CN101237580A - Integer pixel quick mixing search method based on center prediction - Google Patents
Integer pixel quick mixing search method based on center prediction Download PDFInfo
- Publication number
- CN101237580A CN101237580A CN 200810017569 CN200810017569A CN101237580A CN 101237580 A CN101237580 A CN 101237580A CN 200810017569 CN200810017569 CN 200810017569 CN 200810017569 A CN200810017569 A CN 200810017569A CN 101237580 A CN101237580 A CN 101237580A
- Authority
- CN
- China
- Prior art keywords
- prediction
- sad
- point
- search
- opt
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Landscapes
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention discloses an integral pixel quick mixing searching method based on center prediction. The method comprises the following steps: motion vector spatial domain prediction, time domain prediction, UpLayer prediction, multi-reference frame prediction and zero vector prediction are adopted to carry out search window center prediction; the block matching degree of distortion of the candidate points obtained through the five prediction modes is respectively calculated, and the point corresponding to the minimum block matching degree of distortion is selected as a search window central point; a search direction fast definition threshold T1 is set, and a successful judgment threshold T2 and a midway cut-off threshold T3 are predicted; according to the predicted and set thresholds, CPFMS search is carried out to find out a global optimum point. With strong adaptability, the integral pixel quick mixing searching method has less PSNR loss of brightness signal and negligible influence on video reconstruction quality as compared with the FS algorithm and the UMHexagonS algorithm; little increase in bit rate is obtained and coding efficiency is basically unchanged; moreover, time used in integral pixel precise motion estimation is reduced substantially, which increases coding speed.
Description
Technical field
The invention belongs to method for processing video frequency.
Background technology
In recent years, many scholars are carrying out a large amount of research aspect the computation complexity that reduces the block matching motion estimation, mainly concentrate on the exploitation fast search algorithm.These algorithms can be divided into 4 classes substantially:
First kind algorithm adopts equally distributed search dot pattern, as three step search methods (TSS), two dimensional logarithmic method, cross searching algorithm etc., greatly reduces search complexity, but has also reduced the accuracy and the video quality of estimation.For example during the TSS algorithm search, whole process has adopted unified search pattern, makes that the step-length of the first step is excessive, causes misleading easily, thereby lower to little sport efficiency.
The inhomogeneous characteristic of spatial distribution---the center-biased distribution character that the second class algorithm has utilized motion vector to have, typical algorithm have new three step search methods (NTSS), four step search methods (FSS), diamond search method (DS), hexagon search etc.The common high concentration always of motion vector is distributed near the center of search window, for example NTSS adopts the search dot pattern of central tendency not only to improve matching speed, and reduced the possibility that is absorbed in local minimum, adopt the termination discrimination technology then to greatly reduce search complexity, improved search efficiency; The characteristics of DS algorithm are that it has analyzed the basic law of motion vector in the video image, have selected the search pattern of two kinds of shapes of size for use.With the large form search, because step-length is big, the hunting zone is wide, can carry out coarse positioning earlier, and it is local minimum that search procedure can not sunk into; After coarse positioning finishes, can think in the optimum point diamond-shaped area that just 8 points are enclosed around large form, at this moment accurately locate with little template again, make to search for to be unlikely big fluctuating, so its performance is better than other algorithm.In addition, between each step very strong correlation is arranged during the DS search, only need when template moves to mate calculating, so also improved search speed at several new test point places.This type of algorithm is conceived to improve the function of search under the less situation of motion, but predicated error is still bigger when estimating big motion.
More than two class algorithm common characteristic all be that center with search window is as the initial ranging center, and suppose the dull decline of matching error of search point correspondence the time near global optimum's point, when initial search center and global optimum's point distance is big, be absorbed in local minimum easily.
The 3rd class algorithm utilizes the correlation of motion vector on time domain and spatial domain to come motion vectors to select initial ranging center preferably, as motion vectors field self-adaptive search algorithm, self adaptation cross pattern search algorithm and improved estimation range searching algorithm etc.This type of algorithm is intended to utilize the temporal correlation of motion vector to select the future position of a reflection current block movement tendency as initial search point, this future position is than the more close global optimum in the center of search window point, the easier hypothesis that satisfies the matching error monotonicity has improved prediction accuracy.But the performance of this type of algorithm still depends on initial ranging center forecasting reliability and search pattern, and closely related with the local motion feature of piece.
