Rate-GOP Based Rate Control for HEVC SHANSHE WANG, SIWEI MA, SHIQI WANG, DEBIN ZHAO, AND WEN GAO IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING,

Rate-GOP Based Rate Control for HEVCSHANSHE WANG, S IWEI MA, SHIQI WANG, DEBIN ZHAO, AND WEN GAO

IEEE JOURNAL OF SELECTED TOPICS IN S IGNAL PROCESSING, VOL. 7 , NO. 6 , DECEMBER 2013

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Introduction In HEVC, many new coding tools are adopted to improve the coding performance, and they bring great challenges to establish an accurate R-Q model and bit allocation scheme in rate control for HEVC.

To establish a more accurate R-Q model and a more efficient bit allocation scheme, the rate control algorithm must consider the R-D characteristics.

How to use the inter-frame dependency to improve the rate control performance is also a difficult problem which may not be addressed well in the past.

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Hierarchical Rate Control Structure And Problem Formulation

1. Quad-tree Coding Structure And Reference Picture Set (RPS)

2. Proposed Hierarchical Rate Control Structure

3. Problem Formulation

Quad-tree Coding Structure1. CU - LCU size = 64, maximum depth = 42. PU - NxN partition mode is allowed only

when the CU size is equal to minimum size3. TU – TU is defined to represent the basic

unit for transform.• Inter : the size of TU is independent

with the size of PU• Intra : the size of TU cannot exceed the

size of PU

Reference Picture Set (RPS) For RPS determination, it is composed of one frame with nearest temporal distance and three frames with lowest QP in Decoded Picture Buffer.

Proposed Hierarchical Rate Control Structure From Tables I and II and (1) and (2), it can be concluded that different frame in a Rate-GOP has different influence on the following frames.

Generally, the frame with QP+1 with relatively high reconstructed video quality will be referred more than once and thus plays a crucial role, which is called the key frame. (POC %8==0)

*The four frames with the maximum depth will not be used as a reference frame, so these are not included in the proposed structure.

Problem Formulation Based on the proposed structure, the optimization problem can be stated as follows:

Or :

Q and k is the number of frame in a GOP

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Inter-frame Dependency Based Distortion Model and Rate Control

Inter-frame prediction results in the distortion dependency and bit rate dependency between a coding frame and its reference frames. (so we analyze follow issues.) Inter-Frame Dependency Based Distortion Model Inter-Frame Dependency Based Rate Model

Distortion Model The distortion of the coding frame has great dependency with its reference frames. This kind of distortion dependency is one of the main factors affecting the bit allocation scheme for rate control.

Rate Model The bit rate dependency between the coding frame and its reference frame is also investigated aiming at a deeper exploration for R-D relations.

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

R-Q Model For HEVC Generally speaking, R-Q model has great correlation with the distribution of the residual information.

It was stated in [25] that a single Laplacian distribution is not accurate enough to capture the distribution due to the quad-tree coding structure. So a mixed distribution is applied below :

It has been proved in [12] that an accurate linear relation in ρ–domain between the bits and the number of non-zero transformed coefficients holds firmly via the theoretical derivation for a Laplacian distribution as follows:

[12] Z. He and S. Mitra, “Optimum bit allocation and accurate rate control for video coding via -domain source modeling,” IEEE Trans. CircuitsSyst. Video Technol., vol. 12, no. 10, pp. 840–849, Oct. 2002.[25] B. M. Kim, “Modeling rates and distortion based on a mixture of laplacian distribution for inter-predicted residuals in quadtree coding of HEVC,” IEEE Signal Process. Lett., vol. 18, no. 10, pp. 571–574, Oct. 2011.

Nonzero coefficients of CUs in j-th depth, like P.6 mentioned.

R-Q Model For HEVC According to the quantization in HEVC, for a certain j-th depth, the distortion can be calculated by (12) and the percentage of zero coefficients can be computed by (13).

