CN103561263B - Based on motion vector constraint and the motion prediction compensation method of weighted motion vector - Google Patents

Abstract

A kind of based on motion vector constraint and the motion prediction compensation method of weighted motion vector, comprise the following steps: the motion vector of the coded block around present encoding block is carried out motion vector constraint fractional analysis；The encoding block of the Uniform Domains that motion vector meets constraints directly uses fusion mode coding to use the advanced motion vector prediction method of weighting to carry out inter motion compensation predictive coding the encoding block being in the complex region that motion vector is unsatisfactory for constraints.The present invention is by motion vector about beam analysis, inter motion compensation is predicted process refinement, and the A weighting MVP for complex scene predicts, the degree of accuracy of inter prediction is improve under conditions of not introducing extra coding information, thus improve coding efficiency, reduce the code check after video compress.

Description

Based on motion vector constraint and the motion prediction compensation method of weighted motion vector

Technical field

The invention belongs to video coding and decoding technology field, especially one based on motion vector constraint and weighting fortune The motion prediction compensation method of dynamic vector.

Background technology

Digital video technology obtains in communication and broadcast world and is increasingly widely applied, particularly in the nineties After, along with the Internet and the fast development of mobile communication technology, video information and multimedia messages are in the Internet Process and transmission in network and mobile network become the focus in current information technical research.

Video has dominated mankind's information source in the information age, but the original video data of uncompressed is past Toward the hugest, bring the biggest difficulty to storage and transmission.Meanwhile, huge video data exists The redundancy in a lot of information, this is that the compression of video data provides sufficient probability.Video compress is not But pursue high compression ratio (lower code check), also want to farthest keep video image to rebuild simultaneously After quality.But, it is contradiction between these demands, outstanding encryption algorithm will seek these contradictions Between optimum balance.Shannon has successively delivered complete rate distortion theory in 1948 and nineteen fifty-nine, Mathematically analyze obtained how balancing distortion and code check, have found under constant bit rate distortion up to the upper limit Or under given distortion code check up to lower limit, rate distortion theory is one of foundation stone of Video coding.

Video coding, by excavating the spatial coherence in video data and temporal correlation, eliminates in data Time redundancy, spatial redundancy, statistical redundancy and visual redundancy, to reach the purpose of compression.

ADPCM technology (DPCM, the Differential Pulse that nineteen fifty-two is invented by Cutler Code Modulation) it is the research the earliest to image coding technique, the DPCM coding in frame has excavated a frame Dependency on image space is compressed.For video signal, time domain there is also the dependency (time Redundancy), 1969, F.W.Mounts and F.Rocca proposed the DPCM in time domain.The seventies, Along with going deep into of the understanding to relativity of time domain, motor-function evaluation technology is suggested, until today, One of motion compensation technique core remaining video compress.Motion compensation is by Video Image Segmentation in bulk, Mate between temporally adjacent image, then the residual error portion after coupling is encoded, so Can preferably remove the time redundancy between the frame of video in video signal and frame, reach the purpose of compression. Meanwhile, converter technique is paid attention in image and video coding field, and Ahmed et al. proposes famous Block-based discrete cosine transform (DCT, Discrete Cosine Transform), up to now, still So it is used in video compress.The core concept of transition coding is partitioning video data in bulk, utilizes orthogonal The energy of data is focused on less several conversion coefficient by conversion.Predictive coding and the continuous of transition coding send out Exhibition, has promoted the birth of hybrid encoding frame.J.R.Jain and A.K.Jain was the world of 1979 Hybrid encoding frame (MC/DCT based on block motion compensation and transition coding is proposed in picture coding association Hybrid Coding).It also becomes the core of modern almost all of video encoding standard.Hybrid coding The nucleus module of framework, in addition to predictive coding and transition coding, also has and quantifies and entropy code two parts.Depending on The loss of information and the acquisition of compression ratio in frequency coding, largely derive from quantization modules.Quantify basic On all use scalar quantisation technique, certain in predefined code table will be mapped to by the single sample in source signal One fixed level value, forms the fewest mapping, thus reaches the purpose of compression, certainly in the process of compression In just introduce loss.Signal after quantization carries out lossless entropy code again, eliminates the statistical redundancy in signal. The research of entropy code can trace back to the 1950's the earliest, Huffman coding, Run-Length Coding (Run-length coding) and the invention of arithmetic coding (Arithmetic coding).Through decades Development, entropy code application in Video coding is more ripe, more exquisite, makes full use of in video data Contextual information, probabilistic model is estimated more accurate, thus improves the efficiency of entropy code.

