Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
5
,
Octo
be
r
2020
,
pp.
4872
~
4880
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
5
.
pp
4872
-
48
80
4872
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
A
ne
w a
pp
ro
ach
to
im
age c
l
assific
atio
n
b
ased
on a
deep
m
ulti
class
A
da
B
oosting e
nsembl
e
Ha
s
an
Asil
1
, J
amsh
id
Ba
gh
e
rz
ad
eh
2
Depa
rtment
o
f
C
om
pute
r
Engi
n
e
eri
ng,
Facu
lty
of
Elec
tr
ical and
C
om
pute
r
Engi
n
e
eri
ng,
Urm
ia
Un
ive
rsit
y
,
Ir
an
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
20
, 201
9
Re
vised
Ma
r
11
,
2020
Accepte
d
Ma
r
30
, 202
0
In
rec
ent
y
e
ars,
dee
p
learni
ng
m
et
hods
have
been
deve
lope
d
in
orde
r
to
solve
the
probl
ems
.
The
se
m
et
hods
were
eff
ective
in
solving
complex
proble
m
s.
Convolut
ion
is
one
of
th
e
le
ar
ning
m
et
hods.
Thi
s
m
et
hod
is
appl
i
ed
in
cl
assif
y
ing
and
proc
essing
of
i
m
age
s
as
well
.
H
y
brid
m
et
hods
are
anot
her
m
ult
i
-
componen
t
m
ac
hine
le
a
rni
ng
m
et
hod.
The
se
m
et
hods
are
ca
t
egor
ized
int
o
inde
pend
en
t
and
depe
ndent
t
y
pes.
Ada
-
Bo
osting
al
gorit
hm
is
one
of
the
se
m
et
hods.
Today
,
th
e
class
ifi
c
at
ion
of
ima
ges
has
m
an
y
a
ppli
c
at
ions.
So
far
,
seve
ra
l
al
gorit
hm
s
have
bee
n
pre
sen
te
d
for
bina
r
y
and
m
ult
i
-
cl
ass
cl
assifi
ca
t
ion.
M
ost of
the
above
-
m
ent
ione
d
m
et
h
ods ha
ve
a
high
depe
nden
ce
on
the
data.
The
pre
sent
stud
y
in
te
nds
to
use
a
co
m
bina
ti
on
of
de
ep
learni
ng
m
et
hods
and
associate
d
h
y
br
id
m
et
hods
to
cl
assif
y
the
imag
es.
It
i
s
pre
sum
ed
tha
t
thi
s
m
et
ho
d
is
able
to
red
uce
the
err
or
r
a
te
in
images
class
ifi
cation
.
The
proposed
a
lgori
thm
consists
of
the
Ada
-
Boosting
h
y
brid
m
et
hod
and
bi
-
lay
e
r
convol
u
ti
onal
le
arn
ing
m
et
hod.
The
pro
posed
m
et
hod
w
as
ana
l
y
z
ed
aft
er
it
was
implemente
d
on
a
m
ult
i
-
class
Mnist
data
set
and
display
ed
the
resul
t
of
th
e
err
or
ra
te
red
u
c
ti
on.
The
r
esult
s
of
thi
s
stud
y
in
dic
a
te
th
at
the
err
or
ra
te
of
the
propose
d
m
et
hod
is
le
ss
tha
n
Ada
-
Boosting
and
convol
uti
on
m
ethods
.
Also,
the
net
work
has
m
ore
stabi
li
t
y
c
om
par
ed
to
the
o
the
r
m
et
hod
s.
Ke
yw
or
d
s
:
Boo
sti
ng
Conv
olu
ti
on
Deep l
ear
ning
Ensem
ble
m
e
t
hods
Im
age
c
la
ssific
at
ion
Copyright
©
202
0
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Jam
sh
id Bagh
e
rzad
e
h
,
Dep
a
rtm
ent o
f C
om
pu
te
r
E
ng
i
neer
i
ng,
Faculty
of Elec
tric
al
an
d C
om
pu
te
r
E
ng
i
neeri
ng
,
Ur
m
ia
U
niv
e
rs
it
y,
Ur
m
ia
, I
ra
n.
Em
a
il
: j.
baghe
rzad
e
h@u
rm
ia.
ac.ir
1.
INTROD
U
CTION
In
rece
nt
ye
ars
,
dee
p
le
a
rn
i
ng
m
e
tho
ds hav
e
g
ai
ne
d
a
lot of
at
te
ntion
a
nd w
ere
us
e
d
i
n
m
any
ty
pes
of
researc
h.
I
n
de
ep
le
ar
ning,
f
irst,
the
no
nlin
ear
fe
at
ur
e
s
of
seve
ral
la
ye
rs
are
e
xtracte
d
[1
]
.
The
n,
the
y
are
trans
ferred
to
a
cl
assifi
er,
an
d
finall
y,
they
are
sent
to
a
c
om
bin
er
la
ye
r,
in
order
to
perform
the
co
m
b
inati
on
and
pr
e
dicti
on.
The
m
or
e
hierar
chy
of
la
ye
rs
is
(d
eepe
r)
,
th
e
m
or
e
nonline
ar
featu
res
are
ob
ta
ine
d
an
d
be
tt
er
resu
lt
s
a
re
pres
ented.
T
hese
m
et
ho
ds
are
us
ed
t
o
s
olv
e
va
r
iou
s
com
plex
pro
blem
s
[2
]
.
On
e
of
t
hese
ne
tworks
is t
he
co
nvolu
t
ion
m
et
ho
d,
w
hic
h has m
any app
li
cat
io
ns
in im
age p
r
ocessi
ng
[3
]
. T
his
ne
twork was
or
i
gi
nally
desig
ne
d
to
proces
s
i
m
ages,
and
var
i
ous
convo
l
utio
n
-
base
d
m
od
el
s
hav
e
been
de
velo
ped
t
o
cl
assify
the im
ages.
On
the
oth
e
r
hand,
hy
br
id
m
et
ho
ds
a
re
c
omm
on
m
achine
le
ar
ning
te
chn
i
qu
e
s.
T
he
se
m
e
tho
ds
com
bin
e d
if
fere
nt pred
ic
ti
ons
to
pr
ov
i
de bett
er and
m
or
e ac
cur
at
e
pro
blem
-
so
l
ving r
e
su
lt
s
[
4].
These
m
eth
ods
are
us
e
d
f
or
s
ol
vin
g
cl
assifi
ca
ti
on
pro
blem
s
and
hav
e
a
pp
li
ed
in
dif
fer
e
nt
cl
assifi
cat
ion
pro
blem
s
in
ord
er
to
pro
vid
e
var
i
ous
predict
io
ns
[
5].
