Inter
national
J
our
nal
of
Electrical
and
Computer
Engineering
(IJECE)
V
ol.
11,
No.
2,
April
2021,
pp.
1303
1310
ISSN:
2088-8708,
DOI:
10.11591/ijece.v11i2.pp1303-1310
r
1303
Machine
lear
ning-based
ener
gy
consumption
modeling
and
comparing
of
H.264
and
Google
VP8
encoders
Y
ousef
O.
Sharrab
1
,
Mohammad
Alsmirat
2
,
Bilal
Hawashin
3
,
Nabil
Sarhan
4
1,
4
W
ayne
State
Multimedia
Systems
and
Deep
Learning
Research
Laboratory
,
ECE
Department,
Detroit,
USA
2
Department
of
Computer
Science,
Jordan
Uni
v
ersity
of
Science
and
T
echnology
,
Jordan
3
Department
of
Computer
Information
Systems,
Al
Zaytoonah
Uni
v
ersity
of
Jordan,
Jordan
Article
Inf
o
Article
history:
Recei
v
ed
Apr
10,
2020
Re
vised
Jun
7,
2020
Accepted
Jun
23,
2020
K
eyw
ords:
Encoders
comparison
Encoding
ener
gy
consumption
Encoding
time
Feature
engineering
Google
VP8
encoder
Machine
learning
MPEG-4
P
art
10
H.264/A
VC
Percei
v
ed
video
quality
Re
gression,
modeling
V
ideo
encoders
V
ideo
encoding
ABSTRA
CT
Adv
ancement
of
the
prediction
models
used
in
a
v
ariety
of
fields
is
a
result
of
the
contrib
ution
of
machine
learning
approaches.
Utilizing
such
modeling
i
n
feature
en-
gineering
is
e
xceptionally
imperati
v
e
and
required.
In
this
research,
we
sho
w
ho
w
to
utilize
machine
learning
to
sa
v
e
time
in
research
e
xperiments,
where
we
sa
v
e
more
than
fi
v
e
thousand
hours
of
m
easuring
the
ener
gy
consumption
of
encoding
recordings.
Since
measuring
the
ener
gy
consumption
has
got
to
be
done
by
humans
and
since
we
require
more
than
ele
v
en
thousand
e
xperiments
to
co
v
er
all
the
combinations
of
video
sequences,
video
bit
rate,
and
video
encoding
settings,
we
utilize
machine
learning
to
model
the
ener
gy
cons
umption
utilizing
linear
re
gression.
VP8
codec
has
been
of
fered
by
Google
as
a
free
video
encoder
in
an
ef
fort
to
replace
the
popular
H.264
video
en-
coder
standard.
This
research
model
ener
gy
consumption
and
describes
the
major
dif-
ferences
betwee
n
H.264/A
VC
and
VP8
encoders
based
on
of
ener
gy
consumption
and
performance
through
e
xperiments
that
are
machine
le
arning-based
modeling.
T
wenty-
nine
uncompressed
video
se
gments
from
a
standard
data-set
are
used,
with
se
v
eral
sizes,
details,
and
dynamics,
where
the
frame
sizes
ranging
from
QCIF(
176x144)
to
2160p(3840x2160).
F
or
f
airness
in
comparison
analysis,
we
use
se
v
en
settings
in
VP8
encoder
and
fifteen
types
of
tuning
in
H.264/A
VC.
The
settings
co
v
er
v
arious
video
qualities.
The
performance
metrics
include
video
qualities,
encoding
time,
and
encod-
ing
ener
gy
consumption.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Mohammad
Alsmirat
Department
of
Computer
Science
Jordan
Uni
v
ersity
of
Science
and
technology
,
Jordan
Email:
masmirat@just.edu.jo
1.
INTR
ODUCTION
Machine
learning
approaches
ha
v
e
recently
contrib
uted
to
the
adv
ancement
of
the
prediction
model
s
used
for
ener
gy
consumption
[1–3].
In
this
paper
,
we
utilize
machine
learning
to
predict
the
ener
gy
consump-
tion
of
o
v
er
ten
thousand
e
xperiments
to
be
used
for
encoders
performance
comparisons.
The
popularity
of
H.264/A
VC
has
confronted
an
incredible
competence
with
Google
dischar
ging
VP8
and
VP9
as
free
of
char
ge
video
encoders,
mainly
on
Y
ouT
ube
[4].
One
of
the
k
e
y
f
actors
that
w
oul
d
shape
the
encoder
of
the
future
is
the
adequac
y
of
VP8
comparing
to
H.264/A
VC.
Thanks
to
the
a
wesome
o
v
ere
xtend
of
video
compression,
the
viability
of
VP8
comparing
to
with
H.264/A
VC
needs
more
thorough
analysis.
It
is
noted
that
a
fe
w
w
orks
only
compared
the
viability
of
both
encoders
such
as
[5–14].
Study
[9]
has
compared
J
ournal
homepage:
http://ijece
.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
1304
r
ISSN:
2088-8708
H.264/A
VC
and
VP8
encoders
for
only
three
sequences.The
e
v
aluation
measurements
were
only
one
metric;
which
is
the
percei
v
ed
video
quality
.
Besides,
it
tested
only
with
minimum
basic
compression
parameters.
This
research
depi
cts
the
contrasts
between
MPEG-4
P
art
10
(H.264/A
VC)
and
VP8
encoders
in
en-
er
gy
consumption,
video
quality
,
and
encoding
time/speed.
