CN106210747A - A kind of low-complexity video coding method based on quaternary tree probabilistic forecasting - Google Patents

Abstract

The invention belongs to video compression coding field, disclose a kind of low-complexity video coding method based on quaternary tree probabilistic forecasting, based on decoder internal parameter, root node and the tree degree of depth to coded frame quaternary tree are predicted, and periodically the root node of prediction quaternary tree is updated with the tree degree of depth, it is ensured that the precision of prediction of next code frame.Using technical scheme, all can realize the scramble time in various degree in full frame, under the configuration of the lossy compression method such as low delay, Stochastic accessing saves, and code check increases less, rebuilds image subjective quality preferable.

Description

A kind of low-complexity video coding method based on quaternary tree probabilistic forecasting

Technical field

The invention belongs to video compression coding field, particularly relate to a kind of low complex degree based on quaternary tree probabilistic forecasting and regard Frequency coding method.

Background technology

At present, efficient video coding HEVC (High efficiency video coding, HEVC) has become limited depositing Storage resource, with under the network bandwidth, stores and transmits the only selection of high-quality, high-resolution video, but researchers are for efficiently The exploration of video encoding standard is still not stopped at.In order to obtain higher compression performance, HEVC standard have employed increasingly complex Coding techniques.Therefore, encoder complexity is optimized to the target of the always unremitting pursuit of researchers.Low at numerous HEVC In the research work of complexity coding, the most remarkable to the achievement in research of pattern partition problem in predictive coding.These methods exist The most all effectively reduce the amount of calculation during quad-tree partition.But HEVC often to carry out a quad-tree partition equal Travel through the predicting unit of all of pattern.Therefore, if can directly the quad-tree partition process in HEVC be simplified, coding The computation complexity of device end will significantly reduce.Current existing quaternary tree optimized algorithm is mostly by skipping unnecessary pre-depth measurement Degree realizes low encoding complexity, have ignored in a sense different quaternary tree root node also can affect encoder complexity this Key factor.Therefore, the performance towards the HEVC low encoding complexity method of quad-tree structure optimization still has room for promotion.

Summary of the invention

It is an object of the invention to provide a kind of low-complexity video coding method based on quaternary tree probabilistic forecasting, keeping On the premise of video reconstruction picture quality, save the scramble time, improve HEVC code efficiency.

For solving the problems referred to above, the present invention adopts the following technical scheme that:

A kind of low-complexity video coding method based on quaternary tree probabilistic forecasting comprises the following steps:

The coding quaternary tree that step S1, employing are preset is to all volumes in first image sets (Group of picture, GOP) Code frame encodes, and obtains optimum code unit (coding unit, CU) the distribution probability P of every two field picture in initial GOP；

Step S2, obtain prediction quad-tree structure F of encoded GOP by quaternary tree distribution probability P modeling_esta.；

Step S3, GOP to be encoded is carried out quad-tree structure prediction, F_pred.Represent each two field picture in GOP to be encoded Quad-tree structure, F_pred.It is calculated as follows formula:

Wherein, δ (CU, Frame_n) representing whether the coding quadtree's node of n-th frame image in GOP to be encoded exists, δ counts Calculate such as following formula:

$\mathrm{\&delta;}=\left\{\begin{array}{c}0,P<\mathrm{\&sigma;}\\ 1,P\&GreaterEqual;\mathrm{\&sigma;}\end{array}\right.,$

Wherein, whether 1, the 0} node representing coding quaternary tree respectively exists, and σ is empirical value,

As δ (CU, Frame in formula (1)_nDuring)=0, represent that in prediction quaternary tree, this node does not exists；As δ (CU, Frame_n) When=1, it was predicted that the root node size of quaternary tree is defined as δ (CU, Frame_nThe full-size of CU in)；The tree degree of depth is defined as log₂ (Root node)-log₂(leaf node)+1, wherein, Root node represent root node size, leaf node represent δ (CU, Frame_nThe minimum dimension of CU in)；

