基于时空相关性的HEVC帧间模式决策快速算法

doi:10.11959/j.issn.1000-436x.2016073

Abstract

Abstract:

The new generation high efficiency video coding (HEVC) standard adopts coding tree unit (CTU) quadtree partition technology and ten inter prediction unit (PU) modes to improve compression efficiency significantly, but it brings in an enormous encoding computational complexit . A spatio-temporal correlation based fast inter mode decision algorithm for HEVC was proposed to reduce partition times of coding unit (CU) and inter PU mode candidates. First, the spatio-temporal correlation of depth information among the current CTU, the same located CTU in the reference frame and the spatial neighbor CTU in the current frame are utilized to predict depth range of the current CTU. Then, the spatial correlation of optimal PU modes between the current CU and its parent-CU was analyzed. And the rate-distortion (RD) costs of the estimated PU modes of the current CU were also used. Based on these, the redundant inter PU modes of current CU are skipped. The experiment results demonstrate that as compared to the HEVC test model (HM), the proposed algorithm saves about 52% encoding time under different encoding configurations, and it maintains a good RD performance. When compared to HM with fast mode decision algorithms enabled, the proposed algorithm further saves about 30% encoding time.

Key words: high efficiency video coding, coding tree unit, prediction unit, spatio-temporal correlation

Wei ZHU,Xun-hua ZHANG,Cai-sheng WANG,Hua2 ZHANG. Spatio-temporal correlation based fast inter mode decision for HEVC[J]. Journal on Communications, 2016, 37(4): 64-73.

Figures/Tables 14

References 15

[1]	SULLIVAN G J , OHM J R , HAN W J , et al. Overview of the high efficiency video coding (HEVC) standard[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2012,22(12):1649-1668.
[2]	KIM I K , MIN J , LEE T , et al. Block partition structure in the HEVC standard[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2012,22(12):1697-1706.
[3]	周承涛，田翔，陈耀武． HEVC 编码单元尺寸快速选择算法[J]．浙江大学学报（工学版）， 2012,48(8):1451-1460. ZHOU C T , TIAN X , CHEN Y W . Fast coding unit size decision for HEVC[J]. Journal of Zhejiang University(Engineering Science), 2012,48(8):1451-1460.
[4]	GARCIA R , KALVA H . HEVC inter-frame skip enhancement at low bit rates[C]// 2014 IEEE International Conference on Consumer Elec-tronics(ICCE). Las Vegas, c2014:59-60.
[5]	AHN S , LEE B , KIM M . A novel fast CU encoding scheme based on spatiotemporal encoding parameters for HEVC inter codi g[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2015,25(3):422-435.
[6]	蒋刚毅，杨小祥，彭宗举，等．高效率视频编码的快速编码单元深度遍历选择和早期编码单元裁剪[J]．光学精密工程， 2014,22(5):1322-1330. JIANG G Y , YANG X X , PENG Z J , et al. Fast CU depth ra g selec-tion and early CU pruning for HEVC[J]. Optics and Prec ion Engi-neering, 2014,22(5):1322-1330.
[7]	SHEN L Q , LIU Z , ZHANG X P , et al. An effective CU size decision method for HEVC encoders[J]. IEEE Transactions on Mult media, 2013,15(2):465-470.
[8]	齐美彬，陈秀丽，杨艳芳，等．高效率视频编码帧内预测编码单元划分快速算法[J]．电子与信息学报， 2014,36(7):1699-1705. QI M B , CHEN X L , YANG Y F , et al. Fast coding unit splitting algo-rithm for high efficiency video coding intra predictio Journal of Electronics ＆ Information Technology, 2014,36(7):1699-1705.
[9]	SHEN L Q , ZHANG Z Y , LIU Z . Effective CU size decision for HEVC intracoding[J]. IEEE Transactions on Image Processing, 2014,23(10):4232-4241.
[10]	BOSSEN F . Common test conditions and software reference configura-tions[C]// Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC, JCTVC-J1100. Stockholm, c2012:1-3.
[11]	BJONTEGAARD G . Calculation of average PSNR differences between RD curves[C]// ITU-T SG16/Q6, VCEG-M33. Austin, c2001:1-4.
[12]	OHM J R , SULLIVAN G J , SCHWARZ H , et al. Comparison of the coding efficiency of video coding standards—including high efficien-cy video coding (HEVC)[J]. IEEE Transactions on Circuits a Sys-tems for Video Technology, 2012,22(12):1669-1684.
[13]	GWEON R H , LEE Y L , LIM J . Early termination of CU encoding to reduce HEVC complexity[C]// Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC, JCTVC-F045. Torino, c2011:1-4.
[14]	CHOI K , PARK S H , JANG E S . Coding tree pruning based early termination[C]// Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC, JCTVC-F092. Torino, c2011:1-4.
[15]	YANG J , KIM J , WON K , et al. Early SKIP detection for HEVC[C]// Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC, JCTVC-G543. Geneva, c2011:1-6.

