CN101198059A - Integer translation base optimization method in video coding standard - Google Patents

Abstract

The invention relates to the video coding field, in particular to an integer transform basis optimizing method in video coding standards, which is characterized in that the optimizing method comprises the following steps: firstly, an expression of a corresponding integer transformation matrix I is solved according to a DCT transformation coefficient matrix C, wherein, a coefficient marker of the C is ck and a coefficient mark of the I is ik, and K is equal to 0 to N; secondly, a calibration coefficient jk is solved after normalization of ik, and k is equal to 0 to N; thirdly, search finding of the ik is performed by adoption of a search method, and transform basis is composed, wherein, following four constraint conditions must be followed during the search process: elements except for diagonals in a IIT matrix all meet the condition that the value is 0; step size in search is limited to be one half during the process of search founding of the ik; the calibration coefficient jk must meet threshold requirements, and threshold values are subject to transform basis number which is expected to be obtained through search; maximum amplitude of the coefficient ik in the integer transformation matrix I is 15; fourthly, the transform basis obtained through search is scaled to be an integer, and integer transform basis corresponding to the matrix C is obtained. High quality integer transform basis can be comprehensively and quickly obtained by adoption of the invention.

Description

Integer translation base optimization method in a kind of video encoding standard

Technical field

The invention belongs to field of video encoding, the integer translation base optimization method in particularly a kind of video encoding standard.

Background technology

H.264, a new generation's video encoding standard is to be responsible for exploitation by the joint video team (JVT) of Union of International Telecommunication (ITU) and International Standards Organization (ISO), towards the standard of the latest development of practical application.Its target is based on high video resolution, improves picture quality, and the application that can cover all low bandwidth and high bandwidth.H.264 be that what H.26L to release on the basis at ITU-T enhanced multimedia communication standard can be ITU-T and the common single video encoding standard of using of new generation of ISO/IEC, and form system with mpeg standard technically.

H.264 similar to previous standard, residual error data is adopted block-based transition coding, transition coding can be removed the spatial redundancy of original image, and image energy is concentrated on the sub-fraction coefficient, improves compression ratio, strengthens antijamming capability.And the conversion of H.264 adopting is the no multiplication integer transform of approximate DCT, claims the integer dct transform again.It should be noted that conversion herein has not been real DCT, still be called dct transform just in order to say that it is to be derived by DCT, and in order to distinguish mutually with another conversion (Hadamard conversion).H.264 have only integer arithmetic in the integer dct transform, eliminated floating-point operation, reduced operand, and accurate integer has been got rid of the error matching problem between the encoder inverse transformation.

Integer translation base is not unique, and some major companies have all proposed the integer translation base of oneself in the world, and has applied for patent protection, sets up for the own audio and video standard of China and brings obstacle.And from technological layer, although these existing integer translation bases have good conversion characteristics, generally all be to find that by the way of souning out the process of discovery has contingency, efficient is extremely low and not comprehensive.

Summary of the invention

The objective of the invention is to solve the deficiency of existing integer translation base acquisition methods, propose regular integer translation base optimization method.

Integer translation base optimization method provided by the invention may further comprise the steps:

Step 1 according to the dct transform coefficient Matrix C, is found the solution the expression formula of its corresponding integer transform matrix I, and wherein the coefficient of C is labeled as c _k, the coefficient of I is labeled as i _k, k=0～N;

Step 2 is with i _kCalibration coefficient j is tried to achieve in normalization _k, k=0～N;

Step 3 adopts searching method search finding i _kTo form transform-based, follow following 4 constraintss during search,

II ^TIt is 0 that element in the matrix beyond the diagonal all satisfies value, described II ^TMatrix multiply by its transposed matrix by integer transform matrix I and obtains;

Restriction search finding i _kThe time step-size in search be 1/2;

Calibration coefficient i _kNeed to satisfy $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|<T_f$ K=0,1 ..., N, T _fExpression thresholding, threshold value depend on the transform-based number that the expectation search obtains;

Coefficient i among the round numbers transformation matrix I of institute _kAmplitude peak be 15;

Step 4, the transform-based yardstickization that search is obtained arrives integer, obtains the integer translation base of dct transform coefficient Matrix C correspondence.

