CN101309424B - Quick movement estimating method - Google Patents

Abstract

The invention relates to a video compression technology, in particular to a fast motion estimating method. The invention provides a motion estimating method which can reduce the possibility that the local optimum is involved and can not be influence the signal noise ratio of the image basically. The video compression technology adopts the technical proposal that the fast motion estimating method comprises the following steps: firstly, the probability that the motion vector of the block in the image is equal to a predictive motion vector are equal can be counted; secondly, when the motion vector is detected, the predictive motion vector can be detected in sequence from the large vector to the small vector according the counted probability in the first step; thirdly, the termination judgment can performed to the motion searching by adopting a valve value. By the adoption of the fast motion estimating method, the unnecessary searching can be reduced, the searching efficiency can be enhanced, and the fast motion estimating method can be suitable for the high-definition video encoding and the real-time encoding.

Description

A kind of method of fast motion estimation

Technical field

The present invention relates to video compression technology, relate in particular to the rapid motion estimating method in the video compression technology.

Background technology

Estimation is a most important component in the video coding system, and it is to the quality important influence of coded video sequences.Thereby the motion estimation and compensation technology can effectively be removed the redundant efficient coding of realizing of interframe in the image sequence.Motion estimation and compensation technology based on the piece coupling is adopted by a lot of video encoding standards with the realization that is easy to software and hardware owing to algorithm is simple, as MPEG-1/2/4 and ITU-TH.261/263/263+/264 etc.Its basic thought is that coded image is divided into the identical macro block of size, macro block can also be divided littler piece, for each piece, by certain matching criterior search immediate with it match block-prediction piece, the displacement of this prediction piece and current block is exactly motion vector (MV) in its reference frame.It is accurate more to predict, the used bit number of encoding is few more, just economizes on resources more.In order to find match block as well as possible, encoder has spent the estimation of taking exercises of a large amount of time, particularly H.264 adopted multiframe with reference to and the multiple pattern of cutting apart, make the complexity of estimation sharply rise, estimation has occupied a large amount of scramble times.

In present numerous searching algorithms, the estimated accuracy of full-search algorithm (FS) is the highest, and by search each candidate point in the matching window, FS can search out global optimum's motion vector, is unfavorable for real-time application but its computation complexity is too high; Three step search methods (TSS), hexagon search (HEXS), diamond search traditional searching algorithms such as (DS) are absorbed in the local optimum easily, can't find global optimum's point, and search speed also can't be improved.Present searching algorithm of the prior art all is the mode of taking by matching criterior direct search match block basically, and does not adopt the motion vectors statistical probability preferentially to estimate search, has wasted the plenty of time, and efficient can not get improving.And searching algorithm of the prior art adopts the single-stage threshold value to stop differentiating mostly, for example: the MVFAST searching algorithm just is to use is to be exactly fixing single-stage threshold value 512, for some motions with texture complex video source 512 these threshold values may not well meet the demands and owing to can not differentiate signal to noise ratio (PSNR) influence of image very big by premature termination.Particularly for the video sequence of compound movement, although search point reduces, the signal to noise ratio of image has tangible reduction.

Summary of the invention

Technical problem to be solved by this invention is: propose the method for estimating that a kind of minimizing is absorbed in the local optimum possibility and does not influence signal noise ratio (snr) of image substantially.

The technical scheme that the present invention solves the problems of the technologies described above employing is: a kind of method of fast motion estimation may further comprise the steps:

A. add up the probability size that the motion vector of piece in image equates with 4 motion vectors A, B, C, D;

B. when detecting motion vector, according to the descending motion vectors that detects successively of the probability that comes out among the above-mentioned steps a;

C. " if satisfy: D_MV＞Td, then carry out the hexagon iterative search earlier, carry out little diamond search again, otherwise directly carry out little diamond search; Wherein $D\_\mathrm{MV}=\underset{N=A,B,C,D}{\mathrm{\Σ}}|\mathrm{MV}\_N-\mathrm{MVP}|$ The absolute value sum of the motion vector of expression A, B, C, D piece and the difference of median prediction motion vector MVP, Td is a threshold value of judging distance between the motion vector, is a default fixed value;

D. adopt threshold value that motion search is stopped differentiating.

In the described steps d, adopt at least the two-stage threshold value that motion search is carried out premature termination and differentiate, the condition that satisfies any one-level threshold value promptly stops search.

