TELKOM
NIKA
, Vol. 11, No. 10, Octobe
r 2013, pp. 5
703 ~ 5
710
ISSN: 2302-4
046
5703
Re
cei
v
ed Ma
rch 2
9
, 2013;
Re
vised June
30, 2013; Accepte
d
Jul
y
1
4
, 2013
Fast Intra Prediction Mode Decision Algorithm for
HEVC
Mengmeng Zhang*
1
, Jia
n
feng Q
u
1
, Huihui Bai
2
1
Colle
ge of Info
rmation En
gin
e
e
rin
g
, North Ch
ina U
n
ivers
i
t
y
of
T
e
chnol
og
y,
Beiji
ng, Chi
n
a
2
Institute of Informatio
n
Scien
c
e, Be
iji
ng Jia
o
t
ong Un
iversit
y
, Beijin
g, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: muchmen
g
@
126.com
A
b
st
r
a
ct
T
h
is p
aper
pr
o
poses
a
nov
el
fast intra-
pred
iction
al
gorith
m
th
at ex
plo
i
ts the S
o
b
e
l
op
erator
t
o
repl
ace th
e H
a
da
mar
d
transf
o
rm use
d
i
n
th
e Ro
ug
h Mo
de
Decis
i
o
n
(RM
D
) proc
ess of
i
n
tra pr
edicti
on
i
n
High
Efficie
n
cy
Vid
e
o
Co
di
ng
(HEVC). F
i
rst,
the So
be
l
oper
at
or is
us
ed
to
calcul
ate t
he v
e
ctor d
i
rectio
n
o
f
each
pixe
l. A judg
ment is the
n
made
on w
h
i
c
h pred
ic
tio
n
mo
de th
e vect
or bel
on
gs to, and
a histo
g
ra
m i
s
app
lie
d in th
e sche
m
e to g
e
n
e
rate the statis
tics of
predicti
on
mod
e
s for each pr
ed
ictio
n
unit. F
i
na
lly, t
h
e
pred
iction
mo
d
e
ca
ndi
dates
a
r
e p
l
ace
d
in th
e ca
ndi
dat
e
lis
t for the r
a
te-d
istortion
opti
m
i
z
a
t
i
o
n
proc
ess
.
Experi
m
ental
r
e
sults s
how
th
at our
pro
pos
e
d
al
gor
ith
m
for
RMD si
gnific
a
n
t
ly red
u
ce
s the com
p
lexity
of the
enco
der w
i
th a
n
accepta
b
l
e
d
egra
datio
n of
q
uality a
nd BD-r
ate co
mpar
ed
w
i
th HM7.0.
Ke
y
w
ords
: Hi
gh Efficiency V
i
de
o Cod
i
n
g
, Intra, Prediction
Modes,
Sobel
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
High Efficie
n
c
y Video
Co
ding (HEVC) [1] provide
s
signifi
cantly better video
codi
ng
efficien
cy tha
n
the
last
ge
neratio
n vide
o codin
g
, tha
t
is, Advan
c
e
Video
Codin
g
(AV
C
).
Un
de
r
the conditio
n
of maintai
n
i
ng the
same
video
qualit
y, the goal
o
f
HEVC i
s
to
red
u
ce bit
-
rate
deman
d by 5
0
% com
p
a
r
e
d
with AV
C
at the exp
e
n
s
e of i
n
crea
sed comp
utational
com
p
lex
i
ty,
whi
c
h is in a
n
accepted ran
ge.
Although
HE
VC still b
e
lo
ngs to
the b
l
ock-b
a
sed h
y
brid video
codi
ng frame
w
ork, it
provide
s
a h
i
ghly flexible hiera
r
chy o
f
unit
representation [2], which in
clu
des th
ree bl
ock
con
c
e
p
ts, na
mely, codin
g
unit (CU), predictio
n uni
t (PU), an
d tran
sform
unit (T
U). CU i
s
a u
n
it
simila
r to the
macroblo
c
k. CU
can
be
split and is
al
ways
squ
a
re. The dime
nsi
on of CU ra
n
ges
from 8x8 up t
o
the larg
est
codi
ng unit (LCU). The
d
e
fi
nition of CU allo
ws itsel
f
to split into four
equal
-si
z
e
d
PU is a ba
sic u
n
it used for
carryin
g inform
ation relate
d to the predi
cti
on pro
c
e
s
s. PU
can o
n
ly be u
s
ed in
CU. P
U
ha
s two
sizes that
are su
pporte
d in intra pre
d
iction,
namely, 2N×2N
and
N×N. In
addition
to
CU a
nd P
U
, T
U
i
s
a
u
n
it rel
a
ted to t
r
an
sf
ormatio
n
a
n
d
qua
ntizatio
n
and
its size cann
ot exceed tha
t
of
CU. Base
d on the re
cu
rsive
stru
ct
ure, the encod
er mu
st exha
ust
all combi
nati
ons of CU, PU, and T
U
to determin
e
a
n
optimal sol
u
tion. Ho
wev
e
r, this proce
ss i
s
time-con
sumi
ng.