No
al
gorithm
s
are
opti
m
a
l
i
n
al
l
ar
eas.
E
ver
y
le
a
rn
i
ng
al
gorithm
is
lim
it
ed
to
a
s
pe
ci
fic
m
od
el
.
In
fact,
if
dat
a
assum
ption
s
are
no
t
true,
error
s
will
occu
r
[
6].
If
the
m
od
el
par
am
et
ers
are
def
i
ne
d
for
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A n
ew
appro
ac
h
to
im
ag
e
clas
sif
ic
ation
ba
s
e
d on a dee
p m
ul
ti
cl
as
s AdaB
oosti
ng e
ns
em
bl
e
(
Ha
s
an Asil
)
4873
a
trai
nin
g
set
,
the
m
od
el
m
ight
no
t
be
ap
propriat
e
fo
r
s
om
e
data,
f
or
w
hich
ano
the
r
m
od
el
can
produce
be
tt
er
resu
lt
s
[
2].
E
nse
m
ble
cl
assifier
s
are
am
ong
the
m
ulti
co
m
p
on
e
nt
cl
assifi
e
rs
de
fine
d
to
pro
duce
bette
r
r
esults
than
a
si
ng
le
-
c
om
po
ne
nt
cl
assifi
er
[
7].
I
n
s
uch
cl
assi
ficat
ion
s
,
e
ns
em
bl
e
cl
assifi
ers
are
us
e
d
to
ob
ta
in
bette
r
resu
lt
s.
En
sem
ble
m
et
ho
ds
var
y
i
n
ho
w
they
create
diff
e
ren
t
cl
assifi
ers
a
nd
ho
w
they
com
bin
e
basi
c
cl
assifi
ers wit
h res
pect to t
hei
r
w
ei
ghts [
8].
In
fact,
the
re
are
two
e
ns
e
m
ble
fr
am
ewo
rk
s:
de
pe
nd
e
nt
(ser
ia
l)
an
d
ind
e
pe
nd
e
nt
(
pa
rall
el
)
[9
]
.
In
a
de
pende
nt f
ram
ewo
r
k,
t
he
o
ut
pu
t
of
a cl
assifi
er is u
se
d i
n
the n
e
xt classi
fier.
T
her
e
f
ore, the
knowle
dg
e
of
pr
e
vious
it
erati
on
s
can
be
use
d
to
direct
le
arn
i
ng
in
ne
xt
it
erati
on
s
[10
]
.
Boo
sti
ng
is
an
instance
of
su
ch
fr
am
ewo
r
k
s.
I
n
the
sec
ond
f
ram
ewo
rk
(
known
as
i
nd
e
pe
nd
e
nt)
,
eve
ry
cl
assifi
er
is
cr
eat
ed
in
dep
e
ndently
.
Their o
utputs a
re c
om
bin
ed w
it
h
those
of
voti
ng
m
et
ho
ds [
11]
.
So
far,
m
any
te
ch
niques
hav
e
be
en
pr
opos
e
d
for
cl
assify
ing
the
i
m
ages.
S
om
e
of
t
hese
cl
assifi
cat
ion
s
are
bi
-
cl
ass
and
so
m
e
othe
rs
are
m
ulti
-
cl
ass
[8
]
.
It
sh
oul
d
be
note
d
that
m
ulti
-
cl
ass
cl
assifi
cat
ion
s
hav
e
m
any
ap
plica
ti
on
s.
Cre
at
ing
a
r
obus
t
m
od
el
in
li
ne
with
net
work
sta
bili
ty
,
reduci
ng
cl
assifi
cat
ion
e
rror an
d decrea
sing data
de
pe
nd
e
ncy in
the
m
od
e
l are so
m
e of im
age classi
ficat
ion
iss
ue
s
[
12
]
.
The
pr
e
sent
stud
y
pro
poses
a
ne
w
m
od
el
base
d
on
dee
p
le
arn
i
ng
m
et
ho
ds
a
nd
hy
br
i
d
m
et
ho
d
to
i
m
pr
ove
an
d
r
edu
ce
the
er
r
or
rate
of
m
ulti
-
cl
ass
i
m
age
c
la
ssific
at
ion
.
On
the
ot
her
ha
nd,
com
bin
in
g
these
te
chn
iq
ues
res
ults
in
hi
gh
e
r
sta
bili
ty
,
com
par
ed
to
pre
vious
m
et
ho
ds.
This
hybr
i
d
boos
ti
ng
m
eth
od
is
a
bi
-
la
ye
r
co
nv
olu
ti
on r
el
at
ed
hybri
d
m
et
ho
d.
As
yo
u
know,
the g
oal of b
oo
sti
ng
is
to
stre
ngthe
n
a
nd
im
pr
ov
e
poor
le
arn
i
ng,
wh
ic
h
at
te
m
pts
to
stren
gt
hen
it
s
le
arn
i
ng
a
nd
m
ake
the
m
od
el
stron
ger
.
In
thi
s
stud
y,
the b
i
-
la
ye
r
c
onvoluti
on tec
hniq
ue was c
ons
idere
d
a
wea
k
l
earn
i
ng m
et
ho
d.
2.
BACKG
ROU
ND
In
t
he
past,
deep
le
ar
ning
netw
orks
we
re
use
d
to
s
olv
e
dif
fer
e
nt
pro
blem
s
su
ch
as
im
age
cl
assifi
cat
ion
,
obj
ect
rec
ogniti
on
,
im
age
extracti
on,
et
c.
[13].The
se
netw
orks
em
plo
ye
d
var
io
us
m
et
ho
ds
f
or
op
ti
m
iz
ation
.
Conv
olu
ti
onal
neural
netw
ork
s
(CN
Ns)
a
re
a
m
on
g
t
he
m
os
t
i
m
po
rtant
de
ep
le
arn
i
ng
net
works
po
s
sessin
g
se
ver
al
la
ye
rs
a
nd
ser
ving
as
a
ve
ry
po
w
erful
m
et
ho
d
for
m
achine
visi
on
op
ti
m
izati
on
.
These
netw
ork
s
co
ns
ist
of
t
hree
la
ye
rs
nam
ed
c
onvoluti
ona
l,
poolin
g,
an
d
fu
ll
y
c
onnecte
d.
C
NNs
hav
e
been
us
e
d
f
or im
age classi
ficat
ion
.
