CN1777289A - Method for speeding up motion estimation utilizing selective prediction - Google Patents

Abstract

The invention provides a method for speeding up motion estimation utilizing selective prediction, relating to square matching method. At searching most matched prediction block with current block in adjacent frame, firstly using left block MV and upper block MV as current block MV to conduct direct prediction, if direct prediction is unsuccessful, using motion position generating smaller SAD at left block, upper block direct prediction as initial point of thinning searching, then calling simple diamond search pattern to make thinning search. The method has advantages of quick square matching speed, high matching quality, and small integral operand for video compression coding, and can simplifying real time video coding system and reduced cost of actual coding system.

Description

Utilize selective prediction to accelerate the method for estimation

Technical field

The present invention relates to a kind of estimation and compensation method, belong to the compression of digital video coding field.

Background technology

The compression of digital video coding is the research focus in current international community Signal and Information Processing field, is with a wide range of applications and remarkable economical and social benefit.Motion estimation and compensation is one of key technology of compression of digital video, be used to remove the temporal correlation of image sequence, it forms the whole video compressing and coding system with the entropy coding of the orthogonal transform technology of eliminating spatial coherence, elimination statistic correlation, and the performance of orthogonal transform and entropy coding is had a significant impact.

Estimation is the part of video coding stage operand maximum, and the real-time of video compression and the complexity of system are played decisive influence.The method for estimating that exists has block matching algorithm (BMA), PRA, phase correlation method, gridding method and global motion parameter estimation method etc. at present, wherein most widely used is BMA, H.26x, obtained immense success in the serial coding standard such as MPEG-x.BMA is divided into the image block of the fixed size of non-overlapping copies with current encoded frame, and supposes that all pixels are all done identical translational motion in the piece, is generally 16 * 16 pixels as block size.As piece, all search and its picture piece that mates most are as prediction in consecutive frame in the present frame each, and this a pair of picture piece is called motion vector (MV) in the distance of level and vertical process.Only need during coding the error and the MV of current block and prediction piece are encoded, decoding end just can reconstruct present image fully by the former frame image.Because the digital video per second shows 30 two field pictures approximately, the variation of adjacent two two field pictures is very little, and the data volume of residual image is compared with former data and significantly reduced, that is estimation and compensation can remove the temporal correlation of picture signal effectively, realizes Information Compression.

The most direct effective BMA is full search method (FS), and it is point by point search in the reference frame search district, and forecast quality is best, but the operand when image resolution ratio height, need extensive search is huge, realizes that the difficulty of handling in real time is bigger, the system complexity height.Many fast algorithms have been produced in past more than 20 year, basic thought all is to skip the search point that less error is unlikely arranged, under the prerequisite that keeps forecast quality, reduce search point, accelerate the speed of block matching, relatively successful fast algorithm has three-step approach (TSS), diamond search method (DSA) etc.But existing various fast algorithms all come with some shortcomings more or less, and the efficient that speed promotes is not high, and can reduce precision of prediction in some cases, makes video quality relatively poor.

Design a kind of algorithm fast and effectively, reach the purpose that matching speed is fast, precision of prediction is high simultaneously, need to study the fundamental characteristics and the statistical law of digital video sequences sports ground, and be used, design effective algorithm structure according to every attribute proposition correct method.

Summary of the invention

Technical problem to be solved by this invention be propose a kind of can block matching speed very fast, quality of match is higher, the integral operation amount of video compression coding is few, and can simplify the real-time video coded system, reduces the estimation and the compensation method of actual coding system cost.

For addressing the above problem, the present invention proposes the method (being referred to as SPSA herein) that a kind of selective prediction is accelerated estimation, wherein, coded frame is divided into the image block of non-overlapping copies, fixed size, to in the present frame each as piece, all search and its picture piece that mates most adopt certain the prediction piece as piece in the following step searching present frame as the prediction piece in consecutive frame:

(1) selected matching criterior, and selected direct decision threshold T;

(2) suppose motion vectors MV _PdtWith left block motion vector MV _LeftIdentical, calculate this as the piece and the first first error amount SAD of prediction between the piece that supposed _Left

(3) if first error amount less than direct decision threshold T, is this prediction piece as piece with the first prediction piece of being supposed then, finish search; Otherwise, suppose motion vectors MV _PdtWith last block motion vector MV _TopIdentical, calculate this as the piece and the second second error amount SAD of prediction between the piece that supposed _Top

(5) if second error amount less than direct decision threshold T, is this prediction piece as piece with the second prediction piece of being supposed then, finish search; Otherwise, compare the first error amount SAD _LeftWith the second error amount SAD _Top, serve as the prediction search center with the pairing motion vector of reckling among both;

(6) being the search starting point with the prediction search center, adopting simple little diamond search pattern (SDSP) to move dot matrix along the minimal error direction, when minimal error point finishes during at the dot matrix center to search for, is this prediction piece as piece with the dot matrix center.

