Fast Intra Prediction Algorithm with Enhanced Sampling Decision for
H.265/HEVC
Sio-Kei Im
1
, Mohammad Mahdi Ghandi
2
and Ka-Hou Chan
1
1
MPI-QMUL Information Systems Research Centre, Macao Polytechnic Institute, Macao, China
2
University of Essex, Essex, U.K.
{marcusim, chankahou}@ipm.edu.mo, mahdighandi@gmail.com
Keywords:
High Efficient Video Coding, Fast Intra Predication, Weighted Sampling, Adaptive Thresholds.
Abstract:
H.265/HEVC is the latest video coding standard, which offers superior compression performance against
H.264/AVC at the cost of greater complexity in its encoding process. In the intra coding of HEVC, a Coding
Unit (CU) is recursively divided into a quad-tree-based structure from the Largest Coding Unit (LCU). At each
level, up to 35 potential intra modes should be checked. However, examining all these modes is very time-
consuming. In this paper, an intra mode decision algorithm is proposed that reduces the required computations
while having negligible effect on Rate-Distortion (RD) performance. A rough mode decision method based
on image component sampling is proposed to reduce the number of candidate modes for rough mode decision
and RD optimization. To balance the quality and performance, the decision to reduce the full search is made
with a thresholds that is dynamically updated based on the Quantization Parameter (QP) and CU size of each
recursive step. Experiments show that our algorithm can achieve a reasonable trade-off between encoding
quality and efficiency. The saving in encoding time is between 30.0% to 45.0% while BD-RATE may increase
by up to 0.5% for H.265/HEVC reference software HM 16.9 under all-intra configuration.
1 INTRODUCTION
The H.264/AVC video coding standard (Wiegand
et al., 2003) is widely used in currently deployed
video broadcasting systems. However, there is a
need to deliver higher resolution and better quality of
video. To respond to this need, ISO/IEC Moving Pic-
ture Experts Group (MPEG) and ITU-T Video Coding
Experts Group (VCEG) have jointly developed a new
video coding standard named High Efficiency Video
Coding (H.265/HEVC) (Sullivan et al., 2012).
H.265/HEVC realizes better image quality in
comparison with H.264/AVC at the same coding bit-
rate, but also needs much larger complexity in its
encoding process. For intra coding, it has been
known that the recursive quad-tree-based structure
of the Coding Unit (CU) can improve the process-
ing efficiency while the encoding complexity is in-
creased significantly (Kim et al., 2012). The com-
plexity of H.265/HEVC intra coding increases sev-
eral times compared with H.264/AVC intra coding
(Bossen et al., 2012) since the best intra prediction
mode and CU size need to be decided by the Rate-
Distortion Optimization (RDO).
1.1 Intra Prediction in H.265/HEVC
In H.265/HEVC, there are three types of coding
blocks: Coding Unit (CU), Prediction Unit (PU) and
Transform Unit (TU). Coding Tree Unit (CTU) is a
concept of Largest Coding Unit (LCU, size = 64 ×64
, depth = 0). Each CU can be encoded in the best
mode achieving the smallest Rate-Distortion (RD)
cost among all the possible modes or recursively split
into four CUs with equal sizes until the Smallest Cod-
ing Unit (SCU, size = 8 × 8. depth = 3). The parti-
tioning structure of CU and PU is shown in Fig. 1.
2N ·2N
2N ·N
N ·2N
N ·N
PU
LCU
CU CU
CU
CU
CU
CU
CU
CU
CU
CU
PU
PU
PU
PU
PU
PU
PU
PU
PU
PU
PU
PU
Figure 1: Partitioning structure of CU and splitting of PU,
here N is half of the CU size.
After splitting the CU, then the PU and TU can
60
Im S., Mahdi Ghandi M. and Chan K.
Fast Intra Prediction Algorithm with Enhanced Sampling Decision for H.265/HEVC.
DOI: 10.5220/0006131400600067
In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 60-67
ISBN: 978-989-758-225-7
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
be split independently, and selected for the CU if the
sum of the RD costs of four sub-CUs is smaller than
the RD cost of a large CU (Kim et al., 2012; Huang
et al., 2014). In intra mode, there are up to 35 possible
intra prediction modes in H.265/HEVC while only 9
modes are allowed in H.264/AVC. Note that the PU
can only be split into a square shape, and each PU
will choose the best matching prediction mode.