The invention has the beneficial effects as follows: reduced unnecessary search, improved search efficiency, also avoided being absorbed in the possibility of local optimum.

Embodiment

The invention will be further described below in conjunction with embodiment.

Proposition of the present invention mainly according to motion vector in time or have very high correlation on the space.In block-based estimation, because the continuity of object of which movement, be that consecutive frame correspondence position piece has correlation in time, because object and background in the image have all covered considerable, and the motion vector field that belongs to same target or background has very big similitude, and their distortion value (SAD) of correspondence has certain relation in addition.Spatially, current block and adjacent block are in adjacency state, and motion relevance is more tight.The SAD that does the object each several part of rigid motion is similar.Yet the motion of image is complicated on the whole, but since the coding in, image division a lot of fritters, can think that these fritters all are to do rigid motion.Sad value between the adjacent block is approaching more, and the transition between illustrated block and the piece is more level and smooth, and search this moment only adopts little template search to get final product.And, need to use earlier the search of large search template this moment for the big or texture more complicated of the bigger account for motion of the SAD difference between the adjacent block, re-use the search that little template becomes more meticulous, the best matching blocks of piece in searching image.

Embodiment:

Rapid motion estimating method in this example adopts a plurality of motion vectors: intermediate value motion vector (intermediate value MVP), former point motion vector MV (0,0), MV_A (A block motion vector), MV_B (B block motion vector), MV_C (C block motion vector), previous reference frame corresponding blocks motion vector MV_COL _T-1Or the convergent-divergent MV_SC of motion vector on other reference picture.The probability size that the motion vector of statistics piece in image equates with these motion vectors, statistics find, the probability maximum be intermediate value MVP, next is followed successively by MV (0,0), MV_COL _T-1(multiframe with reference to the time non-No. 0 reference frame use

$\mathrm{MV}\_\mathrm{SC}=\mathrm{MVref}0\×\frac{\mathrm{TRn}}{\mathrm{TR}0}$

Replace MV_COL _T-1), MV_A, MV_B and MV_C.Wherein, MVref0 represents the motion vector of current block in first reference picture, and TRn represents the time difference of n reference picture and present image; TR0 represents the time difference of first reference picture and present image.

The first step: detect earlier intermediate value MVP and MV (0,0), if satisfy condition SAD＜minSAD+T1, then premature termination is differentiated and is calculated and store last motion vector MV and sad value.Wherein, minSAD gets SAD_A (A piece distortion value), SAD_B, SAD_C, SADcol _T-1Minimum value in (distortion value of last reference frame corresponding blocks), T1 is the threshold value relevant with block size: a T1=a * s, and wherein, a is a correction factor, and s is that the area of piecemeal is counting of piecemeal.

Second step: make SAD_X=PrevSAD (PrevSAD is the sad value of the corresponding blocks of a nearest frame)

Detect MV_COL _T-1, MV_A, MV_B and MV_C, if satisfy:

SAD＜SAD_X+T2

Then premature termination is differentiated, and calculates and store last motion vector MV and sad value.Wherein, T2=b * s, b are correction factor, and s is that the area of piecemeal is counting of piecemeal.

The 3rd step:, then need to carry out following processing: make PredMV[0 if above-mentioned vector detection does not reach termination ... N-1] be respectively: MV_A, MV_B, MV_C ..., MV_COL _T-1And PredSAD[0 (MV_SC) ... N-1] be respectively the sad value of above motion vector correspondence.