In [12],

R-Q Model For HEVC Assume and denote the number of nonzero transformed coefficients and the number of pixels of j-th depth in frame respectively. Combine with (13), we can get

Taylor expansion

D-Q Model For HEVC As illustrated in the above subsection, the linear relation between the bit rate and non-zero transformed coefficients still holds firmly in HEVC under the mixed Laplacian distribution.

For the first layer, the distortion can be represented by the following as stated in [12].

For the second layer and third layer, we take into account the inter-frame dependency in (7) to model the distortion as follows.

D-Q Model For HEVC Subsequently, the distortion of a Rate-GOP can be presented as follows:

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Proposed Rate Control Algorithm Based on RDO

1. Rate-GOP Level Bit Allocation

2. Optimal Hierarchical Bit Allocation

3. Header Bit Prediction

4. Proposed Rate-GOP Based Rate Control Scheme

Rate-GOP Level Bit Allocation For GOP level, the target bits are allocated according to the bandwidth after coding the previous GOP.

In order to get smoother bit rate, the really allocated bits is modified as

Optimal Hierarchical Bit Allocation For a Rate-GOP, given the bit budget allocated as in the previous subsection, the optimization problem is to choose appropriate QPs to minimize the total distortion of the Rate-GOP.

Combine with previous rate model in (21), distortion model in (26), (30) can be modified as

(21)

(26)

Optimal Hierarchical Bit Allocation By computing partial derivation with respect to ,

From (18)

Combine (20) and (34)

By a Taylor expansion of (33) and (35)

Optimal Hierarchical Bit Allocation Taking (36) into (33), the bit rate for -th frame can be obtain by

Once is determined by the correlation among adjacent frames, and QP can be obtained by (20)

Optimal Hierarchical Bit Allocation To avoid the complicated calculation of in (39), is estimated from that of the previous GOPs. Like (40)

After encoding one Rate-GOP, if , then the weighting factors are calculated and utilized for the next Rate-GOP.

is set to 0.5,0.3,0.2 initially

-th layer in-th GOP

Proposed Rate-GOP Based Rate Control Scheme

Header Bits Allocation Inspired by [10], in the experiments it was found that the header bits has strong linear relationship with the number of PU and the value of MVD, as shown in Fig. 8.

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Experiment Results The experiments are conducted on HEVC test model HM8.0.

The test sequences and 6 testing configurations are detailed in [26].

A. Control accuracy

B. R-D performance

C. Complexity analysis

[26] F. Bossen, “Common test conditions and software reference configurations,” in Document: JCTVC-E700

Control Accuracy The accuracy of bit mismatch is investigated in terms of mismatch error as follows.

R-D Performance The R-D performances in terms of BD-PSNR and BD-Rate [27] results are presented for comparison.

Firstly, experiments are conducted with the proposed bit allocation scheme but without using the proposed R-Q model.

R-D Performance Secondly, experiments are implemented with both the bit allocation scheme and the proposed R-Q model.

As illustrated in Tables V–VII, the R-D performances in terms of BD-PSNR and BD-Rate results are presented for comparison with URQ algorithm and R-λ algorithm.

R-D Performance

Complexity Analysis The parameters and in (20) are obtained by linear regression.

The parameter in (20) are predicted from the previous coded frames as in (38) to avoid the complicated calculations.

The parameters α and β in (43) are also obtained by linear regression.

The parameter in (39) is predicted from the previous Rate-GOPs.

Overview Introduction Hierarchical Rate Control Structure of Rate-GOP And Problem Formulation Inter-frame Dependency Based Distortion Model And Rate Control R-Q Model And D-Q Model For HEVC Proposed Rate Control Algorithm Based on RDO Experiment Results Conclusions

Conclusions The proposed rate control algorithm considered the Rate-GOP as a rate control unit and provided a hierarchical structure for Rate-GOP.

The inter-frame dependency based distortion model and bit rate model for Rate-GOP are derived.

Based on these models and a mixed Laplacian distribution of residual information, a new -domain Rate-GOP based rate control is proposed.

The experimental results demonstrated that the proposed algorithm can achieve much better R-D performance than the two state of the art rate control algorithms for HEVC.