I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
2
,
A
pr
il
2025
, pp.
1518
~
1530
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
2
.pp
1518
-
1530
1518
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
D
e
e
p
l
e
ar
n
i
n
g
-
b
ase
d
t
e
c
h
n
i
q
u
e
s
f
or
vi
d
e
o
e
n
h
an
c
e
m
e
n
t
,
c
om
p
r
e
ss
i
on
an
d
r
e
st
or
at
i
on
R
e
d
ou
an
e
L
h
ia
d
i
,
A
b
d
e
s
s
am
ad
Jad
d
ar
,
A
b
d
e
la
li
K
aaou
ac
h
i
N
a
t
i
ona
l
S
c
hool
of
bus
i
ne
s
s
a
nd
M
a
na
ge
m
e
nt
,
U
ni
ve
r
s
i
t
y
of
M
oha
m
m
e
d
1s
t
,
O
uj
da
,
M
or
oc
c
o
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
J
ul
30, 2024
R
e
vi
s
e
d
O
c
t
29, 2024
A
c
c
e
pt
e
d
N
ov 14, 2024
Video
processing
is
essential
in
entertainment,
surveillance,
and
communi
cation.
This
research
presents
a
strong
framework
that
im
proves
video
clarity
and
decreases
bitrate
via
advanced
restorati
on
and
compressi
on
methods.
The
suggested
framewo
rk
m
erges
various
deep
learning
models
such
as
super
-
resolutio
n,
d
eblurring,
denoising
,
and
frame
int
erpolat
ion,
in
additio
n
to
a
competent
compressi
on
model.
Video
frames
ar
e
first
compressed
using
the
libx26
5
codec
in
order
to
reduce
bitrate
and
s
torage
needs.
After
compression,
restoration
techniques
deal
with
issues
like
noise,
blur,
and
loss
of
detail.
The
video
restoration
transformer
(VRT)
use
s
deep
learning
to
greatly
enhance
video
quality
by
reducing
compression
artifac
ts.
The frame resolution is improved by
the super
-
resolutio
n model,
moti
on blur
is
fixed
by
the
deblurring
model,
and
noise
is
reduced
by
the
de
noising
model,
resulting
in
clearer
frames.
Frame
interpolation
creates
add
itional
frames
between
existing
frames
to
create
a
smoother
video
v
iewing
experience.
Experimen
tal
finding
s
show
that
this
system
succe
ssfully
improves
video
quality
and
decrea
ses
artifac
ts,
providing
better
per
ceptual
quality
and
fidelity.
The
real
-
time
processing
capabilities
of
the
tech
nology
make
it
well
-
suited
for
use
in
video
streaming,
surveillance,
and
digital
cinema.
K
e
y
w
o
r
d
s
:
D
e
e
p
le
a
r
ni
ng
R
e
a
l
-
time
pr
oc
e
s
s
in
g
R
e
s
to
r
a
ti
on
m
ode
ls
S
upe
r
-
r
e
s
ol
ut
io
n
V
id
e
o
pr
oc
e
s
s
in
g
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
R
e
doua
ne
L
hi
a
di
N
a
ti
ona
l
S
c
hool
of
bus
in
e
s
s
a
nd
M
a
na
g
e
m
e
nt
,
U
ni
ve
r
s
it
y
of
M
oha
m
m
e
d
1s
t
O
uj
da
,
M
or
oc
c
o
E
m
a
il
:
lh
ia
di
.r
e
doua
ne
@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
he
a
dve
nt
of
de
e
p
le
a
r
ni
ng
ha
s
r
e
vol
ut
io
ni
z
e
d
vi
d
e
o
r
e
s
to
r
a
t
io
n
by
e
na
bl
in
g
th
e
de
ve
lo
pm
e
nt
of
s
ophi
s
ti
c
a
te
d
m
od
e
ls
c
a
pa
bl
e
of
und
e
r
s
ta
ndi
ng
c
om
pl
e
x
da
ta
r
e
la
ti
ons
hi
ps
a
nd
a
c
hi
e
vi
ng
s
upe
r
io
r
r
e
s
ul
ts
.
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
s
)
a
nd
a
tt
e
nt
io
n
m
e
c
h
a
ni
s
m
s
a
r
e
a
t
th
e
f
or
e
f
r
ont
of
th
e
s
e
a
dv
a
nc
e
m
e
nt
s
,
a
ddr
e
s
s
in
g
va
r
io
us
a
s
pe
c
ts
of
vi
de
o
qua
li
ty
,
in
c
lu
di
ng
r
e
s
ol
ut
io
n
e
nha
nc
e
m
e
nt
,
s
ha
r
pne
s
s
im
pr
ove
m
e
nt
,
a
nd
noi
s
e
r
e
duc
ti
on
[
1]
,
[
2]
.
I
n
c
ont
r
a
s
t,
tr
a
di
ti
ona
l
vi
de
o
r
e
s
to
r
a
ti
on
te
c
hni
que
s
,
w
hi
c
h
r
e
ly
on
h
e
ur
is
ti
c
-
ba
s
e
d
m
e
th
ods
a
nd
m
a
nua
ll
y
c
r
a
f
te
d
f
e
a
tu
r
e
s
,
of
te
n
s
tr
uggl
e
to
e
f
f
e
c
ti
ve
ly
m
a
na
ge
in
tr
ic
a
te
de
gr
a
da
ti
on
pa
tt
e
r
ns
a
nd
c
om
pr
e
s
s
io
n
a
r
ti
f
a
c
ts
[
3]
.
D
e
e
p
l
e
a
r
ni
ng
m
ode
ls
,
le
ve
r
a
gi
ng
C
N
N
s
,
e
x
c
e
l
a
t
c
a
pt
ur
in
g
hi
e
r
a
r
c
hi
c
a
l
r
e
pr
e
s
e
nt
a
ti
ons
a
nd
e
nha
nc
in
g
vi
de
o
qua
li
ty
by
pr
ovi
di
ng
tr
a
ns
la
ti
on
in
va
r
ia
nc
e
a
nd
r
obus
t
p
a
tt
e
r
n
r
e
c
ogni
ti
on
[
4]
,
[
5]
.
F
ig
ur
e
1
il
lu
s
tr
a
te
s
th
e
tr
a
di
ti
ona
l
vi
de
o
c
om
pr
e
s
s
io
n
pr
oc
e
s
s
,
out
li
ni
ng
it
s
k
e
y
c
om
pone
nt
s
a
nd
w
or
kf
lo
w
.
T
hi
s
vi
s
ua
l
r
e
pr
e
s
e
nt
a
ti
on
hi
ghl
ig
ht
s
th
e
li
m
it
a
ti
ons
a
nd
c
ha
ll
e
nge
s
of
c
onve
nt
io
na
l
te
c
hni
que
s
, p
a
r
ti
c
ul
a
r
ly
i
n m
a
na
gi
ng c
om
pr
e
s
s
io
n
a
r
ti
f
a
c
ts
a
nd de
gr
a
da
ti
on pa
tt
e
r
ns
.
D
e
s
pi
te
s
ig
ni
f
ic
a
nt
a
dva
n
c
e
m
e
nt
s
,
not
a
bl
e
ga
ps
r
e
m
a
in
in
pr
e
vi
ous
r
e
s
e
a
r
c
h.
F
or
e
x
a
m
pl
e
,
w
hi
le
s
om
e
s
tu
di
e
s
ha
ve
e
xpl
or
e
d
th
e
im
pa
c
t
of
c
om
pr
e
s
s
io
n
a
r
ti
f
a
c
t
s
on
vi
de
o
qua
li
ty
[
4]
,
th
e
r
e
ha
s
b
e
e
n
li
m
it
e
d
f
oc
us
on
how
a
dva
nc
e
d
r
e
s
to
r
a
ti
on
te
c
hni
que
s
in
f
lu
e
nc
e
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
c
om
pr
e
s
s
io
n
m
ode
ls
.
P
r
e
vi
ous
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
D
e
e
p l
e
ar
ni
ng
-
bas
e
d t
e
c
hni
que
s
f
o
r
v
id
e
o e
nhan
c
e
m
e
nt
, c
om
p
r
e
s
s
io
n and r
e
s
to
r
at
io
n
(
R
e
douane
L
hi
adi
)
1519
w
or
k
ha
s
of
te
n
c
onc
e
nt
r
a
te
d
on
e
it
he
r
r
e
s
to
r
a
ti
on
o
r
c
om
pr
e
s
s
io
n,
w
it
h
a
c
om
pr
e
he
ns
iv
e
f
r
a
m
e
w
or
k
in
te
gr
a
ti
ng
bot
h
a
s
pe
c
t
s
be
in
g
not
a
bl
y
a
bs
e
nt
.
F
ur
th
e
r
m
or
e
,
th
e
gr
ow
in
g
de
m
a
nd
f
or
hi
gh
-
qua
li
ty
di
gi
ta
l
vi
de
o
c
ont
e
nt
ha
s
he
ig
ht
e
n
e
d
th
e
ne
e
d
f
or
r
e
a
l
-
ti
m
e
a
ppl
ic
a
ti
on
of
a
dva
nc
e
d
r
e
s
to
r
a
ti
on
m
ode
ls
in
f
ie
ld
s
s
uc
h
a
s
vi
de
o s
tr
e
a
m
in
g,
s
ur
ve
il
la
nc
e
, a
nd digi
ta
l
c
in
e
m
a
[
5]
,
[
6]
.