It
moreo
v
er
gi
v
es
nitty-gritty
comparati
v
e
e
v
al-
uations
for
more
than
ele
v
en
thousand
tests,
so
as
to
reflect
practical
circumstances
by
wisely
selecting
the
compression
settings,
and
the
video
e
v
aluation
se
gments.
T
wenty-nine
uncompressed
video
se
gments
were
used,
with
a
v
ariety
of
frame
contents,
with
sizes
e
xtending
from
QCIF
(176x144)
to
2160p
(3840x2160)
and
the
contents
changing
significantly
in
detail
and
speed
of
mo
v
ements.
T
o
guarantee
a
reasonable
and
f
air
comparison,
we
utilize
three
encoding
settings
in
H.264/A
VC,
with
fi
v
e
dif
ferent
tunings
for
each,
and
se
v
en
settings
in
VP8
encoder
.
These
settings
co
v
er
an
assortment
of
accomplished
quality
le
v
els.
W
e
change
the
bit
rate
for
each
video
se
gment
to
study
dif
ferent
video
quality
.
The
performance
metrics
include
the
quality
of
the
video,
speed
of
encoding,
and
ener
gy
consumption.
F
or
the
quality
of
the
video,
we
use
the
structur
al
similarity
inde
x
(SSIM)
[15].
The
or
g
anization
of
the
paper
is
as
tak
es
aft
er
.
Section
2.
pro
vides
background
information
of
the
used
encoders
for
comparison.
Subsequently
,
Section
3.
analyzes
and
discusses
the
performance
assessment
and
e
v
aluation
procedures,
techniques,
and
methodology
.
At
last,
Section
4.
illustrate
and
analyzes
the
most
results.
2.
B
A
CKGR
OUND
INFORMA
TION
Most
of
the
researches
that
combine
encoders
and
machine
learning/deep
learning
(ML/DL)
tech-
niques
focused
on
de
v
eloping
rob
ust
ML/DL
algorithms
for
the
detection,
tracking,
and
classification
of
objects
[16,
17]
and
the
detection
and
classification
of
unusual
e
v
ents
[16–21]
(and
reference
within).
The
o
v
erwhelm-
ing
majority
of
research
on
CV
considered
the
de
v
elopm
ent
of
rob
ust
algorithms
to
impro
v
e
accurac
y
[22–26]
(and
references
within).
The
encoding
process
stages
of
an
H.264/A
VC
codec
include
spatial(intra)
and
temporal
(inter)
pre-
diction
[27].
Both
the
spatial
and
the
temporal
prediction
are
utilized
to
decrease
the
duplicates
withi
n
the
video.
V
i
d
e
o
information
contains
temporal
and
spatial
repetition.
Subsequently
,
simil
itudes
can
be
encoded
by
f
air
considering
spatial
(within
a
frame)
and/or
temporal
(between
frames)
residuals.
An
H.264/A
VC
codec
can
select
from
numerous
di
v
erse
intra-prediction
and
motion-estimation
modes
at
the
time
of
encoding
a
macrobl
ock.
The
rate-distortion
optimization
(RDO)
mode
selection
is
an
optimum
algorithm
for
choosing
the
most
e
xcellent
encoding
mode
for
each
macroblock,
based
on
a
cost
that
considers
the
lo
west
v
alue
of
bit
rate
and
distortion
combination.
The
dif
ferences
between
H.264
and
VP8
in
modes
are
sho
wn
in
T
able
1.
T
able
1.
H.264/A
VC
and
VP8
encoding
comparison
Encoding
Step
H.264/A
VC
encoder
VP8
encoder
Intra-Prediction
Nine
modes
of
prediction
were
used
per
(
4
4
)
and
Uses
f
our
intra-prediction
modes
shared
(
4
4
)
and
(
8
8
)
Luma
block
in
high
profiles
prediction
(
16
16
)
Luma,
and
(
8
8
)
Chroma
f
our
modes
of
prediction
were
used
per
(
16
16
)
Luma
block
and
(
8
8
)
Chroma
prediction
modes
Inter
-Prediction
sixteen
reference
frames
three
reference
frames
P
artition
types
are
(
16
16
)
do
wn-to
(
4
4
)
P
artition
types
are
(
16
16
)
do
wn-to
(
4
4
)
Spatial-prediction
is
utilized
to
foresee
the
content
of
a
block
from
its
neighbors’
blocks
with
no
history
utilization.
In
spatial
(intra-mode)
selection
in
H.264/A
VC
,
the
number
of
possible
mode
combinations
for
a
16
16
pix
el
MacroBlock
(MB)
can
be
(16
N
4
+
N
16)
N
8
,
where
N
4
,
N
16
,
and
N
8
represent
the
number
of
modes
of
a
4
4
Luma
block,
and
a
16
16
Luma
block,
an
8
8
Chroma
block,respecti
v
ely
.
T
o
select
the
best
mode
for
one
MacroBlock,
the
encoder
performs
(16
9
+
4)
4
which
is
equals
to
592
RDO
calculations
[28].
In
VP8
codec,
applying
the
same
equation
gi
v
es
(16
4
+
4)
4
which
is
equals
to
272
RDO
calculations.
Hence,
the
comple
xity
of
H.264/A
VC
is
more
than
twice
that
of
VP8
encoder
intra-
prediction
without
optimized
implementation.