The 4th class algorithm is to utilize the temporal correlation of predicated error and the algorithm that locomotor activity is dynamically adjusted the hunting zone or ended to search for, typical algorithm have the irregular template searching algorithm of self adaptation with coupling anticipation, adaptable search scope adjustment algorithm, Motion Adaptive searching algorithm based on temporal correlation and at H.264 asymmetric cross multi-level hexagonal point search algorithm of video compression standard of new generation (Unsymmetrical-Cross Muti-Hexagon Search, UMHexagonS) etc.Wherein, UMHexagonS algorithm (Zhibo Chen, Peng Zhou, Yun He, " Fast Integer Pel and Fractional PelMotion Estimation for JVT " .JVT-F017.doc, JVT of ISO/IEC MPEG﹠amp; ITU-T VCEG 6
ThMeeting:Awaii Island, 2002,5-13) adopt strategies such as search center prediction, hierarchical search and multiple search pattern, and in the process of search, introduced and skipped strategy midway, effectively avoid candidate point to be absorbed in the situation of local optimum, thereby can reach higher matching precision.The UMHexagonS algorithm can increase under few prerequisite with respect to other algorithms at computation complexity, code efficiency improved greatly, and be the algorithm of combination property the best in the above-mentioned algorithm.But this algorithm search path complexity, search pattern variation (having comprised asymmetric cross, rectangle, symmetry/asymmetrical hexagonal-shaped, rhombus), be difficult to the corresponding unified parallel processing structure of design, therefore be not suitable for the requirement of hardware designs, the computational process of thresholding is also too complicated in addition, be not suitable for adopting hardware handles, these have all limited the UMHexagonS algorithm application in the ASIC design field.
Summary of the invention
In order to overcome prior art UMHexagonS algorithm complexity, be not suitable for the deficiency of hardware designs, the invention provides a kind of integer pixel rapid mixing search (CPFMS based on the center prediction, Centered Prediction based Fast Mixed Search) algorithm, at the temporal correlation that utilizes motion vector Optimum Matching point position is carried out on the base of prediction, carry out five point search and 9 uniform rectangular searches that " ten " font distributes successively, and in search procedure, introduce and search for cut-off threshold, under the prerequisite that guarantees coding efficiency, can effectively reduce the H.264 complexity of motion estimation process.
The technical solution adopted for the present invention to solve the technical problems is:
The first step, the prediction of search window center:
Use for reference the thought of search window center prediction in the UMHexagonS algorithm, the prediction of search window center is except spatial domain, the time-domain of motion vector are predicted, in conjunction with change piece size and multi-reference frame motion prediction new features H.264, UpLayer prediction and multi-reference frame prediction have been introduced, in addition according to based on the quick search principle of off-centring, with the zero vector prediction also as a kind of important predictive mode as the candidate.
The candidate point that five kinds of predictive modes are obtained respectively the computing block coupling distortion factor (Block Distortion Measure, BDM), the point of choosing minimum BDM correspondence is as the search window central point.Plant in prediction mode in the first five of above-mentioned motion vector, through experiment showed, that the prediction of spatial domain is more accurate, the best performance of UpLayer prediction wherein is because this mode has made full use of the correlation between the different prediction block mode motion vectors.And the median prediction performance increases along with reducing of prediction piece size, and here because the current block size is more little, correlation is more little.
In actual applications, if these predictive modes are all carried out one time, amount of calculation and and the spent clock cycle all will be a huge numeral.Simultaneously, if 5 kinds of predictive modes are all carried out, the motion vector of corresponding all reference frames of sub-piece under all sub-piece partition modes of all macro blocks in the frame all need be stored, this is again a huge memory space demand undoubtedly.Therefore, it is impossible all carrying out 5 kinds of predictive modes, does not also have such necessity on the other hand.Because in most cases, a plurality of candidate's predictive vectors that obtain under the multiple predictive mode are all very close, adopt the various modes prediction can cause a lot of double countings simultaneously, in fact adopt two or three pattern wherein can reach re-set target.And in actual motion algorithm for estimating process,, be not that every kind of prediction mode all is effective for the piece under the various situations.For example, corresponding piece does not have motion vector, so the prediction of consecutive frame corresponding blocks can't prove effective; For nearest reference frame, there is not the multi-reference frame prediction; Piece Uplayer prediction for 16 * 16 is invalid etc.
Based on above-mentioned analysis, designed the predictive mode selection algorithm, to adapt to the prediction of the motion vector under different situations, selection algorithm is as follows: (1) only adopts zero vector prediction and UpLayer prediction for the piece of 16 * 8,8 * 16,8 * 8 sizes; (2) adopt the median prediction of zero vector prediction, UpLayer prediction and spatial domain for the piece of 4 * 8,8 * 4,4 * 4 sizes; (3) for the piece of 16 * 16 sizes, if corresponding reference frame 1,2,3,4 only adopts zero vector prediction and multi-reference frame to predict; (4) corresponding reference frame 0, predictive mode is selected in the prediction three of zero vector prediction, median prediction, consecutive frame corresponding blocks.
