CN1107413C - Motion vector field coding - Google Patents

Abstract

The present invention relates to an image encoding decoder which comprises an encoder of a motion information group for minimizing the number of motion coefficients of an operating vector information group. A first block in the encoder of a motion information group comprises a device for generating a new matrix representation of an information group of motion vectors. A new information group of coded motion vectors is linear. A second main block is a device for merging adjacent section pairs, the adjacent section pairs are merged by the second main block under the condition that an information group of a combining section can be predicated by an information group of common motion, and merging information is transferred to a decoder. A third main block comprises a device for eliminating basic functions of a motion information group. Square prediction errors are calculated and an elimination process is continued to be executed after each elimination step until the square prediction errors can not be accepted. A final motion coefficient is obtained by computing a linear matrix equation. The present invention has a result that motion coefficients of reducing numbers of each section are obtained. The motion coefficients are transferred to the image encoding decoder.

Description

Motion vector field coding

Technical field

The relevant image compression of the present invention.More precisely, the invention relates to an a kind of estimating motion information sets of coding (motion field) and reach the method that in an image sequence, generates movable information.

Background of invention

Motion compensated prediction is the key link in most of picture coding schemes.Fig. 1 utilizes motion compensation technique at one, is used for the schematic diagram of an encoder of compressed image sequence.Essential elements in this encoder is 1, one motion estimator 2 of a motion-compensated prediction block and a movable information group coding device 3.The operation principle of moving compensation image encoder is the incoming frame I that is encoded that compression is called as present frame _n(x is y) with a predictive frame

Between the predicated error E of difference _n(x, y), wherein: $E_n(x,y)=I_n(x,y)-{\hat{I}}_n(x,y)---\left(1\right)$ Predictive frame

Make up by motion-compensated prediction block 1, and utilize the front or some other be called as reference frame, be designated as The pixel value of coded frame, and the motion vector of the pixel between present frame and reference frame is set up.Motion vector is calculated by movable information group estimation device 2, and then, the Vector Message group of gained was encoded before being transmitted to prediction piece 1 in some way.Predictive frame is: ${\hat{I}}_n(x,y)={\tilde{I}}_{n-1}[x+\mathrm{\Δx}(x,y),y+\mathrm{\Δy}(x,y)]---\left(2\right)$ Several to [x+ Δ x (and x, y), y+ Δ y (x, y)] be called as position in the present frame (x, the motion vector of the pixel of y) locating, and Δ x (x, y) (x y) is the displacement in the horizontal and vertical directions of this pixel with Δ y.Present frame I _n(x, y) in the motion vector of one group of all pixel be called as motion vector information group (motion vector field).The encoding motion vector information sets also is transmitted to decoder as movable information.

Fig. 2 is in decoder, by at reference frame

In obtain the prediction of picture element Reconstruct present frame I _n(x, picture element y).Motion-compensated prediction block 21 is utilized the movable information and the reference frame that receive (in the figure, reference frame is identical with present frame), the generation forecast frame.Then, in predicated error decoder 22, decoded predicated error E _n(x, y) with the predictive frame addition, the result is original present frame I _n

The general purpose of motion compensated prediction is to minimize the amount of information that need send decoder to.Should minimize by such as E _n(the predicated error amount that x, energy y) measure also will minimize the required amount of information of reproduction motion vector information group.

By H.Nguen, " movable information of reproduction picture coding " (" the Representation of motion information for image coding ") that E.Dubois showed, see proceedings, picture coding collection of thesis ' 90, Massachusetts, Cambridge, March nineteen ninety 18-26, the 841-845 page or leaf has provided the commentary of a movable information group coding technology in this article.Rule of thumb, reduce predicated error, just need more complicated movable information group, promptly must have more position to be used for coding.Like this, the general purpose of picture coding is an encoding motion vector information sets as far as possible closely, simultaneously, keeps the value of predicated error low as far as possible.

Fig. 1, movable information group prediction piece 1, the motion vector of all picture elements of a given section of calculating, it minimizes the value of predicated error in this section, for example square prediction error.Movable information group valuation technology is at the model of movable information group and make aspect the minimized algorithm two of selected prediction error value all different.

Because in a frame, the number of picture element is very big, so if will transmit the single motion vector of each picture element, efficient is very low.And in most picture coding schemes, present frame all is divided into big visual section, and like this, motion vector all in the visual section can be described by parameter seldom.The image section can be square block, for example, and according to international standard ISO/IECMPEG-1 or ITU-TH.261, coding decoder uses the piece of 16 * 16 picture elements, and perhaps, they can comprise the district's information sets that is entirely arbitrary shape, for example, the district's information sets that draws by segmentation algorithm.In actual applications, section comprises tens picture elements at least.

