TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 9
85~995
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.1476
985
Re
cei
v
ed
Jan
uary 21, 201
5
;
Revi
sed
Ap
ril 29, 2015; Accepted Ma
y
16, 2015
Pornographic Image Recognition Based on Skin
Probability and Eigenporn of Skin ROIs Images
I Gede Pasek Suta Wija
y
a
*
1
, IBK Widi
artha
2
, Sri Endang Arjar
w
a
n
i
3
Informatics Engineering Dept., Facult
y
of
Engineering, Ma
taram Universit
y
,
Jl. Majapahit 62 Mataram,
Lombok, NT
B-IN
DONESIA
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: gpsuta
w
i
j
a
y
a
@
te.ftunram.a
c.id
1
, w
i
di@ftunram.ac.id
2
,
endang@ti.ftunram.ac.id
2
A
b
st
r
a
ct
T
he p
aper
pr
opos
ed
a p
o
r
nogr
aph
ic i
m
a
ge rec
o
g
n
itio
n
usin
g ski
n
p
r
oba
bil
i
ty an
d
princ
i
pl
e
compo
nent
an
alysis (PCA)
o
n
YCb
C
r col
o
r
space. T
he
p
o
rno
g
rap
h
ic i
m
age r
e
cog
n
iti
o
n is d
e
fine
d a
s
a
process
to cl
a
ssify the
i
m
ag
e co
ntain
i
n
g
a
nd s
how
in
g
g
e
n
ital
el
e
m
e
n
ts
of hu
man
bo
dy
from a
n
y ki
nd
s of
imag
es. T
h
is
p
r
ocess is
har
d
to be
p
e
rfor
med
beca
u
se t
h
e i
m
a
ges
hav
e
larg
e var
i
a
b
il
ity du
e to
pose
s
,
lighti
ng, a
nd b
a
ckgro
und v
a
ri
ations. T
he ski
n pro
bab
ility a
nd h
o
listic fe
atur
e, w
h
ich is e
x
tracted by YC
b
C
r
skin se
g
m
enta
t
ion a
nd P
C
A, is e
m
pl
oye
d
to han
dl
e
tho
s
e vari
abi
lity
prob
le
ms. T
h
e
function
of s
k
in
seg
m
e
n
tatio
n
i
s
to deter
mi
ne
skin Re
gi
on of
Interest (R
OI) ima
ge
and sk
in
prob
abi
lity. W
h
ile th
e functio
n
of
PCA is to extract eigenpor
n of the
skin ROIs images and to
project the
skin ROI vector using the obtained
eig
enp
orns to
holistic fe
ature
s
. T
he ma
in a
i
m
of th
is res
earch is to
op
timi
z
e
the acc
u
racy an
d fals
e
rejecti
on rate
of the skin pr
oba
bil
i
ty and f
u
sio
n
descr
ipt
o
r bas
ed rec
o
gniti
on syste
m
. The experi
m
enta
l
result sh
ow
s that the pr
op
ose
d
metho
d
ca
n i
n
creas
e the
ac
curacy by
ab
o
u
t 4.0%
an
d d
e
creas
es the F
P
R
20.6%
of those
of porn
ogra
p
h
i
c reco
g
n
itio
n u
s
ing fusi
on
des
criptors, resp
e
c
tively. In ad
dit
i
on, the
prop
os
ed
meth
od
is
als
o
ro
bust for
la
rge si
z
e
datas
et that is
sho
w
n by g
i
vin
g
similar
perf
o
rmance
to th
e l
a
test
meth
od
(Multi
l
a
yer-Perc
eptro
n a
nd
Neur
o-
F
u
zz
y
(MP-
N
F
)). T
he pro
pos
ed
met
hod
als
o
w
o
rks fast for
recog
n
itio
n, w
h
ich req
u
ires 0.
12 seco
nds p
e
r
ima
ge.
Ke
y
w
ords
: por
nogr
aph
ic, pca
,
imag
e reco
gn
ition,
skin pr
ob
abil
i
ty, and h
o
li
stic features
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Reje
ction sy
stem for acce
ssi
ng po
rno
g
r
aphi
c
conten
ts (text, image and video)
is a big
issue d
e
velo
ping
count
ry like Indo
ne
sia. It is
cau
s
ed by the weak
prote
c
tio
n
of acce
ssi
ng
porn
o
g
r
aphi
c content
s, wh
ich ca
n lead t
o
so
cial pro
b
l
e
ms in comm
unity. Nowad
a
ys, prote
c
tio
n
for acce
ssing
porn
ographi
c co
ntents i
s
done by blo
cki
ng the
site's dom
ain na
me indi
cating
to
contai
n po
rno
g
rap
h
ic
co
nte
n
ts seen fro
m
the text
information of the
sites.
However, this meth
o
d
has so
me
we
akn
e
sse
s
: firstly, The fal
s
e po
sitive
pe
rforman
c
e
is
high b
e
cau
s
e
text informat
ion
of porno
grap
hic site a
r
e
much
simila
r to medi
cal
sites a
nd sit
e
s rel
a
ted to
human an
atomy;
se
con
d
ly, the gro
w
th of
p
o
rno
g
raphi
c
sites i
s
ve
ry
fast not o
n
ly due to
exch
a
nging th
e sit
e
’s
domain
but
also
big
bu
si
ness; the
re
st, this bl
o
cking m
e
thod
also
ha
s
hig
h
false n
ega
tive
perfo
rman
ce
becau
se th
e
sites containi
ng le
ss te
xt b
u
t many p
o
rn
ogra
phi
c ima
ges an
d vide
os
can
not be re
cog
n
ized a
s
negative site
s. That m
ean
s peo
ple can
easily acce
ss site
s co
ntai
ning
porn
o
g
r
aphi
c image
s or vi
deo
s. More
over, the po
rno
g
rap
h
ic
co
nte
n
t also
can trigger th
e so
ci
al
probl
em
su
ch
as a
ddi
ction
of porn
o
g
r
a
phic
co
ntent, sexual
hara
s
sment, and
early p
r
eg
na
ncy.
These soci
al
probl
em
s will be hap
pe
n to child
re
n
and tee
nag
ers
wh
o al
ways a
c
cess t
h
e
porn
o
g
r
aphi
c content with
out much
kn
owle
dge of
sex education
includi
ng th
e disa
dvanta
ges
free sex.
Therefore, it
is
requi
re
d
a rej
e
ctio
n
syst
em th
at
can
blo
c
k or reje
ct of a
c
ce
ssi
ng
porn
o
g
r
aphi
c content. Th
e system wo
rks like a fire
wall that can
block or rej
e
ct of acce
ssing
porn
o
g
r
aphi
c conte
n
t based on its i
n
formatio
n, su
ch a
s
text, image
s an
d video. The
sy
stem
sho
u
ld
also b
e
run
on
a va
riety of g
adg
e
t
s
with differe
nt platform.