What this section will be realized is, find identical motion vectors PredMV, from these identical motion vectors (the corresponding sad value of each motion vector), find out the sad value and the minimum sad value of the maximum of their correspondences then with current motion vector MVcurr

The distance of definition motion vector is:

|MV1-MV2|＝|MV1_X-MV2_X|+|MV1_Y-MV2_Y|

$D\_\mathrm{MV}=\underset{N=A,B,C,D}{\mathrm{\Σ}}|\mathrm{MV}\_N-\mathrm{MVP}|$

Be that motion vector distance is: A, B, the absolute value of the motion vector of C and D piece and the difference of motion vectors MVP with

If satisfy:

D_MV＞Td

Or

|maxSAD_X-minSAD_X＞T3

Then carry out the hexagon iterative search earlier, carry out little diamond search again.Wherein, Td is a threshold value of judging distance between the motion vector, is a fixed value, can set before coding.

T3＝c×s。C is a correction factor, and s is that the area of piecemeal is counting of piecemeal.

If still do not satisfy above-mentioned condition, then order:

min?MV_X＝MIN(PredMV[0].X，...，PredMV[N-1]X)

max?MV_X＝MAX(PredMV[0].X，...，PredMV[N-1]X)

min?MV_Y＝MIN(PredMV[0].Y，...，PredMV[N-1]Y)

max?MV_Y＝MAX(PredMV[0].Y，...，PredMV[N-1]Y)

Be that minMV_X represents to get MV_A, MV_B, MV_C ..., MV_COL _T-1(MV_SC) minimum value of cross stream component; MaxMV_X represents to get MV_A, MV_B, and MV_C ..., MV_COL _T-1(MV_SC) maximum of cross stream component; Similar, minMV_Y represents to get the minimum value of the longitudinal component of these motion vectors; MaxMV_Y represents to get the maximum of the longitudinal component of these motion vectors.Detect following point then respectively: (MVP_X, minMV_Y), (MVP_X, maxMV_Y), (minMV_X, and MVP_Y) (maxMV_X_, MVP_Y).Carry out little rhombus (or little square) iterative search, fall into template center up to optimum point.

So just finished estimation fast, present embodiment has following advantage with respect to prior art:

(1) use multistage threshold value premature termination to differentiate, compare MVFAST and PMVFAST, to the strictness control of one-level threshold value can premature termination to influence on signal-to-noise ratio (SNR), the secondary threshold value has reduced unnecessary search, has improved the speed of search greatly; For the present invention, can adopt the above threshold value of secondary or two-stage, the many more search of progression are accurate more, also can consume the more time simultaneously, adopt three grades of threshold values proper.

(2) according to the distribution character and the temporal correlation of motion vector, provide how effective motion vectors, avoided three step of picture search method (TSS), hexagon search (HEXS), diamond search (DS) limit to go into local optimum possibility;

(3) multi-mode search: motion is used large search template, the search that becomes more meticulous with little template then than big or texture than the complex image piece.And only need use little template to search for for the image block of little motion or segment smoothing; Effectively differentiate the motion object;

(4) corresponding threshold parameter can be adjusted, the compromise of an optimum can be on coded image and estimation speed, made; Be widely used, can be used for HD video coding and real-time coding.

For motion vectors, can add up three or three above motion vectors equate probability with the motion vector of piece in the image size, successively motion vectors is detected according to probability is descending then.The number that motion vectors adopts is many more, searches for accurate more.

Claims (2)

1. the method for a fast motion estimation is characterized in that: may further comprise the steps:

A. add up the probability size that the motion vector of piece in image equates with 4 motion vectors A, B, C, D;

B. when detecting motion vector, according to the descending motion vectors that detects successively of the probability that comes out among the above-mentioned steps a;

C. " if satisfy: D_MV＞Td, then carry out the hexagon iterative search earlier, carry out little diamond search again, otherwise directly carry out little diamond search; Wherein

The absolute value sum of the motion vector of expression A, B, C, D piece and the difference of median prediction motion vector MVP, Td is a threshold value of judging distance between the motion vector, is a default fixed value;

D. adopt threshold value that motion search is stopped differentiating.

2. the method for a kind of fast motion estimation as claimed in claim 1 is characterized in that: in the described steps d, adopt at least the two-stage threshold value that motion search is carried out premature termination and differentiate, the condition that satisfies any one-level threshold value promptly stops search.