T
hi
s
pa
pe
r
a
im
s
to
f
il
l
th
e
s
e
voi
ds
by
in
tr
oduc
in
g
a
n
in
nova
ti
ve
f
r
a
m
e
w
or
k
th
a
t
c
om
bi
ne
s
c
ut
ti
ng
-
e
dge
r
e
s
to
r
a
ti
on
a
nd
c
om
pr
e
s
s
io
n
te
c
hni
que
s
.
T
hi
s
r
e
s
e
a
r
c
h
e
nha
nc
e
s
vi
d
e
o
qua
li
ty
a
nd
r
e
duc
e
s
c
om
pr
e
s
s
io
n
a
r
ti
f
a
c
ts
by
us
in
g
a
dva
n
c
e
d
m
ode
ls
li
ke
s
up
e
r
-
r
e
s
ol
ut
io
n,
de
bl
ur
r
in
g,
de
noi
s
in
g,
a
nd
f
r
a
m
e
in
te
r
pol
a
ti
on
in
c
om
bi
na
ti
on
w
it
h
th
e
li
bx265
c
om
p
r
e
s
s
io
n
c
ode
c
.
O
ur
m
e
th
od
e
nha
nc
e
s
vi
de
o
qua
li
ty
a
nd
a
c
c
ur
a
c
y w
hi
le
a
l
s
o pr
ovi
di
ng r
e
a
l
-
ti
m
e
pr
oc
e
s
s
in
g f
e
a
tu
r
e
s
, m
a
ki
ng i
t
id
e
a
l
f
or
va
r
io
us
us
e
s
.
F
ig
ur
e
1. B
lo
c
k di
a
gr
a
m
i
ll
us
tr
a
ti
ng t
he
c
onve
nt
io
na
l
m
e
th
od of
vi
de
o c
om
pr
e
s
s
io
n
2.
M
O
T
I
V
A
T
I
O
N
C
onve
nt
io
na
l
vi
de
o
r
e
s
to
r
a
ti
on
te
c
hni
que
s
f
a
c
e
s
ig
ni
f
ic
a
nt
c
ha
ll
e
nge
s
in
m
a
na
gi
ng
c
om
pr
e
s
s
io
n
a
r
ti
f
a
c
ts
a
nd
e
nha
nc
in
g
vi
s
ua
l
qua
li
ty
.
T
r
a
di
ti
ona
l
m
e
th
ods
,
w
hi
c
h
of
te
n
r
e
ly
on
he
ur
is
ti
c
a
ppr
oa
c
he
s
a
nd
m
a
nua
ll
y
c
r
a
f
te
d
f
e
a
tu
r
e
s
,
s
tr
uggl
e
to
a
ddr
e
s
s
th
e
c
om
pl
e
x
de
gr
a
da
ti
on
pa
tt
e
r
ns
in
tr
oduc
e
d
dur
in
g
vi
de
o
c
om
pr
e
s
s
io
n.
R
e
c
ogni
z
in
g
th
e
s
e
li
m
it
a
ti
ons
,
th
is
r
e
s
e
a
r
c
h
in
tr
o
duc
e
s
a
n
in
nova
ti
ve
vi
de
o
r
e
s
to
r
a
ti
on
pi
pe
li
ne
th
a
t
le
ve
r
a
ge
s
t
he
s
tr
e
ngt
hs
of
de
e
p l
e
a
r
ni
ng
m
ode
ls
a
nd c
ut
ti
ng
-
e
dge
c
om
pr
e
s
s
io
n a
lg
or
it
hm
s
.
O
ur
pr
opos
e
d
pi
pe
li
ne
in
te
gr
a
te
s
a
dva
nc
e
d
d
e
e
p
le
a
r
ni
ng
te
c
hni
que
s
,
in
c
lu
di
ng
s
upe
r
-
r
e
s
ol
ut
io
n,
de
bl
ur
r
in
g,
a
nd
de
noi
s
in
g,
w
it
h
a
hi
gh
-
pe
r
f
or
m
a
nc
e
c
om
pr
e
s
s
io
n
a
lg
or
it
hm
,
s
pe
c
if
ic
a
ll
y
th
e
li
bx265
c
ode
c
[
5]
. T
hi
s
i
nt
e
gr
a
ti
on be
gi
ns
w
it
h c
om
pr
e
s
s
in
g t
he
i
nput
vi
de
o f
r
a
m
e
s
us
in
g l
ib
x265, whic
h e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
bi
tr
a
te
a
nd
s
to
r
a
ge
r
e
qui
r
e
m
e
nt
s
.
S
ubs
e
que
nt
ly
,
th
e
c
om
pr
e
s
s
e
d
f
r
a
m
e
s
a
r
e
pr
oc
e
s
s
e
d
th
r
ough
our
vi
de
o
r
e
s
to
r
a
ti
on
m
odul
e
,
w
he
r
e
pr
e
tr
a
in
e
d
de
e
p
le
a
r
ni
ng
m
ode
ls
a
ddr
e
s
s
a
r
ti
f
a
c
ts
a
nd
e
nha
nc
e
vi
de
o
qua
li
ty
.
F
ig
ur
e
2
pr
ovi
de
s
a
vi
s
ua
l
r
e
pr
e
s
e
nt
a
ti
on
of
th
e
tr
a
di
ti
ona
l
vi
de
o
r
e
s
to
r
a
ti
on
w
or
kf
lo
w
,
out
li
ni
ng
it
s
pr
oc
e
s
s
e
s
a
nd
in
he
r
e
nt
li
m
it
a
ti
ons
.
T
hi
s
il
lu
s
tr
a
ti
on
s
e
r
ve
s
a
s
a
f
ounda
ti
on
f
or
unde
r
s
ta
ndi
ng
how
ou
r
a
ppr
oa
c
h
im
pr
ove
s
upon
c
onve
nt
io
na
l
m
e
th
ods
.
B
y
c
om
bi
ni
ng
a
dv
a
nc
e
d
r
e
s
to
r
a
ti
on
m
ode
ls
w
it
h
c
ut
ti
ng
-
e
dge
c
om
pr
e
s
s
io
n
te
c
hni
que
s
,
our
pi
pe
li
ne
a
im
s
to
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
vi
s
ua
l
f
id
e
li
ty
a
nd
pe
r
c
e
pt
ua
l
qua
li
ty
.
M
or
e
ove
r
,
our
f
r
a
m
e
w
or
k
is
de
s
ig
ne
d
to
be
a
da
pt
a
bl
e
a
nd
s
c
a
la
bl
e
,
m
a
ki
ng
it
s
ui
ta
bl
e
f
or
di
ve
r
s
e
vi
de
o
pr
oc
e
s
s
in
g
a
ppl
ic
a
ti
ons
,
in
c
lu
di
ng
vi
de
o
s
tr
e
a
m
in
g,
s
ur
ve
il
la
nc
e
,
a
nd
di
gi
ta
l
e
nt
e
r
ta
in
m
e
nt
[
4]
.
T
he
c
ol
la
bor
a
ti
on
be
twe
e
n
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
r
e
s
to
r
a
ti
on
m
ode
ls
a
nd
e
f
f
ic
ie
nt
c
om
pr
e
s
s
io
n
a
lg
or
it
hm
s
of
f
e
r
s
pr
om
is
in
g
a
dva
nc
e
m
e
nt
s
in
vi
de
o
qua
li
ty
e
nh
a
nc
e
m
e
nt
,
a
ddr
e
s
s
in
g
bot
h
c
ur
r
e
nt
li
m
it
a
ti
ons
a
nd
f
ut
ur
e
ne
e
ds
i
n t
he
f
ie
ld
.
F
ig
ur
e
2. S
c
he
m
a
ti
c
r
e
pr
e
s
e
nt
a
ti
on of
t
r
a
di
ti
ona
l
vi
de
o r
e
s
to
r
a
ti
on pr
oc
e
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
2
,
A
pr
il
2025
:
1518
-
1530
1520
3.
R
E
L
A
T
E
D
WORK
R
e
c
e
nt
ly
,
th
e
r
e
ha
ve
be
e
n
not
a
bl
e
de
v
e
lo
pm
e
nt
s
in
m
e
th
ods
f
or
c
om
pr
e
s
s
in
g
im
a
ge
s
.
C
onvolut
io
na
l
a
ut
oe
nc
ode
r
s
[
5]
s
how
pot
e
nt
ia
l
f
or
e
f
f
e
c
ti
ve
c
om
pr
e
s
s
io
n
w
it
h
pr
e
s
e
r
ve
d
im
a
ge
qua
li
ty
.
F
ur
th
e
r
m
or
e
,
c
om
p
r
e
s
s
io
n
te
c
hni
que
s
th
a
t
a
r
e
opt
im
iz
e
d
f
r
om
one
e
nd
to
a
not
he
r
a
nd
us
e
tr
a
ns
f
or
m
s
ba
s
e
d
on
f
r
e
que
nc
y
ha
ve
s
how
n
be
tt
e
r
r
e
s
ul
ts
in
r
e
duc
in
g
bi
t
r
a
t
e
w
it
hout
c
om
pr
om
is
in
g
pe
r
c
e
pt
ua
l
qua
li
ty
.