Metrics

Recommended 0

No Suggested Reading articles found!

序列		C TU的Depth比例/%
序列	0	12	3
TS0	16.8	7.112.1	63.9
TS1	45.6	10.310.5	33.5
TS2	10.5	15.023.9	50.6
TS3	38.5	15.311.5	34.7
TS4	75.9	7.512.8	3.8
平均值	37.5	11.114.2	37.3

序列	深度				最佳PU模式比例/%
序列	深度	Skip	Merge	Inter 2N×2N	Inter N×2N	Inter 2N×N	Inter AMP	其他
	0	14.1	4.5	7.2	27.6	28.0	13.4	5.2
TS0	1	31.3	6.3	6.2	13.3	14.8	23.3	4.8
TS0	2	56.6	7.1	8.6	5.7	6.4	9.3	6.3
	3	78.7	5.2	5.9	2.2	2.2	0.0	5.8
	0	41.2	4.0	10.6	14.4	17.0	11.0	1.7
TS1	1	64.4	5.6	5.9	5.8	5.7	10.8	1.9
TS1	2	81.0	5.4	3.7	2.3	1.9	3.5	2.1
	3	90.4	3.9	1.7	1.0	0.8	0.0	2.1
	0	9.2	3.3	11.0	33.2	27.1	13.5	2.7
TS2	1	36.6	6.9	9.6	14.0	9.5	20.0	3.5
TS2	2	62.9	8.8	7.2	5.5	3.5	7.8	4.4
	3	80.7	6.4	3.7	2.9	1.4	0.0	4.8
	0	29.7	16.4	9.1	18.4	15.4	8.7	2.3
TS3	1	48.0	16.9	5.7	8.0	6.4	11.3	3.7
TS3	2	70.4	10.5	4.8	2.8	2.2	3.6	5.8
	3	85.8	5.4	2.4	1.0	0.6	0.0	4.7
	0	77.8	3.5	2.6	6.2	5.4	3.5	1.0
TS4	1	88.8	2.4	1.8	2.0	1.4	2.6	1.1
TS4	2	95.1	1.1	1.1	0.6	0.3	0.6	1.1
	3	97.9	0.4	0.4	0.2	0.1	0.0	1.1
	平均值	62.0	6.2	5.5	8.4	7.5	7.1	3.3

序列		Cur-CTU与相邻CTU的Depth差值
序列	L-CTU	LU-CTU	U-CTU	RU-CTU	Col-CTU
TS0	0.56	0.72	0.61	0.67	0.38
TS1	0.73	0.96	0.79	0.99	0.87
TS2	0.73	0.77	0.72	0.88	0.80
TS3	0.93	1.19	0.80	1.06	0.56
TS4	0.32	0.48	0.41	0.51	0.32
平均值	0.65	0.82	0.67	0.82	0.59

QP		CDRP算法的准确率/%
QP	TS0	TS1	TS2	TS3	TS4
22	97.7	96.4	98.5	94.5	96.9
27	97.2	94.9	97.2	95.8	97.9
32	97.2	95.3	94.7	96.4	98.9
37	96.7	96.7	95.1	96.8	99.4
平均值	97.2	95.8	96.4	95.9	98.3

序列	父CU最佳PU模式			当前CU的最佳PU模式比例/%
序列	父CU最佳PU模式	Skip	Merge	Inter 2N×2N	Inter N×2N	Inter 2N×N	Inter AMP	其他
	Skip	98.8	0.4	0.4	0.2	0.2	0.0	0.1
	Merge	53.2	21.9	5.7	3.9	3.7	1.8	9.7
TS0	Inter 2N×2N	60.0	9.1	16.1	3.5	3.6	2.0	5.7
	Inter N×2N	45.8	8.4	15.1	7.7	8.2	8.8	6.1
	Inter 2N×N	45.3	9.0	12.2	9.5	10.1	7.4	6.6
	Skip	98.9	0.4	0.2	0.3	0.1	0.1	0.0
	Merge	52.4	29.1	5.0	3.7	3.1	1.9	4.8
TS1	Inter 2N×2N	62.2	13.1	12.6	3.5	3.0	2.8	2.9
	Inter N×2N	57.9	12.1	10.9	5.4	4.7	6.3	2.7
	Inter 2N×N	52.6	13.2	12.0	6.1	6.1	7.7	2.4

