Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
4, N
o
. 4
,
A
ugu
st
2014
, pp
. 59
3
~
60
2
I
S
SN
: 208
8-8
7
0
8
5
93
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Video Summari
zation B
a
s
e
d on a Fuzzy Bas
e
d In
crement
a
l
Clustering
Monireh P
o
ur
naz
a
ri, F
a
rib
o
rz
Mahm
ou
di,
Amir
Masou
d
Eftekh
ari
Moghadam
Is
lam
i
c Az
ad Un
ivers
i
t
y
of Qa
zvi
n
, Ir
an
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Feb 22, 2014
Rev
i
sed
Jun
2
,
2
014
Accepte
d
J
u
l 24, 2014
The signifi
cant
developm
ent of
m
u
ltim
edia and
digital video pr
oduction i
n
recen
t
y
e
a
r
s has led to th
e m
a
ss production of p
e
r
s
onal and com
m
e
rci
a
l vid
e
o
archiv
es. Th
eref
ore, th
e need fo
r e
fficient
tools
and methods of accessing
video
conten
t
a
nd inform
ation
rapid
l
y is sig
n
ific
antl
y
incr
e
a
sing. Vid
e
o
summ
arizat
ion i
s
the r
e
m
oval
of visual
redun
danc
y
and r
e
pe
titiv
e v
i
de
o
frames, and obtaining
a short summar
y
of th
e whole vid
e
o
so that th
e
summ
ar
y
obtain
e
d effec
tive
l
y
r
e
flec
ts th
e whole video conten
t. Examples of
thes
e s
u
m
m
a
rizations
in rec
e
n
t
y
e
ars
includ
e
STIM
O
and
VSUMM
.
According
to us
ers’ comments, in th
e men
tioned
methods, the summarization
has
a high rate of error in a full report of s
u
mm
ariza
tion and a lo
w accura
c
y
in non-rep
e
titive frames production, as we
ll as a high
com
putation time. I
n
this paper, in or
der to solve the
s
e
problems, we developed a s
y
stem which
modeled users’
and superviso
r
s’
com
m
e
nts
.
W
e
us
ed a
fuzz
y b
a
s
e
d
increm
ent
a
l clus
tering b
y
whi
c
h
the s
e
lection an
d deselection of
frames are
done based on fuzzy
rules. Th
e extrac
ted rules
were determined based on
users’ comments on the video
summari
zation.
Finally
,
we p
e
rf
ormed our
proposed method on the video clips used in the previous methods. Produced
summaries were evaluated b
y
a qua
litative
method to minimize human
interf
eren
ces
.
The res
u
l
t
s
obtain
e
d indi
cat
e the h
i
gh a
ccura
c
y
o
f
summarization and
the
le
ss c
o
mput
a
t
i
o
n ti
me
.
Keyword:
Fuzzy
t
h
res
hol
di
n
g
Inc
r
em
ental clustering
Ma
m
d
ani fuzz
y syste
m
Vi
de
o s
u
m
m
ar
i
zat
i
o
n
Copyright ©
201
4 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Monire
h Pournazari,
Islamic Azad
Uni
v
er
sity of
Qazv
i
n
,
Ir
a
n
Em
ail: m
_pournazar
y@ya
hoo.com
1.
INTRODUCTION
Accessib
ility o
f
t
h
e d
i
g
ital v
i
d
e
o
con
t
en
t o
n
t
h
e web
is in
cred
ib
ly in
creasing
.
Web
s
ites su
ch
as
Yo
uT
u
b
e a
n
d
i
T
u
n
es
Vi
de
o
on
w
h
i
c
h
p
e
opl
e
u
p
l
o
a
d
or
d
o
w
nl
oad
vi
de
os a
r
e s
u
ccessf
ul
an
d t
h
ey
ar
e
devel
opi
ng
ra
p
i
dl
y
.
In t
h
i
s
a
r
t
i
c
l
e
, i
n
st
ead o
f
Tag (
w
hi
ch i
s
not
al
way
s
acc
essi
bl
e, i
n
t
e
g
r
a
t
ed an
d ap
pr
o
p
r
i
a
t
e
),
we
use a
searc
h
t
o
ol
whi
c
h i
s
base
d
on
t
h
e
v
i
deo c
o
nt
ent
.
T
h
i
s
t
o
ol
ca
n
bri
e
fl
y
di
spl
a
y
t
h
e co
nt
ent
t
o
t
h
e use
r
i
n
o
r
de
r t
o
ha
v
e
an i
d
ea
ab
o
u
t
t
h
e vi
deo
co
nt
ent
wi
t
h
o
u
t
wa
t
c
hi
ng
t
h
e
vi
de
o a
nd t
h
e
user
can
deci
de
wh
et
he
r
to
d
o
wn
lo
ad
or watch
th
e
v
i
d
e
o
with
ou
t watch
i
n
g
it firs
t. In
th
is p
a
p
e
r, we o
p
e
rate based
on
in
cremen
tal
cl
ust
e
ri
n
g
. T
h
e
adva
nt
age
of t
h
i
s
ki
n
d
o
f
cl
u
s
t
e
ri
ng i
s
t
h
at
i
t
can fi
nd
key
fram
e
s i
n
a predet
erm
i
ned
m
a
nne
r
with
ou
t d
e
termin
in
g
th
e nu
mb
er of k
e
y
frames.
Bu
t
in
stea
d, t
h
ere is
a need for a
n
acc
urate
determ
ination
of
th
resh
o
l
d
v
a
lue wh
ich
su
mmarizes th
e fil
m
in
to
app
r
op
ri
at
e num
ber of
fra
m
es. Accordi
ngly, as the previous
m
e
thods
were
not accurate in determ
in
ing the thres
h
old a
nd
use
d
a fixe
d
thres
h
old to
sum
m
arize all
vide
os,
we
u
s
ed
an
exp
e
rt
sup
e
rv
iso
r
fo
r d
e
term
in
in
g
t
h
e thre
sh
o
l
d so th
at t
h
e
fu
zzy lo
g
i
c
will determin
e a
v
a
lu
e as
th
e th
resho
l
d
fo
r each
o
f
t
h
e v
i
d
e
o
s
u
n
d
e
r
stu
d
y
an
d
will p
e
rform
th
e su
mmarizatio
n
o
p
e
ration
in
the b
e
st
way. Th
erefo
r
e, th
e d
e
term
i
n
ed thresh
o
l
d
d
e
p
e
nd
s
on
con
d
ition
s
and
t
h
e scen
e ty
p
e
and
it is calcu
lated
separately for each film
.
In this pape
r,
base
d on the
researches conducte
d on feat
ure extraction m
e
thods suc
h
as Vio
l
et, Gab
o
r filters, colo
r h
i
stog
ram
,
ed
g
e
ex
tractio
n, and
con
t
en
t seg
m
en
tatio
n
,
we
u
s
e th
e co
lo
r
hi
st
o
g
ram
as an ap
pr
op
ri
at
e m
e
t
hod f
o
r e
x
t
r
act
i
ng feat
ure
s
. As t
h
e re
qui
red co
nt
i
n
ui
t
y
bet
w
ee
n fram
e
s exi
s
t
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
59
3
–
60
2
59
4
with
ou
t add
itio
n
a
l
calcu
lation
s
, we clu
s
ter fram
e
s and
fi
n
a
lly su
mm
ari
ze th
e
v
i
d
e
o
co
n
t
en
t. Th
e
article is
or
ga
ni
zed as
f
o
l
l
o
ws:
In
sectio
n 2,
we rev
i
ew th
e recen
t
su
mm
a
r
izati
on
m
e
t
h
o
d
s whi
c
h have
t
h
e hi
g
h
est
p
e
rcent
a
ge o
f
summ
arization accuracy so far.
