
b
G
I.
En
c
y
c
l
o
p
e
d
i
a
o
f
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
a
n
d
Dat
a
M
i
n
i
n
g
.
S
p
ri
n
g
e
r.
2
0
1
7
:
7
9
0
-
7
9
1
.
[2
3
]
Ko
k
a
re
M
,
Cha
t
te
rj
i
B
,
Bi
s
w
a
s
P
.
Com
p
a
ri
s
o
n
o
f
s
i
m
i
l
a
r
i
ty
m
e
tr
i
c
s
f
o
r
te
x
tu
r
e
i
m
a
g
e
re
tri
e
v
a
l
.
TENCO
N
2
0
0
3
.
IEEE
,
Co
n
fe
r
e
n
c
e
o
n
Co
n
v
e
r
g
e
n
t
Te
c
h
n
o
l
o
g
i
e
s
fo
r
t
h
e
As
i
a
-
Pa
c
i
f
i
c
Re
g
i
o
n
.
2
0
0
3
;
2
:
571
-
5
7
5
.
[2
4
]
O
j
a
l
a
T
,
Pi
e
ti
k
a
i
n
e
n
M
,
M
a
e
n
p
a
a
T
.
M
u
l
ti
re
s
o
l
u
ti
o
n
g
r
a
y
-
s
c
a
l
e
a
n
d
r
o
ta
t
i
o
n
i
n
v
a
ri
a
n
t
t
e
x
tu
r
e
c
l
a
s
s
i
fi
c
a
ti
o
n
wi
th
l
o
c
a
l
b
i
n
a
ry
p
a
tt
e
rn
s
.
IEEE
Tr
a
n
s
a
c
t
i
o
n
s
o
n
p
a
tt
e
rn
a
n
a
l
y
s
i
s
m
a
c
h
i
n
e
i
n
t
e
l
l
i
g
e
n
c
e
.
2002
;
2
4
(7
)
:
9
7
1
-
9
8
7
.
[2
5
]
Ta
n
X
,
Tr
i
g
g
s
B.
En
h
a
n
c
e
d
l
o
c
a
l
te
x
t
u
re
fe
a
tu
re
s
e
ts
fo
r
fa
c
e
re
c
o
g
n
i
ti
o
n
u
n
d
e
r
d
i
ff
i
c
u
l
t
l
i
g
h
ti
n
g
c
o
n
d
i
ti
o
n
s
.
IEEE
tra
n
s
a
c
ti
o
n
s
o
n
i
m
a
g
e
p
ro
c
e
s
s
i
n
g
.
2
0
1
0
;
1
9
(6
)
:
1
6
3
5
-
1
6
5
0
.
[2
6
]
G
u
o
Z
,
Zh
a
n
g
L
,
Z
h
a
n
g
D
.
A
c
o
m
p
l
e
t
e
d
m
o
d
e
l
i
n
g
o
f
l
o
c
a
l
b
i
n
a
ry
p
a
tt
e
r
n
o
p
e
r
a
to
r
fo
r
te
x
tu
r
e
c
l
a
s
s
i
f
i
c
a
ti
o
n
.
IEEE
Tra
n
s
a
c
ti
o
n
s
o
n
I
m
a
g
e
Pr
o
c
e
s
s
i
n
g
.
2
0
1
0
;
1
9
(6
)
:
1
6
5
7
-
1
6
6
3
.
[2
7
]
Zh
a
n
g
B
,
e
t
a
l
.
L
o
c
a
l
d
e
r
i
v
a
ti
v
e
p
a
tt
e
rn
v
e
r
s
u
s
l
o
c
a
l
b
i
n
a
ry
p
a
tt
e
r
n
:
fa
c
e
re
c
o
g
n
i
t
i
o
n
wit
h
h
i
g
h
-
o
rd
e
r
l
o
c
a
l
p
a
tt
e
rn
d
e
s
c
ri
p
to
r.
IEEE
t
ra
n
s
a
c
ti
o
n
s
o
n
i
m
a
g
e
p
ro
c
e
s
s
i
n
g
.
2
0
1
0
;
1
9
(2
)
:
533
-
544.
[2
8
]
Pra
s
e
t
y
o
H,
Wi
r
a
n
to
W
,
Wi
n
a
r
n
o
W
.
Sta
t
i
s
t
i
c
a
l
M
o
d
e
l
i
n
g
o
f
G
a
b
o
r
Fi
l
te
r
e
d
M
a
g
n
i
tu
d
e
f
o
r
Ba
ti
k
Im
a
g
e
Ret
ri
e
v
a
l
.
J
o
u
r
n
a
l
o
f
T
e
l
e
c
o
m
m
u
n
i
c
a
t
i
o
n
,
E
l
e
c
tro
n
i
c
C
o
m
p
u
t
e
r En
g
i
n
e
e
r
i
n
g
.
2
0
1
8
;
1
0
(2
-
4)
:
85
-
89.
Evaluation Warning : The document was created with Spire.PDF for Python.