T
he
reje
ction
system
con
s
i
s
ts
of many com
p
lex sub pro
c
e
s
ses in
clud
ing free pr
ocessing, age
nt extracto
r, feature extra
c
tion,
and
recogniti
on en
gine. In
addition, the
reje
ction
sy
stem al
so fa
ce
s ma
ny ob
stacl
e
s
su
ch
as
large va
riabili
ty images du
e to lighting, pose, co
lo
r variation
s
. It mean
s, the rejectio
n syst
em
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 985 – 995
986
need
s st
ron
g
re
cog
n
itio
n system th
at can al
mo
st perfe
ctly cla
ssify the
input imag
e
as
porn
o
g
r
aphi
c
image o
r
non
-po
r
no
gra
phi
c image.
In orde
r to solve the men
t
ioned p
r
oble
m
, a porn
ographi
c image
recognitio
n
u
s
ing
ski
n
probability and principl
e compon
ent analysis
(PCA
)
on YCbCr co
lor space i
s
proposed. The
porn
o
g
r
aphi
c image
re
co
g
n
ition is defin
ed a
s
a
pr
ocess to
cla
ssif
y
the image
contai
ning
an
d
sho
w
in
g geni
tal element
s
of human
bo
dy from any
kind
s of ima
ges. Th
e ski
n
pro
bability
and
holisti
c feature, which is ex
trac
ted by
YCbCr skin
seg
m
entation
an
d PCA, is em
ployed to han
dle
large
vari
abili
ty porno
grap
hic im
age
p
r
o
b
lems.
Th
e
fu
nction
of
ski
n
segm
entation
is to
dete
r
mi
ne
ski
n ROI ima
ge and
skin
prob
ability. While the fu
n
c
tion of PCA
is to extract
eigen
porns o
f
the
ski
n ROI
s
image
s and
the holistic features
a
r
e determi
ne
d by using
the eigenp
orn
s
.
Theo
retically, it can be realize
d
becau
se po
rno
g
ra
p
h
ic co
ntent like photo
s
an
d videos can
be
cla
ssifie
d
by sha
pe/po
se, ski
n colo
r,
a
n
d
ge
nita
l info
rmation.
However, p
o
rn
ographi
c
conte
n
t i
s
gene
rally a
color ima
ge th
at has la
rg
e
diversity
in te
rms
of lightin
g,
pose,
and backg
rou
nd. In
addition, tem
p
lates matchi
ng al
so
can
be a
pplied
to
cla
s
sify wh
ether th
e p
o
se
of the o
b
je
cts in
the input im
a
ge contain
s
p
o
rno
g
raphi
c
content.
He
re,
template
s of
pornog
rap
h
i
c
can
be g
eni
ta
l
element
s, su
ch a
s
b
r
ea
st
s, vagina,
an
d/or pe
nis. In
terms of the
text information, input me
dia
su
ch a
s
web
s
ite can b
e
classified a
s
p
o
rno
g
r
aphy from proba
bility occurren
ce
s of word
s
a
n
d
phra
s
e
s
a
s
so
ciated
with po
rnog
ra
phi
c su
ch a
s
a
hard-core, free sex, porn video
s, and so fo
rth.
This pa
per,
whi
c
h i
s
fo
cu
s o
n
p
o
rnog
raphi
c ima
g
e
recognitio
n
,
a po
rn
ograph
ic im
age
recognitio
n
u
s
ing skin p
r
obability and
princi
ple co
mpone
nt ana
lysis (PCA
) on YCb
C
r color
spa
c
e, i
s
o
r
gani
zed
as follows: the fi
rst
se
cti
on d
e
scrib
e
s the
introdu
ction
of this
work;
the
se
con
d
se
cti
on explain
s
the pr
evio
us
works; the third se
ction p
r
ese
n
ts ou
r propo
sed meth
od
inclu
d
ing YCbCr
skin seg
m
entation, Eigenp
orn
ex
traction a
nd
kNN m
a
tchi
ng
pro
c
e
ss; the f
ourth
se
ction explai
ns expe
rimen
t
al re
sult an
d discu
ssi
on; a
nd the re
st prese
n
ts con
c
lu
sion a
nd future
wor
k
.
2. Related Works
This
re
sea
r
ch mostly re
lates to pat
tern recognit
i
on, whi
c
h
con
s
i
s
ts of
lighting
norm
a
lization
,
object dete
c
tion, intelligen
ce sy
stem
s, and matchi
ng
pro
c
e
s
ses.
Some dete
c
ti
on of po
rn
ographi
c imag
es calle
d
a
s
co
ntour-ba
s
e
d
and regio
n
b
a
se
d
[1],
and h
u
man
ski
n proba
bili
ty [2-4] had
been
perfo
rmed by som
e
re
sea
r
che
s
. Those
met
hod
s
perfo
rmed th
e dete
c
tion
based o
n
th
e skin info
rmation, whi
c
h we
re extracted
by skin
segm
entation
.
In addition,
the po
rno
g
ra
phic imag
e re
cog
n
ition u
s
i
ng fu
sion
de
scripto
r
s (FD)
[5]
also
had
be
e
n
propo
se
d. Ho
wever, th
o
s
e m
e
thod
s l
a
ck of a
c
cu
racy, high
fal
s
e
po
sitive a
n
d
negative d
a
ta
due to la
rg
e
variability of
porn
o
g
r
aphi
c image
s. Reg
a
rdin
g to
ski
n
seg
m
entatio
n
,
the skin
cla
s
sificatio
n
mo
dels th
at we
re impl
ement
ed for
se
gm
enting the
skin
regio
n
were
threshold m
o
del in YCb
C
r, HSV, and RGB colo
r
spaces a
nd G
aussia
n
mixture mo
del
s [5-8].
The F
D
-ba
s
e
d
metho
d
i
s
an imp
r
ovem
ent of
ski
n region
an
d contour
ba
se
d
method
s. T
he
eigen
porn of
ski
n
segm
e
n
ted ima
ge
of HSV color cha
nnel
also ha
s be
en
prop
osed
wh
ich
provide b
e
tter achi
eveme
n
t than FD method on t
he HSV colo
r cha
nnel [9]. However, it also
lacks
of a
c
cu
racy
and fal
s
e re
co
gnition
rate. Th
e
po
rnographi
c d
e
tection
usi
ng l
o
cali
zatio
n
skin
ROI [10]
had
bee
n p
r
op
osed a
nd
provi
ded
better
achievement th
an POESIA classifier. In th
a
t
method, the i
m
age featu
r
e
s
that con
s
ist
ed of ratio
of total ski
n to n
on-ski
n pixel
s
within conve
x
hull, mean
s and varia
n
ce
of RGB colo
r cha
nnel, se
ven spatial in
variant mom
en, and ang
e
l
o
f
prin
ciple axi
s
of convex hull versus
hori
s
ont
al ax
is, were extracted fro
m
ROI image.
