Int
ern
at
i
onal
Journ
al of
El
e
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
5
,
Octo
ber
201
9
, pp.
3798
~
38
03
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
pp3798
-
38
03
3798
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Identific
atio
n
of i
nd
ividu
aliza
ti
on techni
qu
es for
cri
min
al
records
in sanctio
n lists
Go
nz
alo M
. A
ri
as
1
,
P
ab
l
o A
.
Pel
áez
2
, Fred
y
E.
H
oyos
3
1
,2
Te
cno
lógi
co
d
e
Antioqu
ia,
Inst
it
uci
ó
n
Unive
rsi
t
ari
a
,
Fa
cul
t
ad
d
e
Inge
ni
ería,
Colo
m
bia
.
3
Univer
sidad
Na
ci
ona
l
de
Colom
bia
-
Sede
Mede
l
lí
n,
Facultad
de
Cie
ncias,
Escuel
a
de
Fís
ica, Colo
m
bia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ve
Dec
31
, 201
8
Re
vised
A
pr 9
,
201
9
Accepte
d
Apr
20
, 201
9
Us
ing
eff
icient
sea
rch
ing
techni
ques
on
sanc
ti
o
ns
li
st
s
and
pr
e
ss
art
ic
l
es
al
lows
a
bet
t
er
f
il
te
r
ing
on
indi
v
idua
ls
and
entities
to
esta
bli
sh
a
comm
erc
ia
l
rel
a
ti
onship
wi
t
h,
in
cl
uding
th
ose
who
ar
e
going
to
hav
e
acce
ss
to
conf
ide
n
ti
a
l
info
rm
at
ion
bel
ongi
ng
to
the
compa
n
y
,
in
orde
r
to
m
ini
m
iz
e
the
risk
of
le
ak
a
ge
or
informati
on
m
ism
ana
gement.
Tha
t
pro
ce
ss
of
fil
t
eri
ng
o
n
indi
viduals
or
ent
ities
cou
ld
be
aut
om
ated
b
y
using
ind
iv
idua
liza
ti
on
al
gorit
hm
s,
sea
r
chi
ng
techniqu
e
s
base
d
on
string
compari
sons
,
art
i
ficial
int
ellige
n
ce,
and
fac
i
al
r
ec
ogni
tion.
Diver
se
m
etho
ds
were
exa
m
ine
d
to
b
e
appl
i
ed
on
ea
ch
m
ent
ione
d
tech
nique
in
orde
r
t
o
ide
nti
f
y
whic
h
ones
are
ide
a
l
to
it
s
app
l
ic
a
ti
on
on
indi
v
idua
liza
ti
on
due
to
the
ir
cha
r
ac
t
eri
sti
cs,
i
n
orde
r
to
obt
ai
n
agi
l
e
and
r
el
i
able
result
s;
ta
king
int
o
a
cc
ount
th
at
diff
ere
n
t
m
et
hods
are
complementa
r
y
an
d
not
exc
lusive,
and
tha
t
their
combinat
ion
al
lows
to
m
ini
m
iz
e
hum
an
intera
c
ti
on
in
th
e
cl
assifi
ca
t
ion
of
informati
on,
avoi
ding
anal
y
s
i
s of
irr
el
ev
ant data
for
th
at
p
artic
ula
r
se
arc
h
.
Ke
yw
or
d
s
:
Cri
m
inal reco
r
ds
sec
ond
F
al
se posi
ti
ves
F
il
te
rs
S
ancti
ons li
st
V
erifica
ti
on m
et
hods
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
:
Fr
e
dy E. H
oyos
,
Faculta
d de Ci
encias
-
Esc
uel
a d
e
Físi
ca
,
Un
i
ver
si
dad Na
ci
on
al
de
C
olo
m
bia
-
Se
de M
edell
ín
,
Ca
rr
era
65 N
o. 59A
-
110
, Me
de
ll
ín,
Colom
bia
.
Em
a
il
: feh
oyosve@
unal
.edu.c
o
1.
INTROD
U
CTION
Currentl
y,
relevan
t
inf
orm
ati
on
f
or
a
ny
s
ubj
ect
of
st
ud
y
is
in
a
la
r
ge
nu
m
ber
of
m
e
dia
f
orm
at
s,
wh
ic
h
can b
e pa
ram
et
erized
and
i
nd
e
xe
d
f
or
sp
eci
al
iz
ed
us
e
in
public
or
pr
ivate
databases
[1]
.
T
hey
can b
e
in
natu
ral
la
ngua
ge,
ca
ptured
i
n
i
m
ages,
or
in
any
m
ediu
m
req
ui
red
t
o
facil
it
at
e
it
s
us
e
and
disclos
ur
e
.
U
pdat
in
g
the
in
form
at
io
n
wen
t
f
ro
m
be
ing
in
ha
nds
of
a
few
com
pa
nies
a
nd
m
edia,
to
be
a
vaila
ble
to
any
pe
rs
on
w
ho
has
an
el
ect
r
onic
de
vice
with
acce
ss
to
t
he
In
te
r
net,
ca
usi
ng
s
ources
of
inf
orm
ation
to
prolife
rate
,
bo
t
h
reli
able an
d o
f dubio
us o
rigin.
Takin
g
int
o
ac
count
this
vast
a
m
ou
nt
of
i
nfo
rm
ation
of
al
l
kinds,
c
om
pani
es
hav
e
f
ound
the
nee
d
to
cl
assify
and
in
div
id
ualiz
e
[
2]
it
,
in
orde
r
to
com
ply
with
the
re
gu
la
ti
ons
that
gove
r
n
them
or
to
im
pr
ov
e
internal
sel
ect
ion
processes
of
perso
nnel
,
associat
ed
com
pan
ie
s,
searc
h
f
or
s
olu
ti
ons
pr
act
ic
es,
a
m
on
g
oth
e
r
obj
ect
ives
.
