CN1234248C - Code rate distribution method for fine grain expansible video coding - Google Patents

Abstract

The present invention belongs to the technical fields of a code and the communication of a figure and an image of a computer, particularly to a code rate allocation method of an expandable video code with fine granularity. The present invention comprises an equal slope code rate allocation method and an equal distortion code rate allocation method. The present invention comprises the specific steps that firstly, a rate distortion function curve of each image is generated; a slope absolute value at an optimal breakpoint site or an equal distortion code rate allocation distortion value of the rate distortion curve is obtained by an iterative computing formula; a bit number corresponding to the slope value at the optimal breakpoint site or a bit number corresponding to the equal distortion value according to a linear interpolation method through a bit number of two adjacent sample points. Optimal video transmission quality can be obtained by the equal slope code rate allocation method of the present invention. Compared with a traditional average code rate allocation method, the present invention is greatly enhanced. The fluctuation of the quality of the transmission image can be reduced at maximum by the equal distortion code rate allocation method. The present invention can be applied to the fields of network flow medium transmission, wireless video transmission, etc.

Description

The code rate allocation method of fine granularity scalable video coding

Technical field the invention belongs to computer graphic image coding and communication technical field, the particularly code rate allocation method of video coding.

The development of the background technology computer network and the third generation (3G) cordless communication network makes video transmission become a kind of important form of MMS (Multimedia Message Service) day by day.Method for video coding also must take into full account the influence of network characteristic to video transmission when pursuing low code check.The challenge that video transmission is at present faced is mainly reflected in following two aspects: the firstth, and the network bandwidth is non-constant, and changes in time, and the secondth, there are problems such as data-bag lost, error code and variable delay in the video transmission.Therefore, all adopted relevant technology in the video encoding standard of being formulated in the world at present, to improve the overall performance that video transmits in complex network environment.

It is a kind of method for video coding that can adapt to network bandwidth dynamic change that fine granularity can be expanded (FGS) video coding, its basic thought is to be basic layer (Base Layer) code stream and an enhancement layer that can block at an arbitrary position (Enhancement Layer) code stream that can decode separately with encoding video pictures, wherein basic layer bit stream uses block motion compensation, the coded system that discrete cosine transform (DCT) conversion and variable-length encoding (Huffman) combine, can satisfy the minimum requirements of Network Transmission bandwidth, can provide basic, lower picture quality, enhancement layer bitstream then uses the Bit-Plane Encoding technology that the DCT residual values is encoded, the network bandwidth that can adapt in the certain limit changes, and provides the low-quality image of decoding from basic layer bit stream to strengthen to a kind of fine-grained quality the high quality graphic.

The fine granularity scalable coding has become the part of fluidisation class (Streaming Profile) in the MPEG-4 standard (ISO/IEC IS-14496) at present.In the process of using the transmission of FGS video code flow, video server also blocks transmission according to the variation of the current time network bandwidth to enhancement layer bitstream in the basic layer bit stream of transmission, uses the Network Transmission bandwidth resources to reach maximal efficiency ground.Although the variation right and wrong of the network bandwidth stably, but the buffer models of video server end can be smoothly and this bandwidth shake of filtering, and make it in the regular hour interval, to tend to be steady, therefore often the frame of video of plurality of continuous (is generally one or several image sets (GOP) as one Basic Transmission Unit in the practical application, relevant with network condition and buffer size) consider, and hypothesis available network bandwidth when these picture frames of transmission is constant.The Data Rate Distribution of FGS is meant and will remains the enhancement layer bitstream that available network bandwidth is distributed to each two field picture in the transmission unit except that basic layer bit rate.The bit number difference that each frame distributed means that the position of enhancement layer point of cut-off is also different, finally can cause the difference of image transmission quality.The purpose of Data Rate Distribution is to satisfy under the network wide constraints condition, select suitable enhancement layer bitstream point of cut-off improving the oeverall quality of Video Transmission Unit, so Data Rate Distribution is a major issue in the MPEG-4 FGS video transmission.

The code rate allocation method of fine granularity scalable coding is to select suitable enhancement layer point of cut-off according to the rate distortion curve of each two field picture, makes the bit number sum of each frame satisfy the requirement of target bit rate.Traditional MPEG-4FGS code rate allocation method is even Data Rate Distribution: accompanying drawing 1 has provided a simple description based on the code rate allocation method of rate distortion function, c among the figure ₁, c ₂, c ₃The rate distortion curve of representing continuous 3 two field pictures.The point of cut-off that is parallel on the straight line SR representation rate distortion curve of longitudinal axis D has identical bit number R _r, (R _r=R/n, n represents frame number, here n=3), corresponding truncated error is D _i

