T
E
L
K
O
M
N
I
K
A
T
elec
o
m
m
un
ica
t
io
n,
Co
m
pu
t
ing
,
E
lect
ro
nics
a
nd
Co
ntr
o
l
Vo
l.
19
,
No
.
3
,
J
u
n
e
2
0
2
1
,
p
p
.
8
5
8
~8
7
1
I
SS
N:
1
6
9
3
-
6
9
3
0
,
ac
cr
ed
ited
First
Gr
ad
e
b
y
Kem
en
r
is
tek
d
i
k
ti,
Dec
r
ee
No
: 2
1
/E/KPT
/2
0
1
8
DOI
: 1
0
.
1
2
9
2
8
/TE
L
KOM
NI
K
A.
v
1
9
i3
.
1
8
1
5
7
858
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//jo
u
r
n
a
l.u
a
d
.
a
c.
id
/in
d
ex
.
p
h
p
/TELK
OM
N
I
K
A
A comp
a
ra
tive a
n
a
ly
sis
of auto
ma
tic
deep
neural n
et
wo
rks for
ima
g
e re
trieva
l
H
a
na
n A.
Al
-
J
ub
o
uri
,
Sa
wsa
n M
.
M
a
hm
m
o
d
Co
m
p
u
ter E
n
g
in
e
e
rin
g
De
p
t.
Co
ll
e
g
e
o
f
E
n
g
i
n
e
e
rin
g
,
M
u
sta
n
siriy
a
h
Un
i
v
e
rsity
,
Ba
g
h
d
a
d
,
I
ra
q
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 1
0
,
2
0
2
0
R
ev
is
ed
Sep
9
,
2
0
2
0
Acc
ep
ted
Sep
2
0
,
2
0
2
0
F
e
a
tu
re
d
e
sc
rip
to
r
a
n
d
sim
il
a
rit
y
m
e
a
su
re
s
a
re
th
e
two
c
o
re
c
o
m
p
o
n
e
n
ts
i
n
c
o
n
ten
t
-
b
a
se
d
ima
g
e
re
tri
e
v
a
l
a
n
d
c
r
u
c
ial
issu
e
s
d
u
e
to
“
se
m
a
n
ti
c
g
a
p
”
b
e
twe
e
n
h
u
m
a
n
c
o
n
c
e
p
tu
a
l
m
e
a
n
in
g
a
n
d
a
m
a
c
h
in
e
l
o
w
-
lev
e
l
fe
a
tu
re
.
Re
c
e
n
tl
y
,
d
e
e
p
lea
rn
i
n
g
tec
h
n
i
q
u
e
s
h
a
v
e
s
h
o
wn
a
g
re
a
t
in
tere
st
in
ima
g
e
re
c
o
g
n
it
i
o
n
e
sp
e
c
iall
y
in
e
x
trac
ti
n
g
fe
a
tu
re
s
in
f
o
rm
a
ti
o
n
a
b
o
u
t
t
h
e
ima
g
e
s.
In
th
is
p
a
p
e
r,
we
in
v
e
stig
a
ted
,
c
o
m
p
a
re
d
,
a
n
d
e
v
a
lu
a
ted
d
iff
e
re
n
t
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s
a
n
d
th
e
ir
a
p
p
li
c
a
ti
o
n
s
f
o
r
ima
g
e
c
las
sifica
ti
o
n
a
n
d
a
u
to
m
a
ti
c
ima
g
e
re
tri
e
v
a
l.
Th
e
a
p
p
ro
a
c
h
e
s
a
re
:
sim
p
le
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
,
Ale
x
Ne
t,
G
o
o
g
l
e
Ne
t,
Re
sN
e
t
-
5
0
,
V
g
g
-
1
6
,
a
n
d
V
g
g
-
1
9
.
We
co
m
p
a
re
d
th
e
p
e
rfo
rm
a
n
c
e
o
f
t
h
e
d
iffere
n
t
a
p
p
ro
a
c
h
e
s
t
o
p
r
io
r
w
o
rk
s
in
th
is
d
o
m
a
in
b
y
u
si
n
g
k
n
o
wn
a
c
c
u
r
a
c
y
m
e
tri
c
s
a
n
d
a
n
a
ly
z
e
d
th
e
d
iffere
n
c
e
s
b
e
twe
e
n
th
e
a
p
p
r
o
a
c
h
e
s.
Th
e
p
e
rfo
rm
a
n
c
e
s
o
f
th
e
se
a
p
p
r
o
a
c
h
e
s
a
re
in
v
e
stig
a
te
d
u
si
n
g
p
u
b
l
ic
ima
g
e
d
a
tas
e
ts
c
o
re
l
1K
,
c
o
re
l
1
0
K,
a
n
d
C
a
lt
e
c
h
2
5
6
.
He
n
c
e
,
we
d
e
d
u
c
e
d
t
h
a
t
G
o
o
g
leN
e
t
a
p
p
r
o
a
c
h
y
ield
s
th
e
b
e
st
o
v
e
ra
ll
re
su
lt
s.
In
a
d
d
it
i
o
n
,
we
i
n
v
e
sti
g
a
ted
a
n
d
c
o
m
p
a
re
d
d
iffere
n
t
sim
il
a
rit
y
m
e
a
su
re
s.
Ba
se
d
o
n
e
x
h
a
u
ste
d
m
e
n
ti
o
n
e
d
in
v
e
stig
a
ti
o
n
s,
we
d
e
v
e
l
o
p
e
d
a
n
o
v
e
l
a
lg
o
rit
h
m
f
o
r
ima
g
e
re
tri
e
v
a
l
.
K
ey
w
o
r
d
s
:
C
o
n
ten
t
-
b
ased
im
ag
e
r
etr
iev
al
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
I
m
ag
e
class
if
icatio
n
I
m
ag
e
r
etr
iev
al
d
ee
p
lear
n
in
g
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r
th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Han
an
A.
Al
-
J
u
b
o
u
r
i
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
E
n
g
i
n
ee
r
in
g
Mu
s
tan
s
ir
iy
ah
Un
iv
er
s
ity
Palest
in
e
Stre
et,
B
ag
h
d
ad
,
I
r
a
q
E
m
ail:
h
an
an
alju
b
o
u
r
i@
u
o
m
u
s
tan
s
ir
iy
ah
.
ed
u
.
iq
1.
I
NT
RO
D
UCT
I
O
N
T
o
d
ay
,
d
i
g
ital
p
h
o
to
g
r
ap
h
ic
d
ev
ices
ar
e
wid
ely
u
s
ed
r
esu
lti
n
g
lar
g
e
v
o
lu
m
es
o
f
d
ig
ital
im
ag
es
h
av
e
b
ein
g
ac
q
u
ir
e
d
an
d
s
to
r
e
d
in
d
atab
ases
in
d
if
f
e
r
en
t
f
ield
s
s
u
c
h
as
s
cien
tific
r
esear
c
h
,
m
ed
ic
al,
f
o
r
e
n
s
ic
an
aly
s
is
,
an
d
s
o
cial
n
etwo
r
k
in
g
.
So
,
th
e
r
etr
iev
al
o
f
t
h
es
e
im
ag
es
s
h
o
u
ld
b
e
d
o
n
e
ef
f
ec
tiv
el
y
an
d
f
ast.
I
n
f
o
r
m
atio
n
r
etr
iev
al
(
I
R
)
attem
p
ts
to
f
in
d
m
ater
ial
s
u
ch
as
im
ag
es
o
r
tex
ts
(
d
o
cu
m
en
ts
)
wh
ich
h
a
v
e
u
n
s
tr
u
ctu
r
ed
f
o
r
m
to
g
et
in
f
o
r
m
atio
n
f
r
o
m
lar
g
e
v
o
lu
m
e
o
f
th
ese
m
ater
ials
[
1
,
2
]
.
I
n
ea
r
l
y
im
ag
e
r
etr
iev
al
s
y
s
tem
s
,
im
ag
es
ar
e
in
d
ex
ed
in
a
d
atab
ase
u
s
in
g
t
ex
tu
al
an
n
o
tatio
n
s
u
ch
as
k
ey
wo
r
d
s
o
r
p
h
r
ases
.
A
u
s
er
ask
s
th
e
s
y
s
tem
to
f
in
d
s
im
ilar
im
ag
es
b
y
en
ter
in
g
th
e
tex
tu
al
an
n
o
tatio
n
an
d
th
e
s
y
s
tem
r
etr
iev
es
im
ag
es
in
o
r
d
er
ac
co
r
d
i
n
g
to
t
h
e
d
eg
r
ee
o
f
m
atch
to
th
e
a
n
n
o
t
atio
n
.
Ho
wev
er
,
s
o
m
e
lim
itati
o
n
s
f
ac
e
s
u
ch
a
m
eth
o
d
.
Fo
r
in
s
tan
ce
,
it
is
tim
e
co
n
s
u
m
in
g
t
o
an
n
o
tate
im
ag
es in
a
lar
g
e
-
s
ca
le
d
atab
ase
m
an
u
ally
an
d
th
e
tex
t m
ay
n
o
t a
v
a
ilab
le
d
u
r
in
g
im
ag
e
ca
p
tu
r
in
g
r
esp
ec
tiv
ely
.
C
o
n
s
e
q
u
en
tly
,
co
n
te
n
t
-
b
ased
im
ag
e
r
etr
iev
a
l
(
C
B
I
R
)
is
a
p
r
o
ce
s
s
th
at
e
x
tr
ac
t
im
a
g
e
f
ea
tu
r
e
(
v
is
u
al
c
o
n
ten
t)
to
r
ep
r
esen
t im
ag
es a
u
to
m
atica
lly
an
d
in
d
ex
th
em
in
a
d
atab
ase
[
3
]
.
Fig
u
r
e
1
illu
s
tr
ates
a
ty
p
ical
d
iag
r
am
o
f
C
B
I
R
s
y
s
tem
th
at
s
t
o
r
es
im
ag
es
in
th
e
d
atab
ase
b
y
ex
tr
ac
tin
g
im
ag
e
f
ea
tu
r
es
at
o
f
f
-
lin
e
p
h
ase
[
4
]
.
Me
an
wh
ile,
t
h
e
s
y
s
tem
e
x
tr
ac
ts
a
f
ea
tu
r
e
v
ec
to
r
f
r
o
m
a
q
u
er
y
im
ag
e
i
n
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
A
co
mp
a
r
a
tive
a
n
a
lysi
s
o
f a
u
t
o
ma
tic
d
ee
p
n
eu
r
a
l n
etw
o
r
ks fo
r
ima
g
e
r
etri
ev
a
l
(
Ha
n
a
n
A
.
A
l
-
Ju
b
o
u
r
i
)
859
s
am
e
way
an
d
co
m
p
a
r
es
it
with
th
e
im
ag
e
f
ea
tu
r
es
in
th
e
d
at
ab
ase
u
s
in
g
a
s
im
ilar
ity
m
ea
s
u
r
e.
T
h
e
m
o
s
t
s
im
ilar
im
ag
es
ar
e
o
r
d
er
ed
in
r
a
n
k
ed
l
is
t
an
d
r
etu
r
n
ed
to
th
e
u
s
er
at
o
n
-
lin
e
p
h
ase.
