CN103546749B - Method for optimizing HEVC (high efficiency video coding) residual coding by using residual coefficient distribution features and bayes theorem - Google Patents

Abstract

The invention relates to a method for optimizing residual coding by using residual coefficient distribution features and bayes theorem. The method comprises the following steps of (1) reading each frame of images in a video sequence according to a sequence set in a control file; (2) performing intra-frame prediction and inter-frame prediction on the brightness and the chromatic value of each frame of images so as to obtain a residual coefficient; (3) performing texture judgment or bayes model judgment on a residual block TU (the maximum is 32*32) according to the residual coefficient so as to judge whether block encoding is required to be stopped ahead of time or not; (4) performing DCT (discrete cosine transform) and quantification on the whole residual block TU so as to obtain a quantification parameter; and (5) performing entropy encoding on the quantification coefficient, and finally outputting the quantification coefficient in a bit stream mode. By the method, the video encoding speed is increased under the condition that the loss on the encoding quality can be ignored, and videos can be conveniently acquired in real time.

Description

Optimize HEVC residual codings using residual error coefficient distribution characteristicss and Bayes theorem Method

Technical field

The present invention relates to high-resolution video coding techniques field, in particular with residual error coefficient distribution characteristicss and Bayes The method that theorem optimizes HEVC residual codings, it is adaptable to which high-resolution video is encoded and real time video collection.

Background technology

In recent years due to the fast development of multimedia technology, HDTV（HDTV）, 3D three-dimensional video-frequencies, video communication etc. are new Technology is gradually well known, and people have been brought into a brand-new video epoch.All these technologies are given people in offer Multimodal various high-quality video while enjoy, also video coding technique is put forward higher requirement.Wherein, one most Important feature is exactly will to process more huge data volume compared to SD two-dimensional video these new techniques before, existing Video encoding standard as H.264 unable to do what one wishes, then the application of video encoding standard HEVC with more high compression efficiency and It is raw.Many new technologies are employed to improve code efficiency HEVC, such as bigger encoding block, circular recursion coding structure is more Intra prediction mode etc., these are improved while code efficiency is improved, also so that whole algorithm becomes considerably complicated, very not Practical application is put into beneficial to which.For this purpose, how to be effectively reduced HEVC encoder complexities becomes a study hotspot now.

Infra-frame prediction, inter prediction occupy most of the time in whole video coding process, become main research Direction, and method has tended to ripe.And also take including the residual coding of integer transform, quantization and inverse transformation and inverse quantization Many scramble times, some scholars are studied to which.Wherein, document [1] is analyzing HEVC residual coding block TU （Transform Unit）Tree structure and circular recursion basis of coding on propose it is a kind of simplicity TU stop scheme in advance. Nonzero coefficient number in counting quantization parameter of the residual block after integer transform and quantization, decides whether to stop Block encoding.The scramble time that the method is saved is a lot, but coding quality is lost also than larger.

The HEVC residual coding optimized algorithms that document [2] is proposed are fairly perfect.Author analyzes optimal coding mode first Criterion：Rate distortion costs function, then finds out the mutual relation of monoblock TU and its four sub-blocks TU on coding bit rate, And analyze mutual derivation relation of the stand growth model type between monoblock TU and its sub-block.Lost by the rate for contrasting tetra- sub-blocks of TU True cost and the size with predetermined threshold value, realization are skipped bulk residual coding and stop fritter residual coding.

Document [3] is proposed with regard to frame in respectively for the residual coding difference of HEVC infra-frame predictions and inter prediction The fast coding algorithm of prediction and the quick residual coding algorithm with regard to inter prediction.The former is original in identifying code HM2.0 The intra prediction mode number of in infra-frame prediction participation rate distortion computation is reduced on the basis of scheme further, but to each pattern The residual tree search of limit is carried out all, so as to also ensure that coding quality on the basis of drop few scramble time.The latter looks for first The maximum gone out in the parameter coefficient of four sub- TU and minima simultaneously make them poor, and then the difference is compared with given threshold So as to decide whether to stop piecemeal residual coding in advance.

Document [1]: Kiho Choi and Euee S.Jang, “Early TU decision method for fast video encoding in high efficiency video coding,”ELECTRONICS LETTERS, Vol. 48, No. 12, 7th June,2012.

