Graphics Reference
In-Depth Information
a
b
First stage of inverse transform
Second stage of inverse transform
−
32768
0
0
0
−
16384
0
0
0
0
0
0
0
−
16384
0
0
0
0
0
0
0
−
16384
0
0
0
0
0
0
0
−
16384
0
0
0
[]
[]
T
×
D
×
D
−2097152
0
0
0
−
1048576
−
1048576
−
1048576
−
1048576
−2097152
0
0
0
−
1048576
−
1048576
−
1048576
−
1048576
−2097152
0
0
0
−
1048576
−
1048576
−
1048576
−
1048576
−2097152
0
0
0
−
1048576
−
1048576
−
1048576
−
1048576
>>
7
S
=
2
−
7
>>
12
IT
1
S
=
2
−
(
20
−
B
)
=
2
−
12
IT
2
−
16384
0
0
0
−
256
−
256
−
256
−
256
−
16384
0
0
0
−
256
−
256
−
256
−
256
−
16384
0
0
0
−
256
−
256
−
256
−
256
−
16384
0
0
0
−
256
−
256
−
256
−
256
Fig. 6.5
Use of the inverse transform scale factors assuming the input to be the final output of
Fig.
6.4
. Video bit depth
B
8(
a
) First stage of the inverse transform, (
b
) Second stage of the
inverse transform. Reproduced with permission from [
6
]. © IEEE 2013
D
the same time limit the dynamic range between the two inverse transform stages to
16 bits, the choice of scale factors for the inverse transform was finally modified as
follows
2
:
After the first inverse transform stage:
S
IT
1
D
2
7
After the second inverse transform stage:
S
IT
2
D
2
(20
B
)
The use of the inverse transform scale factors is illustrated in Fig.
6.5
using the
4
4 inverse transform as an example assuming the input to be the final output of
Fig.
6.4
.
Tab les
6.2
and
6.3
summarize the different scaling factors of the forward and
inverse transform, respectively, when compared to the orthonormal DCT.
The HEVC specification specifies an offset value to be added before scaling to
carry out rounding. This offset value is equal to the scale factor divided by 2. The
offset is not explicitly shown in Figs.
6.3
,
6.4
,and
6.5
.
2
Note that in the final HEVC specification [
16
], a clipping operation is introduced after the first
inverse transform stage, mainly to allow for random quantization that could be used to create “evil”
bitstreams used for stress testing video decoders. With the clipping introduced, the modification to
the inverse transform scale factors is not necessary but has been retained in the HEVC specification
and Test Model software for maturity reasons.