And, further strict constraints of searching for, calibration coefficient j _kNeed to satisfy $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|*f_k<T_f$ K=0,1 ..., N, wherein weight coefficient $f_k=\frac{c_k}{\underset{j=0}{\overset{j=N}{\mathrm{\&Sigma;}}}{c}_j}$ k＝0，1...N。

And, described T _fNumerical value get 0.03.

And it is 1 that coefficient 0.353553 search in the 8x8 dct transform matrix obtains corresponding integer translation base.

And 4 coefficients, 0.490393,0.415735,0.277785,0.097545 search in the 8x8 dct transform matrix obtains 15 groups of corresponding integer translation bases, and 4 coefficients in the integer translation base are respectively: (1) 12.5,10,7,2.5; (2) 12,10.5,7.5,2; (3) 13.5,11,7,3; (4) 7.5,6,4,1.5; (5) 10,9,6,2; (6) 12,10,6,3; (7) 14,12.5,7.5,3.5; (8) 7,6,4.5,1; (9) 9,7,5,1.5; (10) 13,12,7.5,3; (11) 10.5,9,5,3; (12) 8,7.5,4.5,2; (13) 9,7.5,6,1; (14) 4.5,4,2,1.5; (15) 6,5,2.5,2.

And 2 coefficients, 0.461940,0.191342 search in the 8x8 dct transform matrix obtains 11 groups of corresponding integer translation bases, and 2 coefficients in the integer translation base are respectively: (1) 6,2.5; (2) 5,2; (3) 7,3; (4) 6.5,2.5; (5) 4.5,2; (6) 4,1.5; (7) 5.5,2.5; (8) 6.5,3; (9) 5.5,2; (10) 3,1; (11) 3,1.5.

And 2 coefficients, 0.653282,0.270598 search in the 4x4 dct transform matrix obtains 7 groups of corresponding integer translation bases, and 2 coefficients in the integer translation base are respectively: (1) 6,2.5; (2) 5,2; (3) 7,3; (4) 4.5,2; (5) 8,3; (6) 3,1; (7) 2,1.

The present invention is by analyzing the principle of integer dct transform, four constraintss that the integer transform matrix should satisfy have been pointed out, be constrained to starting point to satisfy orthogonality then, derived the quantitative relation between the integer transform matrix element, and be aided with other three constraintss, adopt searching method to seek integer translation base.Adopt technical scheme method of the present invention, only need just can find out all available integer translation bases through after tens search that go on foot, efficient is high and comprehensive.Communication quality when the present invention not only helps improving the practical application video standard has the video compression standard of independent intellectual property right significant for China's formulation.

Embodiment

Integer translation base optimization method provided by the invention may further comprise the steps:

Step 1 according to the dct transform coefficient Matrix C, is found the solution the expression formula of its corresponding integer transform matrix I, and wherein the coefficient of C is labeled as c _k, the coefficient of I is labeled as i _k, k=0～N; The dct transform coefficient Matrix C generally comprises M * M element, M=4,8,16 N represents the number of the coefficient that numerical value in the coefficient matrix is different, and numerical value is decided as the case may be;

Step 2 is with i _kCalibration coefficient j is tried to achieve in normalization _k, k=0～N;

Step 3 adopts searching method search finding i _kTo form transform-based, follow following 4 constraintss during search,

(a) II ^TIt is 0 that element in the matrix beyond the diagonal all satisfies value, described II ^TMatrix multiply by its transposed matrix by integer transform matrix I and obtains;

(b) restriction search finding i _kThe time step-size in search be 1/2;

(c) calibration coefficient j _kNeed to satisfy $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|<T_f$ K=0,1 ..., N, T _fExpression thresholding, threshold value depend on the transform-based number that the expectation search obtains;

