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.
9
, No
.
5
,
Octo
ber
201
9
, pp.
4446
~
44
51
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
9
i
5
.
pp4446
-
44
51
4446
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Handw
ritten di
gits re
cogn
ition wi
th decisi
on tree
classifi
cat
ion:
a machi
ne lea
rning appr
oach
Tseha
y Adm
assu
Asse
gie, P
ram
od
S
ekh
ar
an
N
air
Depa
rtment
o
f
C
om
puti
ng
Techn
olog
y
,
Co
ll
eg
e
o
f
Engi
n
ee
ring
an
d
Technol
og
y
,
A
ksum
Univer
sit
y
,
E
thi
opi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
2
9
, 2
01
9
Re
vised
A
pr
22
, 2
01
9
Accepte
d
Ma
y
4
, 2
01
9
Handwrit
te
n
di
g
it
s
rec
ogni
ti
on
i
s
an
are
a
of
m
a
chi
ne
l
ea
rn
ing,
i
n
which
a
m
ac
hine
is
tr
ai
n
ed
to
ide
n
ti
f
y
h
andwri
tten
digit
s.
One
m
et
hod
of
ac
hi
evi
ng
thi
s
is
with
d
ec
is
ion
tree
c
la
ss
ifi
c
at
ion
m
odel.
A
d
ec
ision
tree
class
ifi
c
at
ion
is
a
m
ac
hine
l
ea
rn
ing
appr
oac
h
th
at
uses
the
pre
d
efi
ned
l
abels
from
the
past
known
sets
to
det
ermine
or
pr
ed
ic
t
th
e
class
es
of
the
future
d
at
a
sets
where
the
class
la
bel
s
a
re
unknown.
In
t
his
pape
r
we
ha
ve
used
the
stan
dar
d
kaggle
digi
ts
dat
ase
t
for
rec
ognition
of
handwri
tt
en
digi
ts
using
a
dec
ision
tr
ee
cl
assifi
ca
t
ion
ap
proa
ch.
And
we
have
eva
lu
at
ed
the
ac
cur
acy
of
the
m
odel
aga
inst
each
d
igit fr
om
0
to
9
.
Ke
yw
or
d
s
:
Decisi
on tree
c
la
ssific
at
ion
Hand
wr
it
te
n di
gits
Hand
wr
it
te
n di
gits rec
ogniti
on
Ma
chine
le
a
rn
i
ng
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Tsehay
Ad
m
assu Asse
gie,
Dep
a
rtm
ent o
f C
om
pu
ti
ng
T
e
chnolo
gy,
Ak
s
um
U
ni
versi
ty
,
PO
B,
1010,
Et
hiopia
Em
a
il
:
tse
haya
dm
assu2
00
6@gm
ail.co
m
1.
INTROD
U
CTION
Hand
wr
it
te
n
dig
it
s
rec
ogniti
on
is
the
proces
s
of
ide
ntifyi
ng
dig
it
s
wr
it
te
n
by
hum
an
us
in
g
a
m
achin
e
le
arn
in
g
a
ppr
oa
ch
[1,
2,
3,
4,
5,
6,
7,
8
,
9
]
.
Ma
chine
t
hat
perform
s
the
r
ecognit
ion
ta
s
k
has
t
o
be
tra
ined
s
o
that
the
m
achine
ca
n
perform
the
ta
s
k
base
d
on
the
t
rainin
g
giv
e
n
with
th
e p
ast
d
at
a
set
s w
it
h
the
kn
own
cl
as
s
la
bels.
Ma
c
hine
le
arn
i
ng
m
od
el
s
s
uch
as,
a
rtific
ia
l
neural
netw
ork
,
c
onvo
l
ution
al
ne
ural
net
work,
de
ci
sion
tree
cl
assifi
er
a
nd
im
pr
ove
d
c
hain
c
od
e h
ist
ogram
featur
e
[
1
, 2
, 3
, 4
]
can also
be
a
ppli
ed
to
ha
ndwr
it
te
n
dig
it
s
recog
niti
on
pr
ob
le
m
,
howe
ve
r,
t
he
recog
ni
ti
on
of
ha
ndwr
it
te
n
dig
it
s
is
sti
ll
a
resea
rch
issue
,
num
ero
us
researc
hes
ha
ve
bee
n
car
ried
ou
t
a
nd
no
ne
of
them
has
pro
po
s
ed
a
pe
rf
ect
appr
oach
to
re
cognize
hand
w
ritt
en
dig
it
s.