Document [2]: Su-Wei Teng, Hsueh-Ming Hang and Yi-Fu Chen, “Fast Mode Decision Algorithm for Residual Quadtree Coding in HEVC,” IEEE Visual Communications and Image Processing.VCIP.2011.6116062, pp. 1-4, 2011.

Document [3]: Yih Han Tan, Chuohao Yeo, Hui Li Tan and Zhengguo Li, “On Residual Quad-tree Coding In HEVC,” IEEE Multimedia Signal Processing, 13th International Workshop. MMSP.2011.6093805, pp. 1–4, 2011.

The content of the invention

The purpose of the present invention is for the distinctive residual coding structures of HEVC and technological deficiency, there is provided a kind of to utilize residual error system The method that number branch's feature and Bayes theorem optimize HEVC residual codings, in the case where ensureing that subjective quality is constant, can have Improve Video coding speed in effect ground.To reach above-mentioned purpose, idea of the invention is that：First, ensureing video encoding quality change On the premise of changing less, using the gaussian distribution characteristic of residual error coefficient, two sub-blocks upper and lower to residual block TU are sub with left and right two Block carries out hypothesis testing twice, so as to judge the inner vein characteristic of current residual block TU, residual with the sub-block for reducing unnecessary Difference coding.Second, using Bayes decision model, subblock coding is stopped in advance to TU so that coding rate is improved.

Conceived according to foregoing invention, the present invention is adopted the following technical scheme that：

A kind of method that utilization residual error coefficient distribution characteristicss and Bayes theorem optimize HEVC residual codings, it is characterised in that Operating procedure is as follows：

（1）Input video sequence：Order according to setting in control file reads each two field picture of video sequence,

（2）Predict within the frame/frames：Brightness and chromatic value to every two field picture carries out infra-frame prediction and inter prediction, so as to Try to achieve residual error coefficient,

（3）TU stops to judge in advance：According to residual error coefficient to residual block TU（It is 32 × 32 to the maximum）Carry out texture judgement or Bayesian model judgement is carried out, so as to decide whether to stop block encoding in advance,

（4）Dct transform and quantization：Dct transform and quantization are carried out to monoblock TU, so as to try to achieve quantization parameter,

（5）Entropy code：Entropy code is carried out to quantization parameter, is finally exported in the form of bit stream.

The present invention is compared with the prior art compared with obviously prominent substantive distinguishing features and notable technology are entered as follows Step：

1）, this HEVC video encoding optimizations method ensure video encoding quality it is constant while so that cataloged procedure exists Residual coding this subprocess can just improve coding rate, and the residual coding time that can be saved in experiment is up to 60%；

2）, the optimized algorithm in this HEVC video encoding optimization schemes with regard to residual error coefficient distribution character be based on the assumption that inspection Survey, it is possible to realize the compromise between coding quality and coding rate by changing significant level according to the actual requirements；

3）, the optimized algorithm in this HEVC video encoding optimization schemes based on Bayes decision model choose forecast error and Mean square deviation MAD of residual error coefficient is characterized the vectorial factor and carries out mode decision, and the two values in frame in and inter prediction just Obtain, so the optimization method will not extraly increase the scramble time.

Description of the drawings

Fig. 1 is the method that utilization residual error coefficient distribution characteristicss and Bayes theorem in the present invention optimize HEVC residual codings Theory diagram.

Fig. 2 is the block diagram of Skip, interframe and infra-frame prediction.

Fig. 3 is based on residual error coefficient distribution characteristicss algorithm block diagram.

Fig. 4 is the structured flowchart of TU.

Fig. 5 is the TU types that can be used for residual coding in different CU, PU and interframe and infra-frame prediction.

Fig. 6 is entropy code schematic block diagram.

Fig. 7 a be resolution be 832 × 480 cycle testss RaceHorses based on residual error coefficient distribution characteristicss RD curve charts under optimized algorithm.

Fig. 7 b are that the cycle testss vidvo1 that resolution is 1280 × 720 is calculated in the optimization based on residual error coefficient distribution characteristicss The RD curve charts of FAXIA.

Fig. 8 is the experimental result based on residual error coefficient distribution characteristicss optimized algorithm compared with original method in HM5.0, main Parameter is wanted to include：The PSNR of brightness, bit rate and residual coding time.

Fig. 9 a be resolution be 832 × 480 cycle testss RaceHorses based on the excellent of Bayes decision model Change the RD curve charts under algorithm.