Spatio-temporal correlation based fast inter mode decision for HEVC

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 14

References 15

Related Articles 2

Metrics

Recommended 0

序列	不同估计阶段	率失真代价最小的PU模式		当前CU最佳PU模式比例/%
序列	不同估计阶段	率失真代价最小的PU模式	Inter N×2N	Inter 2N×N	AMP	其他
TS2	Inter 2N×2N模式估计后	Skip或Merge	2.8	1.5	1.2	94.5
TS2	Inter N×2N和Inter 2N×N模式估计后	Skip、Merge或Inter 2N×2N	0	0	1.2	98.8
TS3	Inter 2N×2N模式估计后	Skip或Merge	1.2	0.7	0.6	97.5
TS3	Inter N×2N和Inter 2N×N模式估计后	Skip、Merge或Inter 2N×2N	0	0	0.7	99.3

判别条件	当前CU跳过的帧间PU模式
父CU最佳PU为Skip	Inter N×2N、Inter 2N×N和Inter AMP
父CU最佳PU为Merge或Inter 2N×2N	Inter AMP
Inter 2N×2N估计后当前CU符合S k ip条件且Flag 1等于1	Inter N×2N、Inter 2N×N和Inter AMP
Inter 2N×2N估计后当前CU不符合S k ip条件且Flag 2等于1	Inter N×2N、Inter 2N×N和Inter AMP
Inter N×2N和Inter 2N×N估计后Flag3等于1	Inter AMP

QP		IPMS算法的准确率/%
QP	TS0	TS1	TS2	TS3	TS4
22	95.2	95.6	91.3	97.4	98.6
27	97.5	98.0	95.7	98.6	99.5
32	98.4	99.1	98.0	99.2	99.8
37	98.8	99.6	99.1	99.5	99.9
平均值	97.5	98.1	96.0	98.7	99.4

序列	LDP配置		RA配置
序列	本文CDRP算法BDBR/BDPSNR/?T (%/dB/%)	本文IPMS算法BDBR/BDPSNR/?T (%/dB/%)	本文CDRP算法BDBR/BDPSNR/?T (%/dB/%)	本文IPMS算法BDBR/BDPSNR/?T (%/dB/%)
Traffic	—	—	0.32/-0.01/26.6	0.75/-0.02/49.4
PeopleOnStreet	—	—	0.21/-0.01/21.2	1.01/-0.04/32.2
Kimono	0.37/-0.01/29.7	0.82/-0.02/40.9	0.56/-0.01/30.8	0.72/-0.02/44.9
ParkScene	0.29/-0.01/23.3	1.46/-0.04/43.7	0.29/-0.01/24.1	0.84/-0.02/47.6
Cactus	0.32/-0.01/26.8	1.39/-0.03/40.7	0.25/0.00/26.2	0.75/-0.01/44.9
BQTerrace	0.40/-0.01/28.1	1.75/-0.02/44.6	0.25/0.00/27.6	0.78/-0.01/49.1
BasketballDrive	0.47/-0.01/26.7	0.94/-0.02/40.3	0.44/-0.01/26.7	0.83/-0.02/43.7
RaceHorsesC	0.30/-0.01/18.0	1.28/-0.04/31.7	0.46/-0.02/19.3	1.27/-0.04/33.4
BQMall	0.46/-0.02/20.8	1.63/-0.06/39.0	0.21/-0.01/20.9	1.10/-0.04/42.8
PartyScence	0.11/0.00/19.8	1.41/-0.05/34.5	0.18/-0.01/20.3	0.78/-0.03/38.6
BasketballDrill	0.43/-0.02/19.9	1.06/-0.04/36.2	0.32/-0.01/19.9	0.73/-0.03/39.1
RaceHorses	0.11/0.00/16.9	1.70/-0.07/29.4	0.25/-0.01/18.9	1.35/-0.06/30.2
BQSquare	0.11/0.00/18.8	1.95/-0.06/38.0	0.63/-0.02/20.8	0.68/-0.02/45.3
BlowingBubbles	0.18/-0.01/17.4	2.01/-0.07/36.6	0.10/0.00/19.5	0.87/-0.03/40.2
BasketballPass	0.11/-0.01/14.9	1.21/-0.05/32.3	0.25/-0.01/16.4	1.08/-0.05/34.6
FourPeople	0.85/-0.03/35.1	1.43/-0.04/51.1	—	—
Johnny	1.02/-0.02/42.2	2.23/-0.05/55.6	—	—
KristenAndSara	0.94/-0.03/38.5	1.58/-0.04/53.1	—	—
平均值	0.40/-0.01/24.8	1.49/-0.04/40.5	0.31/-0.01/22.6	0.90/-0.03/41.1

[1]	Yanchao GONG, Ling WANG, Ying LIU, Kaifang YANG, Qingfan LIN, Fuping WANG. HEVC perceptual intra-frame rate control inspired by video subjective observation experiment [J]. Journal on Communications, 2021, 42(8): 90-102.
[2]	Yue LI,Gaobo YANG,Xiangling DING,Yapei ZHU. Early Merge mode decision algorithm for 3D-HEVC based on learning model [J]. Journal on Communications, 2019, 40(7): 104-113.