As the most ripe coding standard, in March, 2003, H.264 video encoding standard is pushed out.H.264 After releasing, obtained the extensive concern of domestic and international academia and industrial circle, it code efficiency, picture quality, All many-sides such as network adaptability and anti-error code all achieve good effect, but its encryption algorithm has the highest Complexity.Along with terminal and the fast development of network technology, the requirement to Video coding improves constantly, for H.264, the requirement that adaptation is new, needs by the most perfect.From in January, 2010, Video coding is special Family group (VCEG) and Motion Picture Experts Group (MPEG) jointly set up Video coding joint specialist group (JCT-VC, Joint Collaborative Team on Video Coding) formulate new video encoding and decoding standard, High-performance video coding standard (HEVC, High Efficiency Video Coding).HEVC is suggested work For a new generation's video encoding standard, and in January, 2013, ISO/IDE and ITU-T issues HEVC Whole draft internation standard.HEVC has become some major companies in one's power of many Research Centers at International video coding mark The study hotspot of quasi-aspect.The target of HEVC is under the conditions of keeping same video quality, with H.264 standard Comparing, code check saves 50%.But, HEVC up-to-date official identifying code code check compared with H.264 is saved and is 22% to 37%, not yet reach 50% this target.Therefore, HEVC still has bigger proposing in terms of compression performance Rise space and necessity.

Similar with standard before, HEVC still uses hybrid encoding frame, but pre-in data structure, frame The links such as survey, inter prediction, conversion, loop filtering have all carried out careful improvement and optimization.HEVC Key technical feature include the following aspects:

(1) before replacing, standard uses the macro block dividing mode of 16x16 fixed size, uses 4 fork tree flexibly Data structure, introduce coding unit (CU, Coding Unit), predicting unit (PU, Prediction Unit), Converter unit (TU, Transform Unit) concept, uses quaternary tree to carry out coded data block and transform block Express.

Along with the increase of video resolution, image occurring, the probability of larger area flat site improves, based on The motion compensation of more large scale (more than 16x16) can more efficiently improve code efficiency, and HEVC adopts Quad-tree partition mode flexibly, similar with the concept of macro block, divide the image into into several code tree unit (CTU, Coding Tree Unit), each CTU can the most iteratively carry out four points, directly To minimum code unit (SCU, Smallest Coding Unit).

(2) the image division mode of parallel encoding mechanism is supported, parallel including image sheet (Tiles) and row ripple Process (WPP, Wave-front Parallel Processing).Image averaging can be divided by Tiles technology Becoming some row and columns, each Tile can independently decode or independent parsing, so can give parallel computation Offer convenience.WPP technology is to carry out parallel by image according to CTU row, initial two CTU of each CTU row Encoding and decoding complete after, the encoding and decoding of next CTU row can start parallel, has both added degree of concurrence, Keep the context of prediction and entropy code the most as much as possible, compared with Tile and figure slice (Slice), can Preferably to keep code efficiency.

(3) infra-frame prediction have employed Angular technology, prediction direction increases to 35 and (includes direct current DC With plane Planar pattern) and improve smooth (MDIS, Mode in intra prediction value generates frame based on pattern Dependent Intra Smooth).HEVC have employed and the most similar Spatial intra-prediction technology, will The direction of infra-frame prediction increases to 35, makes prediction direction more careful, in order to effectively catch in image Texture, eliminates spatial redundancy, improves compression efficiency.Meanwhile, HEVC in the generation of intra prediction value, root According to difference and the difference of intra prediction mode of infra-frame prediction block size, predictive value is carried out different filtering Process, i.e. MDIS technology, so can improve the accuracy of infra-frame prediction, improve compression efficiency.Colourity is divided The infra-frame prediction of amount has excavated the dependency between colourity and brightness, when colourity is similar with the texture of luminance component Time, chromatic component can directly use the prediction direction of luminance component, to avoid transmitting extra pattern information, Improve code efficiency.

(4) advanced motion vector prediction, the motion vector of HEVC can carry out pre-by spatially and temporally two ways Survey.Compared with standard before, the motion-vector prediction of HEVC make use of more spatial neighboring information, no Use median method to calculate fixing predicted motion vector again, but generate candidate prediction motion vector collection, so The predicted motion vector that rear percent of pass aberration optimizing algorithms selection is optimal, and in code stream, transmit optimum prediction fortune The call number of dynamic vector.

(5) new inter-frame forecast mode Merge pattern exists relatively on room and time due to interframe prediction block Strong concordance, block to be encoded is directly used the movable information of adjacent block, can be eliminated redundancy, reaches coding The lifting of efficiency.Using the block of Merge pattern, the movable information that can directly replicate adjacent block moves Compensate, thus save the encoding overhead of movable information, improve compression efficiency.