A
s
s
e
s
s
in
g
c
om
pr
e
s
s
io
n
a
lg
or
it
hm
s
f
r
e
que
nt
ly
in
c
lu
de
s
s
ubj
e
c
ti
ve
qua
li
ty
e
va
lu
a
ti
ons
[
6]
,
w
hi
c
h
r
e
ve
a
l
im
por
ta
nt
in
f
or
m
a
ti
on
a
bout
th
e
pe
r
c
e
iv
e
d
qua
li
ty
of
c
om
pr
e
s
s
e
d
vi
de
o
s
.
D
e
e
p
le
a
r
ni
ng
te
c
hni
que
s
[
7]
a
r
e
now
be
in
g
us
e
d
e
f
f
e
c
ti
ve
ly
f
or
im
a
ge
c
om
pr
e
s
s
io
n
by
ut
il
iz
in
g
e
nd
-
to
-
e
nd
le
a
r
ni
ng
to
e
nha
nc
e
c
om
pr
e
s
s
io
n
pe
r
f
or
m
a
nc
e
.
S
upe
r
-
r
e
s
ol
ut
io
n
te
c
hni
que
s
in
vi
de
o
pr
oc
e
s
s
in
g
ha
ve
be
c
om
e
popula
r
f
or
im
pr
ovi
ng
th
e
r
e
s
ol
ut
io
n
of
vi
de
o
s
e
que
nc
e
s
in
r
e
a
l
-
ti
m
e
a
ppl
ic
a
ti
ons
[
8]
.
T
r
a
ns
f
or
m
e
r
-
ba
s
e
d
te
c
hni
que
s
s
uc
h
a
s
S
w
in
I
R
ha
ve
di
s
pl
a
ye
d
im
pr
e
s
s
iv
e
out
c
om
e
s
in
im
a
ge
e
nha
nc
e
m
e
nt
dut
ie
s
li
ke
s
upe
r
-
r
e
s
ol
ut
io
n
a
nd
de
noi
s
in
g.
R
e
c
e
nt
pr
ogr
e
s
s
in
vi
de
o
s
upe
r
-
r
e
s
ol
ut
io
n
ha
s
be
e
n
c
onc
e
nt
r
a
te
d
on
e
nha
nc
in
g
f
e
a
tu
r
e
p
r
opa
ga
ti
on
a
nd
a
li
gnm
e
nt
te
c
hni
que
s
,
le
a
di
ng
to
im
pr
ove
d
pe
r
f
or
m
a
nc
e
in
vi
de
o
s
upe
r
-
r
e
s
ol
ut
io
n
a
s
s
ig
nm
e
nt
s
[
8]
.
B
a
s
ic
r
e
s
e
a
r
c
h
on
ne
c
e
s
s
a
r
y
e
le
m
e
nt
s
f
or
im
pr
ovi
ng
vi
de
o
qua
li
ty
[
8]
ha
s
of
f
e
r
e
d
im
por
ta
nt
unde
r
s
ta
ndi
ng
of
th
e
c
r
uc
ia
l
a
s
pe
c
ts
th
a
t
im
pa
c
t
m
od
e
l
e
f
f
e
c
ti
ve
ne
s
s
.
S
ubs
ta
nt
ia
l
a
dva
nc
e
m
e
nt
s
ha
ve
be
e
n
a
c
hi
e
v
e
d
in
th
e
a
r
e
a
of
vi
de
o de
bl
ur
r
in
g t
e
c
hni
que
s
, s
pe
c
if
ic
a
ll
y by uti
li
z
in
g c
a
s
c
a
de
d de
e
p l
e
a
r
ni
ng me
th
ods
t
ha
t
e
xpl
oi
t
te
m
por
a
l
da
ta
to
im
pr
ove
de
bl
ur
r
in
g
e
f
f
ic
ie
nc
y
[
8]
.
D
e
e
p
le
a
r
ni
ng
te
c
hni
que
s
ha
ve
be
e
n
a
ppl
ie
d
to
vi
de
o
de
bl
ur
r
in
g
w
it
h
a
f
oc
us
on
r
e
duc
in
g
m
ot
io
n
bl
ur
a
r
ti
f
a
c
ts
,
w
hi
c
h
l
e
a
ds
t
o
e
nha
nc
e
d
vi
s
ua
l
qua
li
ty
in
ha
ndh
e
ld
vi
de
o
r
e
c
or
di
ngs
.
M
e
th
ods
s
u
c
h
a
s
e
nha
nc
e
d
d
e
f
or
m
a
bl
e
vi
de
o
r
e
s
to
r
a
ti
on
(
E
D
V
R
)
ha
ve
e
f
f
e
c
ti
ve
ly
ut
il
iz
e
d
e
nha
nc
e
d
de
f
or
m
a
bl
e
c
onvolut
io
na
l
ne
twor
ks
to
pr
oduc
e
r
e
m
a
r
ka
bl
e
out
c
om
e
s
in
di
f
f
e
r
e
nt
vi
de
o
r
e
s
to
r
a
ti
on
ta
s
ks
li
ke
s
upe
r
-
r
e
s
ol
ut
io
n
or
de
bl
ur
r
in
g.
M
or
e
ove
r
,
e
xi
s
ti
ng
vi
de
o
de
bl
ur
r
in
g
te
c
hni
que
s
[
8]
ha
ve
in
c
or
por
a
te
d
bl
ur
-
in
va
r
ia
nt
m
ot
io
n
e
s
ti
m
a
ti
on
m
e
th
od
s
to
im
pr
ove
de
bl
ur
r
in
g
a
lg
or
it
hm
e
f
f
e
c
ti
ve
ne
s
s
.
T
o
unde
r
s
ta
nd
th
e
a
ppr
oa
c
h
de
s
c
r
ib
e
d
in
th
is
s
e
c
ti
on,
a
nd
to
il
lu
s
tr
a
te
th
e
pr
oc
e
s
s
e
s
in
vol
ve
d
in
de
bl
ur
r
in
g,
F
ig
ur
e
3
pr
e
s
e
nt
s
a
vi
s
ua
l
de
pi
c
ti
on of
t
he
f
lo
w
a
nd ke
y s
ta
ge
s
ne
c
e
s
s
a
r
y f
or
unde
r
s
ta
ndi
ng t
he
de
bl
ur
r
in
g t
e
c
hni
que
.
F
ig
ur
e
3. F
lo
w
c
ha
r
t
of
i
m
a
ge
de
bl
ur
r
in
g p
r
oc
e
s
s
D
e
bl
ur
r
in
g
a
lg
or
it
hm
:
=
∗
+
w
he
r
e
n
is
t
he
noi
s
e
a
f
f
e
c
ti
ng t
he
i
m
a
ge
f
−
I
nput
:
b
lu
r
r
y w
it
h
n
o
is
y i
m
a
ge
f
.
−
D
e
c
onvolut
io
n:
t
he
pr
oc
e
s
s
in
vol
ve
s
r
e
s
to
r
in
g
th
e
or
ig
in
a
l
i
m
a
ge
g
f
r
om
th
e
obs
e
r
ve
d
im
a
ge
f
us
in
g
th
e
bl
ur
ke
r
ne
l
p
.
−
N
on
-
bl
in
d
de
c
onvolut
io
n:
if
th
e
bl
u
r
ke
r
ne
l
p
is
known
or
ob
ta
in
a
bl
e
,
non
-
bl
in
d
de
c
onvolut
io
n
m
e
th
ods
a
r
e
a
ppl
ie
d.
−
R
e
c
ons
tr
uc
ti
on:
or
ig
in
a
l
im
a
ge
g
is
r
e
c
ons
tr
uc
te
d u
s
in
g s
pe
c
if
ic
de
c
onvolut
io
n ope
r
a
to
r
s
.
−
O
ut
put
:
c
le
a
r
a
nd nois
e
-
f
r
e
e
i
m
a
ge
g
.
4.
M
E
T
H
O
D
4.1.
D
at
a aq
c
u
as
it
io
n
an
d
p
r
e
p
r
oc
e
s
s
in
g
I
n or
de
r
t
o c
ol
le
c
t
th
e
ne
c
e
s
s
a
r
y vi
de
o da
ta
f
or
our
e
xpe
r
im
e
nt
s
, w
e
e
m
pl
oye
d a
P
yt
hon
s
c
r
ip
t
th
a
t
m
a
ke
s
us
e
of
th
e
F
F
m
pe
g
li
br
a
r
y.
T
he
s
c
r
ip
t
is
d
e
s
ig
ne
d
to
w
o
r
k
w
it
h
dyna
m
ic
vi
de
o
da
ta
s
e
t
s
,
in
c
lu
di
ng
th
e
"
your
ow
n
vi
de
o"
,
a
nd
it
e
xt
r
a
c
t
s
s
in
gl
e
f
r
a
m
e
s
a
t
a
s
te
a
dy
f
r
a
m
e
r
a
te
of
15
f
r
a
m
e
s
pe
r
s
e
c
ond.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
D
e
e
p l
e
ar
ni
ng
-
bas
e
d t
e
c
hni
que
s
f
o
r
v
id
e
o e
nhan
c
e
m
e
nt
, c
om
p
r
e
s
s
io
n and r
e
s
to
r
at
io
n
(
R
e
douane
L
hi
adi
)
1521
T
hi
s
f
r
a
m
e
r
a
te
gua
r
a
nt
e
e
s
e
xt
e
ns
iv
e
c
ove
r
a
g
e
of
c
ont
e
nt
a
nd
r
e
s
ol
ut
io
ns
,
w
hi
c
h
in
tu
r
n
e
na
bl
e
s
th
or
ough
te
s
ti
ng of
our
hybr
id
c
om
pr
e
s
s
io
n a
nd r
e
s
to
r
a
ti
on
a
ppr
oa
c
h
[
9]
.