O
ne
of
the
m
ai
n
ne
ed
s
in
c
om
pan
ie
s
at
le
gal
le
vel
is
the
searc
h
of
crim
inal
reco
r
ds
of
natu
r
al
and
le
gal
perso
ns
with
w
hom
they
hav
e
s
om
e
e
m
plo
ym
ent
relat
ion
sh
i
p,
in
order
t
o
m
ini
m
iz
e
the
risk
of
bei
ng
us
e
d
in
m
on
ey
la
un
de
rin
g
an
d
te
rror
ist
fina
ncin
g
operati
ons
(L
AF
T
by
it
s
a
cro
nym
in
Sp
a
nish).
W
it
h
this
pr
em
ise
,
it
is
req
ui
red
t
o
have
the
a
bili
ty
to
validat
e
in
f
orm
at
ion
that
is
releva
nt
to
a
pa
rtic
ular
in
div
i
du
al
i
n
sp
eci
al
iz
ed
da
ta
bases,
as
w
el
l
as
so
ur
ces
of
co
ns
ta
nt
updatin
g
an
d
le
ss
sta
nd
ar
di
zat
ion
su
c
h
a
s
pr
ess
do
c
um
ents
, ar
t
ic
le
s,
natio
nal
and inter
natio
na
l bu
ll
et
ins,
and
oth
e
r on
li
ne
so
urces
.
Usu
al
ly
,
syst
e
m
s
us
ed
f
or
thi
s
pur
po
se
a
pply
te
xt
recogn
it
ion
al
gorithm
s,
com
par
ing
t
he
m
with
their
own
databases
,
w
hich
hav
e
dicti
on
a
ries
of
te
rm
s
si
m
i
la
r
to
th
os
e
us
e
d
in
new
s
co
nce
r
ning
c
rim
inal
act
ivit
ie
s
[3
-
4]
.
The
arisi
ng
pre
m
ise
is
the u
se o
f
these cha
r
act
er r
eco
gn
it
i
on
alg
or
it
hm
s,
com
ple
m
enting
the
m
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
Id
e
ntif
ic
ation
of
ind
iv
id
ua
li
z
ation
tec
hn
i
ques
for crimi
na
l
re
cor
ds
in
s
an
ct
i
on li
sts
(
Gonz
al
o
M.
Arias
)
3799
with
arti
fici
al
intel
li
gen
ce
te
c
hn
i
qu
e
s
that
al
l
ow
us
not
on
ly
to
ide
ntify
the
se
synta
xes
,
but
al
so
hel
p
asse
ssin
g
the
m
os
t
re
li
able
s
ources
ov
er
ti
m
e
and
prov
i
de
resu
lt
s
a
ccordin
g
t
o
w
hat
is
require
d.
At
t
he
sam
e
tim
e
,
visu
al
in
form
a
ti
on
of
in
div
i
dual
s
is
com
pared
with
facial
recog
niti
on
te
c
h
ni
qu
e
s,
e
xp
a
nd
i
ng
or
deli
m
iting
the r
a
nge
of
re
su
lt
s in
ca
ses
wh
e
re t
he
in
for
m
at
ion
in w
ritt
en
m
edia is n
ot correct.
2.
CONCEPT
U
AL F
RAME
WORK
2.1.
Def
ini
ti
on
s
Text
-
sea
rch
al
gorithm
s
:
Text
-
searc
h
al
gorit
hm
s
a
re
te
chn
iqu
e
s
us
e
d
in
orde
r
to
fi
nd
t
he
occurre
nces
of
a
pa
tt
ern
of
cha
racters
in
a
giv
e
n
te
xt
co
rr
e
spo
nd
i
ng
to
a
com
bin
at
ion
of
el
e
m
ents
of
a
de
fine
d
al
ph
a
bet [
5
].
Algorithm
s
of
arti
fici
al
intel
li
gen
ce
f
or
se
arch
i
ng
pe
rson
al
iz
at
ion
:
Pers
on
al
iz
ed
searc
hing
a
re
the
al
gorithm
s
that
us
e
interest
s
of
us
e
rs
to
pro
duce
fast
a
nd
re
le
van
t
searc
h
r
esults.
Am
ong
the
input
pa
ra
m
et
er
s
for
these
al
gor
it
h
m
s
are
us
e
r
prof
il
e,
a
naly
sis
of
hype
rlin
ks
,
a
naly
sis
of
pag
e
s
co
nte
nt,
an
d
valuati
ons
of
colla
borati
ve
s
earches
[
6].
T
he
obj
ect
ive
of
t
he
m
entioned
al
gorithm
s
is
t
o
giv
e
a
weig
ht
of
releva
nce
to
the
resu
lt
s
of
the
us
er'
s
query.
It
[7
]
us
es
s
om
e
cl
assifi
cat
ion
and
weig
htin
g
al
gorithm
s
based
on
the
c
on
t
ent
of
pag
e
s
sel
ect
ed
in
us
e
r'
s
qu
e
r
y,
w
hile
[
8]
it
does
t
he
acc
ordi
ng
w
ork
us
i
ng
aut
om
at
ic
l
earn
i
ng
te
ch
niq
ue
s
-
Re
fer
e
nced
in se
ver
al
s
ources
o
f
co
nsult
at
ion
w
it
h
the
a
ng
l
ic
is
m
"
m
achine learn
in
g"
-
w
hich
ca
n
be
cl
assifi
e
d
as artifi
ci
al
intel
li
gen
ce
work aim
ed
at
the a
utono
m
ou
s
an
a
ly
sis of
data fl
ow
s
.
Faci
al
recogn
it
ion
:
Faci
al
rec
ogniti
on
is
the
us
e
of
al
gorithm
s
that
ta
ke
im
ages
or
m
od
el
s
of
a
fac
e
as
inp
ut
pa
ram
et
ers
in
order
t
o
process
them
and
ge
nerat
e,
as
a
resu
lt
,
a
corres
pondin
g
identit
y
m
at
chi
ng
to
a
database
of
in
div
id
uals.