To imp
r
ove
the effici
en
cy and
a
c
cura
cy of vid
eo
codi
ng,
HEVC p
r
ovide
s
u
p
to
34
predi
ction
mo
des
for i
n
tra
predi
ction
[3], whi
c
h fa
r ex
cee
d
s th
e ni
n
e
predi
ction
mode
s of AV
C.
HEVC use
s
rate-di
s
tortio
n
optimization (RDO
)
te
chni
que to
de
cid
e
the
codi
ng
mode fo
r a
CU.
Figure sh
ows the RDO p
r
o
c
e
ss. As
can
be see
n
fro
m
Figure, to cho
o
se the b
e
st co
ding m
ode
for an
CU,
HEVC en
cod
e
r cal
c
ulate
s
th
e rate
-di
s
torti
on
cost (RD co
st)
of eve
r
y possible m
ode
(34 o
r
17 m
ode
s), after that HEVC choo
se
s
the mode which has the mini
mum value. Thi
s
pro
c
e
ss i
s
re
peatedly
carried out for
al
l the po
ssi
bl
e mode
s fo
r
CU.
Ho
wev
e
r, the en
co
der
can
not affo
rd
the total
co
mputation if
a
ll the
34
predictio
n mo
d
e
s
pa
ss thro
ugh th
e
RDO
pro
c
e
ss. Th
e
r
efore, relea
s
e the
comp
utational bu
rde
n
of RDO in
HEVC is far
more d
e
man
d
ing
than any existing video co
ding alg
o
rith
m. The Ha
d
a
m
ard transf
o
rm is used in
HM7.0, and t
h
ree
or eight can
d
i
dates
a
r
e sel
e
cted
throug
h
a Ro
ugh
M
ode
De
ci
sion
(RMD)
pro
c
e
s
s to
red
u
ce t
he
comp
utation
of the en
co
d
e
r. All p
r
edi
ction mo
d
e
s
i
n
the RM
D pro
c
e
s
s
a
r
e tested usi
ng the
minimum a
b
solute su
m of Hada
mard Tran
sfo
r
med
coeffici
ents o
f
resid
ual si
g
nal [4]. RDO
is
then a
pplie
d t
o
eve
r
y candi
date m
ode
se
lected
by
the
RMD p
r
o
c
e
s
s. Ho
wever,
th
e computatio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 570
3 –
5710
5704
of the encod
er process is still
very huge despite the applicatio
n o
f
the Hadam
ard tra
n
sfo
r
m
in
the RM
D
pro
c
e
ss.
Thu
s
, redu
cing
the
n
u
mbe
r
of
ca
n
d
idate
s
for RDO i
s
ne
ce
ssary to mi
nimi
ze
the com
putat
ional bu
rd
en
of the enco
der. Fo
r ex
a
m
ple, for thi
s
re
ason, a
fast algo
rith
m is
prop
osed i
n
this
study to
re
pla
c
e th
e
Had
a
ma
rd t
r
ansfo
rm i
n
HM 7.0.
HM
7
.
0 mad
e
ma
ny
optimizatio
ns on the foundation of
other HEVC Te
st Model
s. Some of these
optimization
s
o
n
intra predi
ction will be di
scussed in
section 2.
The remain
d
e
r of this
pa
per i
s
o
r
gani
zed
as
follo
ws: Section
2
briefly de
scri
bes i
n
tra
predi
ction
in
HEVC. S
e
ction 3 i
n
trod
uce
s
th
e al
g
o
rithm
and
the p
r
in
ciple
of the fa
st i
n
tra
predi
ction
proce
s
s. Secti
on 4 p
r
e
s
en
ts the expe
rimental resul
t
s. Finally, the la
st se
cti
o
n
pre
s
ent
s the con
c
lu
sio
n
.
