CN106131546A - A kind of method determining that HEVC merges and skips coding mode in advance - Google Patents

Abstract

The invention discloses a kind of method determining that HEVC merges and skips coding mode in advance, the method test coding unit with Skip and Merge pattern-coding after, if the encoding block residual absolute value sum that preferably pattern is corresponding is less than a default threshold value, then ignore Inter and the Intra pattern that HEVC standard is follow-up, and predicting unit partition process, the coding mode determining encoding block in advance is the relatively the superior in Skip and Merge, it is thus possible to exchange the very big reduction of HEVC encoder computation complexity for less performance loss.

Description

A kind of method determining that HEVC merges and skips coding mode in advance

Technical field

The invention belongs to video image compression coding technical field, be specifically related to one and determine that HEVC merges and skips in advance The method of coding mode.

Background technology

High efficiency Video coding (High Efficiency Video Coding, HEVC) is with code tree unit (Coding Tree Unit, CTU) be basic processing unit, CTU can only comprise a coding unit (Coding Unit, CU) or Person is divided into downwards four equal-sized CU with the form of quaternary tree.Further, each size be 2N × 2N (N can be 8, 16,32) CU can continue to be divided into the CU that four sizes are all N × N.In the CTU partition structure of whole quaternary tree form, CU as leaf node gives the predictive mode of certain square area in CTU, it may be assumed that (Intra), interframe in frame (Inter), skip over (Skip) or merge (Merge) pattern.Depend on predictive mode, CU can be divided into one, two or Four predicting unit (Prediction Unit, PU).To intraframe coding, there are two kinds of optional PU dividing mode；Interframe is compiled Code, has eight kinds of optional PU dividing mode；It is combined pattern, then a kind of PU dividing mode.

Merging patterns are a kind of patterns that HEVC is newly introduced, when a predicting unit is encoded into merging patterns, then it and Share identical movable information with its space or temporally adjacent block, spend in encoded motion information side it is thus possible to efficiently reduce Code check cost needed for face.Skip pattern is regarded as a special case of Merge pattern, in such a mode, only Skip by HEVC Mark and the corresponding block index that merges are encoded.

HEVC encoder determines CU partition structure, predictive mode and PU dividing mode in the way of exhaustive, compiles for interframe Code CU, encoder is tested Skip/Merge pattern successively by default order, is not made PU and divide, up and down or the situation such as left and right division Under rate distortion costs, determine above-mentioned project for target minimizing rate distortion costs function.

HEVC can obtain than H.264 etc. the previous more preferable code efficiency of video compression standard and performance, and main cause exists Have employed the coding unit of above-mentioned complexity, predicting unit and converter unit partition structure in it, and solve in the way of exhaustive Rate distortion function optimal solution when different demarcation structure and coding mode.But, " test all of probability, select Excellent " such a coding framework has been simultaneously introduced the biggest calculation cost.According to aforesaid testing sequence, for certain Eventually with the CU of Skip or Merge pattern-coding, if can after the rate distortion costs under completing Skip/Merge pattern calculates i.e. The Skip/Merge pattern forced coding pattern as this coding unit can be judged ahead of time, follow-up inter-frame mode, frame mode Can omit with CU recurrence partition process etc., thus greatly reduce the computation complexity of encoder.

In order to reduce the calculation cost that encoder spends on mode decision, document [1] " Early Merge Mode Decision Based on Motion Estimation and Hierarchical Depth Correlation for HEVC " (Pan etc., IEEE Trans.Broadcasting, 2014.2) disclose a kind of coding mould determining coding unit ahead of time Formula is the method for merging patterns, and coding unit is divided into two big classes: the level degree of depth be 0 root coding unit and level more than 0 Sub-coding unit.To root coding unit, decide whether to determine that it is ahead of time according to stand growth model information and motion estimation information Merging patterns, then utilize the dependency between root coding unit and sub-coding unit to determine merging mould ahead of time for sub-coding unit Whether formula is optimization model.Document [2] " A Novel Fast CU Encoding Scheme Based on Spatiotemporal Encoding Parameters for HEVC Inter Coding " (Ahn etc., IEEE Trans.Circuits Syst.Video Technol., 2015.3) disclosing one determines that coding unit is Skip mould ahead of time Formula and quickly determine that current coded unit is the need of the method continuing to divide downwards.

