CN106713935A - Fast method for HEVC (High Efficiency Video Coding) block size partition based on Bayes decision - Google Patents

Abstract

The invention discloses a fast method for HEVC (High Efficiency Video Coding) block size partition based on a Bayes decision. The fast method comprises the following steps: first of all, dividing a video sequence into an online learning stage and a fast partitioning stage by employing scene change detection based on an average gray scale difference; then, for the online learning stage and a video frame which occurs a scene change, in each partitioning depth, respectively extracting Jinter and Jintra of a CU (Coding Unit) as characteristic values, thereby establishing a mixed Gaussian model, wherein specific parameters of the model are determined according to an EM algorithm initialized by a K-Means algorithm; and for a to-be-partitioned CU in the fast partitioning stage, extracting the characteristic values and finding a conditional probability on whether to partition according to the mixed Gaussian model, and at last, finding the decision with a relatively small risk by employing a Bayes formula of a minimum risk to take as a judgment basis on whether the current CU is partitioned. According to the fast method disclosed by the invention, the algorithm complexity is reduced, and the coding time can be greatly reduced.

Description

A kind of HEVC blocks based on Bayesian decision divide fast method

Technical field

It is the invention belongs to high-definition digital video compression coding and decoding technical field more particularly to a kind of based on Bayesian decision HEVC (High Efficiency Video Coding, efficient video coding) block divides fast method.

Background technology

In the fast development of digital video application industrial chain, in face of Video Applications constantly to fine definition, frame per second high, height The trend that compression ratio direction is developed, previous generation video compression standards agreement limitation H.264/AVC is constantly highlighted.Therefore, by The tissue JCT-VC that International Telecommunication Union (ITU-T) and Motion Picture Experts Group (MPEG) joint are set up proposes video of future generation Encoding and decoding standard --- H.265/HEVC.Its target is that under the premise of identical picture quality, compression ratio is than H.264/AVC high-grade It is secondary to double.

HEVC is in terms of coding principle and basic structure without breakthrough, and H.264/AVC basically identical, i.e. prediction plus conversion Block encoding mode；On Ciphering details and H.264/AVC also very close to comprising infra-frame prediction, inter prediction, estimation With the coding/decoding module such as compensation, orthogonal transformation, quantization, loop filtering, entropy code and reconstruction.But, and H.264/AVC compare Compared with HEVC almost takes important corrective measure in each coding link, and such as infra-frame prediction supports 35 kinds of patterns, frames Between predict that introducing Merge patterns, change quantization supports that being up to 32 × 32 converter units, entropy code uses CABAC (Context Adaptive Binary Arithmetic Coder, the adaptive binary arithmetic coding based on context) algorithm, newly introducing Self adaptation sampling point compensation loop filtering etc..

HEVC encoders employ code tree unit (CTU, Coding Tree Unit), coding unit (CU, Coding Unit), predicting unit (PU, Prediction Unit), converter unit (Transform Unit, TU) structure so that HEVC energy Enough videos to different resolution and applied environment are encoded.When HEVC is encoded, piece image can be divided into multiple mutual Nonoverlapping CTU, each CTU is divided into one or more CU；When frame in or inter prediction encoding is carried out, a CU can To select to be divided into one or more PU；In conversion, quantization operation, a CU can be divided into one or more TU.

CU is the elementary cell of HEVC codings, using recursive dividing mode.HEVC support 8 × 8,16 × 16,32 × 32, 64 × 64 this 4 kinds of CU of size, corresponding division depth is respectively 3,2,1,0.

PU is the elementary cell of HEVC predictions, it is stipulated that the predictive mode of CU, all coded datas related to prediction are all Transmitted by PU.For a size for 2N × 2N CU, its frame in PU can lectotype have 2 kinds：2N × 2N and N × N；Interframe PU can lectotype have 8 kinds, including 4 kinds of symmetric patterns：2N × 2N, 2N × N, N × 2N, N × N and 4 kinds of asymmetric moulds Formula：2N×nD、2N×nU、nR×2N、nL×2N.When using Merge when encoding, residual information is to be coded of, The PU patterns for now using are 2N × 2N skip patterns, and PU partition modes are as shown in Figure 1.

TU is the elementary cell of HEVC transform and quantizations, and its size is also flexible change, is entered in quaternary tree form Row recursive subdivision.When CTU sizes are 64 × 64, the quaternary tree recurrence of TU is divided as shown in Fig. 2 wherein CU's draws in each CU Divide indicated by the solid line, the division of TU is represented by dashed line, and coded sequence is represented with alpha code.

Similar with other international encoding standards, HM encoders are also based on rate-distortion optimization for the decision-making of block size (Rate-Distortion Optimization's, RDO), in HEVC standard encryption algorithm, the division process description of CU is such as Under：

(1) depth for setting current CU is zero；

(2) rate distortion costs (the Rate-Distortion Cost, RD of the Skip patterns under current division depth are calculated Cost)J_skip, and update the rate distortion costs of current optimal partition mode (BestCuMode=Skip) and minimum (MinRDcost=J_skip)；

(3) the rate distortion costs J of the Merge_2Nx2N patterns under current division depth is calculated_{merge_2Nx2N}, and with before The minimum rate distortion costs for going out are compared, if J_{merge_2Nx2N}<MinRDcost, updates current optimal partition mode And minimum rate distortion costs (MinRDcost=J (BestCuMode=Merge_2Nx2N)_{merge_2Nx2N})；