Second goes on foot, is provided with the search thresholding:
In most of the cases, can well drop on after predictive vector is selected the superior near the zone of the SAD dullness of global optimum's point.Yet have under the rare occasion, predicted value and global optimum's point deviation are bigger.If the motion vector predictor to current block is too big with real best vector deviation, in the Local Search window, carry out limited search and can cause being absorbed in local smallest point or border extreme point, influence the estimation precision.Therefore must whether successfully effectively judge motion-vector prediction.The most direct method is that a threshold value is set, point with optimum prediction vector correspondence is that minimum SAD and the thresholding comparison that obtains searched at the center, if less than thresholding then illustrate that SAD has converged to a certain degree, predictive vector and optimum vector are positioned at same monotone area, can judge and predict successfully, otherwise judge that prediction is unsuccessful, need to enlarge the hunting zone and search for again.Under the successful condition of preliminary judgement prediction, the comparison of the SAD by optimum point and time advantage and more strict thresholding can be defined the direction of search fast, thus counting of searching for of minimizing.
In the CPFMS algorithm, the direction of search defines thresholding T1 fast and the successful decision threshold T2 of prediction utilizes time-domain and the spatial domain correlation of piece coupling distortion SAD to calculate:
T1＝SAD
thrh_1＝SAD
pred(1+β
1) (1)
T2＝SAD
thrh_2＝SAD
pred(1+β
2) (2)
SAD wherein
PredThe SAD that adopts Uplayer prediction, space median prediction, the prediction of consecutive frame corresponding blocks and four kinds of prediction mode of adjacent multi-reference frame prediction to obtain selects the superior, and chooses the minimum final predicted value SAD of conduct
PredRegulating the value of parameter beta determines according to the experiment statistics result.7 kinds of block modes for 16 * 16 (mode 1), 16 * 8 (mode 2), 8 * 16 (mode 3), 8 * 8 (mode 4), 8 * 4 (mode5), 4 * 8 (mode 6), 4 * 4 (mode 7) size have respectively:
β1＝0.02，0.02，0.02，0.03，0.04，0.04，0.05；
β2＝0.05，0.10，0.10，0.12，0.125，0.125，0.15。
Cut-off threshold T3 then according to the characteristic of the discrete cosine transform DCT of motion compensated residual coefficient, utilizes the amplitude of direct current DC coefficient to determine thresholding midway.Usually the DC coefficient is all coefficient amplitude maximums in the transformation coefficient block, so if be 0 behind the DC coefficient quantization, then the module of 4 * 4 sizes should be complete zero piece.Therefore can use complete zero piece of DC coefficient prediction.For 4 * 4 block modes (mode 7), the decision gate limit of complete zero piece is defined as:
T
7＝(2
qbits-f)/QE[q
rem][0][0]＜TH
7？(2
qbits-f)/QE[q
rem][0][0]：TH
7 (3)
Qbits=15+QP/6 wherein, and q
Rem=QP%6, f=(1＜＜qbits)/6, QE refers to defined quantization parameter table.QP is a quantization parameter.If 4 * 4 SAD is less than T
7, can judge that then this piece is complete zero piece, can end when motion prediction is searched for to the point of this piece correspondence.For other block mode ( mode 1,2 ..., 6), the corresponding judgment thresholding is defined as:
T
i＝T
7*β
i＜TH
i？T
7*β
i：TH
i?i＝1，...，6 (4)
β wherein
iBe scale factor, the large scale piece is amplified T in proportion
7TH
iBe the thresholding maximum, be used to prevent that the motion search that complete zero piece erroneous judgement causes from ending too early.Therefore the CPFMS algorithm is set at the judgement thresholding of zero piece: T3=T entirely by thresholding T3 midway
i
The 3rd step, CPFMS searching method step:
If current Optimum Matching point is (X with respect to the coordinate of search center point
OPT, Y
OPT), matching error is SAD
OPTCurrent advantage is (X with respect to the coordinate of central point
SOPT, Y
SOPT), matching error is SAD
SOPTSearch step is as follows: Step1: according to the matching error SAD ' of the corresponding match point of current macro motion vector computation optimum prediction motion vector of median prediction, UpLayer prediction, consecutive frame prediction and the prediction of multiframe reference prediction, if SAD '＜T3, then withdraw from search, this match point is global optimum's match point; Otherwise the SAD that the prediction of SAD ' and zero vector is corresponding is relatively, and adopting the less point of error to carry out step-length as the search window central point is five point search that " ten " font of 2 distributes, and Select Error less o'clock as the current Optimum Matching point in the 2nd step.