4.2.
C
om
p
r
e
s
s
io
n
m
od
e
l
T
o
pr
e
s
e
r
ve
a
s
a
ti
s
f
a
c
to
r
y
pe
r
c
e
pt
ua
l
qu
a
li
ty
of
th
e
in
put
vi
d
e
o,
w
e
ha
v
e
ut
il
iz
e
d
a
lo
s
s
y
s
tr
a
te
gy
ba
s
e
d
on
hi
gh
e
f
f
ic
ie
nc
y
vi
de
o c
odi
ng
(
H
E
V
C
)
to
de
c
r
e
a
s
e
it
s
bi
tr
a
te
.
I
n
or
de
r
to
a
c
c
om
pl
is
h
th
is
,
w
e
c
r
e
a
te
d
a
P
yt
hon
f
unc
ti
on
th
a
t
m
a
ke
s
us
e
of
th
e
F
F
m
pe
g
li
br
a
r
y.
T
hi
s
f
unc
ti
on
e
nc
ode
s
th
e
in
put
vi
de
o
ut
il
iz
in
g
th
e
"
li
bx265"
c
ode
c
w
it
h
a
de
s
ig
na
te
d
c
on
s
ta
nt
r
a
te
f
a
c
to
r
(
C
R
F
)
va
lu
e
[
10]
.
F
ur
th
e
r
m
or
e
,
w
e
ha
ve
in
c
lu
de
d
a
r
e
duc
ti
on
in
r
e
s
ol
ut
io
n
of
th
e
vi
de
o
f
r
a
m
e
s
to
one
-
f
our
th
of
th
e
ir
or
ig
in
a
l
s
iz
e
in
or
de
r
to
f
ur
th
e
r
lo
w
e
r
th
e
bi
tr
a
te
.
T
he
f
unc
ti
on
ne
e
d
s
th
e
pa
th
to
th
e
vi
d
e
o
f
il
e
in
put
,
th
e
pa
th
to
th
e
vi
de
o
f
il
e
out
put
f
or
c
om
pr
e
s
s
io
n,
a
nd
opt
io
na
l
pa
r
a
m
e
te
r
s
li
ke
C
R
F
va
lu
e
a
nd
out
put
r
e
s
ol
ut
io
n. T
he
C
R
F
va
lu
e
is
ty
pi
c
a
ll
y
in
th
e
r
a
nge
of
28,
s
tr
ik
in
g
a
ba
la
nc
e
be
twe
e
n c
om
pr
e
s
s
io
n
e
f
f
ic
ie
nc
y
a
nd
vi
s
ua
l
q
ua
li
ty
.
T
he
out
put
r
e
s
ol
ut
io
n
is
dow
ns
c
a
le
d
to
one
-
f
our
th
of
th
e
or
ig
in
a
l
vi
de
o
r
e
s
ol
ut
io
n
to
f
a
c
il
it
a
te
e
f
f
ic
ie
nt
pr
oc
e
s
s
in
g
a
nd
s
to
r
a
ge
.
T
o
a
ppl
y
th
e
de
s
ir
e
d
vi
de
o
s
c
a
li
ng
a
nd
c
om
pr
e
s
s
io
n
s
e
tt
in
gs
,
w
e
c
ons
tr
uc
t
th
e
F
F
m
pe
g
c
om
m
a
nd.
T
h
e
"
li
bx265"
c
ode
c
is
us
e
d
to
e
nc
ode
th
e
vi
de
o
f
r
a
m
e
s
w
it
h
th
e
s
pe
c
if
ie
d
C
R
F
v
a
lu
e
,
r
e
s
ul
ti
n
g
in
a
lo
s
s
y
c
om
pr
e
s
s
io
n
pr
oc
e
s
s
th
a
t
r
e
du
c
e
s
th
e
vi
de
o’
s
bi
tr
a
te
w
hi
le
pr
e
s
e
r
vi
ng
pe
r
c
e
pt
ua
ll
y
r
e
le
v
a
nt
in
f
or
m
a
ti
on.
T
he
c
om
pr
e
s
s
e
d
vi
de
o
is
th
e
n
s
a
ve
d
to
th
e
s
pe
c
if
ie
d f
il
e
pa
th
, r
e
a
dy f
or
s
ubs
e
que
nt
pr
oc
e
s
s
in
g
a
nd e
va
lu
a
ti
on
[
11]
.
4.3.
R
e
s
t
or
at
io
n
m
od
e
l
4.3.1.
O
ve
r
al
l
f
r
a
m
e
w
or
k
T
he
r
e
s
to
r
a
ti
on
m
ode
l
c
om
pr
is
e
s
two
ty
pe
s
of
f
r
a
m
e
s
:
I
L
Q
,
r
e
pr
e
s
e
nt
in
g
a
s
e
que
nc
e
of
lo
w
-
qua
li
ty
in
put
f
r
a
m
e
s
, a
nd
I
H
Q
, i
ndi
c
a
ti
ng high
-
qua
li
ty
t
a
r
ge
t
f
r
a
m
e
s
. W
it
hi
n t
hi
s
c
ont
e
xt
:
−
T
:
to
ta
l
num
be
r
of
f
r
a
m
e
s
,
−
H
:
he
ig
ht
of
e
a
c
h f
r
a
m
e
(
ups
c
a
le
d)
,
−
W
:
w
id
th
of
e
a
c
h f
r
a
m
e
(
ups
c
a
le
d)
,
−
C
in
:
num
be
r
of
i
nput
c
ha
nne
ls
,
−
C
out
:
num
be
r
of
out
put
c
ha
nne
ls
,
−
s
:
ups
c
a
li
ng
f
a
c
to
r
f
or
t
a
s
ks
l
ik
e
vi
de
o
s
upe
r
-
r
e
s
ol
ut
io
n,
−
RT
:
num
be
r
of
f
r
a
m
e
s
i
n t
he
s
e
que
nc
e
.
T
he
pr
opos
e
d
vi
de
o
r
e
s
to
r
a
ti
on
tr
a
ns
f
or
m
e
r
(
V
R
T
)
is
d
e
s
ig
ne
d
to
e
nha
nc
e
T
H
Q
f
r
a
m
e
s
f
r
om
TLQ
f
r
a
m
e
s
,
a
ddr
e
s
s
in
g
va
r
io
us
vi
de
o
r
e
s
to
r
a
ti
on
ta
s
ks
s
uc
h
a
s
s
u
pe
r
-
r
e
s
ol
ut
io
n,
de
bl
ur
r
in
g,
a
nd
de
noi
s
in
g.
T
he
tr
a
ns
f
or
m
a
ti
on
pr
oc
e
s
s
in
vol
ve
s
two
pr
im
a
r
y
c
om
pon
e
nt
s
:
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
a
nd
r
e
c
ons
tr
uc
ti
on.
T
he
goa
l
of
t
he
V
R
T
i
s
t
o r
e
s
to
r
e
T
H
Q
f
r
a
m
e
s
f
r
om
TLQ
f
r
a
m
e
s
e
f
f
e
c
ti
ve
ly
.
∈
ℝ
x
x
x
r
e
pr
e
s
e
nt
s
hi
gh
-
qua
li
ty
t
a
r
ge
t
f
r
a
m
e
s
.
∈
ℝ
x
x
x
r
e
pr
e
s
e
nt
s
a
s
e
que
nc
e
of
lo
w
-
qua
li
ty
i
nput
f
r
a
m
e
s
.
4.3.2.
F
e
at
u
r
e
e
xt
r
ac
t
io
n
S
ha
ll
ow
f
e
a
tu
r
e
s
∈
ℝ
x
x
x
a
r
e
f
ir
s
t
e
xt
r
a
c
te
d
f
r
om
I
LQ
th
r
ough
a
s
in
gl
e
s
pa
ti
a
l
2D
c
onvolut
io
n.
S
ubs
e
que
nt
ly
,
a
m
ul
ti
-
s
c
a
le
ne
twor
k
is
ut
il
iz
e
d
to
s
ync
hr
oni
z
e
f
r
a
m
e
s
a
t
va
r
io
us
r
e
s
ol
ut
io
ns
b
y
in
te
gr
a
ti
ng
dow
ns
a
m
pl
in
g
a
nd
te
m
por
a
l
m
ut
ua
l
s
e
lf
-
a
tt
e
nt
io
n
(
T
M
S
A
)
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
a
t
di
f
f
e
r
e
nt
s
c
a
le
s
.
S
ki
p c
onne
c
ti
ons
a
r
e
i
nt
r
oduc
e
d f
or
f
e
a
tu
r
e
s
a
t
id
e
nt
ic
a
l
s
c
a
le
s
,
pr
oduc
in
g de
e
p f
e
a
tu
r
e
s
∈
ℝ
x
x
x
.
4.3.3.