(6) inter prediction asymmetrical movement divides (AMP, Asymmetric Motion Partitions), removes The 2Nx2N similar with the variable-block in H.264/AVC, 2NxN, Nx2N, NxN (N ∈ 32,16,8, 4}) in addition to such piece divides, HEVC also adopts AMP pattern, can catch image more neatly In moving object, improve code efficiency.

(7) sub-pixel interpolation wave filter based on DCT, the motion vector accuracy of HEVC as H.264/AVC, Remain as 1/4 precision, but have employed separable one-dimensional filtering device based on DCT, further increase sub-picture The accuracy of element interpolation.

(8) quaternary tree shape mapped structure and the conversion of multiple size.The transition coding of the prediction residual in HEVC is adopted With many size change overs 32x32 of quaternary tree shape, the conversion of 16x16,8x8,4x4, transformation kernel is integer DCT. Additionally, the prediction residue block to some intra prediction mode, can use 4x4 discrete sine transform (DST, Discrete Sine Transform, due in infra-frame prediction, the position that the distance upper left corner is the most remote, it was predicted that property Can be the poorest, from border, upper left more away from, residual error is the biggest, converts such data block, and DST is better than DCT。

(9) loop filtering part removes outside block filtering technique, also add self adaptation sample bias (SAO, Sample Adaptive Offset).The block elimination filtering technology of HEVC is with similar, but due to block The difference of dividing mode, needs to process respectively the border of CU, PU, TU etc..Rebuild image block elimination filtering it After, according to pixel value and the difference of location of pixels, pixel is divided into different classifications, according to the difference of classification, Introduce different pixel bias, rebuild distortion, i.e. SAO technology to reduce.

In video data, video interframe temporal redundancy is much larger than spatial redundancies in frame, for condensed frame Between the motion compensated prediction algorithms of temporal redundancy in whole video data compaction coding system, occupy the biggest ratio Weight.Therefore, the quality of motion compensated prediction performance has been largely fixed the compression performance of encoder.

Skip/Merge pattern, as introducing new important coding tools in HEVC interframe encode, can enter one Step promotes interframe encode efficiency.Encode owing to video image divides in bulk, being correlated with between adjacent block Property is the highest, and their texture or motion have the strongest similarity, based on such consideration, before (Skip) pattern of in video standard, skipping, in infra-prediction techniques, has very important status.SKIP The motion vector of mode block is derived by decoding end, and reconstructed value generates entirely by reference to frame, because of This has only to little side information (a usually flag) can express the quantity of information of whole macro block.HEVC Also use similar thought, design Skip and Merge pattern.In order to reduce required for interframe prediction block The expense of the side information of transmission, Skip and Merge fully excavates between the adjacent block on spatial domain and in time domain Similarity, by the way of transmission indexes, expressive movement information, including motion vector, reference key, in advance Surveying direction etc., decoder can recover the movable information of Skip or Merge, saves code check expense.

Skip is (CU) in units of coding unit, the most each coding unit transmission Skip pattern identification and motion letter Breath index, but the most communicating predicted residual error, it was predicted that with the block of 2Nx2N as elementary cell；And Merge is with prediction Unit (PU) is unit, transmits Merge pattern identification and movable information index, in each CU in PU All or part of PU uses Merge pattern, communicating predicted residual error.SKIP with Merge employing is identical Movable information derives mode, and wherein the required movable information derived includes: motion vector, reference key and Prediction direction.

The movable information of Skip/Merge pattern is derived, and uses the adjacent motion information of four types as candidate Generate candidate list, expressed by the index of transmission candidate list.Four class adjacent motion information are respectively as follows:

Spatial domain candidate, time domain candidate, recombinate bi-directional predicted candidate and zero motion vector candidate, and according to institute The order stated, adds in candidate list successively, until meeting the maximum number of candidate list.Skip is to compile Code unit is unit (CU), the most each coding unit transmission Skip pattern identification and movable information index, but not Communicating predicted residual error, it was predicted that with the block of 2Nx2N as elementary cell；And Merge (PU) in units of predicting unit, Transmitting Merge pattern identification and movable information index in PU, all or part of PU in each CU uses Merge pattern, communicating predicted residual error.The movable information derivation mode that SKIP with Merge employing is identical, Wherein the required movable information derived includes, motion vector, reference key and prediction direction.

In HEVC, the selection for Skip/Merge pattern is that traversal search based on rate-distortion optimization compares Method, i.e. the coding for each layer depth CU will carry out the Skip of interframe, Merge, 2Nx2N, Nx2N, The rate distortion of the coding of 2Nx2N, NxN and PCM pattern in 2NxN, NxN, asymmetric AMP segmentation and frame Cost compares, and only just uses this pattern when all patterns all calculate complete and Skip/Merge Least-cost.