As preferred version, preferably select absolute error and criterion (SAD) as matching criterior.

Beneficial effect of the present invention is as follows:

1. only use the left piece of current block, last piece as predictor.Though also there is correlation in time prediction (preceding frame same position picture piece) with current block, in existing other method, employing is arranged more, but the time prediction only prediction under static situation is more accurate, and is also incompatible with the hybrid predicting mode that reality adopts, and increases system complexity.For other space adjacent block in other algorithm also as predictor, but the correlation of left piece, last piece and current block is the strongest, and other adjacent block and left piece, go up piece and also have correlation, there is bigger redundancy when using simultaneously, and can greatly significantly increases operand.

2. propose the selective prediction mode, improved prediction accuracy.See that on the whole all there is correlation in the adjacent picture piece with each of current block, should get the average of each relevant picture piece or weighted mean as prediction.Yet specific to a certain, current block only with the adjacent picture piece kinematic similitude that belongs to same motion object, especially concerning the picture piece that is in motion object bounds place, use mean prediction or median prediction must obtain inaccurate result.In the searching method that the present invention proposes, predicated error when doing prediction according to each predictor selects wherein to predict that the most adjacent picture piece is as predictor, avoided the blindness of prediction, effectively improved prediction accuracy, for the bulk velocity of accelerating block matching lays the foundation.Accuracy and the precision of only using left piece, last piece to make the selection Forecasting Methodology of predictor are higher than mean prediction, median prediction.

3. be illustrated in figure 4 as the motion estimation unit of SPSA, SPSA compares with traditional full search method (FS), only need increase motion vector buffer memory (shown in frame of broken lines among the figure), and required memory span is very little, and system complexity is lower, is easy to hardware and realizes.

4. compare SPSA (T=512) and FS, TSS, representational algorithms such as DSA, PMVFAST, tested the sequence that a plurality of scenes, content and motion power have nothing in common with each other respectively, with Y-PSNR (PSNR) the comparison match quality of luminance component.Experimental result shows:

A.SPSA has on average improved 123 times than FS speed, quality of match decline 0.06dB (only being equivalent to 2 ‰), and each frame estimated performance is all approaching with FS, illustrates that SPSA had both improved speed, had also guaranteed quality and stable performance.

B.SPSA speed is 2.3 times of PMVFAST, and quality on average improves 0.31dB, and performance is the most obvious in comprising the sequence of global motion.As seen the combination property of SPSA is far superior to PMVFAST, and algorithm structure is easy to realize relatively.

The speed of c.SPSA is respectively TSS, DSA 6.0,4.8 times, and quality improves 0.61dB, 0.40dB simultaneously.

Description of drawings

Fig. 1: little diamond search (SDSP) dot matrix

Fig. 2: little diamond search (SDSP) path

Fig. 3: selective prediction searching algorithm (SPSA) flow chart

Fig. 4: the quick block matching algorithm unit of selectivity block diagram

Embodiment

Below in conjunction with embodiment and accompanying drawing the present invention is described in further details.

In the block matching method, in the present frame each as piece (or being called macro block) all need in reference frame, search for one with it one of coupling as the prediction piece, three kinds of matching criterior commonly used are arranged at present:

(1) absolute error and (SAD, Sum of Absolute Difference) criterion;

(2) mean square error (MSE, Mean Square Error) criterion;

(3) Normalized Cross Correlation Function (NCCF, Normalized Cross Correlation Function) criterion.

In above-mentioned three kinds of criterions, the SAD criterion has the multiplying of not needing, realizes advantage simply and easily, thereby searching method of the present invention preferably selects for use SAD as matching criterior.In the block matching algorithm, SAD is defined as:

D is any macro block in the formula, and M, N are that the level and the vertical spline of macro block counted _t(), _T-1() represents present frame and reference frame respectively, and d (p) is the distance of macro block in current macro and the reference frame.