2. Intra Prediction in HEV
C
Figure 1 sh
o
w
s that the in
tra codi
ng to
ol of
HEVC p
r
ovide
s
up to
34 predi
ction
modes,
as
well a
s
the
plana
r mo
de
for the lum
a
comp
one
nt
of different PU
sizes. In
HM7
.
0, PU si
ze
s
of
4×4, 8
×
8, 16
×16, 3
2
×32, and 64
×6
4, correspon
d to 18, 35, 35,
3
5
, and 35
predictio
n mod
e
s.
H
o
w
e
ve
r
,
in
p
r
e
v
io
us
vis
i
on
s
,
th
er
e
ar
e
ju
s
t
4 p
r
e
d
icti
on mod
e
s for 64×64. Thi
s
optimizatio
n i
s
very important for the video
whi
c
h hav
e large blo
cks, at thi
s
case LCU
will be chosen as
the
best CU [5]. So this optimi
z
ation
can off
e
r us
b
e
tter p
r
edi
ction mo
d
e
s, and lo
we
r the BD-rate.
Figure 1. 35 predi
ction m
o
des
Based on the 4×4 predi
ction unit illustrated in
Figure
2, we can
m
a
ke
a rough judgment
from its luma
texture. The
predi
ction ve
ctor i
s
the sa
me as the
arrow. The
RM
D process sh
ould
find a
way
to sele
ct can
d
idate
s
from
the 3
4
pre
d
i
ction
mode
s and
try its
best to
ma
ke
the
optimal mod
e
is one of the
s
e candi
date
s
.
Figure 2. Illustration of a 4x4 PU
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fast Intra Pre
d
iction Mo
de
De
cisi
on Algo
rithm
for HEVC (Me
ngm
en
g Zhang
)
5705
The followi
n
g
intra pre
d
i
c
tion process must
be p
e
rform
ed to sele
ct the best intra
predi
ction m
o
de for the lu
ma co
mpon
e
n
t of each
P
U
. The first step is an
RM
D process, which
will ma
ke
a rough
de
cisi
o
n
on th
e
choi
ce
of t
he b
e
st predi
ction
mode.
HM7.0
then a
pplie
s the
Had
a
ma
rd transfo
rm. Afte
rwa
r
d, th
e p
r
edictio
n
co
st
Jpred,SATD i
s
cal
c
ulate
d
f
o
r all
po
ssible
predi
ction m
ode
s, and se
veral pre
d
icti
on mode
s wi
th lower
co
sts are cho
s
e
n
as ca
ndid
a
t
e
mode
s. The
different pred
iction unit si
zes [6] of
4×4,
8×8, 16
×16,
32×32, and 6
4
×6
4 have 8,
8,
3, 3, and 3 candid
a
te mod
e
s, as liste
d in Table
1. After that, a proce
s
s of MPM is adopted
. In
previou
s
editi
ons, th
e MP
M process h
a
s
two
candi
d
a
tes, h
o
weve
r the
r
e a
r
e
th
ree
ca
ndid
a
te
s in
HM7.0.
What
's mo
re, the
MPM pro
c
e
ss
ha
s
chan
ged tre
m
en
d
ously. In HM7.0, the M
P
M
pro
c
e
s
s d
o
e
s
not j
u
st
sim
p
ly pu
sh
the
intra
dire
ction
s
of
ab
ove a
nd left P
U
i
n
to the
candid
a
te
list. Th
e MP
M process of
HM
7.0 m
a
ke the
mo
st u
s
e
of the
ab
ove an
d left
predi
ction
mo
de
s
throug
h a we
ighted co
efficient
when
ab
ove
an
d
l
e
ft
predi
ction
m
ode
s are eq
ual. Finally, the
sele
cted mo
des a
r
e pl
aced in the ca
ndidate li
st, whi
c
h is u
s
e
d
for the RDO pro
c
e
s
s. The
flowchart of RDO is
sho
w
n
in Figure 3.