In
section 3,
t
h
e propose
d
fuzzy
t
h
res
h
olding is
prese
n
ted. In secti
o
n
4,
we
pr
o
v
i
d
e t
h
e cl
u
s
t
e
ri
ng m
e
t
hod
used i
n
t
h
e art
i
cl
e and i
n
sect
i
on
5, w
e
p
r
ese
n
t
t
h
e p
r
o
p
o
se
d m
e
t
hod’s
di
a
g
ram
.
I
n
sectio
n 6,
we sho
w
th
e
pr
op
o
s
ed
ev
al
u
a
tio
n and
t
h
e
resu
lts ob
tain
ed
.
Th
e last section
o
f
t
h
e article is th
e
gene
ral
c
oncl
u
si
on
.
2.
REVIEW
OF
VIDE
O S
U
M
M
A
R
IZ
ATIO
N
METHO
D
S
A m
e
t
hod
na
m
e
d VS
UM
M
was
p
r
o
p
o
se
d
by
Avi
l
a
et
al
[
1
]
w
h
i
c
h
wi
l
l
ext
r
act
c
o
l
o
r
feat
ure
f
r
o
m
fram
e
s after the pre-sam
p
ling stage of video fram
e
s. Af
ter th
e rem
o
v
a
l o
f
m
ean
in
g
l
ess fram
es, th
e re
main
in
g
fram
e
s will b
e
clu
s
tered
b
a
sed
on
k-m
ean
s clu
s
tering
algorith
m
.
In
o
t
h
e
r word
s, in
th
e
first stag
e, th
e
v
i
d
e
o
will b
e
brok
en
in
to
fram
es. In
th
e second
stage, co
lor
feat
u
r
es will b
e
ex
tracted
fro
m
v
i
d
e
o fram
e
s in
th
e fo
rm
of a c
o
l
o
r
hi
st
og
ram
i
n
HSV col
o
r s
p
ace
;
ho
weve
r, t
h
e
s
e color features are extracte
d
from
the sam
p
led
fram
e
s (1
f
r
am
e pe
r sec
o
nd
),
not
al
l
vi
de
o
fr
am
es. Al
so
, i
n
t
h
i
s
st
age
,
m
e
ani
n
gl
ess
fram
e
s are
rem
oved
fr
om
sam
p
l
e
d ones
.
In t
h
e t
h
i
r
d st
a
g
e,
fram
e
s are
gr
o
upe
d
by
t
h
e
k-m
eans cl
ust
e
ri
n
g
al
g
o
ri
t
h
m
.
In t
h
e n
e
xt
st
age,
after clusteri
ng, one
fram
e
is
selected
from
each cluster as
the key fram
e
.
Furini et al [2] suggeste
d a vide
o
sum
m
ari
zati
o
n
t
echni
q
u
e cal
l
e
d ST
IM
O
bas
e
d o
n
H
S
V c
o
l
o
r hi
st
og
ram
s
cl
ust
e
ri
n
g
. T
h
e
m
a
i
n
argum
ent
of
t
h
i
s
art
i
c
l
e
i
s
stat
i
c
and dy
na
m
i
c su
m
m
ari
z
at
i
on w
h
i
c
h i
s
a sum
m
a
ri
zat
ion t
e
c
hni
que
f
o
r
dy
nam
i
c and o
n
-t
he-
fl
y
sum
m
a
ri
es pr
o
duct
i
o
n.
Th
i
s
m
echani
s
m
i
s
desi
g
n
ed
f
o
r
cust
om
i
zat
i
on pu
r
pose
s
, i
.
e.
users
can sel
e
c
t
t
h
e
length of the
s
u
mmary and t
h
e tim
e they have to
wait
for receiving t
h
e
summ
ary.
STIMO does
not
use all
audi
o a
nd
vi
de
o i
n
fo
rm
at
i
on (
s
uc
h as cl
o
s
ed
vi
ew a
n
d u
s
er
pre
f
ere
n
ce)
an
d i
t
i
s
desi
gne
d
as st
at
i
c
or
dy
nam
i
c
in
o
r
d
e
r
to
summar
i
ze g
e
n
e
ric v
i
d
e
o
s
. D
y
na
m
i
c su
mm
ar
i
e
s w
ith
und
er
stan
d
a
b
l
e sound
s ar
e
p
e
rfo
rmed
to
in
crease read
ab
ility an
d
i
n
form
at
io
n
tran
smissio
n
.
ST
IM
O co
re in
cl
u
d
e
s a
p
r
o
c
ed
ure
wh
ich
calcu
lates HSV
color
space di
stribution for all
vide
o
frames/pers
pectives
and als
o
incl
ude
s a clusteri
ng al
gorithm
whic
h
groups sim
ilar
vide
o fram
es/perspectiv
e
s
and determ
ines the best fram
e/
pers
pective for
each
group ac
cording
to the their color sim
ilarity.
This m
e
thod
has a high
tim
e
com
p
lexity because each input fram
e should be
co
m
p
ared
with all fram
e
s
in
th
e clu
s
ters and it is t
i
m
e
consum
i
ng i
n
m
o
re cl
ust
e
rs an
d
h
a
s a hi
g
h
er m
e
m
o
ry.
Zha
ng et
al
[5]
used an u
n
s
u
p
e
rvi
s
e
d
cl
ust
e
ri
ng f
o
r ext
r
act
i
ng t
h
e key
f
r
a
m
e. At
fi
rst
,
t
h
e vi
deo i
s
di
vi
ded i
n
t
o
shots; t
h
en, a
color
histogra
m
in HDV col
o
r
space
of ea
ch fram
e is calculated.
For at
tribution
of a
fram
e to
clu
s
ter, t
h
e similarity o
f
th
e fra
m
e
with
th
e cen
t
er
o
f
t
h
at
cl
ust
e
r i
s
cal
cul
a
t
e
d base
d
on t
t
h
res
h
ol
d.
Ha
nz
l
i
k
et
al
[6]
prese
n
t
e
d a m
e
t
hod
fo
r
t
h
e vi
de
o sum
m
ary
based o
n
t
h
e cl
ust
e
r-
val
i
di
t
y
anal
y
s
i
s
whi
c
h i
s
per
f
o
r
m
e
d
wi
t
h
o
u
t
hum
an s
u
p
e
r
v
i
s
i
o
n.
At
fi
r
s
t
,
t
h
e
w
hol
e
vi
de
o i
s
gr
o
upe
d i
n
t
o
c
l
ust
e
rs.
Each
f
r
am
e i
s
di
spl
a
y
e
d b
y
colo
r histo
g
ra
m
s
in YUV co
lor space
. A classified cluster
is applied to all video fram
e
s for
n tim
e
s. The fir
s
t
ti
m
e
starts wit
h
a p
r
ed
eterm
i
n
e
d
n
u
m
b
e
r of clu
s
ters an
d
on
e clu
s
ter will b
e
add
e
d
in
each
ti
m
e
th
e clu
s
terin
g
is p
e
rfo
r
m
e
d
.