(d) coefficient i among institute's round numbers transformation matrix I _kAmplitude peak be 15;

Step 4, the transform-based yardstickization that search is obtained arrives integer, obtains the integer translation base of dct transform coefficient Matrix C correspondence.Because the integer transform computing only allows to occur arithmetical operations such as adding, subtract, taking advantage of of integer type and displacement, so the coefficient i that search obtains _kExcept that integer, 0.5 such value (0.5 means 1 the computing of moving to right) can also appear.Can be during concrete enforcement by the calibration coefficient j that obtains in the search procedure _kConstitute normalization matrix, by revising normalization matrix, transform-based can yardstickization arrive integer.

4 fundamental characteristics of integer transform matrix are: 1. can be normalized into orthogonal matrix; 2. transform operation can all be realized with shift operation; 3. the orthogonal matrix after the normalization is close with original dct transform matrix coefficient; 4. the integer transform matrix coefficient is not very big, and transform operation can be with 16 accuracy representings.The present invention is according to above-mentioned 4 principles, and the method for optimizing of design integer transform matrix is converted to the constraints of searching for transform-based with 4 fundamental characteristics, obtains suitable integer translation base under effective search comprehensively.Wherein the corresponding characteristic of constraints (a) 1.; The corresponding characteristic of constraints (b) 2.; The corresponding characteristic of constraints (c) 3.; The corresponding characteristic of constraints (d) 4..Because 1. characteristic is the basis, should be constrained to starting point to satisfy orthogonality during search, has derived the quantitative relation between the integer transform matrix I interior element, is aided with other three constraintss then and adopts searching method to seek transform-based.Owing to provided fully preferred constraints, just can find certain coefficient c in the search of tens steps _kPairing all preferred integer translation base.

Thresholding T during concrete enforcement _fNumerical value depends on the transform-based number that the expectation search obtains, can increase its numerical value, to reduce the transform-based quality is the transform-based that cost obtains more candidates, when the present invention advises that 3. preferred embodiment adopts the 0.03. consideration characteristics, the contribute energy of the bigger coefficient of the amplitude that it is also conceivable that after to conversion is bigger, further strict search constraints condition (c) calibration coefficient j _kNeed to satisfy and be $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|*f_k<T_f$ K=0,1 ..., N, wherein weight coefficient $f_k=\frac{c_k}{\underset{j=0}{\overset{j=N}{\mathrm{\&Sigma;}}}{c}_j}$ k＝0，1...N。After the weight coefficient correction, it is more accurate to ask for the integer translation base that obtains.

With the embodiment that is solved to of 8x8 integer transform matrix, its generative process is as follows: 8x8 dct transform coefficient matrix is

$\left[\begin{array}{cccccccc}0.353553& 0.490393& 0.461940& 0.415735& 0.353553& 0.277785& 0.191342& 0.097545\\ 0.353553& 0.415735& 0.191342& -0.097545& -0.353553& -0.490393& -0.461940& -0.277785\\ 0.353553& 0.277785& -0.191342& -0.490393& -0.353553& 0.097545& 0.461940& 0.415735\\ 0.353553& 0.097545& -0.461940& -0.277785& 0.353553& 0.415735& -0.191342& -0.490393\\ 0.353553& -0.097545& -0.461940& 0.277785& 0.353553& -0.415735& -0.191342& 0.490393\\ 0.353553& -0.277785& -0.191342& 0.490393& -0.353553& -0.097545& 0.461940& -0.415735\\ 0.353553& -0.415735& 0.191342& 0.097545& -0.353553& 0.490393& -0.461940& 0.277785\\ 0.353553& -0.490393& 0.461940& -0.415735& 0.353553& -0.277785& 0.191342& -0.097545\end{array}\right]$