The
cla
ssifi
cation
wa
s pe
rform
ed by
Ran
d
o
m
Forest
tree
model.
Reg
a
r
ding
to robu
st po
rno
g
ra
p
h
ic
image
re
co
gn
ition over la
rg
e si
ze
data
s
e
t, the
multi p
e
r
ce
ptro
n a
nd
neuro fu
zzy
(MP-NF
)
ba
se
d
method
h
ad been pro
p
o
s
e
d
whi
c
h provi
ded
rea
s
on
ab
le
re
sult (abo
ut 87% in
TP
and 5.5% in
FN
on test data
s
et) com
p
a
r
ed
other rel
a
ted
work
s. The
MP-NF m
e
th
od used com
p
lex and 17
kinds
of feature
s
[1
1] for re
cog
n
i
t
ion, which a
nalytic
ally re
quire
d long
computation
a
l
time for featu
r
e
extraction
s.
The i
n
telligence sy
stem t
hat had
succe
ssfully
been devel
oped and impl
em
ented i
n
pattern
re
co
gnition
can
be group
ed i
n
to thre
e m
a
jor g
r
o
u
p
s
, namely: firstl
y, feature-b
a
s
ed
method
whi
c
h is comp
reh
ensive/h
o
listi
c
(feature ex
t
r
acte
d u
s
in
g
statistical a
n
a
l
ysis, texture,
and
freque
ncy
)
; seco
ndly, artifi
cial intelli
gen
ce b
a
s
ed
m
e
thod (Neural Networks,
Ge
netic
Alg
o
rith
ms,
and F
u
zzy L
o
g
ic);
and
the
rest i
s
combi
n
ation of
b
o
th
of them. Mo
st
ly, holistic fe
a
t
ure
whi
c
h
wa
s
extracted
by
sub
s
p
a
ce, ha
d be
en
su
ccessfully imple
m
ented fo
r fa
ce
re
cog
n
itio
n such a
s
fa
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Porno
g
ra
phi
c Im
age Reco
gnition Based
on Skin Prob
ability an
d… (I Gede Pasek Suta Wijaya
)
987
recognitio
n
u
s
ing
PCA
alg
o
rithm [1
2, 1
3
], LD
A [13
-
1
6
], DCT+P
C
A [13, 15], an
d DCT
+
PCA
L
DA
[15, 16]. Among them, th
e PCA and L
D
A are ve
ry popul
ar be
ca
use of not o
n
l
y
their simple
ness
but also thei
r easi
n
e
ss to b
e
impleme
n
te
d.
In
term
s
of feature extra
c
tion, some
te
chni
que
s
known a
s
h
o
l
i
stic fe
ature,
sh
ape
feature,
and f
a
cial
feature
(eyes,
no
se,
and m
outh
)
e
x
traction
had
been
propo
sed a
nd
provi
ded
rea
s
on
able
a
c
hievem
ent. All of these t
e
ch
niqu
es
were widely
i
m
pleme
n
ted to
determi
ne
the
uniqu
e patte
rn of th
e o
b
ject. Th
e i
m
pleme
n
tatio
n
exampl
e
of holi
s
tic fe
ature
extra
c
tion
techni
que
s a
r
e co
ntent-b
a
s
ed fe
ature
u
s
ing frequ
en
cy analysi
s
(F
FT, DCT, an
d Wavel
e
t) a
n
d
feature p
o
int descri
p
tor u
s
i
ng SIFT [13], [15-17]. In
a
ddition, the shape featu
r
e
that is extra
c
ted
by moment a
nalysi
s
co
uld
increa
se the
discrimi
natory power of fa
cial feature.
In this pap
er, the alternat
ive solution
of porn
o
g
r
ap
hic ima
ge re
cog
n
ition u
s
i
ng skin
probability and ski
n ROI
s
i
m
ages
is proposed to sol
v
e many obs
tacles faced in pornographic
image
re
co
gn
ition. In a
ddition, this meth
od i
s
also
p
r
o
posed to
imp
r
ove the
existi
ng m
e
thod
s [
1
-
3], [5] and will be impleme
n
ted for reje
ction syst
em o
f
pornog
rap
h
i
c
co
ntents e
s
peci
a
lly imag
es
and video bo
th in standal
one pe
rsonal
compute
r
, tablet, sma
r
t phon
e and t
he intern
et from
unexpe
cted p
eople.
3. Proposed
Metho
d
s
The p
r
op
ose
d
po
rno
g
ra
ph
ic ima
ge recognition
algo
rithm is p
r
e
s
e
n
ted in th
e F
i
gure
1,
whi
c
h co
nsi
s
t
s
of
trainin
g
and re
cog
n
ition
p
r
o
c
e
s
ses. Both processe
s a
r
e
co
nstructed
by p
r
e-
pro
c
e
ssi
ng
(histog
r
am
eq
ualization a
n
d
skin
segm
entation),
eig
enpo
rn
extra
c
tion
usi
ng P
C
A
and matching
process u
s
in
g kNN sub proce
s
se
s.
The functio
n
of each
sub
-
pro
c
e
ss of p
o
rn
o
g
raphi
c image recogn
iti
on diagram
block is
descri
bed a
s
follows:
a)
Histo
g
ra
m e
q
ualization i
s
a
pro
c
e
s
s to
re
move
no
n u
n
i
f
orm lig
hting
effect on
ima
ge
captu
r
ing
whi
c
h can decrease the large variability of por
nographic images
due to lighting
variation.
b)
Skin segme
n
tation is u
s
e
d
to remove n
on sk
in pixel
of the input i
m
age
s which
can d
e
cre
a
se
the large va
ri
ability porno
g
r
aphi
c ima
g
e
s
due to ba
ckgrou
n
d
s
.
c)
PCA is used
to extract th
e
eigen
po
rn from
skin
seg
m
ented
po
rn
ogra
phi
c ima
ges called
a
s
ski
n ROIs im
age
s. In this
ca
se the
outp
u
t of th
is process in the
eig
enpo
rn a
nd t
he proje
c
tion
matrix whi
c
h i
s
nee
ded in t
he matchi
ng
pro
c
e
ss.
d)
The simila
rity between the
query eig
enp
orn an
d
the trainin
g
set ei
genp
orn i
s
determin
ed by
kN
N alg
o
rit
h
ms.