Step S4, judge that by update cycle T current encoded frame, the need of carrying out model renewal, if need to update, enters Enter step S5, otherwise enter step S6；

Step S5, prediction quaternary tree distribution probability model modification to current GOP, with the quad-tree structure conduct after updating Current or the prediction quaternary tree of follow-up GOP, is calculated as follows formula:

${\tilde{F}}_{esta.}^n=\left\{\begin{array}{cc}{F}_{esta.}^{n-1},& Z\left(F_{esta.}^n\right)=Z\left(F_{esta.}^{n-1}\right)\\ \mathrm{\&rho;}\&CenterDot;F_{esta.}^{n-1}+(1-\mathrm{\&rho;})\&CenterDot;F_{esta.}^n,& Z\left(F_{esta.}^n\right)\&NotEqual;Z\left(F_{esta.}^{n-1}\right)\end{array}\right.$

Wherein,Represent the prediction quad-tree structure after updating,Represent current GOP's it is predicted that quaternary tree knot Structure,Represent that previous GOP's it is predicted that quad-tree structure, Z () represent number and the position of neutral element in F, ρ ∈ [0, 1] it is used for balancing more new capability；WillSubstitutePerform step S3, obtain predicting that the root node size of quaternary tree and tree are deep Degree, then performs step S6；

Step S6, by calculated root node size and tree the degree of depth constitute prediction quaternary tree, from prediction quaternary tree root joint Point begins stepping through all tree nodes, until the tree degree of depth reaches the maximum of prediction, simultaneously by prediction quaternary tree to ought be before and after In continuous GOP, all coded frame carry out node traverses, and percent of pass distortion cost is calculated optimum CU size division result；

Step S7, judge whether current GOP is last GOP, if then using prediction quaternary tree complete coding and tie Bundle, otherwise returns step S4.

As preferably, step S1 particularly as follows: using a tree degree of depth is 4, the coding of root node a size of 64 × 64 pixel All coded frame in first image sets (Group of picture, GOP) are encoded by quaternary tree, obtain in initial GOP every Optimum code unit (coding unit, CU) the distribution probability P of two field picture；

$P=\frac{N\left(CU\right)}{N\left({\mathrm{CU}}_{64}\right)+N\left({\mathrm{CU}}_{32}\right)+N\left({\mathrm{CU}}_{16}\right)+N\left({\mathrm{CU}}_{08}\right)}$

Wherein, CU₆₄、CU₃₂、CU₁₆、CU₀₈Represent respectively a size of 64 × 64 pixels, 32 × 32 pixels, 16 × 16 pixels, 8 The CU of × 8 pixels, N (CU) represent the number of times that a certain size CU is selected.

As preferably, F in step S2_esta.It is calculated as follows formula:

Wherein, P (CU, Frame_n) represent the optimum CU distribution probability of n-th frame image in encoded GOP.

As preferably, in step S4, update cycle T is calculated as follows formula:

$T=\left\{\begin{array}{cc}N,& GOPSize<FrameRate\\ GOPSize,& GOPSize\&GreaterEqual;FrameRate\end{array}\right.$

Wherein, GOPSize represents that GOP length, Frame Rate represent frame per second, within during N presentation code one second, comprises how many Individual integral multiple GOP, N is defined as follows:

$N=GOPSize\&CenterDot;floor\left(\frac{FrameRate}{GOPSize}\right)$

Wherein, floor rounds under representing.

The invention has the beneficial effects as follows, quaternary tree distribution probability forecasting mechanism with reference to different gop structures, to coded frame Root node and the degree of depth of quaternary tree are predicted；Forecast model updates reference video content parameters, periodically predicts new quaternary tree Structure, it is ensured that next code frame precision of prediction.Low-complexity video coding algorithm proposed by the invention effectively prevent completely Quaternary tree travels through, and has broken the pre-geodesic structure of quaternary tree in traditional HEVC standard, reduces encoder and obtains optimum CU division size The calculating time, exchanged the saving of a large amount of scramble time for sacrificing a small amount of Y-PSNR, fundamentally improve HEVC The execution efficiency of encoder prediction coded portion.

