<s>
Video	B-Algorithm
super-resolution	I-Algorithm
(	O
VSR	O
)	O
is	O
the	O
process	O
of	O
generating	O
high-resolution	B-Algorithm
video	O
frames	O
from	O
the	O
given	O
low-resolution	O
video	O
frames	O
.	O
</s>
<s>
—	O
original	O
high-resolution	B-Algorithm
frame	O
sequence	O
,	O
</s>
<s>
—	O
convolution	B-Language
operation	O
,	O
</s>
<s>
Super-resolution	B-Algorithm
is	O
an	O
inverse	O
operation	O
,	O
so	O
its	O
problem	O
is	O
to	O
estimate	O
frame	O
sequence	O
from	O
frame	O
sequence	O
so	O
that	O
is	O
close	O
to	O
original	O
.	O
</s>
<s>
Video	B-Algorithm
super-resolution	I-Algorithm
approaches	O
tend	O
to	O
have	O
more	O
components	O
than	O
the	O
image	O
counterparts	O
as	O
they	O
need	O
to	O
exploit	O
the	O
additional	O
temporal	O
dimension	O
.	O
</s>
<s>
Single	B-Algorithm
image	I-Algorithm
super-resolution	I-Algorithm
methods	O
could	O
be	O
used	O
too	O
,	O
generating	O
high-resolution	B-Algorithm
frames	O
independently	O
from	O
their	O
neighbours	O
,	O
but	O
it	O
's	O
less	O
effective	O
and	O
introduces	O
temporal	O
instability	O
.	O
</s>
<s>
There	O
are	O
a	O
few	O
traditional	O
methods	O
,	O
which	O
consider	O
the	O
video	B-Algorithm
super-resolution	I-Algorithm
task	O
as	O
an	O
optimization	O
problem	O
.	O
</s>
<s>
Last	O
years	O
deep	B-Algorithm
learning	I-Algorithm
based	O
methods	O
for	O
video	O
upscaling	O
outperform	O
traditional	O
ones	O
.	O
</s>
<s>
The	O
high-resolution	B-Algorithm
frame	O
is	O
reconstructed	O
based	O
on	O
both	O
natural	O
preferences	O
and	O
estimated	O
motion	O
.	O
</s>
<s>
The	O
high-resolution	B-Algorithm
frame	O
is	O
estimated	O
in	O
this	O
domain	O
.	O
</s>
<s>
Some	O
methods	O
use	O
Fourier	B-Algorithm
transform	I-Algorithm
,	O
which	O
helps	O
to	O
extend	O
the	O
spectrum	O
of	O
captured	O
signal	O
and	O
though	O
increase	O
resolution	O
.	O
</s>
<s>
There	O
are	O
different	O
approaches	O
for	O
these	O
methods	O
:	O
using	O
weighted	B-Algorithm
least	I-Algorithm
squares	I-Algorithm
theory	I-Algorithm
,	O
total	B-Algorithm
least	I-Algorithm
squares	I-Algorithm
(	O
TLS	O
)	O
algorithm	O
,	O
space-varying	O
or	O
spatio-temporal	O
varying	O
filtering	O
.	O
</s>
<s>
Other	O
methods	O
use	O
wavelet	B-Algorithm
transform	I-Algorithm
,	O
which	O
helps	O
to	O
find	O
similarities	O
in	O
neighboring	O
local	O
areas	O
.	O
</s>
<s>
Later	O
second-generation	O
wavelet	B-Algorithm
transform	I-Algorithm
was	O
used	O
for	O
video	B-Algorithm
super	I-Algorithm
resolution	I-Algorithm
.	O
</s>
<s>
Iterative	O
back-projection	O
methods	O
assume	O
some	O
function	O
between	O
low-resolution	O
and	O
high-resolution	B-Algorithm
frames	O
and	O
try	O
to	O
improve	O
their	O
guessed	O
function	O
in	O
each	O
step	O
of	O
an	O
iterative	O
process	O
.	O
</s>
<s>
Iterative	O
adaptive	O
filtering	O
algorithms	O
use	O
Kalman	O
filter	O
to	O
estimate	O
transformation	O
from	O
low-resolution	O
frame	O
to	O
high-resolution	B-Algorithm
one	O
.	O
</s>
<s>
Some	O
approaches	O