Th
en
, th
e syste
m
au
to
m
a
tica
lly fin
d
s
arb
i
t
r
ary
com
b
i
n
at
i
on(s
)
o
f
cl
ust
e
rs
by
appl
y
i
n
g
cl
ust
e
r
-
validity analysis. Afte
r fi
ndi
ng the
op
tim
i
ze
d num
b
er of cl
usters
, each cl
uster is
display
e
d by a s
p
ecifi
c fram
e
whi
c
h i
s
a new key
fram
e
for t
h
e
vi
de
o. A m
e
t
hod
was p
r
o
p
o
sed
by
Go
ng a
n
d Li
u [
7
]
for
vi
deo
su
mm
arizat
io
n b
a
sed
on
t
h
e sing
u
l
ar
v
a
lue d
e
co
m
p
o
s
itio
n (SVD).
At th
e
b
e
g
i
nn
ing, a set
of i
n
put v
i
d
e
o
fram
e
s is selected (1 out
of 10
fram
e
s); then, t
h
e
c
o
lor histogram
in RGB col
o
r s
p
ace is
used for
vide
o
fram
e
s. In
o
r
der
t
o
c
o
m
b
i
n
e spat
i
a
l
i
n
fo
r
m
at
i
on, eac
h
fram
e
i
s
di
vi
d
e
d i
n
t
o
3
x
3
b
l
ocks
an
d a
t
h
ree
-
dim
e
nsional histogram
is cr
eated for each
bloc
k. T
h
eses
9 histogram
s
are incorporated
to form
a f
eature
vector. Using this
extracte
d
feature
vect
or, the feat
ure
fra
m
e
m
a
trix A (u
s
u
ally spa
r
s
e
) is create
d
for t
h
e
vi
de
o.
The
n
,
S
V
D
i
s
p
e
r
f
o
r
m
e
d
on
m
a
t
r
i
x
A i
n
o
r
der t
o
ob
tain
m
a
trix
V so th
at each
vecto
r
co
lu
m
n
sh
ows a
fram
e
in the
refined feature
space. Af
ter t
h
at, the
nearest
cluster t
o
th
e refined feature
space is found, the
am
ount
o
f
t
h
i
s
cl
ust
e
r’s c
ont
e
n
t
i
s
cal
cul
a
t
e
d and t
h
i
s
val
u
e
i
s
used as a t
h
r
e
sh
ol
d f
o
r cl
u
s
t
e
ri
ng t
h
e
rem
a
i
n
i
n
g
fram
e
s. The
syste
m
selects a fram
e as the ke
y fram
e
from
each
cluster whi
c
h
is near
the c
e
nter of
the cluster.
3.
DETERMINATION OF T
H
E FUZ
Z
Y
BASED THRESHOL
D
VAL
UE
In the
increm
ental clusteri
ng, due
t
o
i
n
com
p
lete inform
atio
n
of the
accura
te num
b
er of c
l
usters, we
have t
o
select a thres
hol
d val
u
e in
order t
o
select or de
select a fram
e
to
place in cluste
rs. T
h
e determination
o
f
th
is th
reshold
with
ou
t consid
ering
th
e natu
re of d
a
ta will lead
to
a
h
i
gh
erro
r in
clu
s
tering
. Also
, th
is
threshol
d
will lead to problem
s in
partitioning clustering
due to the lack of accurate
information of clusters in
i
t
s
i
nput
. T
h
e
det
e
rm
i
n
at
i
on of a l
i
m
it
at
i
on fo
r sel
ect
i
ng
or
reject
i
ng a
fram
e
det
e
rm
ines t
h
e
p
o
we
r
of t
h
e
p
r
od
u
c
ed
su
m
m
arizatio
n
syste
m
. In
th
is article, we d
e
t
e
r
m
i
n
e t
h
e t
h
res
hol
d base
d o
n
i
n
f
o
rm
at
i
on va
l
u
es of
th
e ti
m
e
o
f
in
pu
t fram
e
an
d
also
th
e fram
e
h
i
sto
g
r
am
d
i
stan
ce to
th
e frame wh
ich
is in
th
e cen
ter o
f
clu
s
ters.
The i
ssue
of t
h
res
h
ol
d
det
e
r
m
i
n
at
i
on i
s
an
i
ssue wi
t
h
t
w
o i
n
put
s a
nd
o
n
e o
u
t
p
ut
w
h
i
c
h m
eet
s
t
h
e needs o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Vi
deo
S
u
m
m
a
r
i
zat
i
o
n
Ba
sed
on
a
F
u
zzy B
a
s
e
d
Incre
m
e
n
t
a
l
C
l
ust
e
ri
n
g
(
M
oni
re
h P
o
ur
na
zari
)
59
5
obs
er
vers a
n
d
users
by
de
si
g
n
i
n
g an
d c
r
eat
i
ng
36
r
u
l
e
s.
The
ove
ral
l
st
ruct
u
r
e o
f
t
h
e
pr
o
pose
d
m
e
t
hod
fo
r
det
e
rm
i
n
at
i
on
of t
h
e t
h
res
hol
d
val
u
e i
s
c
o
m
p
ri
se
d
of
5
f
u
z
z
y
l
a
y
e
rs w
h
i
c
h are
s
h
o
w
n i
n
fi
g
u
re
1.
I
n
fi
g
u
re
1,
n
i
s
t
h
e num
ber of i
n
put
s a
nd
R
i
s
t
h
e num
ber of r
u
l
e
s use
d
.
X
i
s are fu
zzy in
pu
ts; in
o
t
h
e
r words, th
e h
i
st
o
g
ram
d
i
fferen
ce an
d ti
m
e
d
i
fference ex
ist
b
e
tween
th
e inpu
t fram
e and the
center
of clust
e
rs.
The
hist
ogram
di
ffe
re
nce i
ndi
cat
es t
h
e pr
oxi
m
i
ty
of
th
e colo
r con
t
en
t of
two
inp
u
t
frames to
th
e cen
ter o
f
clu
s
ter an
d
the
ti
m
e
d
i
fferen
c
e is related
to
in
p
u
t
fram
es an
d
th
e fram
e
wh
ich
is in
th
e cen
ter o
f
clu
s
ter. Accord
ing
l
y, we can
st
udy
vi
de
o fr
a
m
es i
n
t
e
r
m
s of t
w
o a
s
pect
s
of
hi
st
o
g
ram
and t
i
m
e wi
t
h
the p
u
r
p
ose
of
m
a
t
c
hi
ng t
h
e c
r
eat
ed
summ
ary according to users
’
c
o
mments on time and c
o
ntent
.
3.1.
Fuz
z
y
Diagram Struc
t
ure
In the
first lay
e
r (
h
istog
r
am
diffe
re
nce, tim
e diffe
re
nc
e), i
n
puts are st
udi
ed and in the s
econd layer,
th
e m
e
m
b
ersh
ip
fun
c
tio
n
of each
inp
u
t
is calcu
l
ated
. In
th
e n
e
x
t
layer, t
h
e
extracted
rules ar
e applied.
These
rul
e
s ha
ve bee
n
p
r
o
p
o
sed
ba
sed o
n
di
f
f
ere
n
t
t
i
m
e
m
odes and hi
st
og
ra
m
di
ffere
nce i
n
o
r
de
r t
o
l
ead i
n
p
u
t
s
towa
rd a
n
a
p
pr
op
riate thre
sh
o
l
d com
p
reh
e
ns
ively
.
The
n
, i
n
the ne
xt layer, inputs are
aggregate
d
a
nd t
h
e last
l
a
y
e
r i
s
def
u
zz
i
f
i
cat
i
on st
age
whi
c
h i
s
per
f
o
r
m
e
d t
h
ro
u
g
h
an a
v
era
g
i
n
g
m
e
t
hod.