Table 1 Num
ber of predi
ction m
ode
s after the process of RM
D
PU size
Number of
predic
t
ion modes
4x4
8x8
16x16
32x32
64x64
8
8
3
3
3
Since the HM7.0 has t
h
ree MPM [
7
] whic
h is more than
other editi
ons, the
comp
utationa
l burden
is m
o
re
than
othe
r e
d
itions.
Re
lease the
bu
rden
of the
en
cod
e
r i
s
the
key
point of this p
r
opo
se
d algo
rithm.
Figure 3. The
process of RDO
3. Fast Intr
a Prediction Algorithm
Our pro
p
o
s
e
d
algo
rithm
i
n
trodu
ce
s a new method
for
the RM
D pro
c
e
ss, whi
c
h
intend
s
to redu
ce the
number of candid
a
tes an
d the comput
ation load. Th
e flowch
art of the algorithm
is
s
h
ow
n
in
F
i
gu
r
e
4
.
The core
of
t
h
is algorithm ut
ilize the tex
t
ure
of a CU t
o
make out
which predi
ction mode
is mo
st p
r
ob
a
b
ly be the
be
st mod
e
. We
found t
hat th
e pixels (b
oth
luma a
nd
ch
roma
) al
ong t
h
e
dire
ction of l
o
cal e
dge
ne
arly have the
similar
val
u
es. The
r
efo
r
e
by predi
ct the pixels
usi
ng
those
neig
h
b
o
ring
pixels i
n
the same d
i
rectio
n of the
edge, a
pre
d
i
ction mo
de
can be
got. After
that an
edg
e
map
whi
c
h
rep
r
e
s
ent
s th
e vecto
r
of l
o
cal
ed
ge i
s
created,
an
d a l
o
cal e
d
g
e
dire
ction hi
stogra
m
is then
establi
s
he
d for ea
ch CU.
Since texture
is the
core o
f
edge d
e
tect
ion,
the prop
ose
d
alg
o
rith
m will u
s
e th
e Sobel
operator a
s
the co
re alg
o
ri
thm for its excelle
nt perfo
rmance wh
en
use
d
in edg
e detectio
n
[8].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 570
3 –
5710
5706
Figure 4. Algorithm flowch
art
The form
ula of Sobel ope
rator is given
by (1), and m
a
trix A is a 3×3 pixel unit.
1,
1
1
,
1
,
1
,1
,
,
1
1,
1
1
,
1
,
1
10
1
1
2
1
2
0
2
*
,
000
*
,
10
1
1
2
1
ij
ij
ij
x
y
ij
ij
i
j
ij
ij
i
j
pp
p
GA
G
A
A
p
p
p
pp
p
(1)
Two
co
nvolut
ion
kernel
s e
x
ist for the
S
obel
ope
rator, and
ea
ch
kernel
is rel
a
ted to th
e
degree
of dif
f
eren
ce
between th
e h
o
ri
zontal
an
d v
e
rtical
di
re
ctions [9]. The
co
rrespon
di
ng
dire
ction vect
or {dx
i,j
,dy
i,j
} for a luma pix
e
l in a picture
is defined a
s
:
,
1
,1
,1
1
,
1
1
,
1
,
1
1
,
1
2
ij
i
j
ij
i
j
i
j
ij
i
j
dx
p
p
p
p
p
p
(2)
,
1
,
1
1,
1,
1
1
,
1
1
,
1,
1
2
i
j
ij
ij
ij
i
j
i
j
i
j
dy
p
p
p
p
p
p
(3)
The d
e
g
r
ee
of differen
c
e
on the
ho
rizo
ntal an
d verti
c
al
dire
ction
s
are
corre
s
po
ndingly
rep
r
e
s
ente
d
by dx
i,j
and dy
i,j
. The magnitude of th
e vector i
s
a
c
curately p
r
e
s
ente
d
by (4
). In
addition, (5
) i
s
ado
pted wh
en the com
p
u
t
ation load is
con
s
id
ere
d
.
22
,,
,
()
(
)
(
)
ij
ij
ij
Am
p
D
dx
dy
(4)
,,
,
()
'
ij
ij
i
j
A
mp
D
d
x
d
y
(5)
The angl
e of vector is p
r
e
s
ente
d
by (6).
In fact, (7) can be u
s
ed to
pre
s
ent the
angle of
the vecto
r
b
e
ca
use it is a sim
p
ler t
h
re
shol
d techniqu
e for b
u
ilding
up th
e edg
e direction
histog
ram.