(4) the rate distortion costs J of the Inter_2Nx2N patterns under current division depth is calculated_{Inter_2Nx2N}, and with before The minimum rate distortion costs for going out are compared, if J_{Inter_2Nx2N}<MinRDcost, updates current optimal partition mode And minimum rate distortion costs (MinRDcost=J (BestCuMode=Inter_2Nx2N)_{Inter_2Nx2N})；

(5) if the maximum that the depth of current CU is encoder to be allowed divides depth, and the size of CU is bigger than 8 × 8, then ask The rate distortion costs J of Inter_NxN patterns_{lnter_NxN}, and compared with the minimum rate distortion costs for drawing before, if J_{lnter_NxN}<MinRDcost, updates the rate distortion of current optimal partition mode (BestCuMode=Inter_NxN) and minimum Cost (MinRDcost=J_{Inter_NxN}), if being unsatisfactory for above-mentioned CU depth and size requirement, skip step 5；

(6) the rate distortion costs J of the Inter_2NxN patterns under current division depth is calculated_{Inter_2NxN}, and with before The minimum rate distortion costs for going out are compared, if J_{inter_2NxN}<MinRDcost, updates current optimal partition mode And minimum rate distortion costs (MinRDcost=J (BestCuMode=Inter_2NxN)_{Inter_2NxN})；

(7) the rate distortion costs J of the Inter_Nx2N patterns under current division depth is calculated_{Inter_Nx2N}If, J_{Inter_Nx2N}< MinRDcost, updates the rate distortion costs of current optimal partition mode (BestCuMode=Inter_Nx2N) and minimum (MinRDcost=J_{Inter_Nx2N})；

(8) calculate whether current CU needs to calculate the rate distortion costs of AMP patterns, if BestCuMode=Inter_ 2Nx2N or BestCuMode=Inter_2NxN, Test_Hor condition are true, and otherwise Test_Hor is false；If BestCuMode=Inter_2Nx2N or BestCuMode=Inter_Nx2N, Test_Ver condition are true, otherwise Test_ Ver is false.

(9) if Test_Hor conditions are true, next step is continued；Otherwise, step 12 is jumped to；

(10) the rate distortion costs J of the Inter_2NxnU patterns under current division depth is calculated_{Inter_2NxnU}If, J_{Inter_2NxnU}<M inRDcost, update the rate of current optimal partition mode (BestCuMode=Inter_2NxnU) and minimum Distortion cost (MinRDcost=J_{Inter_2NxnU})；

(11) the rate distortion costs J of the Inter_2NxnD patterns under current division depth is calculated_{Inter_2NxnD}If, J_{Inter_2NxnD}<MinRDcost, updates the rate of current optimal partition mode (BestCuMode=Inter_2NxnD) and minimum Distortion cost (MinRDcost=J_{Inter_2NxnD})；

(12) if Test_Ver conditions are true, next step is continued；Otherwise, step 15 is jumped to；

(13) the rate distortion valency J of the Inter_nLx2N patterns under current division depth is calculated_{Inter_nLx2N}If, J_{Inter_nLx2N} ＜ MinRDcost, update the rate distortion costs of current optimal partition mode (BestCuMode=Inter_nLx2N) and minimum (MinRDcost=J_{Inter_nLx2N})；

(14) the rate distortion costs J of the Inter_nRx2N patterns under current division depth is calculated_{Inter_nRx2N}If, J_{Inter_nRx2N}<MinRDcost, updates the rate of current optimal partition mode (BestCuMode=Inter_nRx2N) and minimum Distortion cost (MinRDcost=J_{Inter_nRx2N})；

(15) the rate distortion costs J of the Intra_2Nx2N patterns under current division depth is calculated_{Intra_2Nx2N}If, J_{Intra_2Nx2N}<MinRDcost, updates the rate of current optimal partition mode (BestCuMode=Intra_2Nx2N) and minimum Distortion cost (MinRDcost=J_{Intra_2Nx2N})；

(16) if the current division depth of CU is by the depth capacity for allowing of encoding setting, its rate of lntra_NxN is calculated Distortion cost J_{lntra_NxN}If, J_{lntra_NxN}<MinRDcost, updates current optimal partition mode (BestCuMode=lntra_ ) and minimum rate distortion costs (MinRDcost=J NxN_{lntra_NxN})；

(17) if the maximum that the current division depth ratio PCM patterns of CU are allowed is divided, depth is small, calculates the rate of PCM patterns Distortion cost J_PCMIf, J_PCM＜ MinRDcost, then update current optimal partition mode (BestCuMode=PCM) and minimum Rate distortion costs (MinRDcost=J_PCM)；

(18) if BestCuMode is not more than coding with permitting for centering by the current size of Skip patterns or CU Perhaps minimum dimension, jumps to step 19；Otherwise, current CU will be divided into four sub-blocks, and the depth ratio of each sub-block is current Block is big by 1.Then to the algorithm steps of each Sub-CU iteration (1)-(18), the MinRDcost of sub-block is obtained respectively, after adding up It is J to make_sub-CU。

(19) rate distortion costs for comparing before dividing and being divided into after sub-block, if MinRDcost<J_Sub-CU, it is current to divide deep CU under degree does not continue segmentation, updates minimum rate distortion costs (MinCucost=MinRDcost)；Otherwise current CU is divided It is four Sub-CU, and updates MinCuCost=J_Sub-CU。

The algorithm steps determined by above-mentioned CU sizes understand that encoder will determine a quad-tree partition mode of CTU, Not only to determine the division depth of CU, also to travel through all possible PU, TU division side under a certain depth, finally therefrom select RD Cost minimum coding mode and corresponding division depth, therefore the computation complexity of HEVC block partition process is high.