Hen
ce
,
s
o
m
e
ir
r
elev
an
t
im
ag
es
ar
e
r
etr
iev
ed
in
th
e
r
an
k
e
d
lis
t
d
u
e
to
a
ch
allen
g
e
s
o
-
ca
lled
“sem
an
tic
g
ap
”
wh
i
ch
is
th
e
g
ap
b
etwe
en
h
ig
h
an
d
lo
w
lev
el
f
ea
tu
r
es
in
m
ea
n
in
g
[
4
]
.
T
h
er
ef
o
r
e,
th
e
aim
o
f
r
esear
ch
e
r
s
in
C
B
I
R
is
h
o
w
to
d
ev
el
o
p
a
s
y
s
tem
o
r
alg
o
r
ith
m
th
at
ca
n
b
r
id
g
e
th
e
s
em
an
tic
g
ap
b
etwe
en
h
u
m
an
co
n
ce
p
tu
al
m
ea
n
in
g
f
o
r
im
ag
es
an
d
m
ac
h
in
es
s
u
ch
as
a
co
m
p
u
ter
.
I
n
o
t
h
er
wo
r
d
s
,
h
o
w
th
e
C
B
I
R
s
y
s
tem
ca
n
ex
tr
ac
t
ef
f
ec
tiv
e
f
ea
tu
r
es
th
at
r
e
p
r
esen
t
th
e
im
ag
e
in
th
e
d
atab
ase
an
d
r
etr
ie
v
e
in
ter
m
s
o
f
r
elev
a
n
t
im
ag
es
.
Fig
u
r
e
1
.
T
y
p
ical
d
iag
r
am
o
f
C
B
I
R
s
y
s
tem
T
h
e
m
ain
c
o
n
tr
ib
u
tio
n
s
o
f
th
i
s
p
ap
er
ar
e
as
f
o
llo
ws:
f
i
r
s
t,
c
o
n
v
o
lu
ti
o
n
n
e
u
r
al
n
etwo
r
k
s
(
C
NNs
)
ar
e
in
v
esti
g
ated
to
class
if
y
h
u
g
e
am
o
u
n
t
o
f
im
ag
es.
I
n
o
u
r
in
v
esti
g
atio
n
,
d
if
f
er
e
n
t
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
ar
e
u
s
ed
in
class
if
icatio
n
s
u
ch
im
ag
es.
Seco
n
d
,
th
e
C
NNs a
p
p
r
o
a
ch
es a
r
e
ex
p
lo
ited
t
o
lear
n
f
ea
tu
r
es o
f
im
ag
es f
o
r
im
ag
e
r
etr
iev
al.
T
h
ir
d
,
d
i
f
f
er
e
n
t
d
is
tan
ce
f
u
n
ctio
n
s
ar
e
test
ed
f
o
r
s
im
ilar
ity
m
ea
s
u
r
es.
T
h
e
aim
is
to
ju
d
g
e
wh
ich
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
ca
n
p
r
o
d
u
ce
ef
f
ec
tiv
e
f
ea
tu
r
es
an
d
w
h
ich
d
is
tan
ce
f
u
n
ctio
n
is
m
o
r
e
ac
cu
r
ate
to
r
ed
u
ce
th
e
s
em
an
tic
g
ap
is
s
u
e
in
C
B
I
R
.
C
o
n
s
eq
u
en
tly
,
a
n
o
v
el
alg
o
r
ith
m
f
o
r
im
a
g
e
r
etr
iev
al
is
d
ev
elo
p
ed
.
T
h
e
r
em
ain
d
er
o
f
th
is
p
ap
e
r
is
s
tr
u
ctu
r
ed
as
f
o
llo
ws.
T
h
e
r
elev
an
t
l
iter
atu
r
es
ar
e
p
r
esen
ted
i
n
s
ec
tio
n
2
.
T
h
e
p
r
o
p
o
s
ed
C
NNs
u
s
ed
in
th
is
p
ap
er
ar
e
d
escr
ib
ed
in
s
ec
tio
n
3
wh
ile
s
ec
tio
n
4
d
escr
ib
es
th
e
im
a
g
es
d
atasets
u
s
ed
in
th
e
in
v
esti
g
atio
n
an
d
p
r
esen
ts
th
e
ex
p
er
im
en
tal
r
esu
lts
an
aly
s
is
o
f
o
u
r
e
v
alu
atio
n
s
y
s
tem
.
Fin
a
lly
,
s
ec
tio
n
5
d
r
aws
th
e
f
in
d
in
g
o
f
o
u
r
p
ap
e
r
an
d
g
i
v
es a
r
ec
o
m
m
e
n
d
atio
n
f
o
r
f
u
r
t
h
er
wo
r
k
s
.
2.
RE
L
E
T
AT
E
D
WO
RK
Nu
m
er
o
u
s
s
tu
d
ies
o
f
liter
atu
r
e
h
av
e
in
v
esti
g
ated
C
NNs
in
im
ag
e
r
etr
iev
al.
I
n
th
is
s
ec
tio
n
,
we
will
p
r
esen
t so
m
e
o
f
th
e
liter
atu
r
e
s
u
s
in
g
C
NNs in
th
ese
s
tu
d
ies.
Fo
r
ex
am
p
le,
in
[
5
]
th
r
ee
C
NN
f
ea
tu
r
es f
o
r
I
R
ar
e
p
r
o
p
o
s
ed
b
y
f
u
s
in
g
th
e
p
r
o
d
u
ct
r
u
le
an
d
th
e
we
ig
h
ted
av
er
ag
e
o
f
f
ea
tu
r
es
s
im
ilar
ity
.
T
h
e
au
th
o
r
s
ex
tr
ac
t
th
e
f
ea
tu
r
es
o
f
im
a
g
es
u
s
in
g
t
h
r
ee
k
in
d
s
o
f
C
NNs.
Af
ter
th
at,
b
y
u
s
in
g
p
r
o
d
u
ct
r
u
le,
th
e
weig
h
ted
f
ea
tu
r
e
s
im
ilar
ities
b
etwe
en
th
e
q
u
er
y
an
d
d
atab
a
s
e
im
ag
e
ar
e
ca
lcu
lated
.
Fin
ally
,
th
e
r
etr
ie
v
al
r
e
s
u
lt
is
f
o
u
n
d
b
y
r
etu
r
n
in
g
t
h
e
im
ag
es
with
th
e
h
ig
h
est
to
p
-
n
s
co
r
es.
Als
o
,
in
[
6
]
,
t
h
e
f
ea
tu
r
es
o
f
th
e
im
ag
es
ar
e
e
x
tr
ac
te
d
b
y
an
aly
zin
g
th
e
class
ical
C
NN
an
d
th
en
th
e
r
esu
lts
ar
e
co
m
p
ar
ed
with
th
r
ee
class
ical
alg
o
r
ith
m
s
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
r
etr
iev
al
s
y
s
tem
is
im
p
r
o
v
ed
b
y
co
m
b
i
n
in
g
a
c
o
s
in
e
s
im
ilar
ity
m
ea
s
u
r
em
en
t a
p
p
r
o
ac
h
.
A
d
ee
p
C
NN
m
o
d
el
is
u
tili
ze
d
in
[
7
]
to
e
x
tr
ac
t
th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
f
r
o
m
th
e
ac
tiv
at
io
n
s
o
f
th
e
co
n
v
o
l
u
tio
n
al
lay
er
s
in
a
lar
g
e
im
ag
e
d
ataset
f
o
r
ap
p
licatio
n
s
s
u
ch
as
r
em
o
te
s
en
s
in
g
a
n
d
p
lan
t
b
io
lo
g
y
.
T
h
en
d
atab
ase
in
d
ex
in
g
s
tr
u
ctu
r
e
an
d
r
ec
u
r
s
iv
e
d
en
s
ity
esti
m
atio
n
ar
e
estab
lis
h
ed
to
r
etr
iev
e
th
e
im
ag
es
in
a
f
ast
an
d
ef
f
icien
t
way
.
Als
o
,
to
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
o
f
th
e
im
ag
e
r
etr
iev
al
an
d
p
r
ev
e
n
t
th
e
o
v
er
f
itti
n
g
o
f
tr
ain
in
g
a
C
NN
,
th
e
au
th
o
r
s
i
n
[
8
]
p
r
o
p
o
s
e
a
d
ee
p
C
NN
with
L
1
r
eg
u
la
r
izatio
n
an
d
an
ac
tiv
atio
n
f
u
n
c
tio
n
n
am
e
d
PR
elu
.
T
h
e
d
ee
p
n
etwo
r
k
is
s
u
cc
ess
f
u
lly
u
s
ed
to
s
im
u
late
th
e
b
r
ain
o
f
h
u
m
an
b
y
r
ec
eiv
in
g
an
d
tr
a
n
s
f
er
r
in
g
in
f
o
r
m
atio
n
an
d
it c
o
n
tain
s
a
c
o
n
v
o
lu
tio
n
o
p
er
atio
n
w
h
ich
is
ap
p
r
o
p
r
iate
in
im
ag
e
p
r
o
ce
s
s
in
g
.
I
n
[
1
]
,
d
ee
p
b
elief
n
etwo
r
k
is
in
v
esti
g
ated
a
n
d
tr
ain
ed
to
lear
n
lar
g
e
s
ca
le
r
e
p
r
esen
tatio
n
s
f
r
o
m
th
e
im
ag
es
f
o
r
ap
p
licatio
n
wh
er
e
C
B
I
R
jo
b
s
ar
e
u
s
ed
.
I
n
t
h
at
wo
r
k
,
s
im
ilar
ity
m
ea
s
u
r
es
ar
e
ap
p
lied
f
o
r
C
B
I
R
task
s
.
T
h
e
au
th
o
r
s
in
[
9
]
in
v
esti
g
ate
t
h
e
u
s
in
g
o
f
C
NN
f
o
r
C
B
I
R
jo
b
s
as
well
wh
er
e
d
if
f
er
en
t
s
ettin
g
ar
e
im
p
lem
en
ted
an
d
test
ed
.
A
h
y
b
r
id
o
f
C
NN
a
n
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
m
o
d
el
is
p
r
o
p
o
s
ed
in
[
2
]
u
s
in
g
th
e
m
in
im
u
m
n
u
m
b
er
o
f
m
ater
ials
a
n
d
tim
e
r
eso
u
r
ce
s
.
T
h
e
l
ast
o
u
tp
u
t
lay
e
r
o
f
th
e
p
r
o
p
o
s
ed
C
NN
is
ch
an
g
ed
with
a
class
if
ier
b
ased
o
n
SVM.
T
h
e
r
e
ar
e
t
wo
p
ar
ts
u
s
ed
in
th
at
w
o
r
k
,
co
n
v
o
l
u
tio
n
al
p
a
r
t
an
d
r
ec
o
g
n
itio
n
p
a
r
t.
I
n
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
3
,
J
u
n
e
2
0
2
1
:
8
5
8
-
8
7
1
860
co
n
v
o
l
u
tio
n
al
p
ar
t,
th
e
im
ag
es a
r
e
p
ass
ed
th
r
o
u
g
h
a
s
eq
u
en
c
e
o
f
s
ev
er
al
f
ilter
s
wh
er
e
n
ew
im
ag
es a
r
e
f
o
r
m
in
g
n
am
ed
co
n
v
o
lu
tio
n
m
ap
s
.
I
n
t
h
e
r
ec
o
g
n
itio
n
p
ar
t,
a
SVM
class
if
ier
is
tr
ain
ed
to
au
to
m
atica
lly
ex
tr
ac
t
f
ea
tu
r
es
o
n
test
in
g
im
a
g
es
an
d
tak
e
th
e
f
in
al
d
ec
is
io
n
s
.
