Graphics Reference
In-Depth Information
Fig. 3.7
Example for the partitioning of a 64 64 luma coding tree block (
black
) into coding
blocks (
blue
) and transform blocks (
red
). In the illustration on the
right
,the
blue lines
show the
corresponding coding tree with the coding tree block (
black square
) at its root and the coding
blocks (
blue circles
) at its leaf nodes; the
red lines
show the non-degenerated residual quadtrees
with the transform blocks (
red circles
) as leaf nodes. Note that the transform blocks chosen
identical to the corresponding coding blocks are not explicitly marked in this figure. The numbers
indicate the coding order of the transform blocks
3.2.4.1
Residual Quadtree Structure
Each RQT is restricted by three parameters: The maximum depth d
max
of the tree,
the minimum allowed transform size n
min
and the maximum allowed transform size
n
max
, where the latter two are given as the binary logarithm of the transform block
width. Both n
min
and n
max
can vary within the range from two to five, which means
that transform block sizes from 4
4to32
32 samples are supported in HEVC. The
maximum allowed depth d
max
of the RQT restricts the number of subdivision levels.
For example, a value of d
max
D
1 means that a luma CB can either be coded as a
single TB or it can be split into 4 TBs, but no further splitting is allowed.
As a result of the interaction of these limits on transform block size and tree
depth, there are conditions that imply an internal (i.e., branching) node or a leaf
(i.e., non-branching) node in the RQT. As an example, consider a case, in which the
luma CB of the CU associated with the RQT root has a size of 64
64 samples, the
maximum depth is d
max
D
0 and the maximum transform size is equal to 32
32,
i.e., n
max
D
5. In HEVC, if the luma CB size is larger than the maximum transform
size, as in the mentioned example, the luma CB is forced to be subdivided to comply
with the limitations on the transform size.
If the decorrelating transform is applied across multiple prediction blocks,
the transformed residual signal often includes block edges at the corresponding
prediction boundaries, which increase the energy in the high-frequency transform
coefficients and, hence, decrease the coding efficiency. Due to this reason, HEVC
includes a special condition for implicit subdivision when the maximum RQT depth
is equal to zero. If d
max
D
0, the CU is predicted using motion compensation, and
the associated prediction partitioning consists of more than one PU, the luma CB
is always split into four TBs [
32
]. For d
max
>0and the case of inter-predicted