Coefficient has 7 kinds of values: 0.353553,0.490393,0.415735,0.277785,0.097545,0.461940,0.191342, be designated as successively: c ₀, c ₁, c ₂, c ₃, c ₄, c ₅, c ₆Make its corresponding integer dct transform matrix coefficient be respectively: i ₀, i ₁, i ₂, i ₃, i ₄, i ₅, i ₆, then INTEGER MATRICES I can be written as

$\left[\begin{array}{cccccccc}{i}_0& i_0& i_0& i_0& i_0& i_0& i_0& i_0\\ i_1& i_2& i_3& i_4& {-i}_4& {-i}_3& {-i}_2& {-i}_1\\ i_5& i_6& {-i}_6& {-i}_5& {-i}_5& {-i}_6& i_6& i_5\\ i_2& {-i}_4& {-i}_1& {-i}_3& i_3& i_1& i_4& {-i}_2\\ i_0& {-i}_0& {-i}_0& i_0& i_0& {-i}_0& {-i}_0& i_0\\ i_3& {-i}_1& i_4& i_2& {-i}_2& {-i}_4& i_1& -i_3\\ i_5& {-i}_6& i_6& {-i}_5& {-i}_5& i_6& {-i}_6& i_5\\ i_4& {-i}_3& i_2& {-i}_1& i_1& {-i}_2& i_3& {-i}_4\end{array}\right]$

In order to satisfy orthogonality, consider

${\mathrm{II}}^T=\left[\begin{array}{cccccccc}a& 0& 0& 0& 0& 0& 0& 0\\ 0& b& 0& -d& 0& d& 0& 0\\ 0& 0& c& 0& 0& 0& 0& 0\\ 0& -d& 0& b& 0& 0& 0& d\\ 0& 0& 0& 0& a& 0& 0& 0\\ 0& d& 0& 0& 0& b& 0& d\\ 0& 0& 0& 0& 0& 0& c& 0\\ 0& 0& 0& d& 0& d& 0& b\end{array}\right]$

Wherein,

a＝8i ₀ ²

b＝2(i ₁ ²+i ₂ ²+i ₃ ²+i ₄ ²)

c＝4(i ₅ ²+i ₆ ²)

d＝i ₃(i ₁+i ₄)-i ₂(i ₁-i ₄)

With coefficient i ₀, i ₁, i ₂, i ₃, i ₄, i ₅, i ₆Normalization obtains calibration coefficient j ₀, j ₁, j ₂, j ₃, j ₄, j ₅, j ₆, also,

$j_k=\left\{\begin{array}{cc}\frac{j_k}{\sqrt{a}}& k=0\\ \frac{j_k}{\sqrt{b}}& k=\mathrm{1,2,3,4}\\ \frac{i_k}{\sqrt{c}}& k=\mathrm{5,6}\end{array}\right.$

According to above-mentioned solution procedure (a) d=0, i.e. i are arranged ₃(i ₁+ i ₄)=i ₂(i ₁-i ₄); Step-size in search when seeking integer translation base according to condition (b) restriction is 1/2; According to the coefficient amplitude peak in condition (d) the restriction integer transform matrix is 15; (c) obtains normalization coefficient j according to solution procedure _kNeed satisfy inequality $\underset{k=0}{\overset{6}{\mathrm{\&Sigma;}}}|j_k-c_k|*f_k<T_f$ K=0,1 ..., 6, wherein, weight coefficient $f_k=\frac{c_k}{\underset{j=0}{\overset{j=6}{\mathrm{\&Sigma;}}}{c}_j}$ K=0,1...6, search thresholding T _fGet 0.03 in embodiments of the present invention.According to above-mentioned 4 constraintss, find the solution with searching method and to obtain integer translation base.

For i ₀, from the generating principle of transform-based above as can be seen, i no matter ₀Get what value, the coefficient j after the normalization ₀Be always c0, therefore, i in theory ₀Can get any integer value, but the convenience in order to calculate, embodiment gets i ₀Be 1.For i ₁, i ₂, i ₃, i ₄,, list in table 1 according to one group of transform-based selecting with the deviation order from small to large of original DCT coefficient.j ₀, j ₁, j ₂, j ₃For correspondence is demarcated the back base.In like manner, for i ₅, i ₆, can obtain one group of transform-based as shown in table 2.

