CN101854534B - Fast interframe mode selection method in H. 264 - Google Patents

Abstract

The invention discloses a fast interframe mode selection method in H. 264, which comprises the following steps: preferentially judging a skip mode, and reducing the macroblock level mode judging range on the basis; and according to the relationship between a sub-macroblock coding mode and a quantization parameter QP, carrying out set division for a sub-macroblock level mode, selecting different sub-macroblock mode sets for macroblocks of different quantization parameters QP, and reducing the judging range of the sub-macroblock level mode. With the technical scheme of the invention, through the research and analysis of distribution situations of each prediction mode of different video sequences under different conditions, common grounds are found to reduce the range of prediction modes so as to speed up the mode judgment; in addition, through the research and analysis of sub-macroblock mode distribution of different quantization parameters QP, the mode judging method based on quantization parameters QP is further provided, so that the coding speed is further improved under the condition of very effectively keeping the quality of image compression.

Description

A kind of H.264 fast interframe mode selection method

Technical field

The present invention relates to a kind of inter-frame mode selecting method of natural video frequency sequence, specifically relate to a kind of being applied in H.264 in the coding techniques, can judge the system of selection of the final coding mode of inter macroblocks rapidly and accurately.

Background technology

In the advanced encoder technology that H.264 adopts, the various modes coding techniques of inter macroblocks is to cause one of key factor that complexity significantly increases.For realizing the highest code efficiency, encoder is when each macro block of coding, all to travel through all coding modes, certainly will increase the amount of calculation of encoder so greatly, therefore how to judge the final coding mode of inter macroblocks rapidly and accurately, simultaneously making code efficiency unaffected again, is a problem that is worth research.

For judging the final coding mode of macro block fast, there is following several method now in the technology:

1) utilizes the Sobel operator to obtain the textural characteristics of macro block, dwindle the pattern determination range according to this feature;

2), dwindle the pattern determination range according to the height of level of detail by calculating the level of detail that AC coefficient energy obtains macro block;

3) utilize the time and space correlation to come aero mode to judge in conjunction with threshold decision simultaneously;

4) utilize hierarchical policy, select the different incompatible pattern determination ranges of dwindling of set of patterns according to the image residual error that motion compensation obtains.

Though these above-mentioned methods have also been judged the final coding mode of inter macroblocks, guaranteeing picture quality and do not influencing under the situation of compression efficiency that it is not clearly that speed of coding improves, thereby the shortening of scramble time is also very limited.

Summary of the invention

The purpose of this invention is to provide a kind ofly under the situation that does not influence picture quality and compression efficiency, effectively reduce a kind of H.264 fast interframe mode selection method of scramble time.

For achieving the above object, the present invention is by the following technical solutions:

The present invention judges the skip mode prioritization, and dwindles the pattern determination range of macro-block level on this basis; Secondly according to the relation of inferior macro-block coding pattern and quantization parameter QP, inferior macro-block level pattern is gathered division, the macro block under the different quantization parameter QP is selected different inferior macro block mode set, dwindle inferior macro-block level pattern determination range.

The present invention includes following steps:

1. carry out 16 * 16 macro block mode estimation, and calculated distortion rate cost RD _{16 * 16}

2. be the former frame of present frame as if the optimal reference frame, and it is identical with the motion vector that obtains by prediction to collect the optimal motion vectors that obtains by motion, forwards step to 3.; If the optimal reference frame is not the former frame of present frame, and collect the optimal motion vectors that obtains by motion inequality with the motion vector that obtains by prediction, forwards step to 5.;

3. conversion and quantification are carried out in brightness and colourity residual error;

4. after brightness and colourity residual error are carried out change quantization, all be zero, forward step to 9. if quantize the back coefficient; Coefficient all is not zero if it quantizes the back, then forwards step to 6.;

5. carry out 16 * 8 and 8 * 16 macro block mode estimation, and calculate their distortion rate cost RD respectively _{16 * 8}And RD _{8 * 16}, and the span of judgement quantization parameter QP;

6. as if quantization parameter QP＜=24, each 8 * 8 inferior macro block is carried out 8 * 8 and 4 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{4 * 4},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

7. as if quantization parameter QP＜=36, each 8 * 8 inferior macro block is carried out 8 * 4 and 4 * 8 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 4}And RD _{4 * 8},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

8. as if quantization parameter QP＞36, each 8 * 8 inferior macro block is carried out 8 * 8 macro blocks and 8 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{8 * 4}, each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum, and calculates distortion rate D _{8 * 8}Value;

The pattern of 9. getting distortion rate cost minimum is final pattern.