A
k
in
d
o
f
d
ee
p
lear
n
in
g
is
ap
p
lied
to
class
if
y
im
ag
es
in
[
1
0
]
.
Alex
Net
d
ee
p
l
ea
r
n
in
g
n
etw
o
r
k
is
ef
f
ec
tiv
ely
u
s
ed
o
n
i
m
ag
es
s
elec
ted
f
r
o
m
I
m
ag
e
Net
d
atab
ase.
T
h
e
ex
p
er
im
en
ts
a
r
e
co
n
d
u
cte
d
o
n
th
e
im
ag
es
af
te
r
cr
o
p
p
i
n
g
im
ag
es
f
o
r
d
i
f
f
er
en
t
ar
ea
s
.
I
n
[
1
1
]
,
th
e
s
em
an
tic
f
ea
tu
r
es
o
f
th
e
im
ag
es
ar
e
e
x
tr
ac
ted
u
s
in
g
C
NN
m
o
d
el.
T
h
en
,
a
d
is
tan
ce
f
u
n
ctio
n
is
co
m
p
u
ted
to
f
in
d
th
e
s
im
ilar
ity
b
etwe
en
th
e
s
em
an
tic
f
ea
tu
r
es o
f
t
h
e
im
ag
es.
I
n
[
1
2
]
,
a
C
NN
ca
lled
C
o
n
v
Net
ar
e
tr
ain
e
d
t
o
class
if
y
m
ed
ical
im
ag
es.
T
h
e
m
ed
ical
i
m
ag
es
ar
e
ac
q
u
ir
ed
u
s
in
g
c
o
m
p
u
ted
to
m
o
g
r
ap
h
y
o
f
a
n
o
r
g
an
o
r
b
o
d
y
p
ar
t
-
s
p
ec
if
ic
an
ato
m
ical.
T
h
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
class
if
icatio
n
is
im
p
r
o
v
ed
u
s
i
n
g
d
ata
au
g
m
en
tatio
n
.
Als
o
,
d
ee
p
C
NNs
ar
e
p
r
o
p
o
s
ed
in
[
1
3
]
f
o
r
co
n
ten
t
b
ased
m
ed
ical
im
ag
e
r
etr
iev
al.
Fo
r
r
etr
iev
al
p
r
o
ce
s
s
,
two
ap
p
r
o
ac
h
es
ar
e
p
r
o
p
o
s
ed
.
T
h
e
f
ir
s
t
ap
p
r
o
ac
h
,
th
e
n
etwo
r
k
is
tr
ai
n
ed
to
g
et
th
e
p
r
e
d
ictio
n
o
f
th
e
q
u
er
y
im
ag
e
class
an
d
th
en
th
e
s
p
ec
if
ic
class
is
s
ea
r
ch
ed
f
o
r
r
elev
an
t
im
ag
es.
I
n
t
h
e
s
ec
o
n
d
ap
p
r
o
ac
h
,
th
e
w
h
o
le
d
ataset
is
s
ea
r
ch
ed
f
o
r
th
e
r
ele
v
an
t
im
ag
e
s
with
o
u
t
in
clu
d
i
n
g
in
f
o
r
m
atio
n
r
elate
d
to
th
e
q
u
er
y
im
ag
e
class
.
A
C
B
I
R
s
y
s
tem
is
b
u
ilt
u
s
in
g
a
co
m
b
i
n
atio
n
o
f
d
ee
p
f
ea
tu
r
e
s
g
en
er
ated
b
y
C
NN
an
d
SVM
to
tr
ain
a
lin
ea
r
h
y
p
er
p
la
n
e
in
[
1
4
]
.
T
h
e
au
th
o
r
s
u
s
e
C
NN
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
wh
ile
SVM
is
ap
p
lied
to
f
in
d
th
e
s
im
ilar
ity
b
etwe
en
im
ag
e
p
air
s
.
A
d
ee
p
r
ep
r
esen
tatio
n
f
o
r
im
a
g
e
r
etr
iev
al
ca
lled
r
eg
i
o
n
al
-
m
a
x
im
u
m
ac
tiv
atio
n
s
o
f
co
n
v
o
lu
tio
n
s
(R
-
MA
C
)
is
b
u
ilt
in
[
1
5
]
.
Usi
n
g
R
-
MA
C
,
a
n
u
m
b
er
o
f
im
ag
e
r
eg
io
n
s
ar
e
ag
g
r
eg
ated
in
to
a
s
m
all
an
d
f
ix
ed
len
g
th
f
ea
t
u
r
e
v
ec
to
r
r
o
b
u
s
t
h
e
n
ce
it
is
r
o
b
u
s
t
to
s
ca
le
an
d
tr
an
s
latio
n
.
T
h
i
s
d
ee
p
C
NN
g
iv
es
h
ig
h
ac
cu
r
ac
y
s
in
ce
it
ca
n
d
ea
l
with
im
ag
es
h
av
e
h
ig
h
r
eso
l
u
tio
n
o
f
d
if
f
er
e
n
t
r
atio
s
.
I
n
[
1
6
]
,
a
C
NN
m
o
d
el
is
tr
ain
ed
o
n
I
m
ag
eNe
t
-
2
0
1
2
.
T
h
en
,
f
o
r
C
B
I
R
task
,
th
e
f
o
u
r
lay
er
s
,
wh
ich
a
r
e
ex
tr
a
cted
as
th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
o
f
th
e
d
ata,
ar
e
ev
alu
a
ted
u
s
in
g
th
e
r
etr
iev
a
l
p
er
f
o
r
m
an
ce
.
Fin
ally
,
th
e
o
r
ig
in
al
f
ea
tu
r
es
ar
e
co
m
p
ar
ed
with
th
e
b
in
a
r
ized
f
ea
tu
r
e
r
ep
r
esen
tatio
n
.
Dif
f
er
en
t
C
NNs
with
ap
p
licatio
n
to
C
B
I
R
task
s
ar
e
ex
am
i
n
ed
an
d
co
m
p
ar
ed
u
s
in
g
v
ar
i
ed
s
ettin
g
s
in
[
9
]
.
T
h
e
f
ea
tu
r
es r
ep
r
esen
ta
tio
n
o
f
th
e
im
a
g
es a
n
d
th
e
s
im
ilar
ity
m
ea
s
u
r
es b
etwe
en
im
ag
e
p
air
s
ar
e
lear
n
t to
p
r
o
ce
s
s
th
e
task
s
o
f
C
B
I
R
.
T
h
e
au
th
o
r
s
attem
p
ts
to
ap
p
r
o
v
e
if
C
NNs
ar
e
ef
f
ec
tiv
e
in
lear
n
in
g
th
e
f
ea
tu
r
es
o
f
im
ag
es
wh
en
ap
p
lied
to
C
B
I
R
task
s
.
A
d
ee
p
C
NN
m
o
d
el
is
p
r
o
p
o
s
ed
in
[
1
7
]
to
lear
n
th
e
f
ea
t
u
r
es
r
ep
r
esen
tatio
n
f
r
o
m
th
e
ac
tiv
atio
n
s
o
f
th
e
co
n
v
o
lu
ti
o
n
al
lay
er
s
.
T
h
e
au
th
o
r
s
s
u
g
g
est
th
r
ee
r
etr
ain
in
g
m
eth
o
d
s
in
o
r
d
er
t
o
im
p
r
o
v
e
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
r
etr
iev
al
p
r
o
ce
s
s
an
d
th
e
a
m
o
u
n
t
o
f
th
e
r
e
q
u
ir
e
d
m
em
o
r
y
.
T
h
ese
ar
e:
f
u
lly
u
n
s
u
p
er
v
is
ed
r
etr
ain
in
g
w
h
en
n
o
in
f
o
r
m
atio
n
is
av
ailab
le
b
u
t
o
n
ly
f
r
o
m
th
e
d
ataset
its
elf
,
r
etr
ain
in
g
with
r
elev
an
ce
in
f
o
r
m
atio
n
wh
e
n
th
e
lab
els
o
f
th
e
tr
ain
in
g
d
ata
a
r
e
ex
is
ts
,
an
d
r
elev
an
ce
f
ee
d
b
ac
k
-
b
ased
r
etr
ain
in
g
wh
en
th
er
e
a
r
e
f
ee
d
b
ac
k
s
f
r
o
m
u
s
er
s
.
3.
DE
E
P
CO
NVO
L
U
T
I
O
N
N
E
URAL N
E
T
WO
RK
S
Ov
er
th
e
p
ast
y
ea
r
s
th
er
e
h
a
v
e
b
ee
n
ex
te
n
s
iv
e
s
tu
d
ies
u
s
i
n
g
d
ee
p
lear
n
in
g
n
etwo
r
k
s
(
DL
Ns),
f
o
r
ex
am
p
le,
d
ee
p
b
elief
n
etwo
r
k
,
B
o
ltzm
an
n
m
ac
h
in
es,
r
estricte
d
B
o
ltzm
an
n
m
ac
h
in
es,
d
ee
p
B
o
ltzm
an
n
m
ac
h
in
e,
an
d
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
(
DNN)
[
9
]
.
I
n
th
is
s
tu
d
y
,
we
h
av
e
in
v
esti
g
ated
,
co
m
p
a
r
ed
,
a
n
d
ev
alu
ated
s
o
m
e
co
m
m
o
n
DL
Ns
an
d
th
eir
a
p
p
licatio
n
s
f
o
r
im
ag
e
class
if
icatio
n
an
d
au
to
m
atic
im
ag
e
r
etr
iev
al.
T
h
ese
ar
e:
Alex
Net,
VGG
-
1
6
an
d
VGG
-
1
9
n
etwo
r
k
s
,
Go
o
g
leNe
t,
R
es
Net.
W
e
also
h
av
e
co
m
p
a
r
ed
th
e
p
er
f
o
r
m
an
ce
o
f
th
ese
n
etwo
r
k
s
to
p
r
io
r
wo
r
k
s
in
th
is
d
o
m
ain
b
y
u
s
in
g
k
n
o
wn
ac
cu
r
ac
y
m
et
r
ics
an
d
an
al
y
ze
d
th
e
d
if
f
er
en
ce
s
b
etwe
en
th
e
ap
p
r
o
ac
h
es.
I
n
th
e
f
o
llo
win
g
s
u
b
s
ec
tio
n
s
,
we
w
ill ex
p
lain
th
ese
DL
Ns.
3
.
1
.
Alex
Net
Alex
Net
is
a
k
in
d
o
f
DL
Ns
in
tr
o
d
u
ce
d
b
y
Alex
Kr
iz
h
e
v
s
k
y
[
1
8
]
.
T
h
e
a
r
ch
itectu
r
e
o
f
Alex
Net
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
is
illu
s
tr
ated
in
Fig
u
r
e
2
.
As
s
h
o
wn
in
th
is
f
ig
u
r
e,
c
o
n
v
o
lu
tio
n
an
d
m
ax
p
o
o
lin
g
o
p
er
atio
n
s
ar
e
im
p
lem
en
ted
a
t
th
e
f
ir
s
t
co
n
v
o
lu
tio
n
al
lay
er
with
lo
ca
l
r
esp
o
n
s
e
n
o
r
m
aliza
tio
n
(
L
R
N)
.
T
h
e
co
n
v
o
l
u
tio
n
al
la
y
er
p
ar
am
eter
s
co
n
s
is
t
o
f
a
s
et
o
f
lear
n
ab
le
f
ilter
s
.