Step2: if X
OPT=0 and Y
OPT=0, then:
(1) if SAD
OPT＞T2 illustrates global optimum's point away from current optimum point, need to increase the hunting zone, is that to carry out step-length again be five point search that " ten " font of 2 distributes to search center with current optimum point;
(2) if SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＞T1, global optimum's point near the search window center, with current optimum point be central point to carry out step-length be 9 rectangular searches of 1, as shown in Figure 2;
(3) if SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＜T1, in the zone of global optimum's point between search window center and inferior advantage, with 2 line mid points between search window central point (optimum point) and the inferior advantage be central point to carry out step-length be 9 rectangular searches of 1, as shown in Figure 3.
Step3: if X
SOPT=0 and Y
SOPT=0, then:
(1) if SAD
OPT＞T2 illustrates global optimum's point away from current optimum point, need to increase the hunting zone, with current optimum point be search center to carry out step-length again be five point search that " ten " font of 2 distributes, as shown in Figure 4;
(2) if SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＞T1, global optimum's point near current optimum point, then with current optimum point be central point to carry out step-length be 9 uniform rectangular searches of 1, as shown in Figure 5;
(3) if SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＜T1, in the zone of global optimum's point between current optimum point and inferior advantage, then with 2 line mid points between search window central point and the optimum point be central point to carry out step-length be 9 uniform rectangular searches of 1, as shown in Figure 6.
Step4: if the current search central point is neither optimum point neither time advantage, then:
(1) if SAD
OPT＞T2 or current optimum point and time advantage diagonal angle distribute, illustrate that global optimum's point is away from current optimum point or searched the suboptimum singular point, need to increase the hunting zone, with current optimum point is that to carry out step-length again be five point search that " ten " font of 2 distributes to search center, as shown in Figure 7 and Figure 8, and get back to the circulation that the 2nd, 3,4 steps constituted; Otherwise carry out following steps;
(2) if SAD
OPT＜T1, SAD
SOPT＞T1 illustrates global optimum's point near current optimum point, with current optimum point be central point to carry out step-length be 9 uniform rectangular searches of 1, as shown in Figure 9;
(3) if SAD
OPT＜T1, SAD
SOPT＜T1 illustrates in the zone of global optimum's point between current optimum point and inferior advantage, with 2 line mid points between optimum point and the inferior advantage be central point to carry out step-length be 9 uniform rectangular searches of 1, as shown in figure 10;
Each step in above-mentioned search procedure is if the matching error of optimum point satisfies SAD
OPT＜T3, current optimum point is global optimum's point, ends search and gets final product.
The invention has the beneficial effects as follows: the CPFMS algorithm has very strong adaptability, in little motion, greatly under motion or the camera lens translation situation, can both keep excellent compression performance.Compare with the UMHexagonS algorithm with the FS algorithm, have following characteristics:
1) PSNR of luminance signal loss is less, can ignore substantially the influence of video reconstruction quality;
2) the bit rate increase is very little, and code efficiency is constant substantially;
3) decline consuming time of integer-pel precision estimation is obvious, has improved speed of coding.
The present invention is further described below in conjunction with drawings and Examples.
Description of drawings
Fig. 1 is a motion-vector prediction selection algorithm flow chart.
Fig. 2 is the search step schematic diagram of CPFMS Step2.1.
Fig. 3 is the search step schematic diagram of CPFMS Step2.2.
Fig. 4 is the search step schematic diagram of CPFMS Step3.1.
Fig. 5 is the search step schematic diagram of CPFMS Step3.2.
Fig. 6 is the search step schematic diagram of CPFMS Step3.3.
Fig. 7 is the search step schematic diagram of CPFMS Step4.1.
Fig. 8 is the search step schematic diagram of CPFMS Step4.1.
Fig. 9 is the rapid schematic diagram of the search step of CPFMS Step4.2.
Figure 10 is the rapid schematic diagram of the search step of CPFMS Step4.3.
Embodiment
Method embodiment 1:
In test program JM8.0 code H.264, realize the CPFMS algorithm, and compare with FS algorithm and UMHexagonS algorithm and to analyze the performance of this paper algorithm CPFMS.Choosing 5 typical video sequences, is that coding 100 frames are tested under 18 conditions at quantization parameter QP.Wherein 4 cycle testss are QCIF form (176 * 144), comprise Foreman, Carphone, and Grandma, Mobile, 1 is cycle tests CIF form (352 * 288)-Bus.Selected cycle tests has more intense representativeness, Grandma is little motion or the almost quiet sequence that goes up, and Foreman and Carphone are the middle motion sequence, and the background of Foreman is static, Carphone then is the background of motion, and Mobile and Stefan are big motion sequence.30 frame/seconds of frame frequency are adopted in test, and reference frame is got 5 nearest frames, skips frame number and gets 0, uses whole 7 kinds of block modes to mate.Coding adopts the IPPP... sequence.The maximum possible moving displacement of all whole pixel searching algorithms is all got d=15, and code rate distortion is optimized mode parameter and got RDO=0, and infra-frame prediction adopts full way of search.The motion-estimation encoded time (ME time), mean P SNR and the code check that obtain behind all block matching algorithm codings are added up, as shown in table 1.