Fig. 9 b be resolution be 1280 × 720 cycle testss vidvo1 in the optimized algorithm based on Bayes decision model Under RD curve charts.

Figure 10 is the experimental result based on Bayes decision model optimization algorithm compared with original method in HM5.0, main Parameter is wanted to include：The PSNR of brightness, bit rate and residual coding time.

Specific embodiment

The preferred embodiments of the present invention are described in further detail below in conjunction with accompanying drawing：

Embodiment one：

Optimize method (referring to Fig. 1) bag of HEVC residual codings using the distribution characteristicss and Bayes theorem of residual block coefficient Include following steps：

（1）Input video sequence：Order according to setting in control file reads each two field picture of video sequence,

（2）Within the frame/frames：Brightness and chromatic value to every two field picture carries out infra-frame prediction and inter prediction, so as to try to achieve Residual error

Coefficient,

（3）TU stops to judge in advance：According to residual error coefficient to residual block TU（It is 32 × 32 to the maximum）Carry out texture judgement or

Bayesian model judgement is carried out, so as to decide whether to stop block encoding in advance,

（4）Dct transform and quantization：Dct transform and quantization are carried out to monoblock TU, so as to try to achieve quantization parameter,

（5）Entropy code：Entropy code is carried out to quantization parameter, is finally exported in the form of bit stream.

Embodiment two：The present embodiment is essentially identical with embodiment one, and special feature is as follows：（See Fig. 2 to Figure 10）

Above-mentioned steps (2) are to carry out frame in and inter prediction to input video sequence, and referring to Fig. 2, which comprises the following steps that：

（2-1）Infra-frame prediction is using around encoding block（The left side and top）Encoded reference pixel carries out pre- to current block Survey so as to eliminate video image redundancy spatially；Inter prediction is carried out to which by reference to frame before and after the frame of encoding block place Motion estimation and compensation is so as to eliminating the time redundancy of video sequence.HEVC is according to the optimum criterion of rate distortion from many The frame in and inter-frame forecast mode of optimum are selected in individual frame in and interframe candidate pattern.Rate distortion costs function is：

（1）

WhereinFor rate distortion costs value,For predicted distortion value,It is the ratio exported under different predictive modes Special number,For LaGrange parameter.

（2-2）It is H.264 different from video encoding standard before, it is to improve code efficiency, HEVC adopts bigger coding Block（64×64）, the coded system of circular recursion, intra prediction mode increase to 35 kinds, and these improve and are improving code efficiency Also frame in and inter prediction is caused to become increasingly complex simultaneously.

（2-3）If input is I frames, only infra-frame prediction is carried out to which and rate-distortion optimization is carried out；If P frames or B frames, Skip model predictions and inter prediction are carried out to which first then, infra-frame prediction is then carried out again.Finally press rate-distortion optimization standard from The predictive mode of optimum is selected in Three models.

Above-mentioned steps (3) have two kinds to the method judged by the coefficient of residual block, comprise the following steps that：

（3-1）Hypothesis testing method based on Gauss distribution

Studied from pertinent literature, residual error coefficient obeys the Gauss distribution that expected value is zero, that is, meet following formula：

（2）

WhereinFor average,For variance.

To judge that monoblock TU can be divided to monoblock TU for two kinds of dividing modes, such as Fig. 3 if appropriate for four sub- TU are divided into It is shown.If two kinds of dividing modes all meet, it is believed that TU is adapted to be divided into four sub- TU.If using the division side in Fig. 3 Formula encoding efficiency preferably, then residual error coefficient should Gaussian distributed, the expected value due to obeying same distribution then two partsShould no difference.Therefore, we can be with by assuming that the prediction effect of the test and judge pattern, according to theory of probability Knowledge obtain：

（3）

In above formulaXue Shengshi distributions are distributed as,It is the degree of freedom of the distribution,,It is every kind of dividing mode The average of middle two parts pixel intensity,、It is the respective mathematical expectation in two parts, in above formulaIt is as follows：

（4）

It will be apparent that 32 × 32 macro blocks are divided into equal two part, then the number of pixels of two parts should phase Deng that is,：.Should be met according to hypothesis testing, while taking significance level=0.05, looking into t-distribution table can obtain, then we can obtain as drawn a conclusion：