1.2 Related Work
In order to reduce the complexity of intra coding in
H.265/HEVC, there have been many algorithms pro-
posed in the literature focusing on fast intra mode de-
cision and fast CU size decision. The decision mak-
ing can be modified such as in (Piao et al., 2010) to
speed up the process. The algorithms used in (Yan
et al., 2012) and (Shen et al., 2013a) are based on the
correlation of adjacent prediction costs, which speeds
up the process but the extra computation load is still
high. In (Cho and Kim, 2013), an early CU split-
ting and pruning method is discussed and the method
in (Shen et al., 2013b) replaces the fixed tree-depth
range for each CU with an adaptive CU depth deci-
sion depending on the previously encoded slices and
neighbouring CUs. In (Zhao et al., 2011), a smaller
candidate set is proposed that does not lead to notice-
able quality loss, this compared various settings with
different mode candidates for the full Rate-Distortion
Optimized Quantization (RDOQ).
To reduce the complexity, (Choi and Jang, 2012)
proposed a tree pruning algorithm for fast CU deci-
sion making. It skips the further CU splitting if the
current CU chooses the SKIP mode. Such SKIP-
mode-based early termination cannot be applied to
the intra prediction. Moreover, some papers proposed
alternative thresholds condition: a local edge infor-
mation is obtained by calculating edge feature pa-
rameters in (da Silva et al., 2012) and (Jiang et al.,
2012) employed a gradient based method to speed
up CU decisions. According to (Min and Cheung,
2015; Jamali et al., 2015; Lim et al., 2015), the results
of HEVC intra coding based on different versions of
HMs are similar due to the fact that HEVC intra cod-
ing does not involve significant changes among these
versions. Therefore, it is acceptable to compare these
methods under different HM versions. Then, a gra-
dient based fast intra mode decision is proposed in
(Chen et al., 2013), which based on the gradient direc-
tions for each CU, selects only a small set of modes
for further intra prediction. Hence, a large portion of
modes are removed from the computation.
1.3 Contribution
In this paper, we propose a fast intra mode prediction
algorithm while maintaining the RD performance by
analysing picture feature statistics. An early termi-
nation technique for CU is applied in H.265/HEVC
for experiments. Corresponding to the required en-
coding efficiency and the quality of encoded picture,
we use the subsampling result of deviation values of
Luma/Chroma in different sizes for video sequences.
Since there are several Chroma-Subsampling in the
standard, we propose the weight of Luma/Chroma
will depend on the ratio of common Sampling sys-
tems (explained in Sec. 2.2). Further, to balance
the quality and performance, we define a thresh-
olds (T ) without using static decision thresholds, and
rely on a comparison between the Quantization Pa-
rameter (QP) and CU size of the coding tree un-
der test, which have to be updated at each recur-
sion. The use of adaptive thresholds is fundamental
in H.265/HEVC as the number of possible partition-
ing and coding modes is more appropriate then static
thresholds for each of them.
The rest of the paper is organized as follow: Sec. 2
provides details of the idea of the fast intra prediction
algorithm including the proposed weighed sampling
operator and adaptive thresholds. Sec. 3 gives the de-
tails of the concept about the thresholds and how to
achieve this in different depth layers. The simulation
results are given in Sec. 4, and Sec. 5 concludes the
paper.
2 PROPOSED METHOD
The intra mode prediction defined in the
H.265/HEVC standard allows 35 (including Pla-
nar, DC mode and 33 angular) predictions. Planar
and DC prediction modes can provide predictions
for image areas with smooth and gradually changing
content. The angular intra predictions are designed
to model different directional structures typically
present in picture (see Fig. 2).
2.1 Fast Intra Prediction Algorithm
As a range of sample methods, the angular predictions
are also useful in smooth PU blocks without high fre-
quency components for complex textures that cannot
be properly modelled with any of the directional pre-
dictors. It can easily perceive that a large size CU is
more suitable for some homogeneous or flat regions
in the image. For discriminating texture, CUs that are
not split into sub-CUs are expected to have smaller
Fast Intra Prediction Algorithm with Enhanced Sampling Decision for H.265/HEVC
61
30252015105
0 5 10 15 20 25 30
30
25
20
15
10
5
0
5
10
15
20
25
30
V
H
Figure 2: Angle definitions of angular intra prediction in
H.265/HEVC. (H,V ) is used to indicate the displacement.