As
detai
le
d
by
[
9
],
de
pe
nd
i
ng
on
w
hat
is
go
in
g
to
be
analy
ze
d,
im
ages
or
m
od
el
s,
al
gorithm
s
of
diff
e
re
nt
char
a
ct
erist
ic
s
will
ta
ke
into
acco
un
t
va
rio
us
bio
m
et
ric
aspect
s
for
analy
sis
can
be
us
e
d,
eac
h o
ne havi
ng
di
ff
e
rent
b
ene
fits i
n t
er
m
s o
f
sp
ee
d
a
nd e
ff
ect
ive
ness
.
2.2.
Analysis
Be
low,
a
c
omparati
ve
st
ud
y
betwee
n
al
gori
thm
s
of
the
cat
egories
ex
pose
d
in
def
i
niti
ons
sect
ion
of
this
arti
cl
e
is
presente
d,
ta
king
int
o
acc
ount
ind
ic
at
ors
c
oncern
i
ng
the
m
easur
em
ent
of
e
ff
ect
ive
ness
of
each
on
e
w
it
h res
pe
ct
to
the
pa
rtic
ularit
ie
s of eac
h
cat
eg
ory
the
y belo
ng
to:
a.
Text
-
sea
rch al
gorithm
s
Brute
force
sea
rch
:
T
he
ai
m
of
brute
f
orce
se
arch
is
to
m
ake
a
c
har
act
er
-
by
-
char
act
e
r
c
om
par
ison
i
n
the
te
xt
T
[s
...s
+
m
−1]
fo
r
a
ll
∈
{0,
...,
n−m
+
1}
an
d
t
he
P
[0...
m
−1]
patte
rn.
The
al
gorithm
retur
ns
al
l
valid
m
at
ches.
H
owever,
a
s
[
10
]
points
ou
t,
t
he
pro
blem
with
this
appro
a
ch
is
eff
ect
ive
ness,
since
the
c
om
plexity
of the al
gorith
m
is t
he
w
orst
po
s
sible,
bein
g o
f order
O (M
x N)
.
Knuth
-
Mo
rr
is
-
Pr
at
t
Al
gorith
m
:
KMP
al
gor
it
h
m
is
com
po
sed
of
tw
o
ph
a
ses:
a
te
xt
pr
e
processi
ng
i
n
wh
ic
h
a
bra
nc
h
ta
ble
ba
se
d
on
par
ti
al
f
ai
lures
of
a
brute
f
or
ce
sea
rch
is
ob
ta
ine
d.
Usi
ng
this
ta
ble,
the
al
gorithm
will
scro
ll
thr
ough
the
te
xt
ad
van
ci
ng
t
hro
ugh
it
,
not
cha
rac
te
r
by
cha
racter
as
in
the b
r
ut
e
force
search
,
but
in
t
he
qua
ntit
ie
s
descr
ibe
d
in
the
ta
ble.
The
c
om
plexit
y
of
th
e
al
gorithm
is
giv
e
n
by
the
orde
r
O
(n +
k)
,
whe
re
O
(
n) a
nd O (
k) are
pre
-
proce
ss and s
ubse
quent searc
h
c
om
plexiti
es.
Boye
r
-
M
oore
Algorithm
:
As
descr
ibe
d
by
[10],
the
idea
beh
i
nd
the
B
oy
er
-
Moore
al
gorithm
is
to
perform
a
proc
ess
anal
ogous
to
KM
P
al
gori
thm
,
bu
t
perform
ing
the
sea
r
ch
from
righ
t
t
o
le
ft,
w
hich
a
ll
ow
s
for
la
r
ger
j
um
ps
in
t
he
searc
h
in
the
m
ai
n
te
xt,
beca
us
e
i
f
the
la
st
le
tt
er
of
t
he
patte
r
n
t
o
be
searc
hed
is
no
t
fou
nd,
the
f
ollow
i
ng
n
cha
ra
ct
ers
can
be
discard
e
d,
bein
g
n
the
le
ngth
of
the
patte
rn.
T
he
com
plexity
of
this
al
gorithm
is su
b
-
li
nea
r,
t
hat is
, O (N / M)
.
b.
Com
par
ison be
tween al
gorith
m
s
Ba
sed
on
the
ord
e
r
of
al
gorith
m
s
analy
zed,
it
is
e
vid
e
nt
that t
he
gr
eat
est
effect
iveness
co
rresp
onds
t
o
Boye
r
-
M
oore,
due
t
o
it
s
c
om
plexity
or
de
r
as
sho
wn
i
n
Ta
ble
1
.
I
n
te
sts
co
nduct
ed
by
[
10]
us
ing
an
al
ph
a
nu
m
eric
al
ph
a
bet
an
d
s
ever
al
c
hains
gen
e
rated
rand
om
l
y,
the
fo
ll
ow
i
ng
m
easurem
ents
of
e
xe
cution
sp
ee
d
w
e
re
ob
serv
e
d
in
m
il
l
i
seco
nd
s
,
w
hic
h
co
rro
borate
the
ex
pected
e
f
fici
ency
of
eac
h
al
gorithm
as
sh
ow
n
in Ta
ble 2
.
Table
1.
C
om
plexiti
es o
f
te
xt
search
alg
or
it
hm
s
Alg
o
rith
m
Co
m
p
lex
it
y
Bru
te f
o
rce
O (
M
x
N
)
Kn
u
th
-
Morris
-
P
rat
t
O (
n
+
k)
Bo
y
er
-
Moo
re
O (
N
/
M)
Reco
v
ered f
ro
m
[
1
0
],
algo
rith
m
s f
o
r
String
m
atch
in
g
Table
2.
T
est
re
su
lt
s of te
xt s
earch
alg
or
it
hm
s
Pattern len
g
th
Matches
Bru
te f
o
rce
KMP
BM
3
40
225
221
242
10
0
225
221
82
50
0
224
221
25
Reco
v
ered f
ro
m
[
1
0
],
algo
rith
m
s f
o
r
String
m
atch
in
g
,
1
-
8
.
