A New Rate-Complexity-QP Algorithm for HEVC Intra

Download Report

Transcript A New Rate-Complexity-QP Algorithm for HEVC Intra

A New Rate-Complexity-QP Algorithm for
HEVC Intra-Picture Rate Control
L I N G T I A N , YI MI N ZHOU, A N D X I AOJU N CAO
2 0 1 4 I N T ERNATIONAL CON F E R ENCE ON COM P U T I NG, N E T WORK ING A N D
COMMUN I CATIONS, M U LT I MEDIA COM P U T I NG A N D COM M U N I CATIONS SYM P OSIU M
1
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
2
Introduction
To take advantage of the limited network bandwidth while maintaining optimized visual quality
of video streams, rate control schemes play an indispensable role in video compression and
communication.
Superabundant bits stream may result in network traffic jam or unexpected frame loss.
Over-reduced bits stream may lead to the underutilization of network bandwidth and unnecessary
video quality degradation.
3
Introduction(cont.)
In this paper, different Rate-Complexity-QP model and a RCQA algorithm is proposed.
This algorithm effectively takes the picture content complexity, intra picture rate control and QP
into consideration.
The algorithm is implemented in HM 9.1.
4
Introduction
5
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
6
Related Work
Several classic R-D models and related work will be discussed in this section.
1.
ρ-domain model [6]
2.
Quadratic model (Q-domain) [7]
3.
R𝜆-domain model [4]
4.
Exponential model [9][10]
7
Related Work (cont.)
1.
ρ-domain model
ρ is the percentage of zeros among
the quantized transform coefficient.
2.
Quadratic model
After observing the relationship between Qstep and quantization, the equation is modified by [8]
8
Related Work (cont.)
3.
R𝜆-domain model
4.
Exponential model
intra-picture complexity
similar to MAD, may lead to the model error
9
Related Work (cont.)
[4] B. Li, H. Li, L. Li, and et al. “Rate Control by Rr-lambda Model for HEVC,” in Joint Collaborative
Team on Video Coding (JCT-VC) 11th Meeting, JCTVC-K0103, Shanghai, China, 2012.
[6] Z. He, Y.K. Kim and S.K. Mitra, “Low Delay Rate Control for DCT Video Coding via ρ-domain
Source Modeling,” IEEE Transactions on Circuits and Systems for Video Technology, vol.11,
pp.928–940, 2001.
[7] T. Chiang and Y. Q. Zhang, “A New Rate Control Scheme Using Quadratic Rate-Distortion
Modeling” IEEE Transactions on Circuits System and Video Technology, vol.7, pp 246-250,
Feb.1997.
[9] L. Tian, Y. Sun, Y. Zhou, and et al., “Analysis of Quadratic R-D Model in H.264/AVC Video
Coding,” 17th IEEE International Conference on Image Processing, pp. 2853-2856, China, 2010.
[10] Y. Zhou, Y. Sun, Z. Feng, and et al., “New Rate-Distortion Modeling and Efficient Rate Control
for H. 264/AVC video coding,” Signal Processing: Image Communication, vol. 24, pp 345-356,
2009.
10
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
11
Proposed R-D model for HEVC
The existing R-D models either neglect some factors (like frame content or complexity),the
authors proposed a comprehensive R-D model for HEVC based on practical experiments and
curve fitting.
Without generality, the authors use the function 𝑓𝐷 and 𝑓𝑅 to describe the relationships
between all the components.
Since the intra-picture coding complexity only comes from the individual picture content and
QP has no impact on the picture content, so 𝑓𝑅 can be rewritten as
R ∝ 𝑓𝑐 (𝐶) when QP is fixed
R ∝ 𝑓𝑄 (𝑄) when content complexity is fixed
12
Proposed R-D model for HEVC (cont.)
By the relationship mentioned above, four model should be discussed below
1.
Linear distortion-quantization (D-Q) model
2.
Exponential rate-quantization (R-Q) model
3.
Linear rate-complexity (R-C) model
4.
Uniform rate-complexity-quantization model
13
Proposed R-D model for HEVC (cont.)
1.
Linear distortion-quantization (D-Q) model
By observing Y-axis in Fig.1.a, we can see a linear relationship
between D and Q.
14
Proposed R-D model for HEVC (cont.)
2.
Exponential rate-quantization (R-Q) model
By observing Y-axis in Fig.1.a, based on the curves shape and the
exponential model for H.264 in [9], the model can be as
15
Proposed R-D model for HEVC (cont.)
3.
Linear rate-complexity (R-C) model
Using the variation of picture content to determined the complexity of the picture
The author use the average gradient per pixel to denote the picture content complexity