This
i
m
plies
that
t
he
pro
blem
of
rec
ognizin
g
a
ha
ndw
ritt
en
dig
it
is
an
im
po
rta
nt
researc
h
issu
e
an
d
there a
re
dif
fere
nt app
ro
ac
hes
and alg
or
it
hms p
rop
os
ed
by re
searche
rs f
or
recog
niti
on
s
of
h
a
ndw
ritt
en
di
gits.
The
m
ai
n
pr
ob
lem
s
of
ha
ndw
ritt
en
dig
it
s
re
cogniti
on
us
i
ng
m
achine
le
arn
in
g
ap
proac
he
s
are
li
ste
d
as
f
ollows [
9]
:
1)
Hand
wr
it
in
g
sty
le
s
var
y
base
d
on
the
w
rite
r
-
it
is
us
ually
ver
y
dif
ficult
even
f
or
hum
an
to
rec
ogniz
e
hand
wr
it
te
n di
gits beca
us
e
of
the si
gn
ific
a
nt
d
if
fer
e
nce
of ha
ndw
riti
ng
sty
le
s.
2)
Si
m
il
arity
be
tween
ha
ndw
ritt
en
dig
it
s,
for
exam
ple
six
a
nd
f
our
m
ay
l
ook
t
he
sam
e
dig
it
based
on
th
e
hand
wr
it
in
g
st
yl
e o
f
t
he write
r
sim
i
la
rly
, o
ne
and se
ve
n
m
a
y l
ook
the
sam
e d
igit
.
3)
Ther
e
is
no
pe
rf
ect
m
achine
le
arn
in
g
m
odel
that
is
su
it
able
an
d
ef
fecti
ve
in
recog
nizi
ng
hand
wr
it
te
n
dig
it
s.
Dif
fe
re
nt
researc
he
rs
us
e
dif
fer
e
nt
m
od
el
s
bu
t
no
m
od
el
is
per
f
ect
ly
su
it
ed
for
recog
niti
on
of
hand
wr
it
te
n
di
gits.
This
iss
ue
is
fo
cu
sed
in
t
his
pa
per.
Ef
fici
ency
and
acc
ur
acy
of
pr
opose
d
decisi
on
tr
ee
cl
assifi
cat
ion
m
od
el
is evalu
at
ed.
4)
The
hand
wr
it
te
n
dig
it
s
can
not
be
recog
nized
by
usi
ng
fi
xed
i
m
age
recog
niti
on
patte
rn
s
,
be
cause
th
ey
ar
e
diff
e
re
nt for di
ff
e
ren
t
wr
it
ers
.
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
Han
dwrit
te
n
di
gits reco
gnit
io
n
wi
th
decisi
on
tree cl
as
sif
ic
at
ion:
a ma
c
hin
e
…
(
Tsehay A
dmas
su Asse
gie
)
4447
The
sta
ndar
d
data
set
s
li
ke
kaggle
an
d
MI
NS
T
ha
ve
sim
plifie
d
the
ha
ndw
ritt
en
dig
it
s
reco
gnit
io
n
pro
blem
[
10,
11
]
.
Th
os
e
datas
et
s
con
ta
ins
la
r
ge
num
ber
of
s
a
m
ple
i
m
ages
of
hand
wr
it
te
n
dig
it
s
wh
ic
h
c
an
be
us
e
d
in
m
achine
le
ar
ning
m
od
el
s
li
ke
deci
sion
tree
cl
ass
ific
at
ion
with
out
pre
processi
ng
an
d
the
ef
f
or
t
of
pr
e
par
i
ng
dataset
is
av
oid
e
d.
But
sti
ll
so
m
e
of
the
quest
io
ns
rem
ai
ned
unanswere
d
a
nd
will
co
ntinu
e
to
e
xist
in
the
feat
ur
e,
as
there
is
no
perfect
m
od
el
us
e
d
in
m
achine
le
arn
in
g
in
re
c
ogniti
on
of
ha
ndwr
it
te
n
dig
it
s.
Th
e
pro
blem
s
add
ressed
in
this
pap
e
r
are
,
ho
w
ef
fecti
ve
a
decisi
on
tree
cl
assifi
cat
ion
in
ha
ndw
ritt
en
dig
it
s
recog
niti
on
.
A
nd
t
he
im
ple
m
entat
ion
an
d
accuracy
of
decisi
on
tree
cl
assifi
cat
ion
in
ha
ndw
ritt
en
dig
it
s
recog
niti
on
is
evaluate
d usin
g pyt
hon p
rogra
m
m
ing
la
ngua
ge.
