EP1312220A2

EP1312220A2 - Method for carrying out integer approximation of transform coefficients, and coder and decoder

Info

Publication number: EP1312220A2
Application number: EP01956393A
Authority: EP
Inventors: Claudia Mayer; Mathias Wien
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2000-08-12
Filing date: 2001-07-27
Publication date: 2003-05-21
Also published as: WO2002015584A3; US6856262B2; WO2002015584A2; JP2004506990A; US20040046754A1; CN1242621C; CN1466854A

Abstract

For obtaining cosine transform coefficients that have undergone an integer approximation, a restricted set of values is specified. The coefficients of the base vectors for the submatrices are selected, while satisfying the orthogonality condition, in such a manner that the sum of its squares results in the square of the dc component coefficient. These coefficients are used to derive the coefficients of the ac components. The inventive method has the advantage that a uniform standardization and quantization factor can be used for all coefficients for the purpose of quantization and standardization.

Description

Method for integer approximation of

Transformation coefficients as well as coders and decoders

State of the art

Integer approximations of the DCT (Discrete Cosine Transform) are used in the frequency domain coding of

Moving image sequences, in which pure integer arithmetic is required, application. Integer arithmetic eliminates the divergence between the encoder and decoder due to different computational accuracies.

An integer-cosine transformation of size 4x4 is proposed for the test model of the video coding standard H.26. In [1] a concept is proposed that uses block sizes coupled to the motion compensation for the transformation. In addition, integer

Cosine transformations of sizes 8x8 and 16x16 are necessary, whereby the coefficients can only be selected within a limited range due to the limited range of numbers available, e.g. 32 bits. Orthogonal integer-cosine transformations of size 8x8 are e.g. published in [2].

There are different approaches for approximating the DCT using an integer-cosine transformation. In [3] and [4] shows a method in which the coefficients of different amounts of the DCT are replaced by integer values.

[5] uses a method to approximate DCT using the Hadamard transform. The DCT can be generated by multiplying the Hadamard transformation by an orthogonal conversion matrix. By replacing the real conversion matrix with an orthogonal matrix with integer coefficients, an integer approximation of the DCT is generated.

Advantages of the invention

The method according to the steps of claim 1 and the subclaims provides integer numbers

Transformation matrices, for example for the block sizes 8x8 and 16x16, which approximately have the properties of the DCT for the frequency domain coding of ^' pixel blocks for moving picture sequences and thus enable an increase in coding efficiency.

The coding can be done with the measures of the invention in pure integer arithmetic. Through the choice of the transformation coefficients according to the invention, a maximum of different possible integer coefficients can be used in a limited number range, preferably 32 bits, which represent a good approximation to the non-integer DCT coefficients. The integer coefficients of the transformation are chosen with a view to the quantization and normalization of the transformed coefficients with the above number range less than 2 ³ . Due to the integer arithmetic, these can be implemented using table values. It is particularly advantageous in the invention that the coefficients of all base vectors have the same standard, which means that a uniform standardization and quantization factor can be used in the quantization and normalization. In contrast, in the methods according to [3] and [4], the even and odd basic functions of the specified integer transformations have a • different norm, so that these matrices do not meet the requirement for orthogonality. There are none in [5]

Solutions given that lead to orthogonal integer approximations of the DCT e.g. the big 16x16 lead.

Description of exemplary embodiments

Before the actual invention is explained, some prerequisites are explained beforehand.

An orthogonal, integer approximation of the NxN-DCT must meet the conditions

T rpT π -k Nrv ^■ 'Σ iN _; ; <D

satisfy, where I _N denotes the unit matrix of the size NxN and k _{N is} an integer constant.

The orthogonal integer transformation according to the invention of size 8x8 has within the number range of Coefficients of 2 ⁷ the smallest maximum approximation error

-ma-z msLx {e (t _Q ) ₁ {tι) ₎ . , , , φγ)} (3:

With

where t _D cτ, _m and tχ _CT , _m with m = 0, ...., 7 denote the coefficients of the DCT and ICT (Integer Cosine Transform). So far, there are no indications in the literature of integer approximations of the DCT of the block size 16x16. The integer transformation according to the invention fulfills these conditions, so that, in contrast to the solutions mentioned in [4], all base vectors have the same norm. This means in particular an advantage in the quantization and normalization of the transformed coefficients, since a uniform standardization and quantization factor can be used for all coefficients.