also	O
consider	O
temporal	O
correlation	O
among	O
high-resolution	B-Algorithm
sequence	O
.	O
</s>
<s>
One	O
can	O
also	O
use	O
steepest	O
descent	O
,	O
least	B-Algorithm
squares	I-Algorithm
(	O
LS	O
)	O
,	O
recursive	O
least	B-Algorithm
squares	I-Algorithm
(	O
RLS	O
)	O
.	O
</s>
<s>
Direct	O
methods	O
estimate	O
motion	O
between	O
frames	O
,	O
upscale	O
a	O
reference	O
frame	O
,	O
and	O
warp	O
neighboring	O
frames	O
to	O
the	O
high-resolution	B-Algorithm
reference	O
one	O
.	O
</s>
<s>
To	O
construct	O
result	O
,	O
these	O
upscaled	O
frames	O
are	O
fused	O
together	O
by	O
median	B-Algorithm
filter	I-Algorithm
,	O
weighted	B-Algorithm
median	B-Algorithm
filter	I-Algorithm
,	O
adaptive	O
normalized	O
averaging	O
,	O
AdaBoost	O
classifier	O
or	O
SVD	O
based	O
filters	O
.	O
</s>
<s>
Non-parametric	O
algorithms	O
join	O
motion	O
estimation	O
and	O
frames	O
fusion	B-General_Concept
to	O
one	O
step	O
.	O
</s>
<s>
Weights	O
for	O
fusion	B-General_Concept
can	O
be	O
calculated	O
by	O
nonlocal-means	B-Algorithm
filters	I-Algorithm
.	O
</s>
<s>
Parameters	O
for	O
fusion	B-General_Concept
also	O
can	O
be	O
calculated	O
by	O
kernel	B-Language
regression	I-Language
.	O
</s>
<s>
Another	O
group	O
of	O
methods	O
use	O
maximum	B-General_Concept
a	I-General_Concept
posteriori	I-General_Concept
(	O
MAP	O
)	O
estimation	O
.	O
</s>
<s>
One	O
can	O
align	O
frames	O
by	O
performing	O
motion	O
estimation	O
and	O
motion	B-Algorithm
compensation	I-Algorithm
(	O
MEMC	O
)	O
or	O
by	O
using	O
Deformable	O
convolution	B-Language
(	O
DC	O
)	O
.	O
</s>
<s>
Motion	O
estimation	O
gives	O
information	O
about	O
the	O
motion	O
of	O
pixels	B-Algorithm
between	O
frames	O
.	O
</s>
<s>
motion	B-Algorithm
compensation	I-Algorithm
is	O
a	O
warping	O
operation	O
,	O
which	O
aligns	O
one	O
frame	O
to	O
another	O
based	O
on	O
motion	O
information	O
.	O
</s>
<s>
VSRnet	O
is	O
based	O
on	O
SRCNN	O
(	O
model	O
for	O
single	B-Algorithm
image	I-Algorithm
super	I-Algorithm
resolution	I-Algorithm
)	O
,	O
but	O
takes	O
multiple	O
frames	O
as	O
input	O
.	O
</s>
<s>
VESPCN	O
uses	O
a	O
spatial	O
motion	B-Algorithm
compensation	I-Algorithm
transformer	O
module	O
(	O
MCT	O
)	O
,	O
which	O
estimates	O
and	O
compensates	O
motion	O
.	O
</s>
<s>
DRVSR	O
(	O
detail-revealing	O
deep	O
video	B-Algorithm
super-resolution	I-Algorithm
)	O
consists	O
of	O
three	O
main	O
steps	O
:	O
motion	O
estimation	O
,	O
motion	B-Algorithm
compensation	I-Algorithm
and	O
fusion	B-General_Concept
.	O
</s>
<s>
The	O
motion	B-Algorithm
compensation	I-Algorithm
transformer	O
(	O
MCT	O
)	O
is	O
used	O
for	O
motion	O
estimation	O
.	O
</s>
<s>
The	O
sub-pixel	O
motion	B-Algorithm
compensation	I-Algorithm
layer	O
(	O
SPMC	O
)	O
compensates	O
motion	O
.	O
</s>
<s>
RVSR	O
(	O
robust	O
video	B-Algorithm
super-resolution	I-Algorithm
)	O
have	O
two	O
branches	O
:	O
one	O
for	O
spatial	O
alignment	O
and	O
another	O
for	O
temporal	O
adaptation	O
.	O
</s>
<s>
SOF-VSR	O
(	O
