HEVC编码器复杂性分析与优化策略

下载需积分: 5 | PDF格式 | 11.81MB | 更新于2024-07-16 | 58 浏览量 | 举报

"复杂性意识高效视频编码" 这本书深入探讨了高效视频编码（HEVC）标准在处理计算复杂性方面的问题，特别是在嵌入式系统中的应用。HEVC是视频压缩领域的一项重要进展，相较于之前的H.264/AVC标准，它能提供更高的压缩比，但这也带来了更大的编码计算复杂度，对处理时间和资源的需求增加。书中首先对HEVC标准进行了概述，解释了其工作原理，并回顾了该领域的最新研究进展。作者们特别关注了HEVC编码器中的计算复杂性，这是因为在高分辨率视频编码时，这个问题对于便携设备和实时系统尤为显著。他们提出，理解和分析编码过程中的计算需求是优化HEVC编码器的关键。第一个主要贡献是针对HEVC编码工具的详尽调查和分析。通过这项工作，研究人员能够识别出哪些编码步骤消耗了最多的计算资源，这对于优化编码流程至关重要。了解这些工具的行为有助于找到减少计算复杂性的策略。第二个贡献是一系列新颖的算法，这些算法旨在动态地调整HEVC编码器的复杂度。利用HEVC标准中允许的帧划分结构，如编码单元和预测单元的灵活性，这些算法能有效减少编码计算复杂性。据书中的描述，这些最佳复杂度缩放算法可以将编码复杂度降低至原始值的一半，同时仅造成很小的压缩效率损失，大约在20%以内。第三个部分介绍了基于数据挖掘技术的早期终止方法，这些方法可以进一步减少计算量。数据挖掘技术的应用可以智能地预测编码过程，从而提前结束某些计算密集型的任务，而不显著影响视频质量。这本书不仅对HEVC编码器的复杂性管理提供了深入见解，还为优化嵌入式系统中的视频编码流程提供了实用的解决方案。对于多媒体系统开发者、视频编码研究人员以及对高效视频压缩技术感兴趣的读者来说，这是一部非常有价值的参考著作。通过这些技术，可以在保持视频质量的同时，有效地适应有限的计算资源，这对于移动设备和实时视频流服务等应用场景尤其重要。

展开

List of Figures

Fig. 2.1 Generic hybrid video compression system (encoder) .................... 10

Fig. 2.2 Generic hybrid video compression system (decoder) .................... 11

Fig. 2.3 Rate-distortion (R-D) points and optimal curve ............................ 15

Fig. 2.4 Block diagram of a typical HEVC encoder ................................... 16

Fig. 2.5 Video sequence divided into a number of GOPs ........................... 17

Fig. 2.6 Frame divided into two slices and several CTUs ........................... 18

Fig. 2.7 Coding tree structure of a CTB divided into CBs .......................... 19

Fig. 2.8 Inter-frame and intra-frame PB splitting modes

available in HEVC ......................................................................... 20

Fig. 2.9 RQT structure of a CB divided into TBs ....................................... 21

Fig. 2.10 A frame divided into nine tiles ....................................................... 21

Fig. 2.11 Wavefront parallel processing encoding order ............................... 22

Fig. 2.12 Neighbouring samples used for intra-prediction

in an N × N PB of size 16 × 16 ........................................................ 23

Fig. 2.13 Intra-prediction modes tested for each PB ..................................... 24

Fig. 2.14 Examples of subpixel interpolations from full-pixel samples ....... 26

Fig. 2.15 Multiple reference frame prediction .............................................. 27

Fig. 2.16 Graphical presentation of the AI conﬁ guration ............................. 32

Fig. 2.17 Graphical presentation of the LD conﬁ guration ............................ 33

Fig. 2.18 Graphical presentation of the RA conﬁ guration ............................ 33

Fig. 4.1 Normalised computational complexity for encoding

each video sequence under all 17 conﬁ gurations (QP 32) ............. 67

Fig. 4.2 Normalised bit rate for each video sequence encoded

under all 17 conﬁ gurations (QP 32) ............................................... 67