Figure 1. Pornographi
c im
age re
co
gniti
on diag
ram bl
ock
In the trai
nin
g
process, training
set i
s
norm
a
lized b
y
histog
ram
equali
z
ation,
next th
e
histog
ram eq
ualization out
put is seg
m
e
n
ted to
remo
ve non-ski
n pixels, next from the skin ROIs
image
s of tra
i
ning set, the eigenp
orns
are extra
c
te
d
by PCA. Wh
ile in re
cog
n
ition process, the
query ima
ge i
s
treate
d
a
s
the same a
s
i
n
the tr
aini
ng
pro
c
e
ss
but
until PCA pro
j
ection p
r
o
c
e
ss,
next the similarity of query eigen
porn an
d the trai
ning
set eige
npo
rn
by kNN al
gorithms.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 985 – 995
988
3.1. YCbCr
Based Skin Segmentation
Skin is im
port
ant informatio
n of porn
o
g
r
a
phic im
age
s
becau
se mo
stly the porno
grap
hic
image
s a
r
e
covered
by
ski
n
. The
r
efo
r
e,
non-skin
pi
xe
ls h
a
ve to
be
rem
o
ved f
r
o
m
po
rno
g
ra
p
h
ic
image
s by segmentatio
n
algorith
m
s. Commonly,
the
ski
n se
gme
n
tation algo
rit
h
m wo
rks u
s
i
ng
pixel-ba
se
d skin cl
assification. Several a
l
gorithm
s
reg
a
rdin
g to pixe
l base
d
skin
classificatio
n
[5-
8]
have bee
n
ca
rri
ed out whi
c
h define
a
pixel
s
a
s
skin
or non
-skin u
s
ing
thre
shol
d rule
s.
The
threshold
rule
s were
create
d
ba
sed
on t
he hi
stogram
informatio
n of
colo
ur
spa
c
e
.
For exampl
e,
a pixel in RG
B color
spa
c
e
is defined a
s
skin if it satisfies the following crite
r
ia [5,
7]:
R
>95
and
G
>40
and
B
>
20
and
Max(R,G,B)
-
m
i
n(R,G,B)>15
and
(1)
|
R-G
|
>
15
an
d
R
>
G
an
d
R
>
B
In this ca
se, the R, G, and
B have 256 level in the ra
ge of 0 to 255
.
The skin
cla
s
sificatio
n
also
can b
e
perfo
rmed in Y
C
b
C
r
colo
r space. The mo
st popul
ar
rule for
ski
n classificatio
n
in YCbCr colo
r spa
c
e i
s
def
ined a
s
follows [7].
77
C
b
≤
12
7
an
d
133
C
r
≤
1
7
3
(
2
)
In this case,
Y comp
one
nt has
200 l
e
vels
rangi
ng from 16 to
235
and
Cb, Cr
have 22
5 lev
e
ls
rangi
ng fro
m
16 to 240. T
hose levels
a
r
e extra
c
ted f
o
rm
RGB col
o
r spa
c
e u
s
i
ng YCb
C
r
co
lor
s
p
ac
e
tr
an
s
f
or
ma
tio
n
.
In other ha
nd
, a pixel i
s
cl
assified a
s
skin in
HSV col
o
r
spa
c
e
if it satisfie
s th
e followin
g
c
r
iteria [2, 3].
0 < H < 0.25
and
0.15 < S
< 0.90
and
0
.
2 < V < 0.95
(3)
As mention
e
d
previou
s
ly, these rule
s a
l
so
defin
ed b
a
se
d the hist
ogra
m
inform
ation of
each colo
r compon
ents.
The pixel val
ues of H,
S,
and V
are i
n
the
ran
ge
of 0-1,
whi
c
h
are
determi
ned b
y
the followin
g
equatio
n [2]:
cos
⁄
(
4
)
1
3
,
,
(
5
)
1
3
⁄
(
6
)
The exam
ple
of ski
n seg
m
entation u
s
i
ng the
Equ
a
tion (1
), Equat
ion (2
), an
d
Equation
(3) a
r
e p
r
e
s
e
n
ted in Figu
re 2. From th
ese al
go
rithm
s
, HSV had b
een impl
eme
n
ted to se
gm
ent
ski
n col
o
r a
n
d
to extract skin p
r
ob
ability for
pornog
ra
phic im
age d
e
tection [2] a
nd provided g
ood
enou
gh pe
rfo
r
man
c
e. In a
ddition, the Y
C
b
C
r al
so
ha
s be
en impl
e
m
ented fo
r skin
segm
enta
t
ion
and a
pplie
d
porn
o
g
r
aphi
c image
re
co
g
n
ition [5]
an
d
the YCbCr
skin
seg
m
enta
t
ion gave
bet
ter
perfo
rman
ce.
Therefore, in this
pap
er,
the
YCb
C
r based ski
n segmentatio
n is
em
ployed for
extracting
ski
n
ROIs im
ag
es
of traini
ng
and
que
ryin
g sets. Th
e
diagram bl
ock of
ROI im
a
ge
extraction
is
sho
w
n
in Fi
g
u
re
3. After
ski
n
s
egm
ent
ation, the
ROI extra
c
tion
is
starte
d from
perfo
rming t
he vertical
(rows) dan h
o
risontal
(col
umn) p
r
oje
c
t
i
on pro
babilit
y to know the
coo
r
din
a
tes
having large
of skin a
n
d
non-skin re
gion. Second
ly, the vertical and ho
ri
sontal
projection probability having less
than a defined threshol
d
is removed. In this
case,
by trial and
error th
e b
e
st threshold
can be
defie
n
d
a
s
0.25
of
maximum vertical an
d h
o
ri
sontal
proje
c
tion
prob
ability. T
h
irdly, the
ski
n
tone
is
cro
pped
u
s
ing
the
x
and
y
coordi
nate
s
where
the
verti
c
al
and
ho
riso
nta
l
projectio
n
probability a
r
e
thre
shol
ded.
F
i
nally, the
cro
pped
skin
ton
e
is ma
ppe
d t
o
origin
al imag
e to get the skin ROI imag
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Porno
g
ra
phi
c Im
age Reco
gnition Based
on Skin Prob
ability an
d… (I Gede Pasek Suta Wijaya
)
989
Input
Output of RG
B rules
Output of HS
V rules
Output of YCbCr
rule
s
Figure 2. The
example of skin segme
n
ta
ti
on results u
s
ing
RGB, HSV, and YCbCr rule
s
Figure 3. The
ROI image e
x
traction
Next, from th
e ROI imag
e, the ei
genp
orns
ar
e
extra
c
ted by PCA
for im
provin
g
the ski
n
prob
ability (SP), ski
n re
gion (S
R)
a
nd fusi
on
d
e
scripto
r
[FD] base
d
po
rnographi
c im
age
recognitio
n
. The mai
n
different of thi
s
PCA to t
hat of Referen
c
e
[5] is on its
function
ality. The
traditional P
C
A wa
s imp
l
emented fo
r dimen
s
ion
a
l
redu
ction
o
f
FD to de
crea
se the ti
me
compl
e
xity FD ba
se
d reco
gnition
syste
m
, while th
e
PCA on thi
s
rese
arch i
s
n
o
t
only employ
ed
for dime
nsio
nal re
du
ction
but also fo
r extracti
ng
ei
genp
orn
and
the holisti
c
feature
s
of b
o
th
training
set a
nd qu
eryin
g
i
m
age
s. In a
d
d
ition, the m
a
in differen
c
e
of ou
r eig
e
n
p
orn
to
cla
ssi
cal
PCA is o
n
d
e
termini
ng th
e
global
covari
ance a
s
shown
in
Equatio
n (9), whi
c
h
d
o
e
s not usi
ng
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 985 – 995
990
global m
ean
but usin
g ea
ch cla
ss
mean
. It wa
s ch
osen be
cau
s
e t
he data in
put
only have two
cla
s
ses (po
r
n
ogra
phi
c and
non-po
rnog
ra
phic).