For non-Skip/Merge pattern, need movable information is transmitted, in order to save transmission MV's Code check expense, HEVC, with similar, have employed motion-vector prediction technology, by spatial domain and time Territory neighbor information forms motion-vector prediction candidate list (motion-vector prediction candidate list a length of 2), passes through MV is predicted by transmission candidate list index, and transmits prediction residual MVD of MV.

But, in existing inter prediction compensation process, in a reference frame, only one of which reference block is As the predictive value of present encoding block, and for complicated encoding block or the coding inconsistent with peripheral motor A block then block is difficult to present encoding block is carried out Accurate Prediction.Therefore, some compound movement scene or Person local, the accuracy of traditional motion compensated prediction is unsatisfactory, and this have impact on entering of video compression efficiency One step promotes.

Summary of the invention

It is an object of the invention to overcome the deficiencies in the prior art, it is provided that a kind of reasonable in design, precision is high, fast Spend fast motion prediction compensation method based on motion vector constraint with weighted motion vector.

The present invention solves it and technical problem is that and take techniques below scheme to realize:

A kind of based on motion vector constraint and the motion prediction compensation method of weighted motion vector, including following step Rapid:

Step 1, motion vector to the coded block around present encoding block carry out motion vector about fasciculation and divide Analysis；

Step 2, the encoding block of the Uniform Domains that motion vector meets constraints directly use fusion mode Coding；

Step 3, the height that the encoding block employing being in the complex region that motion vector is unsatisfactory for constraints is weighted Level method of motion vector prediction carries out inter motion compensation predictive coding；

Specifically comprising the following steps that of described step 1

(1) the spatial domain motion vectors reference collection of present encoding block being divided into both of which, a kind of pattern takes currently 3 motion vectors of the encoding block on the left of block, another kind of pattern takes 3 fortune of the encoding block above current block Dynamic vector；

(2) motion vector difference between 2 motion vectors in 3 motion vectors is calculated in each pattern, its Computing formula is as follows:

D=| v_x(A)-v_x(B)|+|v_y(A)-v_y(B)|

D is any two motion vector difference, and A, B represent two motion vectors, v respectively_xRepresent motion vector Horizontal component, v_yRepresent the vertical component of motion vector；

(3) present encoding block and the motion vector of surrounding block are judged according to the result of any two motion vector difference D Field is the most consistent.

And, described step (3) judges the side that present encoding block is the most consistent with the motion vector field of surrounding block Method is: for a length of W, and width is the present encoding block of H, when the horizontal component of D all in two patterns It is respectively less than W/8, and when the vertical component of all D is respectively less than H/8, the most current motion vector field is continuous , it is judged that present encoding block is consistent with the block of surrounding in motion.

And, the fusion mode coded method of described step 2 comprises the following steps:

(1) in merging reference set, by the calculating of absolute difference sum, reference motion vector is most preferably merged, Its computing formula is as follows:

${\mathrm{MV}}_{merge}=\arg \underset{d\∈V_{SMC}}{\min }SAD\left(F_n\right(s),{\tilde{F}}_{n-1}(s+d\left)\right)$

Wherein, MV_mergeRepresenting optimal fusion mode reference motion vector, d represents motion vector, V_SMCTable Showing fusion mode reference set, SAD is the absolute difference sum of all pixels between current block and reference block, for generation The matching degree of two blocks of table, F_nS () represents the pixel value of present frame,Represent the i.e. reference frame of former frame Pixel value；

(2) residual error data after being compensated by optimal fusion mode reference motion vector, through change quantization After the coefficient sum that obtains be zero, then use skip mode to encode, and by the mark of skip mode and The index value of good fusion mode reference motion vector is transferred to decoding end；

(3) residual error data after being compensated by optimal fusion mode reference motion vector, through change quantization After the coefficient sum that obtains be not zero, then by after the index value of fusion mode reference motion vector and change quantization Residual error coefficient encoded be transferred to decoding end.