Fig. 4.3 Delta Y-PSNR for each video sequence encoded

under all 17 conﬁ gurations (QP 32) ............................................... 68

Fig. 4.4 Normalised computational complexity for encoding

each video sequence under all 16 conﬁ gurations (QP 32) ............. 70

Fig. 4.5 Normalised bit rate for each video sequence encoded

under all 16 conﬁ gurations (QP 32) ............................................... 71

xvi

Fig. 4.6 Delta Y-PSNR for each video sequence encoded under

all 16 conﬁ gurations (QP 32) ......................................................... 71

Fig. 4.7 BD-rate values for each conﬁ guration using CFG 1 as reference .. 72

Fig. 4.8 BD-PSNR values for each encoding conﬁ guration

using CFG 1 as reference ............................................................... 72

Fig. 4.9 R-D efﬁ ciency of HEVC with CFG 1, CFG 12, CFG 15

and CFG 16 and H.264 HP for the (a) BQMall, (b) vidyo4,

Fig. 4.10 Encoding time of HEVC with CFG 1, CFG 12, CFG 15

and CFG 16 and H.264 HP for the (a) BQMall, (b) vidyo4,

Fig. 4.11 Normalised computational complexity for encoding

each video sequence under the seven frame partitioning

conﬁ gurations (QP 32) ................................................................... 81

Fig. 4.12 Normalised bit rate for each video sequence encoded

under the seven frame partitioning conﬁ gurations (QP 32) ........... 81

Fig. 4.13 Delta Y-PSNR for each video sequence encoded

under the seven frame partitioning conﬁ gurations (QP 32) ........... 82

Fig. 4.14 BD-rate values for each conﬁ guration using PAR 1

as reference .................................................................................... 82

Fig. 4.15 BD-PSNR values for each conﬁ guration using PAR 1

as reference .................................................................................... 83

Fig. 5.1 Computational complexity of encoding CUs in each coding

tree depth ........................................................................................ 88

Fig. 5.2 Maximum coding tree depth of random co-localised

64 × 64 areas for sequences (a) BQTerrace (QP 27),

(b) BasketballDrive (QP 32) and (c) Cactus (QP 37) .................... 90

Fig. 5.3 Diagram for the FDCS algorithm .................................................. 91

Fig. 5.4 Average R-D performance of the FDCS method ........................... 93

Fig. 5.5 Example of operation of the VDCS strategy ................................. 94

Fig. 5.6 Diagram for the VDCS algorithm .................................................. 95

Fig. 5.7 Adjustment of Nc value according to β ......................................... 95

Fig. 5.8 Nc and CP

variations when encoding sequence

BasketballDrive, QP32 ................................................................... 97

Fig. 5.9 Accuracy of the VDCS complexity scaling method ...................... 97

Fig. 5.10 Average R-D performance of the VDCS method .......................... 98

Fig. 5.11 Example of MTDM motion compensation .................................... 100

Fig. 5.12 Diagram for the MCTDL algorithm .............................................. 101

Fig. 5.13 MTDM fragment and average MVs for the (a) 58th and

(b) 59th frames of the BasketballDrive sequence (QP 32) ............ 102

Fig. 5.14 Accuracy of the MCTDL complexity scaling method ................... 103

Fig. 5.15 Diagram for the CTDE algorithm .................................................. 106

Fig. 5.16 Pseudocode for the maximum coding tree depth decision

in CTDE ......................................................................................... 107

List of Figures

xvii

Fig. 5.17 Accuracy of the CTDE complexity scaling method ...................... 108

Fig. 5.18 Pseudocode for the CCUPU method .............................................. 113

Fig. 5.19 Complexity scalability accuracy for ten video sequences

using QP 27 (a), QP 32 (b), QP 37 (c) and QP 42 (d) ................... 117

Fig. 5.20 Effect of computational complexity scalability

over the R-D performance .............................................................. 118

Fig. 5.21 R-D performance for each target complexity for videos

(a) Trafﬁ c and (b) RaceHorses1 encoded with QPs 27, 32, 37

and 42 ............................................................................................. 119

Fig. 5.22 Average encoding time per frame in each GOP

for different target complexities: (a) PartyScene, QP 27;