The content of the invention

The present invention for technical problem present in above-mentioned prior art, according to video sequence time-space domain correlation, Propose a kind of HEVC blocks based on Bayesian decision and divide fast method.In the present invention, CU division operations are regarded as two points Class problem, i.e. W={ ω_N, ω_P, wherein ω_NRepresent that current CU does not continue to divide, ω_pRepresent that current CU continues to divide, and use Least risk Bayes decision-making solves two classification problems.

The technical solution adopted for the present invention to solve the technical problems comprises the following steps：

Step 1, off-line learning obtain loss function C_NPAnd C_PNAnd prior probability P (ω_N) and P (ω_P)。

When decision-making is done using the Bayesian decision criterion of minimal error rate, if the division result decision error of CU, i.e. shellfish The judgement whether this decision-making of leaf divides to current depth CU is different from the result of canonical algorithm, under will causing distortion performance Drop.Simultaneously for the size CU different with characteristic value, the degree that the distortion performance that the result of decision of mistake causes declines is not yet Together.Therefore, the present invention proposes the decline of the distortion performance for bringing erroneous decision as bayes risk function, and foundation is based on The Bayes decision-making model of minimum risk, the distortion performance brought with reducing erroneous decision is lost, complete rate distortion costs letter Number is shown below：

J_FRD=(SSE_luma+ω_chroma×SSE_chroma)+λ_mode×B_mode

Wherein, SSE_lumaRepresent the pixel difference error sum of squares between the original picture block and reconstruction image block of brightness； SSE_chromaRepresent the pixel difference error sum of squares between the original picture block and reconstruction image block of carrier chrominance signal, ω_chromaIt is color Spend the weight coefficient of component, λ_modeIt is Lagrange's multiplier, B_modeRequired bit number when being to current coding mode.

For 64 × 64, the CU of 32 × 32,16 × 16 sizes is trained by off-line learning obtain loss matrix C respectively, is damaged Matrix is lost to be defined as follows shown in formula：

C herein_PNAnd C_NPIt is defined as follows：

J_{FRD, N}And J_{FRD, P}Be respectively current CU to be divided do not continue to divide with continue to be divided into four sub-blocks it is corresponding completely Rate distortion costs.C_NPRepresent and assume ω_NIt is true, i.e., judges to assume ω when rate distorted result is " not dividing "_PSet up, i.e., CU is adopted Take the value of risk brought during " division " decision-making；C_PNRepresent and assume ω_PIt is true, is brought when taking CU " not dividing " decision-making Value of risk.C_NPAnd C_PNMore than or equal to 0, (C when decision-making is correct_NNOr C_PP), the penalty values of decision-making are 0.

Prior probability P (ω_R) it is by counting under current depth, making ω_RThe proportion that the CU of decision-making accounts for all CU is obtained , i.e.,

P(ω_R)=current depth belongs to ω_RCU numbers/current depth CU sum；

Wherein R is N or P；

Step 2, for incoming sequence of pictures, judge whether whether it there occurs in on-line study stage or scene Change, if present frame is in on-line study stage or scene change frame, depth division is carried out according to HM canonical algorithms to it, The characteristic value of each depth is extracted and stored, step 7 is jumped to；Otherwise perform step 3.

Scene change be in order to judge the severe degree of video image motion, when there is scene change in video sequence, one As for i.e. show video content there occurs acutely variation, will be lost between adjacent frame of video temporal correlation (depending on Temporal redundancy of the frequency at switching disappears).In this case, the block based on Bayesian decision divides fast algorithm and discomfort With, once because scene is converted, the threshold value for judging whether to divide will change, based on offline statistics and on-line study Bayesian decision correct decisions will can not be made according to the characteristic value extracted.Therefore the present invention is proposed based on scene change inspection The on-line study stage of survey and fast alternate mechanism stage by stage.For preceding N frames, Dou Yaochong after N frames before sequence and scene change New characteristic value of extracting sets up new mixed Gauss model.

Because image intensity value can just be drawn before present frame is encoded by simple computation, if adjacent image is in together One scene, adjacent frame correlation is stronger, and gray value is more or less the same；Conversely, adjacent picture frame is in identical bits when occurrence scene is converted The gray value for putting pixel generally differs larger.

2-1. weighs the severe degree of video motion with average gray difference value, i.e., whether occurrence scene is converted, video sequence The average gray difference value of pixel of adjacent two frame in same position is calculated as follows shown in formula in row：

In formula, P, Q are picture traverse and height, G_n(i, j) is the gray value at n-th frame pixel (i, j) place, G_n-1(i, j) It is the gray value at (n-1)th frame pixel (i, j) place.If without scene switching, G is generally smaller.If occurrence scene is converted, G leads to Often it is increased dramatically.

In sum, a detection threshold value for judging whether occurrence scene conversion can be set, if G is more than detection threshold Value λ_sc, then it is assumed that scene is converted, otherwise it is assumed that scene is not changed in；λ_scIt is detection threshold value, its value is often empirical value.Threshold The size of value determines the accuracy of scene change, and crossing conference causes missing inspection and too small, can cause flase drop, therefore λ_scValue should be closed Reason.The frame number in on-line study stage is set to 6 by the present invention, and threshold value is set to 9, that is, work as G>Think that occurrence scene is converted when 9, otherwise There is no occurrence scene to convert.Once scene is converted, the statistical parameter of characteristic vector need be just reevaluated.