A
dee
p
le
ar
ni
ng
m
od
el
was
i
ntr
oduce
d
by
Dan
ie
l
F
rits
et
al
.
for
im
age
cl
assifi
cat
ion
.
This
m
od
el
com
bin
es
a
de
ep
net
wor
k
wit
h
PixelR
NN
a
nd
DCGR
A
N
m
od
el
s
fo
r
im
age
rec
ogniti
on.
T
hese
m
od
e
ls
wer
e
i
m
ple
m
ented
in
PixelR
N
N
a
nd
DCGR
A
N
for
ha
ndw
ritt
en
data
[
14]
.
G
an
Pro
j
ect
was
carried
out
for
i
m
age
cl
assifi
cat
ion
,
too
.
It
is
us
e
d
in
a
k
-
cl
as
s
cat
egorizat
ion
[
15
]
.
Anothe
r
cl
assifi
cat
ion
pro
j
ect
incl
ud
e
d
the
aux
il
ia
ry
de
ep
ge
ner
at
ive
m
od
el
s
(ADGM
s).
A
DG
Ms
consi
st
of
va
ri
ou
s
set
s
of
enc
od
e
rs
an
d
de
c
od
e
rs
.
This
pro
j
ect
w
as
execu
te
d
on
the
MINIST
dataset
.
Acc
ordin
g
to
the
res
ults,
Atla
sRB
F
al
go
rithm
pr
oduc
e
d
1.5 tim
es
m
or
e co
nv
e
r
gen
ce
than dee
p ge
nerat
ive m
od
el
s,
vi
rtual ad
ve
rsar
i
al
, and
la
dde
r m
et
ho
ds [
16
-
18]
.
A
m
ajo
r
ap
pro
ach
to
dee
p
ne
tworks
is
to
cr
eat
e
diff
ere
nt
la
ye
rs
for
le
arni
ng
the
feat
ur
e
s
[19,
20
]
.
An
instanc
e
of
su
c
h
netw
ork
s
is
an
ei
gh
t
-
la
ye
red
netw
ork
desig
ned
by
Kr
iz
hevsky
et
al
.
[
21
]
.
It
is
a
dee
p
netw
ork
arc
hit
ect
ur
e
based
on
the
c
onvolut
ion
of
a
dee
p
ei
gh
t
-
la
ye
re
d
netw
ork.
Give
n
the
s
uccess
of
t
his
netw
ork,
it
is
us
e
d
wi
dely
to
so
lve
m
any
pr
oble
m
s
su
ch
as
vid
e
o
cl
assi
ficat
ion
[22],
f
aci
al
recogn
it
ion,
a
nd
act
ion
detect
io
n
[
20
]
.
Othe
r
r
el
evan
t
pr
oj
ect
s
inclu
de
the
use
of
a
c
onvol
ution
a
wea
k
le
arn
e
r
in
the
e
ns
em
ble
Ad
aB
oosti
ng
a
lgorit
hm
[2
3].
Weak
le
ar
ner
i
s
a
le
ar
ner
t
hat
no
m
at
te
r
what
the
distri
bu
t
ion
over
t
he
tr
ai
nin
g
data
is
will
al
w
ay
s
do
bette
r
t
han
cha
nce
,
w
hen
it
trie
s
to
la
bel
the
data.
This
m
eans
tha
t
the
le
ar
ner
al
gorithm
is
al
ways
goin
g
to
le
a
rn
s
om
et
hin
g,
not
al
ways
com
pletely
accurate.
Th
e
c
om
bin
at
ion
of
a
c
onvo
luti
on
al
netw
ork
an
d
A
daBo
os
ti
ng
m
e
thod
has
al
so
been
us
e
d
in
num
erati
on
.
Th
e
aim
of
this
pa
per
was
to
pr
opos
e
a
m
od
el
based
on
the
com
bin
at
ion
of
A
da
Boo
sti
ng
m
et
h
od
a
nd
a
two
-
la
ye
red
co
nvol
ution
al
net
wor
k
to
pro
du
ce
b
et
te
r
resu
lt
s
.
3.
RESEA
R
CH MET
HO
D
In
t
he
propose
d
a
ppro
ac
h,
en
sem
ble
m
et
hods
a
nd
dee
p
le
anin
g
wer
e
em
plo
ye
d
to
im
pr
ov
e
im
age
cl
assifi
cat
ion
.
Fo
r
this
purpos
e,
a
co
nvol
utio
nal
ne
ur
al
net
work
was
com
bin
e
d
with
the
Ad
aB
oosti
ng
m
et
ho
d.
E
v
e
r
y
e
n
s
e
m
bl
e
p
r
o
b
l
e
m
c
o
n
s
i
s
t
s
o
f
f
o
u
r
s
e
c
t
i
o
n
s
:
a
t
r
a
i
ni
n
g
s
e
t
,
a
b
a
s
i
c
l
e
a
r
n
e
r
,
a
d
r
i
v
e
r
,
a
n
d
a
c
om
p
o
s
e
r
[
1
0
]
.
Likewise,
the
prop
os
ed
m
et
ho
d
incl
ud
e
s the
foll
ow
i
ng secti
on:
a.
Trainin
g
set
:A
trai
ning
set
c
onta
ins
la
belle
d
sam
ples
us
ed
f
or
trai
ni
ng.
In
this
stu
dy,
the
MINIST
sta
nd
ar
d
dataset
w
a
s
em
plo
ye
d f
or im
a
ge
cl
assifi
cat
io
n.
b.
Ba
sic
le
arn
er:
A
ba
sic
le
arn
e
r
is
a
le
ar
ning
al
gorithm
us
ed
f
or
le
ar
ning
a
trai
ning
set
.
In
t
his
pr
oj
ec
t,
the
ba
sic
le
ar
ne
r
was
us
e
d
to
i
m
ple
m
ent
de
ep
le
ar
ning.
F
or
this
pur
pose
,
a
c
onvoluti
on
al
ne
ur
al
netw
ork
was
us
e
d
as
a
le
arn
ing
al
go
rithm
.
Ho
wever,
the
CNN
c
on
sist
e
d
two
c
onvoluti
onal
f
un
ct
io
ns
in
ev
ery
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
72
-
48
80
4874
cl
assifi
er
f
or
m
or
e
converge
nce.
On
t
he
ot
her
ha
nd,
the
conv
olu
ti
onal
functi
ons
are
base
d
on
er
r
or
rate
reducti
on in
learni
ng.
c.
Dr
i
ver
: T
his c
om
po
nen
t i
s
us
e
d
to
creat
e
dif
f
eren
t cl
assifi
e
r
s.
d.