T
h
ese
f
ilter
s
ca
n
b
e
u
s
ed
to
ca
lcu
late
t
h
e
f
ea
tu
r
es
o
f
th
e
im
ag
es
in
class
ificatio
n
.
T
h
e
f
ilter
s
o
f
th
e
c
o
n
v
o
lu
tio
n
al
lay
er
s
a
r
e
u
p
d
ated
b
y
p
er
f
o
r
m
in
g
th
e
f
u
ll
co
n
v
o
l
u
tio
n
al
o
p
e
r
atio
n
o
n
th
e
f
ea
tu
r
e
m
ap
s
b
etwe
en
th
e
co
n
v
o
lu
tio
n
al
lay
er
an
d
its
im
m
ed
iate
p
r
ev
io
u
s
lay
er
.
I
n
th
is
lay
er
,
a
b
o
u
t
9
6
d
i
f
f
er
en
t
r
ec
e
p
tiv
e
f
ilter
s
ar
e
u
s
ed
wh
er
e
th
e
s
izes
o
f
th
ese
f
ilter
s
ar
e
1
1
*
1
1
.
Als
o
,
a
s
tr
id
e
s
ize
o
f
2
an
d
3
*
3
f
ilte
r
s
ar
e
u
s
ed
to
p
er
f
o
r
m
th
e
m
a
x
p
o
o
lin
g
o
p
er
atio
n
.
T
h
e
jo
b
o
f
p
o
o
lin
g
lay
e
r
is
to
r
ed
u
ce
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
wh
en
n
o
n
lin
ea
r
d
o
w
n
s
am
p
lin
g
is
p
er
f
o
r
m
ed
.
T
h
e
s
am
e
o
p
er
atio
n
s
ar
e
im
p
lem
en
ted
b
u
t
with
5
*
5
f
ilt
er
s
in
th
e
s
ec
o
n
d
lay
er
,
3
*
3
f
i
lter
s
with
3
8
4
,
3
8
4
an
d
2
9
6
f
e
a
tu
r
es
m
ap
s
in
th
e
th
ir
d
,
f
o
u
r
th
an
d
f
if
th
co
n
v
o
lu
t
io
n
al
lay
er
s
.
Mo
r
e
im
ag
e
d
etai
ls
an
d
lo
ca
l
f
ea
tu
r
e
im
ag
es
ar
e
ex
tr
ac
ted
s
in
ce
th
e
s
ize
o
f
co
n
v
o
lu
tio
n
al
lay
er
a
n
d
s
tr
id
e
is
s
m
all.
T
wo
la
y
er
s
,
wh
ich
ar
e
f
u
lly
c
o
n
n
ec
ted
(
FC
)
,
ar
e
u
s
ed
with
d
r
o
p
o
u
t.
I
n
Alex
N
et
n
etwo
r
k
,
th
e
p
r
o
b
lem
s
o
f
tr
ain
in
g
tim
e
co
n
s
u
m
in
g
an
d
o
v
er
-
f
itti
n
g
p
r
o
b
lem
s
ar
e
s
o
lv
ed
b
y
d
r
o
p
o
u
t
o
p
er
atio
n
.
Fin
ally
,
a
s
o
f
tm
ax
lay
er
is
u
s
ed
.
Alex
Net
h
as
b
ee
n
u
s
ed
in
a
wid
e
r
an
g
e
o
f
ap
p
licatio
n
s
s
u
ch
as o
b
ject
d
etec
tio
n
,
v
id
eo
class
if
icatio
n
an
d
im
ag
e
s
eg
m
en
tatio
n
[
6
,
1
2
,
1
9
-
2
2
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
A
co
mp
a
r
a
tive
a
n
a
lysi
s
o
f a
u
t
o
ma
tic
d
ee
p
n
eu
r
a
l n
etw
o
r
ks fo
r
ima
g
e
r
etri
ev
a
l
(
Ha
n
a
n
A
.
A
l
-
Ju
b
o
u
r
i
)
861
3
.
2
.
VG
G
-
E
Net
VGG
-
E
n
et
h
as
b
ee
n
p
r
o
p
o
s
ed
b
y
Simo
y
an
et
a
l
.
to
s
im
u
late
th
e
r
elatio
n
o
f
d
ep
t
h
o
f
th
e
n
et
wo
r
k
with
its
ca
p
ac
ity
,
VGG
-
E
n
et
m
ad
e
1
9
d
ee
p
lay
er
s
co
m
p
ar
in
g
with
Alex
Net.
Fig
u
r
e
3
s
h
o
ws th
e
ar
ch
itectu
r
e
o
f
th
e
VGG
n
et.
I
t
co
n
s
is
ts
o
f
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
wh
ich
is
u
s
ed
b
y
two
co
n
v
o
lu
tio
n
al
lay
er
s
.
R
eL
U
is
also
u
s
ed
b
y
a
s
in
g
le
m
ax
p
o
o
lin
g
lay
er
an
d
s
o
m
e
f
u
lly
co
n
n
ec
ted
lay
e
r
s
.
T
h
e
p
u
r
p
o
s
e
b
eh
in
d
p
u
ttin
g
m
ax
p
o
o
lin
g
af
ter
th
e
co
n
v
o
lu
tio
n
al
lay
e
r
is
t
o
t
u
n
e
t
h
e
n
et
wo
r
k
an
d
th
e
p
a
d
d
in
g
is
d
o
n
e
to
p
r
eser
v
e
th
e
s
p
atial
r
eso
lu
tio
n
.
T
h
e
last
lay
er
is
a
s
o
f
tm
ax
lay
er
wh
ich
is
u
s
ed
f
o
r
class
if
icatio
n
.
T
h
e
s
ize
o
f
th
e
co
n
v
o
lu
tio
n
f
il
ter
is
3
x
3
an
d
h
as a
s
tr
id
e
o
f
2
.
B
y
u
s
in
g
s
m
all
s
iz
e
o
f
f
ilter
s
,
it
p
r
o
v
id
es
lo
w
co
m
p
u
tatio
n
al
co
m
p
lex
ity
an
d
r
e
d
u
ce
s
th
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
.
T
h
er
e
a
r
e
d
if
f
er
e
n
t
k
in
d
s
o
f
VGG
-
E
m
o
d
els
wer
e
p
r
o
p
o
s
ed
.
T
h
ese
a
r
e:
VG
G
-
1
1
,
VGG
-
1
6
,
a
n
d
VGG
-
1
9
wh
er
e
th
ese
m
o
d
els
h
av
e
1
1
,
1
6
,
a
n
d
1
9
lay
er
s
r
esp
ec
tiv
ely
.
Alth
o
u
g
h
,
t
h
e
th
r
ee
m
o
d
els
o
f
VGG
-
E
h
av
e
th
r
ee
f
u
lly
co
n
n
ec
ted
at
th
e
en
d
,
VGG
-
1
1
co
n
tain
8
co
n
v
o
lu
tio
n
lay
er
s
,
VGG
-
1
6
h
as
1
3
co
n
v
o
l
u
tio
n
lay
er
s
an
d
VGG
-
1
9
c
o
n
tain
1
3
8
M
w
eig
h
ts
an
d
1
5
.
5
M
MA
C
S [
2
1
,
2
3
]
.
Fig
u
r
e
2
.
Alex
Net
L
a
y
o
u
t w
it
h
its
co
n
v
o
lu
tio
n
an
d
co
n
n
ec
te
d
lay
er
s
Fig
u
r
e
3
.
VGG
n
etwo
r
k
lay
o
u
t w
h
er
e
C
o
n
v
is
th
e
co
n
v
o
l
u
tio
n
lay
er
a
n
d
FC
is
f
u
ll c
o
n
n
ec
ted
3
.
3
.
G
o
o
g
leNe
t
Go
o
g
leNe
t
DL
N
is
p
r
o
p
o
s
ed
b
y
C
h
r
is
tian
Szeg
ed
y
et
a
l.
[
2
2
]
.
Go
o
g
leNe
t
n
etwo
r
k
h
as
b
ee
n
esp
ec
i
ally
d
esig
n
ed
to
r
e
d
u
ce
th
e
co
m
p
u
tatio
n
al
co
s
t
an
d
ac
h
iev
e
h
i
g
h
ac
cu
r
ac
y
co
m
p
ar
ed
with
t
r
ad
itio
n
al
C
NNs.
I
t
p
r
esen
ts
th
e
co
n
ce
p
t
o
f
in
ce
p
t
io
n
b
lo
c
k
.
I
t
h
elp
s
in
co
m
b
i
n
i
n
g
m
u
lti
s
ca
le
co
n
v
o
lu
tio
n
al
t
r
an
s
f
o
r
m
atio
n
s
b
y
ex
p
lo
itin
g
th
e
id
ea
o
f
s
p
lit
m
er
g
e
an
d
t
r
a
n
s
f
o
r
m
o
p
er
atio
n
s
.
T
h
u
s
,
d
if
f
e
r
en
t
ty
p
es
o
f
v
ar
i
atio
n
s
in
th
e
s
am
e
ca
teg
o
r
y
im
ag
es
with
d
iv
e
r
s
e
r
eso
lu
tio
n
s
ar
e
lear
n
t.
I
n
ce
p
ti
o
n
b
lo
ck
s
ar
e
u
s
ed
in
r
ep
laci
n
g
th
e
co
n
v
en
tio
n
al
lay
er
.
T
h
e
y
h
id
e
f
ilter
s
o
f
d
if
f
er
en
t sizes (
1
*
1
an
d
3
*
3
)
t
o
ca
p
tu
r
e
s
p
atial
i
n
f
o
r
m
atio
n
[
2
3
,
2
1
]
.
T
h
e
ar
ch
itectu
r
e
o
f
Go
o
g
leNe
t
is
i
llu
s
tr
ated
in
Fig
u
r
e
4
.
I
n
th
is
n
etwo
r
k
,
n
in
e
in
ce
p
tio
n
m
o
d
u
les
ar
e
u
s
ed
co
n
s
is
ts
o
f
2
2
lay
er
s
.
Al
th
o
u
g
h
,
Go
o
g
leNe
t
h
as
m
an
y
lay
er
s
co
m
p
a
r
ed
t
o
o
t
h
er
n
et
wo
r
k
s
b
ef
o
r
e
it,
th
e
n
u
m
b
er
o
f
th
e
p
ar
am
eter
s
is
m
u
ch
lo
wer
th
an
Alex
Net
an
d
VGG
n
etwo
r
k
s
.
I
t
h
as
7
M
p
ar
am
eter
s
wh
ile
Alex
Net
an
d
VGG
h
av
e
6
0
M
an
d
1
3
8
M
p
ar
am
eter
s
r
esp
ec
tiv
ely
.
Als
o
,
Go
o
g
leNe
t
n
etwo
r
k
h
as
f
o
u
r
m
ax
p
o
o
lin
g
lay
er
s
an
d
o
n
e
av
e
r
ag
e
p
o
o
lin
g
lay
e
r
i.e
.
o
n
ly
lay
er
s
with
p
ar
am
et
er
s
.
T
h
e
av
er
ag
e
p
o
o
lin
g
lay
e
r
h
as
a
f
ilter
with
a
s
ize
o
f
5
*
5
an
d
h
as
th
r
ee
s
tr
id
es
wh
ich
is
u
s
ed
b
ef
o
r
e
th
e
cla
s
s
if
ier
.