Motion-estimation
(Inter
-prediction)
is
the
method
of
predicting
the
a
block
contents
(Luma
and
Chroma)
by
referring
to
past
and/or
future
process
ed
frames.
Refere
nce
frames
(RFs)
and
motion
v
ectors
(MVs)
are
the
essential
parts
of
motion
estimation
(inter
-mode
prediction).
The
RF
is
already
processed
frame
Int
J
Elec
&
Comp
Eng,
V
ol.
11,
No.
2,
April
2021
:
1303
–
1310
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1305
utilized
to
get
similar
blocks
,
and
where
the
MV
sho
w
the
displacement
(in
pix
els)
between
where
the
current
block
is
in
the
current
frame
and
the
comparing
prediction
block
within
the
RF
[29].
H.264/A
VC
and
VP8
encoders
ha
v
e
comparati
v
e
structures
in
te
rms
of
block
sizes.
H.264/A
VC
utilises
the
follo
wing
partitions
16
16
,
16
8
,
and
8
16
,
and
each
8
8
can
be
further
di
vided
into
8
8
,
8
4
,
4
8
,
or
4
4
.
While
VP8
encoder
utilises
16
16
,
16
8
,
8
16
,
8
8
,
and
4
4
partitions.
VP8
encoder
dropped
8
4
and
4
8
partitions
which
is
impossible
to
be
of
a
significant
issue
[30,
29].
Both
H.264/A
VC
and
VP8
encoders
support
v
ariable-size
motion
v
ectors.
3.
PERFORMANCE
EV
ALU
A
TION
METHODOLOGY
3.1.
Used
codecs,
video
sequences,
and
perf
ormance
metrics
As
for
the
e
xperimental
set
up,
we
adopted
vpx
codec
for
encoding
and
decoding
with
re
g
ard
to
the
VP8.
Furthermore,
we
adopted
for
encoding
X264
(r1688)
and
we
adopted
for
decoding
FFmpe
g
(SUN-
r24758)
with
re
g
ard
to
the
H.264/A
VC.
The
e
v
aluation
video
se
gments
were
cate
gorized
into
4
cate
gories
according
to
frame
size,
as
illustrated
in
T
able
2.
pro
vide
more
details
about
the
used
test
sequences
are
in
T
able
3.
T
able
2.
Used
video
se
gments
properties
formatting
frmae
size
(in
Pixles)
Standard
bit
rate
Quarter
common
intermediate
176
144
videoconferencing
20-800
Kbps
format
(QCIF)
Common
intermediate
format
(CIF)
352
288
4
Common
intermediate
format
(4CIF)
704
576
Standard-Definition
T
ele
vision
(SDTV)
100-2000
Kbps
Standard
HD
(720p)
1280
720
High-definition
tele
vision
(HDTV)
500-3000
Kbps
Full
HD
(1080p)
1920
1080
4k
(2160p)
3840
2160
Quad
high-definition
tele
vision
(QHDTV)
2000-8000
Kbps
T
able
3.
Characteristics
of
the
used
standard
video
sequences
V
ideo
Se
gment
Duration
(seconds)
V
ideo
Size
(Frames)
Resolution
F
orman
12
300
176
144,
352
288
Salesman
17
449
176
144
Ne
ws
12
300
176
144,
352
288
Mobile
12
300
176
144
Highw
ay
80
2000
176
144
Stef
an
3
90
352
288
P
aris
42
1065
352
288
MotherDaughter
12
300
352
288
City
24
600
704
576
Cre
w
24
600
704
576
Harbour
24
600
704
576
Soccer
24
600
704
576
Ice
19
480
704
576
DucksT
ak
eOf
f
20
500
1280
720,
3840
2160
InT
oT
ree
5
500
1280
720,
3840
2160
OldT
o
wnCross
5
500
1280
720,
3840
2160
P
arkJo
y
20
500
1280
720,
3840
2160
Mobcal
20
504
1280
720
BlueSk
y
8
645
1920
1080
Pedestrian
Area
15
375
1920
1080
Ri
v
erbed
10
250
1920
1080
RushHour
20
500
1920
1080
Station2
12
313
1920
1080
The
used
e
v
aluation
measurements
are
ener
gy
consumption,
video
quality
,
and
encoding
speed
per
-
formance.
T
o
measure
the
quality
of
the
videos
,
we
utilize
the
structur
al
similarity
inde
x
(SSIM)
[15].
T
o
determine
SSIM,
we
utilized
the
the
brightness
in
the
image
or
Luma
(Y)
part
in
the
YUV
,
as
brightness
w
ould
stimulate
e
ye
more
than
the
color
w
ould.
W
e
split
the
ener
gy
measured
data
to
t
w
o
cate
gories,
training
data
and
test
data
(80
%
of
measured
data
were
for
training).
Mac
hine
learning-based
ener
gy
consumption
modeling
and
comparing
of
...
(Y
ousef
O.
Sharr
ab)
Evaluation Warning : The document was created with Spire.PDF for Python.
1306
r
ISSN:
2088-8708
3.2.
Pr
ocedur
e
Encoding
is
done
using
vpx
enc/vpxdec
(http://www
.webmproject.or
g)
codec.
H.264/A
VC
encoding
is
done
by
X.264
(http://X264.nl),
while
the
decoding
of
H.264/A
VC
is
done
by
using
FFmpe
g.
