<s>
ZPEG	B-Device
is	O
a	O
motion	O
video	O
technology	O
that	O
applies	O
a	O
human	O
visual	O
acuity	O
model	O
to	O
a	O
decorrelated	O
transform-domain	O
space	O
,	O
thereby	O
optimally	O
reducing	O
the	O
redundancies	O
in	O
motion	O
video	O
by	O
removing	O
the	O
subjectively	O
imperceptible	O
.	O
</s>
<s>
The	O
ZPEG	B-Device
company	O
produces	O
modified	O
versions	O
of	O
x264	B-Language
,	O
x265	B-Language
,	O
AV1	B-Application
,	O
and	O
FFmpeg	B-Library
under	O
the	O
name	O
ZPEG	B-Device
Engine	O
(	O
see	O
)	O
.	O
</s>
<s>
Pixel	B-Algorithm
distributions	O
are	O
well-modeled	O
as	O
stochastic	O
process	O
,	O
and	O
a	O
transformation	O
to	O
their	O
ideal	O
decorrelated	O
representation	O
is	O
accomplished	O
by	O
the	O
Karhunen	O
–	O
Loève	O
transform	O
(	O
KLT	O
)	O
defined	O
by	O
the	O
Karhunen	O
–	O
Loève	O
theorem	O
.	O
</s>
<s>
A	O
time-varying	O
Contrast	O
Sensitivity	O
model	O
may	O
be	O
specified	O
,	O
and	O
is	O
applicable	O
to	O
the	O
three-dimensional	O
Discrete	B-General_Concept
cosine	I-General_Concept
transform	I-General_Concept
(	O
DCT	B-General_Concept
)	O
.	O
</s>
<s>
The	O
perceptual	O
strength	O
of	O
the	O
Human	O
Visual	O
Model	O
quantizer	B-Algorithm
generation	O
process	O
is	O
calibrated	O
in	O
(	O
vB	O
)	O
,	O
a	O
logarithmic	O
scale	O
roughly	O
corresponding	O
to	O
perceptibility	O
as	O
measured	O
in	O
screen	O
heights	O
.	O
</s>
<s>
The	O
ZPEG	B-Device
model	O
also	O
includes	O
a	O
temporal	O
component	O
,	O
and	O
thus	O
is	O
not	O
fully	O
described	O
by	O
viewing	O
distance	O
.	O
</s>
<s>
The	O
ZPEG	B-Device
pre-processor	O
optimizes	O
motion	O
video	O
sequences	O
for	O
compression	O
by	O
existing	O
motion	O
estimation-based	O
video	O
compressors	O
,	O
such	O
as	O
Advanced	B-Application
Video	I-Application
Coding	I-Application
(	O
AVC	O
)	O
(	O
H.264	B-Application
)	O
and	O
High	B-Algorithm
Efficiency	I-Algorithm
Video	I-Algorithm
Coding	I-Algorithm
(	O
HEVC	B-Algorithm
)	O
(	O
H.265	B-Algorithm
)	O
.	O
</s>
<s>
The	O
human	O
visual	O
acuity	O
model	O
is	O
converted	O
into	O
quantizers	B-Algorithm
for	O
direct	O
application	O
to	O
a	O
three-dimensional	O
transformed	O
block	O
of	O
the	O
motion	O
video	O
sequence	O
,	O
followed	O
by	O
an	O
inverse	O
quantization	B-Algorithm
(	O
signal	O
processing	O
)	O
step	O
by	O
the	O
same	O
quantizers	B-Algorithm
.	O
</s>
<s>
The	O
application	O
of	O
Human	O
Visual	O
System-generated	O
quantizers	B-Algorithm
to	O
a	O
block-based	O
Discrete	B-General_Concept
Cosine	I-General_Concept
Transform	I-General_Concept
results	O
in	O
increased	O
compressibility	O
of	O
a	O
motion	O
video	O
stream	O
by	O
removing	O
imperceptible	O
content	O
from	O
the	O
stream	O
.	O
</s>
<s>
The	O
quantizers	B-Algorithm
are	O
generated	O
to	O
be	O
imperceptible	O
at	O
a	O
specified	O
viewing	O
distance	O
,	O
specified	O
in	O
.	O
</s>
<s>
Average	O