Evenly code rate allocation method is an enhancement layer bitstream of the current network bandwidth mean allocation being given each frame in the video unit.The advantage of this method is simple, particularly the user when video server asks more for a long time, the average bit rate allocation algorithm can effectively improve the load ability to bear of server, but shortcoming is when video content generation significant change, the average bit rate distribution method will cause the fluctuation by a relatively large margin of picture quality, because people's vision system is relatively more responsive to the variation of picture quality, therefore this code rate allocation method can obviously reduce the subjective quality of image transmission.Microsoft Research, Asia has proposed a kind of local optimum code rate allocation method based on two frames associatings rate distortion function in the enhancement layer bitstream (by Q.Wang, Z.X.Xiong, F.Wu, S.P.Li is published in the article (332-335 page or leaf) of being entitled as in the IEEEInternational Conference on Information Technology:Coding and Computing meeting " Optimal Rate Allocation for Progressive Fine Granularity Scalable Video Coding " April calendar year 2001), this local optimum scheme is exactly to minimize one group of image (adjacent two frames) distortion sum under the certain constraints of total bitrate satisfying, and can represent with (1) formula:

Minimize (D ₁(R ₁)+D ₂(R ₂)), R satisfies condition ₁+ R ₂=S (1)

R in the formula ₁And R ₂The bit number of odd-numbered frame and even frame is distributed in expression respectively, and S is every group of available bit number.

Associating rate distortion function hypothesis information source on this statistical significance satisfies Laplce (Laplacian) or Gauss (Gaussian) distribution and derives, and sees shown in the formula (2), finds the solution the optimal value of Data Rate Distribution thus.

$f\left(R_1\right)=D_1\left(R_1\right)+D_2(S-R_1)={{\mathrm{\σ}}_1}^2{e}^{-\mathrm{\α}{R}_1}+{{\mathrm{\σ}}_2}^2{e}^{-\mathrm{\α}(S-R_1)}----\left(2\right)$

Here f (R ₁) be the distortion factor D of adjacent two frame decoding images ₁(R ₁), D ₂(R ₂) sum, σ ₁The square root of having represented the mean square error of odd-numbered frame basic layer reconstructed image and original image, σ ₂The square root of having represented the mean square error of even frame basic layer reconstructed image and original image, R ₁Be the used bit number of enhancement layer of decoding odd-numbered frame, α is a constant.Just can obtain the Data Rate Distribution of local optimum by the minimum point of solution formula (2).This algorithm can improve the Y-PSNR (PSNR) of video transmission on the whole, but the fluctuation to image transmission quality and code check does not solve fully under the particular network situation.

Summary of the invention the present invention is the weak point that overcomes the code rate allocation method of prior art, propose two kinds of fine granularities and can expand the method for Data Rate Distribution in (FGS) video transmission, these two kinds of methods all obtain based on rate distortion function, this method can be under the constraints of the network bandwidth, rate distortion curve according to the FGS coding can reduce the fluctuation of image transmission quality to greatest extent; Can obtain best video transmission quality, and more traditional average bit rate distribution method there is the raising of 0.2-0.5dB.

The code rate allocation method of a kind of fine granularity scalable video coding that the present invention proposes is characterized in that, this method is the Data Rate Distribution such as slope such as grade based on rate distortion function, specifically may further comprise the steps:

1) at first by to any one encoded point sampling in each bit plane of every two field picture in the enhancement layer bitstream and generate the rate distortion functional curve of every two field picture in the code check allocation units according to the sample point interpolation;

2) by iterative computation formula λ _K+1=λ _k-f (λ _k)/f ' (λ _k), obtain the slope absolute value λ of optimum point of cut-off place of the rate distortion curve of every two field picture in the Data Rate Distribution unit; Wherein, k represents iterations, λ _kRepresent the slope absolute value that the k step calculates, f _kBe by formula $f\left(\mathrm{\λ}\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{R}_i\left(\mathrm{\λ}\right)-R_c$ (N represents number of image frames in the code check allocation units, R _cBe the given total bit number of allocation units) definition functional value f (λ _k), its expression bit number distributes error, f _k' be by formula $f^{\′}\left(\mathrm{\λ}\right)=\underset{i}{\mathrm{\Σ}}\frac{1}{d^2{D}_i/{{\mathrm{dR}}^2}_i}$ Functional value the f ' (λ of definition _k), D _iRepresent that the i frame is at R _iThe distortion value that bit number is corresponding down;

3) bit number by two adjacent sample points of optimum point of cut-off obtains the bit number of the optimum point of cut-off of every two field picture place correspondence according to approach based on linear interpolation at last, with this bit number as distributing to this two field picture bit number.

