Game Development Reference
In-Depth Information
Fig. 9.1
The comparison of
Hadamard-based R-D cost
and the SSE-based R-D cost.
The
horizontal
axis denotes
the index of the
intraprediction mode
modes for any coding unit. Statistical results demonstrate that less intraprediction
mode is enough for the coding unit with less texture details. As depicted in Fig.
9.1
,
for the area with less texture details, the distribution of Hadamard-based R-D cost
is in accordance with the SSE-based R-D cost. The point with yellow color is the
optimal mode for the PU.
Based on gradient variance, the proposed intraoptimization scheme can be
described as follows. Firstly, for a given PU, the gradient of each pixel is calcu-
lated as follows:
I
i
,
j
)
=
I
i
,
j
−
I
i
,
j
+
1
+
I
i
,
j
−
I
i
+
1
,
j
Grad
(
(9.2)
where
I
i
,
j
denotes pixel value with the position as
(
i
,
j
)
. For the PU with size of
M
×
M
, the gradient variance is calculated as follows:
M
2
i
=
1
j
=
1
(
M
1
M
2
I
i
,
j
)
−
G
Va r
g
=
Grad
(
)
(9.3)
where
G
is the average gradient of each pixel. Considering the difference of PU size,
the above formula is rewritten as follows,
1
M
2
Va r
g
_
p
=
·
Va r
g
(9.4)
Thus the number of candidate mode can be represented as a function of Var
g
_
p
.
N
=
f
(
Va r
g
_
p
)
=
[
N
h
·
ʱ
+
0
.
5]
(9.5)
where
ʱ
denotes the shrink factor for the number of prediction mode, which is related
to the Var
g
_
p
. The function
[·]
denotes the rounding operation.
⊧
⊨
1 ar
g
_
p
>
Th
1
3
/
4T
1
≥
Va r
g
_
p
>
Th
2
ʱ
=
(9.6)
2
/
4T
2
≥
Va r
g
_
p
>
Th
3
⊩
1
/
4V
g
_
p
≤
Th
3