Com
po
ser:
T
hi
s
com
po
ne
nt
is
us
e
d
to
c
om
bin
e
cl
assifi
cat
ion
m
et
ho
ds.
I
n
this
pro
je
ct
,
the
boos
ti
ng
te
chn
iq
ue was
us
e
d.
Figure
1
s
hows
a sc
hem
atic v
ie
w of t
he pr
op
os
e
d
m
et
ho
d f
or com
bin
at
ion.
Figure
1. The
g
ene
ral
a
r
chite
ct
ur
e
of the
pro
po
s
ed
m
et
ho
d
The
A
daBo
os
t
ing
net
wor
k
was
create
d
a
nd
c
om
bin
ed
with
the
co
nv
olu
ti
onal
ne
ural
network
to
ob
ta
in
a
po
werfu
l
le
ar
ner
f
rom
weak
c
onvo
luti
on
al
le
ar
ners.
Acc
ordi
ng
ly
,
the
c
onvoluti
on
al
netw
ork
c
on
sist
s
of
t
hr
ee
la
ye
rs
:
conv
olu
ti
ona
l,
poolin
g,
a
nd
fu
ll
y
co
nnect
ed.
T
he
fu
ll
y
connecte
d
la
ye
r
is
res
pons
i
ble
for
com
pu
ti
ng
t
he
cl
assifi
cat
ion
[
23,
24]
.
I
n
the p
r
opos
e
d
al
go
r
it
h
m
,
two
de
ep
conv
olu
ti
onal
lay
ers
of
e
ve
ry
cl
ass
are r
e
spo
ns
ible
for
wea
k
le
ar
ni
ng
.
T
he
m
ulti
c
la
ss boo
sti
ng
ne
twork
is d
e
fin
ed
in
the
f
ollo
wing
way [
13]
.
F: X→
{1…M
}
(1)
In
t
his str
uctu
r
e,X
is all
data in Cla
ss M also
ev
ery
xire
fer
s
to a class
nu
m
ber
of a label
num
ber
of an
M
-
cl
assifi
cat
ion
. F(
X) is trai
ne
d glo
bally
and
dev
el
op
e
d
i
n t
he
f
ollo
wing
way
[
9].
F(
x) =
ar
g
m
ax
<Y
k
,
F(X
)>k=1...M
(2)
In
(
2
)
, YK¬ i
ndic
at
es the la
be
l of a
k
-
cl
assif
ic
at
ion
.In
fact,
the outp
ut
of F
(X) is as
foll
ows
[13].
F(
x) =
[
f
1(x)... fM(
x)
]
∈
RM
(3)
On
the
oth
e
r
ha
nd,
the
m
ulti
cl
ass
boos
t
ing
m
et
ho
d
di
vid
es
t
he
le
a
rn
i
ng
pro
ce
ss
into
se
ve
ral
su
bc
la
sses to
im
pr
ov
e lea
r
ni
ng. T
he
refor
e
, g
i:
X
→RD
sho
ws
trai
ning
fun
ct
ion
s.
F(
X
)
=
∑
∝
(
)
(4)
In
this
functi
on,
t
and
α
sh
ow
the
fr
e
qu
e
ncy
an
d
c
orrelat
ion
c
oe
ff
ic
ie
nt
in
ea
ch
it
erati
on,
resp
ect
ively
.
I
n
this
m
et
ho
d,
GD
-
MC
Bo
os
t
was
em
plo
ye
d
to
im
pr
ov
e
le
a
rn
i
ng
[
13
]
.
Co
ns
ide
rin
g
t
he
l
earn
i
ng
cl
assifi
cat
ion
,
the
pro
pose
d
al
gorithm
us
ed
a
m
ini
m
iz
a
t
ion
poli
cy
in
wh
ic
h
a
desce
nd
i
ng
gradie
nt
was
e
m
plo
ye
d [15].
Th
e
refor
e
, the
m
ini
m
u
m
w
eak lear
ner is sel
ect
ed
in t
he fol
lowing
way:
g=arg
m
in
R
[f
;g]
(5)
The
(
6
)
i
nd
ic
at
es the
functi
on
us
e
d
f
or m
ini
m
iz
at
ion
[7].
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A n
ew
appro
ac
h
to
im
ag
e
clas
sif
ic
ation
ba
s
e
d on a dee
p m
ul
ti
cl
as
s AdaB
oosti
ng e
ns
em
bl
e
(
Ha
s
an Asil
)
4875
[
;
]
=
−
1
2
|
|
|
|
+
∑
<
(
)
(
)
>
,
∈
(6)
The follo
wing
al
gorithm
is a p
art
of the
sour
ce co
de of
the
pro
po
se
d
m
et
ho
d:
Algorithm
1
. T
he
a
lg
ori
thm
s
t
ru
ct
ur
e
cnn.layers
= {
struct(
'type'
,
'i'
)
%input layer
struct(
'type'
,
'c'
,
'outputmaps'
, 6,
'kernelsize'
, 5)
%convolution layer
struct(
'type'
,
's'
,
'scale'
, 2)
%sub sampling layer
struct(
'type'
,
'c'
,
'outputmaps'
, 12,
'kernelsize'
, 5)
%convolution layer
struct(
'type'
,
's'
,
'scale'
, 2)
%sub sampling layer
};
Sample code for any classification
%
opts.alpha = 1;
opts.batchsize = 50;
opts.numepochs = 500;
cnn,er5]= cnntrain(cnn, train_x5, train_y5, opts);
D(1,5), bad4] = cnntest(cnn, test_x5, test_y5);
cnn = cnnsetup(cnn, train_x6, train_y6);
opts.alpha = 1;
opts.batchsize = 50;
opts.numepochs = 500;
[cnn,er6] = cnntrain(cnn, train_x6, train_y6, opts);
[D(1,6), bad5] = cnntest(cnn, test_x6, test_y6);
cnn = cnnsetup(cnn, train_x7, train_y7);
opts.alpha =
1;
opts.batchsize = 50;
opts.numepochs = 500;
[cnn,er7] = cnntrain(cnn, train_x7, train_y7, opts);
[D(1,7), bad6] = cnntest(cnn, test_x7, test_y7);
cnn = cnnsetup(cnn, train_x8, train_y8);
cnn,er10]= cnntrain(cnn, train_x10, train_y10, opts);
D(1,10), bad
9] = cnntest(cnn, test_x10, test_y10);
Increase the result by min
%
er=cat(1,er1,er2,er3,er4,er5,er6,er7,er8,er9,er10);
er=min(er);
4.