Decisi
on
tre
e c
lassif
ic
ation
A
decisi
on
tre
e
cl
assifi
er
is
a
m
achine
le
arni
ng
m
od
el
w
id
el
y
us
ed
to
s
olv
e
cl
assifi
cat
io
n
pro
blem
s,
li
ke
hand
wr
it
te
n
di
gits
recog
ni
ti
on
.
A
s
the
nam
e
decisi
on
tr
ee
su
ggest
s
,
th
e
m
od
el
breaki
ng
dow
n
our
da
ta
by
m
aking
decisi
ons
based
on
as
king
a
se
ries
of
quest
ions.
Let
'
s
co
ns
ide
r
the
f
ollow
i
ng
exam
ple
w
her
e
we
use
a
decisi
on
tree
to
decide
up
on
recog
niti
on
of
a
0
an
d
a
1
digi
t.
The
cl
assifi
er
te
sts
al
l
of
the
possible
s
et
s
wi
t
h
dig
it
s
0
a
nd 1
un
ti
l i
t fin
ds a
m
at
ching
dig
it
as sho
wn in Fi
gure
1.
Figure
1
.
D
eci
s
ion
t
ree
cl
assifi
cat
ion
Th
e
s
ucces
s r
at
e o
f decisi
on tree cl
as
sif
ie
r i
n pyth
on
To
te
st
the
su
c
cess
rate
of
a
de
ci
sion
tree
cl
a
ssific
at
ion
in
r
ecognit
ion
of
ha
ndw
ritt
en
dig
i
ts
we
ha
ve
us
e
d
pyt
hon
pro
gr
am
m
ing
,
a
ve
ry
pow
er
ful
la
nguag
e
f
or
i
m
ple
m
entat
i
on
of
m
achine
le
arn
i
ng
al
gor
it
h
m
s,
m
any
of
the
re
searche
rs
are
de
pendin
g
on
py
tho
n
la
n
gua
ge
fo
r
m
achine
le
arn
in
g
al
gori
thm
i
m
ple
m
entat
ion
s.
The
c
od
e
s
how
n
bel
ow
is
us
e
d
to
dr
a
w
the
grap
h
f
or
su
cce
s
s
rate
of
the
cl
assifi
er,
s
how
n
in
Fig
ur
e
2
a
nd
the
entr
op
y,
the
im
purity
f
or t
he d
eci
sion
tre
e cla
ssific
at
ion
is
s
how
n
in
F
ig
ure
3
.
Figure
2
.
Pyt
hon co
de f
or
m
od
el
ing t
he succ
ess r
at
e
of d
eci
sion t
ree classi
ficat
ion
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.
9
, N
o.
5
,
Oct
ober
201
9
:
4
4
4
6
-
4
4
5
1
4448
Figure
3
.
S
ucc
ess
rate
of d
eci
sion t
ree classi
fier
1.
LIT
ERATUR
E REVIE
W
In
this
sect
io
n,
the
previ
ous
works
relat
ed
to
ha
nd
wr
it
te
n
di
git
s
recogn
it
io
n
will
be
rev
ie
w
ed
.
Th
e
st
ud
ie
s
are
disc
us
se
d
in
the
fo
ll
owin
g
s
ect
ion
s.
Ha
nd
wr
it
te
n
dig
it
s
r
ecognit
ion
ca
n
be
im
ple
m
e
nted
with
scan
ning
the
dig
it
al
i
m
ages
pix
el
s
a
nd
e
xtracti
ng
the
f
eat
ur
es
us
i
ng
a
ne
ur
al
netw
ork
as
cl
assifi
er
s
f
or
hand
wr
it
te
n di
gits rec
ogniti
on
[1
]
.
Hand
wr
it
te
n
di
gits
recog
niti
on
is
beco
m
ing
a
great
er
de
m
and
as
one
of
the
c
om
pu
te
r
vision
[
2]
te
chn
iq
ues
.
Co
m
pu
te
r
vision
has
m
any
app
li
cat
ion
s
f
or
int
erf
aci
ng
hu
m
an
an
d
a
m
achi
ne.
T
he
f
ocu
s
of
the
auth
or
wa
s
ge
tt
ing
bette
r
accuracy
an
d
pe
rfor
m
ance
requirem
ent
with
convo
l
ution
al
neu
ral
net
work
s
f
or
cl
assifi
cat
ion
.