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
20
19
:
3
7
9
8
-
3
8
0
3
3800
c.
Algorithm
s
of
arti
fici
al
intel
lig
ence
for pe
rs
on
al
iz
in
g
sea
rc
hes
LVQ
(Lea
r
ning
ve
ct
or
qua
ntiza
ti
on
)
al
gorithm
s.
Learn
in
g
al
gorithm
s
by
vecto
r
qua
ntiza
ti
on
(
LV
Q
)
are w
i
dely
u
se
d
in
classi
ficat
ion o
f
in
f
or
m
at
i
on tasks
. Acco
rd
i
ng
t
o
[11]
th
e stren
gth o
f
th
is neur
on
al
m
od
el
i
s
the
abili
ty
to
f
or
m
char
act
eri
sti
c
m
aps
in
a
sim
il
ar
way
to
wh
at
ha
ppen
s
in
t
he
br
ai
n,
this
al
go
rithm
us
e
s
reinfo
rced com
petit
ive lea
rn
i
ng, dist
in
gu
is
hin
g a t
rai
ning st
age a
nd an ex
pl
oitat
ion
stage
.
Naive
Ba
ye
sia
n
(N
B
).
Ba
sed
on
Ba
ye
s
conditi
on
al
pro
ba
bili
ty
theor
em
(1
76
3),
it
treat
s
diff
ere
nt
pr
e
dicti
on
va
riables in
dep
e
nd
ently
, while
assu
m
ing
ind
e
pe
nd
e
nce
betwee
n pr
e
dictor a
tt
ribu
te
s. The
al
gorithm
cal
culat
es
conditi
on
al
pr
ob
a
bili
ty
fo
r
co
m
bin
at
ion
s
of
at
tribu
te
s
wi
t
h
the
ob
j
ect
ive.
It
est
ablishes
a
n
ind
e
pende
nt
prob
a
bili
ty
fr
om
pr
e
dicti
ve
data
.
This
pro
ba
bili
ty
pr
ov
i
des
th
e
li
kelihood
of
each
ob
j
ect
ive,
on
ce
the insta
nce
of
each
value
cate
gory is
giv
e
n from
each
input
var
ia
ble.
Decisi
on
trees
(C4.5
).
Using
the
in
du
ct
iv
e
l
earn
i
ng
m
et
ho
do
l
og
y,
decisi
on
tree
al
gorithm
cl
assifi
es
from
a
set
of
t
rainin
g
data.
I
n
eac
h
e
xec
ution
of
t
he
al
gor
it
h
m
,
an
e
valu
at
ion
of
eac
h
node
is
m
ade
a
nd
it
is
determ
ined
w
hich
is
the
be
st
as
a
decisi
on
pa
ram
et
er.
K
-
Nea
rest
N
ei
ghbors
(
KNN)
.
Kno
w
n
a
s
la
zy
le
arn
in
g
[12]
.
The
pa
ram
eter
s
of
cl
assifi
cat
ion
by
nei
ghbor
hood
a
re
base
d
on
the
searc
h
in
a
set
of
prototypes
,
of
k
pr
oto
ty
pes
cl
os
est
to
t
he
prototype
t
o
be
c
la
ssifie
d.
A
m
e
tric
is
sp
eci
fie
d
in
orde
r
to
m
easur
e
pro
xim
it
y, Eucli
dean
distance
is
norm
al
ly
u
sed for c
om
pu
ta
ti
on
al
r
eas
ons.
Suppor
t
vect
or
m
achines
(S
V
M).
As
[
13
]
say
s,
a
Su
pport
Vecto
r
Ma
chin
e
(S
VM)
le
ar
ns
the
su
rf
ac
e
of
tw
o
dif
fer
e
nt
cl
asses
of
e
ntry
po
i
nts.
A
s
a
on
e
-
cl
ass
c
la
ssifie
r,
the
de
scriptio
n
giv
e
n
by
suppo
rt
vectors
data
is
c
apa
bl
e
of
f
or
m
ing
a
decisi
on
bounda
ry
ar
ound
the
le
ar
ning
data
dom
ai
n,
with
ver
y
li
tt
l
e
or
no
knowle
dge
of
data
ou
tsi
de
th
is
boundar
y.
D
at
a
are
m
app
ed
by
m
eans
of
a
Gau
ssia
n
ke
rnel
,
or
an
oth
e
r
ty
pe
of
kernel,
t
o
a
fe
at
ur
e
sp
ace
in
a
hi
gh
e
r
dim
e
ns
io
nal
s
pac
e,
wh
e
re
t
he
m
axi
m
u
m
separ
at
ion
be
twee
n
cl
asses
is
so
ug
ht.
T
his
bor
der
f
un
ct
i
on,
w
hen
bro
ught
bac
k
int
o
in
put
sp
ac
e,
ca
n
sepa
rate
data
by
dif
fere
nt
cl
asses,
each
form
ing
a
grou
ping.
Com
par
ison
be
tween
al
gorithm
s.
Fo
r
c
ompari
ng
t
he
di
f
fer
e
nt
al
gorith
m
s,
we
ta
ke
the
res
ults
ob
ta
ine
d
i
n
previo
us
wor
ks
[
11
]
in
w
hich
a
set
of
arti
cl
es
from
new
s
w
ebsite
s
f
ro
m
6
dif
fer
e
nt
s
our
ces
in
En
glish
is
ev
al
uated.
T
hey
wer
e
pr
e
-
pr
oc
essed
in
ord
er
to
el
im
inate
recurri
ng
te
rm
s
of
the
la
ngua
ge,
uppe
rcase
a
nd
lowe
rcase
w
ere
no
rm
al
iz
e
d,
wor
ds
within
th
e
vect
or
or
docum
ent
giv
e
n
a
weigh
t
or
i
m
po
rtance
,
a
nd
KEEL
sim
ulati
on
to
ol
was
us
e
d
in
orde
r
t
o
ob
ta
in
t
he
ac
cur
acy
per
ce
nt
age
of
eac
h
al
gorith
m
in the ne
ws
c
ol
le
ct
ion
a
s s
ho
wn in T
able
3.