And, described step 3 uses the advanced motion vector prediction method of weighting to carry out the concrete of interframe encode Step includes:

(1) from advanced motion vector prediction reference set, 2 final AMVP predictive values are selected, then to it Averaging, its computing formula is as follows:

Wherein,Represent AMVP mean prediction vector,WithRepresent 2 final AMVP respectively Predictive value；

(2) when 2 final AMVP predictive values do not include zero motion, zero motion vector note power is added, Obtaining the AMVP predictive value of weighting, its computing formula is as follows:

Wherein, A (s) represents the AMVP predictive value of weighting, w_mIt is represented to the weight of AMVP mean prediction vector, w_zeroIt is represented to the weight of zero motion vector；

(3) after the AMVP predictive value obtaining weighting, on the basis of this predictive value, carry out motion search, depend on According to minimum absolute difference sum criterion, obtaining final motion vector, computing formula is as follows:

${\mathrm{MV}}_{weighted}=\arg \underset{d}{min}SAD\left(F\right(s),(A\left(s\right)\·w_{AMVP}+{\tilde{F}}_{n-1}\left(s+d\right)\·w_{search}\left)\right)$

Wherein, MV_weightedRepresenting the motion vector scanning for obtaining after A weighting MVP predicts, d represents All possible motion vector, w in the hunting zone that motion search window i.e. specifies_AMVPIt is represented to weighting The weight of AMVP predictive value, w_searchIt is represented to the weight of the motion vector obtained through motion search；

(4) for final MV_weighted, by it with a motion compensation after the minimum AMVP of absolute difference sum pre- Measured value subtracts each other and obtains residual error motion vector, and its computing formula is as follows:

MVD=MV_weighted-MV_candidate

Wherein, MVD represents residual error motion vector, MV_candidateRepresent the motion vector predictor of present encoding block.

And, in described step (2), when 2 final AMVP predictive values include zero motion, w_mTake 1 And w_zeroTake 0；When 2 final AMVP predictive values do not include zero motion, w_mTake 0.5 and w_zeroTake 0.5.

And, in described step (3), w_AMVPTake 0.2, and w_searchTake 0.8.

Advantages of the present invention and good effect be:

The present invention judges present encoding block and the concordance of surrounding block and complexity by motion vector constrained procedure Property, the inter motion compensation of encoding block is predicted and is divided into two kinds of flow processs the most accurate, thus it is pre-to increase interframe Survey precision；Fusion mode or skip mode are used for the encoding block that motion is simple or concordance is good, saves Go motion search to obtain the process of motion vector, accelerate judgement speed；For complicated or with surrounding The encoding block moving inconsistent uses A weighting MVP Forecasting Methodology the most accurately, can improve complex scene Precision of prediction, simultaneously do not introduce extra coding information (such as: extra reference frame, motion vector and mould Formula mark etc.).

Accompanying drawing explanation

Fig. 1 is the overall process flow figure of the present invention；

Fig. 2 is AMVP prediction and the Candidate Set location drawing of Merge pattern；

Fig. 3 is the flow chart choosing two final AMVP predictive values；

Fig. 4 is that the present invention compares schematic diagram (BasketballDrive sequence) with standard code rate distortion curve；

Fig. 5 is that the present invention compares schematic diagram (RaceHorses sequence) with standard code rate distortion curve.

Detailed description of the invention

Below in conjunction with accompanying drawing, the embodiment of the present invention is further described:

A kind of based on motion vector constraint and the motion prediction compensation method of weighted motion vector, as it is shown in figure 1, Comprise the following steps:

Step 1, carries out motion vector about fasciculation to the motion vector of the coded block around present encoding block and divides Analysis.

In this step, the motion letter of the coded block adjacent with present encoding block time domain and spatial domain is first obtained Breath, then carries out computational analysis to available movable information, finally judges present encoding block and surrounding coding Whether block belongs to same actual object, specifically comprises the following steps that

(1) the spatial domain motion vectors reference collection of present encoding block is divided into both of which: pattern one takes current block 3 motion vectors of the encoding block in left side, pattern two takes 3 motion vectors of the encoding block above current block. A when left side₀In the presence of, choose A₀And B₂As pattern one；Work as A₀Do not exist is to choose A₁And B₂As mould Formula one；B when top₀In the presence of, choose B₀And B₂As pattern two, work as B₀Do not exist is to choose B₁And B₂Make For pattern two.

(2) motion vector difference between 2 motion vectors of 3 motion vectors in each pattern is calculated, Calculate lower formula as follows:

D=| v_x(A)-v_x(B)|+|v_y(A)-v_y(B)|

Wherein, D is any two motion vector difference, and A, B represent two motion vectors, v respectively_xRepresent fortune The horizontal component of dynamic vector, v_yRepresent the vertical component of motion vector.

(3) judge that the motion of present encoding block and surrounding block is vowed according to the result of any two motion vector difference D Whether amount field consistent, or current block is the most complicated.Difference between adjacent two motion vectors is less than Predetermined threshold value then thinks that the motion of the two pixel is consistent, and the two pixel belongs to same reality in other words Border object.