RESU
LT
S
AND A
N
ALYSIS
The
pro
pose
d
al
gorithm
was
i
m
ple
m
ented
on
the
MIN
IS
T
dataset
in
MA
TLAB
f
or
e
valuati
on
[25]
.
The
MNI
ST
da
ta
base
(Mo
dif
ie
d
Nati
on
al
I
nst
it
ute
of
Standa
rd
s
a
nd
Tec
hnol
og
y
data
bas
e)
is
a
la
rg
e
database
of
hand
wr
it
te
n
dig
it
s that is com
m
on
ly
u
sed
f
or
traini
ng
var
i
ou
s im
age p
ro
c
essing
syst
em
s
[26]
. Th
e d
at
a
bas
e
is
al
so
widely
us
e
d
for
trai
ni
ng
a
nd
te
sti
ng
in
the
fiel
d
of
m
achine
le
arn
i
ng
[
27,
28]
.
It
was
creat
ed
by
"re
-
m
ixing
"
th
e
sam
ples
fr
om
NI
ST'
s
or
igi
nal
dataset
s.
F
ur
t
her
m
or
e,
t
he
black
an
d
w
hite
i
m
ages
f
rom
NI
ST
wer
e
norm
al
ized
to
fit
into
a
28x28
pi
xel
boun
ding
bo
x
a
nd
a
nti
-
al
ia
sed
,
w
hich
i
ntrod
uced
gray
scal
e
le
vels.
The
M
NI
S
T
da
ta
base
c
on
ta
i
ns
60,00
0
trai
ning
im
ages
and
10,
000
te
st
ing
im
ages
[
29]
.
Sect
io
n
4.1
to
4.4
sh
ows
t
he
te
sti
ng
resu
lt
s.
I
n
these
sect
io
ns
,
the
CN
N,
AdaB
oo
sti
ng
m
eth
od,
a
nd
the
pro
po
se
d
e
ns
e
m
ble
m
ul
ti
cl
ass
con
vo
l
ution
al
A
da
Boo
sti
ng
m
et
h
od
wer
e
eval
ua
te
d
to
com
par
e
the
resu
lt
s.
On
the
ot
her
hand,
ever
y t
est
was e
valuate
d
wit
h diff
e
re
nt it
erati
on
s
to
an
al
yz
e
the r
es
ults
bette
r.
4.1.
AdaBo
os
t
Figure
2
i
nd
i
cat
es
the
proc
ess
of
le
a
rn
i
ng
the
Ad
aB
oosti
ng
al
gorith
m
fo
r
diff
e
re
nt
it
erati
o
ns.
Accor
dingly
,
the
er
ror
rate
de
creased
w
hen
the
num
ber
of
it
erati
on
s
inc
r
eased.
In
t
his
al
gorithm
,
the
sa
m
ple
sp
ace
was divi
ded into
classe
s.
T
hen the lea
rn
i
ng
proce
ss
was
ca
rr
ie
d
out
on those cla
ss
es in th
e
AdaB
oo
sti
ng
m
et
ho
d.
Nat
ural
ly
,
the
num
ber
of
it
erati
ons
an
d
le
ngth
of
le
arn
in
g
i
ncr
ea
sed.
I
n
oth
e
r
w
ords,
the
re
were
10
0,
500,
an
d
2000
processes
in
10,
50,
an
d
200
it
erati
ons
.T
he
le
ar
ning
cy
cl
e
find
the
s
et
of
par
am
et
ers
to
op
ti
m
iz
e the er
ror fu
nctio
n.
In the e
nd,
the
m
od
el
a
nd
par
a
m
et
ers
are
with t
he
sm
al
le
st
error.
4.2
.
Single
-
l
ay
ere
d
CNN
A
CN
N
co
ns
is
ts
of
c
onvoluti
on
al
,
poolin
g,
and
fu
ll
y
co
nnect
ed
la
ye
rs
[
30]
.
The
refo
re,
a
com
par
iso
n
sh
oul
d
be
m
ad
e
between
this
m
et
ho
d
an
d
the
pro
posed
m
et
ho
d.
T
he
CNN
wa
s
eval
uated
on
the
MINIST
dataset
.
Fig
ur
e
3
sh
ows
the
resu
lt
s
of
the
CNN.
Fi
gure
2
ind
ic
at
es
the
CNN
le
arn
in
g
process
in
10,
50
,
and 20
0
it
erati
on
s
. Acc
ordin
gl
y, the CN
N
is
slow
e
r
tha
n
t
he
AdaBo
os
ti
ng
m
et
ho
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
72
-
48
80
4876
Figure
2.
E
rro
r
r
at
e
c
om
par
iso
n
b
ase
d on the
n
um
ber
of
i
te
r
at
ion
s i
n
the
in
AdaBo
os
ti
ng
Figure
3. Er
ror
r
at
e
c
om
par
ison
b
a
sed
on t
he
n
um
ber
of
i
te
rati
on
s
in
t
he
C
NN
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A n
ew
appro
ac
h
to
im
ag
e
clas
sif
ic
ation
ba
s
e
d on a dee
p m
ul
ti
cl
as
s AdaB
oosti
ng e
ns
em
bl
e
(
Ha
s
an Asil
)
4877
4.3
.
Tw
o
-
la
yered
CNN
Con
si
der
i
ng
th
e
us
e
of
tw
o
-
la
ye
red
CN
Ns
in
weak
le
ar
ning
,
a
conv
olu
ti
on
al
deep
le
ar
ning
net
wor
k
was
form
ed.
T
he
cl
assifi
cat
io
n
resu
lt
s
of
th
at
netw
ork
we
re
e
valuated
t
o
pro
du
ce
bett
er
re
su
lt
s.
Fig
ur
e
4
ind
ic
at
es the
r
e
su
lt
s of t
hat n
et
work.
Figure
4. Er
ror
r
at
e
c
om
par
ison
b
a
sed
on t
he
n
um
ber
of
i
te
rati
on
s
in
t
he
t
wo
-
l
ay
ered C
N
N
4.4
.
Deep m
ulticl
ass A
da
B
oostin
g
The
pr
opos
e
d
m
et
ho
d
was
a
nensem
ble
of
Ad
aB
oosti
ng
and
c
onvoluti
on
al
m
et
ho
ds
.
The
sam
ple
sp
ace
was
di
vid
ed
int
o
te
n
unit
s
in
co
m
bin
at
ion
with
the
co
nvolu
ti
onal
m
e
thod.