,
,
180
(
)
arct
a
n
(
)
o
ij
ij
dy
Ang
D
dx
(6)
,
,
()
'
ij
ij
dy
Ang
D
dx
(7)
After obtainin
g
the vecto
r
in a coo
r
dina
te system, Fi
gure
5
sho
w
s ho
w to m
a
ke the
deci
s
io
n on whi
c
h predi
ction mode th
e vector
bel
ong
s to. The region b
oun
dary of the two
adja
c
ent p
r
ed
iction mo
de v
e
ctors i
s
the
bise
ctrix of th
e two ve
ctors. So if a vector is i
n
the a
r
ea
betwe
en
one
pre
d
ictio
n
m
ode
and
a
bi
se
ctrix, it will
be j
udg
ed t
o
bel
ong
s to
the p
r
edi
ctio
n
mode. Ta
king
Figure 5 a
s
an example,
the vector
we
got is in the area b
e
twe
e
n
the predi
ctio
n
mode 3 a
nd t
he Boun
dary
line of the jud
g
ment, So the vector
belo
ngs to m
ode
3, and (5
) is t
he
corre
s
p
ondin
g
magnitud
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fast Intra Pre
d
iction Mo
de
De
cisi
on Algo
rithm
for HEVC (Me
ngm
en
g Zhang
)
5707
31
8
1
0
26
14
B
o
und
ar
y li
ne of
t
h
e
J
u
d
g
men
t
The v
e
ctor w
e
go
t
Predi
c
t
i
on Mo
de Vect
o
r
s
Figure 5. Illustration of the boun
dar
y line
for predi
ction
mode judg
m
ent
The first ste
p
is calculate t
he vecto
r
of
ev
ery pixel in
a LCU
whi
c
h
will be
ope
ra
ted, and
put the data o
f
these vecto
r
into two arra
y(one a
rray st
ore
s
mag
n
itu
des, an
other
store
s
an
gle
s
).
After that th
e
process
goe
s into
a
re
cu
r
s
iv
e
st
ru
ct
ur
e
,
in t
h
i
s
st
ru
ct
ure,
P
U
wit
h
t
he
siz
e
of 64×64, 32×32, 16×16,
8×8, 4×4 will
get the vector
of every pixel in each
PU from the arrays
whi
c
h we got
in the first ste
p
, and ma
ke the j
udgm
ent
whi
c
h p
r
edi
ction mod
e
the vector b
e
long
s
to.
The next
ste
p
is
creating
a histo
g
ra
m. Figur
e 6
sho
w
s th
at a hi
stogram
wa
s
use
d
to
cou
n
t the
su
m of the
ma
gnitude
of e
a
ch
p
r
edi
ctio
n mo
de i
n
a
PU. T
he X
-
axis
contai
ns the
predi
ction m
ode
s, whe
r
e
a
s the Y-axis contain
s
the
sum of the magnitud
e
of each p
r
edi
ct
ion
mode in
a P
U
. The
highe
st and th
e se
con
d
hig
h
e
s
t predi
ction
mo
des i
n
the hi
stogram
are th
en
cho
s
e
n
(Ta
k
i
ng b
o
th BD-rate an
d
codi
ng time i
n
to
con
s
id
eratio
n
,
two
can
d
id
ates
are
a
b
e
tter
choi
ce
than
one
or three
can
d
idat
es) a
s
the
can
d
idate
s
f
o
r
RDO. Th
ough
it is
not
mathemati
c
al
ly corre
s
po
n
d
ing to
plan
e predi
cti
on
to any di
re
ctional e
dge,
we
can
for
sure
asso
ciate the
predi
ction to
its re
sp
ectiv
e
edg
es. T
h
e
r
efore, it is fairly re
asona
ble for u
s
to
try
plane
predi
ction if it i
s
n
o
t obvio
u
s
ly a
DC predi
ctio
n. In Fig
u
re
5, as an
exa
m
ple, p
r
edi
cti
o
n
mode
s 6 a
n
d
9 are
cho
s
e
n
as the
ca
n
d
idate mo
de
s. Finally, the two mod
e
s
a
r
e pla
c
e
d
in the
can
d
idate li
st.