Test result indicate that if the difference between adjacent two motion vectors is less than 1/8th pixels, recognize Being consistent for the motion always of this pixel, the two pixel belongs to same actual object in other words.So For the present encoding block that a length of W width is H, by the result obtained in step (2) is analyzed, If the horizontal component of all of D in two patterns can be obtained be less than W/8, and the vertical of all of D divides Amount is less than H/8, and current motion vector field is continuous print, then may determine that the block of present encoding block and surrounding exists It is consistent in motion.Whereas if any one component exceedes this threshold value, then it is assumed that present encoding block Around block is inconsistent.

By above 3 steps, present encoding block can be obtained and whether surrounding encoding block belongs to same reality Border object, moves the most consistent in other words, and the selection for step 2 and step 3 provides basis for estimation: according to The result that motion vector constraint fractional analysis obtains is by motion compensated prediction process classification, if at motion vector about The conclusion obtained in fasciculation analysis is that present encoding block is consistent with the motion vector field of surrounding block, then directly adopt Predict by fusion mode；The encoding block being in the complex region that motion vector is unsatisfactory for constraints is used and adds The advanced motion vector prediction method of power carries out interframe encode.

Step 2, the encoding block of the Uniform Domains that motion vector meets constraints directly uses fusion mode Coding.

After motion vector constraint fractional analysis, if present encoding block is consistent with the motion vector field of surrounding block, The most directly using fusion mode to carry out encoding (Merge pattern-coding), it specifically comprises the following steps that

(1) in merging reference set, by the calculating of SAD, obtain optimal fusional movement minimum for SAD and vow Amount, can represent such as below equation.

${\mathrm{MV}}_{merge}=\arg \underset{d\∈V_{SMC}}{\min }SAD\left(F_n\right(s),{\tilde{F}}_{n-1}(s+d\left)\right)$

Wherein MV_mergeRepresenting optimal fusion mode reference motion vector, d represents motion vector, V_SMCRepresent Fusion mode reference set, SAD is the absolute difference sum of all pixels between current block and reference block, is used for representing The matching degree of two blocks, F_nS () represents the pixel value of present frame,Represent the former frame i.e. picture of reference frame Element value.

(2) if by the residual error data after optimal fusion mode reference motion vector compensation, through converted quantity The coefficient sum obtained after change is zero, and namely coded block flag is zero, then use skip mode (skip) Encode, and the mark of skip mode and the index value of optimal fusion mode reference motion vector are transferred to Decoding end.

(3) if by the residual error data after optimal fusion mode reference motion vector compensation, through converted quantity The coefficient sum obtained after change is not zero, then by index value and the change quantization of fusion mode reference motion vector After residual error coefficient encoded be transferred to decoding end.

Step 3, uses the height of weighting to the encoding block being in the complex region that motion vector is unsatisfactory for constraints Level motion-vector prediction (AMVP) method carries out inter motion compensation predictive coding.

After motion vector constraint fractional analysis, if the motion vector field of present encoding block and surrounding block is inconsistent Time, then use the AMVP of weighting to predict, it specifically comprises the following steps that

(1) in order to consistent with two AMVP predictive values of reservation in HEVC standard, this method is also from advanced motion Vector prediction (AMVP) reference set selects 2 final AMVP predictive values, its handling process such as Fig. 3 Shown in, first from 5 spatial domain AMVP predictive values, choose 2, according to A₀, A₁, B₀, B₁, B₂'s Order is chosen, until taking full 2；Then from 2 time domain AMVP predictive values, 1 is chosen, if H Exist and then choose H, otherwise choose C₃.If surrounding motion vectors value is less than 3, can be default.Motion The ordering of vector is that first spatial domain is in time domain.Then, remove the repetition values obtained in AMVP predictive value, If the motion vector number now retained is less than 2, then add zero motion vector i.e. (0,0)；If still 3 motion vectors are also had then to remove the 3rd.So obtain final 2 AMVP predictive value.

(2) then 2 final AMVP predictive values are averaged, carry out as follows.

Represent AMVP mean prediction vector,WithRepresent 2 final AMVP predictive values respectively.

(3) when 2 final AMVP predictive values do not include zero motion, zero motion vector note power is added, So can avoid when the motion vector of surrounding block and the true motion vector difference of current block cause relatively greatly Forecast error is excessive, obtains the AMVP predictive value of weighting, and its detailed process is carried out as follows:

A (s) represents the AMVP predictive value of weighting, w_mIt is represented to the weight of AMVP mean prediction vector, w_zeroTable Show to the weight of zero motion vector.Test result indicate that when 2 final AMVP predictive values include zero motion, w_mTake 1 and w_zeroTake 0；When 2 final AMVP predictive values do not include zero motion, w_mTake 0.5 and w_zeroTake 0.5, higher prediction effect can be obtained.