The code rate allocation method of the another kind of fine granularity scalable video coding that the present invention proposes is characterized in that, this method is the Data Rate Distribution that waits distortion based on rate distortion function, specifically may further comprise the steps:

1) at first by to any one encoded point sampling in each bit plane of every two field picture in the enhancement layer bitstream and generate the rate distortion functional curve of every two field picture in the code check allocation units according to the sample point interpolation;

2) by the iterative computation formula $D_{k+1}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD}),$ Obtain the distortion value that waits the distortion Data Rate Distribution of the rate distortion curve of every two field picture in the Data Rate Distribution unit, wherein, k represents iterations, D _kRepresent the distortion value that the k step calculates, F _kExpression is by formula $F\left(D\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{R}_i\left(D\right)-R_c$ Functional value F (the D of definition _k), R _cBe the given total bit number of allocation units, N represents the frame number of allocation units image, R _iRepresent the bit number that i frame enhancement layer bitstream distributes, D _iRepresent that this frame is at R _iFollowing corresponding distortion, i.e. R _i(D) be the rate distortion function of i two field picture;

3) obtain the pairing bit number of distortion value of distortion Data Rate Distribution such as every two field picture again according to approach based on linear interpolation by the bit number that waits two adjacent sample points of distortion corresponding points, with this bit number as the bit number of distributing to this two field picture.

Effect of the present invention

The slope code rate allocation method that waits of the present invention: equal the total bit number of target (∑ R satisfying each frame bit number sum _i=R, wherein R _iBe that every frame distributes the bit number that obtains, R represents the available bandwidth of current network) under the condition, be family of straight lines SS and the rate distortion curve c of λ with slope among Fig. 1 ₁, c ₂, c ₃The tangent point of contact that obtains is optimum point of cut-off (R ₀, D ₀), can make the total distortion minimum of all frames like this.The benefit of this method is to utilize the network bandwidth to greatest extent, can make the image fault minimum under qualification code check condition.

The distortion code rate allocation method that waits of the present invention: equal the total bit number of target (∑ R satisfying each frame bit number sum _i=R, wherein R _iBe that every frame distributes the bit number that obtains, R represents the available bandwidth of current network) under the condition, with straight line SD that is parallel to trunnion axis R among Fig. 1 and rate distortion curve c ₁, c ₂, c ₃The point that intersects has identical truncated error D _dSatisfying under the bandwidth constraints like this, each frame distortion all be constant, the fluctuation of picture quality can be reduced to minimum.

Adopt the cycle tests of part in the MPEG-4 normal video tissue to compare as follows to code rate allocation method existing and of the present invention: FGS coding frame per second is 10fps, and frame format is I PPP ...For CIF (resolution is 352 * 288) or SIF (resolution is 352 * 240) picture format, the code check of basic layer is 128kbps, and QCIF (resolution is 176 * 144) is 32kbps, and bit rate control method is TM5; The code check of enhancement layer is 768kbps (CIF/SIF) and 160kbps (QCIF), and table 1 has been enumerated experiment sequence and the test condition that is used to test.

Table 1 cycle tests

The cycle tests title	Picture format (original frame per second)	Totalframes	Basic layer bit rate (kbps)	Coding frame per second (fps)
					1.Foreman	CIF(30fps)	300	128	10
2.Coastgurad	CIF(30fps)	300	128	10
					3.Stefan	SIF(30fps)	300	128	10
4.Table-tennis	SIF(30fps)	300	128	10
					5.News	CIF(30fps)	300	128	10
6.Carphone	QCIF(30fps)	300	32	10

Table 2 be the present invention propose wait the slope code check, etc. method (one group of 3 frame) 6 mean flow rate Y-PSNR (PSNR) values that cycle tests experimentizes and obtains in his-and-hers watches 1 of distortion code rate allocation method and average bit rate distribution method, Microsoft Research, Asia's proposition.

The average peak signal to noise ratio (PSNR) of four kinds of code rate allocation methods of table 2

The cycle tests title	Code rate allocation method
					Average bit rate distribution method (dB)	The method (one group of 3 frame) that Microsoft Research, Asia proposes (dB)	Deng slope code rate allocation method (dB)	Deng distortion code rate allocation method (dB)
	1.Foreman CIF	36.8756	36.8823	37.0216					36.5786
2.Coastgurad CIF					33.1285	33.1451	33.1941	33.1933
	3.Stefan SIF	30.2724	30.2684	30.5767					30.1701
4.Tabletennis SIF					35.7186	35.7161	35.7282	35.6162
	5.News CIF	40.9218	40.9369	40.9842					40.9865

6.Carphone CIF

37.1133

37.1117

37.6478

36.6606

Average peak signal to noise ratio Deng the slope code rate allocation method distributes high 0.2-0.5dB than average bit rate.

Table 3 has provided the mean value of difference of the Y-PSNR (PSNR) of the adjacent two frame brightness that four kinds of code rate allocation methods obtain.