R
e
c
on
s
t
r
u
c
t
io
n
T
he
HQ
f
r
a
m
e
s
a
r
e
r
e
c
ons
tr
uc
te
d
th
r
ough
th
e
c
om
bi
na
ti
on
of
s
ha
ll
ow
a
nd
de
e
p
f
e
a
tu
r
e
s
.
G
lo
ba
l
r
e
s
id
ua
l
le
a
r
ni
ng
s
tr
e
a
m
li
ne
s
th
e
pr
oc
e
s
s
of
f
e
a
tu
r
e
le
a
r
ni
ng
by
pr
e
di
c
ti
ng
s
ol
e
ly
th
e
di
f
f
e
r
e
nc
e
be
twe
e
n
th
e
bi
li
ne
a
r
ly
ups
a
m
pl
e
d
LQ
s
e
que
nc
e
a
nd
th
e
a
c
tu
a
l
HQ
s
e
que
n
c
e
.
T
h
e
r
e
c
ons
tr
uc
ti
on
m
odul
e
s
di
f
f
e
r
ba
s
e
d
on
th
e
s
pe
c
if
ic
r
e
s
to
r
a
ti
on
ta
s
ks
;
f
or
in
s
ta
nc
e
,
s
ub
-
pi
xe
l
c
onvolut
io
n
la
ye
r
s
a
r
e
e
m
pl
oye
d
f
or
vi
de
o
s
upe
r
-
r
e
s
ol
ut
io
n
, w
he
r
e
a
s
a
s
in
gl
e
c
onvolut
io
n l
a
ye
r
i
s
a
d
e
qua
te
f
or
vi
de
o de
bl
ur
r
in
g
.
4.3.4.
L
os
s
f
u
n
c
t
io
n
I
s
e
m
pl
oye
d t
o
tr
a
in
th
e
V
R
T
m
ode
l.
I
t
is
de
f
in
e
d a
s
f
ol
lo
w
s
:
=
√
(
−
)
2
+
2
I
R
H
Q
s
ta
nds
f
or
th
e
r
e
c
ons
tr
uc
te
d
HQ
s
e
que
n
c
e
,
w
hi
le
I
HQ
i
s
th
e
gr
ound
-
tr
ut
h
HQ
s
e
que
nc
e
,
w
it
h
be
in
g
a
s
m
a
ll
c
ons
ta
nt
t
ypi
c
a
ll
y
s
e
t
to
10
−3
,
to
pr
e
ve
nt
di
vi
s
io
n by z
e
r
o.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
2
,
A
pr
il
2025
:
1518
-
1530
1522
4.3.5.
T
e
m
p
or
al
m
u
t
u
al
s
e
lf
-
at
t
e
n
t
io
n
I
s
e
m
pl
oye
d
to
to
jo
in
tl
y
a
li
gn
c
ha
r
a
c
te
r
is
ti
c
s
a
c
r
os
s
two
f
r
a
m
e
s
.
G
iv
e
n
a
r
e
f
e
r
e
nc
e
f
r
a
m
e
f
e
a
tu
r
e
X
R
a
nd a
s
uppor
ti
ng f
r
a
m
e
f
e
a
tu
r
e
X
S
, t
he
que
r
y
Q
R
, ke
y
K
S
, a
nd va
lu
e
V
S
a
r
e
c
om
put
e
d i
n t
he
f
ol
lo
w
in
g m
a
nne
r
:
Q
R
=
X
R
·
P
Q
,
K
S
=
X
S
·
P
K
,
V
S
=
X
S
·
P
V
W
he
r
e
P
Q
,
P
K
, a
nd
P
V
r
e
pr
e
s
e
nt
pr
oj
e
c
ti
on ma
tr
ic
e
s
. T
he
c
om
p
ut
a
ti
on of
t
he
a
tt
e
nt
io
n m
a
p
A
is
a
s
f
ol
lo
w
s
:
=
(
√
)
a
n
d
us
e
d
f
o
r
w
e
i
g
h
t
e
d
s
u
m
of
V
S
(
,
,
)
=
(
√
)
4.3.6.
P
ar
al
le
l
w
ar
p
in
g
F
e
a
tu
r
e
w
a
r
pi
ng
is
im
pl
e
m
e
nt
e
d
a
t
th
e
c
onc
lu
s
io
n
of
e
ve
r
y
ne
twor
k
s
ta
ge
to
e
f
f
e
c
ti
ve
ly
a
ddr
e
s
s
s
ig
ni
f
ic
a
nt
m
ove
m
e
nt
s
.
T
he
opt
ic
a
l
f
lo
w
s
of
a
dj
a
c
e
nt
f
r
a
m
e
f
e
a
tu
r
e
s
X
t
-
1
a
nd
X
t
+
1
a
r
e
c
om
put
e
d
f
or
e
a
c
h
f
r
a
m
e
f
e
a
tu
r
e
X
t
,
a
nd
s
ubs
e
que
nt
ly
w
a
r
pe
d
to
w
a
r
ds
f
r
a
m
e
X
t
a
s
̂
t
-
1
a
nd
̂
t
+
1
us
in
g
ba
c
kw
a
r
d
a
nd
f
or
w
a
r
d
w
a
r
pi
ng t
e
c
hni
que
s
. T
h
e
or
ig
in
a
l
f
e
a
tu
r
e
i
s
c
om
bi
ne
d w
it
h t
he
di
s
to
r
te
d f
e
a
tu
r
e
s
a
nd t
he
n pr
oc
e
s
s
e
d t
hr
ough a
m
ul
ti
-
la
ye
r
pe
r
c
e
pt
r
on
(
M
L
P
)
to
m
e
r
ge
th
e
f
e
a
tu
r
e
s
a
nd
r
e
du
c
e
th
e
ir
di
m
e
ns
io
na
li
ty
.
M
or
e
s
pe
c
if
ic
a
ll
y,
a
m
ode
l
f
or
f
lo
w
e
s
ti
m
a
ti
on
pr
e
di
c
ts
th
e
r
e
s
id
ua
l
f
lo
w
,
a
nd
de
f
or
m
a
bl
e
c
onvolut
io
n
is
e
m
pl
oye
d
to
a
c
hi
e
ve
de
f
or
m
a
bl
e
a
li
gnm
e
nt
.
F
ig
ur
e
4
il
lu
s
tr
a
te
s
th
e
f
r
a
m
e
w
or
k
a
r
c
hi
te
c
tu
r
e
of
our
w
or
k
(
l
ib
x265+
V
R
T
)
.
T
hi
s
f
ig
ur
e
pr
ovi
de
s
a
c
om
pr
e
he
ns
iv
e
ove
r
vi
e
w
of
how
our
pr
opos
e
d
vi
de
o
r
e
s
to
r
a
ti
on
te
c
hni
que
in
te
gr
a
te
s
w
it
h
th
e
li
bx265
c
om
pr
e
s
s
io
n
c
ode
c
.
I
t
de
pi
c
ts
th
e
va
r
io
us
c
om
pone
nt
s
in
vol
ve
d
in
th
e
P
a
r
a
ll
e
l
W
a
r
pi
ng
pr
oc
e
s
s
a
nd
th
e
ir
in
te
r
a
c
ti
ons
,
he
lp
in
g
to
vi
s
ua
li
z
e
th
e
w
or
k
f
l
ow
a
nd
th
e
r
ol
e
o
f
e
a
c
h
e
le
m
e
nt
in
e
nha
nc
in
g
vi
de
o r
e
s
to
r
a
ti
on.
F
ig
ur
e
4. T
he
f
r
a
m
e
w
or
k a
r
c
hi
te
c
tu
r
e
of
our
w
or
k (
l
ib
x265+
V
R
T
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
D
e
e
p l
e
ar
ni
ng
-
bas
e
d t
e
c
hni
que
s
f
o
r
v
id
e
o e
nhan
c
e
m
e
nt
, c
om
p
r
e
s
s
io
n and r
e
s
to
r
at
io
n
(
R
e
douane
L
hi
adi
)
1523
5.
E
X
P
E
R
I
M
E
N
T
S
A
N
D
R
E
S
U
L
T
S
5.1.
C
om
p
r
e
s
s
io
n
t
as
k
V
id
e
o
c
om
pr
e
s
s
io
n
of
te
n
in
tr
oduc
e
s
a
r
ti
f
a
c
ts
th
a
t
de
gr
a
de
vi
s
ua
l
qua
li
ty
.
T
o
m
it
ig
a
te
th
e
s
e
is
s
u
e
s
,
w
e
e
m
pl
oye
d
a
dv
a
nc
e
d
de
e
p
le
a
r
ni
ng
m
ode
l
s
to
r
e
s
to
r
e
hi
gh
-
qua
li
ty
f
r
a
m
e
s
f
r
om
c
om
pr
e
s
s
e
d
in
put
s
.
I
ni
ti
a
ll
y, w
e
us
e
d a
c
onvolut
io
na
l
a
ut
oe
nc
ode
r
f
or
i
m
a
ge
c
om
pr
e
s
s
io
n, f
ol
lo
w
in
g t
he
m
e
th
od de
m
ons
tr
a
te
d by
J
o
e
t
al
.
[
2]
.
T
hi
s
m
ode
l
r
e
duc
e
s
f
il
e
s
iz
e
w
hi
le
pr
e
s
e
r
vi
ng
vi
s
ua
l
in
f
or
m
a
ti
on,
s
e
tt
in
g
th
e
f
ounda
ti
on
f
or
th
e
s
ubs
e
que
nt
r
e
s
to
r
a
ti
on t
a
s
ks
.
