CN105721865A - Fast decision algorithm for dividing HEVC inter-frame coding unit - Google Patents

Abstract

The invention provides a fast decision algorithm for dividing a HEVC inter-frame coding unit. Whether division of a current CU is stopped or not is judged by judging whether the optimal mode of the current CU is a SKIP coding mode or not and utilizing coding information of adjacent CUs; according to division probability statistical information of different depths of the CUs, targeted strategies are respectively taken for the CUs in different depths; for the CU, the depth of which is 0, a threshold value is set by utilizing RD-cost of the adjacent CUs; if the coding mode of the current CU is the SKIP mode and the RD-cost is less than or equal to the threshold value, the CU is not continuously divided any longer; for the CU, the depth of which is 1, the threshold value is set by utilizing the depth values of the adjacent CUs; if the coding mode of the current CU is the SKIP mode and the threshold value is less than or equal to 1, the CU is not continuously divided any longer; and for the CU, the depth of which is 2, if the optimal mode is the SKIP mode, the CU is not continuously divided any longer.

Description

A kind of high-speed decision algorithm of HEVC interframe encode dividing elements

Technical field

The present invention relates to high-performance video coding (High Efficiency Video Coding, HEVC) field.

Background technology

In order to meet the demand to HD video and produce more preferable compression effectiveness, HEVC have employed quaternary tree Image partitioning scheme.This partitioning scheme can obtain ratio H.264/AVC more preferable compression efficiency, but simultaneously Due to need the coding unit to each size (Coding Unit, CU), predicting unit (Prediction Unit, And converter unit (Transform Unit, TU) will percent of pass aberration optimizing (Rate-Distortion PU) Optimization, RDO) calculate optimum size so that and the computation complexity of encoder is substantially improved. If able to the size of prediction CU and the PU pattern of CU, then some node in quaternary tree can be carried out The search of beta pruning and in advance termination PU pattern, thus can effectively reduce the computation complexity of encoder.

The size of the CU that the test software (HEVC Test Model, HM) of HEVC is used is 64 × 64, 32 × 32,16 × 16 and 8 × 8, the degree of depth of corresponding CU is 0,1,2,3 respectively.Wherein the degree of depth is 0 CU is designated as LCU (Largest Coding Unit), the degree of depth be 3 CU be designated as SCU (Smallest Coding Unit).The encoder of HM uses the mode of recurrence to divide LCU, and the CU for each layer will Carrying out RDO to determine the division of PU and TU, wherein the division of TU also uses the dividing mode of quaternary tree. The unnecessary computation complexity brought to reduce recurrence to divide, more existing schemes drawing for CU Divide and carried out a series of optimization:

Such as document 1 (sees R.H.Gweon and Y.L.Lee, Early termination of CU encoding To reduce HEVC complexity.document JCTVC-F045,2011), describe at HM 3.1 platform Under, if the optimization model of current CU is the situation of SKIP, then this CU will not continue to divide downwards Probability is 95%.This probability is enough to ensure that the accuracy of prediction, though have should continue divide CU because Select SKIP pattern not continue to divide, compile as SKIP pattern employs less code word Code and be compensated, so this algorithm will not cause the loss that BD-rate is the biggest.But this algorithm is by all of The CU of the degree of depth is unified to be considered, so still there being the space of optimization.

Document 2 (sees Xiong, J.et al, Fast HEVC Inter CU Decision Based on Latent SAD Estimation.IEEE Transactions on Multimedia, 2015:2147-2159) propose based on potential absolutely The high-speed decision algorithm that the CU of difference and (Latent Sum of Absolute Differences, LSAD) is divided. Author employs new method for estimating to obtain the SAD of two-layer CU, and defines motion compensation rate Distortion value (Motion Estimation Rate-distortion, MERD).Exponential model be used for describing LSAD and Relation between MERD, and from exponential model, derive threshold value.If the SAD of two-layer is both less than threshold value, The most do not continue to split downwards.Although the method can save the scramble time of more than 50%, but BD-rate The loss of (Bjontegaard delta bit rate) is bigger.