（5）

Above formula is equivalent to：

（6）

Above formula is the TU calculating for 32 × 32.In HEVC residual codings, TU adopts the quadtree coding similar with CU Structure, i.e., will also carry out traversal formula coding to its sub-block 16 × 16,8 × 8 and 4 × 4 except 32 × 32, then from all patterns In select optimization model, be that we can carry out the process similar with 32 × 32 to 16 × 16 and 8 × 8 TU this, so as to obtain

（7）

（8）

Above three formula is the Rule of judgment of 32 × 32,16 × 16 and 8 × 8 pieces of calculating respectively.If meeting formula（6）Can There is no significant difference with this kind of division thought in 32 × 32TU, if meeting formula（7）、（8）Then it is considered that 16 × 16,8 × This kind of division in 8TU does not have significant difference.Residual error coefficient is relevant with coding mode, simple in order to what is calculated, and we are only.It is based on More than analyze, set forth herein process step it is as follows：

1）First, residual coding is carried out using monoblock TU to 32 × 32.

2）When according to two kinds of dividing mode all no significant differences shown in Fig. 3,3 are redirected）；Otherwise, TU depth adds 1, and return 1）.

3）Piecemeal residual coding is carried out to monoblock TU.

（3-2）Decision algorithm based on bayesian theory

The TU of a certain size carries out block encoding and non-block encoding is event that two pieces is opposed completely, by TU block encodings This event definition is, and TU does not carry out block encoding and is defined as.The vector characteristics of TUIt is used for Help improves the accuracy of classification.Be TU characteristic vectors be F when be divided intoThe posterior probability of class.In residual error In cataloged procedure, if making the decision for making mistake, i.e., should block encoding actually do not have a block encoding, and original should But the block encoding of the piecemeal, this will cause rate distortion to lose.We should piecemeal and actual do not have piecemeal to cause Loss be labeled as, and loss that originally should not be caused by piecemeal actually piecemeal is labeled as.Between them With following relation：

（9）

（10）

Wherein,WithIt is the rate distortion costs produced when TU carries out block encoding and non-block encoding.By Above-mentioned formula understood when TU makes correct selection, i.e., should block encoding in fact also piecemeal or should not piecemeal reality Do not have really piecemeal on border, will not result in any rate distortion loss yet, therefore.The analysis based on more than can obtain artificial situationWhen Bayes risk cost：

（11）

（12）

When<When, selectThe cost for causing is less, should stop to carry out block encoding to TU. And work as>When, selectThe cost for causing is less, should carry out block encoding to TU according to original algorithm. In above formula：

（13）

WhereinWhen representing that TU carries out Splite and None-Splite codings, the probability density of its characteristic vector F Distribution function,It is the priori probability density function of situation Splite and None-Splite.By formula（11）、（12）With（13） The judgement formula that whether block encoding is carried out to TU can be obtained as follows：

（14）

For the probability distribution for effectively predicting, herein by mean square deviation MAD of residual block coefficient and and current TU blocks Forecast error as vector characteristics F key element.It is general that the Non-parameter density estimation method proposed according to D.Chai et al. tries to achieve condition Rate density functionAnd put in a look-up table.To reduce statistical work amount, can be by two characteristic vector key element amounts 10 deciles are melted into, so whole characteristic vector F will be divided into 100 scales.And（14）In formula in the decision threshold of the inequality left side 's、、WithIt is relevant with the depth of the resolution of video, QP sizes and TU.Therefore, herein in advance The video sequence of one group of different resolution is counted under different QP, so as to obtain different point variabilities, different Q P and Decision threshold under different TU depth, and these threshold values are placed in another inquiry table.

In sum, following step is included based on the TU type decision methods of bayesian theory：

1）It is right（14）Decision threshold in formula on the right of inequality is initialized, so as to obtain can to video resolution and

The decision threshold of QP adaptive changes.

2）To a TU block acting in accordance with YIN YANG changes in four seasons node（I.e. 32 × 32）Start, residual coding is carried out to monoblock TU.Obtain coding to miss The MAD. of difference and residual error coefficient

3）Tabled look-up by the MAD of monoblock encoding error and residual error coefficient and find out the vector characteristics F probability-distribution functions of TUWith.Press（14）Formula is judged, if inequality meets requiring, is jumped to（4）.Otherwise, TU is divided into four Identical sub- TU, depth Depth add 1, go to（2）.

4）Next TU is equally processed by progressive scan order.