Adopt the present invention of technique scheme,, find common ground, dwindle the predictive mode scope and add fast mode and judge by researching and analysing the distribution situation of each predictive mode of different video sequence under different condition.Also, the inferior macro block mode under the different quantization parameter QP researchs and analyses in addition by being distributed, mode judging method based on quantization parameter QP has further been proposed, this makes and can further improve coding rate very effectively under the condition that keeps image compression quality.Specifically, to different sequences and different quantization parameter QP, the present invention can both significantly reduce the complexity of encoder, and at picture quality and bit rate connection, performance is suitable.Under identical quantization parameter QP, the complexity difference that different sequence reduced.Have different mode profile because have the sequence of different texture features and kinetic characteristic, thereby cause improving the difference of coding rate.Under the very little situation of QP,, also can effectively reduce encoder complexity to the sequence of texture-rich, complicated movement.Even because for the little sequence of skip pattern ratio, dwindle the model selection scope by twice, promptly macro-block level and inferior macro-block level also can effectively improve coding rate.To same sequence, under different QP, the complexity of reduction is also different.QP is big more, and skip pattern ratio is also corresponding big more, thereby it is many more to make that complexity reduces.

Description of drawings

Fig. 1 is 7 kinds of inter macroblocks predictive coding pattern diagram among the present invention;

Fig. 2 is a flow chart of the present invention;

Table 1 is the shared percentage of not homotactic coding mode;

Table 2 is 16 * 16 when satisfying condition B, C for coding mode among the present invention, each coding mode percentage;

Table 3 is 16 * 16 when satisfying condition B, C, D for coding mode among the present invention, each coding mode percentage;

Table 4 is the distribution of the Central Asia of the present invention macro-block coding pattern;

{ { 8 * 4,4 * 8} coding efficiency relatively for 8 * 8,8 * 4,4 * 8} and set for set among the present invention for table 5;

Table 6 is an experimental result of the present invention.

Embodiment

As shown in Figure 1, for the inter macroblocks coding, H.264 adopt multiple macro block and inferior macroblock partition pattern to carry out estimation, an inter-coded macroblocks can be cut apart according to 16 * 16,16 * 8,8 * 16 and 8 * 8 four kinds of macro-block level modes.Cut apart pattern for 8 * 8, can further cut apart in 8 * 8,8 * 4,4 * 8 and 4 * 4 four kinds of inferior macro-block level modes.Simultaneously, H.264 also allow the P/B frame to adopt intra prediction mode coding, i.e. 16 * 16 predictive modes in the frame of the intra-frame 4 * 4 forecasting model of 9 directions and 4 directions.Except above-mentioned pattern, at background static in the image and global motion part, H.264 defined the skip pattern, the skip pattern does not need coded residual signal and motion vector, and the macro block of the skip pattern of therefore encoding only needs bit number seldom.

In the natural video frequency sequence, exist on the big quantity space evenly smooth or time domain is static and the very little zone of motility, and the skip pattern is occupied quite great proportion in these zones.Table 1 is the shared percentage of not homotactic coding mode.

	Stefan	Mother	Silent
				skip	16.79	58.98	63.87
16×16	20.39	18.49	12.30
				16×8	9.46	6.68	4.31
8×16	7.35	7.57	5.23
				P8×8	38.76	6.80	10.70
Intra4×4	2.35	0.79	2.17
				Intra16×16	4.91	0.70	1.42

Table 1

In table 1, P8 * 8 expression macro-block level are cut apart 8 * 8 patterns of employing.Table 1 data show, the large percentage that the skip pattern is shared.Thereby the skip pattern preferentially judged the complexity that can greatly reduce encoder.Table 1 data also show, for the sequence of kinetic characteristic or textural characteristics complexity, the large percentage that P8 * 8 are shared.In the estimation number of times of whole model selection, occupied 36/41 more than owing to P8 * 8 estimation number of times simultaneously, thereby dwindled inferior macro block mode scope and can further improve code efficiency.