In order to reappear the motion vector of picture element in the section densely, their value can be described preferably with the function that few parameters is arranged.This function is called as the motion vector information group model.A kind of known model family is the linear movement model, and wherein, motion vector is the linear combination of movable information group basic function.In this model, the motion vector of image section is represented by a common-used formula: $\mathrm{\Δx}(x,y)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{c}_i{f}_i(x,y)---\left(3\right)$ $\mathrm{\Δy}(x,y)=\underset{i=N+1}{\overset{N+M}{\mathrm{\Σ}}}{c}_i{f}_i(x,y)$

Wherein, parameters C _iBe called as kinematic coefficient, they are transmitted to decoder.Function f _i(x y) is called as movable information group basic function, and they are fixed, and encoder is all known this function.

When utilization had the linear movement model of above-mentioned formula, the problem of existence was how to minimize the kinematic coefficient C that will be fed to decoder _iNumber, keep predicated error E simultaneously _n(x, value y) is low as far as possible.This process is in encoder, is carried out by movable information group coding piece 3, sees Fig. 1.It is carried out after piece 2 is finished movable information group valuation very complicated in the calculating.Thereby crucial is, the movable information group coding should be simple on calculating, and like this, it could not increase extra burden to encoder.

The number that need be transmitted to the kinematic coefficient of decoder depends on the number of kinematic coefficient in the number of section in the image and each section.Like this, have at least two kinds of approach can reduce the sum of kinematic coefficient.

First method is those can be predicted with a common motion vector information sets, and can not cause that those sections that predicated error increases significantly make up (merging) together, thereby reduce the number of section.The number of section can reduce in the frame, because, usually, be close to, that is, contiguous section can be predicted well by identical kinematic coefficient.The process that makes up this section is called as auxiliary merge (the motion assisted merging) of motion.

Second method is to choose a motion model for each section, and it can obtain gratifying low predicated error, and coefficient is few as much as possible.Because between frame and the frame and between section and the section, it is different that the quantity of motion and complexity all have, so each section is all always used the N+M movement function, effect is also bad.Be necessary for every section minimum number of all finding out the kinematic coefficient that can obtain gratifying low predicated error.This coefficient adaptively selected is called as kinematic coefficient cancellation (motion coefficient removal).

Fig. 3 shows a frame that is divided into section.In the prior art of kinematic coefficient coding, comprise that several motions assist folding.At the motion vector of all sections all after the valuation, carry out that motion is auxiliary to be merged.It is based on every couple of adjacent sections S _iAnd S _jAnd their kinematic coefficient C separately _iAnd C _jFinish.Combination section S _iAnd S _jDistrict's field mark be S _IjIf S _IjDistrict's information sets can be with one group of kinematic coefficient C _IjPredict, and can not cause than by independent prediction S _iAnd S _jWhen section gained error has the predicated error of excessive increase, S then _iAnd S _jMerged.The auxiliary method that merges of motion is at the single kinetic system array C that seeks the good prediction that can provide the section of combining _IjMode on, have essence difference.

A kind of method is arranged, be called as nothing omission motion estimation and merge.This method is every couple of adjacent sections S _iAnd S _jEstimate " starting all over again from the beginning " new kinematic parameter group C _IjIf S _IjPredicated error excessively do not increase section S then _iAnd S _jMerged.Although this method can be selected the merged section of energy well, it is also infeasible in realization, because generally speaking, it will make the complexity of encoder increase by several orders of magnitude.

Another kind method is called as to expand by the movable information group and merges.This method is judged district information sets S _IjWhether can utilize kinematic parameter C _iOr C _jPredict, and predicated error does not have excessive increase.The characteristics of this method are that computational complexity is very low, because it is without any need for new estimating motion.But, this method often can not merge the section of this merging because be motion compensation that a section calculates the gained coefficient seldom can be simultaneously the contiguous section of prediction well.

In addition, also have a kind of method, be called as by the match of movable information group and merge.In this method, calculate kinematic coefficient C by approximation method _IjThis is by estimating in each section that several motion vectors seldom finish.Fig. 3 has described section S _iAnd S _jIn some motion vectors.Utilize some known approximating methods, by these vectors, common motion vector information sets of match, thereby calculation of sector S _IjThe movable information group.The shortcoming of this method is, and is not accurate enough by the movable information group that match obtains, and regular meeting causes the unacceptable increase of predicated error.