Above-mentioned steps (4) carry out integer transform to residual error data（DCT）And quantization, which comprises the following steps that：

(4-1) HEVC carry out integer transform and quantify when TU coding structure as shown in figure 4, TU full-size be 32 × 32, minimum dimension is 4 × 4, similar to the quad-tree structure of CU.

(4-2) for different CU, PU, frame in and inter prediction, TU has different available types, as shown in Figure 5.Its In, red italic represents available TU during infra-frame prediction, and listed all patterns all can be used in inter prediction in table.

Above-mentioned steps（5）Quantization parameter to obtaining through residual coding carries out entropy code, ultimately forms binary bits Stream with

It is easy to transmit in a network.Referring to Fig. 7, which comprises the following steps that：

（5-1）Variable entropy coding is carried out to quantization parameter（VLC）Or arithmetic coding (CABAC), quantify system so as to eliminate Several symbol redundancies, realizes the further compression to video sequence.

（5-2）Entropy-encoded data is finally exported in the form of bit stream.It is capable of achieving by related rate control techniques The adaptively changing of bit rate, which greatly enhances the network friendliness of HEVC encoders.

For checking set forth herein two kinds based on residual error coefficient distribution character and based on bayesian theory quick residual blocks Type is sentenced

Determine the effectiveness of method, below to emulation experiment having been carried out to a large amount of cycle testss.Experiment porch (PC) is configured For：Intel Core2 Duo CPU, 2.53 GHz, 1G Internal Memory, Windows XP Operation System；Criteria check model is HEVC standard identifying code HM 5.0;Mode is encoded separately using YC（It is only right herein Brightness is optimized）；Emulation experiment to resolution for 416 × 240,832 × 480,1280 × 720 and 1920 × 1080 survey Front 20 frame of examination sequence is counted, and wherein BASIC QP are set to 22,27,32,37；The pre- geodesic structure of encoding and decoding selects HHI- IPPP。

Method based on residual error coefficient distribution character（Proposed A）Experimental result is as shown in Fig. 7 a～7b, Fig. 8.Figure 7a, 7b are that video sequence PartyScene and vidyo1 that resolution is 832 × 480 and 1280 × 720 are original in HM5.0 respectively RD curve charts under two methods of residual coding method and Proposed A.It can be seen that in Fig. 7 a and Fig. 7 b two curves are almost It is completely superposed, the coding quality in this explanation present invention based on the method for parameter Coefficients Distribution is almost original with HM5.0 Method is identical.Fig. 8 is that two methods of HEVC raw residuals coding and Proposed A are differentiated to different in the case of different Q P The result tested by the video sequence of rate, refers mainly to indicate DPSNR, DBR and DTime, is respectively customized for：

(15)

(16)

(17)

Wherein, PSNR is Y-PSNR, and BR is video code bit rate, and Time is the residual coding time.It can be seen that In the case where ensureing that coding quality is constant（Coding bit rate averagely reduces by 0.15%, and the PSNR of brightness is averagely reduced 0.024dB）, the present invention in optimization method residual coding time average reduce about 42%.

Quick residual coding method based on bayesian theory（Proposed B）Experimental result such as Fig. 9 a～9b, Tu10Suo Show.Fig. 9 a, 9b are that the video sequence PartyScene and vidyo1 that resolution is 832 × 480 and 1280 × 720 exists respectively RD curve charts under two methods of the coded method of HM5.0 raw residuals and Proposed B.It can be seen that two in Fig. 9 a and Fig. 9 b Bar curve is almost completely superposed, this explanation the present invention in based on Bayes decision theory method coding quality almost with In HM5.0, original method is identical.Figure 10 is the situation of two methods of HEVC raw residuals coding and Proposed B in different Q P Under result that the video sequence of different resolution is tested, it is just the same in leading indicator and Fig. 8.It can be seen that keeping In the case that coding quality is constant（Coding bit rate averagely increases by 0.3%, and the PSNR of brightness averagely reduces 0.02dB）, the present invention In optimization method residual coding time average reduce about 40%.

The present invention can be seen that by hypothesis testing being carried out to residual block coefficient by above-mentioned each chart, and utilize pattra leaves This theory is stopped to judge to residual reference block TU in advance, so as to reduce the scramble time of dct transform and quantizing process, in coding In the case of mass loss is negligible, the time of HEVC residual codings is significantly reduced.