Prediction mode 2 to 17 is index by (H, 32), and mode 18
to 34 is index by (32,V ).
RD costs than would occur if the CUs were split,
while an area with colourful edges or object bound-
aries is usually split into small CUs. Since the feature
of a homogeneous region is that the sampling result
approaches the mean value, the Sum of Absolute Dif-
ference (SAD) can represent the image complexity.
In extreme cases, if (SAD = 0), it means there is full
homogeneity: i.e. there is exactly same sample value
in the whole region. Applying to intra prediction, we
design to calculate the image complexity before the
33 angular intra predictions as:
mean =
1
N × N
N,N
i=1, j=1
C (i, j) (1)
SAD =
N,N
i=1, j=1
|
mean C (i, j)
|
(2)
where N
{
64, 32, 16
}
is the size of the current CU,
C (i, j) is the (sub)-sampling method that is exploited
by a weighted sampling operator and SAD represents
the image complexity and can be used to determine
whether it is necessary to do the angular prediction.
After the texture calculation, if the SAD is smaller
than or equal to the corresponding thresholds, the
image is thought to be smooth enough to terminate
the prediction early. Therefore, this algorithm avoids
checking SCU sizes when they are not likely to be
selected by the brute force RDO process.
2.2 Weighted Sampling Operator
For each CU at a given depth, all the allowed cod-
ing modes are tested and the mode with the minimum
RD cost is selected, then the early termination stops
the CU splitting process if the minimum SAD is al-
ready lower than a given threshold. This mean that a
good coding configuration has already been obtained
and searching for smaller CUs may only slightly im-
prove the overall performance or, indeed, may even
provide a lower performance as smaller CU sizes may
imply less compression in transformation. In order
to enhance the sampling to the reference results, we
must consider the weighted sampling/subsampling of
Luma and Chroma components in prediction block.
According to the efficiency of different configurations
scenario at quantization settings, we used a weighted
sum of the average sampling values for the Luma and
Chroma components, defined by
C =
(w
y
, w
uv
) · (Luma, Chroma)
w
y
+ w
uv
(3)
where the weights value (w
y
, w
uv
) must satisfy the
conditions w
y
> 0, w
uv
> 0 and w
y
+w
uv
6= 0, depend-
ing on the colour sampling modes,
(w
y
, w
uv
) =
(8, 0), type = 4 : 0 : 0
(8, 2), type = 4 : 1 : 1
(8, 2), type = 4 : 2 : 0
(8, 4), type = 4 : 2 : 2
(8, 4), type = 4 : 4 : 0
(8, 8), type = 4 : 4 : 4
(4)
where type means the type of Luma sampling and
Chroma subsampling. To approximate the accurate
quality using the PSNR and the average bit-rate of
generated bitstreams, the weighted sampling better
identifies the homogeneous area which characterizes
the coding efficiency.
2.3 Adaptive Thresholds
Since our algorithm avoids checking the RD costs
when early termination occurs, the values for the
thresholds (T ) play a central role in the trade-off be-
tween encoding quality and efficiency. With different
QP and CU sizes, the use of adaptive thresholds is
fundamental in H.265/HEVC as the number of possi-
ble partitioning and coding modes is so high that it is
impossible to set up appropriate static thresholds for
each of them. In order to find suitable thresholds, we
first analyse the SAD of different sizes of CU in ref-
erence software (see Fig. 3). When the CU texture is
complex, then CU is split into smaller subunits to find
the best size but if the CU texture is flat enough it will
not be divided further into subunits.
As shown in Fig. 3, coding with higher QP will
lose the detail in video reconstruction, which may
VISAPP 2017 - International Conference on Computer Vision Theory and Applications
62
(a) QP = 20
(b) QP = 40
Figure 3: Indicating the CU splitting by intra mode decision
with difference QP in HM reference software.