Accompanying drawing explanation

Fig. 1 is the method for video coding flow chart of the present invention.

Detailed description of the invention

The present invention is described in detail with detailed description of the invention below in conjunction with the accompanying drawings.

In order to avoid HEVC quadtree's node travels through calculating entirely, the invention provides a kind of based on quaternary tree probabilistic forecasting Low-complexity video coding method, flow process is as it is shown in figure 1, specifically follow the steps below:

The first step, using a tree degree of depth is 4, and the coding quaternary tree of root node a size of 64 × 64 pixel is to first image In group (Group of picture, GOP), all coded frame encode, and obtain the optimum code of every two field picture in initial GOP Unit (coding unit, CU) distribution probability P；

$P=\frac{N\left(CU\right)}{N\left({\mathrm{CU}}_{64}\right)+N\left({\mathrm{CU}}_{32}\right)+N\left({\mathrm{CU}}_{16}\right)+N\left({\mathrm{CU}}_{08}\right)}---\left(1\right)$

Wherein, CU₆₄, CU₃₂, CU₁₆, CU₀₈Represent a size of 64 × 64 pixels respectively, 32 × 32 pixels, 16 × 16 pixels, 8 The CU of × 8 pixels, N (CU) represent the number of times that a certain size CU is selected.

Second step, obtains prediction quad-tree structure F of encoded GOP by quaternary tree distribution probability P modeling_esta., F_esta. It is calculated as follows formula:

Wherein, P (CU, Frame_n) represent the optimum CU distribution probability of n-th frame image in encoded GOP.

3rd step, carries out quad-tree structure prediction, F to GOP to be encoded_pred.Represent each two field picture in GOP to be encoded Quad-tree structure, F_pred.It is calculated as follows formula:

Wherein, δ (CU, Frame_n) representing whether the coding quadtree's node of n-th frame image in GOP to be encoded exists, δ counts Calculate such as following formula:

$\mathrm{\&delta;}=\left\{\begin{array}{c}0,P<\mathrm{\&sigma;}\\ 1,P\&GreaterEqual;\mathrm{\&sigma;}\end{array}\right.---\left(4\right)$

Wherein, { whether 1, the 0} node representing coding quaternary tree respectively exists, and σ is empirical value.

As δ (CU, Frame in formula (3)_nDuring)=0, represent that in prediction quaternary tree, this node does not exists；As δ (CU, Frame_n) When=1, it was predicted that the root node size of quaternary tree is defined as δ (CU, Frame_nThe full-size of CU in)；The tree degree of depth is defined as log₂ (Root node)-log₂(leaf node)+1, wherein, Root node represent root node size, leaf node represent δ (CU, Frame_nThe minimum dimension of CU in).

By update cycle T, 4th step, judges that current encoded frame, the need of carrying out model renewal, if need to update, enters Entering the 5th step, otherwise enter the 6th step, T is calculated as follows formula:

$T=\left\{\begin{array}{cc}N,& GOPSize<FrameRate\\ GOPSize,& GOPSize\&GreaterEqual;FrameRate\end{array}\right.---\left(5\right)$

Wherein, GOPSize represents that GOP length, Frame Rate represent frame per second, within during N presentation code one second, comprises how many Individual integral multiple GOP, N is defined as follows:

$N=GOPSize\&CenterDot;floor\left(\frac{FrameRate}{GOPSize}\right)---\left(6\right)$

Wherein, floor rounds under representing.