Document 3 (sees SangsooAhn, Bumshik Lee, and Munchurl Kim, A Novel Fast CU Encoding Scheme Based on Spatiotemporal Encoding Parameters for HEVC Inter Coding,IEEE Transactions on Circuits and System for Video Technology(CSVT), 2015:422-435) propose the high-speed decision algorithm that detection SKIP pattern and CU divide in advance.Utilize and compile The division of CU is instructed by the coefficient produced during Ma.Utilize sampling self adaptation skew (Sample Adaptive Offset, SAO) coefficient that produces determines the direction at main edge.And make use of the size of TU, The size of motion vector and code identification position assist the division determining CU.This algorithm is at Random Access The scramble time of 49.6% can be saved under the configuration of Main, but BD-rate have lost 1.4%.

Summary of the invention

It is an object of the invention to provide the high-speed decision algorithm of a kind of HEVC interframe encode dividing elements, logical Cross CU divides in interframe encode high-speed decision and the termination in advance of model selection, can guarantee that CU terminates drawing Divide the accuracy of prediction, improve the interframe encode efficiency of encoder, and not loss coding quality.

The technical scheme be given:

The present invention proposes the algorithm of interframe encode dividing elements high-speed decision improved, it is characterised in that include as Lower step:

1) after all PU pattern searches of current CU are complete, it is judged that the current CU degree of depth.If the degree of depth is 3 Divide and terminate, if the degree of depth is 0 to go to step (2), if the degree of depth is 1 to go to step (3), if the degree of depth is 2 Go to step (4).

2) if the degree of depth is 0 and optimization model is SKIP.The RDcost for prediction is calculated according to formula (1)_pre, If the RD-cost of current CU is less than RDcost_pre, then divide and terminate.Otherwise continue to divide downwards, And forward step (1) to.

Described formula (1), is for calculating the RDcost of prediction by adjacent CU_{p e}:

${\mathrm{RDcost}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{RDcost}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(1\right)$

Wherein,

RDcost_iValue be the RD-cost of ALC, AC and LC respectively.

k_iWhether the CU representing adjacent exists, if there is and selected pattern be SKIP pattern, then k_iIt is 1 to be otherwise 0.

a_iIt is the weights of 3 CU, the value of i CU each with Fig. 1 (current CU and the adjacent CU in spatial domain thereof) Sequence number corresponding.

3) if the degree of depth is 1 and optimization model is SKIP.The Depth for prediction is calculated according to formula (2)_pre, If Depth_preLess than or equal to 1, then divide and terminate.Otherwise continue to divide downwards, and forward step (1) to. Described formula (2), for the degree of depth of prediction:

${\mathrm{Depth}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{Depth}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(2\right)$

Wherein,

a_iAnd k_iValue identical with formula (1) value.

Depth_iIt is the degree of depth of ALC, AC and LC respectively.

4) if the degree of depth is 2, if the optimization model of current CU is SKIP, then termination is divided.Otherwise continue to Lower division, and forward step (1) to.

Owing to using such scheme, the invention has the beneficial effects as follows:

1. the present invention terminates the difference of the predictablity rate divided according to different depth CU based on SKIP coding mode Feature, the CU that the degree of depth is 0 and 1 is devised different CU divide in advance decision method to ensure CU terminates dividing the accuracy of prediction.

2. the present invention considers the feature of coding information of CU adjacent block, can utilize CU adjacent block accurately Coding information limit the division of CU and terminate, such that it is able in the situation of loss coding quality hardly Under, it is effectively improved the interframe encode efficiency of encoder.

Accompanying drawing explanation

The adjacent CU exemplary plot of the current CU of Fig. 1 and spatial domain thereof

How Fig. 2 obtains adjacent C U exemplary plot from current CU

Fig. 3 interframe encode dividing elements high-speed decision algorithm flow chart

Detailed description of the invention

Below by way of some embodiment supports and expansion technical solution of the present invention.