Table
3.
Pr
e
dic
ti
ve
accu
racy (
%) of
each
alg
or
it
hm
in
the
ne
w
s c
ollec
ti
on
var
yi
ng
the num
ber
of
t
erm
s f
ro
m
2
0
t
o 200
0
#
of
T
er
m
s
LVQ1
LVQ2.1
LVQ3
SVM
°
C
KNN
NB
Av
erage
20
7
7
,5
7
4
,7
7
8
,2
8
4
,5
8
5
,8
7
5
,5
8
6
,9
7
8
,7
40
8
0
,2
7
7
,3
7
9
,2
8
9
,4
9
1
,0
8
2
,5
9
2
,0
8
1
,5
60
7
8
,4
7
7
,5
7
8
,3
8
8
,0
8
9
,1
7
9
,6
9
1
,6
8
0
,6
80
7
9
,0
7
6
,8
7
7
,9
8
8
,8
8
9
,4
7
9
,3
9
2
,3
8
0
,6
100
7
9
,5
7
6
,7
7
9
,1
8
8
,5
8
9
,3
7
8
,4
9
2
,3
8
0
,9
200
8
5
,8
7
8
,9
8
5
,4
9
1
,0
8
7
,4
7
7
,9
9
2
,1
8
5
,3
300
8
7
,1
7
9
,1
8
4
,0
9
2
,6
8
7
,8
6
9
,0
9
1
,9
8
5
,7
400
8
9
,0
8
2
,0
8
7
,7
9
3
,4
8
6
,9
7
4
,6
9
1
,8
8
8
,0
500
8
9
,3
8
3
,1
8
7
,7
9
4
,3
8
6
,8
7
3
,1
9
1
,5
8
8
,6
600
9
0
,0
8
3
,9
8
7
,0
9
5
,5
8
5
,9
7
2
,8
9
1
,4
8
9
,1
700
9
0
,3
8
2
,9
8
7
,3
9
5
,8
8
5
,9
7
3
,3
9
1
,1
8
9
,0
800
9
1
,0
8
3
,4
8
6
,3
9
6
,0
8
5
,9
7
3
,4
9
0
,9
8
9
,1
900
9
0
,3
8
3
,5
8
5
,0
9
6
,6
8
6
,0
7
8
,1
9
0
,5
8
8
,8
1000
9
0
,9
8
2
,1
7
8
,9
9
6
,5
8
5
,9
7
5
,9
9
0
,5
8
7
,1
1100
9
0
,6
8
1
,7
7
3
,3
9
6
,9
8
5
,8
7
4
,5
9
0
,4
8
5
,6
1200
9
0
,5
8
1
,1
6
9
,9
9
7
,1
8
5
,6
7
3
,6
9
0
,1
8
4
,6
1300
9
1
,0
8
1
,3
6
5
,5
9
7
,4
8
5
,6
6
7
,9
8
9
,8
8
3
,8
1400
8
9
,9
8
0
,8
6
4
,4
9
7
,3
8
5
,6
6
5
,9
8
9
,9
8
3
,1
1500
8
9
,8
8
2
,1
6
1
,6
9
7
,4
8
5
,6
6
4
,6
8
9
,9
8
2
,7
1600
9
0
,2
8
0
,8
5
4
,7
9
7
,3
8
5
,6
5
8
,6
8
9
,8
8
0
,7
1700
8
9
,8
8
0
,7
5
2
,8
9
7
,3
8
5
,6
5
1
,5
8
9
,6
8
0
,2
1800
9
0
,3
8
1
,0
4
9
,3
9
7
,6
8
5
,6
4
8
,8
8
9
,6
7
9
,5
1900
8
9
,9
8
1
,6
4
9
,8
9
7
,6
8
5
,6
4
6
,8
8
9
,5
7
9
,7
2000
8
9
,9
8
1
,0
4
6
,3
9
7
,5
8
5
,6
2
5
,0
8
9
,5
7
8
,7
Reco
v
ered f
ro
m
[
1
1
].
An ev
alu
atio
n
o
f
the LVQ
alg
o
rit
h
m
in a text
collect
io
n
.
Cu
b
a
n
Jo
u
rn
a
l of Co
mp
u
ter S
cien
ce
,
10
(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
Id
e
ntif
ic
ation
of
ind
iv
id
ua
li
z
ation
tec
hn
i
ques
for crimi
na
l
re
cor
ds
in
s
an
ct
i
on li
sts
(
Gonz
al
o
M.
Arias
)
3801
d.
Faci
al
r
eco
gnit
ion
Du
e
to
the
fac
t
that
ind
ividual
iz
at
ion
will
be
m
ade
ta
kin
g
into
acco
unt
on
ly
i
m
ages
con
ta
ine
d
i
n
sancti
on
li
sts
da
ta
bases
an
d
i
m
ages
of
we
b
pr
ess
a
rtic
le
s
,
on
ly
m
et
ho
ds
of
im
age
analy
sis
will
be
ta
ke
n
into
account
an
d
t
ho
s
e
base
d
on
3D
m
od
el
s
w
il
l
hav
e
to
be
discar
ded.
Th
e
fo
ll
owin
g
two
facial
recogn
it
i
on
te
chn
iq
ues
are
pro
po
se
d:
Pr
inci
pal
com
pone
nt
analy
sis
(P
CA
).
Kno
wn
as
PC
A
te
chn
i
qu
e
.
This
facial
recogn
it
ion
te
c
hn
i
que
e
m
plo
ys
an
ini
ti
al
pr
ocessi
ng
of
a
f
ace
im
a
ge
to
c
onver
t
t
he
m
at
rix
of
pi
xels
into
a
set
of
vecto
rs,
th
e
n,
they
will
be
pro
j
ect
ed
i
n
a
sp
ace
of
sm
al
le
r
value
s.