3.2. Eigenporn Extrac
tio
n
Suppo
se, the
r
e a
r
e t
w
o d
a
t
a cla
s
ses (p
orn
and
non
-porn
)
d
enote
d
by X
1
and
X
2
. The X
1
has N
1
imag
es an
d X
2
has N
2
image
s.
In this case, the images
are the ROIs having different
sizes. The
r
ef
ore, the ROI
s
have to be normali
zed t
he size to be 64 pixels with kee
p
ing
the
asp
e
ct
ratio.
The
ratio
is define
d
a
s
64 divid
ed
b
y
max of im
age
width
a
nd ima
ge
he
ight
(64/max(i
m
Width,imHeig
ht)). By re
si
zin
g
the im
a
ge usin
g
the def
ined ratio,
if the
input
im
a
ge
width i
s
wid
e
r
tha
n
the
hei
ght, the im
ag
e wi
dth i
s
re
sized
into
64
and th
e im
ag
e hei
ght follo
ws
the ratio, and
vice versa. From these d
a
ta, the eigenp
orn is extract
ed as follo
ws.
Step
1.
Conve
r
ting e
a
ch im
age o
f
the data cl
asse
s into column ve
ctor using
ro
w-o
r
de
ring
p
r
oc
es
s
.
Step
2.
Re
sizi
ng the
colu
mn ve
ct
or into
si
ze
4096
elem
en
ts by zero p
addin
g
in th
e end
of
element
s the
vector to provide the sa
me vect
or
si
ze f
o
r P
C
A
.
For in
st
an
ce,
t
he zer
o
paddi
ng i
s
d
one fro
m
ind
e
x 4032 to
4
095 of in
put
ROI havin
g 6
4
x 63 pixel
s
to rea
c
h
defined
colu
mn vector 4
0
96 eleme
n
ts. In this
ca
se, assumin
g
the
index is start
ed from 0
Step
3.
Determinin
g t
he me
an
of e
a
ch
data
cl
ass
a
nd m
ean
of all d
a
ta
sa
mples u
s
ing
Equation
(7) a
nd (8
), re
spe
c
tiv
e
ly
.
∑
,
k=1,
2
(
7
)
∑∑
(
8
)
Step
4.
By using the
mean of ea
ch
cla
ss
(
k
) and data
cla
s
ses, the gl
o
bal cova
ria
n
ce matrix
(
C
g
) i
s
cal
c
ul
a
t
ed by Equation (9
).
∑∑
(
9
)
Step
5.
Performi
ng t
he eig
en a
n
a
lysis
of
C
g
for obtai
nin
g
projectio
n
matrix, W, u
s
ing th
e
Equation (10).
, i = 1, 2, 3,
… ,
n
(
1
0
)
The
w
i
an
d
λ
i
is the
i
-th ei
genve
c
tor
an
d eige
nvalue
of
C
g
respectively. While
n
is the
numbe
r of eig
en value
s
whi
c
h is alm
o
st t
he sam
e
as d
i
mensi
onal in
put vector.
Step
6.
Selecting sm
all
numbe
r
of eigenve
c
tors (
m
) rep
r
e
s
ent
ing eige
npo
rn
of data cla
s
ses an
d
put them
into
matrix
W
=[
w
1
,w
2
,w
3
,…,w
m
]
n x
m
w
h
ic
h h
a
v
e to
s
a
tis
f
y the cr
iter
ia as
pre
s
ente
d
in Euatiuon (11).
W
C
W
max
arg
J
g
T
W
PCA
(
1
1
)
In order to satis
f
y this
c
r
it
eria, s
m
all
m
eigenvecto
rs which
corre
s
po
nd to the large
s
t
eigenvalu
e
s (i.e.
m<
n
) are
sele
cted a
s
ei
genp
orn.
Step
7.
Proje
c
ting th
e ea
ch
inp
u
t
vectors
usin
g
obtain
ed
eig
enpo
rn
as ho
listic fe
ature
t
hat can
be perfo
rme
d
by Equation (12
)
:
)
x
(
W
y
k
k
j
T
k
j
(
1
2
)
Step 8.
Save sele
ctin
g eigen
porn a
nd proj
ectio
n
vector for
re
cognition p
r
o
c
ess.
The exampl
e of eigenp
orn t
hat is
extra
c
ted by this alg
o
rithm is
sho
w
n in Fig
u
re
4. In this
ca
se, the inp
u
t data classes are 687 p
o
rno
g
r
aphi
c image
s and t
he sele
cted e
i
genp
orn
s
are 18.
Theo
retically, eigen
po
rn e
x
traction
usi
n
g PCA i
s
la
ck of the
po
wer di
scrimi
nat
ory compa
r
e
to
LDA (Li
nea
r
Discri
mina
nt Analysis) for
large
sam
p
le
size
data. It also
req
u
ire
s
retraini
ng of
all
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Porno
g
ra
phi
c Im
age Reco
gnition Based
on Skin Prob
ability an
d… (I Gede Pasek Suta Wijaya
)
991
s
a
mples to
obtain the mo
st favorabl
e p
r
ojectio
n
mat
r
i
x
. However, t
he d
a
ta of
thi
s
re
sea
r
ch h
a
v
e
two cl
asse
s
(po
r
n a
nd n
on-p
o
rn), the
power
discriminatory d
o
e
s n
o
t give
much
effect
to
recognitio
n
p
e
rform
a
n
c
e.
The small
sa
mple si
ze
dat
a, the betwe
en cl
ass
scatter matrix of
LDA
clo
s
e to o
r
be
sing
ular,
whi
c
h me
an
s the
proje
c
tion
m
a
trix is not o
p
timum or
can
not be o
b
tain
ed
by eigen an
al
ysis.
Figure 4. The
18 eigen
porn
s
of PCA of the
training set
Figure 5. The
matching p
r
o
c
e
ss u
s
in
g kNN
3.3. Matchin
g
Process
The mat
c
hin
g
pro
c
e
s
s is perfo
rmed
by
k-nea
re
st
neigh
bors (kNN), whi
c
h
can be
illustrate
d
usi
ng Fi
gure 5.