Step 3, judge whether picture is fast the first two field picture stage by stage, if it is not, performing step 4；Otherwise carry The characteristic value of on-line study stage or the storage of scene change frame is taken, and feature based value sets up mixed Gauss model, the system of model Meter parameter is estimated to expectation maximization (EM) algorithm of K-Means algorithm initializations.

If the observation information of each CU n dimensional feature vector X=[x₁, x₂, x₃…x_n] describe, p (X | ω_R), R ∈ N, P }, it is ω_RThe conditional probability density of feature vector, X is observed under state.State ω_RPosterior probability p (ω_R| X) can be by pattra leaves This formula is tried to achieve

For the risk that the decision-making for weighing mistake brings, present invention introduces loss function C_RQ, represent rate-distortion optimization result It is decision-making R, mistakenly classifies as the risk in the case of decision-making Q.Consider that the Bayes rule in the case of risk is changed into：

R_N(ω_N| X)=C_NN p(X|ω_N)P(ω_N)+C_PNp(X|ω_P)P(ω_P)

R_P(ω_P| X)=C_PPp(X|ω_P)P(ω_P)+C_NP p(X|ω_N)P(ω_N)

R_R(X) it is conditional risk, illustrates characteristic value X and be judged to ω_RThe average lost during class, C_NPRepresent and assume ω_NFor Very, i.e., judge to assume ω when rate distorted result is " not dividing "_PSet up, i.e., take CU the risk brought during " division " decision-making Cost.

In order to ensure expected risk minimum, when must ensure to do each decision-making, all cause that conditional risk is minimum, it is such to determine Plan is referred to as Least risk Bayes decision-making, and the rule of decision-making is as follows：If R_N(ω_N|X)<R_P(ω_P| X), then ω=ω_N, decision-making It is ω_N, represent the decision-making that " not dividing " is taken current CU；The decision-making of " division " is otherwise taken current CU.

3-1. feature extractions and selection

As described above, it is necessary to extract characteristic vector from observation to cause decision-making rationally, for the space-time of reflecting video Feature, our selected characteristic vectors are J_interAnd J_intra, obtained by following formula：

J_LRD=SAD+ λ_pred×B_pred

J_inter=argminP ∈ Inter { J_LRD(P)}

J_intra=argminP ∈ Intra { J_LRD(P)}

λ herein_predIt is Lagrange's multiplier, B_predTo all information under current prediction mode, (such as dividing mode, pattern are believed Breath, conversion coefficient etc.) bit number required when being encoded, P represents that the candidate of PU divides.The calculating of characteristics extraction is answered Miscellaneous degree is almost negligible, because the information of rate distortion costs can be obtained directly from cataloged procedure.

Feature vector, X is obtained from the picture in on-line study stage, if training picture is I frames, the dimension of characteristic vector Just drop to 1, X=J_intra, because I frames can only use intra prediction mode.

The probability density function of 3-2. on-line studies

In order to Bayesian decision is applied to quick CU partitioning algorithms, it is necessary to each on-line study stage estimate feature to The statistical parameter of amount.In the method for proposing, we are according to R_N(ω_N| X) and R_P(ω_P| X) value determine whether to divide current CU to be divided；From Least risk Bayes formula, to compare R_R(ω_R| X) size, it is necessary to first try to achieve p (X | ω_N)P (ω_N) and p (X | ω_P)P(ω_P) value.

Each is divided in the case of " division " or " not dividing " of depth, and likelihood function is all obeyed dimensional Gaussian and divided at random Cloth, i.e. p (X | ω_R) by two Gaussian probability-density function weighted sum approximate representations of independent probability, therefore can have：

Wherein i is i-th single Gauss model in mixed Gauss model.The parameter of each Gauss model is θ_i=(μ_i, σ_i), μ_iAnd σ_iRepresent class ω_iCharacteristic vector mean vector and variance, the mixing probability function of feature vector, X is

Wherein Θ=(θ₁,θ₂) be mixed Gauss model distributed constant collection, P=(p (ω₁),p(ω₂)) it is corresponding height The prior probability collection of this model.The parameter of mixed Gauss model is determined by the EM algorithms with K-Means algorithm initializations.

Step 4, the characteristic value for extracting current CU, try to achieve p (X | ω_R), R ∈ { N, P }.

The characteristic value X of CU to be divided is extracted, the letter on the Gaussian mixtures of same depth is tried to achieve according to residing depth Numerical value, i.e. p (X | ω_N) and p (X | ω_P)。

Step 5, decide whether to divide current CU, if dividing, mistake described in step 1- steps 5 is performed to each Sub-CU Journey；Do not divide current CU otherwise, jump to step 6.

Calculating is comparedWithSize, ifIt is larger, current CU is divided into four Sub-CU, And step 1-5 is performed to each Sub-CU；Otherwise, current CU is not divided.

Step 6, the CU is encoded, if current frame image CU divided completions, perform step 7；Otherwise, to next Individual CU performs process described in step 1-6.

Fast algorithm is divided according to the block based on Bayes and encodes current CU, then to next CU traversal steps 1-6.If working as All of CU has completed to divide in preceding picture frame, then perform step 7.Step 7,2-7 steps institutes is performed to next two field picture Process is stated, until having encoded all of frame of video.