H.Nicolas and C.labit showed " utilize the prejudgementing character scheme of approaching one by one to do estimating motion (" Region-based motion estimationusing deterministic relaxation schemes for image sequencecoding "); to see proceedings; 1994 annual acoustics international symposiums; language and signal processing; No. 265～268, III based on the interval for coding sequence of pictures; and the symmetrical directed split plot design of the estimating motion and the coding sequence of pictures in interval " efficiently based on " that P.Cicconi and H.Nicolas showed, see the IEEE collection of thesis, the Circuits and Systems of image technology, No. the 3rd, the 4th volume, in June, 1994, the 357th～364 page.In these two pieces of documents, proposed to carry out estimating motion and select only a kind of method with different models.These methods attempt to rely on the complexity of motion to adopt motion model by carrying out estimation and select only one with different models.The major defect of these methods is complexity that it calculates and in fact, can be for the quantity of the different motion information sets model that detects seldom.

How to make the kinematic coefficient C that transmits toward decoder _iNumber minimum, keep simultaneously predicated error En (x, is value y) as far as possible little? in the foregoing method, neither one can be independently solved this problem.

Summary of the invention

An object of the present invention is to generate a movable information group coding device, it can reduce the quantity of the movable information group Vector Message that is generated by some known motion estimator largely, and can not cause the bigger increase of predicated error.The complexity of movable information group coding device should be lower, so that in practice, can realize on available signal processor or general purpose microprocessor.

By the present invention, movable information group coding device comprises three main pieces.

First main piece is called as a QR motion analyzer.Its task is to find out the new expression of the input motion district information sets that is produced by movable information group estimator.This new expression is fed to second main piece.The operation of this first main piece comprises the many steps that contain matrix manipulation: in the first step, utilize some known approximation method linearisation predictive frames, so that predictive frame becomes the linear function about motion vector.In second step, structural matrix E _iAnd Y _i, be used for minimizing of square prediction error.In the 3rd step, use a known QR factoring algorithm, with matrix E _iBe decomposed into two matrix Q _iAnd R _iProduct.In addition, by coefficient matrix Q _iWith matrix Y _i, calculate an auxiliary vector Z _iMatrix R _iWith auxiliary vector Z _iA part be sent to second main piece.

Second main piece is called a section and merges piece.This piece is by judging that whether the right combination region information sets of adjacent sections can utilize a common motion district information sets prediction, carries out union operation.In matrix manipulation, at first form a Matrix Formula, surplus with known matrix computations method then, handle the factor matrix.The gained result is a matrix-style, and here, a matrix includes some, on its basis, can calculate the square prediction error in merging section district information sets at an easy rate.If according to a selected rule, the variation of square prediction error is can be received, and then these two sections are merged.

To after having been investigated, the output that section merges section is at all sections:

I. the new division of the image that reduced of sector number,

Ii. to each new section, this piece is output matrix R all _Ij ¹With vector Z _Ij ¹

Iii. be sent to the pooling information of decoder, it can help decoder to determine merged.

The 3rd main piece is called as coefficient cancellation piece, and this piece receives the present frame of importing of repartitioning with the form of section, and to each section, its receives by section and merges the matrix R that piece produces _k ¹, Z _k ¹And C _kEvery section motion vector is represented by some kinematic coefficients.Kinematic coefficient cancellation piece is every section and determines whether to simplify the movable information group model, and predicated error is excessively increased.Some basic functions are by cancellation from motion model, thereby only need less coefficient to describe the movable information group model of such simplification.

Operation in this 3rd main piece is a matrix manipulation, and wherein, at first cancellation delegation and row from coefficient matrix are revised Matrix Formula, then, Matrix Formula are carried out trigonometric ratio.Cancellation delegation and row are equivalent to basic function of cancellation from motion model.The variation of the square prediction error of the section that causes by basic function of cancellation equal in the formula one square.

If according to a selected rule, the variation of predicated error is can be received, then coefficient of cancellation from coefficient sets.Then repeat these matrix manipulations, can also reduce the more coefficient of this section.Behind the coefficient of cancellation sufficient amount,, calculate the final kinematic coefficient of this section by finding the solution final linear formula.This formula can utilize a kind of the finding the solution in the algorithm known, for example back-substitution algorithm.

The 3rd main piece is each processed section output selection information, and it shows has by cancellation from the movable information group model for which basic function.In addition, its output is corresponding to the new kinematic coefficient of residue basic function.Selection information and kinematic coefficient all are transmitted to decoder.

Accompanying drawing is described

Fig. 1 is the schematic diagram of a known encoder;

Fig. 2 is the schematic diagram of a known decoder;

Fig. 3 has described the adjacent sections that is used to merge;

Fig. 4 shows the merging of being done by motor area information sets approximation method;

Fig. 5 be one according to movable information group coding device of the present invention;

Fig. 6 is the sketch of a QR motion analyzer.