Summary of the invention

The invention provides a kind of method determining that HEVC merges and skips coding mode in advance, in order to reduce HEVC coding Device spends in the calculation cost determined on the partition structure of code tree unit and coding mode, thus reduces the meter of HEVC encoder Calculate complexity.

A kind of method determining that HEVC merges and skips coding mode in advance is as follows:

For current coded unit, determine its optimal candidate kinematic parameter under merging patterns, and record by calculating Rate distortion costs J that this optimal candidate kinematic parameter is corresponding_mergeAnd prediction residual absolute value sum S_merge；In like manner by meter Calculate and determine its optimal candidate kinematic parameter under skip mode, and record the rate distortion generation that this optimal candidate kinematic parameter is corresponding Valency J_skipAnd prediction residual absolute value sum S_skip；

Judge following condition: J_skip＜ J_mergeAnd S_skip＜ T_skip, T_skipThreshold value for current QP and N of correspondence；

If setting up, judge the most in advance the coding mode of current coded unit as skip mode, and it is follow-up pre-to make it skip Survey the rate distortion costs test of dividing elements and other coding mode；

If being false, then determine whether following condition: S_merge＜ T_merge, T_mergeThreshold value for current QP and N of correspondence； QP is quantization parameter, and N is the size of coding unit；

Judge that the coding mode of current coded unit, as merging patterns, and makes it skip follow-up prediction the most in advance The rate distortion costs test of dividing elements and other coding mode；

If it is not, then make current coded unit continue HEVC coding standard specify that follow-up predicting unit divides and other is compiled The rate distortion costs test of pattern.

The process that implements of described step (1) is: first determine current coded unit candidate's fortune under merging patterns Dynamic parameter sets, this candidate motion parameter set close by with current coded unit space or temporally adjacent and encoded respectively predict list The kinematic parameter composition of unit；Then under merging patterns, calculate motion compensated image corresponding to each candidate motion parameter and prediction is residual Difference, and calculate, by prediction residual being carried out two-dimension discrete cosine transform, the rate that each candidate motion parameter is corresponding with quantification treatment Distortion cost and prediction residual absolute value sum, with minimize rate distortion costs for criterion from candidate motion parameter set close choose Optimal candidate motion parameter, and record rate distortion costs J that this optimal candidate kinematic parameter is corresponding_mergeAnd prediction residual is exhausted To value sum S_merge；

In like manner determining that current coded unit candidate motion parameter set under skip mode is closed, this candidate motion parameter set is closed It is made up of the kinematic parameter with current coded unit space or temporally adjacent and encoded each predicting unit；Then mould is being skipped Calculate motion compensated image corresponding to each candidate motion parameter and prediction residual under formula, and by prediction residual is carried out two dimension from Dissipate cosine transform and calculate rate distortion costs corresponding to each candidate motion parameter and prediction residual absolute value sum with quantification treatment, From candidate motion parameter set is closed, choose optimal candidate motion parameter for criterion minimizing rate distortion costs, and record this Rate distortion costs J that good candidate motion parameter is corresponding_skipAnd prediction residual absolute value sum S_skip。

Described threshold value T_skipDetermination method as follows:

First with the quantization parameter QP set, HEVC encoder is used to encode, the test video of input for institute Have a coding unit of N × N size, record wherein with prediction residual absolute value corresponding to the coding unit of skip mode coding it With, form observation sample set { K+}；The prediction residual that coding unit that record wherein encodes with non-skip mode is corresponding is absolute Value sum, forms observation sample set { K-}；

Take and meet prediction residual absolute value sum S under following condition premise_iMinima as threshold value T_skip；

p(S_i| K-) P (K-) ＞ p (S_i|K+)P(K+)