A
blob
base
d
cl
assifi
cat
ion
of
im
age
is
ano
t
her
al
go
rithm
us
ed
to
so
lve
hand
w
ritt
en
dig
it
s
recog
niti
on
prob
le
m
s.
The
al
gorithm
reco
gn
iz
es
dig
it
s
by
cl
ass
ify
ing
them
as
blo
bs
wit
h
an
d
without
stem
[3
]
.
In
t
his
m
e
tho
d
t
he
m
at
he
m
atical
m
or
phology
was
use
d
to
c
on
st
ru
ct
the
cl
assifi
cat
ion
m
od
el
s
.
This
is
a
te
ch
nique
of
rec
og
niti
on
of
dig
it
s
by
i
den
ti
fica
ti
on
of
bl
ob
s
and
ste
m
s.
Th
e
pro
blem
wit
h
this
m
et
ho
d
is
that
it
is
no
t a
ble to
r
ec
ognize the
bro
ken d
i
gits o
f
la
r
ge gap
and
d
igit
s
with e
xt
ra s
te
m
s.
Mult
iple
pe
rc
eption
la
ye
rs
with
neural
ne
twork
i
s
an
ot
her
m
od
el
use
d
in
ha
ndw
r
it
te
n
dig
it
s
recog
niton
[
4].
In
t
heir
st
ud
y,
the
aut
hors
use
d
MN
IS
T
dat
aset
for
ha
ndw
ritt
en
dig
it
s
rec
ogniti
on,
the
da
ta
set
was
trai
ned
with
gr
a
dient
desce
nt
ba
ck
-
pro
pa
gation
a
lgorit
hm
and
the
ve
rified
with
the
fee
d
-
f
orward
al
gorithm
.
A
Decisi
on
tre
e
cl
assifi
cat
ion
is
on
e
of
the
si
m
plest
m
achine
le
arn
in
g
m
od
el
use
d
in
ha
ndw
ritt
en
dig
it
s
recog
niti
on
[
5]
an
d
a
com
par
at
ive
stu
dy
perform
ed
on
decisi
on
tree
cl
assifi
cat
ion
and
ra
ndom
fo
rests
.
Accor
ding
to
a
uthors,
t
he
r
an
do
m
forest
perform
ed
superi
or
tha
n
decisi
on
tree
cl
assifi
cat
ion
.
O
ne
of
t
he
m
os
t
com
m
on
ly
us
ed
m
achine
le
arn
in
g
m
od
el
in
hand
wr
it
te
n
dig
it
s
identific
at
ion
is
K
NN
(K
-
Nea
rest
Nei
ghbo
r
)
[6
]
. In
t
his stu
dy
MIN
ST
data
set
w
as used
for
te
sti
ng
t
he
al
gorithm
, 5
0,0
00 im
ages w
ere
us
e
d
f
or
trai
ning and
550
im
ages
fo
r
te
sti
ng
the
al
gorithm
.
Accord
i
ng
to
a
uthor
s,
the
pr
eci
si
on
le
vel
of
the
syst
e
m
is
70
.
9
%
and
the accu
racy
of the syst
em
w
ill i
m
pr
ove
base
d on the
qual
it
y of dataset
.
A
num
ber
of
m
achine
le
arn
i
ng
m
od
el
s,
s
uc
h
as
K
N
N,
K
Near
est
cl
assi
ficat
ion
,
Ar
ti
fi
ci
al
Neu
ral
Netw
ork,
ANN
Ba
ye
sia
n
Cl
assifi
cat
ion
and
Decisi
on
tree
cl
assifi
cat
ion
ca
n
al
so
be
us
e
d
to
id
entify
hand
wr
it
te
n
di
gits
as
m
achines
le
arn
in
g
te
chn
i
qu
e
s.
The
s
tud
ie
s
that
we
hav
e
perform
e
d
on
accu
racy
of
eac
h
of
th
e
trai
ning
m
od
el
s
dif
fers,
the
A
NN
[7
,
12
,
14
]
is
bette
r
perform
ed
al
tho
ug
h
it
has
it
s
ow
n
dr
a
w
backs.
The
accu
racy
of
the
m
achine
le
arn
in
g
m
od
el
s
are
al
so
de
pends
on
the
qual
it
y
of
the
im
age
to
be
pr
edict
ed
and
t
he
a
vaila
bi
li
t
y
of
the
in
fe
rior
it
y
of
di
git
i
m
ages
in
the
trai
ning
set
s.