W
e
de
v
eloped
C++
program
to
computerize
the
measurements
of
the
encoding
time
and
the
encoded
video
bit
rate.
W
e
utilize
MA
TLAB
program
to
measure
SSIM
metrics,
to
find
the
quality
of
the
distorted
decoded
video
sequences
relati
v
e
to
the
original
distortion-free
video
sequence.
The
ratio
of
the
tar
get
bit
rate
to
the
accomplished
bit
rate
is
calculated
by
di
viding
the
input
bit
rate
that
we
specify
in
the
command
line
by
the
accomplished
bit
rate,
which
is
the
a
v
erage
of
the
real
bit
rate
for
the
encoded
sequences.
T
o
tradeof
f
the
encoding
speed
and
the
accomplished
bit
rate
for
certain
SSIM
quality
,
we
dra
w
the
point
whi
ch
corresponds
to
the
interpolation
tuple
of
that
SSI
M
quality
.
In
these
figures,
the
higher
speed
and
lo
wer
bit
rate
are
better
encoders
and
parameters
combinations.
The
encoding
setti
ngs
for
both
H.264/A
VC(X264)
and
VP8
encoders
are
se
lected
to
include
a
v
ariety
of
ener
gy
consumption
and
video
quality
.
The
encoding
settings
of
VP8
encoder
are
selected
from
the
W
ebM
website.
Additionally
,
H.264/A
VC
(X264)
parameter
settings
are
selected
from
the
de
v
elopers
website
[8].
The
utilized
encoding
settings
are
e
xplained
in
[11].
The
formats
of
the
encoded
videos
are
W
ebM
and
mp4
for
VP8
encoder
and
H.264/A
VC,
respecti
v
ely
.
The
ener
gy
consumption
e
xperiments
were
conducted
on
a
dual-core
processor
laptop.
The
ener
gy
w
as
measured
by
Graphic
T
imer
W
att
meter
.
The
number
of
needed
e
xperiments
are
(
29
video
sequence
6
bitrates
22
settings
3
repetition
for
accurac
y
which
is
equals
to
11
;
484
).
Each
e
xperiment
needs
in
a
v
erage
half
an
hour
to
be
setup
and
conducted
manually
,
which
is
v
ery
time
consuming
and
prone
to
error
.
F
or
that
reason,
we
pick
a
sample
as
follo
ws,
for
each
sequence
we
measure
the
ener
gy
for
one
bit
rate
out
of
six,
and
13
settings
out
of
22
which
reduces
the
number
of
needed
e
xperiments
to
522
e
xperiments
(
29
1
6
3
repetitions
for
accurac
y).
W
e
use
this
collected
data
for
training,
to
model
the
ener
gy
consumption
based
on
the
encoding
time
feature.
From
the
de
v
eloped
model,
we
predicted
the
remaining
10
;
962
(
11
;
484
522
)
ener
gy
consumption
predictions.
The
block
diagram
of
the
ener
gy
consumption
modeling
including
training
and
testing
is
sho
wn
in
Figure
1.
Figure
1.
Machine
learning-based
ener
gy
consumption
modeling
block
diagram
4.
RESUL
T
PRESENT
A
TION
AND
AN
AL
YSIS
In
this
sec
tion,
we
analyze
the
results
of
H.264/A
VC
and
VP8
encoders
Comparison
in
ener
gy
con-
sumption,
encoding
time,
and
video
quality
.
F
or
space,
we
do
not
sho
w
all
figures.
F
or
clear
figures,
we
do
not
sho
w
all
setting
results
in
all
figures.
4.1.
Ener
gy
consumption
and
encoding
speed
against
per
ceptual
quality
at
certain
bit
rate
tradeoff
In
all
frame
sizes
e
xcluding
high
definition
standar
ds
(1080p
and
2160p),
H.264/A
VC
accomplishes
lo
wer
ener
gy
consumption
and
higher
performance
than
VP8
encoder
at
the
same
bitrate.
In
High
Definition
Standards
resolutions,
VP8
encoder
yields
higher
performance
b
ut
more
encoding
time
and
higher
ener
gy
con-
sumption
than
H.264/A
VC
at
the
same
performance.
Notes
that
we
use
dif
ferent
settings
for
high
definition
standar
ds
(720p,
1080p,
and
2160p)
for
H.264/A
VC
based
on
the
de
v
elopers
recomm
endations.
Upon
the
Int
J
Elec
&
Comp
Eng,
V
ol.
11,
No.
2,
April
2021
:
1303
–
1310
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1307
e
v
aluation
of
VP8
encoder
settings
under
QCIF
,
CIF
,
and
4CIF
,
with
“Best”
indicating
the
most
ef
ficient
per
-
formance,
the
maximum
ener
gy
consumption,
and
the
sl
o
west
speed
of
encoding,
settings
“Good
0”,
“Good
1”,
“Good
2”,
“Good
3”,
“Good
4”,
and
“Good
5”
come
ne
xt,respecti
v
ely
.
The
setting
“Good
5”
produces
the
w
orst
performance,
the
minimum
ener
gy
intak
e,
and
the
f
astest
encoding
among
all
the
VP8
encoder
settings.
On
the
other
hand,
settings
“Good
1”,
“Good
2”,
and
“Good
3”
(720p,
1080p,
and
2160p)
were
more
ef
ficient
in
terms
of
encoding
speed,
performance,
and
ener
gy
intak
e.