If macro block is encoded with the skip pattern, then this must satisfy following condition:

The best macroblock partitions of condition A. motion compensation is 16 * 16;

Condition B. reference frame is the former frame of present frame;

Condition C. the optimal motion vectors that obtains by motion search is identical with the motion vector that obtains by prediction;

Condition D. residual error is 0 through behind the change quantization.

Obviously, the condition above macro block satisfies is many more, and then the final coding mode of this macro block is that the possibility of skip pattern is just big more.

When satisfying the present encoding pattern is 16 * 16 when condition B, the C, and each coding mode percentage is as shown in table 2.

	Stefan	Mother	Silent
				skip	44.15	85.51	85.62
16×16	22.98	8.95	8.21

16×8	4.04	1.59	0.96
				8×16	3.31	1.86	1.22
P8×8	16.57	1.40	3.25
				Intra4×4	0.46	0.13	0.23
Intra16×16	8.50	0.56	0.51

Table 2

Afterwards residual error is carried out transform and quantization, judge whether quantization parameter QP all is 0, i.e. condition D.When satisfying the present encoding pattern is 16 * 16 during with condition B, C, D, and each coding mode percentage is as shown in table 3.

	Stefan	Mother	Silent
				Skip	82.93	98.10	98.61
16×16	0.00	0.00	0.00
				16×8	1.11	0.62	0.21
8×16	0.81	0.71	0.28
				P8×8	0.28	0.09	0.64
Intra4×4	0.00	0.00	0.00
				Intra16×16	14.88	0.48	0.25

Table 3

The data of table 3 show, with this understanding, the shared ratio sum in skip and Intra16 * 16 is more than 91%, and along with motion gradually slowly, gradually simple of texture, itself and big more.Also can find out from table 2 and table 3 simultaneously, satisfy condition B and C and do not satisfy under the situation of D, final coding mode concentrate on 16 * 16, P8 * 8, Intra16 * 16}.

The distribution of inferior macro-block coding pattern is as shown in table 4:

Table 4

The shared ratio of 8 * 8 patterns constantly increases in P8 * 8 patterns as can be seen from Table 4, and 4 * 4 constantly reduce, and 8 * 4 and 4 * 8 variations are little.When QP is less, the mode profile of inferior macro block mainly concentrate on set 8 * 8,4 * 4}; When QP is big, mainly concentrate on set 8 * 8,8 * 4}; When QP gets medium numerical value, mainly concentrate on set 8 * 8,8 * 4,4 * 8}.Thereby according to this phenomenon P8 * 8 are divided into plurality of sub set and correspond respectively to different Q P.Weigh code efficiency and speed simultaneously, the inferior macro block mode subclass of interpolation under medium QP 8 * 4,4 * 8}.Set 8 * 8,8 * 4,4 * 8} and set 8 * 4,4 * 8} coding efficiency is more as shown in table 5:

Table 5

Table 5 is in conjunction with above-mentioned thought, and inferior macro block mode is got the coding efficiency table that different mode subsets is compared with FMS.{ 8 * 8,8 * 4,4 * 8} is mid_b, and { 8 * 4,4 * 8} is mid_l in set in the note set.

Wherein "-" expression the present invention compares reduction with FMS, and "+" increases, and the Y-PSNR of " Δ psnr " expression brightness changes the dB of unit; " Δ br " represents bit rate variation, and its computing formula is suc as formula (2); " Δ t " the presentation code time changes, and its computing formula is suc as formula (3).

$\mathrm{\Δbr}=\frac{\mathrm{bitrate}\left(\mathrm{FMS}\right)-\mathrm{bitrate}\left(\mathrm{proposed}\right)}{\mathrm{bitrate}\left(\mathrm{FMS}\right)}\×100\%---\left(2\right)$

$\mathrm{\Δt}=\frac{\mathrm{time}\left(\mathrm{FMS}\right)-\mathrm{time}\left(\mathrm{proposed}\right)}{\mathrm{time}\left(\mathrm{FMS}\right)}\×100\%---\left(3\right)$

Draw easily from table 5, under medium QP, the set of inferior macro block mode gets that { 8 * 4,4 * 8} effect is better.

Since above-mentioned demonstration, thereby the present invention can realize based on such principle: the skip mode prioritization is judged, and dwindled the pattern determination range of macro-block level on this basis; Secondly according to the relation of inferior macro-block coding pattern and quantization parameter QP, inferior macro-block level pattern is gathered division, the macro block under the different quantization parameter QP is selected different inferior macro block mode set, dwindle inferior macro-block level pattern determination range.