For a two field picture of new incoming, the same block based on Bayes divides fast algorithm and processes, until having encoded All of frame of video.

So far, the whole HEVC blocks based on Bayesian decision divide fast method and terminate, for the division of a certain depth CU.

Beneficial effect

This method advantage is as follows：

(1) present invention does high-speed decision by Bayesian decision come the dividing mode to current depth CU, reduces CU and draws Divide algorithm complex, the HEVC scramble times can be greatly reduced.

(2) present invention employs the scene change detection mechanism based on average gray difference value, it is ensured that the stability of method And accuracy.

(3) using of the invention test result indicate that, compared with the original algorithms of HM13.0, announcement method of the present invention is reduced 44.5% scramble time, and coding bit rate BD-Rate only increases by 0.94%.

Brief description of the drawings

Fig. 1：H.265/HEVC the partition mode of PU.

Fig. 2：The quad-tree partition of CU, TU.

Fig. 3：On-line study stage and fast replace schematic diagram stage by stage.

Fig. 4：HEVC blocks based on Bayesian decision divide fast algorithm main-process stream.

Specific embodiment

By taking the block partition process of 16 sequences such as People on street as an example, the inventive method to proposing is carried out in detail Explanation.

The step of HEVC blocks based on Bayesian decision divide fast algorithm is as follows：

Step 1, off-line learning obtain loss function C_NPAnd C_PNAnd prior probability P (ω_N) and P (ω_P)。

According to below equation respectively to 64 × 64, the CU of 32 × 32,16 × 16 sizes is trained by off-line learning and damaged Lose Matrix C and prior probability：

J_FRD=(SSE_luma+ω_chroma×SSE_chroma)+λ_mode×B_mode

Wherein, SSE_luma, SSE_chromaRepresent the picture between the original picture block and reconstruction image block of luminance and chrominance information Element difference error sum of squares, ω_chromaIt is the weight coefficient of chromatic component, λ_modeIt is Lagrange's multiplier, B_modeIt is to present encoding Required bit number during pattern.J_{FRD, N}And J_{FRD, P}It is respectively that current CU to be divided does not continue to divide and continue to be divided into four The corresponding complete rate distortion costs of sub-block.

Prior probability P (ω_R) it is by counting under current depth, making ω_RThe proportion that the CU of decision-making accounts for all CU is obtained , i.e.,

P(ω_R)=current depth belongs to ω_RCU numbers/current depth CU sum

Step 2, for incoming sequence of pictures, judge it whether in on-line study stage or whether scene there occurs Change, if present frame is in on-line study stage or scene change frame, depth division is carried out according to HM canonical algorithms to it, The characteristic value of each depth is extracted and stored, step 7 is jumped to；Otherwise perform step 3.

Video motion is weighed using the average gray difference value of pixel of adjacent two frame in same position in video sequence Severe degree, is calculated as follows shown in formula：

In formula, P, Q are picture traverse and height, G_n(i, j) is the gray value at n-th frame pixel (i, j) place, G_n-1(i, j) It is the gray value at (n-1)th frame pixel (i, j) place.If without scene switching, G is generally smaller.If occurrence scene is converted, G leads to Often it is increased dramatically.Once scene is converted, the statistical parameter of characteristic vector need be just reevaluated.Judge video according to following formula Whether occurrence scene changes frame：

Scene change be in order to judge the severe degree of video image motion, when there is scene change in video sequence, one As for i.e. show video content there occurs acutely variation, will be lost between adjacent frame of video temporal correlation (depending on Temporal redundancy of the frequency at switching disappears).In this case, the block based on Bayesian decision divides fast algorithm and discomfort With, once because scene is converted, the threshold value for judging whether to divide will change, based on offline statistics and on-line study Bayesian decision correct decisions will can not be made according to the characteristic value extracted.Therefore the present invention is proposed based on scene change inspection The on-line study stage of survey and fast alternate mechanism stage by stage.Schematic diagram is as shown in Figure 3.Become for N frames before sequence and scene N frames, will again extract characteristic value and set up new mixed Gauss model before after changing.

Step 3, judge whether picture is fast the first two field picture stage by stage, if it is not, performing step 4；Otherwise carry The characteristic value of on-line study stage or the storage of scene codes frame is taken, mixed Gauss model, the system of model are set up based on these data Meter parameter is estimated to the EM algorithms of K-Means algorithm initializations.

Assuming that the observation information of each CU n dimensional feature vector X=[x₁, x₂, x₃…x_n] describe, our selected characteristics Vector is J_interAnd J_intra, obtained by following formula：

J_LRD=SAD+ λ_pred×B_pred

J_inter=argminP ∈ Inter { J_LRD(P)}

J_intra=argminP ∈ Intra { J_LRD(P)}

λ herein_predIt is Lagrange's multiplier, B_predTo all information under current prediction mode, (such as dividing mode, pattern are believed Breath, conversion coefficient etc.) bit number required when being encoded, P represents that the candidate of PU divides.

Feature vector, X is obtained from the picture in on-line study stage, if training picture is I frames, the dimension of characteristic vector Just drop to 1, X=J_intra, because I frames can only use intra prediction mode.

Each is divided in the case of " division " or " not dividing " of depth, and likelihood function is all obeyed dimensional Gaussian and divided at random Cloth, i.e. p (X | ω_R) by two Gaussian probability-density function weighted sum approximate representations of independent probability, therefore can have：

Wherein i is i-th single Gauss model in mixed Gauss model.The parameter of each Gauss model is θ_i=(μ_i, σ_i), μ_iAnd σ_iRepresent class ω_iCharacteristic vector mean vector and variance, the mixing probability function of feature vector, X is

Wherein Θ=(θ₁,θ₂) be mixed Gauss model distributed constant collection, P=(p (ω₁),p(ω₂)) it is corresponding height The prior probability collection of this model.