T
hese
val
ue
s
are
com
par
e
d
with
t
ho
s
e
st
or
e
d
i
n
a
data
ba
se
of
fac
ia
l
inf
or
m
ation
ta
king i
nto
account a t
oler
ance
v
al
ue
.
Locali
ty
pr
ese
rv
i
ng
pro
j
ect
io
ns
(L
PP
).
T
his
al
go
rit
hm
is
known
as
L
PP
.
LPP
perform
s
the
sam
e
reducti
on
of
init
ia
l
data
that
PCA
pe
rfor
m
s,
but
in
ad
diti
on
,
it
pe
rfor
m
s
ano
t
her
proc
ess
wh
ic
h
res
ul
ts
in
alm
os
t
identic
al
values
in
the
sm
a
ll
pr
oject
e
d
s
pace
of
val
ues
w
hen
deali
ng
with
the
fac
e
of
the
sam
e
per
s
on
in
co
ns
ecuti
ve
i
m
ages
ta
ken
fr
om
the
sa
m
e
vid
e
o
source
.
Su
c
h
ad
diti
onal
processi
ng
m
ay
resu
lt
in
lowe
r
com
pu
ta
ti
on
s
pe
ed,
but t
he a
c
cur
acy
i
n
re
su
l
ts wil
l be
gr
e
at
er.
e.
Com
par
ison be
tween al
gorit
hm
s
To
ver
i
fy
the
su
ccess
rate
a
nd
s
peed
with
w
hich
both
m
et
ho
ds
posit
ively
identify
i
nd
i
viduals
by
feed
i
ng
al
go
rithm
s
with
face
i
m
ages,
resu
lt
s
ob
ta
ine
d
by
[
9
]
will
be
ta
ken
,
w
ho
de
velo
ped
the
s
oft
wa
re
te
sts
ta
kin
g
into
acc
ount
return
val
ues
of
insta
ntane
ous
res
ults
a
nd
accum
ulate
d
r
esults;
that
i
s,
th
os
e
t
hat
re
qu
i
red
m
or
e p
r
ocessin
g
ti
m
e b
efo
r
e
ge
ner
at
in
g
a
pos
it
ive r
esult
as
s
how
n
in
Ta
ble
4
.
Table
4.
T
est
re
su
lt
s w
it
h faci
al
r
eco
gnit
ion
al
gorithm
s
Metho
d
Ins
tan
t r
esu
lts
Accu
m
u
lated
resu
lts
DCT
7
8
.74
4
%
8
1
.7%
LPP
7
7
.45
6
%
8
5
.4%
Reco
v
ered f
ro
m
[
9
].
Estu
d
io
de técnicas
de re
co
n
o
ci
m
i
e
n
to
f
acial,
8
6
.
f.
Ba
ckgrou
nd
As
ca
n
be
obs
erv
e
d
i
n
t
he
th
eor
et
ic
al
f
ram
e
work
a
nd
in
th
e
arti
cl
es
re
fere
nced,
al
gorith
m
s
intende
d
to
be
us
e
d
f
or
a
m
or
e
ef
fici
ent
ind
i
viduali
zat
ion
pr
ocess
hav
e
al
read
y
been
widely
de
velo
ped
by
va
rio
us
exp
e
rts
in
c
ompu
te
r
sci
ence.
So
it
can
be
sa
id
that
the
de
ve
lop
m
ent
of
th
is
work
is
base
d
on
the
c
om
pilat
ion
of
previ
ou
s
w
orks
res
ults
rath
er
tha
n
in
de
ve
lop
m
ent.
O
n
th
e
oth
e
r
hand,
de
sp
it
e
the
a
ntiq
uity
of
s
om
e
of
the
al
gorithm
s
discuss
e
d
in
this
arti
cl
e,
they
are
wi
dely
use
d
no
wad
ay
s
,
because
they
hav
e
pro
ve
n
their
eff
ect
ive
ness
over
ti
m
e,
su
ch
as
KMP
al
gorithm
fo
r
sea
r
chin
g
te
xt
patte
rn
s
,
w
hic
h
is
sti
ll
us
ed
in
c
urrent
browsers
w
he
n
the
us
e
r
wa
nts
to
sea
rc
h
for
a
te
xt
i
n
a
we
b
doc
ume
nt.
T
he
op
ti
m
iz
at
ion
of
processes
descr
i
bed in t
hi
s ar
ti
cl
e w
oul
d
3.
RESU
LT
S
Fr
om
the
anal
ysi
s
of
te
xt
se
arch
al
gorithm
com
plexiti
es,
as
well
as
te
sts
pro
p
os
e
d
a
nd
de
velo
pe
d
by
[
10
]
,
the
resu
lt
is
that
Boye
r
-
Mo
ore
al
go
rithm
is
the
m
os
t
rec
omm
end
ed
t
o
car
ry
ou
t
s
earches
,
bein
g
s
up
e
rio
r
to
the
oth
e
rs
in
te
rm
s
of
exec
ution
sp
ee
d.
I
n
the
analy
sis
of
in
div
id
ualiz
e
d
al
gorithm
s
r
esults,
we
ap
pr
eci
at
e
that
SV
M
pe
rfo
rm
ance
pr
esent
s
a
con
sist
ency
su
pe
rior
to o
th
er
al
gorithm
s,
al
ways
exh
i
biti
ng
a
beh
a
vior a
bove
the av
e
rag
e
, r
e
gardless
of
the
nu
m
ber
of
ter
m
s
o
f
the sam
ple, an
d sh
ow
i
ng a b
e
ha
vior wi
tho
ut
neg
at
ive
f
l
uctu
at
ion
s i
n
cases
of great
er
num
ber o
f
te
rm
s.
Wh
e
n
c
om
par
ing
facial
rec
ogniti
on
al
gorith
m
s
pr
opos
e
d
a
nd
eval
uated
by
[
9
]
,
it
is
co
nc
lud
e
d
that
the
m
e
tho
d
ge
ner
at
in
g
the
m
os
t
posit
ive
res
ults
was
local
it
y
pr
ese
rv
i
ng
proj
ect
io
ns,
reac
hing
m
or
e
than
85
%
identific
at
ions,
so
it
would
be
the
m
os
t
rec
omm
end
ed
one
w
hen
im
ple
m
enting
a
syst
e
m
.