Embodiment describes

Fig. 5 shows according to movable information group coding device of the present invention.It is corresponding to the piece among Fig. 13, but its input also has reference frame and present frame.The 3rd input of this piece is the motion vector information group [Δ x (), Δ y ()] that is generated by movable information group valuation piece among Fig. 12.

Our hypothesis, the output of picture coding device is the condensed frame that is divided into section, and each section has kinematic coefficient, comprising P coordinate for one so is (x _i, y _i); I=1,2 ... the section of the picture element of P, the task of movable information group coding device are to find out a compressing motion vector information sets Kinematic coefficient, wherein motion vector is by a linear movement model representation, the form of motion vector information group is: $\widetilde{\mathrm{\Δ}}x(x,y)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\tilde{c}}_i{f}_i(x,y)---\left(4a\right)$ $\widetilde{\mathrm{\Δ}}y(x,y)=\underset{i=N+1}{\overset{N+M}{\mathrm{\Σ}}}{\tilde{c}}_i{f}_i(x,y)---\left(4b\right)$ Like this, it minimizes square prediction error: $\underset{i=1}{\overset{P}{\mathrm{\Σ}}}{(I_n(x_i,y_i)-{\tilde{I}}_{n-1}(x_i+\widetilde{\mathrm{\Δx}}(x_i,y_i),y_i+\widetilde{\mathrm{\Δy}}(x_i,y_i)))}^2.---\left(5\right)$

For finishing described task, movable information group coding device comprises three main pieces that make up, and they are QR motion analyzer pieces, and section merges piece and kinematic coefficient cancellation piece.Section merges the quantity of piece and kinematic coefficient cancellation piece minimizing movable information, consequently causes the increase of square prediction error.

The purpose of QR motion analyzer is to find out the new expression of motor area information sets.Be used in two pieces that are illustrated in the back that this is new so that fast and flexible find out kinematic coefficient that merges section and the cancellation that is used for coefficient.

The operation of QR motion analyzer comprises the steps:

The first step is the linearisation of error.In this step, utilize some known approximation methods, the predictive frame I in the approximation formula (7) _N-1() so that it with respect to

Linearisation.Like this, every coefficient C that becomes of summation in the formula (5) _iLinear combination $\underset{j=1}{\overset{P}{\mathrm{\Σ}}}{(e_{j,1}{\tilde{c}}_1+e_{j,2}{\tilde{c}}_2+\·\·\·+e_{j,N+M}{\tilde{c}}_{N+M}-y_j)}^2---\left(6\right)$

The 2nd step was a structural matrix.It is based on such fact, and promptly formula (6) minimizes and matrix expression Minimize fully and be equal to, E here _iAnd Y _iFor: $E_i=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{1,1}}& {\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">e</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& <figure-callout\; id="1"\; label="...\; e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">.</figure-callout>..& e_{}{}_{1,N+\mathrm{<figure-callout\; id="2"\; label="M\; e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">M</figure-callout>}}& \\ e_{}{}_{\mathrm{2,1}}& {\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">e</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& <figure-callout\; id="2"\; label="...\; e"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">.</figure-callout>..& e_{}{}_{2,N+M}\\ \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}\\ e_{P,1}& e_{P,2}& ...& e_{P,N+M}\end{array}\right],y=\left[\begin{array}{c}{y}_1\\ {\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">y</figure-callout>}}_2\\ \underset{.}{\overset{.}{.}}\\ y_P\end{array}\right]---\left(7\right)$

The 3rd step was the QR factorization." matrix computations " second edition that G.H.Golub and C.Van Loan are shown, famous QR factoring algorithm 1989, has been described by Johns Hopkins university press in this book.This algorithm is used for E _iBe decomposed into the product of two matrixes

E _i=Q _iR _i(8) in this step, also calculated a companion matrix Z _i, wherein

z _i＝Q ^T _iy _i (9)

In the 4th step, the output of calculating QR motion analyzer piece.Output packet contains a matrix R _i ¹, it contains matrix R _iPreceding N+M capable, also comprise a vector Z _i ¹, it contains Z _iPreceding N+M element.

In section merges piece, for adjacent sections to S _iAnd S _jCarry out union operation, see Fig. 4, this operation is by judging their combination region information sets S _IjWhether can utilize by kinematic coefficient C _IjA common motion information sets describing is predicted and is finished.Union operation may further comprise the steps:

The 1st step was matrix computations, and the present invention has used a unknown in advance parameter, kinematic coefficient C _IjCan obtain by finding the solution a system of linear equations: $\left[\begin{array}{c}{R^1}_i\\ {R^1}_j\end{array}\right]c_{\mathrm{ij}}=\begin{array}{c}\left[\begin{array}{c}{z^1}_i\\ {z^1}_j\end{array}\right]---\left(10\right)\end{array}$

R wherein _i ¹, Z _i ¹And R _j ¹Z _j ¹By QR analyzer piece respectively to section S _iAnd S _jFind the solution and obtain.