cause lots of homogeneous area that can be split into
bigger CU then coding with lower QP case. By this
reason, we consider the SAD as a thresholds (T ) to
determine the CU splitting and find the relation about
the thresholds as T (QP,CU
size
). According to statis-
tics from various videos, the thresholds is adaptive to
video content, CU size and QP, given by,
T (QP, CU
size
) =
E (C)
QP
2
· CU
2
size
(5)
where E (C) is the expected value of sampling, and
the partial differential equation of Eq. 5 about QP and
CU
size
as follows:
T
QP
=
CU
2
size
2
(6)
T
CU
size
= 2 ·
E (C)
QP
2
· CU
size
(7)
Therefore, the relationship between the SAD and CU
size is not very sensitive to QP. As the depth in-
creases, the CU will be evenly split into four sub-
CUs, while the thresholds are scaled down into ap-
proximately one quarter. Using large thresholds will
achieve a highly accurate decision but the complexity
reduction for intra mode decisions will be very small,
and using small thresholds will reduce the complexity
but lose accuracy. Thresholds should be determined
to provide a good trade-off between encoding qual-
ity and efficiency. In order to make the thresholds(T )
more adaptive to recent changes, a weighting function
is adopted which, similar to Eq. 4, corresponds to the
(sub)-sampling method. Thus, it smoothly allows up-
dating the expected values to better take into account
the most recent Luma and Chroma components, de-
fined by
E (C) =
(w
y
, w
uv
) · (E (Luma) , E (Chroma))
w
y
+ w
uv
(8)
3 IMPLEMENTATION
To verify our method, experiments were conducted
by using various test sequences under different QP
settings for investigating the RD cost distribution
for CUs resulting from the exhaustive search of the
H.265/HEVC Test Model (HM 16.9) and using our
modified HM software with the fast intra prediction
added.
3.1 Fast Intra Predication Walkthrough
The algorithm presented in this paper was imple-
mented in the HM 16.9 reference software. Since
the number of intra modes available for the predic-
tion of each CU can be reduced, the number of can-
didate modes selected in the HM 16.9 intra prediction
method is also modified. The overall proposed fast in-
tra predication process for each CU is now presented
as pseudo code, integrating the weighting sampling
and adaptive thresholds.
We first start from the Planar and DC mode since
these modes occur with high probability. Then we
calculate the SAD of CU that is used to compare with
the thresholds (T ) by considering the current QP and
CU
size
. If the SAD is smaller than the thresholds,
it means that the texture is flat enough so that there
needs to be no further CU splitting, just finding the
smallest RD cost between intra Planar, DC and 33 an-
gular modes in the current depth layer. Otherwise,
it means that the texture is too complex to be intra
predicated in the current Depth, so we can skip the
remaining angular predication and there is no need to
check the intra mode in the current depth layer any-
more. We then consider the next Depth to find the
smallest RD cost between intra Planar, DC and sum
of sub-CUs RD cost.
Fast Intra Prediction Algorithm with Enhanced Sampling Decision for H.265/HEVC
63
Algorithm 1: FastIntraPred(CU, QP, Depth).
Data: CU: object
Data: QP: int
Data: Depth: int
Result: RD
cost
: unsigned int
1 /* determine the best intra modes
and provide smallest RD cost */
2 begin
3 /* first find better intra mode
between Planar an DC */
4 RD
cost
smaller RD cost between mode 0, 1;
5 SAD calculate the SAD of CU;
6 /* compare with thresholds
condition */
7 if SAD < T (QP, CU
size
) then
8 /* if condition is verified
that means there are lots
of homogeneous area */
9 for all the CU to be coded at Depth do
10 /* find best intra angular
modes */
11 RD
cost
find the smallest RD cost;
12 end
13 else
14 /* CU is evenly split into
four units as CU
0,1,2,3
*/
15 RD
cost
sum of FastIntraPred(CU
0,1,2,3
,
QP, Depth + 1);
16 end
17 return RD
cost
;
18 end
4 EXPERIMENT RESULTS AND
DISCUSSION
The H.265/HEVC reference software HM 16.9 was
modified to handle our proposed method, and a se-
ries of simulations were carried out to measure the
improvement offered by accurate bit estimation. For
most of the comparisons, the Bjontegaard PSNR and
bit-rate differences (BD-PSNR and BD-RATE) be-
tween the proposed and the reference method are
shown in the following.
Table 1: Mean of SAD of difference CU size and QP in HM
reference software.
QP
64 × 64, 32 × 32, 16 × 16, 8 × 8,
depth = 0 depth = 1 depth = 2 depth = 3
20 481236.87 120007.25 30044.54 7479.20
24 472784.87 118100.85 29540.11 7030.81
28 464040.13 115765.73 29057.78 5899.01
32 455679.14 113168.00 28498.35 6827.27
36 448174.94 111326.90 28006.23 6987.15
40 437079.86 108763.22 27449.52 5667.41
As expected in Tab. 1 corresponding to Fig. 3,
QP values are taken into consideration as the QP will
scale the range of sample values. Note that these re-
sults are matched to Eq. 6 and 7, so the difference
between each sample and the mean of sampling CU
becomes smaller as QP increases.