16
Proposed R-D model for HEVC (cont.)
3.
Linear rate-complexity (R-C) model
By observing Y-axis in Fig.1.b, based on the dot on the figure,
the model can be as
17
Proposed R-D model for HEVC (cont.)
4.
Uniform Rate-Complexity-Quantization (RCQ) Model
Since C is independent of Q, we propose a uniform RCQ model to depict the relationships among rate,
complexity and quantization
18
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
19
QP calculation for rate control
To get more efficient and fine-granulate rate control, the authors further derive the total
differential of the RCQ model to obtain the incremental QP for the encoding process in HEVC.
1.
Total differential of the uniform RCQ model
2.
Incremental QP calculation
3.
RCQ model update
4.
RCQ based rate control algorithm
20
QP calculation for rate control (cont.)
1.
Total differential of the uniform RCQ model
(全微分 dR 為對應於 c 和 q 小變化時 R 變化量的近似值。)
21
QP calculation for rate control (cont.)
1.
Total differential of the uniform RCQ model
For a small variation range of bit-rate, means △ 𝑅 = 𝑑𝑅, by integrating Eq.(14) into Eq. (15) at a specific
(r, c, q)
22
QP calculation for rate control (cont.)
2.
Incremental QP calculation
Since the content and motion complexity of the successive pictures have a high correlation, the bitrate
difference between two successive pictures can be
Hence, QP value of the current frame can be obtained according to the feedback of rate control buffer
23
QP calculation for rate control (cont.)
3.
RCQ model update
After current picture encoding is finished, RCQ model parameter should be updated.
𝑆𝑡∗ : = 𝑆𝑡 when 𝑞𝑖 = 𝑄𝑚 , 𝑆𝑡# ≔ 𝑆𝑡 otherwise;
Because of the content correlation and context similarity between the successive pictures, one can see
that the QP value has more influence on the output bit-rate than the picture content complexity does.
The equation (13) can be rewritten as :
𝛾
For the parameter γ , we convert the set 𝑆𝑡# to 𝑆𝑡 as in Eq. (21) and adopt the linear regression with the
least square method to update γ .
24
QP calculation for rate control (cont.)
3.
RCQ model update
Given a specific value γ , Eq. (13) can be rewritten as Eq. (22).
For the parameter α , we convert the set 𝑆𝑡∗ to 𝑆𝑡𝛼 as Eq. (23), then adopt the linear regression with the
least square method to update α
25
QP calculation for rate control (cont.)
4.
RCQ based rate control algorithm
26
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
27
Experimental Result
The authors implement RCQA to control bit-rate under the platform of HEVC reference
software HM 9.1, which is compared with HM 9.1 RC and the normal HM 9.1 fixed QP (FIXQP)
rate control schemes.
Video sequences are coded in an all intra configuration with high efficiency coding parameters
as described in “encoderintra-main.cfg”.
28
Experimental Result (cont.)
The proposed RCQ model is implemented in HM 9.1, which is compared with HM 9.1 RC and
the normal HM 9.1 fixed QP rate control schemes.
29
Experimental Result (cont.)
Table III shows the results of rate control accuracy, which
is calculated by |TBR − ABR | /TBR , where TBR is obtained
from the FIXQP scheme.
30
Experimental Result (cont.)
Fig.3 shows the simulation results of the buffer size during the coding process.
The X-axis is the picture index and Y-axis denotes the relative buffer size minus the standard
buffer threshold.
Buffer occupancy for fixed QP, HM9.1 rate control and RCQA rate control.
31
(a): B : Cactus, target bit rate 105.423mbps, initial QP = 22. (b): C : RaceHorsesC, target bit rate 5.106Mbps, initial QP = 32.
Outline
1.
Introduction
2.
Related Work (Rate distortion models for rate control)
3.
Proposed R-D model for HEVC
4.
QP calculation for rate control
5.
Experimental Result
6.
Conclusion
32
Conclusion
Pros :
1.
This paper has proposed a linear distortion-quantization model, an exponential rate
quantization model and a linear rate complexity model, which are integrated into a novel
Rate-Complexity-QP (RCQ) model for HEVC intra-picture rate control.
2.
Experimental results show that RCQA outperforms the HM 9.1 RC and the normal HM 9.1
fixed QP (FIXQP) rate control schemes.
Cons :
1.
The proposed method only apply rate control on frame level.
2.
The proposed method only deals with “encoderintra-main.cfg” configuration.
33