I
t
also
u
s
es
d
r
o
p
o
u
t
lay
er
wh
ich
h
as
a
r
atio
o
f
7
0
% f
r
o
m
d
r
o
p
p
ed
o
u
tp
u
ts
.
All c
o
n
v
o
lu
tio
n
al
lay
er
s
an
d
i
n
ce
p
tio
n
m
o
d
u
les u
s
e
R
eL
u
[
2
1
, 2
2
].
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
3
,
J
u
n
e
2
0
2
1
:
8
5
8
-
8
7
1
862
Fig
u
r
e
4
.
Go
o
g
leNe
t a
r
ch
itect
u
r
e
3
.
3
.
ResNet
Deep
r
esid
u
al
n
etwo
r
k
s
o
r
ca
ll
ed
R
esNet
i
s
p
r
o
p
o
s
ed
b
y
Kai
m
in
g
He
et
a
l
.
[
2
4
]
.
I
t
is
o
n
e
o
f
th
e
s
tates
o
f
ar
t
a
n
d
g
r
ea
test
C
NNs
u
s
ed
f
o
r
im
a
g
e
r
ec
o
g
n
itio
n
.
I
n
I
m
a
g
eNe
t
L
ar
g
e
Scale
Vis
u
al
R
ec
o
g
n
itio
n
C
h
allen
g
es
in
2
0
1
5
(
I
L
SVR
C
-
1
5
)
,
R
esNet
wo
n
th
at
ch
allen
g
e
with
a
to
p
5
er
r
o
r
o
f
3
.
5
7
%.
Fo
r
in
s
tan
ce
,
R
esNet
-
5
0
h
as
r
ea
ch
ed
a
n
av
e
r
ag
e
o
f
5
.
2
5
% o
f
to
p
-
5
e
r
r
o
r
wh
en
it
is
tr
ain
e
d
o
n
1
.
2
8
m
illi
o
n
tr
ain
in
g
im
a
g
es
in
1
0
0
0
class
es.
I
t
h
as
s
h
o
wn
a
h
ig
h
ac
cu
r
ac
y
i
n
co
m
p
u
te
r
v
is
io
n
.
Fig
u
r
e
5
s
h
o
ws
th
e
ar
ch
itectu
r
e
o
f
R
esNet
-
5
0
.
I
n
t
h
is
s
tu
d
y
,
R
esNet
-
5
0
h
as
b
ee
n
u
s
ed
f
o
r
im
ag
e
class
if
icatio
n
.
I
n
th
is
n
etwo
r
k
,
5
co
n
v
o
lu
tio
n
al
lay
er
s
ar
e
u
s
ed
an
d
th
e
in
p
u
t
im
ag
es
ar
e
o
f
s
ize
2
2
4
*
2
2
4
*
3
.
R
esNet
-
5
0
,
wh
ic
h
h
as
5
0
-
lay
er
C
NN
ar
ch
itectu
r
e
,
is
c
o
n
s
id
er
ed
to
b
e
th
e
f
ir
s
t
d
ee
p
C
NN
th
at
ap
p
lied
r
esid
u
al
lear
n
in
g
[
2
4
,
2
5
]
.
Fig
u
r
e
5
.
Re
sN
e
t
a
rc
h
it
e
c
tu
r
e
4.
P
RO
P
O
SE
D
M
E
T
H
O
D
I
n
th
is
wo
r
k
,
two
s
ce
n
ar
io
s
ar
e
f
o
llo
wed
:
im
ag
e
class
if
icatio
n
an
d
im
ag
e
r
etr
iev
al.
Fig
u
r
e
6
s
h
o
ws
th
e
s
tag
es
o
f
th
e
f
r
am
ewo
r
k
,
tr
ai
n
in
g
,
C
NN
m
o
d
el
tr
ain
in
g
,
im
ag
e
class
if
icatio
n
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
s
im
ilar
ity
m
ea
s
u
r
e,
an
d
im
a
g
e
r
etr
iev
al.
Fo
r
im
ag
e
class
if
icatio
n
,
C
NNs
ar
e
in
v
esti
g
ated
to
class
if
y
h
u
g
e
am
o
u
n
t
o
f
im
ag
es.
I
n
o
u
r
in
v
esti
g
atio
n
,
d
if
f
er
en
t
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
ar
e
u
s
ed
i
n
class
if
icatio
n
s
u
ch
im
ag
es.
T
h
e
C
NNs
ap
p
r
o
ac
h
es
ar
e
e
x
p
lo
it
ed
to
lea
r
n
f
ea
tu
r
es
o
f
im
a
g
es.
I
m
ag
e
class
if
icatio
n
is
ac
h
ie
v
ed
b
y
two
s
tag
es.
First,
a
s
et
o
f
tr
ain
in
g
im
a
g
es
th
at
ass
o
ciate
d
with
class
lab
e
l
ar
e
u
s
ed
t
o
tr
ain
a
class
if
ier
.
Seco
n
d
,
t
h
e
tr
ai
n
ed
class
if
ier
is
u
s
ed
to
p
r
ed
ict
th
e
class
lab
el
o
f
a
q
u
er
y
im
ag
e
b
ased
o
n
its
tr
ain
ed
k
n
o
wled
g
e
ab
o
u
t
th
e
class
.
Hen
ce
,
th
e
ac
cu
r
ac
y
o
f
th
e
class
if
ier
is
ev
alu
ated
ac
co
r
d
in
g
to
co
r
r
ec
t
p
r
e
d
ictio
n
.
I
m
ag
e
r
etr
iev
al
is
im
p
lem
en
ted
u
s
in
g
f
ea
tu
r
es
th
at
ar
e
lear
n
ed
b
y
th
e
C
NNs
ap
p
r
o
ac
h
es
an
d
th
e
n
r
esu
lts
ar
e
co
m
p
ar
ed
.
B
ased
o
n
o
u
tco
m
es a
n
d
an
aly
s
es a
n
ew
alg
o
r
ith
m
f
o
r
im
ag
e
r
etr
iev
al
i
s
d
ev
elo
p
ed
(
s
ee
s
u
b
s
ec
tio
n
4
.
2
.
3
)
.
4
.
1
.
Da
t
a
s
et
s
Dif
f
er
en
t
d
atasets
h
av
e
b
ee
n
u
s
ed
f
o
r
test
in
g
alg
o
r
ith
m
s
o
r
a
p
p
r
o
ac
h
es in
C
B
I
R
.
T
h
e
d
atasets
u
s
ed
in
th
is
p
ap
er
to
ev
alu
ate
th
e
p
e
r
f
o
r
m
a
n
ce
o
f
C
NNs
ar
e
d
ata
s
ets
with
a
h
ig
h
q
u
ality
wh
e
r
e
th
e
im
a
g
es
ar
e
non
-
lab
ele
d
an
d
c
o
m
p
r
ess
ed
.
Data
s
ets
co
r
el
1
K
[
2
6
]
,
co
r
el
5
0
K
[
2
6
]
a
n
d
C
altec
h
2
5
6
[
2
7
]
ar
e
u
s
ed
in
th
is
wo
r
k
to
v
alid
ate
th
e
p
r
o
p
o
s
ed
s
y
s
tem
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
A
co
mp
a
r
a
tive
a
n
a
lysi
s
o
f a
u
t
o
ma
tic
d
ee
p
n
eu
r
a
l n
etw
o
r
ks fo
r
ima
g
e
r
etri
ev
a
l
(
Ha
n
a
n
A
.
A
l
-
Ju
b
o
u
r
i
)
863
4
.
1
.
1
.
Co
re
l 1
K
C
o
r
el
1
K
d
ataset
[
2
6
]
co
n
s
is
ts
o
f
1
0
0
0
im
ag
es
with
1
0
0
f
o
r
ea
ch
class
.
T
h
e
s
ize
o
f
im
ag
es
is
(
2
5
6
x
3
8
4
)
o
r
(
3
8
4
x
2
5
6
)
ea
ch
im
a
g
e
m
ay
b
e
o
n
e
o
f
th
e
ten
class
lab
els
(
Af
r
ican
p
ea
p
o
le,
b
ea
c
h
,
b
u
id
in
g
s
,
b
u
s
es,
d
in
o
s
au
r
s
,
elep
h
an
ts
,
f
lo
wer
s
,
h
o
u
r
s
es,
m
o
u
n
tain
s
,
an
d
f
o
o
d
s
)
.
T
h
ese
lab
els
ar
e
an
n
o
tated
m
an
u
ally
u
s
in
g
an
E
x
ce
l
f
ile.
A
s
am
p
le
o
f
2
0
im
ag
es is
s
h
o
wn
in
Fig
u
r
e
7
with
th
eir
lab
el
s
.
4
.
1
.
2
.
Co
re
l 1
0
K
C
o
r
el
1
0
K
d
ataset
[
2
6
]
co
n
s
i
s
ts
o
f
1
0
0
0
0
im
ag
es
with
1
0
0
f
o
r
ea
c
h
class
.
T
h
e
s
ize
o
f
im
ag
es
is
(
1
2
6
x
1
8
7
)
o
r
(
1
8
7
x
1
2
6
)
.
W
e
s
lecte
d
5
0
class
es,
ar
t,
wem
an
,
d
o
g
,
clo
u
d
,
m
ac
h
r
o
o
m
,
ca
s
tle,
g
lass
,
b
ea
r
,
f
ig
h
tin
g
p
eo
p
le,
a
n
d
f
r
u
it
.
Fig
u
r
e
8
s
h
o
ws a
s
am
p
le
o
f
im
ag
es.
4
.
1
.
3
.
Ca
lt
ec
h
2
5
6
C
altec
h
2
5
6
d
ataset
[
2
7
]
co
n
s
i
s
ts
o
f
3
0
,
6
0
7
im
ag
es
o
f
o
b
ject
s
with
d
if
f
er
en
t
s
izes.
I
m
ag
es
ar
e
d
iv
id
ed
in
to
2
5
6
class
s
.
R
esear
ch
er
s
s
e
lect
s
o
m
e
class
e
s
to
ev
alu
ate
th
eir
ap
p
r
o
ac
h
es
o
r
alg
o
r
ith
m
s
.
I
n
o
u
r
ex
p
e
r
im
en
t,
we
ch
o
s
e
5
0
class
es with
1
0
0
f
o
r
ea
ch
class
.
Fig
u
r
e
9
s
h
o
ws
s
am
p
le
o
f
s
o
m
e
im
a
g
es.
Fig
u
r
e
6
.
Fra
m
ewo
r
k
o
f
im
ag
e
class
if
icatio
n
an
d
r
etr
iev
al
(
a)
(
b
)
(
c
)
(
d
)
(
e
)
(
f
)
(
g
)
(
h
)
(
i)
(
j)
Fig
u
r
e
7
.
Sam
p
le
o
f
c
o
r
el
1
K
i
m
ag
es
;
(
a)
Af
r
ican
Peo
p
le
1
,
(
b
)
B
ea
ch
2
,
(
c)
B
u
ild
in
g
s
3
,
(
d
)
B
u
s
es 4
,
(
e)
Din
o
s
au
r
s
5
,
(
f
)
E
lep
h
an
ts
6
,
(
g
)
Flo
wer
s
7
,
(
h
)
Ho
r
s
es 8
,
(
i)
Mo
u
n
tain
s
9
,
an
d
(
j)
Fo
o
d
s
1
0
Fig
u
r
e
8
.