C
om
pu
ta
ti
on
al
e
ff
ic
ie
ncy
is
an
oth
e
r
par
am
et
er to
se
le
ct
b
et
te
r
m
od
el
an
d t
he K
N
N
is
bette
r
in
e
ff
ic
ie
ncy c
om
par
ed
w
it
h ot
her
al
gorithm
s [
13
]
.
2.
DA
T
AS
ET
A
ND FEAT
URE
SELE
CTIO
N
2.1.
Kagg
le
dat
as
e
t
In
t
his
pa
pe
r,
we
hav
e
us
e
d
kaggle
dat
aset
.
The
ka
gg
le
da
ta
set
has
sam
ple
ha
ndw
ritt
en
dig
it
s
f
or
evaluati
ng
m
a
chine
le
ar
ning
m
od
el
s
on
th
e
hand
wr
it
te
n
dig
it
rec
ogniti
on
pro
blem
.
It
has
been
c
omm
on
ly
us
e
d
in
resea
r
ch
to
im
ple
m
e
nt
ha
ndw
ritt
en
dig
it
re
co
gnit
ion
syst
em
s.
The
kaggle
dataset
co
ntain
s
21,
000
trai
ning
an
d
21,
000
te
sti
ng
of
ha
ndw
ritt
en
dig
it
s
(
0
to
9)
.
Each
di
git
is
sta
nd
a
rd
iz
e
d
a
nd
ca
ntere
d
in
a
gr
ay
-
scal
e
(0
-
255)
i
m
age
with
si
ze
28
×
28
pixe
l.
Each
im
age
consi
sts
of
78
4
pi
xels
that
re
pr
ese
nt
the
st
r
uctu
res
of the
dig
it
s. S
om
e o
f
the
exa
m
ples f
ro
m
the k
a
gg
le
dataset
are
s
how
n
i
n Fi
gure
4.
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
Han
dwrit
te
n
di
gits reco
gnit
io
n
wi
th
decisi
on
tree cl
as
sif
ic
at
ion:
a ma
c
hin
e
…
(
Tsehay A
dmas
su Asse
gie
)
4449
Figure
4
.
Ex
a
m
ples o
f
ka
ggle
d
at
aset
2.2.
Digits re
co
gn
i
tion
Decisi
on
tree
c
la
ssific
at
ion
m
od
el
is
us
e
d
to
trai
n
4200
0
dat
aset
;
by
sp
li
tt
i
ng,
half
of
t
he
dataset
s
for
the
trai
ni
ng
se
t
and
the
rem
ai
nin
g
21
000
data
set
s
for
t
est
ing
.
W
e
ha
ve
fo
ll
owe
d
th
e
f
ollow
i
ng
ste
ps
t
o
perform
the
classificat
ion
on
kaggle
ha
ndw
r
it
t
en
dig
it
s
dataset
and
the
pse
udo
c
od
e
of
t
he
ha
ndw
ritt
en
dig
it
s
recog
niti
on
is
s
how
n
in
F
ig
ure
5
.
a.
Loa
d
the
k
a
ggle
h
a
ndw
ritt
en digit
s
dataset
s
for
cl
assifi
cat
ion.
b.
Sp
li
t t
he data
s
et
s into
t
wo sets, one
for t
he
tr
ai
nin
g an
d
t
he rest
for t
est
ing.
c.
Learn
i
ng
a
nd
pr
e
dicti
ng,
in
t
he
dig
it
s
datas
et
s,
the
rec
ogni
zer
is
trai
ne
d
to
pre
dict,
gi
ve
n
a
n
im
age
of
a
hand
wr
it
te
n
di
git.
W
e
ha
ve
us
e
d
the
s
am
ples
of
each
of
the
10
possibl
e
cl
asses
of
ha
ndw
ritt
en
di
gi
t
s
from
kag
gle
da
ta
set
(the
dig
it
s
zero
to
nin
e
)
on
w
hich
we
fi
t
an
est
i
m
at
or
to
predict
the
cl
asses
to
w
hich
un
s
een
sam
ple
s b
el
ong
t
o.
d.
Test
ing
t
he
acc
ur
acy
of the cl
assifi
er.
Figure
5
.
Pse
udo
co
de o
f digi
ts reco
gnit
ion
3.
E
X
PERI
MEN
TAL RES
UL
TS
The
ex
per
im
e
nt
was
co
nduc
te
d
on
hand
wr
it
te
n
di
gits
of
the
sta
ndar
d
ka
gg
le
data
set
us
ing
th
e
decisi
on
tree
cl
assifi
er
for
tra
ining
the
m
achine.