5th step, the prediction quaternary tree distribution probability model modification to current GOP, with the quad-tree structure conduct after renewal Current or the prediction quaternary tree of follow-up GOP, is calculated as follows formula:

${\tilde{F}}_{esta.}^n=\left\{\begin{array}{cc}{F}_{esta.}^{n-1},& Z\left(F_{esta.}^n\right)=Z\left(F_{esta.}^{n-1}\right)\\ \mathrm{\&rho;}\&CenterDot;F_{esta.}^{n-1}+(1-\mathrm{\&rho;})\&CenterDot;F_{esta.}^n,& Z\left(F_{esta.}^n\right)\&NotEqual;Z\left(F_{esta.}^{n-1}\right)\end{array}\right.---\left(7\right)$

Wherein,Represent the prediction quad-tree structure after updating,Represent current GOP's it is predicted that quaternary tree knot Structure,Represent that previous GOP's it is predicted that quad-tree structure, Z () represent number and the position of neutral element in F, ρ ∈ [0, 1] it is used for balancing more new capability.

WillSubstitutePerform the 3rd step, obtain predicting the root node size of quaternary tree and the tree degree of depth, then perform 6th step.

6th step, is constituted prediction quaternary tree by calculated root node size and the tree degree of depth, from prediction quaternary tree root joint Point begins stepping through all tree nodes, until the tree degree of depth reaches the maximum of prediction.With prediction quaternary tree in current and follow-up GOP All coded frame carry out node traverses, and percent of pass distortion cost is calculated optimum CU size division result；

7th step, it is judged that whether current GOP is last GOP, if then using prediction quaternary tree complete coding and tie Bundle, otherwise returns the 4th step.

Present invention low-complexity video coding based on quaternary tree probabilistic forecasting method is entered with international standard algorithm HM15.0 Row contrast, usesDelta peak signal-to noise ratio (BDPSNR) andDelta bit rate (BDBR) weighs coding quality gain.Scramble time saving TS represents, Encoding Statistical result can be shown in Table 1, table 2 and table 3, wherein, the official that Class A-F formulates tissue for International video coding standard and provides surveys Examination sequence.

This paper algorithm and HM15.0 performance comparison result under the conditions of in table 1 full frame

This paper algorithm and HM15.0 performance comparison result under the conditions of table 2 low latency

This paper algorithm and HM15.0 performance comparison result under the conditions of table 3 Stochastic accessing

Statistical result shows, compared with a new generation international video standard HEVC (HM15.0), the present invention is based on quaternary tree The low-complexity video coding method of probabilistic forecasting presents outstanding low encoding complexity performance.For polymorphic type, difference point The cycle tests of resolution, the scramble time of the highest saving 28%, under reconstructed image quality loses relatively minor premise, coding is multiple Miscellaneous degree is optimized, and improves the compression efficiency of HEVC encoder on the whole.

Above example is only the exemplary embodiment of the present invention, is not used in the restriction present invention, protection scope of the present invention It is defined by the claims.The present invention can be made respectively in the essence of the present invention and protection domain by those skilled in the art Planting amendment or equivalent, this amendment or equivalent also should be regarded as being within the scope of the present invention.

Claims (4)

1. a low-complexity video coding method based on quaternary tree probabilistic forecasting, it is characterised in that comprise the following steps:

The coding quaternary tree that step S1, employing are preset is to all coded frame in first image sets (Group of picture, GOP) Encode, obtain optimum code unit (coding unit, CU) the distribution probability P of every two field picture in initial GOP；

Step S2, obtain prediction quad-tree structure F of encoded GOP by quaternary tree distribution probability P modeling_esta.；

Step S3, GOP to be encoded is carried out quad-tree structure prediction, F_pred.Represent four forks of each two field picture in GOP to be encoded Tree construction, F_pred.It is calculated as follows formula:

Wherein, δ (CU, Frame_n) representing whether the coding quadtree's node of n-th frame image in GOP to be encoded exists, δ calculates such as Following formula:

$\mathrm{\&delta;}=\left\{\begin{array}{c}0,P<\mathrm{\&sigma;}\\ 1,P\&GreaterEqual;\mathrm{\&sigma;}\end{array}\right.,$

Wherein, whether 1, the 0} node representing coding quaternary tree respectively exists, and σ is empirical value,