(4) after the AMVP predictive value obtaining weighting, on the basis of this predictive value, carry out motion search, depend on According to minimum SAD criterion, obtaining final motion vector, computing formula is as follows.

${\mathrm{MV}}_{weighted}=\arg \underset{d}{min}SAD\left(F\right(s),(A\left(s\right)\·w_{AMVP}+{\tilde{F}}_{n-1}\left(s+d\right)\·w_{search}\left)\right)$

Wherein MV_weightedRepresenting the motion vector scanning for obtaining after A weighting MVP predicts, d represents in fortune All possible motion vector, w in the hunting zone that dynamic search window i.e. specifies_AMVPIt is represented to the AMVP of weighting The weight of predictive value, w_searchIt is represented to the weight of the motion vector obtained through motion search.Experimental result table Bright, w_AMVPTake 0.2, and w_searchTake 0.8 and can obtain optimal prediction effect.

(5) for final MV_weighted, by AMVP predictive value minimum for sad value after itself and a motion compensation Subtract each other and obtain residual error motion vector (MVD), it is therefore an objective to remove redundancy further, and meet main flow encoder Code stream specification.This process is calculated as follows.

MVD=MV_weighted-MV_candidate

Wherein, MV_candidateRepresenting the motion vector predictor of present encoding block, this predictive value is in (1) With MV in two the final AMVP predictive values obtained_weightedMinimum one of MVD value.

In sum, the feature of the inventive method is to first pass through Part I motion vector is carried out about fasciculation Analyze, when present encoding block is consistent with the motion vector field of surrounding block, then the method using step 2, carry out Fusion mode encodes；When the motion vector field of present encoding block and surrounding block is inconsistent, then use step 3 Method, the AMVP being weighted predicts.

In order to the effect of this motion prediction compensation method is described, uses and verify as follows:

Experiment use HEVC official standard video test sequence as test object, in the process of the present invention and HEVC Official's code compares and carries out experiment simulation.Coded quantization parameter QP is set to: 22,27,32,37.This The Y-PSNR (BD-PSNR) of method and HEVC official identifying code and code check (BD-rate) performance ratio Relatively such as following table.Fig. 4 is that the inventive method of BasketballDrive sequence is bent with the rate distortion of standard code Line compares, and Fig. 5 is that the inventive method of RaceHorses sequence compares with the rate distortion curve of standard code.

Table 1 BD-PSNR and BD-rate Performance comparision

BD-PSNR represents under conditions of same video code check, the difference of Y-PSNR, and positive represents this Invention Y-PSNR compared with standard code increases, and unit is dB.BD-rate represents at identical peak Under conditions of value signal to noise ratio, the present invention is the difference of video code rate compared with standard code, and negative value represents this Bright compared with standard code video code rate declined.Unit is percentage ratio.For cycle tests Vidyo1, Three individual sports in video are very simple and slow, occupy the large-area background parts of video and keep static, So the frequency ratio that the advanced motion vector prediction of weighting proposed by the invention occurs is relatively low, coding gain is relatively Little.For cycle tests BasketballDrill, the complicated movement of the basket baller in video acutely and Quickly, significant portion of motion vector is discontinuous, therefore the senior fortune of the weighting that the inventive method proposes The frequency that dynamic vector prediction occurs is higher, and coding gain is bigger.

It is emphasized that embodiment of the present invention is illustrative rather than determinate, therefore The present invention is not limited to the embodiment described in detailed description of the invention, every by those skilled in the art according to this Other embodiments that the technical scheme of invention draws, also belong to the scope of protection of the invention.

Claims (6)

1., based on motion vector constraint and a motion prediction compensation method for weighted motion vector, its feature exists In, comprise the following steps:

Step 1, motion vector to the coded block around present encoding block carry out motion vector about fasciculation and divide Analysis；

Step 2, the encoding block of the Uniform Domains that motion vector meets constraints directly use fusion mode Coding；

Step 3, the height that the encoding block employing being in the complex region that motion vector is unsatisfactory for constraints is weighted Level method of motion vector prediction carries out inter motion compensation predictive coding；

Specifically comprising the following steps that of described step 1

(1) the spatial domain motion vectors reference collection of present encoding block being divided into both of which, a kind of pattern takes currently 3 motion vectors of the encoding block on the left of block, another kind of pattern takes 3 fortune of the encoding block above current block Dynamic vector；

(2) motion vector difference between 2 motion vectors in 3 motion vectors is calculated in each pattern, its Computing formula is as follows:

D=| v_x(A)-v_x(B)|+|v_y(A)-v_y(B)|

D is any two motion vector difference, and A, B represent two motion vectors, v respectively_xRepresent motion vector Horizontal component, v_yRepresent the vertical component of motion vector；

(3) present encoding block and the motion vector of surrounding block are judged according to the result of any two motion vector difference D Field is the most consistent.