In order to obtain the parameter of mixed Gauss model, defining log-likelihood function is

Wherein m represents the dimension of characteristic vector.Parametric solution is to need to obtain model parameter (Θ₀, P₀) cause formula logarithm Likelihood function value is maximum, then the statistical parameter of mixed model is estimated meet following formula：

The present invention is estimated the statistical parameter of mixed Gauss model with EM algorithms.Solution is divided into two using EM algorithms Individual step：The parameter for assuming initially that Gauss model is known, goes to estimate the weights of each Gauss model using parameter；Second Step, obtains after weights, recycles the weights of Gauss model to go to estimate its parameter.

The initial value for assuming initially that the mixed Gauss model parameter of the characteristic value of CU blocks is Θ⁽⁰⁾, prior probability initial value is P⁽⁰⁾, the mixed Gauss model estimates of parameters for then setting t steps is Θ^(t)And P^(t), then the calculating process of t+1 steps is as follows：

(1) E steps (calculate and expect)：The estimate Θ of iteration is walked according to current mixed Gauss model parameter t^(t)And P^(t), calculating in known current characteristic value, the decision-making of CU belongs to ω_iPosterior probability estimate：

Obviously, if characteristic value x_m(m=1,2) ω_iP (the ω of class_i|x_m,θ_i)≥P(ω_i|x_m,θ_j), j ≠ i, then x_m Belong to ω_iThe possibility of class is just very big.

(2) M steps (maximize and expect)：The posterior probability that each data sample belongs to each subclass is obtained, this step profit The statistical parameter estimation formulas of formula mixed model are solved with gradient descent method.So as to obtain Θ, estimations of the P in the step of kth+1.Will be right Number likelihood function brings the posterior probability that above formula is calculated into respectively to Θ and P derivations, can be obtained after arrangement：

It is thereinBelong to the probability number of the characteristic value of the i-th class when representing the step iteration of kth+1.EM algorithms are by continuous The parameter of iteration E steps and M step improved models, until convergence can obtain the parameter of mixed Gauss model.The present invention The initial value of middle EM algorithms has K-Means algorithms to determine, and it is the metric algorithm of distinctiveness ratio to use Euclidean distance, Its formula is as follows：

Wherein A={ a₁,a₂,…,a_n, B={ a₁,a₂,…,a_nIt is two data element entries, each have n individual mensurable Characteristic attribute.

K-Means algorithms arbitrarily choose center of the k object as initial clustering first from sample.Then according to sample Distinctiveness ratio between cluster centre judges which cluster each sample belongs to.Then the cluster centre of each cluster is updated, is repeated Iteration said process, until criterion function is restrained, comprises the following steps that shown：

The initial cluster center point for randomly selecting k cluster first isThen repeat procedure below until Convergence：

(1) calculate each conditional probability p (X | ω_i) belonging to class：

(2) in each class j, such barycenter is recalculated：

K is the number of cluster, it is necessary to be specified before cluster.c⁽ⁱ⁾Represent conditional probability p (X | ω_i) with distance in k class most That near class, c⁽ⁱ⁾∈{1,2,3,…,k}.Barycenter μ_jRepresent j-th cluster centre position of class, iteration step (1) With step (2) until cluster centre no longer changes or varies less, in the present invention, k=2 is taken.

Step 4, the characteristic value for extracting current CU, try to achieve p (X | ω_R), R ∈ { N, P }.

The characteristic value X of CU to be divided is extracted, the letter on the Gaussian mixtures of same depth is tried to achieve according to residing depth Numerical value, i.e. p (X | ω_N) and p (X | ω_P)。

Step 5, decide whether to divide current CU, if dividing, process described in step 1-5 is performed to each Sub-CU；It is no Do not divide current CU then, jump to step 6.

Calculating is comparedWithSize, ifIt is larger, current CU is divided into four Sub-CU, And step 1-5 is performed to each Sub-CU；Otherwise, current CU is not divided.

Step 6, the CU is encoded, if current frame image CU divided completions, perform step 7；Otherwise, to next Individual CU performs process described in step 1-6.

Fast algorithm is divided according to the block based on Bayes and encodes current CU, then to next CU traversal steps 1-6.If working as All of CU has completed to divide in preceding picture frame, then perform step 7.

Step 7, to next two field picture perform 2-7 steps described in process, until having encoded all of frame of video.

For a two field picture of new incoming, the same block based on Bayes divides fast algorithm and processes, until having encoded All of frame of video.

So far, the whole HEVC blocks based on Bayesian decision divide fast method and terminate, for the division of a certain depth CU, Algorithm flow chart is as shown in Figure 4.

Experiment test environment of the present invention is compiled using HEVC standard identifying code HM13.0 in the system VS2010 of windows 7 16 cycle tests are counted and tested by operation.With full I frames (All Intra, AI), arbitrary access (Random Access, RA), low latency-P frames (LP), low latency-B frames (LB) be coding environment, set 22,27,32,37 as QP values, it is sharp The block that will be proposed with the Bjontegaard-Delta bit rates (YBD-rate) of brightness divides fast algorithm and is drawn with HEVC standard block Divide algorithm to be contrasted, and count the encoding and decoding time, as a result as shown in table 1.