4.1 Performance Verification
The procedure detailed in (Senzaki et al., 2010) is
used to calculate BD-PSNR and BD-RATE with QP
{
20, 24, 28, 32, 36, 40
}
. For all tests the the main-
profile features of HM in progressive format are en-
abled, also 100 frames are coded (frame rates given
in the table captions) with one intra update every sec-
ond.
Table 2: H.265/HEVC, proposed vs. reference HM 16.9,
QCIF@30Hz, all Intra pictures with BD-RATE, BD-PSNR
and time reduction.
QCIF IIIII
BD-RATE BD-PSNR Time
(%) -Y (dB) Reduction (%)
akio +0.441 -0.015 -44.99
carp +0.471 -0.014 -45.12
coast +0.219 -0.007 -43.98
fore +0.042 -0.091 -43.86
mobile +0.123 -0.059 -43.21
news +0.404 -0.099 -42.71
paris +0.253 -0.077 -41.88
silent +0.108 -0.044 -42.90
Table 3: H.265/HEVC, proposed vs. reference HM 16.9,
4CIF@30Hz, all Intra pictures with BD-RATE, BD-PSNR
and time reduction.
4CIF IIIII
BD-RATE BD-PSNR Time
(%) -Y (dB) Reduction (%)
aspen +0.118 -0.005 -30.44
blue sky +0.447 -0.028 -49.90
controlled burn +0.176 -0.010 -42.42
crowd run +0.352 -0.021 -44.59
in to tree +0.103 -0.004 -42.35
life +0.176 -0.010 -47.66
park joy +0.181 -0.012 -44.56
riverbed +0.095 -0.004 -31.22
rush hour +0.086 -0.003 -46.70
sunflower +0.264 -0.015 -31.63
Tab. 2 and 3 results of the proposed method for
all intra coding, 4CIF and QCIF videos respectively.
As expected, since the proposed fast intra-prediction
mode decision and fast intra-transform skip-mode
decision have been added into the H.265/HEVC
scheme, and the weighted sampling and adaptive
thresholds condition can find an approximate result
with HM reference software, so the BD-Rate just in-
creases by up to 0.5%.
VISAPP 2017 - International Conference on Computer Vision Theory and Applications
64
Table 4: H.265/HEVC, proposed vs. reference HM 16.9, QCIF@30Hz, all Intra pictures with difference QP and coding time
reduction.
QCIF IIII QP = 20 (%) QP = 24 (%) QP = 28 (%) QP = 32 (%) QP = 36 (%) QP = 40 (%) AVERAGE
akio -46.40 -45.83 -45.60 -44.06 -44.12 -43.92 -44.99
carp -45.88 -46.24 -44.67 -45.83 -44.77 -43.30 -45.12
coast -45.17 -46.08 -45.89 -41.61 -41.38 -43.74 -43.98
fore -46.62 -47.30 -46.14 -40.24 -41.99 -40.89 -43.86
mobile -44.88 -45.85 -44.45 -42.55 -40.11 -41.39 -43.21
news -44.22 -45.38 -45.07 -41.20 -41.14 -39.27 -42.71
paris -41.68 -44.37 -42.49 -42.92 -40.09 -39.72 -41.88
silent -44.57 -45.37 -41.21 -44.47 -43.30 -38.45 -42.90
Table 5: H.265/HEVC, proposed vs. reference HM 16.9, 4CIF@30Hz, all Intra pictures with difference QP and coding time
reduction.