Re
g
arding
H.264/A
VC,
the
setting
“higher
qual-
ity”
has
clearly
t
he
slo
west
encoding,
the
maximum
ener
gy
intak
e,
and
the
most
ef
ficient
performance,
mainly
for
smaller
resolution
v
alues,
whereas
adopting
“Higher
Speed”
pro
vides
the
w
orst
quality
,
the
smallest
ener
gy
intak
e,
and
the
f
astest
encoding.
“Normal
quality”
is
in
between
the
pre
vious
tw
o
settings.
4.2.
P
erf
ormance
at
v
arious
bit
rate
Figure
2
compares
the
H.264/A
VC
and
VP8
encoding
time
(speed)
and
video
quali
ty
performance
P
according
to
the
follo
wing
model
P
=
a
Q=Q
h
+
(1
a
)
S
=S
h
;
(1)
where
a
is
weighting
f
actor
,
Q
,
Q
h
,
S
,
S
h
represents
quality
,
highest
quality
,
encoding
speed,
highest
encoding
speed,
respecti
v
ely
.
The
weighting
f
actor
tak
e
v
alues
fro
0
to
1
,
its
v
alue
can
be
set
based
on
which
we
care
more
about;
the
quali
ty
or
the
encoding
speed.
F
or
e
xampl
e
in
real
time
streaming
we
care
more
about
encoding
speed,
while
in
of
f
line
recording
we
care
more
about
the
quality
.
This
model
tak
es
the
mean
v
alue
of
the
normalized
percei
v
ed
quali
ty
and
encoding
speed
for
a
=
0
:
5
.
The
figure
sho
ws
that
‘VP8
Good
5
setting’
and
‘H.264/A
VC
High
Speed
setting’
ha
v
e
the
best
performance,
where
‘VP8
Best
setting’
and
‘H.264/A
VC
High
Quality
setting’
ha
v
e
the
w
orst
performance.
This
e
v
aluation
sho
ws
that
high
encoding
speed
(less
quality)
settings
perform
better
as
the
y
increase
the
encoding
speed
without
significant
impact
on
the
percei
v
ed
video
quality
.
H.264/A
VC
encoder
generally
has
higher
performance
than
Vp8
encoder
based
on
(1).
0
200
400
600
800
0.4
0.5
0.6
0.7
0.8
0.9
Performance
Accomplished Bit Rate (Kbit/s)
(a)
0
500
1000
1500
2000
0.4
0.5
0.6
0.7
0.8
0.9
Performance
Accomplished Bit Rate (Kbit/s)
(b)
0
500
1000
1500
2000
0.4
0.5
0.6
0.7
0.8
0.9
Performance
Accomplished Bit Rate (Kbit/s)
(c)
0
1000
2000
3000
4000
0.4
0.5
0.6
0.7
0.8
0.9
Performance
Accomplished Bit Rate (Kbit/s)
(d)
2000
4000
6000
8000
10000
0.5
0.6
0.7
0.8
0.9
1
Performance
Accomplished Bit Rate (Kbit/s)
(e)
VP8 Best Configurations
VP8 Good0 Configurations
VP8 Good1 Configurations
VP8 Good2 Configurations
VP8 Good3 Configurations
VP8 Good4 Configurations
VP8 Good5 Configurations
H.264 High Quality No Tunning Configurations
H.264 High Quality SSIM Tunning Configurations
H.264 High Speed No Tunning Configurations
H.264 High Speed SSIM Tunning Configurations
H.264 Normal No Tunning Configurations
H.264 Normal SSIM Tunning Configurations
(f)
Figure
2.
Codecs
comparison
in
performance=(
0
:
5
quality/highest
quality
+0
:
5
encoding
speed/highest
encoding
speed)
Vs.
bit
rate,
(a)
Quality
QCIF
and
CIF
,
(b)
Quality
4CIF
,
(c)
Quality
720p,
(d)
Quality
1080p,
(e)
Quality
2160p
Mac
hine
learning-based
ener
gy
consumption
modeling
and
comparing
of
...
(Y
ousef
O.
Sharr
ab)
Evaluation Warning : The document was created with Spire.PDF for Python.
1308
r
ISSN:
2088-8708
As
we
see
in
Figures
3
and
4,
H.264/A
VC
generally
consumes
less
ener
gy
than
VP8
encoder
,
only
VP8
“Good
5”
setting
consumes
less
ener
gy
than
H.264/A
VC
“High
Quality”
setting.
0
200
400
600
800
0
0.1
0.2
0.3
0.4
Energy Consumed (Joul/Frame)
Accomplished Bit Rate (Kbit/s)
(a)
0
500
1000
1500
2000
0
0.5
1
1.5
2
2.5
Energy Consumed (Joul/Frame)
Accomplished Bit Rate (Kbit/s)
(b)
0
500
1000
1500
2000
0
1
2
3
4
5
6
Energy Consumed (Joul/Frame)
Accomplished Bit Rate (Kbit/s)
(c)
0
1000
2000
3000
4000
0
1
2
3
4
5
6
Energy Consumed (Joul/Frame)
Accomplished Bit Rate (Kbit/s)
(d)
2000
4000
6000
8000
10000
0
50
100
150
200
Energy Consumed (Joul/Frame)
Accomplished Bit Rate (Kbit/s)
(e)
VP8 Best Configurations
VP8 Good0 Configurations
VP8 Good1 Configurations
VP8 Good2 Configurations
VP8 Good3 Configurations
VP8 Good4 Configurations
VP8 Good5 Configurations
H.264 High Quality No Tunning Configurations
H.264 High Quality SSIM Tunning Configurations
H.264 High Speed No Tunning Configurations
H.264 High Speed SSIM Tunning Configurations
H.264 Normal No Tunning Configurations
H.264 Normal SSIM Tunning Configurations
(f)
Figure
3.