(b) vidyo1, QP 32; (c) BasketballDrive, QP 37;

and (d) Trafﬁ c, QP 42 .................................................................... 121

Fig. 5.23 Average encoding time per frame in each GOP

for different target complexities, considering video

RaceHorses1 and QP 32 ................................................................. 122

Fig. 5.24 Target complexity versus running complexity

for the complexity scaling methods ............................................... 122

Fig. 5.25 BD-rate increase for the ﬁ ve complexity scaling methods ............ 123

Fig. 5.26 BD-PSNR for the ﬁ ve complexity scaling methods ...................... 123

Fig. 6.1 Example of a binary classiﬁ cation tree .......................................... 126

Fig. 6.2 Example of an ARFF ﬁ le ............................................................... 128

Fig. 6.3 Occurrence of (a) 64 × 64 CUs, (b) 32 × 32 CUs

and (c) 16 × 16 CUs split and not split into smaller CUs

according to the PU mode chosen (i.e. Partition) .......................... 131

Fig. 6.4 Neighbouring CTUs used in the calculation of ΔNeighDepth ...... 132

Fig. 6.5 Occurrence of (a) 64 × 64 CUs, (b) 32 × 32 CUs

and (c) 16 × 16 CUs split and not split into smaller CUs

according to the average of CU depths in neighbouring CTUs ..... 132

Fig. 6.6 Occurrence of (a) 64 × 64 CUs, (b) 32 × 32 CUs

and (c) 16 × 16 CUs split and not split into smaller CUs

according to the ratio between the 2N × 2N

and MSM R-D costs ....................................................................... 134

Fig. 6.7 CTUs divided into CUs in the 100th frame

of the BasketballDrill video sequence encoded with QP 32

by (a) the original HM encoder and the (b) HM encoder

with the coding tree early termination ........................................... 135

Fig. 6.8 Frequency of occurrence of each inter-PU splitting

mode in the BasketballDrive sequence encoded

with (a) QP 27, (b) QP 32, (c) QP 37 and (d) QP 42 ..................... 136

Fig. 6.9 Inter-frame PU splitting mode decision with the proposed

early termination ............................................................................ 137

List of Figures

xviii

Fig. 6.10 Occurrence of (a) 64 × 64 CUs, (b) 32 × 32 CUs,

into PUs smaller than 2N × 2N according to the ratio

between the R-D costs of 2N × 2N and MSM modes..................... 138

Fig. 6.11 Occurrence of (a) 32 × 32 CUs, (b) 16 × 16 CUs

and (c) 8 × 8 CUs split and not split into PUs smaller

than 2N × 2N according to the splitting decision

at the upper coding tree depth ........................................................ 139

Fig. 6.12 PU boundaries in the 100th frame of the BasketballDrill

video sequence encoded with QP 32

by (a) the original HM encoder and the (b) HM encoder

with the PU early termination ........................................................ 141

Fig. 6.13 Occurrence of (a) 32 × 32 and (b) 16 × 16 TUs split

and not split into smaller TUs according to the absolute

sum of luminance residues ............................................................. 142

Fig. 6.14 Occurrence of (a) 32 × 32 and (b) 16 × 16 TUs split

and not split into smaller TUs according to the number

of nonzero luminance coefﬁ cients after the transform .................. 143

Fig. 6.15 Occurrence of (a) 32 × 32 and (b) 16 × 16 TUs split

and not split into smaller TUs according to the R-D cost

for the current TU depth ................................................................ 144

Fig. 6.16 TUs in the 100th frame of the BasketballDrill

video sequence encoded with QP 32

by (a) the original HM encoder and the (b) HM encoder

with the RQT early termination ..................................................... 145

Fig. 6.17 Number of bits per frame in the SlideEditing sequence

(QP 32) encoded with the original HM and the modiﬁ ed HM

with the RQT early termination ..................................................... 149