Claims (5)

1. a kind of method that utilization residual error coefficient distribution characteristicss and Bayes theorem optimize HEVC residual codings, it is characterised in that：

(1) input video sequence：Order according to setting in control file reads each two field picture of video sequence；

(2) predict within the frame/frames：Brightness and chromatic value to every two field picture carries out infra-frame prediction and inter prediction, so as to try to achieve Residual error coefficient；

(3) TU stops to judge in advance：Texture judgement is carried out to the residual block TU for being 32 × 32 to the maximum according to residual error coefficient or is carried out Bayesian model judges, so as to decide whether to stop block encoding in advance；

(4) dct transform and quantization：Dct transform and quantization are carried out to monoblock TU, so as to try to achieve quantization parameter；

(5) entropy code：Entropy code is carried out to quantization parameter, is finally exported in the form of bit stream.

2. utilization residual error coefficient distribution characteristicss according to claim 1 and Bayes theorem optimize the side of HEVC residual codings Method, it is characterised in that what the step (2) was predicted within the frame/frames comprises the following steps that：

(2-1) infra-frame prediction is current block to be predicted so as to eliminate using the encoded reference pixel in the encoding block left side and top Video image redundancy spatially；Inter prediction by reference to frame before and after the frame of encoding block place which is carried out estimation and Motion compensation is so as to eliminating the time redundancy of video sequence；HEVC is according to the optimum criterion of rate distortion from multiple frame ins and frame Between the frame in and inter-frame forecast mode of optimum are selected in candidate pattern；Rate distortion costs function is：

Cost_RD=D+ λ R (1)

Wherein Cost_RDFor rate distortion costs value, D is predicted distortion value, and R is the bit number exported under different predictive modes, and λ is LaGrange parameter；

(2-2) it is H.264 different from video encoding standard before, it is to improve code efficiency, HEVC adopts bigger 64 × 64 Encoding block, the coded system of circular recursion, intra prediction mode increase to 35 kinds, and these are improved while code efficiency is improved Also so that frame in and inter prediction become increasingly complex；

If (2-3) input is I frames, only infra-frame prediction is carried out to which and rate-distortion optimization is carried out；If P frames or B frames, then first Skip model predictions and inter prediction are carried out to which, infra-frame prediction is then carried out again, finally by rate-distortion optimization standard from three kinds The predictive mode of optimum is selected in pattern.

3. utilization residual error coefficient distribution characteristicss according to claim 1 and Bayes theorem optimize the side of HEVC residual codings Method, it is characterised in that step (3) TU stops comprising the following steps that for judgement in advance：

1. the hypothesis testing method based on Gauss distribution

Studied from pertinent literature, residual error coefficient obeys the Gauss distribution that expected value is zero, that is, meet following formula：

X～N (μ, σ²) (2)

Wherein X represents residual error coefficient, and μ=0 is average, σ²For variance, N represents Gauss distribution；

To judge that monoblock TU can be divided to monoblock TU for two kinds of dividing modes if appropriate for four sub- TU are divided into, if two kinds are drawn Point mode all meets, then it is believed that TU is adapted to be divided into four sub- TU, if the dividing mode encoding efficiency is preferably, residual error coefficient Should Gaussian distributed, due to obey it is same distribution then two parts expected value μ should no difference, therefore, Can be with by assuming that test and judge prediction effect, obtains according to the knowledge of theory of probability：

$\frac{(\stackrel{\&OverBar;}{X}-\stackrel{\&OverBar;}{Y})-({\mathrm{\&mu;}}_1-{\mathrm{\&mu;}}_2)}{S_w\sqrt{\frac{1}{n_1}+\frac{1}{n_2}}}~t(n_1+n_2-2)---\left(3\right)$

N in above formula (3)₁And n₂The number of residual error coefficient in difference two sub-blocks of A and B, t-distribution are that Xue Shengshi is distributed, n₁+n₂- 2 are The degree of freedom of the distribution,It is the average of two parts pixel intensity in every kind of dividing mode, μ₁、μ₂It is that two parts are respective Mathematical expectation, S_wIt is as follows：

$S_w=\sqrt{\frac{\underset{i=1}{\overset{n_1}{\&Sigma;}}{(X_i-\stackrel{\&OverBar;}{X})}^2+\underset{i=1}{\overset{n_2}{\&Sigma;}}{(Y_i-\stackrel{\&OverBar;}{Y})}^2}{n_1+n_2-2}}---\left(4\right)$