4CIF IIII QP = 20 (%) QP = 24 (%) QP = 28 (%) QP = 32 (%) QP = 36 (%) QP = 40 (%) AVERAGE
aspen -31.91 -33.30 -33.72 -29.19 -27.80 -26.72 -30.44
blue sky -49.37 -48.45 -48.02 -51.52 -51.07 -50.97 -49.90
controlled burn -43.70 -36.18 -39.90 -43.19 -44.76 -46.76 -42.42
crowd run -42.58 -43.17 -48.39 -45.05 -44.92 -43.40 -44.59
in to tree -48.70 -37.62 -38.43 -42.19 -43.97 -43.17 -42.35
life -50.68 -43.32 -45.20 -48.94 -48.62 -49.18 -47.66
park joy -51.35 -47.75 -36.46 -42.11 -45.59 -44.08 -44.56
riverbed -38.73 -31.00 -29.38 -28.74 -28.80 -30.64 -31.22
rush hour -51.26 -49.97 -47.90 -47.87 -45.44 -37.78 -46.70
sunflower -36.34 -30.84 -30.49 -30.21 -30.77 -31.11 -31.63
As shown in Tab. 4 and 5, the encoding bit stream
size for different values of QP with QCIF and 4CIF
samples, and encoding time reduction is derived by
t =
t
proposed
t
reference
t
reference
× 100.00% (9)
where t
reference
and t
proposed
are the encoding time of
the original intra mode decision and the proposed fast
intra mode decision scheme. Results show that pro-
posed method can generate a speed-up on average of
40.0% and up to 45.0% with the BD-PSNR degrada-
tion less than 0.1dB in QCIF, with 4CIF the average
speed-up is 30.0% up to 50.0% with BD-PSNR degra-
dation of less than 0.03dB. As discussed in Sec. 2.3,
these simulation results are not very sensitive to the
QP value, but the percentage of time reduction should
decrease with increasing QP value. Because the tex-
tures could be more complex when using small QP
(which often requires more CU splitting) our pro-
posed approach can be effective to skip lots of an-
gular predication when the coding block is complex
enough.
As shown in Fig. 4 to 7, the performance of the
proposed method gives a more significant gain in the
higher resolution sequence with all inter mode. Our
method performs better in terms of BD-RATE and
thresholds, and our results approximate very well to
the HM reference. It can be seen that the proposed
method achieves a reduction in encoding time of be-
tween 30.0% to 45.0%. According to the result for
each sequence, the average BD-RATE increase for
our method is 0.5% while the BD-RATE increase is
0.9% in (Zhang et al., 2016). Both (Zhang et al.,
2016) and our method provide a good trade-off be-
tween BD-PSNR and encoding time saving.
5 CONCLUSION
This paper proposed a fast intra prediction algorithm
to lead to faster encoding time that more accurately
approximates the original H.265/HEVC standard re-
sults. Our algorithm introduces the weighted sam-
pling and adaptive thresholds condition to determine
CU-splitting. The proposed algorithm offers no di-
vergence from the H.265/HEVC standards and can
be used in current systems. The result of the per-
formance is shown to offer a more significant gain
in the higher resolution sequence with all inter mode.
Simulations also show that the proposed method can
speed up processing by 30.0% to 45.0% with little
BD-PSNR degradation and only 0.5% BD-RATE in-
crease while incurring a controlled precision and ad-
justable complexity overhead.
Fast Intra Prediction Algorithm with Enhanced Sampling Decision for H.265/HEVC
65
20 24 28 32
36
40
0
5
10
15
20
25
30
QP
coding time (second)
reference
proposed
Figure 4: H.265/HEVC, coding time curves proposed vs.
reference HM 16.9, paris QCIF@30Hz, all Intra pictures.
20 24 28 32
36
40
0
5
10
15
20
25
30
QP
coding time (second)
reference
proposed
Figure 5: H.265/HEVC, coding time curves proposed vs.
reference HM 16.9, carp QCIF@30Hz, all Intra pictures.
REFERENCES
Bossen, F., Bross, B., Suhring, K., and Flynn, D. (2012).
Hevc complexity and implementation analysis. IEEE
Transactions on Circuits and Systems for Video Tech-
nology, 22(12):1685–1696.
Chen, G., Pei, Z., Sun, L., Liu, Z., and Ikenaga, T. (2013).
Fast intra prediction for hevc based on pixel gradi-
ent statistics and mode refinement. In 2013 IEEE
China Summit & International Conference on Signal
and Information Processing (ChinaSIP), pages 514–
517. IEEE.
Cho, S. and Kim, M. (2013). Fast cu splitting and prun-
ing for suboptimal cu partitioning in hevc intra cod-
ing. IEEE Transactions on Circuits and Systems for
Video Technology, 23(9):1555–1564.
Choi, K. and Jang, E. S. (2012). Fast coding unit
decision method based on coding tree pruning for
20 24 28 32
36
40
0
50
100
150
200
250
300
QP
coding time (second)
reference
proposed
Figure 6: H.265/HEVC, coding time curves proposed vs.
reference HM 16.9, aspen 4CIF@30Hz, all Intra pictures.