Fig. 6.18 Fragment of the 11th frame in the SlideEditing sequence

(QP 32) encoded with (a, b) the original HM and (c, d)

the modiﬁ ed HM with the RQT early termination ......................... 150

Fig. 6.19 Rate-distortion efﬁ ciency for the (a) SlideEditing,

(b) BQTerrace, (c) Kimono and (d) PeopleOnStreet

video sequences encoded with the original HM

and the three early terminations implemented separately .............. 151

Fig. 6.20 Rate-distortion efﬁ ciency for the (a) BQTerrace,

(b) BQSquare, (c) Kimono and (d) PeopleOnStreet video

sequences encoded with the original HM and the four joint

early termination schemes implemented ........................................ 154

Fig. 7.1 Encoding time control system integrated

with the HEVC encoder ................................................................. 160

Fig. 7.2 Example of R-D-C space and 3-D Pareto frontier

for a set of 240 encoding parameter conﬁ gurations ....................... 161

List of Figures

xix

Fig. 7.3 (a) R-C and (b) D-C average results for the 240 conﬁ gurations

tested. Squares represent Pareto-efﬁ cient conﬁ gurations

in the R-C space, circles represent Pareto-efﬁ cient

conﬁ gurations in the D-C space and × symbols

represent the remaining conﬁ gurations .......................................... 165

Fig. 7.4 Encoding time control system integrated

with the HEVC encoder ................................................................. 168

Fig. 7.5 Pseudocode for the encoding time limitation system .................... 171

Fig. 7.6 (a) Encoding times, (b) conﬁ guration index k,

c(i)

parameter and (d) N

c(i)

parameter

for each GOP in the Beauty sequence (QP 22) .............................. 173

Fig. 7.7 (a) Encoding times, (b) conﬁ guration index k,

c(i)

parameter and (d) N

c(i)

parameter

for each GOP in the YachtRide sequence (QP 32) ........................ 174

Fig. 7.8 Encoding times for each GOP in the concatenation

of the Beauty and YachtRide sequences (QP 32),

with the scene change in GOP 150 ................................................ 175

Fig. A.1 BaskeballDrillText ......................................................................... 185

Fig. A.2 BaskeballDrill ................................................................................ 185

Fig. A.3 BasketballDrive .............................................................................. 186

Fig. A.4 BasketballPass ................................................................................ 186

Fig. A.5 Beauty ............................................................................................ 187

Fig. A.6 BlowingBubbles ............................................................................. 187

Fig. A.7 Bosphorus ...................................................................................... 188

Fig. A.8 BQMall .......................................................................................... 188

Fig. A.9 BQSquare ....................................................................................... 189

Fig. A.10 BQTerrace ...................................................................................... 189

Fig. A.11 Cactus ............................................................................................. 190

Fig. A.12 ChinaSpeed .................................................................................... 190

Fig. A.13 FourPeople ..................................................................................... 191

Fig. A.14 HoneyBee ....................................................................................... 191

Fig. A.15 Jockey............................................................................................. 192

Fig. A.16 Johnny ............................................................................................ 192

Fig. A.17 Kimono .......................................................................................... 193

Fig. A.18 KristenAndSara .............................................................................. 193

Fig. A.19 NebutaFestival ................................................................................ 194

Fig. A.20 ParkScene ....................................................................................... 194

Fig. A.21 PartyScene ...................................................................................... 195

Fig. A.22 PeopleOnStreet .............................................................................. 195

Fig. A.23 Poznan_CarPark ............................................................................. 196

Fig. A.24 Poznan_Hall1 ................................................................................. 196

Fig. A.25 PoznanStreet .................................................................................. 197

Fig. A.26 RaceHorses1 .................................................................................. 197

Fig. A.27 RaceHorses2 .................................................................................. 198

List of Figures

剩余245页未读，继续阅读

身份认证购VIP最低享 7 折!

30元优惠券

weixin_39841882

粉丝: 447

HEVC编码器复杂性分析与优化策略

轻松解密PDF：一键解锁加密文件

关系数据库设计系列：从基础到优化

ROBT4451_Lab6_OptEnc：C语言优化编码实验室实践

bengio2021.pdf

基于深度学习的显著性检测方法模型——SCS.pdf

音视频-编解码-设计中字体的情感传达.pdf

java实习报告.pdf

ICML 2019 笔记.pdf

Wlan网络安全.pdf

电子政务信息安全体系.doc.pdf

最新资源