X in above formula (4)_iAnd Y_iThe value of residual error coefficient in difference two sub-blocks of A and B；It will be apparent that 32 × 32 macro blocks are divided into Two equal parts, then the number of pixels of two parts should be equal, i.e.,：n₁=n₂=512, should be met according to hypothesis testing, Take significance level=0.05 simultaneously, look into t-distribution table and can obtain t_0.025(510)=1.96, then obtain as drawn a conclusion：

$|\stackrel{\&OverBar;}{X}-\stackrel{\&OverBar;}{Y}|<t_{0.025}\left(1022\right)\&times;S_w/8=0.245\&times;S_w---\left(5\right)$

Above formula (5) is equivalent to：

$|\underset{i=1}{\overset{512}{\&Sigma;}}{X}_i-\underset{i=1}{\overset{512}{\&Sigma;}}{Y}_i|<125.44\&times;S_w=125.44\&times;\sqrt{\frac{\underset{i=1}{\overset{512}{\&Sigma;}}{(X_i-\stackrel{\&OverBar;}{X})}^2+\underset{i=1}{\overset{512}{\&Sigma;}}{(Y_i-\stackrel{\&OverBar;}{Y})}^2}{1022}}---\left(6\right)$

Above formula (6) is the TU calculating for 32 × 32；In HEVC residual codings, residual block TU is using similar with encoding block CU Quadtree coding structure, i.e., to also carry out traversal formula coding to its sub-block 16 × 16,8 × 8 and 4 × 4 except 32 × 32, so Optimization model is selected from all patterns afterwards, is that we can carry out the place similar with 32 × 32 to 16 × 16 and 8 × 8 TU this Reason, so as to obtain

$|\underset{i=1}{\overset{128}{\&Sigma;}}{X}_i-\underset{i=1}{\overset{128}{\&Sigma;}}{Y}_i|<31.36\&times;S_w=31.36\&times;\sqrt{\frac{\underset{i=1}{\overset{128}{\&Sigma;}}{(X_i-\stackrel{\&OverBar;}{X})}^2+\underset{i=1}{\overset{128}{\&Sigma;}}{(Y_i-\stackrel{\&OverBar;}{Y})}^2}{254}}---\left(7\right)$

$|\underset{i=1}{\overset{32}{\&Sigma;}}{X}_i-\underset{i=1}{\overset{32}{\&Sigma;}}{Y}_i|<15.92\&times;S_w=15.92\&times;\sqrt{\frac{\underset{i=1}{\overset{32}{\&Sigma;}}{(X_i-\stackrel{\&OverBar;}{X})}^2+\underset{i=1}{\overset{32}{\&Sigma;}}{(Y_i-\stackrel{\&OverBar;}{Y})}^2}{62}}---\left(8\right)$

Above-mentioned (6), (7), (8) three formulas are the Rule of judgment of 32 × 32,16 × 16 and 8 × 8 pieces of calculating respectively；If met Formula (6) it is considered that 32 × 32TU in this kind of division there is no significant difference, if formula (7), (8) are met can consider 16 × 16th, in 8 × 8TU this kind of division does not have significant difference；Residual error coefficient is relevant with coding mode, simple in order to what is calculated, is based on The process step of analysis is as follows above：

1) residual coding is carried out using monoblock TU to 32 × 32 first,；

2) when two kinds of dividing mode all no significant differences based on residual error coefficient branch characteristics algorithm, redirect 3)；Otherwise, TU Depth adds 1, and returns 1)；

3) piecemeal residual coding is carried out to monoblock TU；

2. the decision algorithm based on bayesian theory：It is that two pieces is complete that the TU of a certain size carries out block encoding and non-block encoding The event of opposition, by TU block encodings, this event definition is w_S, and TU does not carry out block encoding and is defined as w_N；The vector of TU is special Levy F={ f₁,f₂,f₃, for helping improve the accuracy of classification, wherein f₁,f₂,f₃Deng expression, whether piecemeal is closed with TU It is some close eigenvalues；P(w_i| F) be TU characteristic vectors be F when be divided into w_iThe posterior probability of class；In residual coding mistake Cheng Zhong, if making the decision for making mistake, i.e., should block encoding actually do not have a block encoding, and it is original should not piecemeal But block encoding, this will cause rate distortion lose；We are should piecemeal and the actual loss for not having piecemeal to cause It is labeled as C_ns, and loss that originally should not be caused by piecemeal actually piecemeal is labeled as C_sn；Have between them following Relation：