Wherein: S_iFor the prediction residual absolute value sum that i-th coding unit is corresponding, p (S_i| K+) represent S_iRelative to sight Survey sample set { distribution density of K+}, p (S_i| K-) represent S_iRelative to observation sample set { distribution density of K-}, P (K+) For the ratio that the coding unit encoded with skip mode is shared in the coding unit of all N × N sizes, P (K-) is with non-jumping Crossing the coding unit of pattern-coding ratio shared in the coding unit of all N × N sizes and P (K-)=1-P (K+), i is Natural number and 1≤i≤m, m is the coding unit quantity of all N × N sizes；

According to above-mentioned, travel through all possible quantization parameter QP and size N, obtain each group of QP with N determined corresponding Threshold value T_skip。

Described threshold value T_mergeDetermination method as follows:

First with the quantization parameter QP set, HEVC encoder is used to encode, the test video of input for institute Have a coding unit of N × N size, record wherein with prediction residual absolute value corresponding to the coding unit of merging patterns coding it With, form observation sample set { M+}；Record is wherein absolute with prediction residual that the coding unit of unconsolidated pattern-coding is corresponding Value sum, forms observation sample set { M-}；

Take and meet prediction residual absolute value sum S under following condition premise_iMinima as threshold value T_merge；

p(S_i| M-) P (M-) ＞ p (S_i|M+)P(M+)

Wherein: S_iFor the prediction residual absolute value sum that i-th coding unit is corresponding, p (S_i| M+) represent S_iRelative to sight Survey sample set { distribution density of M+}, p (S_i| M-) represent S_iRelative to observation sample set { distribution density of M-}, P (M+) For the ratio that the coding unit encoded with merging patterns is shared in the coding unit of all N × N sizes, P (M-) is with non-conjunction And the ratio shared in the coding unit of all N × N sizes of the coding unit of pattern-coding and P (M-)=1-P (M+), i is Natural number and 1≤i≤m, m is the coding unit quantity of all N × N sizes；

According to above-mentioned, travel through all possible quantization parameter QP and size N, obtain each group of QP with N determined corresponding Threshold value T_merge。

Described prediction residual absolute value sum S_iRelative to the observation sample set { distribution density p (S of L}_i| L) calculating Method is as follows, wherein L=K+, K-, M+ or M-；

First, observation sample set { each separation A of L} is determined₀,A₁,…,A_m+1, wherein A₀=0, A_i-A_i-1=d, A_m+1 More than or equal to observation sample set { maximum in L} and less than this maximum+d；D be default group away from；

Then, distribution density p (S is calculated according to below equation_i| L):

$p\left(S_i\right|L)=\frac{C_i}{{\mathrm{\Σ}}_{i=1}^m{C}_i}$

Wherein: C_iFor observation sample set, { L} declines at interval [A_i,A_i+1) number of samples.

The integer that span is 0～51 of described quantization parameter QP, size N=8 of described coding unit, 16,32 or 64。

The present invention test coding unit with Skip and Merge pattern-coding after, if the volume that preferably pattern is corresponding Code block residual absolute value sum is less than a default threshold value, then ignore Inter and the Intra pattern that HEVC standard is follow-up, with And predicting unit partition process, determine that the coding mode of encoding block is the relatively the superior in Skip and Merge in advance, it is thus possible to relatively Few performance loss exchanges the very big reduction of HEVC encoder computation complexity for.

Accompanying drawing explanation

Fig. 1 is that the present invention determines that HEVC merges and skips the method flow schematic diagram of coding mode in advance.

Fig. 2 is the conditional probability density distribution schematic diagram of the prediction residual absolute value sum of 16 × 16 encoding blocks.

Detailed description of the invention

In order to more specifically describe the present invention, below in conjunction with the accompanying drawings and detailed description of the invention is to technical scheme It is described in detail.