Sam
p
le
o
f
co
r
el
1
0
K
im
ag
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
3
,
J
u
n
e
2
0
2
1
:
8
5
8
-
8
7
1
864
Fig
u
r
e
9
.
Sam
p
le
o
f
C
altch
2
5
6
im
ag
es
4
.
2
.
E
x
perim
ent
a
l r
esu
lt
s
a
nd
a
na
ly
s
is
I
n
th
is
s
ec
tio
n
,
we
p
r
esen
t
th
e
r
esu
lts
o
f
th
e
ex
p
er
im
en
ts
co
n
d
u
cted
to
ev
alu
ate
th
e
ac
cu
r
ac
y
o
f
I
R
an
d
co
m
p
u
tatio
n
al
ef
ficien
cy
b
ase
d
o
n
p
r
p
o
s
ed
C
NNs
in
ter
m
s
o
f
im
a
g
e
class
i
f
icatio
n
a
n
d
im
ag
e
r
er
ie
v
al.
I
m
ag
e
class
if
icatio
n
is
ac
h
iev
ed
b
y
t
wo
s
tag
es.
First,
a
s
et
o
f
tr
ain
in
g
im
ag
es th
at
ass
o
ciate
d
with
class
lab
el
ar
e
u
s
ed
to
tr
ain
a
class
if
ier
.
Seco
n
d
,
th
e
tr
ain
ed
class
if
ier
is
u
s
ed
to
p
r
ed
ict
th
e
class
lab
el
o
f
a
q
u
er
y
im
ag
e
b
ased
o
n
its
tr
ain
ed
k
n
o
wled
g
e
a
b
o
u
t
t
h
e
class
.
Hen
ce
,
th
e
ac
cu
r
ac
y
o
f
th
e
class
if
ier
is
ev
alu
ated
ac
co
r
d
in
g
to
co
r
r
ec
t
p
r
ed
ictio
n
.
I
R
r
etu
r
n
s
to
p
T
im
ag
es
as
a
r
an
k
ed
lis
t
f
r
o
m
d
ata
b
ase
im
ag
es
th
at
ar
e
m
o
s
t
s
im
ilar
to
a
q
u
er
y
im
ag
e
b
y
u
s
in
g
a
s
im
ilar
ity
m
ea
s
u
r
e
with
o
u
t
u
s
in
g
class
lab
els.
T
h
e
ac
cu
r
ac
y
is
ev
alu
ated
ac
c
o
r
d
in
g
to
h
o
w
m
a
n
y
co
r
r
ec
t
im
ag
es
o
u
t
o
f
th
e
T
im
ag
es
in
th
e
r
an
k
ed
lis
t.
All
ex
p
er
im
en
ts
ar
e
p
er
f
o
r
m
ed
u
s
in
g
MA
T
L
AB
2
0
1
8
a,
o
n
a
co
m
p
u
ter
with
a
p
r
o
ce
s
s
o
r
I
n
tel
co
r
e
i7
C
PU 2
.
5
GHz
2
.
6
GHz
an
d
8
GB
R
AM
.
4
.
2
.
1
.
E
v
a
lua
t
io
n o
f
t
he
perf
o
rm
a
nce
I
n
im
ag
e
class
if
icatio
n
,
a
co
n
f
u
s
io
n
m
atr
ix
is
u
s
u
ally
u
s
ed
to
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
a
class
if
ier
.
T
ab
le
1
s
h
o
ws
a
co
n
f
u
s
io
n
m
atr
ix
f
o
r
two
class
es
an
d
it
ca
n
b
e
ex
ten
d
e
d
in
to
m
class
es
(
i.e
.
m
x
m
)
.
T
r
u
e
p
o
s
itiv
e
(
T
P),
tr
u
e
n
eg
ativ
e
(
T
N)
,
f
alse
n
eg
ativ
e
(
FN)
,
an
d
f
alse
p
o
s
i
tiv
e
(
FP
)
ar
e
th
e
ter
m
s
g
iv
en
to
a
n
im
ag
e
class
if
icatio
n
test
[
2
8
]
.
Pre
cisi
o
n
o
r
ac
cu
r
ac
y
is
ca
lcu
lated
as f
o
llo
ws:
=
(
,
)
∑
(
,
)
=
1
⁄
(
1
)
=
∑
=
1
⁄
(
2
)
wh
er
e,
AC
is
th
e
p
r
ec
is
io
n
o
r
ac
cu
r
ac
y
.
T
ab
le
1
.
C
o
n
f
u
s
io
n
M
atr
ix
P
r
e
d
i
c
t
e
d
c
l
a
ss
A
c
t
u
a
l
c
l
a
ss
C
1
C
2
C
1
TP
FN
C
2
FP
TN
I
n
im
ag
e
r
etr
ie
v
al,
a
m
ea
n
a
v
e
r
ag
e
p
r
ec
is
io
n
(
MA
P)
is
u
s
ed
f
o
r
ev
alu
atio
n
b
ased
o
n
p
r
ec
is
i
o
n
(
P)
an
d
av
er
ag
e
p
r
ec
is
io
n
(
AP)
[
2
8
]
.
=
(
3
)
wh
er
e,
is
th
e
p
r
ec
is
io
n
o
f
im
a
g
e
r
etr
iev
al,
is
n
u
m
b
er
o
f
r
ele
v
an
t
r
etr
ie
v
ed
im
a
g
es
an
d
is
to
tal
n
u
m
b
e
r
o
f
r
etr
iev
e
d
im
ag
es
,
=
∑
=
1
(
4
)
wh
er
e,
is
av
er
ag
e
p
r
ec
is
io
n
o
f
im
ag
e
r
etr
iev
al,
is
p
r
ec
is
io
n
o
f
im
ag
e
i
n
th
e
class
,
an
d
is
to
tal
n
u
m
b
er
o
f
im
ag
es in
th
e
class
.
=
∑
=
1
(
5
)
wh
er
e,
is
m
ea
n
av
er
ag
e
p
r
ec
is
io
n
o
f
im
a
g
e
r
etr
iev
al
,
is
av
er
ag
e
p
r
ec
is
io
n
o
f
class
im
ag
e,
an
d
is
to
tal
n
u
m
b
er
o
f
class
es in
th
e
d
atab
ase.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
A
co
mp
a
r
a
tive
a
n
a
lysi
s
o
f a
u
t
o
ma
tic
d
ee
p
n
eu
r
a
l n
etw
o
r
ks fo
r
ima
g
e
r
etri
ev
a
l
(
Ha
n
a
n
A
.
A
l
-
Ju
b
o
u
r
i
)
865
4
.
2
.
2
.
I
m
a
g
e
cla
s
s
if
ica
t
io
n
Ma
n
y
ex
p
e
r
im
en
ts
ar
e
co
n
d
u
c
ted
o
n
th
e
im
ag
e
d
atasets
.
T
h
e
tr
ain
in
g
m
o
d
els
o
f
t
h
e
n
etwo
r
k
s
ar
e
s
et
u
p
as
f
o
llo
ws:
th
e
d
atasets
ar
e
d
iv
id
ed
in
t
o
7
0
%
f
o
r
tr
ain
in
g
,
1
5
%
f
o
r
v
alid
atio
n
a
n
d
1
5
%
f
o
r
test
in
g
d
ata.
I
n
ad
d
itio
n
,
t
h
e
tr
ain
in
g
p
ar
am
et
er
s
f
o
r
th
e
C
NNs
ar
e
s
et
as
f
o
llo
ws:
th
e
lear
n
in
g
r
ate
is
s
et
to
0
.
0
0
0
0
1
;
t
h
e
m
ax
im
u
m
e
p
o
ch
n
u
m
b
er
is
4
3
5
.
Als
o
,
th
e
weig
h
t
o
f
th
e
lea
r
n
in
g
r
ate
f
ac
to
r
an
d
b
ias
lear
n
in
g
r
ate
f
ac
to
r
ar
e
s
et
to
2
0
f
o
r
th
e
lay
er
o
f
f
u
lly
co
n
n
ec
ted
.
T
h
e
m
o
s
t
co
m
m
o
n
C
NNs
u
s
e
d
in
th
e
p
ap
e
r
as
m
en
tio
n
ed
i
n
th
e
p
r
ev
io
u
s
s
ec
tio
n
ar
e:
s
i
m
p
le
C
NN,
Alex
Net,
Go
o
g
leNe
t,
R
esNet
-
5
0
,
Vg
g
-
1
6
an
d
Vg
g
-
1
9
.
T
h
ese
m
o
d
els
ar
e
co
m
p
a
r
ed
wit
h
th
e
co
n
v
en
tio
n
al
m
eth
o
d
s
u
s
ed
f
o
r
I
R
s
u
ch
as
t
h
e
h
u
e
s
atu
r
atio
n
v
alu
e
(
HSV)
co
lo
u
r
f
ea
tu
r
e,
g
r
ay
lev
el
c
o
-
o
cc
u
r
r
e
n
ce
m
atr
ix
(
GL
C
M)
f
ea
tu
r
es
an
d
s
ca
le
in
v
ar
ian
t
f
ea
tu
r
e
tr
an
s
f
o
r
m
(
SIFT
)
[
6
]
.
T
h
e
ac
cu
r
ac
y
o
f
th
e
r
e
s
u
lts
o
f
th
e
test
in
g
an
d
v
alid
atio
n
d
ata
s
ets
is
u
s
e
d
o
n
im
ag
e
d
ata
to
ev
al
u
ate
th
e
p
er
f
o
r
m
a
n
c
e
o
f
th
ese
m
eth
o
d
s
.
T
h
e
r
esu
lts
o
f
th
e
co
n
v
en
tio
n
al
m
eth
o
d
s
an
d
C
NNs m
o
d
els as
f
ea
tu
r
e
ex
tr
ac
to
r
s
b
ased
o
n
co
r
el
1
K
d
ataset
ar
e
s
h
o
wn
in
T
ab
le
2
with
d
ata
a
u
g
m
en
tatio
n
.
As ca
n
b
e
s
ee
n
f
r
o
m
th
is
tab
le,
th
e
b
est ac
cu
r
ac
y
ar
e
9
9
%,
9
7
% a
n
d
9
5
% a
ch
iev
e
d
b
y
C
NNs
m
o
d
els
wh
en
th
e
tr
ain
i
n
g
,
v
ali
d
atin
g
a
n
d
test
in
g
d
ata
ar
e
au
g
m
en
ted
co
m
p
a
r
ed
wit
h
th
e
c
o
n
v
e
n
tio
n
al
ap
p
r
o
ac
h
es.
T
ab
le
2
.
Acc
u
r
ac
y
o
f
C
NNs m
o
d
els as f
ea
tu
r
e
ex
tr
ac
to
r
s
b
ased
o
n
c
o
r
el
1
K
d
ata
au
g
m
e
n
tatio
n
d
ataset
M
e
t
h
o
d
A
c
c
u
r
a
c
y
(
%)
Tr
a
i
n
i
n
g
T
i
me
H
S
V
G
LC
M
S
I
F
T
C
N
N
[
5
]
A
l
e
x
N
e
t
G
o
o
g
l
e
N
e
t
R
e
sN
e
t
-
50
V
g
g
-
16
V
g
g
-
19
0
.
4
3
0
.
3
9
0
.
5
3
0
.
7
9
0
.
9
9
0
.
9
7
0
.
9
5
0
.
9
8
0
.