Algorithm proposed by the present invention is average under the configuration of AI compared to HM13.0 canonical algorithms as shown in Table 1 saves 46% Scramble time above, and BD-Rate only have lost 1.08%；45.9% scramble time, BD- are averagely saved under RA configurations Rate losses 1.12%；Scramble time under LB configurations reduces 42.8%, and BD-Rate only have lost 0.83%；LP is configured Averagely save for 42.8% scramble time down, BD-Rate have lost 0.73%, and coding efficiency declines very little.

The block that table 1 is based on Bayesian decision divides fast algorithm and HEVC standard algorithm effect comparing result

Certainly, those of ordinary skill in the art is it should be appreciated that above example is intended merely to illustrate this hair Bright, and limitation of the invention is not intended as, as long as within the scope of the invention, change, modification to above example are all Protection scope of the present invention will be fallen into.

Claims (4)

1. a kind of HEVC blocks based on Bayesian decision divide fast method, it is characterised in that regard CU division operations as two classification Problem, i.e. W={ ω_N, ω_P, wherein ω_NRepresent that current CU does not continue to divide, ω_pRepresent that current CU continues to divide, and using most Small risk Bayesian decision solves two classification problems, implements step as follows：

Step 1, off-line learning obtain loss function C_NPAnd C_PNAnd prior probability P (ω_N) and P (ω_P)；

Step 2, for incoming sequence of pictures, judge whether whether it there occurs change in on-line study stage or scene Change, if present frame is in on-line study stage or scene change frame, depth division is carried out to it according to HM canonical algorithms, carry The characteristic value of each depth is taken and stored, step 7 is jumped to；Otherwise perform step 3；

Step 3, judge whether picture is fast the first two field picture stage by stage, if it is not, performing step 4；Otherwise it is extracted in Line learns the characteristic value of stage or the storage of scene change frame, and feature based value sets up mixed Gauss model, the statistics ginseng of model Number is estimated to the EM algorithms of K-Means algorithm initializations；

Step 4, the characteristic value for extracting current CU, try to achieve p (X | ω_R)R∈{N,P}；

The characteristic value X of CU to be divided is extracted, the functional value on the Gaussian mixtures of same depth is tried to achieve according to residing depth, I.e. p (X | ω_N) and p (X | ω_P)；

Step 5, decide whether to divide current CU, if dividing, process described in step 1- steps 5 is performed to each Sub-CU；It is no Do not divide current CU then, jump to step 6；

The judgement divided for current CU is as follows：Calculating is comparedWithSize, ifIt is larger, ought Preceding CU is divided into four Sub-CU, and performs step 1-5 to each Sub-CU；

Step 6, the CU is encoded, if current frame image CU divided completions, perform step 7；Otherwise, to next CU Perform process described in step 1-6；

Step 7, to next two field picture perform 2-7 steps described in process, until having encoded all of frame of video.

2. a kind of HEVC blocks based on Bayesian decision according to claim 1 divide fast method；It is characterized in that step Rapid 1 detailed process is as follows：

The decline of the distortion performance that proposition brings erroneous decision is set up based on minimum risk as bayes risk function Bayes decision-making model, the distortion performance brought with reducing erroneous decision is lost, complete rate distortion costs function such as following formula institute Show：

J_FRD=(SSE_luma+ω_chroma×SSE_chroma)+λ_mode×B_mode

Wherein, SSE_luma, represent the pixel difference error sum of squares between the original picture block and reconstruction image block of brightness；SSE_chroma Represent the pixel difference error sum of squares between the original picture block and reconstruction image block of carrier chrominance signal, ω_chromaIt is chromatic component Weight coefficient, λ_modeIt is Lagrange's multiplier, B_modeRequired bit number when being to current coding mode；

For 64 × 64, the CU of 32 × 32,16 × 16 sizes is trained by off-line learning obtain loss matrix C respectively, loses square Battle array is defined as follows shown in formula：

$C=\left(\begin{array}{cc}{C}_{NN}& C_{NP}\\ C_{PN}& C_{PP}\end{array}\right)$

C herein_PNAnd C_NPIt is defined as follows：

$C_{PN}=\frac{J_{FRD,P}-J_{FRD,N}}{J_{FRD,N}}$

$C_{NP}=\frac{J_{FRD,N}-J_{FRD,P}}{J_{FRD,P}}$

J_{FRD, N}And J_{FRD, P}It is respectively that current CU to be divided does not continue to divide rate completely corresponding with continuing to be divided into four sub-blocks and loses True cost；C_NPRepresent and assume ω_NIt is true, i.e., judges to assume ω when rate distorted result is " not dividing "_PSet up, i.e., CU is taken The value of risk brought during " division " decision-making；C_PNRepresent and assume ω_PIt is true, takes CU what is brought during " not dividing " decision-making Value of risk；C_NPAnd C_PNMore than or equal to 0, (C when decision-making is correct_NNOr C_PP), the penalty values of decision-making are 0；

Prior probability P (ω_R) it is by counting under current depth, making ω_RWhat the proportion that the CU of decision-making accounts for all CU was obtained, i.e.,

P(ω_R)=current depth belongs to ω_RCU numbers/current depth CU sum；

Wherein R is N or P.