Fig
ur
e
5
sh
ows
the
res
ul
ts
of
le
arn
in
g proces
s in dif
fer
e
nt it
erati
on
s
.
4.5
.
Results
and
discussi
on
In
order
t
o
eva
luate
the
pro
po
sed
m
et
ho
d,
t
he
i
m
age
cl
assifi
cat
ion
was
c
onduct
ed
on
the
Mnist
data
set
with
dif
fere
nt
al
gorithm
s
.
The
m
et
ho
ds
us
ed
f
or
im
ple
m
enting
thes
e
exp
e
rim
ents
wer
e
pr
ese
nt
ed
i
n
sect
ion
s
4.1
to
4.4.
The
m
on
o
-
an
d
bi
-
la
ye
r
convo
l
ution
m
et
ho
ds
wer
e
car
ried
out
to
eva
luate
the
A
da
Bo
os
ti
ng
m
et
hods
.
T
he
er
ror
rates wer
e d
et
erm
ined
ba
sed
on
th
e
nu
m
ber
o
f
re
petit
ion
s.
The
r
esults
of
these
e
xp
e
rim
ents
are
s
how
n
in
Ta
ble
1.
Fo
r
e
xam
ple,
in
500
re
petit
ion
s,
the
e
rror
r
at
e
of
A
da
-
bo
os
ti
ng,
bila
te
r
conv
olut
ion
,
an
d
de
ep
Ada
-
bo
os
ti
ng
are
0.008,
0.0
089,
a
nd
0.0
067
r
especti
vely
.
On
the
oth
e
r
ha
nd,
if
the
num
ber
of
rep
et
it
io
ns
increase
d
t
hen
a
bette
r
conve
rg
e
nce
c
an
be
ac
hie
ve
d
in
the
pro
po
s
ed
m
et
ho
d.Acc
ordin
g
to
th
e
res
ults,
the
im
age
cl
assifi
cat
ion
error
rate
in
th
e
pro
p
os
e
d
m
et
hod
wa
s
8%
bette
r
than
the
pr
e
vi
ou
s
m
et
ho
d
a
nd
ha
d
bette
r
sta
bili
ty
as
well
.
Table
1
s
how
s
the
erro
r
rat
es
in
three
dif
fer
e
nt
sit
uations
.
Fig
ur
e
6 sho
ws
a
com
par
ison d
r
awn b
et
ween t
he pr
opos
e
d
al
gorithm
an
d ot
her m
et
ho
ds
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
72
-
48
80
4878
Figur
e
5. Er
ror
r
at
e
c
om
par
ison
b
a
sed
on t
he
n
um
ber
of
i
te
rati
on
s
in
d
ee
p
m
ul
ti
cl
ass A
da
Boo
sti
ng
Table
1.
E
rro
r
r
at
e
c
om
par
iso
n
b
ase
d on
a
lg
or
it
hm
s an
d
i
te
rati
on
s
Ro
w
Iter
atio
n
s
Ad
aBo
o
stin
g
Sin
g
le
-
Layered
CNN
Two
-
La
y
e
red
CNN
Co
n
v
o
l
u
tio
n
al
Multiclas
s Ad
aBo
o
stin
g
1
10
0
.01
3
5
0
.46
8
0
0
.03
2
3
0
.45
0
1
2
50
0
.00
9
4
0
.50
0
0
0
.00
3
7
0
.04
4
6
3
200
0
.00
9
0
.01
9
7
0
.00
8
9
0
.01
6
8
4
500
0
.00
8
0
.08
7
0
.00
8
1
0
.00
6
7
Figure
6. The
e
rror
r
at
e
of
a
lg
or
it
hm
s
b
ase
d on
i
te
rati
ons
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A n
ew
appro
ac
h
to
im
ag
e
clas
sif
ic
ation
ba
s
e
d on a dee
p m
ul
ti
cl
as
s AdaB
oosti
ng e
ns
em
bl
e
(
Ha
s
an Asil
)
4879
5.
CONCL
US
I
O
N
Deep
le
arn
i
ng
is
now
us
e
d
f
or
so
l
ving
var
i
ous
c
om
plex
prob
le
m
s.
Conv
olu
ti
on
is
on
e
of
t
he
deep
le
arn
in
g
m
et
ho
ds,
w
hich
i
s
us
e
d
f
or
c
la
ssifyi
ng
im
a
ges.
Th
e
h
yb
rid
m
et
ho
ds
are
c
onsidere
d
a
s
a
m
ulti
-
com
po
nen
t
le
ar
ning
m
et
ho
d
t
hat
inten
ds
to
st
re
ng
t
hen
t
he
le
a
rn
i
ng
process
by
us
i
ng
c
ombine
d
appr
oach
es
.
B
oo
sti
ng
i
s
one
of
t
hese
m
et
hods
.
This
stu
dy
seeks
t
o
fi
nd
a
m
et
ho
d
f
or
optim
izing
a
nd
re
du
ci
ng
error
s
w
hen
cl
assify
ing
im
ages.
So
far,
var
i
ou
s
m
eth
ods
a
nd
m
od
el
s
ha
ve
bee
n
prese
nted
f
or
t
his
cl
assifi
cat
ion
.
The
pro
posed
m
et
ho
d
is
bas
ed
on
a
c
om
bin
at
ion
of
bo
ost
ing
m
et
ho
d
a
nd
de
ep
co
nvol
ution
te
chn
iq
ue.
B
oosti
ng
ha
s
a
set
of
basic
le
arn
i
ng
t
hat
i
m
pr
ov
es
and
st
rength
ens
the
poor
le
arn
i
ng
in
eac
h
ste
p.
In
t
he
pro
po
se
d
m
et
ho
d,
the
bi
-
la
ye
r
c
onvo
luti
on
netw
ork
is
co
ns
ide
red
as
basic
le
ar
ni
ng.
T
his
m
et
ho
d
was
i
m
ple
m
ented
us
in
g
M
ATL
AB
s
of
twa
re
and
eval
uation
was
ca
rr
ie
d
ou
t
us
in
g
ha
ndw
ritt
en
data.
In
this
al
gorithm
,
the
Mnist
data
s
et
was
app
li
ed.
A
fter
eval
ua
ti
ng
the
res
ults
of
the
pro
pose
d
m
et
ho
d,
it
was
analy
zed
us
i
ng
convo
l
ution
a
nd
boos
ti
ng
m
et
hods
.