W
it
h
a
naly
sis
of
resu
lt
s
a
nd
ob
t
ai
nin
g
the
best
al
gorithm
ic
m
et
ho
ds,
the
f
uture
inc
orp
orat
ion
of
dif
fe
r
ent
m
et
ho
dolo
gies
is
consi
der
e
d
as
a
fu
tu
re
w
ork
in
orde
r
to
op
ti
m
iz
e
resu
lt
s
in
ind
ivid
ualiz
at
ion
process,
le
a
ding
to
ideal
resu
lt
s
and
with
a
lower
er
ror
coe
f
fici
ent
and
fa
l
se
po
sit
ives.
Ar
ti
fici
al
intel
l
igence
al
gorith
m
s
allow
not
on
ly
to
i
m
pr
ove
res
ult
s
searc
hing
databases
a
nd
in
docum
ents
of
ne
ws
web
sit
es
,
bu
t
al
s
o
to
cl
as
sify
dif
fer
e
nt
s
ources
in or
der
t
o give
prior
it
y t
o
t
ho
se that
hav
e
gr
eat
er r
el
e
van
ce
in
the
in
div
i
du
al
iz
at
ion
of s
ubj
ect
s
.
4.
CONCL
US
I
O
N
In
t
his
pa
per,
three
cat
eg
or
ie
s
of
al
gorithm
s
to
be
use
d
in
t
he
pro
cess
of
im
ple
m
enting
a
n
ind
ivi
du
al
iz
at
ion
syst
em
fo
r
crim
inal
recor
ds
sea
rch
e
s
we
re
exam
ined.
Th
a
lg
ori
thm
ic
cat
egories
e
xa
m
ined
her
e
we
re
sea
rch
i
n
te
xt
s,
arti
fici
al
intel
l
igence
for
personal
iz
at
ion
of
searc
hes
,
an
d
facial
rec
og
niti
on.
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
20
19
:
3
7
9
8
-
3
8
0
3
3802
They
we
re
c
om
par
ed
us
i
ng
t
he
m
et
rics
pr
opose
d
i
n
pre
vio
us
w
orks
,
su
c
h
as
Herná
nd
e
z
,
G
ou
,
a
nd
Be
ta
ncou
r
in
order
t
o
obta
in
the
best
te
chn
i
qu
e
s
from
e
ach
cat
eg
or
y.
Finall
y,
it
was
fou
nd
that
the
m
os
t
reco
m
m
e
nd
a
ble
al
gorithm
s
fo
r
us
e
in
an
in
divi
du
al
iz
at
ion
sy
stem
are
Boyer
-
Mo
ore
f
or
te
xt
search
,
vect
or
s
upport
m
ac
hin
e
s
for
a
rtific
ia
l i
ntell
igence,
a
nd l
ocali
ty
p
rese
rvi
ng
pro
j
ect
io
ns f
or
facial
r
ec
ogniti
on.
ACKN
OWLE
DGE
MENTS
This
w
ork
wa
s
su
pp
or
te
d
by
the
Un
ive
rsi
dad
Nacio
nal
de
Colom
bia,
Sede
Me
dellí
n
under
t
he
pro
j
ect
s
HER
MES
-
34
671
a
nd
HERMES
-
36911.
The
a
ut
hors
tha
nk
the
School
of
P
hy
sic
s
fo
r
their
valua
ble
su
pp
or
t t
o
c
ondu
ct
t
his r
e
sear
ch.
REFERE
NCE
S
[1]
A
F.
Z.
Salmam,
A.
Mada
ni,
an
d
M.
Kiss
i,
“
Em
oti
on
rec
ogni
ti
o
n
from
fac
ia
l
ex
pre
ss
ion
base
d
on
fiduc
ia
l
poin
ts
det
e
ct
ion
and
usi
ng
Neura
l
Netw
ork,
”
Int. J. Elec
tr.
Comput.
Eng
.
,
vol
.
8
,
no
.
1
,
p
.
52,
Feb
.
2018
.
[2]
B
L.
Deshpande
and
M.
N.
Rao,
“
Conce
pt
Drift
I
dent
if
ic
a
ti
on
using
Cla
ss
ifi
er
En
sem
ble
Approac
h,
”
Int.
J
.
Elec
tr.
Comput.
Eng
.
,
v
ol.
8
,
no
.
1
,
p
.
19
,
Feb
.
2018
.
[3]
C
A.
L.
H.
P
.
S
and
U.
Er
anna,
“
An
Eff
ic
ie
nt
Acti
vi
t
y
Det
ec
t
ion
S
y
stem
base
d
o
n
Skele
ton
Join
t
s
Ide
nti
fi
ca
t
ion,”
Int.
J. Ele
c
tr.
Co
mput.
Eng
.
,
vol
.
8,
no
.
6
,
p
.
4995
,
Dec
.
2018
.
[4]
V.
Balajicha
ndr
ase
khar
M
.
,
T.
S.
Rao,
and
G.
Sriniva
s,
“
An
I
m
provised
Meth
odolog
y
to
Unbar
Android
Mobile
Phone
for
Foren
sic
Ex
aminat
ion
,
”
Int
.
J. E
l
ectr.
Comput.
Eng
.
,
v
ol.
8
,
no
.
4
,
p
.
22
39,
Aug.
2018.
[5]
Pande
y
,
S
.
K.
,
D
ub
e
y
,
N
.
K.
,
&
S
har
m
a,
S,
“
A St
ud
y
on
String
Matc
h
ing
Methodo
logi
es,
”
5(3)
,
47
32
-
4735
,
2014
.
[6]
Jos
hi,
C.
,
Jaiswa
l,
T., &
Gaur
,
H,
“
An Overvi
ew S
tud
y
of
Person
al
i
ze
d
W
eb
Sear
ch
,
”
3(1), 1
-
3
,
2
013.