compression	O
savings	O
for	O
6Mbs	O
HD	O
video	O
using	O
the	O
x.264	B-Language
codec	O
when	O
processed	O
at	O
−12vB	O
is	O
21.88	O
%	O
.	O
</s>
<s>
Average	O
compression	O
savings	O
for	O
16Mbs	O
Netflix	O
4K	O
test	O
suite	O
video	O
using	O
the	O
x.264	B-Language
codec	O
processed	O
at	O
−12	O
vB	O
is	O
29.81	O
%	O
.	O
</s>
<s>
While	O
the	O
effects	O
of	O
ZPEG	B-Device
pre-processing	O
are	O
imperceptible	O
to	O
the	O
average	O
viewer	O
at	O
the	O
specified	O
viewing	O
distance	O
,	O
edge	O
effects	O
introduced	O
by	O
block-based	O
transform	O
processing	O
still	O
affect	O
the	O
performance	O
advantage	O
of	O
the	O
video	O
optimization	O
process	O
.	O
</s>
<s>
While	O
existing	O
deblocking	B-General_Concept
filters	I-General_Concept
may	O
be	O
applied	O
to	O
improve	O
this	O
performance	O
,	O
optimal	O
results	O
are	O
obtained	O
through	O
use	O
of	O
a	O
multi-plane	O
deblocking	B-General_Concept
algorithm	O
.	O
</s>
<s>
Pixels	B-Algorithm
values	O
are	O
then	O
chosen	O
according	O
to	O
their	O
distance	O
to	O
the	O
block	O
edge	O
,	O
with	O
interior	O
pixel	B-Algorithm
values	O
being	O
preferred	O
to	O
boundary	O
pixel	B-Algorithm
values	O
.	O
</s>
<s>
While	O
some	O
transform-domain	O
video	B-Application
codec	I-Application
technologies	O
exist	O
,	O
ZPEG	B-Device
is	O
based	O
on	O
the	O
three-dimensional	O
Discrete	B-General_Concept
Cosine	I-General_Concept
Transform	I-General_Concept
(	O
DCT	B-General_Concept
)	O
,	O
where	O
the	O
three	O
dimensions	O
are	O
pixel	B-Algorithm
within	O
line	O
,	O
line	O
within	O
frame	O
,	O
and	O
temporal	O
sequence	O
of	O
frames	O
.	O
</s>
<s>
The	O
extraction	O
of	O
redundant	O
visual	O
data	O
is	O
performed	O
by	O
the	O
computationally-efficient	O
process	O
of	O
quantization	B-Algorithm
of	O
the	O
transform-domain	O
representation	O
of	O
the	O
video	O
,	O
rather	O
than	O
the	O
far	O
more	O
computationally	O
expensive	O
process	O
of	O
searching	O
for	O
object	O
matches	O
between	O
blocks	O
.	O
</s>
<s>
Quantizer	B-Algorithm
values	O
are	O
derived	O
by	O
applying	O
a	O
Human	O
Visual	O
Model	O
to	O
the	O
basis	O
set	O
of	O
DCT	B-General_Concept
coefficients	O
at	O
a	O
pre-determined	O
perceptual	O
processing	O
strength	O
.	O
</s>
<s>
Compression	O
is	O
then	O
performed	O
by	O
an	O
entropy	B-Algorithm
removal	O
process	O
.	O
</s>
<s>
Once	O
the	O
viewing	O
conditions	O
has	O
been	O
chosen	O
under	O
which	O
the	O
compressed	O
content	O
is	O
to	O
be	O
viewed	O
,	O
a	O
Human	O
Visual	O
Model	O
generates	O
quantizers	B-Algorithm
for	O
application	O
to	O
the	O
three-dimensional	O
Discrete	B-General_Concept
Cosine	I-General_Concept
Transform	I-General_Concept
(	O
DCT	B-General_Concept
)	O
.	O
</s>
<s>
These	O
quantizers	B-Algorithm
are	O
tuned	O
to	O
remove	O
all	O
imperceptible	O
content	O
from	O
the	O
motion	O
video	O
stream	O
,	O
greatly	O
reducing	O