Y
i
s
are outputs of t
h
e fuz
z
y
sy
st
em
whi
c
h s
h
o
w
t
h
e desi
re
d
t
h
res
hol
d val
u
e.
Fi
gu
re
1.
F
u
zz
y
Di
ag
ram
3.
2. Fuz
z
i
fi
cati
on of
Inp
u
ts
The first step in fuzzy infe
rence system
s
is
receiving
inputs and de
te
rmining thei
r de
gree
of
me
m
b
ersh
ip
. In
th
e
fu
zzy l
o
gic, inputs are
always num
e
rical va
lu
es
wh
ich
are limit
ed
to
t
h
e referen
c
e
col
l
ect
i
on, i
.
e
.
t
a
ngi
bl
e
num
bers
fo
r i
n
di
vi
dual
s
a
n
d wi
t
h
o
u
t
n
o
r
m
a
li
zat
i
on.
In
or
de
r
t
o
det
e
rm
i
n
e fuzzy
i
n
t
e
rval
s,
we
u
s
ed t
h
e
kn
o
w
l
e
dge
o
f
m
a
xi
m
u
m
and m
i
nim
u
m
of i
n
p
u
t
di
f
f
e
rences
(
h
i
s
t
o
g
r
am
di
ffere
nce
o
f
0
– 5
,
an
d t
i
m
e
di
ffe
re
nce o
f
0
– 2
0
)
.
The
re
aso
n
o
f
det
e
rm
i
n
i
ng t
h
ese val
u
es wa
s t
h
at
t
h
e hi
ghest
hi
st
og
ram
di
ffe
re
nce b
e
t
w
een
i
n
put
fra
m
e
s i
s
not
m
o
re t
h
an
5 a
n
d al
so,
by
i
n
vest
i
g
at
i
ng t
h
e
vi
deo
of
col
l
ect
i
o
n
s
use
d
, i
t
was
d
e
term
in
ed
th
at
th
e m
a
x
i
m
u
m
t
i
m
e o
f
selectin
g
a
fram
e
in
con
s
ecu
ti
ve fram
e
s is less th
an
2
0
seco
nd
s. In
th
is article, we u
s
ed
th
e trian
g
u
l
ar
fu
zzy me
m
b
ersh
ip
fu
n
c
tion
.
Relatio
n
(1
) is t
h
e triang
u
l
ar m
e
mb
ersh
i
p
fun
c
tion
in
which
a, b
and
c are th
e lo
cation
s
of trian
g
u
l
ar fun
c
tion
on
X in
pu
ts. In
this relatio
n
,
th
e fu
zzy
i
n
fere
nce sy
st
e
m
reci
eves i
n
p
u
t
s
an
d det
e
rm
i
n
s t
h
e m
e
m
b
ershi
p
de
gree
of
i
nput
s
fo
r eac
h f
u
zzy
set
.
Fi
gu
res 2
an
d 3 sh
ow
t
h
e d
i
gr
am
s o
f
f
u
zzy
m
e
m
b
er
sh
ip in
pu
t
f
o
r
th
e t
w
o fu
zzy inpu
s.
(
;
,
,
)
m
a
x
(m
i
n
(
,
),
0)
xa
c
x
ux
a
b
c
ba
c
b
(1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
59
3
–
60
2
59
6
Fi
gu
re
2.
M
e
m
b
ers
h
i
p
f
u
nct
i
o
n
of
hi
st
o
g
r
am
di
ffe
re
nce
bet
w
een
t
w
o
fram
e
s
Fi
gu
re
3.
M
e
m
b
ers
h
i
p
f
u
nct
i
o
n
of
t
i
m
e
di
ffe
r
e
nce
bet
w
ee
n t
w
o
f
r
am
es
Al
l
i
nput
va
ri
abl
e
s sh
oul
d becom
e
fuzzy
usi
ng m
e
m
b
ershi
p
fu
nct
i
o
ns.
W
e
use t
r
i
a
ng
ul
ar m
e
mbers
h
i
p
fun
c
tion
s
for
t
h
e fu
zzification
o
f
two
inpu
ts.
3.
3. Func
ti
on
of
Fuz
z
y
Oper
at
ors
Aft
e
r
f
u
zzi
fi
ca
t
i
on
of
i
n
put
s
,
t
h
e acc
uracy
d
e
gree
o
f
eac
h
com
pone
nt
of
assum
e
d sect
i
o
ns
(i
n
put
s
)
will be
determ
i
n
ed. Relation (2) is the c
o
m
putation function of
fuzzy
interface.
α
j
=
(2)
N
adj
is th
e equiv
a
len
t
co
efficien
t wh
ich
in th
is article we
co
nsid
er it equ
a
l to
n
/
4
.
µ
ij
is the fuzzy
me
m
b
ership coefficient.
3.
4. Ap
pl
yi
n
g
the
Im
pl
i
c
ati
o
n
Me
th
od
Th
e
resu
lt sect
io
n
is a d
e
termin
ed
fuzzy set
b
y
th
e m
e
m
b
ersh
ip
fu
n
c
tion
.
Th
e
p
r
o
cess inp
u
t
sho
w
s a
n
u
m
b
e
r and
its ou
tpu
t
ind
i
cates a fu
zzy set.
Accord
ing
to relatio
n
(3
),
90
th
j
ou
tpu
t
is equ
a
l to
:
(3
)
3.
5. Ou
tpu
t
s Ag
gre
g
a
t
i
o
n
As in a fuzzy interface
, decisi
ons a
r
e m
a
de b
a
sed on
th
e evalu
a
tio
n
of all
ru
les,
we in
tegrate th
em in
th
is layer and acco
rd
ing
t
o
rel
a
tio
n
(4),
90
th
k
out
put
i
s
eq
u
a
l
t
o
:
(4
)
whe
r
e W
jk
is th
e
weigh
t
of each
ru
le
wh
ich is app
lied
to
t
h
e
value
obtained from
the assum
e
d section.
We
consider the
weight
of each rule equal t
o
1.
3.
6.
Ou
tpu
t
Di
agr
am
a
nd T
h
reshol
d
De
ter
m
i
n
ati
o
n
In
or
der t
o
re
g
u
l
a
t
e
sum
m
arizat
i
on r
u
l
e
s, w
e
det
e
rm
i
n
e a
speci
fi
ed
hi
st
o
g
ram
di
ffere
nc
e and t
i
m
e
di
ffe
re
nce f
o
r
ext
r
act
i
n
g t
h
e
bo
r
d
er o
f
sh
ot
s or di
ffe
rent
clu
s
ter
s
(
t
he clu
s
ter
i
ng
bor
de
r is used
fo
r fr
am
es).
Bo
rd
er d
e
term
in
atio
n is
d
o
n
e
as fo
llo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Vid
e
o Sum
ma
riza
tio
n Based
o
n
a Fu
zzy Ba
sed
I
n
cre
m
e
n
t
a
l
C
l
ust
e
ri
n
g
(
M
oni
re
h P
o
ur
na
zari
)
59
7
The
hi
st
o
g
ram
di
ffe
re
nce
of
v
a
ri
o
u
s cl
u
s
t
e
rs
i
s
st
udi
e
d
a
n
d
t
h
e m
a
xim
u
m
di
ffe
re
nt
bet
w
e
e
n cl
u
s
t
e
rs
i
s
sel
ect
ed. Al
so, det
e
rm
i
n
at
ion
of a val
i
d
t
i
m
e
l
i
m
i
t i
s
done usi
n
g t
h
e t
i
m
e di
ffere
nce bet
w
ee
n t
h
e cent
r
al
fram
e
an
d
t
h
e l
a
st on
e. Th
e m
e
th
od
o
f
reg
u
l
atin
g
inpu
t limit
s is shown in
fo
rm
u
l
a 5
.