T
he
c
om
pr
e
s
s
io
n
ta
s
k
in
vol
ve
s
e
nc
odi
ng
vi
de
o
f
r
a
m
e
s
u
s
in
g
th
e
li
bx265
c
ode
c
to
r
e
duc
e
bi
tr
a
te
a
nd
s
to
r
a
ge
r
e
qui
r
e
m
e
nt
s
[
3]
.
I
ni
ti
a
ll
y,
in
put
f
r
a
m
e
s
a
r
e
pa
r
ti
ti
one
d
in
to
c
odi
ng
tr
e
e
uni
ts
(
C
T
U
s
)
a
nd
unde
r
go
in
tr
a
or
in
te
r
pr
e
di
c
ti
on
f
o
r
e
f
f
ic
ie
nt
da
ta
r
e
pr
e
s
e
nt
a
ti
on.
T
r
a
ns
f
or
m
a
nd
qua
nt
iz
a
ti
on
pr
oc
e
s
s
e
s
a
r
e
a
ppl
ie
d
to
s
pa
ti
a
ll
y
a
nd
te
m
por
a
ll
y
c
or
r
e
la
te
d
da
ta
.
E
nt
r
opy
c
odi
ng
te
c
hni
que
s
li
ke
c
ont
e
xt
a
da
pt
iv
e
bi
na
r
y
a
r
it
hm
e
ti
c
c
odi
ng
(
C
A
B
A
C
)
a
r
e
th
e
n
e
m
pl
oye
d
f
or
e
f
f
ic
ie
nt
bi
ts
tr
e
a
m
ge
ne
r
a
ti
on.
A
de
bl
oc
ki
ng
f
il
te
r
is
a
ppl
ie
d
t
o
r
e
duc
e
a
r
ti
f
a
c
ts
.
F
ig
ur
e
5
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
of
th
e
c
om
pr
e
s
s
io
n
ta
s
k,
s
how
in
g
th
e
or
ig
in
a
l
f
r
a
m
e
a
lo
ngs
id
e
th
e
c
om
pr
e
s
s
e
d
f
r
a
m
e
.
T
he
li
bx265
c
ode
c
a
c
hi
e
v
e
d
a
pe
a
k
s
ig
na
l
-
to
-
noi
s
e
r
a
ti
o
(
P
S
N
R
)
of
31.469
dB
,
s
tr
uc
tu
r
a
l
s
im
il
a
r
it
y
in
de
x
(
S
S
I
M
)
of
0.801,
a
nd
m
ul
ti
-
s
c
a
le
s
tr
uc
tu
r
a
l
s
im
il
a
r
it
y
in
de
x
(
MS
-
S
S
I
M
)
of
0.801.
T
hi
s
r
e
pr
e
s
e
nt
s
a
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
ove
r
pr
e
vi
ous
m
e
th
o
ds
, w
it
h a
P
S
N
R
i
nc
r
e
a
s
e
of
+
1.4 dB
.
F
ig
ur
e
5. C
om
pr
e
s
s
io
n t
a
s
k output
T
he
P
S
N
R
a
nd
S
S
I
M
m
e
tr
ic
s
pr
ovi
de
in
s
ig
ht
s
in
to
th
e
vi
s
ua
l
qua
li
ty
of
th
e
c
om
pr
e
s
s
e
d
f
r
a
m
e
c
om
pa
r
e
d
to
th
e
or
ig
in
a
l.
T
he
c
a
lc
ul
a
ti
ons
f
or
th
e
s
e
m
e
tr
ic
s
r
e
ve
a
l
th
a
t
th
e
c
om
pr
e
s
s
io
n
pr
oc
e
s
s
m
a
in
ta
in
s
a
hi
gh
le
ve
l
of
vi
s
ua
l
f
id
e
li
ty
de
s
pi
te
th
e
r
e
duc
ti
on
in
f
il
e
s
iz
e
.
T
a
bl
e
1
il
lu
s
tr
a
te
s
th
a
t
our
a
ppr
oa
c
h
de
m
ons
tr
a
te
s
s
ubs
ta
nt
ia
l
im
pr
ove
m
e
nt
s
a
c
r
os
s
ke
y
m
e
tr
ic
s
,
w
it
h
a
not
a
bl
e
in
c
r
e
a
s
e
in
P
S
N
R
(
+
1.4
dB
)
a
nd
e
nha
nc
e
m
e
nt
s
in
S
S
I
M
a
nd
M
S
-
S
S
I
M
by
+
0.12
on
a
ve
r
a
ge
.
A
lt
hough
our
bi
tr
a
te
r
e
duc
ti
on
is
s
li
ght
ly
le
s
s
th
a
n t
ha
t
of
pr
e
vi
ous
m
e
th
ods
, t
he
ove
r
a
ll
ga
in
s
i
n vi
s
u
a
l
qua
li
ty
a
r
e
s
ig
ni
f
ic
a
nt
.
T
a
bl
e
1. C
om
pa
r
is
on of
v
id
e
o
c
om
pr
e
s
s
io
n
m
e
th
ods
M
e
t
hod
P
S
N
R
S
S
I
M
MS
-
S
S
I
M
B
I
T
R
A
T
E
C
V
Q
E
27
0.72
0.71
2,300
S
I
C
28
0.74
0.73
2,100
T
I
U
28
0.75
0.76
2,100
B
V
C
29
0.78
0.77
2,000
S
I
R
30
0.79
0.78
2,200
L
i
bx265
31.469
0.801
0.801
1
,
903.95
T
hi
s
gr
a
ph
a
s
s
how
n
in
F
ig
ur
e
6
pr
ov
id
e
s
a
c
le
a
r
a
nd
c
om
pr
e
he
ns
iv
e
vi
s
ua
l
c
om
pa
r
is
on
of
th
e
pe
r
f
or
m
a
nc
e
of
va
r
io
us
vi
de
o c
om
pr
e
s
s
io
n m
e
th
ods
:
−
T
he
l
ib
x265 model a
c
hi
e
ve
s
t
he
be
s
t
r
e
s
ul
ts
i
n t
e
r
m
s
of
P
S
N
R
,
S
S
I
M
, a
nd M
S
-
S
S
I
M
, w
hi
le
m
a
in
ta
in
in
g
a
r
e
la
ti
ve
ly
l
ow
B
I
T
R
A
T
E
.
−
T
he
in
c
r
e
a
s
e
of
+
1.4
d
B
in
P
S
N
R
c
om
pa
r
e
d
to
th
e
pr
e
vi
ou
s
m
e
th
od
is
c
le
a
r
ly
vi
s
ib
le
,
a
s
a
r
e
th
e
im
pr
ove
m
e
nt
s
i
n S
S
I
M
a
nd M
S
-
S
S
I
M
.
−
T
hi
s
hi
ghl
ig
ht
s
t
he
e
f
f
e
c
ti
ve
ne
s
s
of
our
a
ppr
oa
c
h i
n e
nha
nc
in
g
vi
s
ua
l
qua
li
ty
, de
s
pi
te
a
s
li
ght
i
nc
r
e
a
s
e
i
n
B
I
T
R
A
T
E
c
om
pa
r
e
d t
o ot
he
r
m
e
th
ods
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
2
,
A
pr
il
2025
:
1518
-
1530
1524
F
ig
ur
e
6
. G
r
a
ph of
c
om
pa
r
a
ti
ve
a
na
ly
s
is
of
v
id
e
o
c
om
pr
e
s
s
io
n
m
e
th
ods
5.2.
R
e
s
t
or
at
io
n
t
as
k
s
5.2.1.
S
u
p
e
r
-
r
e
s
ol
u
t
io
n
t
as
k
F
or
th
e
s
upe
r
-
r
e
s
ol
ut
io
n
ta
s
k,
w
e
ut
il
iz
e
d
th
e
B
a
s
ic
V
S
R
m
ode
l,
de
s
ig
ne
d
to
e
nh
a
nc
e
s
pa
ti
a
l
r
e
s
ol
ut
io
n i
n vi
de
o
f
r
a
m
e
s
[
12]
, [
13]
.
T
he
pr
oc
e
s
s
i
nvol
ve
d
:
−
P
r
e
pr
oc
e
s
s
in
g:
f
r
a
m
e
s
w
e
r
e
dow
ns
a
m
pl
e
d a
nd r
e
s
iz
e
d t
o f
a
c
il
it
a
te
e
nha
nc
e
m
e
nt
.
−
M
ode
l
a
ppl
ic
a
ti
on:
th
e
B
a
s
ic
V
S
R
m
ode
l
w
a
s
a
ppl
ie
d t
o up
s
c
a
le
f
r
a
m
e
s
by a
f
a
c
to
r
of
4.
−
P
os
tp
r
oc
e
s
s
in
g:
e
nha
n
c
e
d
f
r
a
m
e
s
w
e
r
e
r
e
s
iz
e
d t
o t
he
ir
or
ig
in
a
l
di
m
e
ns
io
ns
.
O
ur
a
ppr
oa
c
h a
c
hi
e
ve
d s
ubs
ta
nt
ia
l
e
nha
n
c
e
m
e
nt
s
i
n P
S
N
R
a
nd S
S
I
M
m
e
tr
ic
s
w
he
n c
om
pa
r
e
d t
o c
ut
ti
ng
-
e
dge
m
e
th
ods
,
a
s
de
m
ons
tr
a
te
d
in
T
a
bl
e
2
a
nd
F
ig
ur
e
6
.