Thereby it may further comprise the steps:

1. 16 * 16 macro block modes are carried out estimation, and calculated distortion rate cost RD _{16 * 16}

Estimation is a technology well known to those of ordinary skill in the art, the estimation of 16 * 16 patterns is cut apart pattern is sought best matching blocks in reference frame process at 16 * 16 exactly, the foundation of selecting is exactly to select the pattern of cost minimum according to the cost formula, and the cost formula is:

RD(mode)＝D(S，C，mode|QP)+λ _mode×R(S，C，mode|QP)

(S, C are under prescribed coding mode m ode and quantization parameter QP mode|QP) to D in the formula, the distortion factor between primary signal S and the reconstruction signal C; λ _ModeBe Lagrangian Lagrange multiplier factor, by quantization parameter decision, λ _Mode=0.85 * 2 ^Qp/3(S, C are under designated mode mode and quantization parameter mode|QP) to R, the bit number that macroblock coding is required.

Wherein the calculating of distortion factor employing SATD (Sum of Absolute Transformed Difference) is absolute value summation again after the hadamard conversion.

The hadamard transformation for mula is as follows:

$F=\left[\begin{array}{cccc}1& 1& 1& 1\\ 1& 1& -1& -1\\ 1& -1& -1& 1\\ 1& -1& 1& -1\end{array}\right]\left[\begin{array}{cccc}{c}_{00}& c_{01}& c_{02}& c_{03}\\ c_{10}& c_{11}& c_{12}& c_{13}\\ c_{20}& c_{21}& c_{22}& c_{23}\\ c_{30}& c_{31}& c_{32}& c_{33}\end{array}\right]\left[\begin{array}{cccc}1& 1& 1& 1\\ 1& 1& -1& -1\\ 1& -1& -1& 1\\ 1& -1& 1& -1\end{array}\right]$

That the cost among the present invention is calculated employing is SATD, and its value has reflected the size that generates code stream to a certain extent.

2. be the former frame of present frame as if the optimal reference frame, and it is identical with the motion vector that obtains by prediction to collect the optimal motion vectors that obtains by motion, forwards step to 3.; If the optimal reference frame is not the former frame of present frame, and collect the optimal motion vectors that obtains by motion inequality with the motion vector that obtains by prediction, forwards step to 5.;

3. conversion and quantification being carried out in brightness and colourity residual error, promptly mainly is that the original pixels of brightness and colourity and the difference of predict pixel are carried out conversion and quantification.

H.264 the unique characteristic of another one is that it has utilized 4 * 4 integer space conversion, is not that the former coding standard of picture adopts 8 * 8 floating-point dct transform.If original sample point is χ _Ij, the matrix behind the dct transform is Y, then the dct transform formula is as follows:

$Y=\left[\begin{array}{cccc}1& 1& 1& 1\\ 2& 1& -1& -2\\ 1& -1& -1& 1\\ 1& -2& 2& -1\end{array}\right]\left[\begin{array}{cccc}{x}_{00}& x_{01}& x_{02}& x_{03}\\ x_{10}& x_{11}& x_{12}& x_{13}\\ x_{20}& x_{21}& x_{22}& x_{23}\\ x_{30}& x_{31}& x_{32}& x_{33}\end{array}\right]\left[\begin{array}{cccc}1& 2& 1& 1\\ 1& 1& -1& -2\\ 1& -1& -1& 2\\ 1& -2& 1& -1\end{array}\right]$

Quantitative formula

Z _Ij=round (Y _IjMF/2 ^Qbits)---formulate

Round () is bracket function (the nearest integer of its input value position in-real), Y _IjFor being the conversion coefficient in the above-mentioned Y matrix, MF is a quantizing factor, by H.264 standard code, and qbits=15+floor (QP/6), wherein floor () is a bracket function, and its output valve is the maximum integer that is not more than in-real, and QP is a quantization parameter.