super-resolution	B-Algorithm
optical	O
flow	O
for	O
video	B-Algorithm
super-resolution	I-Algorithm
)	O
calculate	O
high-resolution	B-Algorithm
optical	O
flow	O
in	O
coarse-to-fine	O
manner	O
.	O
</s>
<s>
TecoGAN	O
(	O
the	O
temporally	O
coherent	O
GAN	B-Algorithm
)	O
consists	O
of	O
generator	O
and	O
discriminator	O
.	O
</s>
<s>
RBPN	O
(	O
the	O
recurrent	B-Algorithm
back-projection	O
network	O
)	O
.	O
</s>
<s>
Another	O
way	O
to	O
align	O
neighboring	O
frames	O
with	O
target	O
one	O
is	O
deformable	O
convolution	B-Language
.	O
</s>
<s>
While	O
usual	O
convolution	B-Language
has	O
fixed	O
kernel	O
,	O
deformable	O
convolution	B-Language
on	O
the	O
first	O
step	O
estimate	O
shifts	O
for	O
kernel	O
and	O
then	O
do	O
convolution	B-Language
.	O
</s>
<s>
Some	O
methods	O
align	O
frames	O
by	O
calculated	O
homography	B-Algorithm
between	O
frames	O
.	O
</s>
<s>
TGA	O
(	O
Temporal	O
Group	O
Attention	B-General_Concept
)	O
divide	O
input	O
frames	O
to	O
N	O
groups	O
dependent	O
on	O
time	O
difference	O
and	O
extract	O
information	O
from	O
each	O
group	O
independently	O
.	O
</s>
<s>
VSRResNet	O
like	O
GAN	B-Algorithm
consists	O
of	O
generator	O
and	O
discriminator	O
.	O
</s>
<s>
MRMNet	O
(	O
the	O
multi-resolution	O
mixture	O
network	O
)	O
consists	O
of	O
three	O
modules	O
:	O
bottleneck	O
,	O
exchange	O
,	O
and	O
residual	B-Algorithm
.	O
</s>
<s>
STMN	O
(	O
the	O
spatio-temporal	O
matching	O
network	O
)	O
use	O
discrete	O
wavelet	B-Algorithm
transform	I-Algorithm
to	O
fuse	B-General_Concept
temporal	O
features	O
.	O
</s>
<s>
Non-local	O
matching	O
block	O
integrates	O
super-resolution	B-Algorithm
and	O
denoising	O
.	O
</s>
<s>
While	O
2D	O
convolutions	B-Language
work	O
on	O
spatial	O
domain	O
,	O
3D	O
convolutions	B-Language
use	O
both	O
spatial	O
and	O
temporal	O
information	O
.	O
</s>
<s>
DUF	O
(	O
the	O
dynamic	O
upsampling	O
filters	O
)	O
uses	O
deformable	O
3D	O
convolution	B-Language
for	O
motion	B-Algorithm
compensation	I-Algorithm
.	O
</s>
<s>
3DSRnet	O
(	O
The	O
3D	O
super-resolution	B-Algorithm
network	O
)	O
uses	O
3D	O
convolutions	B-Language
to	O
extract	O
spatio-temporal	O
information	O
.	O
</s>
<s>
Recurrent	B-Algorithm
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
perform	O
video	B-Algorithm
super-resolution	I-Algorithm
by	O
storing	O
temporal	O
dependencies	O
.	O
</s>
<s>
STCN	O
(	O
the	O
spatio-temporal	O
convolutional	B-Architecture
network	I-Architecture
)	O
extract	O
features	O
in	O
the	O
spatial	O
module	O
,	O
pass	O
them	O
through	O
the	O
recurrent	B-Algorithm
temporal	O
module	O
and	O
final	O
reconstruction	O
module	O
.	O
</s>
<s>
BRCN	O
(	O
the	O
bidirectional	O
recurrent	B-Algorithm
convolutional	B-Architecture
network	I-Architecture
)	O
has	O
two	O
subnetworks	O
:	O
with	O
forward	O
fusion	B-General_Concept
and	O
backward	O
fusion	B-General_Concept
.	O
</s>
<s>
RISTN	O
(	O
the	O
residual	B-Algorithm
invertible	O
spatio-temporal	O
network	O
)	O
consists	O