The selective prediction that utilizes that the present invention proposes is accelerated the method for estimation, belong to a kind of block matching algorithm, before current block is carried out estimation, its left piece and last piece finished estimation, two all found MV (when left piece and last piece think that its MV is 0 when exceeding image boundary).When in consecutive frame, searching for the prediction piece that mates most with current block, at first with left piece MV as current block MV, predicated error SAD on this position when calculating this MV, if this SAD is less than default thresholding T, the motion that current block and left piece are described is identical, just no longer do other search, directly with the motion of left piece motion as current block.If not so, illustrate that the current block motion is inequality with left piece motion, above again piece MV does identical computing relatively.

If above direct prediction of failure illustrates that then left piece, last piece motion and current block motion are all inequality, need do further search (search refinement) and just can provide the most similar position.Though direct prediction of failure, the motion of current block is different with adjacent block, and their motion is similar.Still above piece, left piece produce the starting point of the position of less SAD person motion when directly predicting as search refinement, call simple little diamond search pattern (SDSP) again and make search refinement.

The little diamond search dot matrix (SDSP) that the present invention adopts is made up of 5 search points, promptly at every turn by these 5 positions of dot matrix distribution form search, seeks the position that wherein produces minimum SAD, referring to Fig. 1 and Fig. 2.If minimum SAD, illustrates best match position at the dot matrix center and is current dot matrix center, because all bigger error can occur toward the words of any direction mobile search point; If minimum SAD, then moves to the dot matrix center current minimum SAD position not in the center and continues to check that whether minimum SAD is in the dot matrix center.So search finishes search until minimum SAD during at the dot matrix center, with last dot matrix center as best match position.

Fig. 3 is selective prediction searching algorithm of the present invention (SPSA) flow chart.If MV _Left, MV _Top, MV _PdtAnd MV _OptBe respectively left piece MV, go up piece MV, prediction MV and optimum MV, SAD _Left, SAD _TopBe with MV _Left, MV _TopError during for prediction, T is the direct decision threshold of MV.The present invention adopts the SAD search criteria, and direct decision threshold T is set earlier before the search.At first suppose motion vectors MV _PdtWith left block motion vector MV _LeftIdentical, calculate the prediction piece directly predicted according to left piece MV and the residual error of current block.If SAD _LeftLess than direct decision threshold T, serve as the prediction piece then with this piece, finish search; Otherwise, suppose motion vectors and last block motion vector MV again _TopIdentical, calculate the prediction piece directly predicted according to last piece MV and the residual error of current block.If SAD _TopLess than direct decision threshold T, be this prediction piece then as piece with this piece, finish search; Otherwise relatively two residual values serve as the prediction search center with the pairing motion vector of reckling among both, as the search starting point of further carrying out simple little rhombus search refinement.

Distance between this moment and the optimum MV is generally less, can satisfy central distribution better, moves dot matrix with simple little diamond search pattern along the minimal error direction, and when minimal error point finishes during at the SDSP center to search for, the dot matrix center is optimum MV.

Claims (2)

1, a kind of method of utilizing selective prediction to accelerate estimation, wherein, coded frame is divided into the image block of the fixed size of non-overlapping copies, to in the present frame each as piece, all in consecutive frame the search with its mate most the picture piece as the prediction piece, it is characterized in that, adopt certain the prediction piece in the following step searching present frame as piece:

(1) selected matching criterior, and selected direct decision threshold (T);

(2) suppose motion vectors (MV _Pdt) and left block motion vector (MV _Left) identical, calculate this as the piece and the first first error amount (SAD of prediction between the piece that supposed _Left);

(3) if first error amount less than direct decision threshold (T), is this prediction piece as piece with the first prediction piece of being supposed then, finish search; Otherwise, suppose motion vectors (MV _Pdt) and last block motion vector (MV _Top) identical, calculate this as the piece and the second second error amount (SAD of prediction between the piece that supposed _Top);

(5) if second error amount less than direct decision threshold (T), is this prediction piece as piece with the second prediction piece of being supposed then, finish search; Otherwise, compare the first error amount (SAD _Left) and the second error amount (SAD _Top), serve as the prediction search center with the pairing motion vector of reckling among both;

(6) being the search starting point with the prediction search center, adopting simple little diamond search pattern (SDSP) to move dot matrix along the minimal error direction, when minimal error point finishes during at the dot matrix center to search for, is this prediction piece as piece with the dot matrix center.

2, the method for utilizing selective prediction to accelerate estimation according to claim 1 is characterized in that selected matching criterior is absolute error and criterion (SAD).