Game Development Reference
In-Depth Information
SAO, Adaptation Parameter Set (APS), is also proposed, which supports to save
ALF and SAO parameters in multiple frames and reuse for subsequent frames by
only transmitting an APS index.
7.4.1 Filter Shape and Coefficient Derivation
Adaptive loop filter is the last process stage for the reconstructed pictures in AVS-2
after the SAO stage. In AVS-2, region-based filtering (RA) method (Zhang et al.
2014a
) is adopted to support local adaptation. Multiple adaptive filters are derived
by minimizing the mean square errors between pixels in original image and recon-
structed image by SAO and merged according to the rate-distortion cost. The detail
techniques of ALF are introduced in the following.
The filter shape adopted in AVS-2 is a 7
3 square
shape, just as illustrated in Fig.
7.10
for both luminance and chroma components,
which are much more simplified compared with those three shapes, 5
×
7 cross shape superposing a 3
×
×
5, 7
×
7, and
9
9 square shapes, used in key technical (KTA) software (KTA 2.7). Each square in
Fig.
7.10
corresponds to a sample. Therefore, a total of 17 samples are used to derive
a filtered value for the sample of position C8. Considering overhead of transmitting
the coefficients, a point-symmetrical filter is utilized with only nine coefficients left,
{C0, C1,
×
, C8}, which reduces the number of filter coefficients to half as well
as the number of multiplications in filtering. The point-symmetrical filter can also
reduce half of the computation for one filtered sample, e.g., only 9 multiplications
and 14 add operations for one filtered sample.
For ALF, the input pixels are the output of SAO, and the desired pixels are the
original ones, and the filtered pixels are the output of the ALF. For convenience of
later discussion, we introduce some notations as follows:
...
1. Sample location
r
=
(
x
,
y
)
belongs to the to-be-filtered region
ʩ
and
ʩ
is the
number of samples in
ʩ
;
2. Desired pixel:
s
;
3. To-be-filtered pixel:
t
[
r
]
;
4. Wiener filter with
N
coefficients:
c
[
r
]
T
;
=[
c
0
,
c
1
,...,
c
N
−
1
]
5. Filter tap position offsets:
{
p
0
,
p
1
,...,
p
N
−
1
}
, where
p
n
denotes the sample
location offset of the
n
th filter tap;
6. Filtered pixel:
f
[
r
]
and is derived by
N
−
1
c
n
t
r
p
n
.
f
[
r
]
=
+
(7.24)
n
=
0
The ALF coefficients can be derived by minimizing the sum of square errors
(SSE) between the filtered pixels and the original pixels and setting derivatives of
SSEwith respect to
c
equal to zero, which leads to theWiener-Hopf equations shown
in Eq. (
7.25
).