Suppo
se
we
have two
cl
a
s
se
s
(A a
nd B
)
re
pre
s
e
n
ted
by blue
delta
(
∆
)
and re
d squ
a
r
e (
), if the query (g
ree
n
diamond,
) ent
er to the syst
em the distan
ce bet
ween t
he
query
and
th
e trai
ning
cl
a
s
s a
r
e
determined
by Eu
clide
an
dista
n
ce.
Wh
en th
e
k
pa
ram
e
te
r i
s
3,
the query i
s
concl
uded as
class
B, because the
nearest
neighbor
probability to
class B i
s
hi
hger
(2/3) tha
n
to cla
ss A (1/3
). Howeve
r, if the k pa
ramet
e
r is set-up b
y
7,
the query
is con
c
lud
e
d
as
the cl
ass A
becau
se th
e
nearest
nei
g
hbor p
r
ob
abil
i
ty
to cla
s
s B
is le
ss (3/7) than
to
cla
s
s A
(4/7).
4. Results a
nd Analy
s
is
Several expe
riments were
carried out to
know the
performance
of proposed
metho
d
usin
g data
s
et
con
s
i
s
ting
of 1400
imag
es [5, 19].
This dataset
is co
mpos
ed
of 6
87 po
rn
ograp
hic
image
s an
d
715 n
on-po
rnographi
c im
age
s, whi
c
h
were do
wnl
o
aded from th
e intern
et using
download
er t
ools. Th
e po
rnog
ra
phi
c image
s cont
a
i
n naked
sin
g
le, cou
p
le
s, triples p
e
rsons
sho
w
in
g hu
man bo
dy genitals a
nd
sexual a
c
tivities. The p
o
rnographi
c im
age
s are m
o
stly
females and
some
of ima
ges
have
ski
n like
ba
ck
ground
s
su
ch a
s
sand,
woo
d
, etc. Whil
e n
on-
porn
o
g
r
aphi
c image
s co
ntain obje
c
ts t
hat are
simil
a
r to hum
an
ski
n su
ch flo
w
er,
woo
d
, tiger,
desse
rt, etc. The expe
rime
nts we
re pe
rf
or
me
d und
er
the followin
g
circum
stan
ce
s [5]:
1)
50%
of
e
a
ch porn
o
g
r
aphi
c and non
-po
r
n
ogra
phi
c
ima
ges were
ran
domly sele
cte
d
a
s
trai
nin
g
s
e
t,
2)
testing imag
e
s
we
re ove
r
l
appe
d with traini
ng
set, beca
u
se the testing ima
g
e
s
highly com
e
from the sa
m
e
person a
s
the trainin
g
se
t.
3)
the accu
ra
cy, false
neg
ative ra
te
(F
NR),
and fal
s
e
po
sitive rate
(F
PR) p
a
ramet
e
rs were
u
s
e
d
for perfo
rma
n
c
e indi
cato
rs,
and
4)
the evaluatio
n wa
s carried
out on pc
wi
th spe
c
ificatio
n Intel Core i3-23
70M, 2.4
GHz, 8 G
B
RAM.
The a
c
curacy
, false ne
gati
v
e rate
(FNR), and fal
s
e
p
o
sitive rate (FPR)
we
re
calcul
ated
usin
g the followin
g
formul
a:
100%
(
1
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 985 – 995
992
100%
(
1
4
)
100%
(
1
5
)
Whe
r
e the
T
P
(true
po
sitive) is th
e p
o
rno
g
raphi
c t
e
sting im
age
that is truly
re
cogni
ze
d
as
porn
o
g
r
aphi
c imag
e, T
N
(true
ne
gative) i
s
th
e n
o
n
-po
r
n
ograph
ic te
sting
im
age th
at i
s
t
r
uly
recogni
ze
d a
s
non
-po
r
n
o
g
r
aphi
c ima
ge,
FN (false
ne
gative) is the
porno
g
raphi
c testing ima
ge
that is fal
s
e
l
y re
cogni
ze
d a
s
n
o
n
-
p
o
rno
g
raphi
c
image, FP
(false
po
sitive) i
s
th
e n
on-
porn
o
g
r
aphi
c testing imag
e that is fals
ely reco
gni
ze
d as po
rno
g
raphi
c image,
N
P
is
total of
porn
o
g
r
aphi
c
testing imag
e
s
, and N
N
is t
o
tal of non-p
o
r
nog
ra
phi
c testing imag
es.
The first exp
e
rime
nt was perfo
rme
d
t
o
find the
b
e
st
k
p
a
ra
m
e
ter of
k
N
N
for the
prop
osed me
thod. The ex
perim
ent wa
s done u
s
in
g
18 eige
npo
r
n
s
an
d the 0.1
5
skin p
r
ob
ab
ility
threshold, which means that t
he images having skin probability le
ss than given threshold were
con
c
lu
ded
as false d
e
tecti
on data. T
h
e
experim
enta
l
data (Fi
g
u
r
e 6)
sh
ow t
hat the be
st
k
para
m
eter of
kNN fo
r p
e
rfo
r
ming
the re
cognition
i
s
9 whi
c
h
i
s
in
di
cated by the
hi
ghe
st a
c
cura
cy
(86.99%
) a
n
d
small
e
st FP
R a
nd F
N
R (18.
81% a
nd
9.05%, re
sp
e
c
tively). Thi
s
k
pa
r
a
me
te
r is
impleme
n
ted for the perfo
rming the further evalu
a
tio
n
to prove that the eigenp
orn can be u
s
ed
for re
cog
n
izi
n
g the porn
o
g
r
aphi
c image
s.
The second
experim
ent was pe
rformed
to fi
nd the best num
ber
o
f
eigenpo
rn
s
of PCA
that is suita
b
le for the reco
gnition. In th
is ca
se, the skin p
r
o
bability threshold and the
k
para
m
eter of kNN we
re setup 0.15 and 9
resp
ective
ly. The experi
m
e
n
tal data sho
w
that the best
numbe
rs of
e
i
genp
orn
s
th
at is suitable
for
pe
rformi
ng the
recog
n
ition i
s
2
0
e
i
genp
orn
s
,
which
are
shown
by high
en
o
ugh
accu
ra
cy and
sm
all
FPR and
FNR (87.95
%, 13.71%, and
10.33%respe
c
tively), as shown in Fi
gu
re 7.
From thi
s
a
c
hieve
m
e
n
t, the 20
eig
enpo
rn
s a
nd
k=
9
will be used f
o
r further pe
rformances ev
aluation in
the next experi
m
ents.