Embodiment

The present invention first consider when the optimization model of CU be SKIP pattern be the division i.e. terminating CU time different The predictablity rate of degree of depth CU.The present invention have selected 3 representative standard video sequence (its titles Be respectively BQMall, FourPeople and BasketballDrive) analyze CU segmentation distribution, BQMall Having abundant texture, FourPeople video content motion ratio is shallower and texture smooths, The video content of BasketballDrive moves the most violent, Texture complication between BQMall and Between FourPeople.Table 1 give on HM 16.4 platform test the optimization model of proper CU be The result of the predictablity rate of different depth CU when SKIP pattern is the division i.e. terminating CU, wherein QP is Referring to quantization step (Quantization Step, QP), LD refers to that test configurations is Low Delay main, RA Refer to that test configurations is Random Access Main.If the degree of depth is the CU selection of 2 as can be seen from Table 1 If SKIP pattern, the probability of 97.1%～99.9% is had to will not continue to divide.It is the CU of 2 for the degree of depth, in advance Survey accuracy rate and be little affected by the impact of QP.The degree of depth is that the CU probability that will not continue to divide downwards of 0 is 73.3%～97.7%, it can be seen that the division proportion of the CU of this degree of depth is affected bigger by QP.Work as QP especially When being 22, the predictablity rate of BasketballDrive sequence is only 73.3%.The degree of depth is the CU continuation stroke of 1 The probability divided is 92.6%～99.5%, and the predictablity rate of this degree of depth is the highest, but not as good as the degree of depth is still The CU of 2.So the loss on the coding quality i.e. loss of BD-rate of this algorithm can be speculated mainly Prediction from the CU that the degree of depth is 0 and 1.So can terminate dividing by limiting the CU of the degree of depth 0 and 1 Condition improve predictablity rate.

Table 1. when the optimization model of CU be SKIP pattern i.e. terminate CU division time different depth CU predictablity rate

For reaching above-mentioned purpose, the solution of the present invention is: choose upper left side CU (Above Left CU, ALC), the CU (Above CU, AC) of upside and the CU (Left CU, LC) in left side predicts currently Whether CU (Current CU) continues to divide in the case of optimization model is SKIP.Fig. 1 be current CU with The schematic diagram of the position relationship of its adjacent C U.When choosing CU, owing to may be simultaneously present in the same direction Multiple CU, invention provides for the position of adjacent C U that this algorithm is used.It is concrete that adjacent C U obtains Way is first to obtain the block ALB that upper left side 4 × 4 in current CU is little_currWith 4 × 4 fritters of upper right side ARB_currThe coordinate of ' Z ' scanning sequency in LCU, then according to ALB_currCoordinate derive Fig. 2 Shown size is the Block of 4 × 4_ALCAnd Block_LCThe coordinate of fritter, according to ARB_currCoordinate derive Block_ACCoordinate.Block_ALCThe CU at place is i.e. used ALC, Block_LCThe CU at place is i.e. It is used LC, Block_ACThe CU at place be i.e. used AC.Due to Block_ALCWith Block_LCChoosing The CU taken is overlapping greatly, removes after repeating, the weights of both sums and AC with quite, institute The weights of AC to be set to 0.5 and ALC is for 0.2, LC is 0.3.For the CU of the 0th layer, choose Predict by rate distortion costs (Rate-Distortion cost, RD-cost).If selected depth is predicted, Owing to the situation that the adjacent C U degree of depth is 0 is little, whether should so being difficult to predict the CU that current depth is 0 When continuing to divide.The RDcost of prediction is calculated by adjacent CU_pre:

${\mathrm{RDcost}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{RDcost}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(1\right)$

Wherein, RDcost_iValue be the RD-cost of ALC, AC and LC respectively.k_iRepresent adjacent CU whether Exist, if there is and selected pattern be SKIP pattern, then k_iIt is 1 to be otherwise 0.a_iIt is 3 The weights of individual CU, the sequence number of the value of i CU each with Fig. 1 is corresponding.If all of CU is the most not Exist and the most do not carry out any prediction.If current optimization model selected by CU is SKIP pattern, then RD-cost and RDcost by current CU_preCompare.If less than RDcost_preIf, then just Do not continue to divide downwards.Even if otherwise the optimization model of current CU is SKIP pattern, CU is also required to continue Divide downwards.For the CU of the 1st layer, the method using depth prediction, compared by experiment and adopt at the 1st layer More more effective than RD-cost by the method for depth prediction.As follows for the depth calculation of prediction:

${\mathrm{Depth}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{Depth}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(2\right)$

Wherein, a_iAnd k_iValue identical with formula (1) value.Depth_iIt is ALC, AC and LC respectively The degree of depth.In the case of the optimization model of current CU is SKIP pattern, if the Depth calculated_preLittle In equal to 1.0, then CU does not continues to divide downwards.Otherwise continue to divide, though the optimum of current CU Pattern is SKIP.Table 2 lists under LD and RA configures, and limits the CU that the degree of depth is 0 and 1 After predictablity rate.By table 2 it can be seen that the degree of depth and RD-cost by adjacent C U are come optimum Pattern is that the division of the CU of SKIP pattern carries out restriction and can effectively improve predictablity rate.The most right In texture than more rich video such as BQMall, it was predicted that the raising of accuracy rate is clearly.More for having Smooth region and the most violent video that moves, the predictablity rate before being not unrestricted is the highest, so carrying Rise is not very obvious.As can be seen from the table, along with the increase of QP, it was predicted that the lifting of accuracy rate

Gradually reduce.The predictablity rate of the 0th layer can be promoted to the 87.3%～99.1%, the 1st by limiting The predictablity rate of layer is promoted to 94.8%～99.6%.