Average quality shake between adjacent two two field pictures of table 3

The cycle tests title	Code rate allocation method (dB)
									The average bit rate distribution method		The method (one group of 3 frame) that Microsoft Research, Asia proposes		Deng the slope code rate allocation method		Deng the distortion code rate allocation method
	Average jitter	Maximum difference	Average jitter	Maximum difference	Average jitter	Maximum difference	Average jitter	Maximum difference
									1.Foreman CIF	0.41	2.21	0.25	2.16	0.16	0.84	0.05	0.39
2.Coastgurad CI	0.22	4.90	0.16	1.59	0.09	1.65	0.02	0.35
									3.Stefan SIF	0.19	0.80	0.13	1.13	0.10	0.72	0.03	0.17
4.Tabletennis S	0.17	1.62	0.14	1.62	0.11	0.64	0.04	0.42
									5.News CIF	0.09	0.74	0.06	0.67	0.03	0.17	0.02	0.09
6.Carphone QCIF	0.25	1.71	0.19	2.47	0.14	0.47	0.05	0.17

From table 3, biggest quality shake is greater than 1.5dB between adjacent two two field pictures that the average bit rate distribution method produces, and the distortion code rate allocation method that waits that the present invention proposes has only 0.3dB, from the whole video fragment, the picture quality of average bit rate distribution method is dithered as 5-10dB, method that Microsoft Research, Asia proposes and average bit rate distribution are more or less the same, and wait the distortion code rate allocation method shake can be limited within the 1dB.

Description of drawings

Fig. 1 is that three kinds of MPEG-4 fine granularities can be expanded the schematic diagram that (FGS) encoder bit rate distributes.

The code rate allocation method based on the fine granularity scalable video coding of rate distortion function that embodiment the present invention proposes comprises and waits slope code rate allocation method and code rate allocation method such as distortion such as grade, is described in detail as follows respectively in conjunction with the embodiments:

What the present invention proposed waits slope code rate allocation method embodiment concrete steps as follows:

1) before Data Rate Distribution, generate the rate distortion functional curve of every two field picture, concrete grammar is: each bit plane starting point place sampling of every two field picture in enhancement layer bitstream obtains the sample point (R of this image in the R-D space _j, D _j), and these sample points are coupled together R-D function curve as this two field picture with the segmentation straight line, distortion value on non-sampled point is that the distortion value linear interpolation by adjacent nearest two sample points with it obtains, and is to obtain by extrapolation for the point outside first sample point and last sample point; And the slope dD/dR on each sample point is the left and right sides slope mean value that this point and its adjacent sample point constitute, the slope of non-sample point is that the slope value linear interpolation by adjacent nearest two sample points with it obtains, for the sample point slope value replacement nearest with it of the slope value of non-sample point outside first sample point and last sample point;

2) after obtaining the rate distortion curve, by iterative computation formula λ _K+1=λ _k-f (λ _k)/f ' (λ _k), obtaining the slope absolute value λ of optimum point of cut-off place, this iterative computation specifically may further comprise the steps:

(1) step:

Each variate-value of initialization: k=0, λ _k=λ 0, f _k=f (λ ₀), f _k'=f ' (λ ₀);

K represents iterations, λ _kRepresent the slope absolute value that the k step calculates, f _kBy formula $f\left(\mathrm{\λ}\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{R}_i\left(\mathrm{\λ}\right)-R_c$ (N represents code check allocation units number of image frames, R _cBe the given total bit number of allocation units) definition functional value f (λ _k), f _k' by formula $f^{\′}\left(\mathrm{\λ}\right)=\underset{i}{\mathrm{\Σ}}\frac{1}{d^2{D}_i/d{R^2}_i}$ The functional value f ' (λ of definition _k);

For CIF/SIF image in different resolution R _cGet 7680 kilobits (total bit number of N two field picture, the i.e. product of code check and corresponding time of N two field picture), for QCIF image in different resolution R _cGet 1600 kilobits, λ ₀For bit number is R _c/ N o'clock on the rate distortion curve of first two field picture the absolute value of corresponding slope (result of this iterative computation is not subjected to λ ₀The influence of value), set greatest iteration step number MaxNum=30, bit number distributes error threshold MinDiff=8N, N=100;

(2) step:

Calculate the temporary variable λ of (k+1) step λ value _Temp, λ _Temp=λ _k-f (λ _k)/f ' (λ _k);

Calculate the temporary variable f of (k+1) step λ value _Temp, $f_{\mathrm{temp}}=\underset{i}{\mathrm{\Σ}}{R}_i\left({\mathrm{\λ}}_{\mathrm{temp}}\right)-R_c;$

(3) step:

If Sign is (f _k) value and Sign (f _Temp) the value difference, so just according to λ _K+1=(λ _k+ λ _Temp)/2 obtain (k+1) the slope absolute value λ in step _K+1, according to f _K+1=f (λ _K+1) obtain the bit number in (k+1) step and distribute error amount f _K+1Otherwise just according to λ _K+1=λ _TempObtain (k+1) the slope absolute value λ in step _K+1, according to f _K+1=f _TempThe bit number that obtains (k+1) step distributes error amount f _K+1

Sign (f wherein _k) the is-symbol function, be defined as $\mathrm{Sign}\left(v\right)=\left\{\begin{array}{cc}1& ,\mathrm{ifv}\&GreaterEqual;0\\ -1& ,\mathrm{ifv}<0\end{array}\right.;$

The iteration step number is added 1, i.e. k=k+1;

(4) step:

If iteration step number k surpasses greatest iteration step number MaxNum or current bit number distributes the absolute value of error to distribute error threshold MinDiff less than bit number, so just stop computational process, (3) slope absolute value that obtains that goes on foot as optimum point of cut-off place slope absolute value, is gone on foot the continuation iterative computation otherwise just jump to (2);

3) bit number of back by its two adjacent sample points obtains by 2 according to approach based on linear interpolation) in calculate the pairing bit value of slope absolute value of optimum point of cut-off.Concrete grammar is: establish Being the 2nd) iterative computation obtains slope absolute value, can find adjacent nearest such two sample point (R with point of cut-off ₁, D ₁) and (R ₂, D ₂), their slope absolute value is λ respectively ₁, λ ₂, make ${\mathrm{\&lambda;}}_1\&GreaterEqual;\widehat{\mathrm{\&lambda;}}\&GreaterEqual;{\mathrm{\&lambda;}}_2,$ The allocation bit number of this two field picture is so $R=(\widehat{\mathrm{\&lambda;}}-{\mathrm{\&lambda;}}_2)R_1/({\mathrm{\&lambda;}}_1-{\mathrm{\&lambda;}}_2)+({\mathrm{\&lambda;}}_1-\widehat{\mathrm{\&lambda;}})R_2/({\mathrm{\&lambda;}}_1-{\mathrm{\&lambda;}}_2);$ If Slope absolute value than all sample points is all big, and R just gets the pairing bit value of sample point of slope absolute value maximum so, if Slope absolute value than all sample points is all little, and R just gets the pairing bit value of sample point of slope absolute value minimum so.

The above-mentioned the 2nd) the iterative computation derivation of equation process in the step is as follows:

The target that waits the slope code rate allocation method of fine granularity scalable coding is exactly to make interior each two field picture of allocation units identical at the rate-distortion slope of point of cut-off at one under the constant bit rate, promptly satisfies (3) formula, establishes R _iRepresent the bit number that i frame enhancement layer bitstream distributes, D _iRepresent that this frame is at R _iThe distortion value that bit number is corresponding down, λ is the absolute value of point of cut-off slope, N represents code check allocation units number of image frames.

dD _i/dR _i＝-λ，i＝0，…，N (3)

As long as known the λ value that waits slope, just can obtain the bit number R of every two field picture _iThe equation that will find the solution Deng the slope code rate allocation method is:

f(λ)＝0 (4)

Wherein $f\left(\mathrm{\&lambda;}\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(\mathrm{\&lambda;}\right)-R_c----\left(5\right)$

R _i(λ) be the monotonic function of λ, R _cIt is given total bit number.Can use newton (Newton) iterative search method to carry out numerical solution by following iterative computation.If λ _kBe the k step to calculate the λ value, f (λ) is at λ so _kNear functional value can be represented with first approximation:

$f\left(\mathrm{\&lambda;}\right)\&ap;f\left({\mathrm{\&lambda;}}_k\right)+(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left({\mathrm{\&lambda;}}_k\right)/\mathrm{d\&lambda;})(\mathrm{\&lambda;}-{\mathrm{\&lambda;}}_k)----\left(6\right)$

Formula (6) brought into formula (5) and find the solution constraint equation (4) just can obtain

λ _k+1＝λ _k-f(λ _k)/f′(λ _k) (7)

$f^{\&prime;}\left(\mathrm{\&lambda;}\right)=\underset{i}{\mathrm{\&Sigma;}}\frac{{\mathrm{dR}}_i}{\mathrm{d\&lambda;}}=\underset{i}{\mathrm{\&Sigma;}}\frac{{\mathrm{dR}}_i}{d({\mathrm{dD}}_i/{\mathrm{dR}}_i)}=\underset{i}{\mathrm{\&Sigma;}}\frac{1}{d({\mathrm{dD}}_i/{\mathrm{dR}}_i)/{\mathrm{dR}}_i}=\underset{i}{\mathrm{\&Sigma;}}\frac{1}{d^2{D}_i/{{\mathrm{dR}}^2}_i}$

Carry out above iterative computation repeatedly, distribute the absolute value of error f (λ) less than certain setting threshold, optimal solution so up to bit number Also just tried to achieve, each frame distributes the bit number that obtains to represent with following formula:

$R_i=R_i\left(\widehat{\mathrm{\&lambda;}}\right),i=0,...,N.$

Code rate allocation method such as slope such as grade of the present invention is the code rate allocation method of distortion minimum, is described as follows:

If R _iRepresent the bit number that i frame enhancement layer bitstream distributes, D _iRepresent that this frame is at R _iFollowing corresponding distortion, the Data Rate Distribution of distortion minimum is that problem is found the solution in an optimization, just can be expressed as:

Satisfy constraint $\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{R}_i=R_c$

Wherein N is the frame number that Data Rate Distribution unit (GOP) is comprised, R _cIt is total bit number of available code check.Utilize Lagrange (Lagrange) multiplier method, constrained optimization problem can be converted into unconfined optimization problem.Introduce Lagrange multiplier λ, the target function of optimization can be expressed as:

$J=\underset{i}{\mathrm{\&Sigma;}}{D}_i\left(R_i\right)+\mathrm{\&lambda;}(\underset{i}{\mathrm{\&Sigma;}}{R}_i-R_C)$

It is exactly the extreme point of finding the solution above-mentioned function that desired distortion minimum code rate is distributed, order $\frac{\&PartialD;J}{\&PartialD;R_i}=0,$ Can obtain:

dD _i/dR _i，＝-λ

DD _i/ dR _iBe at bit number point of cut-off R at the i two field picture _iThe slope at place, above equation represent that the Data Rate Distribution of distortion minimum requires the slope (dD/dR) of all images point of cut-off in the Data Rate Distribution unit all to equate, and this just and etc. the requirement of slope code rate allocation method consistent.

The code rate allocation method embodiment concrete steps of distortion that wait of the present invention are as follows:

1) before the Data Rate Distribution, generate the rate distortion functional curve of every two field picture, concrete grammar is: each the bit plane starting point place sampling to every two field picture in the enhancement layer bitstream obtains the sample point (R of this image in the R-D space _j, D _j), and these sample points are coupled together R-D function curve as this two field picture with the segmentation straight line, distortion value on non-sampled point is that the distortion value linear interpolation by adjacent nearest two sample points with it obtains, and is to obtain by extrapolation for the point outside first sample point and last sample point.And the dR/dD on each sample point is the mean value of the left and right sides dR/dD value of this point and its adjacent sample point formation, but not being dR/dD value linear interpolation by adjacent nearest two sample points with it, the dR/dD value at sample point place obtains, for the sample point dR/dD value replacement nearest of the dR/dD value of non-sample point outside first sample point and last sample point with it:

2) after obtaining the rate distortion curve, by the iterative computation formula $D_{k+1}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD}),$ Obtain waiting the distortion value of distortion Data Rate Distribution, this iterative computation specifically may further comprise the steps:

(1) step:

Initializing variable value: k=0, D _k=D ₀, F _k=F (D ₀);

K represents iterations, D _kRepresent the distortion value that the k step calculates, F _kExpression is by formula $F\left(D\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(D\right)-R_c$ (N represents the frame number of allocation units image, R _cBe the given total bit number of allocation units) definition functional value F (D _k), for CIF/SIF image in different resolution R _cGet 7680 kilobits (total bit number of N two field picture, the i.e. product of code check and corresponding time of N two field picture), for QCIF image in different resolution R _cGet 1600 kilobits, D ₀Be R _c/ N o'clock on the first two field picture rate distortion curve corresponding distortion value (final result of this iterative computation is not subjected to D ₀The influence of value):

Set greatest iteration step number MaxNum=15, bit number distributes error threshold MinDiff=8xN, N=100;

(2) step:

Calculate (k+1) step distortion value D _K+1Temporary variable D _Temp, $D_{\mathrm{temp}}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD});$

Calculate (k+1) step distortion value D _K+1Temporary variable F _Temp, $F_{\mathrm{temp}}=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(D_{\mathrm{temp}}\right)-R_c;$

(3) step:

If Sign is (F _k) and Sign (F _Temp) difference, so just according to D _K+1=(D _k+ D _Temp)/2 obtain (k+1) the distortion value D in step _K+1, according to F _K+1=F (D _K+1) obtain (k+1) step bit number and distribute error F _K+1Otherwise just according to D _K+1=D _TempObtain (k+1) the distortion value D in step _K+1, according to F _K+1=F _TempObtain (k+1) step bit number and distribute error F _K+1

Sign (f wherein _k) the is-symbol function, be defined as $\mathrm{Sign}\left(v\right)=\left\{\begin{array}{cc}1& ,\mathrm{ifv}\&GreaterEqual;0\\ -1& ,\mathrm{ifv}<0\end{array}\right.$

The iteration step number is added 1, i.e. k=k+1;

(4) step:

If the iteration step number surpasses MaxNum or current bit number Error Absolute Value less than MinDiff, then stop computational process, the distortion value that obtains in (3) step as last result of calculation, otherwise is just jumped to (2) step continuation iterative computation.

3) find the solution obtain distortion value after, the bit number by its two adjacent sample points obtains the pairing bit value of this distortion value according to approach based on linear interpolation again.Concrete grammar is: establish

Being the 2nd) iterative computation obtains distortion value, can find adjacent nearest such two sample point (R with point of cut-off ₁, D ₁) and (R ₂, D ₂), their distortion value is D respectively ₁, D ₂, make $D_1\&GreaterEqual;\hat{D}\&GreaterEqual;D_2,$ The allocation bit number of this two field picture is so $R=(\hat{D}-D_2)R_1/(D_1-D_2)+(D_1-\hat{D})R_2/(D_1-D_2);$ If

Distortion value than all sample points is all big, and R just gets the pairing bit value of sample point of distortion value maximum so, if Distortion value than all sample points is all little, and R just gets the pairing bit value of sample point of distortion value minimum so.