Table 1 8x8 integer translation base (i ₁~ i ₄)

Sequence number	Transform-based				Demarcate the back base				Deviation
										i 1	i 2	i 3	i 4	j 1	j 2	j 3	j 4
	0	12.5	10.5	7	2.5	0.492760	0.413919	0.275946										0.098552	0.001971
1									12	10.5	7.5	2	0.478471	0.418662	0.299045	0.079745	0.011475
	2	13.5	11	7	3	0.502244	0.409236	0.260423										0.111610	0.011478
3									7.5	6	4	1.5	0.504505	0.403604	0.269069	0.100901	0.011481
	4	10	9	6	2	0.475651	0.428086	0.285391										0.095130	0.011481
5									12	10	6	3	0.499134	0.415945	0.249567	0.124784	0.011604
	6	14	12.5	7.5	3.5	0.482615	0.430906	0.258544										0.120654	0.013828
7									7	6	4.5	1	0.480196	0.411597	0.308697	0.068599	0.014149
	8	9	7	5	1.5	0.507495	0.394719	0.281942										0.084583	0.015251
9									13	12	7.5	3	0.472649	0.436292	0.272682	0.109073	0.015443
	10	10.5	9	5	3	0.494700	0.424029	0.235571										0.141343	0.016823
11									8	7.5	4.5	2	0.470588	0.441176	0.264706	0.117647	0.020198
	12	9	7.5	6	1	0.482104	0.401754	0.321403										0.053567	0.020511
13									4.5	4	2	1.5	0.488094	0.433861	0.216930	0.162698	0.024911
	14	6	5	2.5	2	0.502625	0.418854	0.209427										0.167542	0.025839

Wherein, No. 4 base is used by the AVS reference software, and No. 5 bases just in time are the integer translation bases of using in the JVT reference software.

Table 2 8x8 integer translation base (i ₅, i ₆)

Sequence number	Transform-based		Demarcate the back base		Deviation
						i 5	i 6	j 5	j 6
	0	6	2.5	0.461538						0.192308	0.001367
1					5	2	0.464238	0.185695	0.007945
	2	7	3	0.459573						0.196960	0.007985
3					6.5	2.5	0.466673	0.179490	0.016585
	4	4.5	2	0.456906						0.203069	0.016761
5					4	1.5	0.468165	0.175562	0.022005
	6	5.5	2.5	0.455183						0.206901	0.022316
7					6.5	3	0.453980	0.209529	0.026147
	8	5.5	2	0.469897						0.170872	0.028427
9					3	1	0.474342	0.158114	0.045630
	10	3	1.5	0.447214						0.223607	0.046991

The base of 4x4 integer transform also can calculate with similar method, and the integer transform coefficient of DCT coefficient 0.653282,0.270598 correspondence sees Table 3.

Table 3 4x4 integer translation base

Sequence number	Transform-based		Demarcate the back base		Deviation
						DCT 0.653282	DCT 0.270598	DCT 0.653282	DCT 0.270598
	0	6	2.5	0.652714						0.271964	0.000802
1					5	2	0.656532	0.262613	0.004637
	2	7	3	0.649934						0.278543	0.004695
3					4.5	2	0.646162	0.287183	0.009892
	4	8	3	0.662085						0.248282	0.012761
5					3	1	0.670820	0.223607	0.026165
	6	2	1	0.632456						0.316228	0.028091

JVT reference software and AVS reference software are all used No. 6 bases wherein.This coincideing can prove the reliability of method provided by the present invention completely.

More than in each table the candidate transformation basic group that provides belong to and do not obtain as yet in the prior art; because it is great for the coding and decoding video quality influence; has important market value; for China in that to occupy superiority aspect the international video encoding standard significant, the present invention's protection that claims especially.