1
11
((
1
/
)
*
(
(
)
(
)))
cc
I
n
t
er
v
a
l
h
i
s
t
o
g
r
a
m
M
A
X
N
cen
t
e
r
i
cen
t
e
r
j
1
(_
_
_
_
)
)
c
T
i
m
e
f
r
a
m
e
m
o
v
i
e
M
A
X
C
E
N
T
E
R
T
I
M
E
F
I
N
AL
F
R
AM
E
C
L
U
S
T
E
R
(5
)
In
fi
g
u
r
e
s
4 a
n
d
5, m
e
m
b
ersh
i
p
f
u
n
c
t
i
o
n
o
f
t
h
e t
h
resh
ol
d
v
a
l
u
e an
d t
h
e
re
l
a
t
i
onshi
p
bet
w
een
i
n
put
s
(tim
e, histo
g
ra
m
difference
)
a
n
d
the t
h
res
h
old a
r
e selected,
respectively.
Fi
gu
re
4.
M
e
m
b
ers
h
i
p
f
u
nct
i
o
n
of
t
h
e t
h
res
h
ol
d
val
u
e
Fi
gu
re
5.
R
u
l
e
d c
o
n
s
i
d
er
e
d
for
the fuzzy syste
m
In
or
de
r t
o
ac
hi
eve t
h
e
desi
red
res
u
l
t
fo
r
det
e
rm
i
n
at
i
on of t
h
e t
h
re
sh
ol
d an
d a
p
p
r
op
r
i
at
e out
p
u
t
s
whi
c
h ha
ve l
e
s
s
err
o
rs a
n
d a
hi
g
h
er s
u
m
m
ar
i
zat
i
on pe
r
cent
a
ge,
we use rules and principl
es selected by
users
fo
r a
fram
e
. Th
ere are
di
ffe
ren
t
sty
l
es
in
th
ese prin
cip
l
es as fo
llo
ws:
The sel
ect
e
d
f
r
a
m
e
s
m
u
st
not
fol
l
o
w eac
h
ot
her
.
Fram
es
m
u
st have a c
o
m
p
lete concept a
n
d m
eaning.
All selected
fra
m
e
s
m
u
st show th
e
film
’s concept a
n
d m
eaning.
The selecte
d
fra
m
e
s
m
u
st have less col
o
r sim
ilarity.
Th
ey m
u
st g
u
e
ss sim
i
lar lo
catio
n
s
in
d
i
fferent situ
atio
n
s
.
Ti
m
e
with
co
l
o
r sim
i
larity
m
u
st no
t b
e
rev
i
ewed in
a lin
ear m
a
n
n
e
r.
C
ont
e
n
t
l
e
ss fl
a
s
hes a
n
d
fram
e
s m
u
st
be rem
oved
.
The selecte
d
fra
m
e
s with m
i
nor differe
n
ces
m
u
st
not
be
pl
a
ced i
n
an
ot
h
e
r
cl
ust
e
r.
According to t
h
ese
pri
n
ciples
, we
re
gulate a
rela
tions
hip
between tim
e a
n
d col
o
r
differences
of t
h
e
in
pu
t fram
e
an
d
fram
e
s in
th
e cen
ters of clusters in
or
d
e
r to
satisfy g
e
n
e
ral co
mmen
t
s. Resu
lts o
f
th
ese ru
les
i
ndi
cat
e t
h
at
t
h
e t
h
res
h
ol
d
det
e
rm
i
n
at
i
on
si
gni
fi
ca
nt
l
y
sim
u
l
a
t
e
s users
’
c
o
m
m
e
nt
s whi
c
h a
r
e s
h
o
w
n
i
n
Fi
gu
re 5.
4.
CLUSTE
RI
N
G
In
order to cl
uster fram
es
whose feat
ure
s
were
e
x
trac
ted, we
use the increm
ental clustering
alg
o
rith
m
u
s
ed in
referen
ce
[2
] wh
ich
h
a
s
group
ed
simila
r fram
e
s; then,
one
re
prese
n
tat
i
ve fram
e is se
lected
in each cluster as the key frame. Th
e extrac
ted feature
from color hist
ogram fram
e
s
is
base
d on c
o
nverting
RGB color s
p
ace into
HSV one. T
h
is conve
r
sion
oc
curs beca
use
HSV col
o
r s
p
ace is close to the
un
de
rst
a
n
d
i
n
g
of t
h
e
h
u
m
a
n vi
sual
sy
st
em
. Thi
s
o
p
erat
i
o
n st
art
s
wi
t
h
a
n
i
n
p
u
t
f
r
am
e. Ne
xt
, t
h
e
hi
st
og
ram
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
59
3
–
60
2
59
8
diffe
re
nce and tim
e
of the ne
w fram
e
are ca
lculated by
t
h
e hi
st
og
ram
di
fference a
nd t
i
m
e of t
h
e fram
e
whi
c
h
is in the ce
nter of
form
ed clusters. T
h
e
n
, these two
data
a
r
e inse
rted i
n
t
h
e rules
form
ed in t
h
e
fuzzy
s
y
ste
m
an
d th
ese ru
les estab
lish a
nu
m
b
er fo
r u
s
.
Th
is
n
u
m
b
e
r
wh
ich
sh
ows th
e thresh
o
l
d
valu
e sp
ecifies
th
at th
e
i
n
p
u
t
fram
e
shoul
d be i
n
se
rt
ed i
n
w
h
i
c
h cl
ust
e
r. I
f
t
h
e n
u
m
b
er i
s
m
o
re t
h
an ot
he
r cl
ust
e
rs
’ l
i
m
i
t
s
,
a ne
w
clu
s
ter is fo
rm
ed
and
it is inserted
i
n
t
h
at clu
s
ter. Th
en, t
h
e tim
e
and hist
ogram
of
t
h
e
fram
e are rec
o
rded in
the center of t
h
e new cl
uster. If a fr
am
e belongs to
one
of the
pre
v
ious
clusters, a
ne
w center is c
r
eated for
th
at clu
s
ter acco
rd
ing
to
the h
i
sto
g
ram
rate o
f
cluste
rs a
nd t
h
e aim
is
that cluster ce
nters
have t
h
e
leas
t
diffe
re
nce wit
h
ot
her existing fram
es in the sam
e
cluste
r so that at the
e
nd
of s
u
mmarization,
we ca
n select
key fram
es without a
dditiona
l
calculations
. In this i
n
crem
ental cluste
ri
ng m
e
thod,
we
st
ore central fra
m
e
s
in
th
e m
e
m
o
ry u
s
in
g
i
n
d
e
x
i
ng
(h
istog
r
am
rate an
d ti
m
e
o
f
th
e in
pu
t fram
e), we co
m
p
are the h
i
stog
ram
rate and
th
e tim
e o
f
inp
u
t
fram
es with
th
e inform
a
tio
n
o
f
cen
t
ral fram
e
and e
v
aluate their di
ffe
r
ence
in order to
det
e
rm
i
n
e t
h
at t
h
ey
are i
n
sert
ed i
n
whi
c
h f
u
zzy
i
n
t
e
rval
.