The above-mentioned the 2nd) the iterative computation derivation of equation process in the step is as follows:

If the code rate allocation method Deng distortion is exactly the distortion unanimity that makes every two field picture, its target function is

$D_i\left(R_i\right)=\hat{D},$ Satisfy constraints $\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i=R_c$

Wherein

Be the distortion value that waits the distortion Data Rate Distribution, N is the frame number of Data Rate Distribution unit, R _cIt is given total bit number.Distortion D can obtain by solving equation (8):

F(D)＝0 (8)

Wherein, $F\left(D\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(D\right)-R_c----\left(9\right)$

Because D _i(R _i) be monotonic function, can find the solution by newton (Newton) iterative algorithm

If D _kBe the distortion value that calculates in k step, so with F (D) at D _kDoing single order launches approximate:

$F\left(D\right)\&ap;F\left(D_k\right)+(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD})(D-D_k)----\left(10\right)$

(10) formula is brought in the equation (8), can obtain (k+1) distortion value in step:

$D_{k+1}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD})----\left(11\right)$

The calculating that so iterates, after the absolute value of F (D) was less than certain threshold value, iteration finished, and just can obtain optimum distortion value $\hat{D}=D_{k+1}.$

Claims (6)

1, a kind of code rate allocation method of fine granularity scalable video coding is characterized in that, this method is the Data Rate Distribution such as slope such as grade based on rate distortion function, specifically may further comprise the steps:

1) at first by to any one encoded point sampling in each bit plane of every two field picture in the enhancement layer bitstream and generate the rate distortion functional curve of every two field picture in the code check allocation units according to the sample point interpolation;

2) by iterative computation formula λ _K+1=λ _k-f (λ _k)/f ' (λ _k), obtain the slope absolute value λ of optimum point of cut-off place of the rate distortion curve of every two field picture in the Data Rate Distribution unit, wherein, k represents iterations, λ _kRepresent the slope absolute value that the k step calculates, f _kThe bit number of representing the k step distributes error, and it is by formula $f\left(\mathrm{\&lambda;}\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(\mathrm{\&lambda;}\right)-R_c$ Functional value f (the λ of definition _k), N represents number of image frames in the code check allocation units, R _cBe the given total bit number of allocation units, f _k' be by formula $f^{\&prime;}\left(\mathrm{\&lambda;}\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\frac{1}{d^2{D}_i/d{R^2}_i}$ Functional value the f ' (λ of definition _k), D _iRepresent that the i frame is at R _iThe distortion value that bit number is corresponding down;

3) bit number by two adjacent sample points of optimum point of cut-off obtains the bit number of the optimum point of cut-off of every two field picture place correspondence according to approach based on linear interpolation at last, with this bit number as distributing to this two field picture bit number.

2, the code rate allocation method of fine granularity scalable video coding as claimed in claim 1, it is characterized in that, the concrete grammar of the rate distortion functional curve of the every two field picture of said generation is: the starting point place of each Bit-Plane Encoding of every two field picture sampling in enhancement layer bitstream obtains a plurality of sample point (Rs of this image in code check and distortion space R-D _j, D _j), and these sample points are connected in turn as the R-D function curve of this two field picture with the straightway of segmentation, distortion value on non-sampled point is that the distortion value linear interpolation by adjacent nearest two sample points with it obtains, and is to obtain by linear extrapolation for the point outside first sample point and last sample point; Slope dD/dR on each sample point is the mean value of the left and right sides slope of this point and its adjacent sample point formation, but not the slope of sample point is to use the slope value linear interpolation of adjacent two sample points recently with it to obtain, and uses sample point slope value nearest with it to replace for the slope value of non-sample point outside first sample point and last sample point.

3, the code rate allocation method of fine granularity scalable video coding as claimed in claim 1 is characterized in that, this iterative computation specifically may further comprise the steps:

(1) each variate-value of initialization: k=0, λ _k=λ 0, f _k=f (λ ₀), f _k'=f ' (λ ₀);

λ ₀It is the initial slope absolute value;

Set greatest iteration step number MaxNum, bit number distributes error threshold MinDiff;

(2) calculate (k+1) step λ _K+1The temporary variable λ of value _Temp: λ _Temp=λ _k-f (λ _k)/f ' (λ _k);

Calculate (k+1) step λ _K+1The temporary variable f of value _Temp: $f_{\mathrm{temp}}=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left({\mathrm{\&lambda;}}_{\mathrm{temp}}\right)-R_c;$

(3) if Sign is (f _k) value and Sign (f _Temp) the value difference, then according to λ _K+1=(λ _k+ λ _Temp)/2 obtain (k+1) the slope absolute value λ in step _K+1, according to f _K+1=f (λ _K+1) obtain the bit number in (k+1) step and distribute error amount f _K+1Otherwise just according to λ _K+1=λ _TempObtain (k+1) the slope absolute value λ in step _K+1, according to f _K+1=f _TempThe bit number that obtains (k+1) step distributes error amount f _K+1