Figure 6. The
k
para
m
eter
of kNN versu
s
perfo
rman
ce
of the propo
sed method
Figure 7. The
performan
ce
s of our meth
od
versu
s
nu
mb
er of eigen
po
rns
The third exp
e
rime
nt wa
s carrie
d out to kno
w
the effect of ski
n probability thre
shol
d to
porn
o
g
r
aphi
c re
co
gnition
perfo
rman
ce.
In thi
s
ca
se,
the exp
e
rim
ent was do
n
e
u
s
ing
the
same
data set a
s
i
n
the first an
d se
co
nd
experim
ent,
k=9
, 20 ei
gen
p
o
rn
s of P
C
A, and th
e
sa
me
perfo
rman
ce i
ndicators. Th
e experime
n
tal results sh
o
w
that the best thre
shold o
f
skin proba
bi
lity
that provid
e h
i
gh a
c
cura
cy
and
small fal
s
e
re
cog
n
iti
on rate i
s
0.1
5
.
By using thi
s
skin
proba
b
ility
threshold, th
e propo
sed
method
s p
r
o
v
ide ac
cu
ra
cy by about 8
7
.02% and
F
P
R and
FNR by
about
14.13
% and
11.7
9
%, re
spe
c
ti
vely (se
e
Fi
gure
8
)
. Th
ese
re
sult
s
sup
port th
at the
eigen
porns from P
C
A
of YCb
C
r ROI
s
im
age
s i
s
re
asonabl
e
metho
d
fo
r re
co
gni
zing
the
porn
o
g
r
aphi
c image
s. In
addition, fro
m
the first a
nd second
e
x
perime
n
ts a
c
hievem
ent (20
eigenporn and skin probability 0.15)
, the next experiment i
s
perf
ormed for
knowin
g the further
perfo
rman
ce
s.
The fourth
e
x
perime
n
t was carried o
u
t to kno
w
the pe
rform
a
nce of the p
r
opo
se
d
algorithm
compared to the existi
ng m
e
thods (skin
probability (S
P), skin
regi
on (S
R), fusi
on
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Porno
g
ra
phi
c Im
age Reco
gnition Based
on Skin Prob
ability an
d… (I Gede Pasek Suta Wijaya
)
993
descri
p
tor
(F
D) m
e
thod
s [
1
-3], [5], and
eigen
porn of
HSV skin
se
g
e
mentaio
n
im
age [16]. In t
h
is
experim
ent,
k=
9
, 20
eige
np
orn fe
ature
s
and 0.1
5
thresh
old of
ski
n probability
were em
ploy
ed.
The exp
e
rim
ental results (Fig
ure
9)
sho
w
th
at propo
sed
meth
od ten
d
s to
provid
e b
e
tter
perfo
rman
ce
s than all exist
i
ng metho
d
s.
In detail,
the accuracy of
prop
osed me
thod in
cre
a
se
s
by about 4% (from 84 to 8
8
%) and de
creases the
F
N
R 20.6% (fro
m 28.8 to
8.2%) of those of
the be
st met
hod
(FD on
YCbCr),
re
sp
ectively. Ho
wever, the FP
R in
crea
se
s
by about
12.
9% of
that of FD method. The
accuracy in
cre
m
ent an
d
FNR d
e
cre
m
ent are
hi
gher th
an F
P
R
increment, which
means t
hat the proposed
method
still gives better perform
a
nce than that
of FD
method [5]. It can b
e
a
c
h
i
eved be
ca
use the eig
enp
orn of
ski
n
ROIs i
m
age
s provide
holi
stic
informatio
n of porno
graphi
c data cla
ss
which i
s
re
p
r
e
s
ented by som
e
eigen
porns
(se
e
Figu
re 4
)
.
In addition, the holisti
c informatio
n its self does
not
only come from eigen
porns but al
so come
from
ski
n inf
o
rmatio
n, wh
ich
are
rep
r
ese
n
ted
by
s
k
i
n
ROIs images
. It is k
n
own from the
experim
ental
results that th
e a
c
curacy
of the
pro
p
o
s
ed
metho
d
i
s
hi
gher tha
n
tha
t
of with
out
ski
n
probability (S
P), and the
F
N
R and FP
R
is less than t
hose of
withou
t SP, as
shown i
n
Figure
10.
From the
s
e p
e
rform
a
n
c
e
s
, it can be co
nclu
ded that
the pro
p
o
s
ed
method is alt
e
rnative solut
i
on
for pornog
rap
h
ic imag
e re
cognition.
Figure 8. The
performan
ce
s of the our
method for
so
me skin prob
ability variations
Figure 9. The
propo
s
e
d
me
thod com
p
a
r
e
d
to exis
ting methods
Figure 10. Th
e comp
utatio
nal time of our
method comp
ared to the ex
isting metho
d
s
Figure 11. Th
e perfo
rman
ce of our
method comp
ared to the lat
e
st method
The fifth experime
n
t was
done to kn
o
w
com
p
utatio
nal time of the pro
p
o
s
ed
method
comp
are to that of existin
g
method
s.
The
comp
uta
t
ional time is one imp
o
rta
n
t para
meter for
evaluation
s
. It means the
best recogniti
on syste
m
m
u
st be
sho
w
n
by the highe
s
t accu
ra
cy, the
smalle
st F
NR and
FPR,
a
nd the
sho
r
te
st comp
ut
atio
nal time
for
reco
gnition. T
he id
eal
syst
em
must have 1
00% of accuracy, 0%
of F
NR, 0% of FPR, and al
m
o
st 0 se
con
d
s
com
putatio
nal
time. In this
ca
se, the
co
mputational
time i
s
d
e
fi
ned
as the tim
e
t
hat is requi
re
d for cl
assifying
the input ima
ge start
s
fro
m
loading th
e input im
ag
e. The experi
m
ental re
sult
s sh
ows that the
90.
68
5.
52
13.
03
90.
13
4.
93
14.
61
0
10
20
30
40
50
60
70
80
90
Acurracy
FPR
FNR
Rate (%)
Param
e
ters
MP+NF
[
11]
SP.+
E
P (
S
EP)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 985 – 995
994
prop
osed me
thod nee
ds
much l
e
ss co
mputational
ti
me than F
D
method [5] a
nd almo
st as fast
as SP and SR based met
hod
s (see Fi
gure 1
0
). It
mean
s that the pro
p
o
s
ed
method give
s no
t
only high e
n
ough a
c
cu
ra
cy and sm
all FNR
and FP
R but al
so n
eed
s sh
ort
computation
a
l
time
(0.12
se
con
d
s
). Thi
s
a
c
hi
evement al
so su
ppo
rts t
he previou
s
achi
evement
s that propo
sed
method i
s
alternative
solu
tion for po
rn
ogra
phi
c ima
ge re
co
gnitio
n
. In additio
n
, the pro
p
o
s
ed
method
also
an im
prove
m
ent of th
e
FD-ba
s
e
d
p
o
rno
g
raphi
c i
m
age
re
co
gn
ition which
can
improve the a
c
cura
cy, false
reco
gnition,
and computat
ional time.