In the 2nd step, to the 1st step gained matrix trigonometric ratio.Matrix R _i ¹, R _i ¹For upper triangular matrix and equation group (10) have the form described in the document as previously mentioned:

Wherein symbol X represents nonzero element.According to what said in the above-mentioned document, by scalar each row is carried out a series of multiplying each other by the heel additional row, come the trigonometric ratio equation group, that is, it is converted into this form:

In the 3rd step, estimate pooled error, by section S _iAnd S _jMerge caused district information sets S _IjIn the change of square prediction error, calculate according to the method for being said in the above-mentioned document, wherein: ${\mathrm{\&Delta;E}}_{\mathrm{ij}}=\underset{k=1}{\overset{N+M}{\mathrm{\&Sigma;}}}{q^2}_k---\left(13\right)$

At last, in the 4th step, if according to selected rule, the change of square prediction error is can be received in the formula (13), then section S _iAnd S _jMerged, for the new section S of gained _Ij, capable by the preceding N+M that gets equation group (12), structural matrix R _Ij ¹With vector Z _Ij ¹, that is, provide by following formula:

Sections all in frame are obtained the output that section merges piece to after all investigating.Output comprises three category informations.At first, it provides the new division of the image that the number of section reduced.Secondly, it provides the piece output matrix R of each new section _Ij ¹, vector Z _Ij ¹Once more, it provides pooling information, and this information is sent in the decoder, helps decoder to discern merged section.Now, can be by the solving equation group $R_{\mathrm{ij}}^1{C}_{\mathrm{ij}}={\mathrm{<figure-callout\; id="1"\; label="Z\; ij"\; filenames="C9519800000201.png,C9519800000212.png"\; state="\{\{state\}\}">Z</figure-callout>}}_{\mathrm{ij}}^{},$ Calculation of sector S _IjKinematic coefficient C _Ij=(C ₁, C ₂... C _N+M), but, if use next piece, i.e. coefficient cancellation piece then there is no need to calculate them.

Now, we have a look the operation of coefficient cancellation piece.This piece receives the new division of the present frame of input with the form of section, and for each section S _k, it receives by front section and merges the matrix R that piece produced _k ¹, Z _k ¹And C _kEvery section motion vector is represented by the N+M kinematic coefficient.

To a given section S _k, kinematic coefficient cancellation piece determines whether to simplify the movable information group model, and does not have the excessive increase of predicated error.When some basic functions by from should with background technology in the model of the formula (3) told about during cancellation, the movable information group model that can obtain simplifying.The movable information group model of describing such simplification only needs less coefficient.

To each section, carry out following processes, its objective is and find out that i basic function (and i coefficient) whether can be by cancellation from the movable information group model:

The 1st step comprised matrix modifications, here, and by from R _k ¹In remove i row and from C _kIn remove i element, revise system of linear equations

R ¹ _kc _k＝z ¹ _k (15)

The 2nd step comprised the matrix trigonometric ratio, here, with a kind of known way, by the scalar by the heel additional row a series of row was multiplied each other, and made equation group (15) trigonometric ratio, that is, it is converted into such form:

The 3rd step comprised error estimation, the variation of the square prediction error of the section that is caused by i coefficient of cancellation and the q in the formula (16) _i ²Item equates.

The 4th step comprised the cancellation of coefficient.If according to a given rule, the change of predicated error is can be received, then coefficient C _iBy cancellation from coefficient sets.New coefficient number is N+M-1.Matrix R _k ¹With vector Z _k ¹Be modified to form:

Utilize the matrix (17) in the formula (15) and repeated for 1～4 step, also can further reduce coefficient number.

The 5th step comprised coefficient calculations.Behind the coefficient of enough numbers of cancellation, begin this step.In this step,, calculate a section S by finding the solution system of linear equations _kLast kinematic coefficient:

R ¹ _kc _k＝z ¹ _k (18)

Matrix R wherein _k ¹With vector Z _k ¹Be the front result in 1～4 step.Can utilize an algorithm known, for example back-substitution algorithm solving equation group.

Kinematic coefficient cancellation piece is each processed section output selection information, and it can tell decoder which basic function all to be arranged by cancellation from the movable information group model.In addition, its output is to being applied to the new kinematic coefficient of remaining basic function.Selection information and kinetic system number average are fed to decoder.

As in above all pieces result in steps, generate as encoder of the present invention: pooling information, which section quilt the notice decoder has merged; Selection information, which basic function the notice decoder has by cancellation; And kinematic coefficient information.