of	O
spatial	O
,	O
temporal	O
and	O
reconstruction	O
module	O
.	O
</s>
<s>
Spatial	O
module	O
composed	O
of	O
residual	B-Algorithm
invertible	O
blocks	O
(	O
RIB	O
)	O
,	O
which	O
extract	O
spatial	O
features	O
effectively	O
.	O
</s>
<s>
RRCN	O
(	O
the	O
residual	B-Algorithm
recurrent	B-Algorithm
convolutional	B-Architecture
network	I-Architecture
)	O
is	O
a	O
bidirectional	O
recurrent	B-Algorithm
network	O
,	O
which	O
calculates	O
a	O
residual	B-Algorithm
image	O
.	O
</s>
<s>
BTRPN	O
(	O
the	O
bidirectional	O
temporal-recurrent	O
propagation	O
network	O
)	O
use	O
bidirectional	O
recurrent	B-Algorithm
scheme	O
.	O
</s>
<s>
The	O
key	O
idea	O
is	O
to	O
use	O
all	O
possible	O
positions	O
as	O
a	O
weighted	B-Algorithm
sum	I-Algorithm
.	O
</s>
<s>
PFNL	O
(	O
the	O
progressive	O
fusion	B-General_Concept
non-local	O
method	O
)	O
extract	O
spatio-temporal	O
features	O
by	O
non-local	O
residual	B-Algorithm
blocks	I-Algorithm
,	O
then	O
fuse	B-General_Concept
them	O
by	O
progressive	O
fusion	B-General_Concept
residual	B-Algorithm
block	I-Algorithm
(	O
PFRB	O
)	O
.	O
</s>
<s>
The	O
result	O
of	O
these	O
blocks	O
is	O
a	O
residual	B-Algorithm
image	O
.	O
</s>
<s>
MSHPFNL	O
also	O
incorporates	O
multi-scale	O
structure	O
and	O
hybrid	O
convolutions	B-Language
to	O
extract	O
wide-range	O
dependencies	O
.	O
</s>
<s>
The	O
common	O
way	O
to	O
estimate	O
the	O
performance	O
of	O
video	B-Algorithm
super-resolution	I-Algorithm
algorithms	O
is	O
to	O
use	O
a	O
few	O
metrics	O
:	O
</s>
<s>
FSIM	O
(	O
Feature	O
Similarity	O
Index	O
for	O
Image	O
Quality	O
)	O
uses	O
phase	B-Algorithm
congruency	I-Algorithm
as	O
the	O
primary	O
feature	O
to	O
measure	O
the	O
similarity	O
between	O
two	O
corresponding	O
frames	O
.	O
</s>
<s>
Currently	O
,	O
there	O
are	O
n't	O
so	O
many	O
objective	O
metrics	O
to	O
verify	O
video	B-Algorithm
super-resolution	I-Algorithm
method	O
's	O
ability	O
to	O
restore	O
real	O
details	O
.	O
</s>
<s>
Another	O
way	O
to	O
assess	O
the	O
performance	O
of	O
the	O
video	B-Algorithm
super-resolution	I-Algorithm
algorithm	O
is	O
to	O
organize	O
the	O
subjective	B-Device
evaluation	I-Device
.	O
</s>
<s>
While	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
of	O
video	B-Algorithm
super-resolution	I-Algorithm
outperform	O
traditional	O
ones	O
,	O
it	O
's	O
crucial	O
to	O
form	O
a	O
high-quality	O
dataset	B-General_Concept
for	O
evaluation	O
.	O
</s>
<s>
A	O
few	O
benchmarks	O
in	O
video	B-Algorithm
super-resolution	I-Algorithm
were	O
organized	O
by	O
companies	O
and	O
conferences	O
.	O
</s>
<s>
The	O
NTIRE	O
2019	O
Challenge	O
was	O
organized	O
by	O
CVPR	O
and	O
proposed	O
two	O
tracks	O
for	O
Video	B-Algorithm
Super-Resolution	I-Algorithm
:	O
clean	O
(	O
only	O
bicubic	B-Algorithm
degradation	O
)	O
and	O
blur	O
(	O
blur	O
added	O
firstly	O
)	O
.	O
</s>
<s>
Dataset	B-General_Concept
REDS	O
was	O
collected	O
for	O
this	O
challenge	O
.	O
</s>
<s>
To	O
evaluate	O
models	O
 '	O