The block matching algorithm performance relatively during table 1 QP=18
| QP＝18 | The ME scramble time (s) | PSNR(dB) | Code check (kbits/s) | ||||||
| FS | UMHS | CPFMS | FS | UMHS | CPFMS | FS | UMHS | CPFMS | |
| Grandma | 124.83 | 11.54 | 9.21 | 43.96 | 43.94 | 43.94 | 210.74 | 209.50 | 210.22 |
| Foreman | 125.40 | 26.34 | 16.85 | 43.57 | 43.54 | 43.53 | 504.75 | 495.28 | 499.48 |
| Carphone | 162.13 | 23.02 | 14.97 | 44.34 | 44.30 | 44.30 | 404.43 | 400.71 | 401.27 |
| Mobile | 128.59 | 30.22 | 18.32 | 42.13 | 42.13 | 42.12 | 1617.7 | 1612.0 | 1612.50 |
| Bus | 530.05 | 153.61 | 87.31 | 42.59 | 42.58 | 42.58 | 4353.9 | 4197.8 | 4218.95 |
Method embodiment 2:
In test program JM8.0 code H.264, realize the CPFMS algorithm, and compare with FS algorithm and UMHexagonS algorithm and to analyze the performance of this paper algorithm CPFMS.Choosing 5 typical video sequences, is that coding 100 frames are tested under 28 conditions at quantization parameter QP.Wherein 4 cycle testss are QCIF form (176 * 144), comprise Foreman, Carphone, and Grandma, Mobile, 1 is cycle tests CIF form (352 * 288)-Bus.Selected cycle tests has more intense representativeness, Grandma is little motion or almost static sequence, and Foreman and Carphone are the middle motion sequence, and the background of Foreman is static, Carphone then is the background of motion, and Mobile and Stefan are big motion sequence.30 frame/seconds of frame frequency are adopted in test, and reference frame is got 5 nearest frames, skips frame number and gets 0, uses whole 7 kinds of block modes to mate.Coding adopts the IPPP... sequence.The maximum possible moving displacement of all whole pixel searching algorithms is all got d=15, and code rate distortion is optimized mode parameter and got RDO=0, and infra-frame prediction adopts full way of search.The motion-estimation encoded time (ME time), mean P SNR and the code check that obtain behind all block matching algorithm codings are added up, as shown in table 2.
The block matching algorithm performance relatively during table 2 QP=28
| QP＝28 | The ME scramble time (s) | PSNR(dB) | Code check (kbits/s) | ||||||
| FS | UMHS | CPFMS | FS | UMHS | CPFMS | FS | UMHS | CPFMS | |
| Grandma | 125.71 | 11.18 | 8.4 | 36.57 | 36.56 | 36.55 | 38.00 | 37.61 | 37.45 |
| Foreman | 129.11 | 23.01 | 14.95 | 36.50 | 36.48 | 36.42 | 138.42 | 137.15 | 138.97 |
| Carphone | 152.92 | 20.88 | 11.60 | 37.34 | 37.31 | 37.23 | 99.87 | 98.78 | 100.54 |
| Mobile | 129.55 | 29.91 | 16.82 | 33.33 | 33.33 | 33.33 | 436.9 | 437.45 | 436.77 |
| Bus | 522.52 | 152.38 | 85.39 | 34.70 | 34.69 | 34.69 | 1174.9 | 1175.6 | 1185.42 |
Method embodiment 3:
In test program JM8.0 code H.264, realize the CPFMS algorithm, and compare with FS algorithm and UMHexagonS algorithm and to analyze the performance of this paper algorithm CPFMS.Choosing 5 typical video sequences, is that coding 100 frames are tested under 38 conditions at quantization parameter QP.Wherein 4 cycle testss are QCIF form (176 * 144), comprise Foreman, Carphone, and Grandma, Mobile, 1 is cycle tests CIF form (352 * 288)-Bus.Selected cycle tests has more intense representativeness, Grandma is little motion or almost static sequence, and Foreman and Carphone are the middle motion sequence, and the background of Foreman is static, Carphone then is the background of motion, and Mobile and Stefan are big motion sequence.30 frame/seconds of frame frequency are adopted in test, and reference frame is got 5 nearest frames, skips frame number and gets 0, uses whole 7 kinds of block modes to mate.Coding adopts the IPPP... sequence.The maximum possible moving displacement of all whole pixel searching algorithms is all got d=15, and code rate distortion is optimized mode parameter and got RDO=0, and infra-frame prediction adopts full way of search.The motion-estimation encoded time (ME time), mean P SNR and the code check that obtain behind all block matching algorithm codings are added up, as shown in table 3.