Claims (7)

1. the integer translation base optimization method in the video encoding standard is characterized in that may further comprise the steps:

Step 1 according to the dct transform coefficient Matrix C, is found the solution the expression formula of its corresponding integer transform matrix I, and wherein the coefficient of C is labeled as c _k, the coefficient of I is labeled as i _k, k=0～N;

Step 2 is with i _kCalibration coefficient j is tried to achieve in normalization _k, k=0～N;

Step 3 adopts searching method search finding i _kConstituting transform-based, during search according to following 4 constraintss,

II ^TIt is 0 that element in the matrix beyond the diagonal all satisfies value, described II ^TMatrix multiply by its transposed matrix by integer transform matrix I and obtains;

Restriction search finding i _kThe time step-size in search be 1/2;

Calibration coefficient j _kNeed to satisfy $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|<T_f$ K=0,1 ..., N, T _fExpression thresholding, threshold value depend on the transform-based number that the expectation search obtains;

Coefficient i among the round numbers transformation matrix I of institute _kAmplitude peak be 15;

Step 4, the transform-based yardstickization that search is obtained arrives integer, obtains the integer translation base of dct transform coefficient Matrix C correspondence.

2. integer translation base optimization method as claimed in claim 1 is characterized in that: further strict constraints of searching for, calibration coefficient j _kNeed to satisfy $\underset{k=0}{\overset{N}{\mathrm{\&Sigma;}}}|j_k-c_k|*f_k<T_f$ K=0,1 .., N, wherein weight coefficient $f_k=\frac{c_k}{\underset{j=0}{\overset{j=N}{\mathrm{\&Sigma;}}}{c}_j}$ k＝0，1...N。

3. integer translation base optimization method as claimed in claim 1 is characterized in that: described T _fNumerical value get 0.03

4. as claim 1 or 2 or 3 described integer translation base optimization methods, it is characterized in that: it is 1 that coefficient 0.353553 search in the 8x8 dct transform matrix obtains corresponding integer translation base.

5. as claim 1 or 2 or 3 described integer translation base optimization methods, it is characterized in that: 4 coefficients, 0.490393,0.415735,0.277785,0.097545 search in the 8x8 dct transform matrix obtains 15 groups of corresponding integer translation bases, 4 coefficients in the integer translation base are respectively

(1)12.5、10、7、2.5；

(2)12、10.5、7.5、2；

(3)13.5、11、7、3；

(4)7.5、6、4、1.5；

(5)10、9、6、2；

(6)12、10、6、3；

(7)14、12.5、7.5、3.5；

(8)7、6、4.5、1；

(9)9、7、5、1.5；

(10)13、12、7.5、3；

(11)10.5、9、5、3；

(12)8、7.5、4.5、2；

(13)9、7.5、6、1；

(14)4.5、4、2、1.5；

(15)6、5、2.5、2。

6. as claim 1 or 2 or 3 described integer translation base optimization methods, it is characterized in that: 2 coefficients, 0.461940,0.191342 search in the 8x8 dct transform matrix obtains 11 groups of corresponding integer translation bases, and 2 coefficients in the integer translation base are respectively:

(1)6、2.5；

(2)5、2；

(3)7、3；

(4)6.5、2.5；

(5)4.5、2；

(6)4、1.5；

(7)5.5、2.5；

(8)6.5、3；

(9)5.5、2；

(10)3、1；

(11)3、1.5。

7. as claim 1 or 2 or 3 described integer translation base optimization methods, it is characterized in that: 2 coefficients, 0.653282,0.270598 search in the 4x4 dct transform matrix obtains 7 groups of corresponding integer translation bases, and 2 coefficients in the integer translation base are respectively:

(1)6、2.5；

(2)5、2；

(3)7、3；

(4)4.5、2；

(5)8、3；

(6)3、1；

(7)2、1。