9
8
N
o
t
A
v
a
i
l
a
b
l
e
N
o
t
A
v
a
i
l
a
b
l
e
N
o
t
A
v
a
i
l
a
b
l
e
N
o
t
A
v
a
i
l
a
b
l
e
1
0
mi
n
.
7
m
i
n
.
8
9
mi
n
.
4
3
mi
n
.
1
3
1
m
i
n
.
Fo
r
th
e
co
r
el
1
K
d
atasets
,
th
e
m
o
d
els
b
ased
o
n
th
e
C
NNs
m
o
d
els
d
id
co
n
v
er
g
e
t
o
ex
ce
llen
t
ac
cu
r
ac
y
an
d
d
em
o
n
s
tr
ate
h
ig
h
p
er
f
o
r
m
an
ce
in
tr
ain
in
g
s
tag
e
with
th
e
least
n
u
m
b
er
o
f
ep
o
ch
s
.
Alth
o
u
g
h
,
th
er
e
ar
e
n
o
s
ig
n
if
ican
t
d
if
f
er
e
n
ce
s
in
th
e
co
n
v
er
g
en
ce
s
o
f
th
e
m
o
d
els
(
m
o
r
e
th
an
9
5
%),
th
ey
to
o
k
m
o
r
e
tr
ain
in
g
tim
e
f
o
r
co
n
v
er
g
en
ce
as
th
e
co
m
p
le
x
ity
o
f
th
e
C
NNs
ar
e
in
cr
ea
s
ed
.
O
n
th
e
o
t
h
er
h
a
n
d
,
th
e
co
n
v
er
g
e
n
ce
ac
cu
r
ac
y
r
esu
lts
f
o
r
th
e
s
am
e
d
atasets
with
o
u
t
d
ata
au
g
m
en
tatio
n
h
av
e
n
o
t
g
iv
en
g
o
o
d
ac
cu
r
ac
y
.
Fo
r
e
x
am
p
le,
th
e
test
in
g
ac
cu
r
ac
y
is
1
0
%,
4
6
%
an
d
6
8
%
f
o
r
t
h
e
s
im
p
le
C
NN,
Alex
N
et
an
d
Go
o
g
leNe
t r
esp
ec
tiv
ely
.
I
t
is
s
h
o
w
n
th
at
to
im
p
r
o
v
e
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
C
NNs,
d
ata
au
g
m
en
tatio
n
ca
n
s
u
cc
ess
f
u
lly
b
e
u
s
ed
.
A
s
a
m
p
l
e
o
f
3
0
c
l
as
s
p
r
o
b
a
b
i
li
t
ie
s
r
e
s
u
l
t
s
f
o
r
b
o
t
h
A
le
x
N
e
t
a
n
d
G
o
o
g
l
e
N
e
t
c
o
n
v
o
l
u
t
i
o
n
al
n
e
u
r
a
l
n
e
t
w
o
r
k
a
s
f
e
a
t
u
r
e
e
x
t
r
a
c
t
o
r
s
w
i
t
h
a
u
g
m
e
n
t
a
t
i
o
n
i
s
s
h
o
w
n
i
n
F
i
g
u
r
e
1
0
.
F
r
o
m
t
h
e
r
e
s
u
l
t
s
,
it
i
s
o
b
s
e
r
v
e
d
t
h
a
t
m
o
s
t
c
l
as
s
es
h
a
v
e
h
i
g
h
a
c
c
u
r
a
c
y
,
t
h
e
c
l
as
s
i
f
i
c
a
t
i
o
n
is
a
l
m
o
s
t
s
u
c
c
es
s
f
u
l
.
Al
s
o
,
i
t
i
s
s
h
o
w
n
t
h
a
t
t
h
e
C
N
Ns
m
o
d
e
l
s
r
e
s
u
l
ts
a
r
e
s
u
p
e
r
i
o
r
t
o
t
h
e
k
n
o
w
n
t
h
r
e
e
m
e
t
h
o
d
s
.
O
n
t
h
e
o
t
h
e
r
h
a
n
d
,
t
h
e
r
e
s
u
l
ts
o
f
t
h
e
C
N
N
s
m
o
d
e
ls
as
f
e
a
t
u
r
e
e
x
t
r
a
ct
o
r
s
b
a
s
e
d
o
n
c
o
r
e
l
5
0
K
a
n
d
C
a
lt
e
ch
2
5
6
d
a
t
a
s
e
t
s
a
r
e
s
h
o
w
n
i
n
T
ab
l
e
3
a
n
d
T
a
b
l
e
4
w
i
t
h
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
.
F
r
o
m
t
h
e
e
x
p
e
r
i
m
e
n
t
s
,
i
t
i
s
a
p
p
a
r
e
n
t
t
h
a
t
t
h
e
c
o
r
e
l
1
K
i
m
a
g
e
s
d
a
t
a
a
r
e
c
l
a
s
s
i
f
ie
d
c
o
r
r
e
c
tl
y
u
s
i
n
g
C
N
Ns
m
o
d
e
ls
wi
t
h
h
i
g
h
a
c
c
u
r
a
c
y
w
h
i
l
e
t
h
e
C
a
lt
e
c
h
2
5
6
d
a
t
a
w
i
t
h
5
0
c
l
a
s
s
es
h
a
s
l
o
w
a
c
c
u
r
a
c
y
.
I
t
i
s
c
o
n
c
l
u
d
e
d
t
h
a
t
t
h
e
f
e
a
t
u
r
e
s
o
f
t
h
e
i
m
a
g
e
s
a
r
e
l
e
a
r
n
t
f
r
o
m
t
h
e
p
r
e
-
t
r
a
i
n
e
d
m
o
d
e
l
s
a
n
d
it
d
o
e
s
n
o
t
n
e
e
d
t
o
s
e
a
r
c
h
f
e
at
u
r
e
s
m
a
n
u
a
ll
y
.
Fig
u
r
e
1
0
.
A
s
am
p
le
o
f
class
p
r
o
b
ab
ilit
ies o
f
Alex
Net
an
d
Go
o
g
leNe
t f
ea
tu
r
e
ex
tr
ac
tio
n
o
f
co
r
el
1
K
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
3
,
J
u
n
e
2
0
2
1
:
8
5
8
-
8
7
1
866
T
ab
le
3
.
Acc
u
r
ac
y
o
f
C
NNs m
o
d
els as f
ea
tu
r
e
ex
tr
ac
to
r
s
b
ased
o
n
c
o
r
el
5
0
K
d
ata
au
g
m
en
t
atio
n
d
ataset
M
e
t
h
o
d
A
c
c
u
r
a
c
y
(
%)
Tr
a
i
n
i
n
g
T
i
me
A
l
e
x
N
e
t
G
o
o
g
l
e
N
e
t
R
e
sN
e
t
-
50
V
g
g
-
16
V
g
g
-
19
0
.
9
6
0
.
9
7
0
.
9
3
0
.
9
7
0
.
9
8
1
9
8
m
i
n
.
3
2
4
m
i
n
.
3
0
6
m
i
n
.
5
6
4
m
i
n
.
5
9
1
m
i
n
.
T
ab
le
4
.
Acc
u
r
ac
y
o
f
C
NNs m
o
d
els as f
ea
tu
r
e
ex
tr
ac
to
r
s
b
ased
o
n
C
altec
h
2
5
6
d
ata
a
u
g
m
e
n
tatio
n
d
ataset
M
e
t
h
o
d
A
c
c
u
r
a
c
y
(
%)
Tr
a
i
n
i
n
g
T
i
me
A
l
e
x
N
e
t
G
o
o
g
l
e
N
e
t
R
e
sN
e
t
-
50
V
g
g
-
16
V
g
g
-
19
0
.
4
3
0
.
4
4
0
.
4
5
0
.
4
2
0
.
4
4
4
8
mi
n
.
4
8
mi
n
.
8
8
9
m
i
n
.
1
7
5
9
m
i
n
.
5
5
5
m
i
n
.
4
.
2
.
3
.
I
m
a
g
e
re
t
riev
a
l
As
m
en
tio
n
ed
ea
r
lier
,
f
ea
tu
r
e
r
ep
r
esen
tatio
n
is
o
n
e
ch
allen
g
e
o
f
s
em
in
tac
g
ap
in
C
B
I
R
.
R
ec
en
tly
,
C
NN
h
as
b
ee
n
u
s
ed
t
o
lear
n
f
ea
tu
r
es
to
b
e
m
o
r
e
ac
cu
r
ate.
Hen
ce
,
o
u
r
aim
in
t
h
ese
ex
p
er
im
e
n
ts
is
u
s
in
g
ab
o
v
e
C
NNs
ap
p
r
o
ac
h
es
to
lear
n
f
ea
t
u
r
es
an
d
h
an
d
le
im
ag
e
r
et
r
iev
al
with
o
u
t
u
s
in
g
class
lab
els
b
y
u
s
i
n
g
th
em
.
R
esu
lted
f
ea
tu
r
es
f
r
o
m
f
iv
e
C
NNs
(
Alex
Net,
Go
o
g
leNe
t,
R
esNet
-
5
0
,
Vg
g
-
1
6
,
a
n
d
Vg
g
-
1
9
)
ar
e
s
ep
er
ately
test
ed
ac
co
r
d
in
g
to
th
e
f
r
am
ewo
r
k
in
Fig
u
r
e
6
.
Firstl
y
,
ex
p
er
im
en
ts
o
f
i
m
ag
e
r
e
tr
iev
al
ar
e
co
n
d
u
cte
d
o
n
co
r
el
1
K
s
tan
d
ar
d
d
atab
ase
to
ju
d
g
e
wh
ic
h
d
ee
p
lear
n
i
n
g
ap
p
r
o
ac
h
ca
n
p
r
o
d
u
ce
e
f
f
ec
tiv
e
f
ea
t
u
r
e
th
a
n
o
th
er
s
.
L
ea
v
e
-
o
n
e
-
o
u
t
m
an
n
er
i
s
u
s
ed
to
ca
lcu
late
Pre
cisi
o
n
s
(
P
)
f
o
r
im
ag
es
an
d
th
en
MA
Ps
ar
e
co
m
p
u
ted
.
C
ity
-
b
lo
ck
(
L
1
)
d
is
tan
ce
f
u
n
ctio
n
is
u
s
ed
to
co
m
p
u
te
th
e
s
im
ilar
ity
b
etwe
en
a
q
u
er
y
im
ag
e
v
ec
to
r
(
f
ea
t
u
r
e)
an
d
d
a
tab
ase
im
ag
es
v
ec
to
r
s
.
R
esu
lte
d
s
im
ilar
ity
v
alu
es
ar
e
r
an
k
e
d
in
ascen
d
in
g
o
r
d
e
r
.
T
o
p
(
5
-
1
0
0
)
r
etr
iev
e
d
im
a
g
es
in
ter
m
s
o
f
MA
Ps
f
o
r
C
NN
ap
p
r
o
ac
h
es
a
r
e
ca
lcu
lated
an
d
illu
s
tr
ated
in
Fi
g
u
r
e
1
1
.
I
t
is
clea
r
th
at
th
e
p
er
f
o
r
m
a
n
ce
o
f
u
s
in
g
f
ea
tu
r
e
(
1
0
D)
t
h
at
is
p
r
o
d
u
ce
d
f
r
o
m
Go
o
g
leNe
t
with
2
2
-
lay
e
r
s
is
m
o
r
e
ef
f
ec
tiv
e
an
d
r
o
b
u
s
t
th
an
o
th
er
s
as
lo
n
g
as
T
o
p
(
5
-
1
0
0
)
r
et
r
iev
ed
im
ag
es.