In
this
da
ta
set
i
m
ages
wer
e
pr
e
proce
ssed
so
i
n
this
pap
e
r
pr
e
processi
ng
was
no
t
i
nc
lud
e
d.
T
his
da
ta
set
con
ta
in
s
42,
000
im
a
ge
s
of
s
am
pl
e
ha
ndwr
it
te
n
dig
it
s
.
F
or
m
these
im
ages;
we
ha
ve
us
e
d
21,
000
in
trai
ning
a
nd
the
rest
21,00
0
im
ages
in
te
sti
ng
.
W
e
te
ste
d
t
he
decisi
on
tree
c
la
ssific
at
ion
m
od
el
on
al
l
of
the
di
gits
(
0
-
9)
.
Ta
ble
1,
s
ho
ws
t
he
te
st
r
es
ults
after
trai
ni
ng
the
m
achine
on
th
is
dataset
.
In
Table
1,
it
is
s
how
n
that
so
m
e
of
the
dig
it
s
are
not
recogni
zed
by
the
m
achi
ne
le
arn
in
g
m
od
el
,
the
decisi
on
t
ree
cl
assifi
er.
Fo
r
e
xam
ple
8
wer
e
pr
e
dicte
d
as
5
a
nd
the
reason
beh
i
nd
this
is
the
sim
il
arity
of
hand
wr
it
in
g
sty
le
between
5
an
d
8.
I
n
Ta
ble
1,
t
he
sam
ple
ha
ndwr
it
te
n
di
gits
recog
ni
zed
by
the
decisi
on
tr
ee
cl
assifi
er
is
sh
ow
n
an
d
th
e
accuracy
of
the
cl
assifi
er
in
rec
ognizin
g
each
di
git
(0
t
o
9)
i
s
sh
ow
n
in
Ta
ble 2
.
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.
9
, N
o.
5
,
Oct
ober
201
9
:
4
4
4
6
-
4
4
5
1
4450
Table
1.
T
he
w
or
est
rec
ognize
d digit
s li
sts
Han
d
written Dig
it
s
i
m
ag
es
/Test
Data
Sy
ste
m
Digit
Predi
ctio
n
Actu
al E
x
p
ected P
redictio
n
Co
rr
ectly
Pr
ed
icte
d
0
Co
rr
ectly
Pr
ed
icte
d
0
Co
rr
ectly
Pr
ed
icte
d
0
W
ron
g
ly
P
redicted
0
W
ron
g
ly
P
redicted
6
Co
rr
ectly
Pr
ed
icte
d
6
Co
rr
ectly
Pr
ed
icte
d
6
Co
rr
ectly
Pr
ed
icte
d
6
Table
2.
Acc
uracy
of the
deci
sion t
ree classi
fier in ha
ndw
ri
tt
en
dig
it
s
rec
ogniti
on
Han
d
W
ritten
Dig
its(0
-
9)
Accurac
y
of
the Cl
ass
if
er
(10
0
%)
0
8
3
.56
1
9
3
.73
2
8
3
.69
3
8
3
.73
4
8
3
.8
1
5
8
3
.65
6
8
3
.47
7
8
3
.81
8
8
4
.12
9
8
3
.75
Figure
6,
sho
ws
the
acc
ur
a
cy
of
t
he
deci
sion
t
ree
cl
ass
ifie
r
(i
n
per
ce
ntage
)
in
rec
ognizin
g
eac
h
hand
wr
it
te
n
di
git.
A
nd
the
di
git
1
is
recog
ni
zed
with
bette
r
accu
racy
by
the
cl
assifi
er,
8
is
recog
nized
with
le
ast
accuracy
and
t
his
is
bec
ause
the
di
git
8
resem
bles
wi
th
m
any
oth
er
dig
it
s
f
or
e
xam
ple
de
pe
nd
i
ng
on
t
he
hand
w
riti
ng
sty
le
,
it
m
ay
l
ook
fi
ve.
T
his
sh
ows
that,
t
he
dif
fer
e
nce
in
ha
ndwr
it
in
g
sty
le
hav
e
a
gr
eat
influ
e
nce
on th
e accu
racy o
f
a
m
achine learni
ng
m
od
el
s
us
e
d
in
h
a
ndw
ritt
en digit
s r
ec
ogni
ti
on
.
Figure
6
.
The
a
ccur
acy
of d
e
ci
sion t
ree classi
fier
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
Han
dwrit
te
n
di
gits reco
gnit
io
n
wi
th
decisi
on
tree cl
as
sif
ic
at
ion:
a ma
c
hin
e
…
(
Tsehay A
dmas
su Asse
gie
)
4451
4.