20 24 28 32
36
40
0
50
100
150
200
250
300
350
QP
coding time (second)
reference
proposed
Figure 7: H.265/HEVC, coding time curves proposed vs.
reference HM 16.9, blue sky 4CIF@30Hz, all Intra pic-
tures.
high efficiency video coding. Optical Engineering,
51(3):030502–1.
da Silva, T. L., Agostini, L. V., and da Silva Cruz, L. A.
(2012). Fast hevc intra prediction mode decision
based on edge direction information. In 2012 Pro-
ceedings of the 20th European Signal Processing Con-
ference (EUSIPCO), pages 1214–1218. IEEE.
Huang, C.-T., Tikekar, M., and Chandrakasan, A. P. (2014).
Memory-hierarchical and mode-adaptive hevc intra
prediction architecture for quad full hd video decod-
ing. IEEE Transactions on Very Large Scale Integra-
tion (VLSI) Systems, 22(7):1515–1525.
Jamali, M., Coulombe, S., and Caron, F. (2015). Fast hevc
intra mode decision based on edge detection and satd
costs classification. In 2015 Data Compression Con-
ference, pages 43–52. IEEE.
Jiang, W., Ma, H., and Chen, Y. (2012). Gradient based fast
mode decision algorithm for intra prediction in hevc.
VISAPP 2017 - International Conference on Computer Vision Theory and Applications
66
In 2012 2nd International Conference on Consumer
Electronics, Communications and Networks (CEC-
Net), pages 1836–1840. IEEE.
Kim, I.-K., Min, J., Lee, T., Han, W.-J., and Park, J.
(2012). Block partitioning structure in the hevc stan-
dard. IEEE transactions on circuits and systems for
video technology, 22(12):1697–1706.
Lim, K., Lee, J., Kim, S., and Lee, S. (2015). Fast pu skip
and split termination algorithm for hevc intra predic-
tion. IEEE Transactions on Circuits and Systems for
Video Technology, 25(8):1335–1346.
Min, B. and Cheung, R. C. (2015). A fast cu size decision
algorithm for the hevc intra encoder. IEEE Transac-
tions on Circuits and Systems for Video Technology,
25(5):892–896.
Piao, Y., Min, J., and Chen, J. (2010). Encoder improve-
ment of unified intra prediction. JCTVC-C207.
Senzaki, K., Chono, K., Aoki, H., Tajime, J., and Senda,
Y. (2010). Bd-psnr/rate computation tool for five data
points. In Proc. of the Meeting of Joint Collaborative
Team on Video Coding. Geneva, Switzerland:[sn].
Shen, L., Liu, Z., Zhang, X., Zhao, W., and Zhang, Z.
(2013a). An effective cu size decision method for
hevc encoders. IEEE Transactions on Multimedia,
15(2):465–470.
Shen, L., Zhang, Z., and An, P. (2013b). Fast cu size
decision and mode decision algorithm for hevc intra
coding. IEEE Transactions on Consumer Electronics,
59(1):207–213.
Sullivan, G. J., Ohm, J.-R., Han, W.-J., and Wiegand, T.
(2012). Overview of the high efficiency video coding
(hevc) standard. IEEE Transactions on circuits and
systems for video technology, 22(12):1649–1668.
Wiegand, T., Sullivan, G. J., Bjontegaard, G., and Luthra,
A. (2003). Overview of the h. 264/avc video coding
standard. IEEE Transactions on circuits and systems
for video technology, 13(7):560–576.
Yan, S., Hong, L., He, W., and Wang, Q. (2012). Group-
based fast mode decision algorithm for intra predic-
tion in hevc. In 2012 Eighth International Conference
on Signal Image Technology and Internet Based Sys-
tems (SITIS), pages 225–229. IEEE.
Zhang, T., Sun, M.-T., Zhao, D., and Gao, W. (2016). Fast
intra mode and cu size decision for hevc.
Zhao, L., Zhang, L., Ma, S., and Zhao, D. (2011). Fast
mode decision algorithm for intra prediction in hevc.
In IEEE Visual Communications and Image Process-
ing (VCIP 2011), pages 1–4. IEEE.
Fast Intra Prediction Algorithm with Enhanced Sampling Decision for H.265/HEVC
67