T
he
res
ults
ind
ic
at
ed
that
in
500
re
pe
ti
ti
on
s,
the
erro
r
rate
was decrease
d by 8%
. On
the
oth
e
r
ha
nd, s
in
ce the boost
ing
m
et
ho
d was
use
d,
t
he
pro
pos
ed
m
od
el
h
as
hi
gh
e
r
sta
bili
ty
.
In
the
fu
tu
re,
it
is
possible
to
im
pr
ov
e
the
propose
d
al
gorithm
by
stren
gth
e
ning
the
basic
al
gor
i
th
m
and creati
ng s
om
e connecti
on
s in
t
he wei
ghti
ng pr
ocess.
REFERE
NCE
S
[1]
H
.
F.
Parde
de
,
e
t
al
.
,
“
Convolu
ti
o
nal
Neura
l
Netw
ork
and
Feat
ur
e
Tra
nsform
at
ion
f
or
Distant
Spee
c
h
Rec
ogni
ti
on,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
v
ol
.
8
,
n
o
.
6,
pp.
5381
-
538
8,
2018
.
[2]
S
.
Khan
and
Kiruba
nand
V.B
.
,
“
Com
par
ing
m
a
chi
ne
l
ea
rn
ing
a
nd
ense
m
ble
lea
rning
in
the
f
ie
l
d
of
footba
ll,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
v
ol
.
9
,
n
o
.
5,
pp.
4321
-
432
5,
2019
.
[3]
R
.
Carua
na
and
A
.
N
.
Mizi
l,
“
An
E
m
piri
cal
Com
par
ison
of
Supervised
Le
arn
ing
Algorit
hm
s,
”
ICML
'06
Proce
ed
ings o
f
t
he
23rd i
n
te
rnati
onal
con
fe
ren
ce
on
Mac
hin
e le
ar
ning,
USA
,
pp
.
1
61
-
168,
2006
.
[4]
E.
Alp
a
y
din
,
“
In
troduc
ti
on
to
Ma
chi
ne
Learni
ng
,
”
Mass
ac
husett
s
,
US
A
,
MIT
Press
,
2004
.
[5]
C
.
W
.
Hs
u
and
C
.
J
.
Lin
,
“
A
compari
son
of
m
et
hods
for
m
ult
ic
la
ss
support
vec
tor
m
ac
hine
s,
”
I
EE
Et
rans
action
on
neural
ne
tworks
,
vol
.
13,
no.
2,
p
p.
415
-
425
,
200
2.
[6]
T
.
T
.
Khuat
an
d
M
.
H
.
Le,
“
Ensemble
le
arn
in
g
for
software
f
aul
t
pre
di
ction
proble
m
with
imbala
nce
d
da
ta,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
v
ol
.
9
,
n
o
.
4,
pp.
3241
-
324
6,
2019
.
[7]
M
.
Ghos
h
and
P
rab
u
P
.
,
“
Empiric
al
Ana
l
y
s
is
of
Ensemble
Methods
for
the
Cla
ss
ifi
c
at
io
n
of
Roboca
ll
s
in
Te
l
ec
om
m
unic
ations
,
”
In
te
rnati
onal
Journal
of
Elec
tri
cal
and
Computer
Eng
ine
ering
(
IJE
CE)
,
v
ol.
9,
n
o.
4,
pp.
3108
-
3114
,
2019
.
[8]
S
.
Ravi
das
and
M.
A.
Ans
ari
,
“
Dee
p
learni
ng
f
or
pose
-
inva
ri
an
t
face
de
tecti
on
in
unconstraine
d
envi
ronm
ent,
”
Inte
rnational
Jo
urn
al
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
v
ol
.
9
,
n
o
.
1,
pp.
577
-
584,
2019
.
[9]
J
. T
anh
a,
“
Ense
m
ble
appr
oa
che
s
to
sem
i
-
superv
i
sed
learni
ng
,
”
Thesis
,
Unive
rsit
y
of
Am
sterda
m
,
2013
.
[10]
A
.
Rasm
us,
et
al
.
,
“
Sem
i
-
Supervised
Le
a
rning
with
La
dder
N
et
works
,
Neura
l
and
Evol
uti
on
a
r
y
Com
puti
ng,
”
arXiv
:1507.
0267
2
,
2015
.
[11]
J
.
Ta
nha
,
et
al
.
,
“
An
AdaBoost
Algorit
hm
for
Multi
xla
ss
Sem
i
-
supervise
d
Le
a
rning,
”
IE
EE
1
2th
Inte
rnationa
l
Confe
renc
e
on
Data
Mini
ng
(
ICDM)
,
vol.
1
,
pp
.
1116
-
1121,
201
2
.
[12]
M
.
S
.
H
.
Al
-
Ta
m
imi,
“
Co
m
bini
ng
convol
uti
on
a
l
neur
al
ne
tworks
and
slant
le
t
t
ran
sform
for
an
eff
ec
t
ive
imag
e
ret
ri
eva
l
sche
m
e,
”
Int
ernati
ona
l
Journal
of
E
le
c
tric
al
and
C
omputer
Engi
n
e
ering
(
IJE
CE)
,
v
ol.
9,
n
o.
5,
pp.
4382
-
4395
,
2
019
.
[13]
M
.
Moghim
i,
et
al
.
,
“
Boosted
C
onvolut
ional
Ne
ura
l
Networks,
”
Proce
ed
ings
of
the
27th
Bri
ti
sh
Mac
hine
Vi
sion
Confe
renc
e
,
pp.
24.
1
-
24.
13
,
201
6
.
[14]
A
.
van
Oord
,
et
al
.
,
“
Pixel
Re
cur
ren
t
Neura
l
Net
works
,
”
Proce
ed
ings
of
the
I
nte
rnational
Confe
r
enc
e
on
Mac
hin
e
Learning,
vol
.
4
8,
pp
.
1747
-
175
6,
2016
.
[15]
A
.
Oden
,
“
Sem
i
-
Supervised
L
earning
with
Gene
rat
iv
e
Adversar
i
al
Ne
tworks,
”
D
ata
E
ff
i
ci
en
t
Ma
chi
ne
Learning
workshop at
IC
ML 2016
,
2016
.
[16]
H
.
Valpol
a
,
“
From
neur
al
PC
A
to
dee
p
unsuper
vised
le
arn
ing
,”
i
n
Adv
anc
es
in
I
ndepe
nden
t
Component
Anal
ysi
s
and
Learning
M
achi
nes
,
p
p.
143
-
1
71
,
2015
.
[17]
L
.
Maa
lø
e,
e
t
al
.