[7]
Salmela
,
L
.
,
&
T
arh
io, J., “
Algori
thms
for
W
ei
ght
ed
Ma
t
chi
ng
,
”
W
ork,
276
-
286
,
2007
.
[8]
Hapfe
lmeie
r
,
A.
,
Merte
s
,
C.
,
Sch
m
idt
,
J.,
&
Kra
m
er,
S,
“
Tow
ard
s
Rea
l
-
Ti
m
e
M
ac
hin
e
Learni
ng
,
”
Department
of
Computer
Scien
ce
,
Techn
ische
Unive
rs
it
ät
Mün
che
n
,
85748
Ga
r
chi
ng,
Germ
an
y
,
2012
.
[9]
Herná
ndez,
R
.
,
“
Estudi
o
d
e
t
éc
n
i
ca
s de
re
conoci
m
ie
nto
fa
cial,
”
8
6.
Re
cupe
r
ado
d
e
htt
p://te
o
com.go
ogle
cod
e.com/svn
-
histor
y
/r39/
tru
nk/doc
s/pap
ers/
PF
C_RogerGim
eno.
pdf
,
2010
.
[10]
Gou,
M
, “
Algori
thms
for
String m
at
chi
ng,
”
1
-
8
,
2014
.
[11]
Guerre
ro
En
amor
ado,
A
.
,
&
C
eb
al
los
Gast
el
l
,
D,
“
Una
evalua
c
ió
n
del
al
gor
it
m
o
LVQ
en
una
co
l
ec
c
ión
de
t
ext
o,
”
Re
v
ista
Cubana
de
Ci
enc
ias
Info
rm
áti
cas
,
10(4)
,
154
-
170
,
2016
.
[12]
Garc
ía,
C
.
,
&
G
óm
ez
,
I
,
“
Algorit
m
os
de
apr
end
iz
a
je
:
knn
&
k
m
ea
ns,”
Uni
ve
r
sidad
Carlos
III
de
Madrid
,
1
-
8
.
Ret
ri
eve
d
f
rom
htt
p://ww
w.i
t.
u
c
3m
.
es/j
villena/ir
c/
pra
ct
i
ca
s/08
-
0
9/06.
pdf
,
2006
.
[13]
Bet
an
cour
,
G
,
“
La
s
m
áqui
n
as
de
soport
e
vec
to
rial
(S
VM
s),”
Sci
en
tia
Et
Te
chni
c
a
,
(27),
67
-
7
2.
htt
ps://
do
i.
org/1
0.
22517/234472
14.
6895
,
2005
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Go
nz
alo
M
.
Arias
:
Inform
a
ti
cs
Engi
ne
er
gr
adua
t
ed
from
Po
li
técni
co
Colom
bia
no
Jaime
Isaz
a
Cada
vid
,
Mede
l
lí
n,
Co
lombia.
W
orke
d
as
profe
ss
or
of
C#
in
the
T
ec
no
lógi
c
o
de
Antioquia,
Mede
ll
ín
,
Colo
m
bia
.
He
is
a
software
progra
m
m
er
with
expe
ri
enc
e
in
C#
,
Visual
Basic
and
C
++
.
His re
sea
r
ch
in
terests c
over
m
ostl
y
s
ec
uri
t
y
topi
cs.
E
-
Mai
l:
M
aur
ici
o.
arias@outlook
.
com
Pab
l
o
A.
Pel
áe
z
:
Syst
e
m
s
En
gi
neer
gr
a
dua
te
d
from
U
niv
e
rsida
d
de An
ti
oquia, Medell
ín,
Colom
bia. H
e i
s a
do
c
um
entation
m
ana
ge
r
a
nd d
at
a
base a
dm
inist
rator
.
E
-
Ma
il
: andres
gr
is
82@
gm
ail.co
m
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
Id
e
ntif
ic
ation
of
ind
iv
id
ua
li
z
ation
tec
hn
i
ques
for crimi
na
l
re
cor
ds
in
s
an
ct
i
on li
sts
(
Gonz
al
o
M.
Arias
)
3803
Fr
ed
y
Ed
ime
r
Hoy
os
:
r
ec
e
ive
d
his
BS
and
MS
degr
ee
from
the
Nati
ona
l
Univer
sit
y
of
Colom
bia,
at
Man
iz
a
le
s,
C
olombia,
in
El
e
c
tri
c
al
Engi
ne
eri
n
g
and
Industr
ia
l
Autom
at
ion,
in
2006
and
2009,
respe
ctively
,
an
d
Industrial
Au
tomati
on
Ph.D.
in
2012
.
Dr.
Ho
y
os
is
cur
re
ntly
an
As
sociate
Profess
or
of
the
Scie
n
ce
Facu
l
t
y
,
School
of
P
h
y
sics
,
at
N
a
ti
o
nal
Univer
sit
y
of
Colom
bia
,
at
Mede
ll
in
,
Co
lo
m
bia
.
His
rese
ar
ch
intere
sts
in
clude
nonli
n
ea
r
co
ntrol
,
s
y
st
em
m
o
del
li
ng
,
non
li
ne
a
r
d
y
nami
cs
anal
y
s
is,
con
trol
of
no
nsm
ooth
sy
st
ems
,
and
power
ele
ct
roni
cs,
with
ap
pli
c
at
ion
wi
thi
n
a
broa
d
ar
ea
of
t
e
chnol
ogi
c
al
pro
ce
ss
.
Dr.
Ho
y
os
is
an
As
socia
t
e
Resea
r
che
r
in
Colc
ie
n
cias
and
m
ember
of
the
Applie
d
Techno
logi
es
Resea
r
ch
Group
-
GITA
at
the
Univer
si
dad
Nac
ional
de
Colom
bia
.
h
tt
ps:
//
orc
id
.
org/0000
-
0001
-
8766
-
5192
Evaluation Warning : The document was created with Spire.PDF for Python.