As δ (CU, Frame in formula (1)_nDuring)=0, represent that in prediction quaternary tree, this node does not exists；As δ (CU, Frame_n)=1 Time, it was predicted that the root node size of quaternary tree is defined as δ (CU, Frame_nThe full-size of CU in)；The tree degree of depth is defined as log₂ (Root node)-log₂(leaf node)+1, wherein, Root node represent root node size, leaf node represent δ (CU, Frame_nThe minimum dimension of CU in)；

Step S4, judge that by update cycle T current encoded frame, the need of carrying out model renewal, if need to update, enters step Rapid S5, otherwise enters step S6；

Step S5, prediction quaternary tree distribution probability model modification to current GOP, by the quad-tree structure after updating as currently Or the prediction quaternary tree of follow-up GOP, it is calculated as follows formula:

${\tilde{F}}_{esta.}^n=\left\{\begin{array}{cc}{F}_{esta.}^{n-1},& Z\left(F_{esta.}^n\right)=Z\left(F_{esta.}^{n-1}\right)\\ \mathrm{\&rho;}\&CenterDot;F_{esta.}^{n-1}+\left(1-\mathrm{\&rho;}\right)\&CenterDot;F_{esta.}^n,& Z\left(F_{esta.}^n\right)\&NotEqual;Z\left(F_{esta.}^{n-1}\right)\end{array}\right.$

Wherein,Represent the prediction quad-tree structure after updating,Represent current GOP's it is predicted that quad-tree structure,Representing previous GOP's it is predicted that quad-tree structure, Z () represents number and the position of neutral element in F, and ρ ∈ [0,1] uses In balancing more new capability；WillSubstitutePerform step S3, obtain predicting the root node size of quaternary tree and the tree degree of depth, so Rear execution step S6；

Step S6, by calculated root node size and tree the degree of depth constitute prediction quaternary tree, from prediction quaternary tree root node open Begin to travel through all tree nodes, until the tree degree of depth reaches the maximum of prediction, simultaneously by prediction quaternary tree to current and follow-up GOP In all coded frame carry out node traverses, percent of pass distortion cost is calculated optimum CU size division result；

Step S7, judge whether current GOP is last GOP, if then using prediction quaternary tree complete coding and terminate, no Then return step S4.

2. low-complexity video coding method based on quaternary tree probabilistic forecasting as claimed in claim 1, it is characterised in that step Rapid S1 is particularly as follows: using a tree degree of depth is 4, and the coding quaternary tree of root node a size of 64 × 64 pixel is to first image sets In (Group of picture, GOP), all coded frame encode, and obtain the optimum code list of every two field picture in initial GOP Unit (coding unit, CU) distribution probability P；

$P=\frac{N\left(CU\right)}{N\left({\mathrm{CU}}_{64}\right)+N\left({\mathrm{CU}}_{32}\right)+N\left({\mathrm{CU}}_{16}\right)+N\left({\mathrm{CU}}_{08}\right)}$

Wherein, CU₆₄、CU₃₂、CU₁₆、CU₀₈Represent respectively a size of 64 × 64 pixels, 32 × 32 pixels, 16 × 16 pixels, 8 × 8 The CU of pixel, N (CU) represent the number of times that a certain size CU is selected.

3. low-complexity video coding method based on quaternary tree probabilistic forecasting as claimed in claim 1, it is characterised in that step F in rapid S2_esta.It is calculated as follows formula:

Wherein, P (CU, Frame_n) represent the optimum CU distribution probability of n-th frame image in encoded GOP.

4. low-complexity video coding method based on quaternary tree probabilistic forecasting as claimed in claim 1, it is characterised in that step Update cycle T in rapid S4 is calculated as follows formula:

$T=\left\{\begin{array}{cc}N,& GOPSize<FrameRate\\ GOPSize,& GOPSize\&GreaterEqual;FrameRate\end{array}\right.$

Wherein, GOPSize represents that GOP length, Frame Rate represent frame per second, within during N presentation code one second, comprises how many individual whole Several times GOP, N is defined as follows:

$N=GOPSize\&CenterDot;floor\left(\frac{FrameRate}{GOPSize}\right)$

Wherein, floor rounds under representing.