,
“
Auxili
ar
y
D
ee
p
Gene
ra
ti
v
e
Models,
”
Proceedi
ngs
of
the
33rd
Inte
rnational
Confe
renc
e
on
Mac
hine Learni
ng
,
vol
.
48
,
pp
.
1
445
-
1453,
2016
.
[18]
R.
Jensen
and
Q.Shen,
“
New App
roa
che
s to
Fuzz
y
-
Rough
Feat
ure
Sele
c
ti
on
,
”
IEEE
Tr
an
sacti
ons
on
FuzzySy
stems
,
vol.
17
,
no
.
4
,
pp
.
824
-
838
,
2009
.
[19]
K
.
Sim
ony
an
a
nd
A
.
Zi
ss
erman
,
“
Ver
y
d
ee
p
convol
uti
on
al
net
works
fo
r
large
sca
l
e
imag
e
rec
ogn
it
ion
,”
arXiv
:
1409
.
155
6
,
2014
.
[20]
C
.
Szeg
ed
y
,
e
t
al
.,
“
Going
dee
p
er
with convol
ut
io
ns
,”
arXi
v:
1409
.
4842
,
2014.
[21]
A
.
Krizh
evsk
y
,
et
al
.
,
“
Im
age
n
et
cl
assifi
ca
t
ion
with
de
e
p
convo
lut
ional
n
eur
al
n
et
works
,
”
i
n
A
d
vanc
es
in
neural
inf
orm
ati
on
proc
essing
systems
, p
p.
1097
-
1105
,
2
012.
[22]
A
.
Karpa
th
y
,
et
al
.,
“
La
rg
e
-
sca
le
vide
o
cl
assifi
cat
ion
with
convolutiona
l
n
eur
al
n
e
tworks,
”
i
n
IEEE
Confe
renc
e
on
Computer
Vi
sion
and
Pa
tt
ern
Rec
ognit
ion
(
CVPR)
,
pp
.
1725
-
1732
,
2014.
[23]
J
.
Zhu
,
e
t
a
l
.
,
“
Multi
-
class
ada
b
oost
,”
S
tat
ist
ic
s
and
Its
Int
erface
,
vol
.
2
,
pp
.
349
-
36
0,
2009
.
[24]
N
.
Kari
ana
kis,
e
t
al
.
,
“
Boosting
c
onvolut
ional
f
ea
t
u
res
for
robust
o
bje
c
t
proposals
,
”
arXiv
:
1503.
063
50
,
2015
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
72
-
48
80
4880
[25]
M
.
J.
Saber
ia
n
and
N
.
Vasconc
el
os,
“
Multi
class
Boosting:
The
or
y
and
Al
gorit
hm
s,
”
Advance
s
in
Neural
Information
Pro
ce
ss
ing
Syst
em
,
2011
.
[26]
Qiao,
Yu
,
“
T
he
MN
IST
D
at
aba
s
e
of
h
andwri
t
te
n
digi
ts,
”
Re
trieve
d
18
Augus
t
201
3.
[27]
J.
C.
Pl
at
t
,
“
Us
ing
ana
l
y
t
i
c
QP
and
sparsene
ss
t
o
spee
d
tr
ai
n
ing
of
support
ve
ctor
m
ac
hine
s,
”
A
dvanc
es
in
Neur
al
Information
Pro
ce
ss
ing
Syst
ems
,
pp.
557
-
563
,
1
9
99
.
[28]
T
.
A
.
As
segi
e
a
nd
P
.
S
.
Nair
,
“
Handwrit
te
n
d
ig
it
s
rec
ogn
it
ion
with
decision
tr
e
e
class
ifi
c
at
ion
:
a
m
ac
hin
e
l
ea
rni
ng
appr
oac
h
,
”
In
te
r
nati
onal
Journal
of
El
e
ct
ri
cal
an
d
Computer
Engi
nee
ring
(
IJE
CE
)
,
v
ol
.
9,
n
o
.
5,
pp.
4446
-
4451
,
2019
.
[29]
Y.
Le
Cun
,
et al
.
,
“
MN
IST
handwri
tt
en
dig
it
d
atab
ase
,
”
[Onlin
e]
.
Avail
ab
le
:
htt
p
:/
/
y
an
n
.
l
ec
un
.
com/
exdb/
m
nist/
.
[30]
K
.
Kum
ar
N
.
,
e
t
al
.
,
“
An
ense
m
ble
m
ult
i
-
m
odel
t
ec
hniqu
e
for
pre
dicting
chr
o
nic
kidn
e
y
diseas
e,
”
Int
ernati
on
al
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
(
IJE
CE)
,
v
ol
.
9,
n
o
.
2
,
pp
.
1321
-
1326,
2019
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Hasan
As
il
is
a
l
ec
tur
er
in
th
e
Facu
lty
computer
of
Islamic
Aza
d
Univer
si
t
y,
Aza
rsh
ahr
Branc
h.
He
is
al
so
PH
D
Student
a
t
Urm
ia
Univer
sit
y
.
His
rese
ar
ch
Fiel
d
s
are
Quer
y
Optimiza
ti
o
n
in
Data
base
,
Soft
ware
Deve
lopm
ent
Methodol
og
ie
s,
Dat
a
Minin
g,
Mac
hin
e
Le
arn
ing
.
J
amsh
id
B.
Mohasefi
rec
ei
v
ed
his
B.
S.
in
computer
engi
n
ee
r
in
g
from
Sharif
U
nive
rsit
y
of
Te
chno
log
y
in
Ira
n
at
1996
a
nd
M.S.
in
computer
engi
ne
e
ring
from
Ta
rbia
t
Modare
s
Univer
sit
y
in
Ira
n
at
1999.
He
got
his
PhD
in
c
om
pute
r
engi
ne
e
ring
from
India
n
Instit
ute
of
Te
chno
log
y
Del
hi
(IIT
D)
in
Ind
ia
a
t
2006.
H
e
i
s
cur
ren
tly
associate
prof
essor
and
dea
n
of
computer
and
e
le
c
tri
c
al
engi
n
eering
fa
cul
t
y
a
t
Urm
ia
Univer
sit
y
,
Ir
an.
He
is
a
lso
hea
d
of
Com
pute
r
Emerge
nc
y
Response
Te
am
(CER
T)
ce
nt
er
a
t
Urm
ia
Univer
sit
y
,
I
ran
.
His
cur
ren
t
rese
arc
h
in
te
rest
s inc
lud
e
n
et
wor
k
sec
uri
t
y
,
m
ac
h
ine
le
arn
ing, a
nd
interne
t
of thing
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.