The present invention compares with prior art, and biggest advantage is that it has the ability that reduces movable information quantity largely, and can not cause the increase that predicated error is bigger.In addition, the complexity of whole system is very low, and in actual applications, it can be realized on existing signal processor or general microprocessor.

Section merges piece a distinctive ability, and it can be from given for finding out the motion vector of combination section the single section estimated motion vectors.Can prove that in fact, the motion vector that is produced by its is best in the square error that keeps combined segment aspect minimizing.This has just explained why this piece has and can significantly reduce the number of section, and only causes the ability that very little square prediction error increases.

For the actual quantity and the type of the motion in the instantaneous adaptation image of the motion model scene, kinematic coefficient cancellation piece is a strong tool.This piece can utilize very a large amount of models to detect at an easy rate and predict the outcome (the square prediction error value of a section), for example, utilize all possible combination of movable information group basic function.Known method without any other has big like this flexibility.A very big benefit of this scheme is, it does not need the repeating motion estimation process, thereby, on calculating, be very simple.

By after the QR motion analyzer, using estimating motion, movable information group coding device can be by finding the solution very simple system of linear equations, for the movable information group model in the combination of any desired image section or any desired section is found out new kinematic coefficient.

Most preferred embodiment is described

In most preferred embodiment, used quadratic polynomial motion vector information group model with 12 coefficients: Δ x (x, y)=c ₁+ c ₂X+c ₃Y+c ₄Xy+c ₅x ²+ c ₆y ²Δ y (x, y)=c ₇+ c ₈X+c ₉Y+c ₁₀Xy+c ₁₁x ²+ c ₁₂y ²

In actual applications, this model can be provided in the image sequence by very complicated motion and predicting the outcome of providing well.

In QR motion analyzer piece, by each the picture element (x around point _i, y _i) (i=1,2 ... P) locate to use I _N-1The Taylor expansion of () is finished the linearisation in the 1st step: x ' _i=x _i+ Δ x (x _i, y _i) y ' _i=y _i+ Δ y (x _i, y _i) utilize ∑ a ²=∑ (a) ²Characteristic, predicated error is: $\underset{i=1,}{\overset{P}{\mathrm{\&Sigma;}}}{({\tilde{I}}_{n-1}(x_i^{\&prime;},y_i^{\&prime;})+\left(\widetilde{\mathrm{\&Delta;}}x(x_i,y_i)-\mathrm{\&Delta;x}(x_i,y_i)\right)G_x(x_i^{\&prime;},y_i^{\&prime;})+(\widetilde{\mathrm{\&Delta;}}y(x_i,y_i)-\mathrm{\&Delta;y}(x_i,y_i))G_y(x_i^{\&prime;},y_i^{\&prime;})-I_n(x_i,y_i))}^2$ Utilize following formula to calculate instrumental value: g _j(x _i, y _i): Here, function f _j(x _i, y _i) be basic function as definition at formula (4a) and (4b).Matrix E in the formula (9) and vector Y are utilized the following formula structure: $\left[\begin{array}{cccc}{g}_1(x_1,y_1)& g_2(x_1,y_1)& ...& g_{N+M}(x_1,y_1)\\ g_1(x_2,x_2)& g_2(x_2,y_2)& ...& g_{N+M}(x_2,y_2)\\ \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}& \underset{.}{\overset{.}{.}}\\ g_1(x_P,y_P)& g_2(x_P,y_P)& ...& g_{N+M}(x_P,y_P)\end{array}\right],$ $y=\left[\begin{array}{c}{I}_n(x_1,y_1)-{\tilde{I}}_{n-1}(x_1^{\&prime;},y_1^{\&prime;})+G_x(x_1^{\&prime;},y_1^{\&prime;})\mathrm{\&Delta;x}(x_1,y_1)+G_y(x_1^{\&prime;},y_1^{\&prime;})\mathrm{\&Delta;y}(x_1,y_1)\\ I_n(x_2,y_2)-{\tilde{I}}_{n-1}(x_2^{\&prime;},y_2^{\&prime;})+G_x(x_2^{\&prime;},y_2^{\&prime;})\mathrm{\&Delta;x}(x_2,y_2)+G_y(x_2^{\&prime;},y_2^{\&prime;})\mathrm{\&Delta;y}(x_2,y_2)\\ .\\ .\\ .\\ I_n(x_P,y_P)-{\tilde{I}}_{n-1}(x_P^{\&prime;},y_P^{\&prime;})+G_x(x_P^{\&prime;},y_P^{\&prime;})\mathrm{\&Delta;x}(x_P,y_P)+G_y(x_P^{\&prime;},y_P^{\&prime;})\mathrm{\&Delta;y}(x_P,y_P)\end{array}\right]$

G _x(x, y) and G _y(x y) is reference frame I _N-1(they are calculated by following formula for x, level y) and vertical slope value: $G_x(x,y)={\tilde{I}}_{n-1}(x+1,y)-{\tilde{I}}_{n-1}(x-1,y),$ $G_y(x,y)={\tilde{I}}_{n-1}(x,y+1)-{\tilde{I}}_{n-1}(x,y-1).$

Fig. 6 shows the sketch of a QR motion analyzer.Row selects piece only to select the preceding N+M of input matrix capable.