CONCL
US
I
O
N
In
t
his
pa
per,
we
ha
ve
us
e
d
a
decisi
on
tree
c
la
ssific
at
ion
m
od
el
for
m
achine
le
ar
ning
tha
t
recogn
iz
es
hand
wr
it
te
n
di
gits.
The
m
ac
hin
e
w
as
trai
ne
d
with
a
ka
ggle
dataset
w
hi
ch
co
ntains
4200
0
r
ow
s
a
nd
72
0
colum
ns
an
d
r
esult
sh
ows
83.4
%
acc
ur
acy
.
W
e
ha
ve
use
d
dig
it
i
m
ages
pix
el
s
as
featu
re
s
vecto
r
an
d
D
eci
sion
Tree
as
cl
assif
ie
rs
for
hand
w
ritt
en
dig
it
s
re
cogniti
on.
W
e
hav
e
us
e
d
the
Kaggle
re
posit
or
y
f
or
t
rainin
g
a
nd
te
sti
ng
datase
t
s,
the
e
xperim
ent
res
ult
sho
ws
that
t
he
de
ci
sion
tree
cl
assifi
er
is
ef
fec
ti
ve
in
rec
ogni
ti
on
of
hand
wr
it
te
n di
gits
.
REFERE
NCE
S
[1]
K
h
Tohi
dul
Isl
a
m
,
Ghulam
Mujta
ba,
“
Handwr
it
te
n
Digit
s
Re
cog
nit
ion
wi
th
Art
if
i
ci
al
N
eural
Ne
t
work
,”
Proc.
o
f
t
he
Inte
rnational
Co
nfe
renc
e
on
Eng
ine
ering
Technologies
and
Te
chn
oprene
urs
hip
(
ICE2T 201
7)
,
Sep
2017
.
[2]
Haide
r
A.
Alwz
waz
y
,
Ha
y
d
er
M.
Albeha
d
il
i
,
Younes
S.
Alwan,
Na
z
E
.
Isl
a
m
.
“
Handwrit
ten
Digit
R
ec
ogni
ti
on
Us
ing
Convolut
iona
l
Neur
al
Networks,”
Int
ernati
onal
Jour
nal
of
Inn
ovative
R
ese
arch
i
n
Computer
a
n
d
Comm
unic
ati
on
Engi
ne
ering
(
An
ISO 3297:
2007
Certi
f
ie
d
Or
ganizati
on)
,
v
ol
.
4
,
no.
2
,
Feb
2016
.
[3]
Vijay
a
Kum
ar,
A
Siriki
rshna,
R
ave
endr
a
Babu
and
Radhi
ka
M
ani
,
“
Cla
ss
ifica
t
i
on
and
rec
ogni
tion
of
handwri
tten
digi
ts b
y
using
m
at
hemati
c
al m
orpholog
y
,
”
Indi
an
Ac
ad
emy
o
f
S
ci
en
ce
s
,
2010
[4]
Ism
ai
l
M.
Kesh
ta
.
“
Handwrit
ten
Digit
R
ec
ogn
it
ion
b
ase
d
on
Output
Inde
p
en
dent
Mult
i
-
L
a
yer
Perc
ept
ron’s,
”
(
IJA
CSA)
Inte
rnational
Journal of
Ad
vanc
ed
Co
mputer
Scienc
e and A
ppl
ic
at
ions,
vol
.
8
,
n
o
.
6
,
20
17
.
[5]
La
van
y
a
K,
Sha
ur
y
a
Bajaj
,
Prat
ee
k
T
ank
Shash
wat
Jain
,
“
Handwrit
te
n
Dig
it
R
ec
ogni
ti
on
using
Hoeffdi
ng
tr
e
e
,
Dec
ision
tr
ee
a
nd
Random
fore
sts
Com
par
at
iv
e
Approac
h
,
”
2
017
Inte
rnation
al
Confe
renc
e
on
Computati
on
al
Inte
lligen
ce i
n
D
ata
Sc
ie
nc
e(
ICCIDS)
,
201
7.