3. a kind of HEVC blocks based on Bayesian decision according to claim 2 divide fast method；It is characterized in that step Rapid 2 are implemented as follows：

The severe degree of video motion is weighed with average gray difference value, i.e., whether occurrence scene is converted, adjacent in video sequence The average gray difference value of pixel of two frames in same position is calculated as follows shown in formula：

$G=\frac{1}{P\&CenterDot;Q}\underset{i=0}{\overset{P-1}{\&Sigma;}}\underset{j=0}{\overset{Q-1}{\&Sigma;}}|G_n(i,j)-G_{n-1}(i,j)|$

In formula, P, Q are picture traverse and height, G_n(i, j) is the gray value at n-th frame pixel (i, j) place, G_n-1(i, j) is the The gray value at n-1 frame pixel (i, j) place；If without scene switching, G is generally smaller；If occurrence scene is converted, G is generally anxious Increase severely big；

If G is more than detection threshold value, then it is assumed that scene is converted, otherwise it is assumed that scene is not changed in, i.e.,

λ_scIt is detection threshold value, its value is often empirical value, for judging whether that occurrence scene is converted；Detection threshold value is set to 9, that is, work as G> Think that occurrence scene is converted when 9, otherwise converted without occurrence scene.

4. a kind of HEVC blocks based on Bayesian decision according to claim 3 divide fast method；It is characterized in that step Rapid 3 concrete implementation process is as follows：

If the observation information of each CU n dimensional feature vector X=[x₁, x₂, x₃…x_n] describe, p (X | ω_R), R ∈ { N, P } are ω_RThe conditional probability density of feature vector, X is observed under state；State ω_RPosterior probability p (ω_R| X) can be by Bayesian formula Try to achieve

$p\left({\mathrm{\&omega;}}_R\right|X)=\frac{p\left(X\right|{\mathrm{\&omega;}}_R)P\left({\mathrm{\&omega;}}_R\right)}{P\left(x\right)},R\&Element;\{N,P\}$

Introduce loss function C_RQ, represent that rate-distortion optimization result is decision-making R, mistakenly classify as the risk in the case of decision-making Q；Examine Consider the Bayes rule in the case of risk to be changed into：

R_N(ω_N| X)=C_NNp(X|ω_N)P(ω_N)+C_PNp(X|ω_P)P(ω_P)

R_P(ω_P| X)=C_PPp(X|ω_P)P(ω_P)+C_NPp(X|ω_N)P(ω_N)

R_R(X) it is conditional risk, illustrates characteristic value X and be judged to ω_RThe average lost during class, C_NPRepresent and assume ω_NBe it is true, i.e., Judge to assume ω when rate distorted result is " not dividing "_PSet up, i.e., take CU the value of risk brought during " division " decision-making；

In order to ensure that expected risk is minimum, when must ensure to do each decision-making, all cause that conditional risk is minimum, such decision-making quilt Referred to as Least risk Bayes decision-making, the rule of decision-making is as follows：If R_N(ω_N|X)<R_P(ω_P| X), then ω=ω_N, decision-making is ω_N, represent the decision-making that " not dividing " is taken current CU；The decision-making of " division " is otherwise taken current CU；

3-1. feature extractions and selection

Characteristic vector J is extracted from observation_interAnd J_intra, obtained by following formula：

J_LRD=SAD+ λ_pred×B_pred

J_inter=argminP ∈ Inter { J_LRD(P)}

J_intra=argminP ∈ Intra { J_LRD(P)}

λ herein_predIt is Lagrange's multiplier, B_predTo all information under current prediction mode；Required bit when being encoded Number, P represents that the candidate of PU divides；

Feature vector, X is obtained from the picture in on-line study stage, if training picture is I frames, the dimension of characteristic vector just drops To 1, X=J_intra, because I frames can only use intra prediction mode；

The probability density function of 3-2. on-line studies

According to R_N(ω_N| X) and R_P(ω_P| X) value determine whether to divide current CU to be divided；By Least risk Bayes formula Understand, to compare R_R(ω_R| X) size, it is necessary to first try to achieve p (X | ω_N)P(ω_N) and p (X | ω_P)P(ω_P) value；

Each is divided in the case of " division " or " not dividing " of depth, and likelihood function all obeys dimensional Gaussian random distribution, i.e. p (X|ω_R) by two Gaussian probability-density function weighted sum approximate representations of independent probability, therefore have：

$\begin{array}{c}p\left(X\right|{\mathrm{\&omega;}}_i,{\mathrm{\&theta;}}_i)=\frac{1}{\sqrt{2{\mathrm{\&pi;\&sigma;}}_i}}\exp (-\frac{1}{2{\mathrm{\&sigma;}}_i^2}{\left(X-{\mathrm{\&mu;}}_i\right)}^2)\\ i=1,2\end{array}$

Wherein i is i-th single Gauss model in mixed Gauss model；The parameter of each Gauss model is θ_i=(μ_i, σ_i), μ_iAnd σ_i Represent class ω_iCharacteristic vector mean vector and variance, the mixing probability function of feature vector, X is

$p\left(X\right|\mathrm{\&Theta;},P)=\underset{i=1}{\overset{2}{\&Sigma;}}P\left({\mathrm{\&omega;}}_i\right)p\left(X\right|{\mathrm{\&omega;}}_i,{\mathrm{\&theta;}}_i)$

Wherein Θ=(θ₁,θ₂) be mixed Gauss model distributed constant collection, P=(p (ω₁),p(ω₂)) it is corresponding Gaussian mode The prior probability collection of type；The parameter of mixed Gauss model is determined by the EM algorithms with K-Means algorithm initializations.