Motion compensation based on motion vector constraint and weighted motion vector the most according to claim 1 is pre- Survey method, it is characterised in that: described step (3) judges that the motion vector field of present encoding block and surrounding block is No consistent method is: for a length of W, and width is the present encoding block of H, as D all in two patterns Horizontal component be respectively less than W/8, and when the vertical component of all D is respectively less than H/8, the most current motion is vowed Amount field is continuous print, it is judged that present encoding block is consistent with the block of surrounding in motion.

Motion compensation based on motion vector constraint and weighted motion vector the most according to claim 1 is pre- Survey method, it is characterised in that: described step 2 fusion mode coded method comprises the following steps:

(1) in merging reference set, by the calculating of absolute difference sum, reference motion vector is most preferably merged, Its computing formula is as follows:

${\mathrm{MV}}_{merge}=\arg \underset{d\∈V_{SMC}}{\min }SAD\left(F_n\right(s),{\tilde{F}}_{n-1}(s+d\left)\right)$

Wherein, MV_mergeRepresenting optimal fusion mode reference motion vector, d represents motion vector, V_SMCTable Showing fusion mode reference set, SAD is the absolute difference sum of all pixels between current block and reference block, for generation The matching degree of two blocks of table, F_nS () represents the pixel value of present frame,Represent the i.e. reference frame of former frame Pixel value；

(2) residual error data after being compensated by optimal fusion mode reference motion vector, through change quantization After the coefficient sum that obtains be zero, then use skip mode to encode, and by the mark of skip mode and The index value of good fusion mode reference motion vector is transferred to decoding end；

(3) residual error data after being compensated by optimal fusion mode reference motion vector, through change quantization After the coefficient sum that obtains be not zero, then by after the index value of fusion mode reference motion vector and change quantization Residual error coefficient encoded be transferred to decoding end.

Motion compensation based on motion vector constraint and weighted motion vector the most according to claim 1 is pre- Survey method, it is characterised in that: described step 3 uses the advanced motion vector prediction method of weighting to carry out interframe The concrete steps of coding include:

(1) from advanced motion vector prediction reference set, 2 final AMVP predictive values are selected, then to it Averaging, its computing formula is as follows:

Wherein,Represent AMVP mean prediction vector,WithRepresent 2 final AMVP respectively Predictive value；

(2) when 2 final AMVP predictive values do not include zero motion, zero motion vector note power is added, Obtaining the AMVP predictive value of weighting, its computing formula is as follows:

Wherein, A (s) represents the AMVP predictive value of weighting, w_mIt is represented to the weight of AMVP mean prediction vector, w_zeroIt is represented to the weight of zero motion vector；

(3) after the AMVP predictive value obtaining weighting, on the basis of this predictive value, carry out motion search, depend on According to minimum absolute difference sum criterion, obtaining final motion vector, computing formula is as follows:

${\mathrm{MV}}_{weighted}=\arg \underset{d}{min}SAD\left(F\right(s),(A\left(s\right)\·w_{AMVP}+{\tilde{F}}_{n-1}\left(s+d\right)\·w_{search}\left)\right)$

Wherein, MV_weightedRepresenting the motion vector scanning for obtaining after A weighting MVP predicts, d represents All possible motion vector, w in the hunting zone that motion search window i.e. specifies_AMVPIt is represented to weighting The weight of AMVP predictive value, w_searchIt is represented to the weight of the motion vector obtained through motion search；

(4) for final MV_weighted, by it with a motion compensation after the minimum AMVP of absolute difference sum pre- Measured value subtracts each other and obtains residual error motion vector, and its computing formula is as follows:

MVD=MV_weighted-MV_candidate

Wherein, MVD represents residual error motion vector, MV_candidateRepresent the motion vector predictor of present encoding block.

Motion compensation based on motion vector constraint and weighted motion vector the most according to claim 4 is pre- Survey method, it is characterised in that: in described step (2), when 2 final AMVP predictive values include zero motion Time, w_mTake 1 and w_zeroTake 0；When 2 final AMVP predictive values do not include zero motion, w_mTake 0.5 and w_zero Take 0.5.

Motion compensation based on motion vector constraint and weighted motion vector the most according to claim 4 is pre- Survey method, it is characterised in that: in described step (3), w_AMVPTake 0.2, and w_searchTake 0.8.