In order to know the robust
performance of ou
r proposed method over large
variability
porn
o
g
r
aphi
c image
s, the
last expe
rim
ent wa
s
ca
rri
ed out o
n
large si
ze d
a
ta
set [11, 18]
and
usin
g the be
st variation of eigen
porns.
This d
a
taset con
s
i
s
ts of 1
8354 im
age
s whi
c
h 92
95
and
9059 ima
g
e
s
are pornog
raphi
c and no
n-po
rn
ograph
ic,
respe
c
tively. The images of this data
s
et
were al
so d
o
w
nlo
aded f
r
o
m
Internet u
s
i
ng some d
o
wnload
er tool
s.
The po
rn
ographi
c imag
es of
have larg
e variability in terms of peo
pl
e, pose,
skin.
Similar to UNRAM data
s
et[5,18], the
non-
porn
o
g
r
aphi
c images
cont
ain obje
c
ts
which a
r
e
ski
n
like su
ch a
s
flower, woo
d
, tiger, dessert,
etc. The trai
ning ima
g
e
s
also
setu
p
the sa
me a
s
ca
rrie
d
out
in the Refe
ren
c
e [11].
The
experim
ental
re
sult indi
cates that
ou
r propo
se
d method
te
nd
s
to
give al
most
the sa
me
achi
evement
as th
e late
st
existing m
e
th
od (M
P+NF) about almo
st
90.13% of
a
c
cura
cy,
4.93
%
and 14.6
1
% of FPR and FNR, resp
ectively, as
sho
w
n in Fi
gure 1
1
. It re-p
rove
s that the
prop
osed me
thod can
giv
e
go
od eno
u
gh
a
c
hi
evem
ents and
the eigen
porn
of YCbCr skin ROI
image i
s
sui
t
able co
ncep
t for recogni
zing p
o
rnog
raphi
c image
s. The
s
e ca
n be a
c
hiev
ed,
becau
se the
eigenpo
rn p
r
ovide
s
holi
s
tic inform
atio
n of porno
g
pahi
c image
such as ge
nital
informatio
n, sexual activityies, etc a
s
sh
own in Fig
u
re
4.
5. Conclusio
n
and Futu
r
e
Works
The p
r
op
ose
d
method
can imp
r
ove
the
existing method
s
of porn
o
g
r
aphi
c
image
recognitio
n
such a
s
skin
prob
ability, skin re
gi
on a
nd the fusio
n
descri
p
tor.
This metho
d
is
alternative so
lution for dev
elopin
g
the rejectio
n of porno
graphy i
m
age
s.
The holisti
c
features
that is extracted by eig
enpo
rn
s of ski
n ROI
s
image
s is suitable featu
r
es
con
c
e
p
t for
porn
o
g
r
aphi
c image re
cog
n
ition. It is known by
better accu
ra
cy and less F
N
R and FPR t
han
existing m
e
th
ods. In
detail
,
the propo
se
d metho
d
in
crea
se
s the
a
c
cura
cy of th
e late
st existi
ng
method
(F
D)
by about
4.0
%
and
de
cre
a
se
s th
e F
N
R a
nd
com
p
utational time
by abo
ut 20.
6%
and 2
08 milli
se
con
d
s,
re
spectively. Ho
wever, th
e F
P
R in
cre
a
ses by about 1
2
.9%. In additi
on,
the prop
osed
method ha
s a
l
most the sa
me perfo
rma
c
e a
s
MP-NFfor large si
ze
dataset.
This m
e
thod
need
s to b
e
improve
d
by adding
n
o
t only ROIs image
s fro
m
intensity
comp
one
nts
but al
so f
r
om
the chromina
nce
comp
one
nt su
ch
a
s
Cb an
d
Cr com
pone
nts i
n
o
r
der
to inc
r
ease the acc
u
rac
y
.
Ackn
o
w
l
e
dg
ements
This re
se
arch
is
su
ppo
rted
and fu
nde
d b
y
Mi
niste
r
of
Comm
uni
cati
on a
nd Info
rmation
of
Rep
ubli
c
of In
done
sia. In
a
ddition, o
u
r great t
han
k is a
l
so to
Imag
e
Media
Lab
ora
t
ory Kumam
o
to
Univers
i
ty for disc
us
s
i
ons
to this
res
e
arch.
Referen
ces
[1
]
H
u
W, Wu
O,
C
h
en
Z, Fu
Z,
May
b
an
k S. R
e
cog
n
i
tio
n
o
f
Po
rn
og
ra
ph
i
c
We
b
Pa
ge
s by
Cl
a
ssi
fy
ing
T
e
xts and Ima
ges.
IEEE Transactio
n
s on
Pattern Ana
l
ys
is and M
a
chi
n
e Intelli
ge
nce
. 29(6): 10
19-
103
4.
[2]
Marcial-B
a
sil
i
o
JA, Agui
lar-
T
o
rres G, Sa
nchez-Pr
e
z G, T
o
scano-Med
i
na
LK, Prez-
M
ean
a HM
,
Hern
adez
E. E
x
p
licit
Cont
ent
Image D
e
tecti
on.
Si
gna
l & I
m
a
ge
Process
i
ng: An
Internat
ion
a
l J
ourn
a
l
(SIPIJ)
. 2010; 1(2): 47-5
8
.
[3]
Marcial-B
a
sil
i
o
JA, Agui
lar-
T
o
rres G, Sa
nchez-Pr
e
z G, T
o
scano-Med
i
na
LK, Prez-
M
ean
a HM
.
Detectio
n of Porno
g
rap
h
ic Di
gital Imag
es.
Internatio
nal J
o
u
r
nal of Co
mput
ers
. 2011; 5(
2): 298-3
05.
[4]
Mustafa R, Z
hu D. Objectio
n
abl
e Image
Det
e
ction i
n
Cl
ou
d
Computi
ng P
a
radi
gm-a R
e
vie
w
.
Jo
urn
a
l of
Co
mp
uter Scie
nce
. 201
3, 9(1
2
): 1715-
17
21.
[5]
Wijay
a
I GPS,
Widiartha
I BK, Uchim
u
ra K, K
outaki G.
Por
n
ogra
phic
Imag
e Rec
o
g
n
itio
n
Using
F
u
sio
n
of Scal
e Invar
i
ant D
e
scriptor
s
. Procee
din
g
s
of the
21st K
o
rea-J
apa
n j
o
i
n
t W
o
rksho
p
o
n
F
r
onti
e
rs o
f
Comp
uter Visi
on (F
CV 201
5)
. Mokpo- South
Korea. 20
15.
Evaluation Warning : The document was created with Spire.PDF for Python.