May include that as it is shown in figure 1, the present invention determines that HEVC merges and skips the method flow of coding mode in advance

Step 101, for the coding unit of input, calculates current coded unit corresponding optimal candidate fortune under merging patterns The rate distortion costs of dynamic parameter, the residual absolute value sum of record coding unit.Specifically, the kinematic parameter of candidate respectively from In the predicting unit adjacent in room and time territory with current prediction unit, encoder calculates corresponding each Candidate Motion ginseng respectively The motion compensated image of number and prediction residual, carry out two-dimension discrete cosine transform and quantification treatment to prediction residual, be calculated as follows Rate distortion costs:

J=D+ λ R

Wherein: D is the coding distortion represented with the error between original image and reconstruction image, and R is to represent present encoding Bit number needed for block, λ is the multiplier determined by quantization parameter QP.Encoder to minimize rate distortion costs function J as criterion, Choosing optimal kinematic parameter from candidate collection, the rate distortion costs of note correspondence is the rate distortion costs under merging patterns J_merge, and the prediction residual absolute value sum after corresponding motion-compensated of record is S_merge。

Step 102, calculates current coded unit rate distortion costs J under SKIP pattern_skip, and calculate corresponding with record Prediction residual absolute value sum S_skip。

Step 103, it is judged that J_skip＜ J_merge？The most then perform step 104, otherwise go to step 105.

Step 104, it is judged that residual absolute value sum S in block_skipWhether less than default threshold value T_skip, S_skip＜ T_skip？If so, go to step 106, otherwise continue the condition test of step 105.

Step 105, it is judged that residual absolute value sum S in block_mergeWhether less than default threshold value T_merge, S_merge＜ T_merge？The most then go to step 107, otherwise turn 108.

Step 106, is judged to eligible S to by step 104_skip＜ T_skipCoding unit, skip follow-up PU and divide Test with the rate distortion costs of other coding mode, judge that the coding mode of this coding unit is as SKIP in advance.

Step 107, is judged to eligible S to by step 105_merge＜ T_mergeCoding unit, skip follow-up PU and draw Point and the rate distortion costs test of other coding mode, judge that the coding mode of this coding unit is as MERGE ahead of time.

Step 108, for not meeting step 104 and the encoding block of 105 conditions, carry out follow-up PU block divide and INTER, Rate distortion costs test under INTRA pattern, determines the rate distortion costs of optimal PU partition structure, coding mode and correspondence.

Threshold value T in said method, described in step 105_mergeDetermine with the method for study, specifically, the volume of N × N size Code unit, has a corresponding threshold value to the QP value that each is possible, and these threshold values are saved in one with N and QP value for index In form.

The method of a list item in computation sheet is: with corresponding QP value, uses the test to input of the HEVC encoder Video encodes, and the coding mode of the coding unit of N × N size regards as random event, and is set by the sample space of event For: {, with MERGE pattern-coding, CU is with non-MERGE pattern-coding for CU }.CU is with MERGE pattern-coding for event M+ for note, and CU is with non- MERGE pattern-coding is event M-.The block of those CU with MERGE pattern-coding it is readily obtained in the cataloged procedure of encoder The distribution of interior residual absolute value sum, and with in the block of those CU of non-MERGE pattern-coding residual absolute value sum point Cloth, remembers the conditional probability density that p (S | M+) and p (S | M-) is above-mentioned two class event S values respectively；Calculate S value relative to overall Conditional probability density is as follows:

First, to sample set, { observation sample in L} is ranked up, and determines the separation often organized away from d with default group A₀,A₁,...,A_K, wherein A₀=0, A_i-A_i-1=d, A_KAdd more than or equal to the maximum in sample set and less than this maximum Group away from and value；

{ L} declines at every class interval [A in statistical sample set_i,A_i+1) number of samples, be denoted as C_i；

It is calculated as follows the corresponding overall Density Distribution of sample:

$p\left(S_i\right|L)=C_i/\underset{i}{\Σ}{C}_i$

Wherein: L represents one in classification M+ and M-；Fig. 2 is shown, and a large amount of tests are regarded by HEVC reference model encoder The conditional probability density of the S value of 16 × 16 encoding blocks of frequency experiment gained.