序列	LDP配置		RA配置
	文献[7]算法	本文总体算法	文献[7]算法	本文总体算法
	BDBR/BDPSNR/?T (%/dB/%)	BDBR/BDPSNR/?T (%/dB/%)	BDBR/BDPSNR/?T (%/dB/%)	BDBR/BDPSNR/?T (%/dB/%)
Traffic	—	—	0.65/-0.02/49.1	1.12/-0.03/60.1
PeopleOnStreet	—	—	0.89/-0.03/31.4	1.25/-0.05/46.9
Kimono	0.36/-0.01/37.3	1.28/-0.04/55.2	0.56/-0.02/42.6	1.26/-0.03/59.0
ParkScene	1.05/-0.03/39.8	1.79/-0.05/53.8	0.71/-0.02/46.3	1.19/-0.03/58.3
Cactus	0.97/-0.02/40.5	1.69/-0.03/53.2	0.66/-0.01/44.8	0.96/-0.02/56.8
BQTerrace	1.73/-0.02/43.7	2.02/-0.03/55.8	1.13/-0.02/48.9	1.04/-0.01/60.2
BasketballDrive	0.69/-0.02/39.4	1.39/-0.03/52.8	0.75/-0.02/43.1	1.30/-0.03/55.9
RaceHorsesC	0.87/-0.03/29.2	1.51/-0.05/43.7	0.84/-0.03/30.0	1.88/-0.06/46.7
BQMall	1.14/-0.04/34.7	2.00/-0.07/49.1	0.77/-0.03/39.9	1.88/-0.05/53.7
PartyScence	0.96/-0.03/28.5	1.59/-0.06/46.0	0.77/-0.03/35.7	0.91/-0.03/51.0
BasketballDrill	0.89/-0.03/34.9	1.32/-0.05/46.6	0.67/-0.03/37.5	1.03/-0.04/49.6
RaceHorses	0.81/-0.03/23.0	1.81/-0.07/41.6	0.79/-0.03/24.4	1.70/-0.07/44.7
BQSquare	1.61/-0.04/28.9	2.05/-0.06/47.6	0.69/-0.02/40.8	0.80/-0.03/55.7
BlowingBubbles	1.15/-0.04/26.3	1.99/-0.07/45.3	0.78/-0.03/35.4	1.01/-0.04/51.1
BasketballPass	0.75/-0.03/28.2	1.49/-0.07/42.7	0.74/-0.03/31.2	1.31/-0.06/45.1
FourPeople	1.92/-0.06/54.7	2.50/-0.07/63.4	—	—
Johnny	1.42/-0.03/58.3	2.84/-0.06/68.7	—	—
KristenAndSara	1.02/-0.03/54.3	2.59/-0.07/65.8	—	—
平均值	1.08/-0.03/37.6	1.87/-0.05/52.0	0.76/-0.02/38.8	1.21/-0.04/53.0

序列	LDP配置	RA配置
序列	BDBR/BDPSNR/(%/dB/%) ?T	BDBR/BDPSNR/(%/dB/%) ?T
Traffic	—	1.72/-0.05/29.5
PeopleOnStreet	—	1.80/-0.07/35.7
Kimono	1.53/-0.04/37.5	1.58/-0.04/36.1
ParkScene	1.99/-0.05/29.4	0.71/-0.02/30.1
Cactus	2.29/-0.05/32.6	2.01/-0.04/31.5
BQTerrace	2.01/-0.03/28.2	1.17/-0.02/27.4
BasketballDrive	1.53/-0.03/33.4	1.65/-0.03/32.2
RaceHorsesC	1.64/-0.05/33.3	2.90/-0.10/34.2
BQMall	2.00/-0.07/29.4	1.89/-0.07/29.6
PartyScence	1.61/-0.06/30.4	1.27/-0.05/32.2
BasketballDrill	1.64/-0.06/30.8	1.44/-0.05/30.2
RaceHorses	1.99/-0.08/31.7	3.14/-0.13/34.0
BQSquare	2.03/-0.06/26.3	1.00/-0.03/27.4
BlowingBubbles	1.83/-0.06/28.8	1.50/-0.06/30.2
BasketballPass	1.60/-0.07/29.1	1.91/-0.09/29.3
FourPeople	2.65/-0.08/25.4	—
Johnny	3.21/-0.06/23.8	—
KristenAndSara	3.15/-0.08/27.8	—
平均值	2.04/-0.06/29.9	1.71/-0.06/31.3