S
pe
c
if
ic
a
ll
y,
th
e
P
S
N
R
in
c
r
e
a
s
e
d
by
+
2.3
dB
,
in
di
c
a
ti
ng
a
s
ig
ni
f
ic
a
nt
e
nha
nc
e
m
e
nt
i
n vi
s
u
a
l
qua
li
ty
.
A
na
ly
s
is
a
nd
D
is
c
us
s
io
n:
T
he
r
e
s
ul
ts
f
r
om
T
a
bl
e
2
a
nd
F
ig
ur
e
7
in
di
c
a
te
th
a
t
th
e
B
a
s
i
c
V
S
R
m
ode
l
s
ubs
ta
nt
ia
ll
y
out
pe
r
f
or
m
s
ot
he
r
m
e
th
ods
in
te
r
m
s
of
P
S
N
R
a
nd
S
S
I
M
.
N
ot
a
bl
y,
our
pr
opos
e
d
m
e
th
od
us
in
g
l
ib
x265+
V
R
T
a
c
hi
e
ve
d
a
P
S
N
R
of
34.457
dB
,
w
hi
c
h
is
+
2.
067
dB
hi
ghe
r
th
a
n
th
e
s
e
c
ond
-
be
s
t
m
e
th
od,
B
a
s
ic
V
S
R
+
+
.
T
hi
s
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
de
m
ons
tr
a
te
s
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
our
a
ppr
oa
c
h
in
e
nha
n
c
in
g
vi
s
ua
l
qua
li
ty
.
T
he
u
s
e
of
de
e
p
l
e
a
r
ni
ng
m
ode
ls
,
pa
r
ti
c
ul
a
r
ly
tr
a
ns
f
or
m
e
r
s
li
ke
V
R
T
[
14]
,
[
15]
,
in
c
om
bi
na
ti
on w
it
h a
dva
nc
e
d
c
om
pr
e
s
s
io
n t
e
c
hni
que
s
, pr
ove
s
t
o
be
hi
ghl
y be
ne
f
ic
ia
l
f
or
s
upe
r
-
r
e
s
ol
ut
io
n t
a
s
ks
.
T
a
bl
e
2
. S
upe
r
r
e
s
ol
ut
io
n (
A
vg
m
e
tr
ic
s
)
M
e
t
hod
P
S
N
R
S
S
I
M
B
I
T
R
A
T
E
B
i
c
ubi
c
26.14
0.729
-
S
w
i
nI
R
29.05
0.826
-
S
w
i
nI
R
-
ft
29.24
0.831
-
T
O
F
l
ow
27.98
0.799
-
DUF
28.60
0.825
-
P
F
N
L
29.63
0.850
-
R
B
P
N
30.09
0.859
-
M
uC
A
N
30.88
0.875
-
E
D
V
R
31.09
0.880
-
V
S
R
T
31.19
0.881
-
B
a
s
i
c
V
S
R
31.42
0.890
-
I
c
onV
S
R
31.67
0.894
-
B
a
s
i
c
V
S
R
++
32.39
0.906
-
V
R
T
32.19
0.900
-
L
i
bx265+V
R
T
(
O
ur
s
)
34.457
0.902
7
,
499.671
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
D
e
e
p l
e
ar
ni
ng
-
bas
e
d t
e
c
hni
que
s
f
o
r
v
id
e
o e
nhan
c
e
m
e
nt
, c
om
p
r
e
s
s
io
n and r
e
s
to
r
at
io
n
(
R
e
douane
L
hi
adi
)
1525
F
ig
ur
e
7. S
upe
r
-
r
e
s
ol
ut
io
n pe
r
f
o
r
m
a
nc
e
5.2.2.
D
e
b
lu
r
r
in
g
t
as
k
T
o
a
ddr
e
s
s
m
ot
i
on
bl
ur
,
w
e
e
m
p
lo
y
e
d
t
he
r
e
c
ur
r
e
nt
vi
d
e
o
d
e
b
lu
r
r
in
g
m
o
de
l
[
1
6]
.
T
h
e
p
r
o
c
e
s
s
in
c
lu
de
d:
−
I
nput
pr
e
pa
r
a
ti
on:
f
r
a
m
e
s
f
r
om
th
e
s
upe
r
-
r
e
s
ol
ut
io
n
ta
s
k
w
e
r
e
r
e
s
iz
e
d
to
f
it
th
e
de
bl
ur
r
in
g
m
ode
l’
s
r
e
qui
r
e
m
e
nt
s
.
−
D
e
bl
ur
r
in
g
a
ppl
ic
a
ti
on:
t
he
m
ode
l
r
e
s
to
r
e
d s
ha
r
pne
s
s
i
n t
he
bl
ur
r
e
d f
r
a
m
e
s
.
−
P
a
r
a
m
e
te
r
c
onf
ig
ur
a
ti
on:
w
e
f
ol
lo
w
e
d
r
e
c
om
m
e
nde
d
s
e
tt
in
gs
to
e
ns
ur
e
c
on
s
is
te
nc
y
O
ur
m
e
th
od s
how
e
d
a
s
ubs
ta
nt
ia
l
in
c
r
e
a
s
e
in
P
S
N
R
(
+
3.4
dB
)
a
nd
a
m
ode
s
t
im
pr
ov
e
m
e
nt
in
S
S
I
M
,
de
m
ons
tr
a
ti
ng
e
f
f
e
c
ti
ve
r
e
s
to
r
a
ti
on of
s
ha
r
pne
s
s
, a
s
de
ta
il
e
d i
n
T
a
bl
e
3
a
nd F
ig
ur
e
8
.
A
na
ly
s
is
a
nd
di
s
c
u
s
s
io
n:
th
e
r
e
s
ul
ts
in
T
a
bl
e
3
a
nd
F
ig
ur
e
8
s
how
th
a
t
our
pr
opos
e
d
m
e
th
od
(
l
ib
x265+
V
R
T
)
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
s
P
S
N
R
,
a
c
hi
e
vi
ng
39.21
dB
,
w
hi
c
h
is
+
2.42
dB
hi
ghe
r
th
a
n
th
e
V
R
T
m
ode
l
a
lo
ne
.
T
he
S
S
I
M
a
ls
o
im
pr
ove
d,
in
di
c
a
ti
ng
be
tt
e
r
p
e
r
c
e
pt
ua
l
qua
li
ty
a
nd
s
ha
r
pne
s
s
r
e
s
to
r
a
ti
on.
T
hi
s
im
pr
ove
m
e
nt
c
a
n
be
a
tt
r
ib
ut
e
d
to
th
e
s
yne
r
gy
be
tw
e
e
n
th
e
r
e
c
ur
r
e
nt
a
r
c
hi
te
c
tu
r
e
a
nd
a
dva
nc
e
d
c
om
pr
e
s
s
io
n
[
17]
, w
hi
c
h e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
m
ot
io
n bl
ur
a
nd e
n
ha
nc
e
s
t
he
vi
de
o’
s
c
la
r
it
y.
T
a
bl
e
3
. D
e
bl
ur
r
in
g (
A
vg
m
e
tr
ic
s
)
M
e
t
hod
P
S
N
R
S
S
I
M
B
I
T
R
A
T
E
D
e
e
pD
e
bl
ur
26.16
0.824
-
S
R
N
26.98
0.814
-
D
B
N
26.55
0.806
-
E
D
V
R
34.80
0.948
-
V
R
T
36.79
0.964
-
L
i
bx265+V
R
T
39.21
0.986
78
,
960.82
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
2
,
A
pr
il
2025
:
1518
-
1530
1526
F
ig
ur
e
8. D
e
bl
ur
r
in
g
p
e
r
f
o
r
m
a
nc
e
5.2.3.
D
e
n
oi
s
in
g
t
as
k
W
e
ut
il
iz
e
d
th
e
S
w
in
I
R
m
ode
l
f
or
de
noi
s
in
g,
known
f
or
it
s
e
f
f
e
c
ti
ve
noi
s
e
r
e
duc
ti
on
[
18]
.
T
he
pr
oc
e
s
s
i
nc
lu
de
d
:
−
P
a
r
a
m
e
te
r
tu
ni
ng:
w
e
s
e
le
c
te
d a
s
ig
m
a
l
e
ve
l
of
10 ba
s
e
d on pr
e
vi
ous
r
e
s
e
a
r
c
h a
nd our
ow
n e
xp
e
r
im
e
nt
s
.
−
M
ode
l
a
ppl
ic
a
ti
on:
t
he
S
w
in
I
R
m
ode
l
w
a
s
a
ppl
ie
d t
o
de
noi
s
e
f
r
a
m
e
s
w
hi
le
pr
e
s
e
r
vi
ng i
m
por
ta
nt
de
ta
il
s
.
R
e
s
ul
ts
s
how
e
d
our
a
ppr
oa
c
h
a
c
hi
e
ve
d
s
im
il
a
r
ga
in
s
to
a
dva
nc
e
d
m
e
th
ods
,
w
it
h
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
s
in
P
S
N
R
a
nd P
S
N
R
Y
m
e
tr
ic
s
, a
s
s
how
n i
n T
a
bl
e
4
a
nd F
ig
ur
e
9
.