One
of t
h
e ad
vant
a
g
es o
f
t
h
i
s
m
e
t
hod i
s
when t
h
e
t
i
m
e
of fi
l
m
s is l
o
n
g
an
d we
have a l
i
m
i
t
e
d m
e
m
o
ry
. Figu
re 6 s
h
o
w
s
t
h
e st
udy
o
f
cl
ust
e
ri
n
g
.
It
i
s
wo
rt
h
men
tio
n
i
ng
th
at in
th
is in
cremen
tal clu
s
terin
g
, th
ere is
no
need to c
o
m
p
are the i
n
pu
t frame with
all clu
s
tered
fram
e
s an
d
it is co
m
p
ared
ju
st with
on
e fram
e wh
ich
is
in
th
e cen
ter
o
f
t
h
e clu
s
ter.
In
fact, th
is fra
m
e
is
in
serted
i
n
on
e o
f
th
e ex
isting
clu
s
ters u
s
i
n
g
th
e
d
e
term
in
atio
n
of fu
zzy
h
i
gh
prob
ab
ility o
r
a n
e
w cluster is
creat
ed.
Wi
t
h
t
h
i
s
cl
ust
e
ri
ng
whi
c
h i
s
base
d
on t
h
e f
u
zzy
l
ogi
c, c
o
m
put
at
i
onal
spe
e
d i
s
i
n
creasi
n
gl
y
hi
ghe
r
t
h
an
STIM
O
m
e
t
hod
w
h
i
c
h
i
s
base
d
o
n
t
h
e i
n
c
r
em
ental
cl
ust
e
ri
n
g
a
n
d
l
eads t
o
a
hi
g
h
e
r
acc
ur
acy
i
n
su
mm
arizat
io
n in
less tim
e,
and
th
is clai
m
is d
i
scu
ssed
in th
e ev
al
u
a
tio
n section
.
Th
e
reaso
n
o
f
t
h
e
co
m
p
u
t
atio
n
a
l co
m
p
lex
ity an
d
th
e greater time o
f
STIM
O alg
o
rith
m
i
m
p
l
e
m
en
tatio
n
is th
at it
m
a
k
e
s a
fram
e
to
b
e
co
m
p
ared
with all o
t
her fram
es d
u
e
to
th
e
p
u
r
pose
of cl
ust
e
ri
ng
wi
t
h
l
e
ss f
r
am
e err
o
r
w
h
i
c
h
i
s
no
t
per
f
o
r
m
e
d i
n
t
h
e
pr
o
pose
d
m
e
t
h
o
d
.
Fig
u
r
e
6
.
Study o
f
in
ser
tin
g a n
e
w
f
r
a
m
e
in
th
e cluster
4.
1. Pseud
o Code of
Fuz
z
y
and
In
cre
mental Clus
tering
Algorithm
In
fi
gu
re
7,
t
h
e
pse
u
do
co
de
r
e
l
a
t
i
ng t
o
t
h
e
e
x
t
r
action
o
f
fuzzy th
resho
l
d
an
d in
crem
en
tal alg
o
rith
m
i
s
obse
r
vabl
e.
In t
h
e i
m
pl
ement
a
t
i
on se
q
u
e
n
ce, t
h
e t
h
re
sho
l
d
is ex
tracted
fo
r each
cl
uster; th
en
, t
h
e
li
mits
specified in t
h
e
increm
en
tal algorithm
are us
ed.
Fig
u
re
7
.
Pseud
o
cod
e
of th
e
in
crem
en
tal cl
u
s
tring
al
g
o
rithm
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Vi
deo
S
u
m
m
a
r
i
zat
i
o
n
Ba
sed
on
a
F
u
zzy B
a
s
e
d
Incre
m
e
n
t
a
l
C
l
ust
e
ri
n
g
(
M
oni
re
h P
o
ur
na
zari
)
59
9
5.
D
I
AGRAM
OF THE PR
OP
OSED METH
OD
The
di
ag
ram
in
fi
g
u
re
8
sh
o
w
s t
h
e
pr
op
os
ed m
e
t
hod
. I
n
t
h
i
s
m
e
t
hod, t
h
e
fi
lm
i
s
con
v
ert
e
d i
n
t
o
a
fram
e
;
t
h
en, fe
at
ures are e
x
t
r
act
ed fr
om
fram
e
s and t
h
e m
e
t
hod o
f
f
u
z
z
y
based t
h
res
hol
d det
e
rm
i
n
at
i
on i
s
im
ple
m
ented.
At this sta
g
e, t
h
e
num
be
r
of c
l
usters a
n
d thre
shol
d
of each c
l
uster a
r
e s
p
eci
fied, and
finall
y, the
in
crem
en
tal al
g
o
rith
m
is ap
p
l
ied
to
th
e
ex
tr
ac
te
d
f
e
a
t
u
r
e fr
ame
s
.
Fi
gu
re
8.
Di
a
g
r
a
m
of t
h
e
pr
op
ose
d
m
e
t
hod
In
o
r
de
r t
o
rem
ove
m
eani
ngl
e
ss f
r
am
es, t
h
e
st
anda
rd
d
e
viat
ion
of the
feat
ure
vector is
use
d
s
o
t
h
at if
t
h
e feat
u
r
e
vec
t
or
of
on
e f
r
a
m
e has t
h
e st
a
nda
r
d
de
vi
at
i
o
n o
f
ze
ro
or cl
ose t
o
ze
r
o
, i
t
sho
u
l
d
be
rem
ove
d a
n
d
sho
u
l
d
not
e
n
t
e
r t
h
e
st
age
o
f
cl
ust
e
ri
ng
. M
eani
n
gl
ess
fra
m
e
s are
obse
r
ved
as
o
p
aq
ue
vi
de
o
f
r
am
es, fram
e
s
wi
t
h
fl
as
h,
bl
a
c
k o
u
t
p
ut
s
or
whi
t
e
fl
as
hes
and al
s
o
, t
h
ey are create
d
in
im
ages in
th
e
form
o
f
no
ise wh
ich
create som
e
problem
s for clustering an
d i
n
c
r
ease the
cluste
ring error.
6.
EVAL
UATI
O
N
In
order t
o
eva
l
uate the proposed m
e
thod,
we use
the e
v
aluation m
e
thod used in
refe
renc
e [1] whic
h
per
f
o
r
m
s
t
h
e eval
uat
i
o
n
base
d
on
us
ers
’
c
o
m
m
e
nt
s. Thi
s
e
v
al
uat
i
o
n m
e
t
hod
can
be
o
b
se
rve
d
i
n
fi
gu
re
9.
Fi
gu
re
9.
The
val
u
e e
v
al
uat
i
o
n m
e
t
hod
use
d
(f
rom
v s
u
m
m
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
59
3
–
60
2
60
0
mA
S
A
US
n
CUS
n
In t
h
is e
v
aluation,
we consi
d
e
r
a fram
e as the
m
a
t
c
hed f
r
a
m
e whi
c
h
has
l
e
ss differe
n
ce
with the
ke
y
fram
e
su
mmarized
b
y
t
h
e
u
s
er.
In th
e related
articles, th
i
s
rate is con
s
idered to
b
e
0
.
5
wh
ich
is t
h
e
best rate
fo
r c
o
n
f
i
r
m
i
ng
t
h
e di
ffe
rence
det
e
rm
i
n
at
i
on.