4. after brightness and colourity residual error are carried out change quantization,, forward step to 9. if quantization parameter QP is zero; If its quantization parameter QP is non-vanishing, then forward step to 6.;

5. 16 * 8 macro block modes and 8 * 16 macro block modes are carried out estimation, and calculate their distortion rate cost RD respectively _{16 * 8}And RD _{8 * 16}, judge the span of quantization parameter QP then;

6. each 8 * 8 inferior macro block is carried out 8 * 8 and 4 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{4 * 4},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

7. as if quantization parameter QP＜=36, each 8 * 8 inferior macro block is carried out 8 * 4 and 4 * 8 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 4}And RD _{4 * 8},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

8. as if quantization parameter QP＞36, each 8 * 8 inferior macro block is carried out 8 * 8 macro blocks and 8 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{8 * 4},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

The pattern of 9. getting distortion rate cost minimum is final pattern.

The present invention realizes on the JM10.1 basis, tests with 100 frames of 4 different standard test sequences.

The encoder major parameter is provided with: image format size: CIF;

Frame per second: 30fps;

GOP：IPP…；

Frameskip：1；

Hunting zone: 16;

Reference frame: 5;

Entropy coding:: CABAC;

Rdo：off。

Hardware platform: Pentium (R) 4CPU 3.00GHZ, the 512MB internal memory.

Its experimental result is as shown in table 6:

Table 6

As can be seen from Table 6, to different sequences and QP, this paper algorithm can both significantly reduce the complexity of encoder, at picture quality and bit rate connection, though performance slightly descend, suitable substantially.

Under identical QP, the complexity difference that different sequence reduced.Have different mode profile because have the sequence of different texture features and kinetic characteristic, thereby cause improving the difference of coding rate.

Under the very little situation of QP,, also can effectively reduce encoder complexity to the sequence of texture-rich, complicated movement.Even because for the little sequence of skip pattern ratio, dwindle the model selection scope by twice, promptly macro-block level and inferior macro-block level also can effectively improve coding rate.

To same sequence, under different QP, the complexity of reduction is also different.QP is big more, and skip pattern ratio is also corresponding big more, thereby it is many more to make that complexity reduces.

Claims (1)

1. a fast interframe mode selection method H.264 is characterized in that: the skip mode prioritization is judged, and dwindled the pattern determination range of macro-block level on this basis; Secondly according to the relation of inferior macro-block coding pattern and quantization parameter QP, inferior macro-block level pattern is gathered division, the macro block under the different quantization parameter QP is selected different inferior macro block mode set, dwindle inferior macro-block level pattern determination range; It may further comprise the steps:

1. carry out 16 * 16 macro block mode estimation, and calculated distortion rate cost RD _{16 * 16}, the order execution in step 2. then;

2. be the former frame of present frame as if the optimal reference frame, and it is identical with the motion vector that obtains by prediction to collect the optimal motion vectors that obtains by motion, forwards step to 3.; If the optimal reference frame is not the former frame of present frame, and collect the optimal motion vectors that obtains by motion inequality with the motion vector that obtains by prediction, forwards step to 5.;

3. conversion and quantification are carried out in brightness and colourity residual error, the order execution in step 4.;

4. after brightness and colourity residual error are carried out change quantization, all be zero if quantize the back coefficient, the final coding mode that then can preferentially judge this macro block is the skip pattern; Coefficient all is not zero if it quantizes the back, then forwards step to 6.;

5. carry out 16 * 8 and 8 * 16 macro block mode estimation, and calculate their distortion rate cost RD respectively _{16 * 8}And RD _{8 * 16}, and the span of judgement quantization parameter QP;

6. as if quantization parameter QP＜=24, each 8 * 8 inferior macro block is carried out 8 * 8 and 4 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{4 * 4},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value; If quantization parameter QP＞24, then execution in step 7.;

7. as if quantization parameter QP＜=36, each 8 * 8 inferior macro block is carried out 8 * 4 and 4 * 8 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 4}And RD _{4 * 8},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value; If quantization parameter QP＞36, then execution in step 8.;

8. as if quantization parameter QP＞36, each 8 * 8 inferior macro block is carried out 8 * 8 macro blocks and 8 * 4 inferior macro block mode estimation, and calculate their distortion rate cost RD respectively _{8 * 8}And RD _{8 * 4},, and calculate distortion rate D for each 8 * 8 inferior macro block is selected the pattern of distortion rate minimum _{8 * 8}Value;

9. according to step 6., 7., 8. result of calculation, the pattern of getting distortion rate cost minimum is final pattern.

Images