Codecs
comparison
in
encoding
ener
gy
consumption,
(a)
Quality
,
QCIF
and
CIF
,
(b)
Quality
4CIF
,
(c)
Quality
720p,
(d)
Quality
1080p,
(e)
Quality
2160p
0.96
0.97
0.98
0.99
1
0
0.1
0.2
0.3
0.4
Energy (Joul/Frame)
SSIM Quality
(a)
0.9
0.92
0.94
0.96
0.98
0
0.5
1
1.5
Energy (Joul/Frame)
SSIM Quality
(b)
0.93
0.94
0.95
0.96
0.97
0.98
0
1
2
3
4
5
Energy (Joul/Frame)
SSIM Quality
(c)
0.92
0.94
0.96
0.98
1
0
1
2
3
4
5
Energy (Joul/Frame)
SSIM Quality
(d)
0.95
0.96
0.97
0.98
0.99
1
0
50
100
150
200
Energy (Joul/Frame)
SSIM Quality
(e)
H.264 High Quality No Tunning Configurations
H.264 High Quality PSNR Tunning Configurations
H.264 High Quality SSIM Tunning Configurations
H.264 High Speed No Tunning Configurations
H.264 High Speed PSNR Tunning Configurations
H.264 High Speed SSIM Tunning Configurations
H.264 Normal No Tunning Configurations
H.264 Normal PSNR Tunning Configurations
H.264 Normal SSIM Tunning Configurations
(f)
Figure
4.
Codecs
comparison
in
encoding
ener
gy
consumption
and
quality
at
settled
Bit
rate,
(a)
Quality
,
Bitrate=300
Kbps,
QCIF
and
CIF
,
(b)
Quality
,
Bitrate=500
Kbps,
4CIF
,
(c)
Quality
,
Bitrate=800
Kbps,
720p,
(d)
Quality
,
Bitrate=2000
Kbps,
1080p,
(e)
QQuality
,
Bitrate=7000
Kbps,
2160p
Int
J
Elec
&
Comp
Eng,
V
ol.
11,
No.
2,
April
2021
:
1303
–
1310
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1309
5.
CONCLUSIONS
Progression
of
the
prediction
models
utilized
in
a
v
ariety
of
areas
is
a
result
of
the
contrib
ution
of
machine
learning
approaches.
W
e
sho
w
that
we
can
e
xtend
our
e
xperiments
and
ha
v
e
more
data
to
compare
by
predicting
ener
gy
consumption
bas
ed
on
a
machine
learning
model.
The
model
sa
v
ed
us
more
than
fi
v
e
thou-
sand
hours
of
ener
gy
measurements.
It
is
e
xceptionally
imperati
v
e
to
ha
v
e
a
w
ay
of
preparing
data
through
a
machine
learning
model
to
sa
v
e
tim
e
and
it
is
immune
to
error
.
Implementation
computation
comple
xities
ha
v
e
been
compared
of
H.264/A
VC
and
VP8
encoders,
and
pro
v
ed
that
VP8
encoder
prediction
is
more
straight-
forw
ard.
Such
easiness
of
prediction
results
in
speedier
and
lo
wers
ener
gy
encoding
utilization
at
the
encoder
.
which
can
simplify
the
encoding
by
an
ef
ficient
implementation.
Furthermore,
we
ha
v
e
compared
and
ana-
lyzed
the
e
x
ecution
of
H.264/A
VC
and
VP8
encoding
through
broad
tests.
In
terms
of
the
ener
gy
consumption,
H.264/A
VC
is
predominant
to
VP8
encoder
for
most
frame
sizes.
The
results
illustrate
that
H.264/A
VC
by
and
lar
ge
accomplish
superior
to
VP8
Encoder
in
terms
of
performance
and
ener
gy
utilization.
REFERENCES
[1]
Amir
Mosa
vi
and
Abdullah
Bahmani,
”Ener
gy
consumption
prediction
using
machine
learning:
A
re-
vie
w
,
”
Qut
ePrints,
2019,
https://doi.or
g/10.20944/preprints201903.0131.v1.
[2]
Y
ousef
O.
Sharrab
and
Nabil
J.
Sarhan,
”Aggre
g
ate
po
wer
consumption
modeling
of
li
v
e
video
streaming
systems,
”
Proceedings
of
the
A
CM
Multimedia
Systems
Conference,
pp.
60-71,
2013,
https://doi.or
g/10.1145/2483977.2483983.
[3]
Y
aser
Jararweh,
Manar
Bani
Issa,
Mustaf
a
Daraghmeh,
Mahmoud
Al-A
yyoub,
and
Mohammad
A
Alsmi-
rat,
”Ener
gy
ef
ficient
dynamic
resource
management
in
cloud
computing
based
on
logistic
re
gression
model
and
median
absolute
de
viation,
”
Sustainable
Computing:
Informatics
and
Systems,
v
ol.
19,
pp.
262–274,
2018,
https://doi.or
g/10.1016/j.suscom.2018.07.005.