6.1. Summ
a
ri
z
a
tion Accur
a
cy Rate
In
th
is article, th
e ev
alu
a
tion
was
p
e
rfo
r
med
b
a
sed
on th
e si
m
ilarit
y
rate b
e
tween su
mm
aries
pr
o
duce
d
by
u
s
ers a
n
d t
h
e
o
u
t
p
ut
o
f
c
o
m
p
ared
al
g
o
ri
t
h
m
s
. Accord
i
n
g to
th
e co
m
p
ariso
n
of
resu
lts ex
tracted
fr
om
previ
o
us
m
e
t
hods
wi
t
h
t
h
e pr
o
pose
d
m
e
t
hod, t
h
e m
e
t
hod
of t
h
i
s
art
i
c
l
e
sho
w
s t
h
e hi
g
h
est
rat
e
of
summ
arization accuracy
(bec
ause it wa
s ra
nked base
d
on
users’ ou
tput). The
e
v
aluation
process
is pe
rform
e
d
so that the
propos
ed al
gor
ithm is co
m
p
ared with
th
e
fra
m
e
summ
arized by use
r
s acc
o
r
d
i
ng
to
t
h
e conf
or
m
i
ty
of t
h
e
out
put
fram
e
. The c
o
nf
orm
i
t
y
i
s
det
e
rm
i
n
ed t
h
ro
u
g
h
t
r
i
a
l
an
d e
r
r
o
r a
n
d
we c
oncl
ude
d t
h
at
i
f
t
h
e
hi
st
o
g
ram
di
ff
erence
bet
w
ee
n
out
put
fram
e
s o
f
t
h
e
al
g
o
r
i
t
h
m
and
use
r
s’
sum
m
ary
is l
e
ss t
h
a
n
0.
5, t
h
e
sum
m
ari
zati
o
n
has c
o
n
f
o
r
m
i
ty
;
ot
her
w
i
s
e, t
h
e s
u
m
m
a
ri
zati
on
was
do
ne
wi
t
h
er
ro
r.
Fi
gu
re
10 s
h
ow
s
t
h
i
s
process
.
The a
ccuracy rate is CUS
A
whi
c
h
i
s
obser
va
bl
e i
n
rel
a
t
i
on 6
.
N
mA
S
i
s
t
h
e num
ber o
f
m
a
t
c
hed key
fram
e
s i
n
t
h
e pro
d
u
ced s
u
m
m
ary
t
h
ro
u
gh t
h
e pr
op
ose
d
m
e
t
h
o
d
an
d n
US
i
s
t
h
e num
ber of
key
fram
e
s i
n
user
s
’
sum
m
ary.
(6
)
Fi
gu
re 1
0
. Di
f
f
e
rence
bet
w
ee
n pr
o
duce
d
s
u
m
m
a
ri
ce
Acco
r
d
i
n
g t
o
f
i
gu
re 1
0
, we h
a
ve 2 sum
m
arizat
i
on err
o
rs am
ong 11
out
p
u
t
fram
e
s, i
.
e. 2 key
fram
e
s
p
r
od
u
c
ed
b
y
the pr
opo
sed m
e
t
h
od
ar
e
no
t availab
l
e in
th
e su
mmar
y
p
r
o
duced
b
y
user
s an
d ar
e no
t co
n
s
id
er
ed
by
user
s.
Acc
o
r
d
i
n
g
t
o
res
u
l
t
s
of
pre
v
i
o
us
m
e
t
hods
(
S
TI
M
O
,
V
S
UM
M
,
OV
,
an
d
DT
) a
n
d t
h
e
p
r
o
p
o
s
e
d
m
e
t
hod
o
n
si
m
i
l
a
r dat
a
set
s
, t
h
e p
r
op
ose
d
m
e
t
hod
has
1
5
%
su
peri
ori
t
y
i
n
t
h
e
fi
nal
s
u
m
m
a
ri
zati
on
w
h
i
c
h i
s
sho
w
n i
n
t
a
bl
e
1. The
vi
de
os
un
der st
u
d
y
a
r
e 5
0
vi
de
os f
r
o
m
vi
deo seri
e
s
of O
p
e
n
Vi
d
e
o. Al
l
vi
deos
are i
n
th
e for
m
at o
f
MPEG
-1
, an
d h
a
v
e
t
h
e
r
e
solu
tio
n of
35
2x4
0 an
d th
e
speed
o
f
3
0
fr
ames p
e
r seco
nd
.
Th
e
selected vide
os are divi
ded i
n
to
di
ffe
rent
groups (doc
um
e
n
tary, instruct
i
onal,
daily, hi
storical and speech
vi
de
os) a
n
d t
h
ei
r t
i
m
e
i
s
vari
abl
e
fr
om
1 t
o
4 m
i
nut
es. 50
users
ha
ve p
r
o
duce
d
user s
u
m
m
a
ri
zati
ons
and eac
h
user
dealt with
5 vide
os
,
i.e. each vide
o has
5
s
u
mm
arie
s
whic
h
a
r
e produce
d
by 5
differe
n
t users
.
In othe
r
wo
rd
s, 2
5
0
vi
deo s
u
m
m
ari
e
s were p
r
od
uc
ed m
a
nual
l
y
.
Yo
u can see a
l
l
vi
deos an
d
sum
m
ari
e
s of users at
h
ttp
://www.n
pd
i.d
c
c.ufm
g
.b
r/
VSUMM
.
Table
1. T
h
e
a
v
era
g
e
of s
u
mmarization acc
uracy
rate
Pr
oposed
M
e
thod
STIMO
VSUM
M
DT
OV
Su
mm
a
r
ization
M
e
thods
91.
47%
72%
85%
53%
70%
Mean accur
a
c
y
rate
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Vid
e
o Sum
ma
riza
tio
n Based
o
n
a Fu
zzy Ba
sed
I
n
cre
m
e
n
t
a
l
C
l
ust
e
ri
n
g
(
M
oni
re
h P
o
ur
na
zari
)
60
1
Fi
gu
re
1
1
.
Vi
d
e
o s
u
m
m
ari
e
s
of
di
f
f
ere
n
t
a
p
pr
oac
h
es
of
t
h
e
vi
de
o
D
r
i
f
t
Ice
as a
Ge
ol
o
g
i
c
Age
n
t
,
segm
ent
8
(available at
Open Vi
deo Project)
6.
2. Summ
a
ri
z
a
ti
on
E
r
r
o
r Ra
te
In
or
de
r t
o
det
e
rm
i
n
e t
h
e su
m
m
a
ri
zati
on e
r
r
o
r
,
we
di
vi
d
e
u
n
m
a
t
c
hed f
r
am
es of t
h
e
p
r
o
p
o
sed
m
e
t
hod a
nd
users
’
s
u
m
m
arizat
i
on i
n
t
o
t
h
e
num
ber
of
key
fram
e
s.
In
ot
he
r
wo
rd
s, we
sh
o
u
l
d
have
t
h
e m
i
nim
u
m
nu
m
b
er
of
u
n
m
a
t
c
hed fram
e
s, i
.
e. o
u
r
al
g
o
ri
t
h
m
’
s out
pu
t
sho
u
l
d
n
o
t
hav
e
m
a
ny
key
fram
e
s and fram
e
s pr
ovi
ded
fo
r t
h
e o
u
t
p
ut
m
u
st
ha
ve t
h
e m
a
xi
m
u
m
conf
o
r
m
i
t
y
wi
t
h
user
s’
co
m
m
e
nt
s. Acco
rdi
n
g
t
o
t
h
e e
r
ro
r
of
p
r
e
v
i
o
us
m
e
t
hods a
n
d
t
h
e
pr
o
pose
d
o
n
e, i
n
t
a
bl
e
2
we ca
n
see t
h
at
t
h
e er
ro
r o
f
t
h
e p
r
o
pos
ed m
e
t
hod
i
s
10% l
e
ss t
h
an t
h
e p
r
e
v
i
o
u
s
m
e
t
hods
. Th
e err
o
r rat
e
i
s
cal
l
e
d
CUS
E
wh
ich is
d
e
fi
n
e
d as
fo
ll
o
w
s:
(7
)
whe
r
e
݊
ௌ
i
s
t
h
e
n
u
m
b
er o
f
unm
at
ched
key
fram
e
s i
n
t
h
e s
u
m
m
ary
p
r
o
d
u
ced
b
y
t
h
e p
r
op
ose
d
m
e
t
hod.
