KSII TRANSACTIONS ON INTERNET AND INFORMATION SYSTEMS VOL. 8, NO. 3 Mar. 2014 1095
Copyright ⓒ 2014 KSII
During the transform and quantization process, a CU can also be recursively split into a
quadtree of transform unit (TU). The intra prediction is applied for each TU sequentially
instead of applying the intra prediction at the PU level [2]. In this way, the nearest neighboring
reference samples from the already reconstructed TU are used for prediction. The HEVC
defines 35 intra prediction modes, namley Intra_DC, Intra_Planar and 33 angular prediction
modes denoded as Intra_Angular[k], k = 2,...,34. In comparison with H.264/AVC, the intra
prediction modes defined in HEVC are able to capture more directional structures.
In the HEVC test mode (HM), a rate-distortion (RD) optimized bottom-up pruning
algorithm [3] is adopted to find the best CU partition of a CTU. Given a CU, the best PU
partition type, prediction mode and TU partition are found by minimizing the Lagrangian cost.
While the flexible block structure and more choices of prediction mode provide better coding
performance, it dramatically increases the encoder complexity. Even for intraframe coding, it
is still far away from real-time application [4]. The computational complexity mainly comes
from two folds. First, an optimal CU partition within a CTU can consist of various sizes of CU
ranging from 8x8 to 64x64. Second, finding an optimal prediction mode requires a RD
optimization process. Therefore, fast algorithms in both aspects are desirable. Research works
regarding fast CU size decision include [5-8]; works targeting fast intra mode decision include
[9-15]. In our previous work [16], a novel fast bottom-up pruning (FBUP) algorithm was
proposed and proved to be efficient. In this paper, three techniques are proposed for fast intra
mode decision and combined with the FBUP algorithm. Starting from the smallest CU size, a
full quadtree is processed in a bottom-up manner. First, we check whether mode decision
process is needed for a current CU node. If the mode decision process is performed, then a set
of intra candidate modes are derived from the coding information obtained from its sub-blocks
(sub-CUs or PART_NXN PUs), followed by a rough decision process to choose a smaller set
of candidates. The number of candidates in the set is adaptively chosen based on the low
complexity RD cost. In this way, we can either skip the mode decision process or reduce the
number of candidate modes. Thus, the computational complexity is reduced.
In Section 2, the previous works in the literature are reviewed. Then the proposed method is
described in Section 3. In Section 4, the experimental results are presented to prove the
efficiency of the proposed algorithm. Compared to the original HM encoder, averaging over
all the test sequences and rate points, 41.39% encoding time saving is achieved while the
bitrate increase is only 0.77%. Finally, the conclusion is given in Section 5.
2. Related Works
2.1 Rate-Distortion Optimized Bottom-up Pruning
In HEVC, the CTU size is 64x64 and the maximum number of depth levels of the quadtree
is
. The HM encoder adopts the rate-distortion optimized bottom-up pruning algorithm
to obtain the optimal CU quadtree partition. Consider the full quadtree as shown in Fig. 3, we
locate a node by its depth level from top to bottom and its position from left to right. Let
be the index of
th node at depth level
, then
. Denote
as a CU at node
without further splitting and
as the optimal tree after rate-distortion optimized pruning.
Let
be the operation to calculate the best RD cost. The bottom-up pruning algorithm