[6]
Nurul
Ilmi,
T
jo
korda
Agung
Budi
W
,
Kurnia
wan
Nur
R,
“
Handwr
it
ing
Digi
t
Recogni
t
ion
u
sing
Local
Bi
na
ry
Pat
te
rn
Variance
and
K
-
Nearest
Nei
ghbor
Classifi
ca
ti
on
,”
I
EE
E
,
Fourth
Inte
rnat
ional
Confe
r
ence
on
Informatio
n
and
Comm
unic
a
ti
on
Te
chnol
ogi
e
s (
ICoICT
)
,
2016
[7]
Stuti
As
tha
na
1,
Am
it
kant
Pandit,
Dinesh
Go
y
a
,
“
A
Li
te
rat
ur
e
Surve
y
On
Be
tt
er
S
y
stem
Eff
icien
c
y
Of
Handwrit
t
en
Num
era
l
Rec
og
nit
ion,”
Suresh
Gyan
Vi
har
Unive
rs
it
y
Int
ernati
onal
Journ
al
of
Env
ironm
ent
,
Sc
ie
n
ce
a
nd
Technol
ogy
,
vo
l. 3, no.
1,
pp.
1
-
5
,
Jan
2017
.
[8]
Shengfe
ng
Chen
,
Rabia
Alm
amlook,
Yuw
en
Gu,
Dr.
L
ee
well
s,
“
Offl
ine
Hand
writt
en
Dig
it
s
R
ec
ogni
ti
on
Us
in
g
Mac
hine
l
earni
ng
,”
Proc
ee
din
gs
of
the
Inte
r
nati
onal
Conf
ere
nce
on
Indust
rial
Engi
ne
erin
g
and
Operations
Manage
ment
Wa
shington
DC
,
U
SA
,
Sep
27
-
29,
2018.
[9]
Anuj
Dutt,
A
ashiDutt
,
“
Handwri
tt
en
Dig
it
R
ec
og
nit
ion
Us
ing
De
ep
L
ea
rning
,
”
In
te
rnational
Jour
nal
of
Adv
anc
ed
Re
search in Co
mputer
Engi
n
ee
r
ing
&
Technol
ogy,
v
ol
.
6,
no.
7,
Jul 20
17
.
[10]
Pooja
Yada
v
,
N
idhi
ka
Yad
av,
“
Handwrit
ing
Re
cogni
ti
on
S
y
st
e
m
-
A
Revi
ew,
”
Inte
rnational
Jo
urnal
of
Computer
Appl
ic
a
ti
ons
,
v
ol
.
114
, n
o
.
19
,
pp
.
46
-
40
,
Mar
201
5.
[11]
Aree
j
Alsa
afi
n,
As
hra
f
El
nag
ar,
“
A
Minim
al
S
ubset
of
Feat
ur
e
s
Us
ing
Feat
ure
Sele
c
ti
on
for
Handwri
tt
en
Di
git
Rec
ognition,
”
Jo
urnal
of
In
te
l
li
g
e
nt
Learning
S
yst
ems and
Applica
ti
ons
,
vo
l.
9,
pp.
55
-
68,
2017
.
[12]
J.
J.W
ang,
S.G.
Hu1,
X.T.
Zhan,
Q.Yu1,
Z.
Liu,
T
.
P.
Chen
,
Y.Yin,
Sum
io
Hos
aka
,
Y.
Li
u,
"
Handwrit
te
n
-
Digi
t
Rec
ognition
b
y
H
y
brid
Convol
uti
onal
Neura
l
Network
base
d
on
HfO
2
Mem
risti
ve
Spik
ing
-
Neuron,
"
Sc
ei
nt
if
i
Re
ports
,
Aug
20
18
[13]
Stuti
As
tha
na
,
Am
it
kant
Pandi
t,
Dinesh
Go
y
a
l,
"
A
Li
te
r
at
ur
e
Surve
y
On
Bet
t
er
S
y
stem
Eff
ic
i
ency
Of
Handwrit
ten
Num
era
l
Rec
og
nit
ion
,
"
Suresh
Gyan
Vi
har
Unive
rs
it
y
Inte
r
nati
o
nal
Journal
of
Env
iron
ment,
S
ci
en
ce
and
Technol
ogy
,
v
ol
.
3,
no.
1,
Jan
20
17
[14]
Raul
R.
T
iwar
i
,
A
par
navi
s
H
wana
,
D
hana
shre
e
W
adhone
,
"
H
an
dwritt
en
dig
it
s
rec
ogni
ti
on
using
bac
k
propa
ga
tion
nuer
al
n
et
work
and
KN
N
cl
assifie
r
,
"
In
te
rnatio
nal
Journal
of
El
e
ct
rica
l,
El
e
ctr
onic
s
and
Data
Comm
unic
ati
o
n
,
vol.
1
,
2013
.
Evaluation Warning : The document was created with Spire.PDF for Python.