Merge in the piece at section, used following section to merge skill:

A. select a thresholding T, it is corresponding to the square prediction error added value that allows in the entire frame,

B. the adjacent sections to all is right, utilizes formula (13) to calculate Δ E _Ij

C. Δ E _IjMinimum section is to being merged.

D. repeat b-c, up to corresponding to the Δ E of all merged sections to sum _IjGreater than T.

Δ E to sum _IjGreater than T.

For the trigonometric ratio of equation group (11), a series of Givens circulations have been used.

In kinematic coefficient cancellation piece, used the skill of following elimination factor: a. to select a thresholding T, it is corresponding to the permission added value of square prediction error in the entire frame.B. to all sections and all functions, utilize formula (16) to calculate q _i ²C. the minimum q of a section of cancellation _i ²The basic function of value.D. repeat b-c, up to in different sections all by all corresponding q of the basic function of cancellation _i ²And greater than T.

Use a series of Givens circulations, equation group (16) is realized trigonometric ratio.

Utilize back-substitution algorithm, by solution formula (18), the final kinematic coefficient of calculation of sector.

Pixel value I _N-1(x, y), G _x(x, y) and G _y(x y), only is integer coordinate points X and Y definition.At many X or Y is not under the integer-valued situation, comes calculating pixel values by the nearest picture element that the rounded coordinate value is arranged is carried out bilinear interpolation.

Under the situation that does not break away from the spirit and scope of the present invention, system can realize in a different manner.For example, in formula (3), can use different linear movement models.Can with diverse ways go in the linearization equations (5) the item.In addition, can remove to determine whether to merge two sections with different rules.Determine the skill whether a given basic function is somebody's turn to do cancellation from model that difference is also arranged.Can carry out the trigonometric ratio of matrix in formula (10) and (15) with different algorithms.And can separate the algorithm known of system of linear equations with some, solution formula (18) solves final coefficient.At last, can ask the I at non-integer coordinates point place with different interpolation algorithms _N-1(x, y), G _x(x, y) and G _y(x, value y).

Claims (11)

1. image coding-decoding device, the present frame I of the image sequence that is used to encode _n, so that the predictive frame of a coding to be provided This image coding-decoding device comprises a movable information group coding device, and this movable information group coding device comprises:

A motion analyzer is used for receiving

-present frame I _n,

-reference frame I _N-1, and

-motion vector information group is to the present frame I of multi-region section _nEach section, this motion vector information group have one group of motion vector [x+ Δ x (and x, y), y+ Δ y (x, y)], each motion vector is with the reference frame of a section Be mapped to a section present frame I corresponding to this vector _n,

This motion analyzer comprises:

A kind of device is used for approaching a prediction error functions between each present frame section and reference frame section, the reference frame section then passes through with one group of N+M kinematic coefficient C _iWith one group of basic function f _iApproach corresponding motion vector [x+ Δ x (and x, y), y+ Δ y (x, y)] and be mapped to the present frame section, this approximated function comprises a kinematic coefficient C _iLinear combination, and according to E _iC _i-y _iCan be expressed as a matrix expression and

The matrix processing unit is used for being each present frame section, by described matrix expression, forms one first matrix R _i ¹With the first vector Z _i ¹, wherein $R_i^1{C}_i=Z_i^1,$ And the first matrix R _i ¹Line number equal the number of kinematic coefficient N+M;

Section merges device

Be connected on the motion analyzer, to receive the described first matrix R _i ¹With the first vector Z _i ¹, and be used for

To present frame I _nAdjacent sections to (S _i, S _j), by each each right matrix R _i ¹, R _j ¹With the first vector Z _i ¹, Z _j ¹, determine that merges a matrix R _Ij ¹With combined vector Z _Ij ¹, and

With merging matrix R _Ij ¹With combined vector Z _Ij ¹Replace the first paired matrix R _i ¹With the first vector Z _i ¹, wherein, merges matrix and vector and in predetermined threshold, provide predicated error, and it is right to merge corresponding section; The coefficient destructor is connected to section and merges the first remaining matrix R of reception on the device _i ¹With the first vector Z _i ¹And the merging matrix R that replaces _Ij ¹With combined vector Z _Ij ¹, and be used for

Each section or merging section to present frame utilize corresponding matrix R _k ¹With vector Z _k ¹, determine one group of kinematic coefficient C that simplifies _kThereby, the variation that makes the prediction error functions of approaching in a preset range, and

A present frame I is provided _nEncoding scheme, merge section for each section or each, it comprises the kinematic coefficient that described number reduces and shows the mark which coefficient has been omitted.