[4]
Holmes
W
ilson,
”
Open
letter
to
google:
Free
vp8,
and
use
it
on
youtube,
”
2010.
[5]
Chouitek
Mama,
Benouzza
Noureddine,
and
Bekk
ouche
Benaissa,
”Control
of
v
ariable
reluctance
ma-
chine
(8/6)
by
artificiel
intelligence
techniques,
”
International
Journal
of
Electrical
and
Computer
Engi-
neering,
v
ol.
10,
no.
2,
pp.
1893-1904,
2020,
https://doi.or
g/10.11591/ijece.v10i2.pp1893-1904.
[6]
Pei
Xie,
Hong
Zhang,
W
eik
e
Y
ou,
Xianfeng
Zhao,
Jianchang
Y
u,
and
Y
i
Ma,
”Adapti
v
e
vp8
ste
g
anog-
raph
y
based
on
deblocking
filtering,
”
Proceedings
of
the
A
CM
W
orkshop
on
Information
Hiding
and
Multimedia
Security
,
pp.
25-30,
2019,
https://doi.or
g/10.1145/3335203.3335711.
[7]
Mauricio
Altamirano
Silv
a,
Raul
Bertone,
and
Jor
g
Schafer
,
”T
opology
distrib
ution
for
video-
conferencing
applica-
tions,
”
IEEE
10th
International
Conference
on
Netw
orks
of
t
h
e
Future
(NoF),
pp.
90–97,
2019,
https://doi.or
g/10.1109/NoF47743.2019.9014957.
[8]
Dmitriy
V
atolin,
Dmitriy
K
ulik
o
v
,
and
Ale
xander
P
arshin,
”Sixth
MPEG-4
A
VC/
H.264
video
codecs
comparison-short
v
ersion,
”
2010.
[9]
P
atrick
Seeling,
Frank
H.P
.
Fitzek,
Ger
go
Ertli
,
Akshay
Pulipaka,
and
Martin
Reisslein.
”V
ideo
netw
ork
traf
fic
and
quality
comparison
of
vp8
and
H.264
svc,
”
Proceedings
of
the
3rd
w
ork-
shop
on
Mobile
video
deli
v
ery
.
Conf.
on
Human
F
actors
in
Computing
Systems,
pp.
33-38,
2010,
https://doi.or
g/10.1145/1878022.1878031.
[10]
Christian
Feller
,
Jur
gen
Wunschmann,
Thorsten
Roll,
and
Albrecht
Rothermel,
”The
vp8
video
codec
-
o
v
ervie
w
and
comparison
to
h.2
64/
a
vc,
”
IEEE
International
Conference
on
Consumer
Electronics
-Berlin
(ICCE-Berlin),
pp.
57–61,
2011,
https://doi.or
g/10.1109/ICCE-Berlin.2011.6031852.
[11]
Y
ousef
O.
Sharrab
and
Nabil
J.
Sarhan,
”Detailed
comparati
v
e
analysis
of
VP8
and
H.264,
”
IEEE
Inter
-
national
Symposium
on
Multimedia,
pp.
133–140,
2012,
https://doi.or
g/10.1109/ISM.2012.33.
[12]
Y
ousef
O.
Sharrab
and
Nabil
Sarhan,
”Accurac
y
and
po
wer
consumption
tradeof
fs
in
video
rate
adaptation
for
computer
vision
applications,
”
IEEE
International
Conference
on
Multimedia
and
Expo,
pp.
410–415,
2012,
https://doi.or
g/10.1109/ICME.2012.77.
[13]
Muhammad
Shafique,
Bastian
Molk
enthin,
and
Jor
g
Henk
el,
”An
HVS-based
adapti
v
e
computational
comple
xity
reduction
scheme
for
H.264/A
VC
video
encoder
using
prognostic
early
mode
e
xclusion,
”
Proceedings
of
Design,
Automation
and
T
est
in
Europe
Conference
and
Exhibi
tion,
pp.
1713–1718,
2010,
https://doi.or
g/10.1109/D
A
TE.2010.5457091.
[14]
Y
ousef
O
Sharrab
and
Nabil
Sarhan,
”Modeling
and
analysi
s
of
po
wer
consumption
in
li
v
e
video
streaming
systems,
”
A
CM
T
ransactions
on
Mul
timedia
Computing,
Communications,
and
Applications
Mac
hine
learning-based
ener
gy
consumption
modeling
and
comparing
of
...
(Y
ousef
O.
Sharr
ab)
Evaluation Warning : The document was created with Spire.PDF for Python.
1310
r
ISSN:
2088-8708
(T
OMM),
v
ol.
13,
no.
4,
pp.
1–25,
2017,
https://doi.or
g/10.1145/3115505.
[15]
Zhou
W
ang,
Alan
C.
Bo
vik,
Hamid
R.
Sheikh,
and
Eero
P
.
Simoncelli,
”Image
quality
assessment:
From
error
visibili
ty
to
structural
similarity
,
”
IEEE
TRANSA
CTIONS
ON
IMA
GE
PR
OCESSING,
v
ol.
13,
no.
4,
pp.
600–612,
2004.
[16]
Omar
Ja
v
ed
and
Mubarak
Shah,
”T
racking
and
object
classification
for
automated
surv
eillance,
”
Proceed-
ings
of
the
7th
European
Conference
on
Computer
V
ision-P
art
IV
,
pp.