performance	O
PSNR	O
and	O
SSIM	B-Algorithm
were	O
used	O
.	O
</s>
<s>
The	O
proposed	O
dataset	B-General_Concept
consists	O
of	O
1000	O
videos	O
,	O
each	O
length	O
is	O
4	O
–	O
6	O
seconds	O
.	O
</s>
<s>
PSNR	O
and	O
VMAF	B-Algorithm
metrics	O
were	O
used	O
for	O
performance	O
evaluation	O
.	O
</s>
<s>
The	O
challenge	O
was	O
held	O
by	O
ECCV	O
and	O
had	O
two	O
tracks	O
on	O
video	O
extreme	O
super-resolution	B-Algorithm
:	O
first	O
track	O
checks	O
the	O
fidelity	O
with	O
reference	O
frame	O
(	O
measured	O
by	O
PSNR	O
and	O
SSIM	B-Algorithm
)	O
.	O
</s>
<s>
Dataset	B-General_Concept
consists	O
of	O
328	O
video	O
sequences	O
of	O
120	O
frames	O
each	O
.	O
</s>
<s>
The	O
MSU	O
Video	B-Algorithm
Super-Resolution	I-Algorithm
Benchmark	O
was	O
organized	O
by	O
MSU	O
and	O
proposed	O
three	O
types	O
of	O
motion	O
,	O
two	O
ways	O
to	O
lower	O
resolution	O
,	O
and	O
eight	O
types	O
of	O
content	O
in	O
the	O
dataset	B-General_Concept
.	O
</s>
<s>
To	O
evaluate	O
models	O
 '	O
performance	O
PSNR	O
and	O
SSIM	B-Algorithm
were	O
used	O
with	O
shift	O
compensation	O
.	O
</s>
<s>
The	O
MSU	O
Super-Resolution	B-Algorithm
for	O
Video	O
Compression	O
Benchmark	O
was	O
organized	O
by	O
MSU	O
.	O
</s>
<s>
The	O
dataset	B-General_Concept
consists	O
of	O
9	O
videos	O
,	O
compressed	O
with	O
different	O
Video	B-Application
codec	I-Application
standards	O
and	O
different	O
bitrates	O
.	O
</s>
<s>
5	O
video	B-Application
codecs	I-Application
were	O
used	O
to	O
compress	O
ground-truth	O
videos	O
.	O
</s>
<s>
Top	O
combinations	O
of	O
Super-Resolution	B-Algorithm
methods	O
and	O
video	B-Application
codecs	I-Application
are	O
performed	O
in	O
the	O
table	O
:	O
</s>
<s>
The	O
resolution	O
of	O
video	O
can	O
be	O
degraded	O
because	O
of	O
imperfections	O
of	O
measuring	O
devices	O
,	O
such	O
as	O
optical	O
degradations	O
and	O
limited	O
size	B-Algorithm
of	I-Algorithm
camera	I-Algorithm
sensors	I-Algorithm
.	O
</s>
<s>
Super	B-Algorithm
Resolution	I-Algorithm
techniques	O
help	O
to	O
restore	O
the	O
original	O
video	O
.	O
</s>
<s>
It	O
also	O
helps	O
to	O
solve	O
task	O
of	O
object	B-General_Concept
detection	I-General_Concept
,	O
face	O
and	O
character	B-Application
recognition	I-Application
(	O
as	O
preprocessing	O
step	O
)	O
.	O
</s>
<s>
The	O
interest	O
to	O
super-resolution	B-Algorithm
is	O
growing	O
with	O
the	O
development	O
of	O
high	O
definition	O
computer	B-Device
displays	I-Device
and	O
TVs	O
.	O
</s>
<s>
Video	B-Algorithm
super-resolution	I-Algorithm
finds	O
its	O
practical	O
use	O
in	O
some	O
modern	O
smartphones	O
and	O
cameras	O
,	O
where	O
it	O
is	O
used	O
to	O
reconstruct	O
digital	O
photographs	O
.	O
</s>
<s>
A	O
process	O
called	O
demosaicing	B-Algorithm
is	O
used	O
to	O
reconstruct	O
the	O
photos	O
from	O
partial	O
color	O
information	O
.	O
</s>
<s>
You	O
can	O
observe	O
these	O
motions	O
on	O
Google	O
Pixel	B-Algorithm
3	O
phone	O
by	O
holding	O
it	O
perfectly	O
still	O
(	O
e.g.	O
</s>