0
200
400
600
800
1000
1200
012
3456
789
1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
3
1
3
2
Figure 6. Edge dire
ction hi
stogram
4. Experimental Re
sults
In this expe
ri
ment, 300 fra
m
es of e
a
ch
seq
uen
ce a
r
e
cod
ed to test the perfo
rm
ance of
prop
osed
alg
o
rithm, an
d e
v
ery frame i
s
intra
cod
ed.
A comp
uter
whi
c
h h
a
s 2.
8GHz
co
re i
s
use
d
to do this
experiment.
The exp
e
rim
ental results [
10] of thi
s
st
udy
are p
r
e
s
ented i
n
Ta
bl
es
2, whi
c
h
show that
the pro
p
o
s
ed
algorith
m
sa
ves 16.5% o
n
avera
ge
coding time
a
nd the B
D
-ra
t
e increa
se
s
by
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 570
3 –
5710
5708
approximatel
y 1.122% in high efficie
n
cy (HE) te
st con
d
ition
s
, resp
ectivel
y
. The resul
t
s in
Cla
s
ses A a
nd B
are
be
tter than
tho
s
e i
n
oth
e
r test
seq
uen
ces. T
herefore, the p
r
o
p
o
s
ed
algorith
m
perf
o
rm
s better u
nder a hi
gh g
r
aphi
cs co
ndi
tion.
To calculate
the time saving of the
fast
intra
-
p
r
edi
ction al
g
o
rithm, the f
o
llowin
g
cal
c
ulatio
n is defined to
e
v
aluate the ti
me differe
nces.
We ta
ke
T
HEVC
prese
n
t the coding
time
use
d
by HM
7.0 encode
r,
corre
s
po
ndi
ngly T
pro
be the time take
n by the faster intra
p
rediction
algorith
m
, an
d is define
d
a
s
:
%
100
T
T
T
Time
HEVC
Pro
HEVC
(8)
Table 2. Result in HE test condition
All-Intra HE
Y
BD-rat
e
U
BD-rat
e
V
BD-rat
e
△
Ti
m
e
Class A
0.73%
0.0
0.3%
-15.7%
Class B
0.4%
0.0%
-0.15%
-16.9%
Class C
1.245%
0.45%
0.775%
-16.7%
Class
D
2.025%
1.125%
0.575%
-16.8%
Class
E
1.21%
1.1%
0.8%
-16.1%
All 1.122%
0.535%
0.46%
-16.44
%
The Rate
-Distortion (RD) curves of ParkSc
ene a
nd Race
Ho
rses a
r
e shown in Figures
7, 8. The cu
rve of the pro
posed alg
o
rit
h
m is very
cl
ose to the
cu
rve of HM7.0
.
In addition, the
efficien
cy of ParkS
c
en
e is be
tter than that of Race
Horses.
Figure 7. RD
Curve
s
of ParkSce
ne (Cla
ss B 1920
×10
80HE
)
Figure 8. RD
Curve
s
of Ra
ce
Horse
s
_
8
3
2
×4
80_
30 (Cl
a
ss C 83
2×4
80 HE)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fast Intra Pre
d
iction Mo
de
De
cisi
on Algo
rithm
for HEVC (Me
ngm
en
g Zhang
)
5709
Figure 9
sho
w
s that the
re
con
s
tru
c
ted
i
m
age
quality
differen
c
e
bet
wee
n
HM7.0
and th
e
prop
osed alg
o
rithm is alm
o
st negli
g
ible
unde
r the HE
test conditio
n
.
(a)
Re
con
s
tru
c
ted by HM7.
0(HE
)
(c) Re
co
nstru
c
ted by
pro
p
o
s
ed al
gorith
m
(HE)
Figure 9. The
difference be
tween
HM4.0
and pro
p
o
s
e
d
algorith
m
(Ra
c
e
H
orse
s)
5. Conclusio
n
This pap
er propo
sed
a fa
st intra
pre
d
ict
i
on mo
de
de
cisi
on
algo
rithm for HEV
C
, whi
c
h
wa
s achieve
d
usin
g the Sobel ope
rat
o
r and
by redu
cing the
numbe
r of p
r
edi
ction mo
de
can
d
idate
s
fo
r
RDO. The
experim
ental results
i
n
Se
ction
4
sho
w
that the p
r
o
posed
algo
rithm
can
red
u
ce the co
mplexit
y
of the enco
der a
nd t
hat
the video qu
ality loss i
s
a
l
most ne
gligi
b
le
comp
ared
wit
h
HM7.0. In
addition, th
e
prop
osed
alg
o
rithm
achiev
ed u
p
to
19.