The block matching algorithm performance relatively during table 3 QP=38
| QP＝38 | The ME scramble time (s) | PSNR(dB) | Code check (kbits/s) | ||||||
| FS | UMHS | CPFMS | FS | UMHS | CPFMS | FS | UMHS | CPFMS | |
| Grandma | 132.48 | 11.15 | 5.21 | 30.69 | 30.69 | 30.61 | 8.94 | 8.90 | 8.98 |
| Foreman | 130.27 | 22.80 | 10.89 | 29.83 | 29.78 | 29.69 | 42.50 | 41.99 | 43.95 |
| Carphone | 187.02 | 20.37 | 6.15 | 30.59 | 30.54 | 30.43 | 26.19 | 25.87 | 27.02 |
| Mobile | 128.28 | 31.74 | 15.32 | 25.62 | 25.59 | 25.57 | 81.77 | 81.59 | 81.95 |
| Bus | 560.23 | 161.76 | 66.84 | 27.50 | 27.48 | 27.42 | 272.89 | 272.86 | 282.96 |
From table 1～table 3 as can be seen, the CPFMS algorithm has very strong adaptability, in little motion, greatly under motion or the camera lens translation situation, can both keep excellent compression performance.Compare with FS, have following characteristics:
1) PSNR of luminance signal loss average out to 0.056dB, maximum is no more than 0.16dB, can ignore substantially the influence of video reconstruction quality;
2) the bit rate increase is very little, and maximum is no more than 3.66%, and mean value is 0.84%, and code efficiency is constant substantially;
3) decline consuming time of integer-pel precision estimation is obvious, and for various sequences, its speed can improve speed of coding near 12 times of FS.Compare with the UMHexagonS algorithm in addition, under the approaching situation of reconstructed image quality and code check, speed has improved 1.87 times.
Claims (2)
1, based on the integer pixel quick mixing search method of center prediction, it is characterized in that comprising the steps:
(a) the spatial domain prediction of employing motion vector, time-domain prediction, UpLayer prediction, multi-reference frame prediction and zero vector are predicted and are carried out the prediction of search window center, to the candidate point difference computing block coupling distortion factor that five kinds of predictive modes obtain, the point of choosing smallest blocks coupling distortion factor correspondence is as the search window central point;
(b) direction of search is set and defines thresholding T1, T1=SAD fast
Thrh_1=SAD
Pred(1+ β
1), the prediction successful decision threshold T2, T2=SAD
Thrh_2=SAD
Pred(1+ β
2) and cut-off threshold T3 midway, T3=T
iWherein: SAD
PredAdopt Uplayer prediction, space median prediction, the prediction of consecutive frame corresponding blocks and four kinds of minimum predicted values that prediction mode obtains of adjacent multi-reference frame prediction, 7 kinds of block modes for 16 * 16,16 * 8,8 * 16,8 * 8,8 * 4,4 * 8,4 * 4 sizes, β 1=0.02,0.02,0.02,0.03,0.04,0.04,0.05; β 2=0.05,0.10,0.10,0.12,0.125,0.125,0.15; T
7=(2
Qbits-f)/QE[q
Rem] [0] [0]＜TH
7(2
Qbits-f)/QE[q
Rem] [0] [0]: TH
7Qbits=15+QP/6 wherein, q
Rem=QP%6, f=(1＜＜qbits)/6, QE refers to defined quantization parameter table, QP is a quantization parameter;
(c) carry out the CPFMS search, may further comprise the steps:
(1) according to the matching error SAD ' of the corresponding match point of current macro motion vector computation optimum prediction motion vector of median prediction, UpLayer prediction, consecutive frame prediction and the prediction of multiframe reference prediction, if SAD '＜T3, then withdraw from search, this match point is global optimum's match point; Otherwise the SAD that the prediction of SAD ' and zero vector is corresponding is relatively, and adopting the less point of error to carry out step-length as the search window central point is five point search that " ten " font of 2 distributes, and Select Error less o'clock as the current Optimum Matching point in the 2nd step;
(2) if X
OPT=0 and Y
OPT=0, then: if 1. SAD
OPT＞T2 is that to carry out step-length again be five point search that " ten " font of 2 distributes to search center with current optimum point; If 2. SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＞T1 is that to carry out step-length be 9 rectangular searches of 1 to central point with current optimum point; If 3. SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＜T1 is that to carry out step-length be 9 rectangular searches of 1 to central point with 2 line mid points between search window central point and the inferior advantage;
(3) if X
SOPT=0 and Y
SOPT=0, then: if 1. SAD
OPT＞T2 is that to carry out step-length again be five point search that " ten " font of 2 distributes to search center with current optimum point; If 2. SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＞T1 is that to carry out step-length be 9 uniform rectangular searches of 1 to central point with current optimum point then; If 3. SAD
OPT＜T2, and SAD
OPT＜T1, SAD
SOPT＜T1 is that to carry out step-length be 9 uniform rectangular searches of 1 to central point with 2 line mid points between search window central point and the optimum point then;
(4) if the current search central point neither optimum point neither time advantage, then: if 1. SAD
OPT＞T2 or current optimum point and time advantage diagonal angle distribute, with current optimum point be search center to carry out step-length again be five point search that " ten " font of 2 distributes, and got back to for (2) step and constitute circulation; Otherwise carry out following steps; If 2. SAD
OPT＜T1, SAD
SOPT＞T1 is that to carry out step-length be 9 uniform rectangular searches of 1 to central point with current optimum point; If 3. SAD
OPT＜T1, SAD
SOPT＜T1 is that to carry out step-length be 9 uniform rectangular searches of 1 to central point with 2 line mid points between optimum point and the inferior advantage;
Each step in above-mentioned search procedure is if the matching error of optimum point satisfies SAD
OPT＜T3, current optimum point is global optimum's point, ends search and gets final product.