343–357,
2002.
[17]
Xiaojing
Y
uan,
Zehang
Sun,
Y
aak
o
v
V
arol,
and
Geor
ge
Bebis,
”A
dist
rib
uted
visual
surv
eillance
system,
”
Proceedings
of
the
IEEE
Conference
on
Adv
anced
V
ideo
and
Signal
Bas
ed
Surv
eillance
(A
VSS),
pp.
199-204,
2003,
https://doi.or
g/10.1109/A
VSS.2003.1217922.
[18]
W
.
Niu
and
J.
Long,
”Human
acti
vity
detection
and
recognition
for
video
surv
eillance,
”
Proceedings
of
the
IEEE
Multimedia
and
Expo
Conference
(ICME),
pp.
719–722,
2004,
https://doi.or
g/10.1109/ICME.2004.1394293.
[19]
Robert
Collins,
Alan
Lipton,
T
ak
eo
Kanade,
Hironob
u
Fujiyoshi,
Da
vid
Duggins,
Y
anghai
Tsin,
Da
vid
T
olli
v
er
,
Nob
uyoshi
Enomoto,
and
Osamu
Hase
g
a
w
a,
”A
system
for
video
surv
eillance
and
monitor
-
ing,
”
T
echnical
Report
CMU-RI-TR-00-12,
Robotics
Institute,
Carne
gie
Mellon
Uni
v
ersity
,
Pittsb
ur
gh,
P
A,
pp.
1-68,
2000.
[20]
Mohammad
Alsmirat
and
Nabil
J
Sarhan,
”Cross-layer
optimization
for
automated
video
surv
eillance,
”
IEEE
International
Symposium
on
Multimedia
(ISM),
pp.
243–246,
2016,
https://doi.or
g/10.1109/ISM.2016.0055.
[21]
Mahmoud
Al-A
yyoub,
Nusaiba
Al-Mnayyis,
Mohammad
A
Alsmirat,
Khaled
Ala
wneh,
Y
aser
Jararweh,
and
Brij
B
Gupta,
”Sift
based
roi
e
xtraction
for
lumbar
disk
herniation
cad
system
from
mri
axial
scans,
”
Journal
of
Ambient
Intelligence
and
Humanized
Computing,
pp.
1-9,
2018.
[22]
Christian
Sze
gedy
,
Ser
ge
y
Iof
fe,
and
V
incent
V
anhouck
e,
”Inception-v4,
i
nception-resnet
and
the
im-
pact
of
residual
connections
on
learning,
”
Proceedings
of
the
Thirty-First
AAAI
Conference
on
Artificial
Intelligence
(AAAI-17),
2017.
[23]
Kaiming
He,
Xiangyu
Zhang,
Shaoqing
Ren,
a
n
d
Jian
Sun,
”Deep
residual
learning
for
image
recog-
nition,
”
Proceedings
of
the
IEEE
Conference
on
Computer
V
i
sion
and
P
attern
Recognition
(CVPR),
pp.
770–778,
2016.
[24]
W
eiyang
Liu,
Y
andong
W
en,
Zhiding
Y
u,
Ming
Li,
Bhiksha
Raj,
and
Le
Song,
”Spheref
ace:
Deep
h
yper
-
sphere
embedding
for
f
ace
recognition,
”
Proceedings
of
The
IEEE
Conference
on
Computer
V
ision
and
P
attern
Recognition
(CVPR),
v
ol.
1,
pp.
1,
2017.
[25]
W
aqas
Sultani,
Chen
Chen,
and
Mubarak
Shah,
”
Real-w
orld
anomaly
detection
in
surv
eillance
videos,
”
Proceedings
of
the
IEEE
Conference
on
Computer
V
ision
and
P
attern
Recognition,
2018.
[26]
Lukas
Esterle
and
Peter
R
Le
wis,
”Online
multi-object
k-co
v
erage
with
mobile
smart
cameras,
”
Proceed-
ings
of
the
11th
International
Conference
on
Distrib
uted
Smart
Cameras
(ICDSC),
pp.
107–112,
2017.
[27]
Iain
E.
G.
Richardson.
The
H.264
Adv
anced
V
ideo
Compression
Standrad,
second
edition,
WILEY
,
Pub-
lished
Online.
Vcode
x
Limited,
UK,
2010.
[28]
Jongho
Kim,
Dongh
yung
Kim,
and
Jechang
Jeong,
”Comple
xity
reduction
algorithm
for
intra
mode
se-
lection
in
H.264/A
VC
video
coding,
”
Adv
anced
Concepts
for
Intelligent
V
isi
on
Systems
(A
CIVS’06),
v
ol.
4179,
pp.
454–465,
2006.
[29]
Jason
Garrett-Glaser
,
”The
first
in-depth
technical
analysis
of
VP8,
”
http://x264de
v
.multimedia.cx/archi
v
es/377
.
[30]
Jim
Bank
oski,
P
aul
W
ilkins,
and
Y
ao
wu
Xu,
”T
echnical
o
v
ervie
w
of
VP8,
an
open
source
video
codec
for
the
web,
”
IEEE
International
Conference
on
Multimedia
and
Expo,
pp.
1-6,
2011.
Int
J
Elec
&
Comp
Eng,
V
ol.
11,
No.
2,
April
2021
:
1303
–
1310
Evaluation Warning : The document was created with Spire.PDF for Python.