C_s,n=(RD_S-RD_N)/RD_N (9)

C_n,s=(RD_N-RD_S)/RD_S (10)

Wherein, RD_sAnd RD_nIt is the rate distortion costs produced when TU carries out block encoding and non-block encoding；By above-mentioned formula (9) understand with (10) when TU makes correct selection, i.e., should block encoding in fact also piecemeal or should not piecemeal Actually do not have really piecemeal yet, will not result in any rate distortion loss, therefore C_nnAnd C_ssIt is 0；Based on more than, analysis can be with Obtain artificial situation i, Bayes risk cost R (w during i ∈ (N, S)_i|F)：

R(w_N| F)=C_n,nP(w_N|F)+C_n,sP(w_S| F)=C_n,sP(w_S|F) (11)

R(w_S| F)=C_s,sP(w_S|F)+C_s,nP(w_N| F)=C_s,nP(w_N|F) (12)

As R (w_N|F)<R(w_S| F) when, select w_NThe cost for causing is less, should stop to carry out block encoding to TU；And work as R (w_N|F) >R(w_S| F) when, select w_SThe cost for causing is less, should carry out block encoding to TU according to original algorithm；In above formula：

$R\left(w_i\right|F)=\frac{P\left(F\right|w_i)P\left(w_i\right)}{P\left(F\right)},i\&Element;\{N,S\}---\left(13\right)$

Wherein P (F | w_i) represent TU when carrying out piecemeal and non-block encoding, the probability density function of its characteristic vector F, P (w_i) it is situation piecemeal and the not priori probability density function of piecemeal；Whether can be obtained to TU by formula (11), (12) and (13) The judgement formula for carrying out block encoding is as follows：

For the probability distribution for effectively predicting, herein by mean square deviation MAD of residual block coefficient and and current TU blocks it is pre- Error is surveyed as the key element of vector characteristics F；It is close that the Non-parameter density estimation method proposed according to D.Chai et al. tries to achieve conditional probability Degree function P (F | w_i) and put in a look-up table；To reduce statistical work amount, two characteristic vector key elements can be quantized into 10 Decile, so whole characteristic vector F will be divided into 100 scales；And the C in (14) formula in the decision threshold of the inequality left side_ns、 C_sn、p(w_N) and p (w_S) with the resolution of video, the size of quantization parameter QP (Quantization Parameter) and TU Depth it is relevant；Therefore, in advance the video sequence of one group of different resolution is counted under different QP, so as to obtain Decision threshold under different points of variabilities, different Q P and difference TU depth, and these threshold values are placed in another inquiry table；

In sum, following step is included based on the TU type decision methods of bayesian theory：

1) decision threshold on the right of inequality in (14) formula is initialized, so as to obtain can to video resolution and QP from Adapt to the decision threshold of change；

2) a TU block acting in accordance with YIN YANG changes in four seasons node 32 × 32 is started, residual coding is carried out to monoblock TU, encoding error and residual error is obtained The MAD of coefficient；

3) tabled look-up by the MAD of monoblock encoding error and residual error coefficient find out TU vector characteristics F probability-distribution function P (F | w_S) and P (F|w_N)；Judged by (14) formula, if inequality meets requiring, jumped to (4)；Otherwise, TU is divided into four identical sub- TU, depth Depth adds 1, goes to (2)；

4) next TU is equally processed by progressive scan order.

4. utilization residual error coefficient distribution characteristicss according to claim 1 and Bayes theorem optimize the side of HEVC residual codings Method, it is characterised in that：Step (4) dct transform and comprising the following steps that for quantifying：

1., when HEVC carries out integer transform and quantifies, the full-size of TU is 32 × 32, and minimum dimension is 4 × 4, similar to CU's Quad-tree structure；

2. different CU, PU, frame in and inter prediction are directed to, TU there are different available types.

5. utilization residual error coefficient distribution characteristicss according to claim 1 and Bayes theorem optimize the side of HEVC residual codings Method, it is characterised in that：Step (4) entropy code is comprised the following steps that：

1. variable entropy coding (VLC) or arithmetic coding (CABAC) are carried out to quantization parameter, so as to eliminate the symbol of quantization parameter Redundancy, realizes the further compression to video sequence；

2. entropy-encoded data is finally exported in the form of bit stream；Bit rate is capable of achieving by related rate control techniques Adaptively changing, which greatly enhances the network friendliness of HEVC encoders.