2, according to the integer pixel quick mixing search method based on the center prediction of claim, it is characterized in that: described step (a) only adopts zero vector prediction and UpLayer prediction for the piece of 16 * 8,8 * 16,8 * 8 sizes; Adopt the median prediction of zero vector prediction, UpLayer prediction and spatial domain for the piece of 4 * 8,8 * 4,4 * 4 sizes; For the piece of 16 * 16 sizes, if corresponding reference frame 1,2,3,4 only adopts zero vector prediction and multi-reference frame prediction; Corresponding reference frame 0, predictive mode is selected in the prediction three of zero vector prediction, median prediction, consecutive frame corresponding blocks.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 200810017569 CN100551071C (en) | 2008-02-29 | 2008-02-29 | Integer pixel quick mixing search method based on the center prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 200810017569 CN100551071C (en) | 2008-02-29 | 2008-02-29 | Integer pixel quick mixing search method based on the center prediction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101237580A true CN101237580A (en) | 2008-08-06 |
| CN100551071C CN100551071C (en) | 2009-10-14 |
Family
ID=39920887
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 200810017569 Expired - Fee Related CN100551071C (en) | 2008-02-29 | 2008-02-29 | Integer pixel quick mixing search method based on the center prediction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN100551071C (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101815218A (en) * | 2010-04-02 | 2010-08-25 | 北京工业大学 | Method for coding quick movement estimation video based on macro block characteristics |
| CN101835047A (en) * | 2010-04-30 | 2010-09-15 | 中山大学 | Fast UMHexagonS motion estimation algorithm based on residual error reduction rate |
| WO2013107028A1 (en) * | 2012-01-19 | 2013-07-25 | Mediatek Singapore Pte. Ltd. | Methods and apparatuses of amvp simplification |
| CN105282557A (en) * | 2015-09-28 | 2016-01-27 | 天津大学 | H264 rapid movement estimation method for prediction movement vector |
| CN106034238A (en) * | 2015-03-11 | 2016-10-19 | 杭州海康威视数字技术股份有限公司 | Methods and systems for calculating coding distortion degree and coding mode control |
| CN111264061A (en) * | 2019-03-12 | 2020-06-09 | 深圳市大疆创新科技有限公司 | Method and apparatus for video encoding, and method and apparatus for video decoding |
| CN112492314A (en) * | 2020-11-25 | 2021-03-12 | 杭州微帧信息科技有限公司 | Method for selecting dynamic motion estimation algorithm based on machine learning |
| WO2021163862A1 (en) * | 2020-02-18 | 2021-08-26 | 深圳市大疆创新科技有限公司 | Video encoding method and device |
-
2008
- 2008-02-29 CN CN 200810017569 patent/CN100551071C/en not_active Expired - Fee Related
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101815218A (en) * | 2010-04-02 | 2010-08-25 | 北京工业大学 | Method for coding quick movement estimation video based on macro block characteristics |
| CN101835047A (en) * | 2010-04-30 | 2010-09-15 | 中山大学 | Fast UMHexagonS motion estimation algorithm based on residual error reduction rate |
| CN105898325B (en) * | 2012-01-19 | 2019-11-19 | 寰发股份有限公司 | The derivation method and device of motion vector predictor candidate item set |
| WO2013107028A1 (en) * | 2012-01-19 | 2013-07-25 | Mediatek Singapore Pte. Ltd. | Methods and apparatuses of amvp simplification |
| CN104054349A (en) * | 2012-01-19 | 2014-09-17 | 联发科技（新加坡）私人有限公司 | Method and apparatus for simplified motion vector predictor derivation |
| CN105898325A (en) * | 2012-01-19 | 2016-08-24 | 联发科技（新加坡）私人有限公司 | Method and device for deriving sub-candidate set of motion vector prediction |
| US9912952B2 (en) | 2012-01-19 | 2018-03-06 | Hfi Innovation Inc. | Method and apparatus for simplified motion vector predictor derivation |
| CN104054349B (en) * | 2012-01-19 | 2017-04-26 | 寰发股份有限公司 | Method and apparatus for deriving motion vector predictor candidate set |