By Bayesian formula, two states M+ of encoding block and the posterior probability of M-it is:

$P(M+|S)=\frac{p\left(S\right|M+)P(M+)}{P\left(S\right)}$

$P(M-|S)=\frac{p\left(S\right|M-)P(M-)}{P\left(S\right)}$

Wherein: P (M+) and P (M-) is the prior probability of state M+ and M-respectively, P (S) be one for normalized because of Son；P is calculated with the ratio that the coding unit of merging patterns coding is shared in the coding unit of all N × N sizes by statistics (M+) P (M-), it is calculated as follows:

P (M-)=1-P (M+)

Use Bayes decision rule based on minimal error rate: if following formula is set up, then prediction residual absolute value it It is MERGE with the coding mode decision of the encoding block for S.

P (S | M+) P (M+) ＞ p (S | M-) P (M-)

By reference model encoder, test video is tested N × N encoding block of gained with MERGE pattern and non-MERGE mould The conditional probability density of the S value of the encoding block of formula coding, and prior probability P (M+) and P (M-), take the minimum S meeting following formula Value is as threshold value T_merge。

p(S_i| M-) P (M-) ＞ p (S_i|M+)P(M+)

With T_mergeAsk for be similar to, threshold value T can be obtained_skip。

The span of quantization parameter QP is the integer of 0 to 51, and N can be 8,16,32 or 64.To each possible QP value, N Corresponding two threshold values T of the coding unit of × N size_skipAnd T_merge, these threshold values are stored in two forms, they respectively with N and QP is index, and one preserves corresponding T_skip, another preserves corresponding T_merge。

The above-mentioned description to embodiment is to be understood that for ease of those skilled in the art and apply the present invention. Above-described embodiment obviously easily can be made various amendment by person skilled in the art, and described herein typically Principle is applied in other embodiments without through performing creative labour.Therefore, the invention is not restricted to above-described embodiment, ability Field technique personnel should be in protection scope of the present invention according to the announcement of the present invention, the improvement made for the present invention and amendment Within.

Claims (6)

1. the method determining that HEVC merges and skips coding mode in advance, it is characterised in that:

For current coded unit, determine its optimal candidate kinematic parameter under merging patterns by calculating, and record this Rate distortion costs J that good candidate motion parameter is corresponding_mergeAnd prediction residual absolute value sum S_merge；In like manner by calculating really Its optimal candidate kinematic parameter under skip mode fixed, and record the rate distortion costs that this optimal candidate kinematic parameter is corresponding J_skipAnd prediction residual absolute value sum S_skip；

Judge following condition: J_skip＜ J_mergeAnd S_skip＜ T_skip, T_skipThreshold value for current QP and N of correspondence；

If setting up, judge that the coding mode of current coded unit, as skip mode, and makes it skip follow-up prediction list the most in advance Unit divides and the rate distortion costs test of other coding mode；

If being false, then determine whether following condition: S_merge＜ T_merge, T_mergeThreshold value for current QP and N of correspondence；QP is Quantization parameter, N is the size of coding unit；

Judge that the coding mode of current coded unit, as merging patterns, and makes it skip follow-up predicting unit the most in advance Divide and the rate distortion costs test of other coding mode；

If it is not, then make current coded unit continue HEVC coding standard specify that follow-up predicting unit divides and other coding mould The rate distortion costs test of formula.

Method the most according to claim 1, it is characterised in that: the process that implements of described step (1) is: first determine Current coded unit candidate motion parameter set under merging patterns is closed, this candidate motion parameter set close by with current coded unit The kinematic parameter composition of space or temporally adjacent and encoded each predicting unit；Then under merging patterns, calculate each candidate fortune Move motion compensated image corresponding to parameter and prediction residual, and by prediction residual is carried out two-dimension discrete cosine transform and quantization Process and calculate rate distortion costs corresponding to each candidate motion parameter and prediction residual absolute value sum, with minimum rate distortion generation Valency is that criterion chooses optimal candidate motion parameter from candidate motion parameter set is closed, and records this optimal candidate kinematic parameter pair Rate distortion costs J answered_mergeAnd prediction residual absolute value sum S_merge；