A
na
ly
s
is
a
nd
di
s
c
us
s
io
n:
T
a
bl
e
4
a
nd
F
ig
ur
e
9
il
lu
s
tr
a
te
th
e
d
e
noi
s
in
g
pe
r
f
or
m
a
nc
e
th
e
m
e
th
od
w
e
s
ugge
s
t
. T
h
e
r
e
s
ul
ts
s
ho
w
a
s
li
ght
de
c
r
e
a
s
e
i
n P
S
N
R
w
he
n
c
om
pa
r
e
d t
o t
he
V
R
T
m
ode
l
but
w
it
h a
hi
gh S
S
I
M
of
0.983.
T
he
P
S
N
R
Y
im
pr
ove
m
e
nt
to
41.77
dB
hi
ghl
ig
ht
s
our
m
e
th
od’
s
e
f
f
e
c
ti
ve
ne
s
s
in
m
a
in
ta
in
in
g
lu
m
in
a
nc
e
de
ta
il
,
c
r
uc
ia
l
f
or
hi
gh
-
qua
li
ty
vi
de
o
r
e
s
to
r
a
ti
on.
T
he
s
li
ght
tr
a
de
-
of
f
in
P
S
N
R
is
ba
la
nc
e
d
by
s
ig
ni
f
ic
a
nt
pe
r
c
e
pt
ua
l
qua
li
ty
ga
in
s
a
s
i
ndi
c
a
te
d by the
S
S
I
M
m
e
tr
ic
s
.
T
a
bl
e
4
. D
e
noi
s
in
g (
S
ig
m
a
=
10)
(
A
vg
m
e
tr
ic
s
)
M
e
t
hod
P
S
N
R
S
S
I
M
B
I
T
R
A
T
E
P
S
N
R
Y
S
S
I
M
Y
V
L
N
B
38.785
-
-
-
-
D
V
D
ne
t
38.13
-
-
-
-
F
a
s
t
D
V
D
ne
t
38.71
-
-
-
-
P
a
c
ne
t
39.97
-
-
-
-
V
R
T
40.82
-
-
-
-
(
x265+V
R
T
)
P
r
opos
e
d
40.00
0.983
91
,
772
41.77
0.987
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
D
e
e
p l
e
ar
ni
ng
-
bas
e
d t
e
c
hni
que
s
f
o
r
v
id
e
o e
nhan
c
e
m
e
nt
, c
om
p
r
e
s
s
io
n and r
e
s
to
r
at
io
n
(
R
e
douane
L
hi
adi
)
1527
F
ig
ur
e
9. D
e
noi
s
in
g pe
r
f
or
m
a
nc
e
5.2.4.
F
r
am
e
i
n
t
e
r
p
ol
at
io
n
F
r
a
m
e
in
te
r
pol
a
ti
on
(
T
a
bl
e
5)
w
a
s
in
c
or
por
a
te
d
to
im
pr
ove
te
m
por
a
l
c
ohe
r
e
nc
e
,
ut
il
iz
in
g
a
dva
nc
e
d
te
c
hni
que
s
[
19]
,
[
20]
.
A
lt
hough
th
e
in
te
r
pol
a
te
d
f
r
a
m
e
s
w
e
r
e
not
di
r
e
c
tl
y
us
e
d
due
to
in
te
gr
a
ti
on
c
ha
ll
e
nge
s
,
th
e
ir
m
e
tr
ic
s
w
e
r
e
e
va
lu
a
te
d
a
nd
in
c
lu
de
d
in
our
r
e
s
ul
ts
.
F
ut
ur
e
w
or
k
w
il
l
f
oc
us
on
r
e
f
in
in
g
th
e
s
e
te
c
hni
qu
e
s
to
e
nha
nc
e
t
he
r
e
s
to
r
a
ti
on pr
oc
e
s
s
.
A
na
ly
s
is
a
nd
di
s
c
us
s
io
n:
th
e
in
te
r
pol
a
ti
on
r
e
s
ul
ts
pr
e
s
e
nt
e
d
in
F
ig
ur
e
10
in
di
c
a
te
th
a
t
our
a
ppr
oa
c
h,
us
in
g
th
e
c
om
bi
na
ti
on
of
li
bx265
a
nd
V
R
T
,
s
ho
w
e
d
not
a
bl
e
i
m
pr
ove
m
e
nt
s
in
f
r
a
m
e
in
te
r
pol
a
ti
on.
A
s
s
how
n
in
F
ig
ur
e
10,
th
e
f
r
a
m
e
in
te
r
pol
a
ti
on
qua
li
ty
is
de
m
on
s
tr
a
te
d
b
y
a
P
S
N
R
of
27.32
dB
a
nd
a
S
S
I
M
of
0.867.
T
hi
s
f
ig
ur
e
hi
ghl
ig
ht
s
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
our
m
e
th
od
in
e
nha
nc
in
g
te
m
por
a
l
r
e
s
ol
ut
io
n
a
nd
ove
r
a
ll
vi
de
o
qua
li
ty
c
om
pa
r
e
d
to
s
ta
te
-
of
-
th
e
-
a
r
t
te
c
hni
que
s
.
S
pe
c
if
ic
a
ll
y,
m
e
th
ods
li
ke
th
os
e
pr
e
s
e
nt
e
d
in
[
21]
,
[
22]
ha
ve
de
m
ons
tr
a
te
d
s
ig
ni
f
ic
a
nt
a
dva
nc
e
s
in
vi
de
o
s
upe
r
-
r
e
s
ol
ut
io
n
a
nd
in
te
r
pol
a
ti
on,
w
hi
c
h
a
li
gn
w
it
h
th
e
im
pr
ove
m
e
nt
s
obs
e
r
ve
d
in
our
f
r
a
m
e
w
or
k.
O
ur
r
e
s
ul
ts
a
r
e
c
o
ns
is
te
nt
w
it
h
r
e
c
e
nt
s
tu
di
e
s
th
a
t
hi
ghl
ig
ht
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
de
e
p
le
a
r
ni
ng
m
ode
ls
in
vi
de
o
pr
oc
e
s
s
in
g
ta
s
ks
.
F
or
in
s
ta
nc
e
,
[
23]
s
how
c
a
s
e
a
dva
n
c
e
m
e
nt
s
in
vi
de
o
de
bl
ur
r
in
g
a
nd
f
r
a
m
e
in
te
r
pol
a
ti
on
th
a
t
a
r
e
c
om
pa
r
a
b
le
to
our
f
in
di
ngs
.
T
he
pe
r
f
or
m
a
nc
e
in
f
r
a
m
e
in
te
r
pol
a
ti
on
de
m
ons
tr
a
te
s
th
e
pot
e
nt
ia
l
of
our
f
r
a
m
e
w
or
k
to
de
li
ve
r
s
upe
r
io
r
r
e
s
ul
ts
in
vi
de
o
r
e
s
to
r
a
ti
on
ta
s
ks
,
e
c
hoi
ng
th
e
a
dva
n
c
e
m
e
nt
s
not
e
d
in
[
24]
‒
[
26]
.
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
unde
r
s
c
or
e
th
a
t
our
c
om
pr
e
he
ns
iv
e
vi
de
o
r
e
s
to
r
a
ti
on
f
r
a
m
e
w
or
k
a
c
hi
e
ve
s
not
a
bl
e
im
pr
ove
m
e
nt
s
a
c
r
os
s
v
a
r
io
us
qua
li
ty
m
e
tr
ic
s
,
in
c
lu
di
ng
P
S
N
R
a
nd
S
S
I
M
.
T
he
c
om
bi
na
ti
on
of
a
dv
a
nc
e
d
de
e
p
le
a
r
ni
ng
m
ode
ls
w
it
h
e
f
f
e
c
ti
ve
c
om
pr
e
s
s
io
n
te
c
hni
que
s
h
a
s
c
ont
r
ib
ut
e
d
s
ig
ni
f
ic
a
nt
ly
to
th
e
s
e
e
nha
nc
e
m
e
nt
s
.
S
im
il
a
r
im
pr
ove
m
e
nt
s
ha
ve
b
e
e
n
r
e
por
te
d
in
th
e
li
te
r
a
tu
r
e
,
s
uc
h
a
s
in
[
27
]
,
[
28
]
,
w
hi
c
h
f
oc
us
on
hi
gh
-
qua
li
ty
f
r
a
m
e
ge
ne
r
a
ti
on
a
nd
r
e
a
l
-
ti
m
e
f
lo
w
e
s
ti
m
a
ti
on.
F
ut
ur
e
e
f
f
or
ts
w
il
l
be
de
di
c
a
te
d
to
e
nha
nc
in
g
th
e
s
e
m
e
th
ods
a
nd
in
te
gr
a
ti
ng
th
e
m
m
or
e
s
uc
c
e
s
s
f
ul
ly
in
to
a
s
e
a
m
le
s
s
r
e
s
to
r
a
ti
on
pr
oc
e
s
s
f
or
r
e
a
l
-
li
f
e
s
c
e
na
r
io
s
,
w
it
h
th
e
goa
l
of
a
dva
nc
in
g
th
e
s
ta
nda
r
ds
of
vi
de
o r
e
s
to
r
a
ti
on i
n t
e
r
m
s
of
qua
li
ty
a
nd e
f
f
ic
ie
nc
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.