Sign (f wherein _k) the is-symbol function, be defined as $\mathrm{Sign}\left(v\right)=\left\{\begin{array}{cc}1,& \mathrm{ifv}\&GreaterEqual;0\\ -1,& \mathrm{ifv}<0\end{array}\right.;$

The iteration step number is added 1, i.e. k=k+1;

(4) if iteration step number k surpasses greatest iteration step number MaxNum or current bit number distributes the absolute value of error to distribute error threshold MinDiff less than bit number, then stop computational process, (3) slope absolute value that obtains that goes on foot as optimum point of cut-off place slope absolute value, is gone on foot the continuation iterative computation otherwise just jump to (2).

4, a kind of code rate allocation method of fine granularity scalable video coding is characterized in that, this method is the Data Rate Distribution that waits distortion based on rate distortion function, specifically may further comprise the steps:

1) at first by to any one encoded point sampling in each bit plane of every two field picture in the enhancement layer bitstream and generate the rate distortion functional curve of every two field picture in the code check allocation units according to the sample point interpolation;

2) by the iterative computation formula $D_{k+1}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD}),$ Obtain the distortion value that waits the distortion Data Rate Distribution of the rate distortion curve of every two field picture in the Data Rate Distribution unit, wherein, k represents iterations, D _kRepresent the distortion value that the k step calculates, F _kExpression is by formula $F\left(D\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(D\right)-R_c$ Functional value F (the D of definition _k), R _cBe the given total bit number of allocation units, N represents the frame number of allocation units image, R _iRepresent the bit number that i frame enhancement layer bitstream distributes, D _iRepresent that this frame is at R _iFollowing corresponding distortion, i.e. R _i(D) be the rate distortion function of i two field picture;

3) obtain the pairing bit number of distortion value of distortion Data Rate Distribution such as every two field picture again according to approach based on linear interpolation by the bit number that waits two adjacent sample points of distortion corresponding points, with this bit number as the bit number of distributing to this two field picture.

5, the code rate allocation method of fine granularity scalable video coding as claimed in claim 4, it is characterized in that, the rate distortion functional curve of the every two field picture of said generation, concrete steps are: each the Bit-Plane Encoding starting point place sampling to every two field picture in the enhancement layer bitstream obtains the sample point (R of this image in code check and distortion R-D space _j, D _j), and these sample points are coupled together R-D function curve as this two field picture with the straightway of segmentation, distortion value on non-sampled point is that the distortion value linear interpolation by adjacent nearest two sample points with it obtains, and is to obtain by linear extrapolation for the non-sample point distortion value outside first sample point and last sample point; And the dR/dD value on each sample point is the mean value of the left and right sides dR/dD of this point and its adjacent sample point formation, but not sample point is that dR/dD value linear interpolation by adjacent nearest two sample points with it obtains, for the sample point dR/dD value replacement nearest with it of the dR/dD value of non-sample point outside first sample point and last sample point.

6, the code rate allocation method of fine granularity scalable video coding as claimed in claim 4 is characterized in that, said this iterative computation specifically may further comprise the steps:

(1) initializing variable value: k=0, D _k=D ₀, F _k=F (D ₀);

D ₀It is initial distortion value;

Set greatest iteration step number MaxNum, bit number distributes error threshold MinDiff;

(2) calculate (k+1) step distortion value D _K+1Temporary variable D _Temp: $D_{\mathrm{temp}}=D_k-F\left(D_k\right)/(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{dR}}_i\left(D_k\right)/\mathrm{dD});$

Calculate (k+1) step distortion value D _K+1Temporary variable F _Temp: $F_{\mathrm{temp}}=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{R}_i\left(D_{\mathrm{temp}}\right)-R_c;$

(3) if Sign is (F _k) and Sign (F _Temp) difference, so just according to D _K+1=(D _k+ D _Temp)/2 obtain (k+1) the distortion value D in step _K+1, according to F _K+1=F (D _K+1) obtain (k+1) step bit number and distribute error F _K+1Otherwise just according to D _K+1=D _TempObtain (k+1) the distortion value D in step _K+1, according to F _K+1=F _TempObtain (k+1) step bit number and distribute error F _K+1

Sign (f wherein _k) the is-symbol function, be defined as $\mathrm{Sign}\left(v\right)=\left\{\begin{array}{cc}1,& \mathrm{ifv}\&GreaterEqual;0\\ -1,& \mathrm{ifv}<0\end{array}\right.;$

The iteration step number is added 1, i.e. k=k+1;

(4) if the iteration step number distributes Error Absolute Value less than MinDiff above MaxNum or current bit number, then stop computational process, (3) distortion value that obtains that goes on foot as last result of calculation, is gone on foot the continuation iterative computation otherwise just jump to (2).

Images