| US9807396B2 (en) | 2012-01-19 | 2017-10-31 | Hfi Innovation Inc. | Method and apparatus for simplified motion vector predictor derivation |
| CN106034238A (en) * | 2015-03-11 | 2016-10-19 | 杭州海康威视数字技术股份有限公司 | Methods and systems for calculating coding distortion degree and coding mode control |
| CN106034238B (en) * | 2015-03-11 | 2019-04-02 | 杭州海康威视数字技术股份有限公司 | The method and its system of the calculation code distortion factor and coding mode control |
| CN105282557B (en) * | 2015-09-28 | 2018-05-18 | 天津大学 | A kind of H.264 rapid motion estimating method of predicted motion vector |
| CN105282557A (en) * | 2015-09-28 | 2016-01-27 | 天津大学 | H264 rapid movement estimation method for prediction movement vector |
| CN111264061A (en) * | 2019-03-12 | 2020-06-09 | 深圳市大疆创新科技有限公司 | Method and apparatus for video encoding, and method and apparatus for video decoding |
| WO2020181504A1 (en) * | 2019-03-12 | 2020-09-17 | 深圳市大疆创新科技有限公司 | Video encoding method and apparatus, and video decoding method and apparatus |
| WO2021163862A1 (en) * | 2020-02-18 | 2021-08-26 | 深圳市大疆创新科技有限公司 | Video encoding method and device |
| CN112492314A (en) * | 2020-11-25 | 2021-03-12 | 杭州微帧信息科技有限公司 | Method for selecting dynamic motion estimation algorithm based on machine learning |
| CN112492314B (en) * | 2020-11-25 | 2024-05-14 | 杭州微帧信息科技有限公司 | Dynamic motion estimation algorithm selection method based on machine learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN100551071C (en) | 2009-10-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN100551071C (en) | Integer pixel quick mixing search method based on the center prediction | |
| Ahn et al. | A novel fast CU encoding scheme based on spatiotemporal encoding parameters for HEVC inter coding | |
| Shen et al. | Effective CU size decision for HEVC intracoding | |
| CN101699865B (en) | Method for fast searching mass movement self-adapting sub pixel | |
| CN101394566B (en) | Cross rhombic motion estimation searching method | |
| KR101344200B1 (en) | Dynamic selection of motion estimation search ranges and extended motion vector ranges | |
| US9479786B2 (en) | Complexity allocation for video and image coding applications | |
| US8040948B2 (en) | Method and system for coding moving image signals, corresponding computer program product | |
| Shen et al. | Low-complexity mode decision for MVC | |
| Liu et al. | An efficient intermode decision algorithm based on motion homogeneity for H. 264/AVC | |
| CN101420617A (en) | Motion estimation searching method for cross hexagon | |
| CN103237222B (en) | The method for estimating of multi-mode search | |
| Ma et al. | Low complexity rate distortion optimization for HEVC | |
| CN108076347A (en) | A kind of acquisition methods and device for encoding starting point | |
| Huang et al. | Rate-distortion-complexity optimized coding mode decision for HEVC | |
| Shi et al. | A motion estimation algorithm based on Predictive Intensive Direction Search for H. 264/AVC | |
| WO2007089916A2 (en) | Dynamic reference frame decision method and system | |
| US20120128059A1 (en) | Method of adaptive motion estimation in search windows for video coding | |
| Lee et al. | Fast H. 264/AVC motion estimation algorithm using adaptive search range | |
| KR20120110795A (en) | Device and method for adjusting search range | |
| CN103458237B (en) | The determination method and apparatus of Video coding mode and method for video coding and device | |
| KR20080022843A (en) | Method for adjusting search area using motion vector | |
| Yang et al. | A fast inter-frame encoding scheme using the edge information and the spatiotemporal encoding parameters for HEVC | |
| Hong et al. | Optimization method of integer pixel motion estimation algorithm based on HEVC | |
| Lou et al. | Motion vector search window prediction in memory-constrained systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C17 | Cessation of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20091014 Termination date: 20120229 |