Me
an
wh
ile,
Alex
Net
with
20
-
lay
er
s
e
x
tr
ac
ted
f
ea
tu
r
e
(
4
0
9
6
D)
th
at
h
as
l
o
west
ac
h
iev
em
en
t.
Vg
g
-
1
6
an
d
-
1
9
p
r
o
d
u
ce
d
f
ea
tu
r
es
wh
ich
ar
e
th
e
s
am
e
as
th
at
o
f
Alex
Net
in
len
g
th
b
u
t
th
ey
p
er
f
o
r
m
e
d
h
ig
h
er
.
R
esNet
-
5
0
ex
tr
ac
ted
a
s
m
aller
d
im
en
s
io
n
o
f
f
ea
tu
r
e
wh
ich
is
2
0
4
8
co
m
p
ar
ed
to
th
e
Alex
Net
,
Vg
g
-
1
6
an
d
-
1
9
ap
p
r
o
ch
es
b
u
t
th
e
f
ea
tu
r
es
ar
e
m
o
r
e
r
o
b
u
s
t
esp
ec
ially
at
T
o
p
3
0
-
1
0
0
r
an
k
e
d
lis
t
o
f
im
ag
es.
T
h
er
ef
o
r
e,
it
is
in
ter
est
to
an
aly
s
in
d
iv
id
u
al
cl
ass
im
ag
es
b
etwe
en
Go
o
g
len
et
an
d
R
esen
et
-
5
0
at
T
o
p
1
0
0
r
etr
iev
e
d
im
ag
es.
Hen
ce
,
APs
ar
e
clar
if
ied
in
Fig
u
r
e
1
2
.
At
th
e
f
i
r
s
t
v
iew,
th
er
e
is
a
b
ig
d
if
f
e
r
en
ce
b
etwe
en
two
ap
p
r
o
ac
h
es
wh
er
e
th
e
p
er
f
o
r
m
an
ce
o
f
Go
o
g
leNe
t
is
h
ig
h
e
r
th
a
n
R
esNet
-
5
0
o
v
er
all
class
es
ex
ce
p
t
f
o
r
th
e
b
u
s
class
,
th
e
r
ate
is
e
q
u
al.
I
n
o
r
d
er
t
o
ju
d
g
e
h
o
w
th
e
d
if
f
er
en
ce
is
s
ig
n
if
ica
n
t,
a
t
-
test
s
tatis
tica
l m
eth
o
d
is
u
s
ed
th
at
ca
n
b
e
ca
lcu
lated
as
[
2
9
]
:
=
̅
1
−
̅
2
√
(
(
1
−
1
)
1
2
+
(
2
−
1
)
2
2
1
+
2
−
2
)
(
1
1
+
1
2
)
(
6
)
wh
er
e
̅
1
an
d
̅
2
ar
e
th
e
s
am
p
le
p
r
e
cisi
o
n
r
ates
(
)
,
S
1
an
d
S
2
ar
e
s
tan
d
ar
d
s
d
ev
iatio
n
s
,
an
d
1
an
d
2
ar
e
th
e
s
am
p
le
s
izes.
T
wo
h
y
p
o
th
eses
ar
e
r
eg
ar
d
ed
an
d
d
eter
m
in
e
d
b
ased
o
n
t
-
test
,
th
e
n
u
ll
h
y
p
o
th
esis
(
H
0
)
wh
er
e
̅
1
−
̅
2
=
0
an
d
alter
n
ativ
e
h
y
p
o
th
esis
(
H
A
)
wh
er
e
̅
1
−
̅
2
≠
0
.
P
-
v
alu
e
o
f
th
e
t
est
is
th
e
p
r
o
b
ab
ilit
y
o
f
o
b
s
er
v
in
g
a
test
.
Sm
all
v
alu
es o
f
p
r
e
f
er
s
to
th
at
th
e
n
u
ll h
y
p
o
th
esis
is
r
ejec
ted
at
s
ig
n
if
ican
ce
lev
el
0
.
0
5
.
F
o
r
e
a
c
h
c
l
as
s
i
n
t
h
e
c
o
r
e
l
1
K
d
a
t
a
b
a
s
e
,
t
h
e
t
e
s
t
w
as
c
o
m
p
u
t
ed
.
T
h
i
s
m
e
a
n
s
t
h
e
s
i
ze
o
f
e
a
c
h
s
a
m
p
l
e
is
1
0
0
e
l
e
m
e
n
t
s
(
i
.
e
.
p
r
e
c
i
s
i
o
n
v
al
u
e
s
)
.
H
e
n
c
e
,
t
h
e
f
i
r
s
t
s
a
m
p
l
e
(
S
1
)
a
n
d
s
e
c
o
n
d
s
a
m
p
l
e
(S
2
)
h
a
v
e
p
r
e
c
i
s
i
o
n
r
a
t
e
s
o
f
T
o
p
1
0
0
r
e
t
r
i
e
v
e
d
i
m
a
g
e
s
f
r
o
m
u
s
i
n
g
G
o
o
g
l
e
n
e
t
f
e
a
t
u
r
e
a
n
d
R
es
N
e
t
-
5
0
f
e
a
t
u
r
e
r
e
s
p
e
ct
i
v
e
l
y
.
T
h
e
t
-
t
e
s
t
p
r
o
v
e
d
t
h
a
t
a
l
l
d
i
f
f
r
e
n
c
e
s
b
et
w
e
e
n
p
r
e
c
es
io
n
v
a
l
u
e
s
a
r
e
s
i
g
n
i
f
i
g
a
n
t
e
v
e
n
f
o
r
B
u
s
e
s
c
l
as
s
.
F
i
g
u
r
e
1
3
s
h
o
w
s
t
h
e
tw
o
s
a
m
p
l
e
s
w
h
e
r
e
t
h
e
m
o
s
t
v
al
u
e
s
o
f
S
1
9
9
%
c
o
m
p
a
r
e
d
t
o
S
2
.
W
e
c
a
n
c
o
n
c
l
u
d
e
t
h
a
t
G
o
o
g
l
e
N
et
l
ea
r
e
n
ed
a
f
e
a
t
u
r
e
w
i
t
h
l
o
w
d
i
m
e
n
s
i
o
n
(
1
0
)
m
e
a
n
s
l
e
s
s
c
o
m
p
u
t
a
t
i
o
n
a
n
d
h
i
g
h
a
c
c
u
r
a
c
y
d
u
e
t
o
t
h
e
i
n
c
e
p
ti
o
n
b
l
o
c
k
t
h
at
e
x
p
l
o
i
t
s
s
p
li
t
,
m
e
r
g
e
a
n
d
t
r
a
n
s
f
o
r
m
o
p
e
r
a
t
i
o
n
s
t
o
co
m
b
i
n
e
m
u
l
t
i
s
c
a
l
e
c
o
n
v
o
l
u
ti
o
n
a
l
t
r
a
n
s
f
o
r
m
a
ti
o
n
s
.
T
h
e
r
e
f
o
r
e,
d
i
f
f
e
r
e
n
t
t
y
p
e
s
o
f
v
a
r
i
a
t
i
o
n
s
i
n
t
h
e
s
a
m
e
ca
t
e
g
o
r
y
i
m
a
g
e
s
w
it
h
d
i
v
e
r
s
e
r
es
o
l
u
ti
o
n
s
a
r
e
l
e
a
r
n
t
.
I
n
o
t
h
e
r
w
o
r
d
s
,
G
o
o
g
l
e
n
e
t
h
a
s
a
b
i
li
ty
t
o
e
x
t
r
a
c
t
m
o
r
e
d
is
c
r
i
m
i
n
at
i
v
e
in
f
o
r
m
a
t
i
o
n
a
b
o
u
t
i
n
t
e
r
e
s
t
e
d
o
b
je
c
t
s
t
h
a
n
R
es
n
e
t
-
5
0
a
t
l
a
y
e
r
2
2
.
W
e
co
n
d
u
cted
o
th
er
r
etr
iev
al
ex
p
er
im
en
ts
to
in
v
esti
g
ate
th
e
s
ec
o
n
d
is
s
u
e
in
C
B
I
R
wh
ich
is
s
im
ilar
ity
m
ea
s
u
r
es.
I
n
th
e
liter
atu
r
e,
d
if
f
er
en
t
m
ea
s
u
r
es
h
av
e
b
ee
n
u
s
ed
to
co
m
p
u
te
th
e
s
im
ilar
ity
b
etwe
en
a
q
u
er
y
im
ag
e
an
d
d
atab
ase
im
ag
e
s
d
ep
e
n
d
i
n
g
o
n
im
ag
e
d
escr
ip
to
r
.
Fo
r
i
n
s
tan
ce
,
th
e
d
escr
ip
to
r
is
r
e
p
r
esen
ted
as
a
s
in
g
le
v
ec
to
r
o
r
a
s
et
o
f
v
ec
to
r
,
in
lin
ea
r
s
p
ac
e
o
r
n
o
n
-
lin
ea
r
m
an
if
o
ld
.
[
2
9
,
3
0
]
.
He
n
ce
,
c
o
r
r
elatio
n
(
D
1
)
,
co
s
in
e
(
D
2
)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
A
co
mp
a
r
a
tive
a
n
a
lysi
s
o
f a
u
t
o
ma
tic
d
ee
p
n
eu
r
a
l n
etw
o
r
ks fo
r
ima
g
e
r
etri
ev
a
l
(
Ha
n
a
n
A
.
A
l
-
Ju
b
o
u
r
i
)
867
an
d
E
u
clid
ea
n
(D
3
)
wer
e
ap
p
l
ied
r
ath
er
th
an
city
-
b
l
o
ck
(D
4
)
in
o
u
r
s
y
s
tem
s
ep
ar
ately
.
Su
p
p
o
s
e
an
d
r
ef
er
e
to
q
u
er
y
im
ag
e
an
d
d
at
ab
ase
im
ag
e
f
ea
tu
r
e
v
ec
to
r
s
r
e
s
p
ec
tiv
ely
with
th
e
d
im
en
s
io
n
,
th
en
D
1
,
D
2
,
D
3
,
an
d
D
4
ar
e
d
ef
in
ed
as f
o
llo
ws [
3
1
]
.
1
=
1
−
(
−
̅
)
(
−
̅
)
√
(
−
̅
)
(
−
̅
)
√
(
−
̅
)
(
−
̅
)
(
7
)
wh
er
e
̅
=
1
∑
an
d
̅
=
1
∑
2
=
1
−
√
(
)
(
)
(
8
)
3
=
√
∑
(
−
)
2
=
1
(
9
)
4
=
∑
|
−
|
=
1
(
1
0
)
Fig
u
r
e
1
1
.
MA
Ps
f
o
r
C
NN
ap
p
r
o
ac
h
es u
s
in
g
co
r
el
1K
Fig
u
r
e
1
2
.
APs
o
f
T
o
p
1
0
0
r
etr
iev
ed
im
ag
es f
o
r
co
r
el
1
K
d
at
ab
ase
im
ag
e
class
es
Fig
u
r
e
1
3
.
Pre
cisi
o
n
v
alu
es a
l
o
n
g
b
u
s
class
im
ag
es
Evaluation Warning : The document was created with Spire.PDF for Python.