In like manner determine that current coded unit candidate motion parameter set under skip mode is closed, this candidate motion parameter set close by with The kinematic parameter composition of current coded unit space or temporally adjacent and encoded each predicting unit；Then under skip mode Calculate motion compensated image corresponding to each candidate motion parameter and prediction residual, and by prediction residual being carried out more than two-dimensional discrete String transform and quantization processes and calculates rate distortion costs corresponding to each candidate motion parameter and prediction residual absolute value sum, with Little rate distortion cost is criterion chooses optimal candidate motion parameter from candidate motion parameter set is closed, and records this and most preferably wait Select rate distortion costs J that kinematic parameter is corresponding_skipAnd prediction residual absolute value sum S_skip。

Method the most according to claim 1, it is characterised in that: described threshold value T_skipDetermination method as follows:

First with set quantization parameter QP, use HEVC encoder to input test video encode, for all N × The coding unit of N size, record is wherein with the prediction residual absolute value sum that the coding unit of skip mode coding is corresponding, composition Observation sample set { K+}；The prediction residual absolute value sum that coding unit that record wherein encodes with non-skip mode is corresponding, Composition observation sample set { K-}；

Take and meet prediction residual absolute value sum S under following condition premise_iMinima as threshold value T_skip；

p(S_i| K-) P (K-) ＞ p (S_i|K+)P(K+)

Wherein: S_iFor the prediction residual absolute value sum that i-th coding unit is corresponding, p (S_i| K+) represent S_iRelative to observation sample This set { distribution density of K+}, p (S_i| K-) represent S_iRelative to observation sample set the distribution density of K-}, and P (K+) be with The coding unit of skip mode coding ratio shared by the coding unit of all N × N sizes, P (K-) is for skip mould with non- Ratio that the coding unit of formula coding is shared in the coding unit of all N × N sizes and P (K-)=1-P (K+), i is nature Number and the coding unit quantity that 1≤i≤m, m are all N × N sizes；

According to above-mentioned, travel through all possible quantization parameter QP and size N, obtain each group of threshold corresponding for QP with N determined Value T_skip。

Method the most according to claim 1, it is characterised in that: described threshold value T_mergeDetermination method as follows:

First with set quantization parameter QP, use HEVC encoder to input test video encode, for all N × The coding unit of N size, record is wherein with the prediction residual absolute value sum that the coding unit of merging patterns coding is corresponding, composition Observation sample set { M+}；Record wherein with the prediction residual absolute value sum that the coding unit of unconsolidated pattern-coding is corresponding, Composition observation sample set { M-}；

Take and meet prediction residual absolute value sum S under following condition premise_iMinima as threshold value T_merge；

p(S_i| M-) P (M-) ＞ p (S_i|M+)P(M+)

Wherein: S_iFor the prediction residual absolute value sum that i-th coding unit is corresponding, p (S_i| M+) represent S_iRelative to observation sample This set { distribution density of M+}, p (S_i| M-) represent S_iRelative to observation sample set the distribution density of M-}, and P (M+) be with The coding unit of merging patterns coding ratio shared by the coding unit of all N × N sizes, P (M-) is with unconsolidated mould Ratio that the coding unit of formula coding is shared in the coding unit of all N × N sizes and P (M-)=1-P (M+), i is nature Number and the coding unit quantity that 1≤i≤m, m are all N × N sizes；

According to above-mentioned, travel through all possible quantization parameter QP and size N, obtain each group of threshold corresponding for QP with N determined Value T_merge。

5. according to the method described in claim 3 or 4, it is characterised in that: described prediction residual absolute value sum S_iRelative to sight Survey the sample set { distribution density p (S of L}_i| L) computational methods as follows, wherein L=K+, K-, M+ or M-；

First, observation sample set { each separation A of L} is determined₀,A₁,...,A_m+1, wherein A₀=0, A_i-A_i-1=d, A_m+1It is more than Equal to observation sample set { maximum in L} and less than this maximum+d；D be default group away from；

Then, distribution density p (S is calculated according to below equation_i| L):

$p\left(S_i\right|L)=\frac{C_i}{{\mathrm{\Σ}}_{i=1}^m{C}_i}$

Wherein: C_iFor observation sample set, { L} declines at interval [A_i,A_i+1) number of samples.

6. according to the method described in claim 3 or 4, it is characterised in that: the span of described quantization parameter QP is 0～51 Integer, size N=8 of described coding unit, 16,32 or 64.