2. image coding-decoding device as claimed in claim 1, wherein, for each section in the present frame, to its prediction error functions approximation, this function is by motion analyzer: $\underset{j=i}{\overset{P}{\mathrm{\&Sigma;}}}{(e_{j,1}{\tilde{c}}_1+e_{j,2}{\tilde{c}}_2+\&CenterDot;\&CenterDot;\&CenterDot;+e_{j,N+M}{\tilde{c}}_{N+M}-y_i)}^2$ $E_i=\left[\begin{array}{cccc}{e}_{\mathrm{1,1}}& e_{\mathrm{1,2}}& ...& e_{1,N+M}\\ e_{\mathrm{2,1}}& e_{\mathrm{2,2}}& ...& e_{2,N+M}\\ \underset{\&CenterDot;}{\overset{\&CenterDot;}{\&CenterDot;}}& \underset{\&CenterDot;}{\overset{\&CenterDot;}{\&CenterDot;}}& \underset{\&CenterDot;}{\overset{\&CenterDot;}{\&CenterDot;}}& \underset{\&CenterDot;}{\overset{\&CenterDot;}{\&CenterDot;}}\\ e_{P,1}& e_{P,2}& ...& e_{P,N+M}\end{array}\right]$ With $y_i=\left[\begin{array}{c}{y}_1\\ y_2\\ \underset{.}{\overset{.}{.}}\\ y_P\end{array}\right]$

3. image coding-decoding device as claimed in claim 1, wherein, described matrix processing unit comprises QR factorization device, is used for matrix E _iBe decomposed into a matrix Q _iAnd R _iProduct, and be used to form an auxiliary vector Z _i, Z wherein _i=Q ^T _iY _i, and, by matrix R _iThe capable formation of preceding N+M R _i ¹, by vector Z _iPreceding N+M element form vector Z _i ¹

4. image coding-decoding device as claimed in claim 1, wherein, section merges device and comprises the trigonometric ratio device, to described adjacent sections to (S _i, S _j), it is with formula $\left[\begin{array}{c}{R_i}^1\\ {R_j}^1\end{array}\right]c_{\mathrm{ij}}=\left[\begin{array}{c}{z_i}^1\\ {z_j}^1\end{array}\right]$ Triangle turns to following form

C wherein _IjBe to merge one group of right N+M kinematic coefficient of adjacent sections.

5. image coding-decoding device as claimed in claim 4, wherein section merges device according to formula ${\mathrm{\&Delta;E}}_{\mathrm{ij}}=\underset{k=1}{\overset{N+M}{\mathrm{\&Sigma;}}}{q_k}^2.$

By being every couple of adjacent sections (S _i, S _j) calculating square prediction error Δ E _IjVariation, determine that predicated error is whether in preset range.

6. as claim 4 or 5 described image coding-decoding devices, wherein, section merges device and utilizes the preceding N+M of trigonometric ratio Matrix Formula capable, determines described merging matrix R _Ij ¹With combined vector Z _Ij ¹, like this

7. image coding-decoding device as claimed in claim 1 wherein, includes the mark that merges section in the encoding scheme of described present frame In.

8. image coding-decoding device as claimed in claim 1, wherein, described coefficient destructor is by selectively from matrix R _k ¹In cancellation i row and from vector Z _k ¹In i element of cancellation, thereby the Matrix Formula of a simplification is provided $R_K^1{C}_K=Z_K^1,$ From then on, determine the kinematic coefficient C of number minimizing one by one for each section _k

9. image coding-decoding device as claimed in claim 8, wherein, the coefficient destructor turns to following form with the Matrix Formula triangle of simplifying:

And by formula ${\mathrm{\&Delta;E}}_K=q_i^2$ Calculate the variation of square prediction error.

10. image coding-decoding device as claimed in claim 9, if the variation of square prediction error is in preset range, then the coefficient destructor is with kinematic coefficient C _iFrom coefficient matrix C _kMiddle cancellation, and, by cancellation last column from second described trigonometric ratio Matrix Formula, represent this formula again.

11. image coding-decoding device as claimed in claim 1, wherein, the coefficient destructor is by finding the solution Matrix Formula $R_K^1{C}_K=Z_K^1$ Or the reduced form of this formula, determine kinematic coefficient C _k

Images