Table
2. T
h
e
a
v
era
g
e
of s
u
mmarization error rate
Pr
oposed
m
e
thod
STIMO
VSUM
M
DT
OV
Su
mm
a
r
ization
m
e
thods
25%
58%
38%
29%
57%
Mean
er
ro
r
rate
According
to results of
t
h
e pre
v
ious
m
e
thods
,
the proposed
m
e
thod ha
s
a
significant
superiority
whi
c
h i
s
res
u
l
t
e
d f
r
o
m
usi
n
g t
h
e f
u
zzy
e
xpe
rt
m
e
t
hod
fo
r t
h
e t
h
res
h
ol
d
det
e
rm
i
n
at
i
on.
T
h
e
key
fram
e
s
pr
o
duce
d
by
t
h
e p
r
o
p
o
sed
m
e
tho
d
a
n
d
ot
he
r t
echni
que
s are
s
h
o
w
n i
n
fi
gu
re
11
o
n
t
h
e
vi
de
o
of
i
ce d
r
i
f
t
.
6.
3. C
o
mp
ari
s
on of
Impl
em
ented
Res
u
l
t
s B
a
sed on
di
f
f
e
rent Met
h
o
d
s
The
propose
d
m
e
thod has
a significa
nt superiority com
p
ared to
ot
he
r m
e
thods due to t
h
e accuracy
and
er
r
o
r
rat
e
s.
Thi
s
com
p
ari
s
on
wa
s
do
ne
b
e
t
w
een
t
h
e
p
r
o
pos
ed
m
e
t
hod
and
e
x
i
s
t
i
n
g
m
e
t
h
o
d
s i
n
refe
r
e
nces
[1]
,
[
2
]
,
[
8
]
and
[9]
.
T
h
e ac
curacy
a
nd e
r
r
o
r
rat
e
s o
f
t
h
e
afo
r
em
ent
i
oned m
e
t
hod are
fi
gu
res
12 a
n
d 1
3
,
respectively.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
59
3
–
60
2
60
2
Figure
12. T
h
e
com
p
arison
be
tween t
h
e a
v
erage
perc
e
n
t of the
accuracy ra
te
of propose
d
m
e
thods
and
pre
v
i
o
us t
e
c
hni
que
s
7.
CO
NCL
USI
O
N
Vi
de
o sum
m
ari
zat
i
on i
s
on
e of t
h
e m
o
st
i
n
t
e
rest
i
ng
issu
es in
th
e film
an
d
adv
e
rtising
i
n
du
stry and
di
ffe
re
nt
m
e
t
h
ods
w
e
re
desi
g
n
ed
f
o
r
t
h
i
s
i
s
s
u
e.
Acc
o
r
d
i
n
g
t
o
res
u
l
t
s
obt
ai
ned
t
h
ro
u
g
h
p
r
evi
o
us m
e
t
hod
s, w
e
cann
o
t
pe
rf
o
r
m
sum
m
a
ri
zat
ion
wi
t
h
ce
rt
ai
nt
y
.
The
r
ef
or
e,
we nee
d
a
n
e
xpe
rt
m
e
t
hod
whi
c
h can
su
m
m
a
ri
ze
vi
de
os acc
urat
el
y
and cl
ose t
o
users
’
c
o
m
m
e
nt
s a
n
d
t
h
i
s
i
s
p
o
ssi
bl
e t
h
r
o
u
g
h
t
h
e
f
u
zzy
m
e
t
h
o
d
.
Fu
rt
he
r
m
ore,
we
need
a
n
a
p
pr
o
p
ri
at
e c
o
m
put
at
i
o
nal
s
p
e
e
d i
n
fi
l
m
sum
m
a
ri
zati
on
whi
c
h
we
ha
v
e
achi
e
ved
t
h
i
s
p
u
r
p
ose
t
h
r
o
u
g
h
i
m
pro
v
i
n
g t
h
e
cl
ust
e
ri
n
g
m
e
t
hod.
Figure
13. T
h
e
com
p
arison
be
tween t
h
e a
v
erage
perce
n
t
of
t
h
e
er
ro
r rat
e
o
f
pr
op
ose
d
m
e
tho
d
s an
d pre
v
i
ous
techniques
REFERE
NC
ES
[1]
Avila S., Lop
e
s
A., Lu
z Jr. A
.,
Araujo A., 2011
,
VSUMM:
A mechanism design
to produce sta
t
ic vid
e
o summaries
and a novel evaluation method
, Pattern Recognition Lette
rs
, Vol (
32), pp
. 56-68
.
[2]
Furini M., Geraci F., Monta
nger
o
M., Pellegrini
M., 2010,
STIM
O:STIll and MOving vid
e
o storyboard for the web
scenario
, Multimedia Too
l
s Ap
pl., Vol (46)
, pp
. 47-69.
[3]
Yang Wu, Wu
Yuan, Wu Zhon
gru, 2004,
Forecast Model Study Base
d on Fuzzy
Neural Network
, Water Resources
and Power, vol (
22), pp
. 63-65
.
[4]
Zhang X.Y.
, Wang P., 2009
,
I
m
proved T-S Fuzzy Neural Netw
ork in
Appl
icat
i
on of Spee
c
h R
e
cognition
Syste
m
,
Computer Engin
eering
and
Appl
ications, vol (
45)
, pp. 246-248.
[5]
Zhuang, Y., Rui, Y., Huang,
T.S., Mehrotra, S., 1
998,
Adaptive
key frame ex
traction using unsupervised clustering
,
In: Proc. IEEE I
n
ternet. Conf. on
Imag
e Processing (ICIP), vo
l. 1
,
pp
. 866–870
.
[6]
Hanjalic, A.,
Zhang, H., 1999,
An integrated
scheme for automated
vid
e
o a
b
straction based on unsupervised
clus
ter
-
va
lidi
t
y a
nalys
is
, I
E
EE Tr
ans. Cir
c
uits
S
y
s
t
ems Video
Technolog
y
9 (8)
,
1
280–1289.
[7]
Gong, Y.,
Liu
,
X
., 2000
,
V
i
deo
summarization using singular valu
e decomposition
, In: Proc. IEEE I
n
ternat. Conf
. o
n
Com
puter Vision and Pattern Recogni
tion (CVPR). IEEE Co
m
puter Societ
y
,
L
o
s Alam
itos, C
A
, USA, pp. 21
74–
2180.
[8]
Mundur P., Rao Y., Yesha Y.,
2006,
Keyframe-
based video summariza
tion using Delaunay clus
tering
, Int J
D
i
g
i
t
Lib, Vol (6)
,
pp
.
219-232.
[9]
DeMenthon, D., Kobla, V.
, Do
ermann, D., 19
98,
Vid
e
o summarization by cur
ve simpli
fica
tion
. In: P
r
oc
. A
C
M
Internat. Conf. o
n
Multim
edi
a
. NY, USA, pp. 211
–218.
Evaluation Warning : The document was created with Spire.PDF for Python.