8% re
du
ction
in
codi
ng time with negligible
degradatio
n o
f
quality and BD-rate.
Ackn
o
w
l
e
dg
ements
This
wo
rk wa
s
supp
orte
d i
n
pa
rt by National
Natu
ral
Scien
c
e
Fou
ndation
of Ch
ina (No.
6110
3113,
No. 6127
2051
), Jiang
su Provincial
Na
tu
ral Sci
e
n
c
e
Found
ation (BK20114
55)
and
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 570
3 –
5710
5710
Beijing Mu
ni
cipal Ed
ucation Commi
ssion Scie
nce
and T
e
chn
o
logy Devel
opment P
r
og
ram
(KM201
310
0
0900
4).
Referen
ces
[1]
Kemal Ug
ur, Kenn
eth
A
nde
rsson.
Hi
gh
Perf
ormanc
e, Lo
w
Com
p
le
xi
t
y
Vi
de
o
C
odi
ng and
the
Emergi
ng HEV
C
Stand
ard.
IEEE Transactio
n
s on C
i
rcuits
and Syste
m
s f
o
r Vide
o Tech
nol
ogy
. 20
10
;
20(1
2
): 168
8-1
697.
[2]
W
oo-Jin H
an.
Improved Vi
de
o Compr
e
ssio
n
Effi
cienc
y T
h
rou
gh F
l
e
x
ibl
e
Unit R
epres
entatio
n an
d
Corresp
on
din
g
Extens
ion
of Codi
ng T
ools.
IEEE Transactions on C
i
rc
uits and Syste
m
s for Vid
e
o
T
e
chno
log
. 2
0
10; 20(1
2
): 170
9-17
20.
[3]
Lia
ng Z
hao. F
a
st Mode Deci
sion Al
gorithm
for Intra Prediction in HEVC.
Visua
l
Co
mmu
n
icati
ons an
d
Imag
e Process
i
ng (VCIP).
20
11; 1-4.
[4]
F
r
ank Bosse
n,
Virgi
n
ie
Dru
g
eon. Vi
de
o C
o
din
g
Usi
ng
a
Simplifi
ed B
l
oc
k Structure a
n
d
Adva
nce
d
Codi
ng T
e
chni
ques.
IEEE Transactions on
Circuits a
nd S
ystem
s for
Video Technology
. 2010; 2(
12)
:
166
7-16
75.
[5]
Van W
a
ll
end
ae
l G. Improved intra mode si
gn
alin
g for HEVC
.
Multime
d
i
a
an
d Expo (ICME)
. 2011; 1-6.
[6]
Benj
amin
Bros
s. High
efficie
n
c
y
v
i
de
o
co
di
n
g
(HEVC)
te
xt specific
ation
dr
aft 7.
JCT
-
VC 5th
Me
eting
.
201
1; 56-7
1
.
[7]
Il-Koo K
i
m. H
M
7: Hig
h Effici
enc
y V
i
de
o
Co
din
g
(HEV
C) T
e
st Mod
e
l
7 E
n
cod
e
r D
e
scri
p
tion.
JC
T-
V
C
9th Meetin
g
. 2
012; 31-
46.
[8]
Anil
K J
a
in,
A
d
it
ya V
a
il
a
y
a.
Image r
e
triev
a
l
usi
n
g
col
o
r
a
nd s
h
a
pe.
P
a
ttern Reco
gn
it
. 29(8):
12
33-
124
4.
[9]
F
eng Pan. F
a
st Mode Deci
sion Al
gorithm
for Intrapredi
ction in H.2
6
4
/
AVC Video C
odi
ng.
IE
EE
T
r
ansactio
n
s o
n
Circuits a
nd
Systems for Vi
deo T
e
ch
no
log
y
. 2005; 15(
7): 813-
822.
[10]
F
r
ank Bosse
n. Commo
n test
cond
itions
an
d
soft
w
a
re ref
e
r
ence c
onfi
gura
t
ions.
JCT
-
VC 6th
Meeti
n
g
.
201
1; 14-2
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.