Table 2. CU that the degree of depth is 0 and 1 is limited after predictablity rate

Embodiment

Below in conjunction with attached embodiment illustrated in fig. 3, the present invention is further illustrated.

As it is shown on figure 3, an embodiment of the present invention comprises the following steps:

Step 1: based on test platform HM general for HEVC, starts a LCU and divides, and carries out current All PU pattern searches of CU, obtain optimal prediction modes, go to step 2.

Step 2: judge the current CU degree of depth.If the degree of depth is 3 to go to step 6, otherwise go to step 3.

Step 3: judge whether the optimization model that PU search obtains is SKIP pattern, if going to step 4, Otherwise go to step 2.

Step 4: whether the degree of depth judging current CU is 0.If 0, then according to formula of the present invention (1) Calculate the RDcost for prediction_preIf the RD-cost of current CU is less than RDcost_pre, then go to step 6, otherwise continue to divide downwards, go to step 2.If the degree of depth of current CU is not 0, then go to step 5.

Step 5: whether the degree of depth judging current CU is 1.If 1, then according to formula of the present invention (2) Calculate the Depth for prediction_preIf, Depth_preLess than or equal to 1, then go to step 6, otherwise continue Divide downwards, go to step 2.If the degree of depth of current CU is not 1, then go to step 6.

Step 6:LCU divides and terminates.

Embodiment

(effect example)

Table 3 inventive algorithm (Ours) and the original CU of HEVC divide termination algorithm (ECU in HEVC) in advance and join at LD Experimental result contrast (experiment porch is HM 16.4) under putting

Embodiment

(effect example)

Table 4. inventive algorithm (Ours) and the original CU of HEVC divide and shift to an earlier date termination algorithm (ECU in HEVC) at RA Experimental result contrast (experiment porch is HM 16.4) under Pei Zhi

Claims (1)

1. the algorithm of a HEV interframe encode dividing elements high-speed decision, it is characterised in that include walking as follows Rapid:

1) after all PU pattern searches of current CU are complete, it is judged that the current CU degree of depth；If the degree of depth is 3 Divide and terminate, if the degree of depth is 0 to go to step (2), if the degree of depth is 1 to go to step (3), if the degree of depth is 2 Go to step (4)；

2) if the degree of depth is 0 and optimization model is SKIP；The RDcost for prediction is calculated according to formula (1)_pre, If the RD-cost of current CU is less than RDcost_pre, then divide and terminate；Otherwise continue to divide downwards, And forward step (1) to；

Described formula (1), is for calculating the RDcost of prediction by adjacent CU_pre:

${\mathrm{RDcost}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{RDcost}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(1\right)$

Wherein,

RDcost_iValue be the RD-cost of ALC, AC and LC respectively；

k_iWhether the CU representing adjacent exists, if there is and selected pattern be SKIP pattern, then k_iIt is 1 to be otherwise 0；

a_iIt is the weights of 3 CU, the value of i CU each with Fig. 1 (current CU and the adjacent CU in spatial domain thereof) Sequence number corresponding；

3) if the degree of depth is 1 and optimization model is SKIP；The Depth for prediction is calculated according to formula (2)_pre, If Depth_preLess than or equal to 1, then divide and terminate；Otherwise continue to divide downwards, and forward step (1) to；

Described formula (2), for the degree of depth of prediction:

${\mathrm{Depth}}_{pre}=({\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i\×{\mathrm{Depth}}_i)/{\mathrm{\Σ}}_{i=1}^3{a}_i\×k_i---\left(2\right)$

Wherein,

a_iAnd k_iValue identical with formula (1) value,

Depth_iIt is the degree of depth of ALC, AC and LC respectively；

4) if the degree of depth is 2, if the optimization model of current CU is SKIP, then termination is divided；Otherwise continue to Lower division, and forward step (1) to.