the	O
entropy	B-Algorithm
of	O
the	O
representation	O
.	O
</s>
<s>
The	O
viewing	O
conditions	O
expressed	O
in	O
and	O
the	O
correlation	O
of	O
pixels	B-Algorithm
before	O
transformation	O
are	O
generated	O
for	O
reference	O
by	O
the	O
entropy	B-Algorithm
encoding	I-Algorithm
.	O
</s>
<s>
While	O
quantized	O
DCT	B-General_Concept
coefficients	O
have	O
traditionally	O
be	O
modeled	O
as	O
Laplace	O
distributions	O
,	O
more	O
recent	O
work	O
has	O
suggested	O
the	O
Cauchy	O
distribution	O
better	O
models	O
the	O
quantized	O
coefficient	O
distributions	O
.	O
</s>
<s>
The	O
ZPEG	B-Device
entropy	B-Algorithm
encoder	O
encodes	O
quantized	O
three-dimensional	O
DCT	B-General_Concept
values	O
according	O
to	O
a	O
distribution	O
that	O
is	O
completely	O
characterized	O
by	O
the	O
quantization	B-Algorithm
matrix	O
and	O
the	O
pixel	B-Algorithm
correlations	O
.	O
</s>
<s>
Each	O
DCT	B-General_Concept
band	O
is	O
separately	O
entropy	B-Algorithm
coded	I-Algorithm
to	O
all	O
other	O
bands	O
.	O
</s>
<s>
These	O
coefficients	O
are	O
transmitted	O
in	O
band-wise	O
order	O
,	O
starting	O
with	O
the	O
DC	O
component	O
,	O
followed	O
by	O
the	O
successive	O
bands	O
in	O
order	O
of	O
low	O
resolution	O
to	O
high	O
,	O
similar	O
to	O
Wavelet	B-Algorithm
packet	I-Algorithm
decomposition	I-Algorithm
.	O
</s>
<s>
In	O
an	O
analogy	O
to	O
the	O
single-stimulus	O
continuous	O
quality-scale	O
(	O
SSCQS	O
)	O
normalized	O
metric	O
Mean	O
Opinion	O
Score	O
(	O
MOS	O
)	O
,	O
the	O
overall	O
DSCQS	O
score	O
is	O
normalized	O
to	O
the	O
range	O
( −100	O
,	O
100	O
)	O
and	O
is	O
termed	O
the	O
Differential	O
Mean	O
Opinion	O
Score	O
(	O
DMOS	O
)	O
,	O
a	O
measure	O
of	O
subjective	B-Device
video	I-Device
quality	I-Device
.	O
</s>
<s>
ZPEG	B-Device
extends	O
the	O
list	O
of	O
available	O
techniques	O
by	O
providing	O
a	O
subjective	O
quality	O
metric	O
generated	O
by	O
comparing	O
the	O
Mean	B-Algorithm
Squared	I-Algorithm
Error	I-Algorithm
metric	O
of	O
the	O
difference	O
between	O
the	O
reference	O
and	O
impaired	O
videos	O
after	O
pre-processing	O
at	O
various	O
perceptual	O
strengths	O
(	O
in	O
)	O
.	O
</s>
<s>
Statistically	O
ideal	O
format	O
conversion	O
is	O
done	O
by	O
interpolation	O
of	O
video	O
content	O
in	O
Discrete	B-General_Concept
Cosine	I-General_Concept
Transform	I-General_Concept
space	O
.	O
</s>
<s>
The	O
conversion	O
process	O
,	O
particularly	O
in	O
the	O
case	O
of	O
up-sampling	O
,	O
must	O
consider	O
the	O
ringing	B-Algorithm
artifacts	I-Algorithm
that	O
occur	O
when	O
abrupt	O
continuities	O
take	O
place	O
in	O
a	O
sequence	O
of	O
pixels	B-Algorithm
being	O
re-sampled	O
.	O
</s>
<s>
The	O
resulting	O
algorithm	O
can	O
down-sample	O
or	O
up-sample	O
video	O
formats	O
by	O
changing	O
the	O
frame	O
dimensions	O
,	O
pixel	B-Device
aspect	I-Device
ratio	I-Device
,	O
and	O
frame	O
rate	O
.	O
</s>