The intεger approximation T _{8 of} the size 8x8 is generated by replacing the 7 different real amounts of the coefficients of the 8x8-DCT with 7 integer values taking into account the orthogonality conditions ^■ (!) And (2) for N = 8, IGT, 8 =

According to [5] the DCT can be generated with the help of a conversion matrix of size 16x16 from the Hadamard transformation H _BR0 of size 16x16. BRO stands for bit-reversed order of the base vectors of the Hadamard transformation known, for example, from [6]. The real conversion matrix is replaced by an integer orthogonal conversion matrix C ₁₆ :, so that

T 16.BRO = C 1lK6 ' ^■ Ξ- ⁰ -'BR.O- applies

C _ιe is a sparse matrix with a block diagonal structure,

C ₂ denotes a single, integer coefficient, C _n is an integer orthogonal nxn matrix. The construction instruction follows for the sub-matrices of the matrix C _n

2n - (7:

J _n is the mirrored unit matrix, e.g.

And Bι, _n and B ₂ , _n describe integer nxn matrices whose coefficients approximate the property of the real conversion matrix as closely as possible.

For Ci = 17 different integer orthogonal approximations of the conversion matrix can be created and thus orthogonal integer-cosine transformations of size 16x16 with k _ιe = 16. 17 ² determine. The coefficients of the constant base vectors - constant component - this integer (integer) - cosine transformation (ICT) each have the value 17. The coefficients of the submatrices for the alternating components can be derived from this value 17 by squaring the value 17 Sum of the squared coefficient values of each row of the sub-matrices results. The values for the coefficients of the lowest alternating component are 15 and 8, since 17 ² = 15 ² + 8 ² . The values of the coefficients for the second lowest alternating component result accordingly for example to 12, 9, and 8 and for the third lowest Share of changes, for example, to 13, 10, 4 and 2. For illustration, one of the possible solutions for the matrix is below

C ₁₆ and thus also given for T _IC τ, i6. The matrices C _n can

to get voted. After the row vectors of T ₁₆ , _{BR0 have been rearranged accordingly} , the orthogonal integer transformation matrix then results

The cell vectors are continued symmetrically according to the DCT, alternating between even and odd.

Another possible orthogonal 16 x 16 integer transformation matrix has the following form:

17 17 17 17 17. 17 17 17 17 17 17 17 17 17 17 17

21 21 21 21 15 7 15 7 - -7 - -15 -7 • -15 -21 -21 -21 -21

24 20 12 β ^' - -6 - ^• 12- ^■ 20 - -24- -24- -20 - -12 -6 6 12 20 24

15 7 15 7 - -21- -21- -21 -21 23 21 21 21 -7 -15 -7 -15

23 7 - -7 -23 -23 -7 7 23 23 7 - -7 - -23 -23 -7 7 23

21 21 - -21 ^■ -21 -15 -7 15 7 -7 -15 7 15 21 21 - -21 -21

20 -6 -24 -12 12 ^'6 24 - 6 -20 -20 24 12 - -12 - 24 -6 20 ^■

15 7 - -15 ^• -7 21 21 - -21 -21 21 21- -21 -21 7 15 -7 -15

1 16 17- -17- -17 17 17 -17 -17 17 17 -17 -17 17 17 -17 -17 17

21 '-21 -21 21 7 -15 -7 15 -15 7 15 -7 -21 21 21 -21

12- -24 6 20- -20 -6 24 ^• -12 -12 24 ^■ -6 - -20 20 6 - -24 12

7 - -15 -7 15 - -21 21 21 ^■ -21 21 -21 -21 21 -15 7 15 -7

7 - -23 23 -7 -7 23- -23 7 7 -23 23 -7 -7 23 -2- 5 7

21 -21 21 -21 -7 15 -7 15 -15 7 -15 7 21 -21 21 -21

6 - -12 ^■ 20- -24 24 - -20 12 - -6 -6 12 -20 24- -24 20 - -12 6

7 ^• -15 7 -15 21 ^■ -21 21 -21 21 -21 21 -21. 15 -7 15 -7

It is constructed as follows:

the even lines 0: 2: 14 correspond to the even lines of the TJCT g already shown above, the continuation of the line length 8 to 16 taking place by mirroring the coefficients. The odd lines 1: 2: 15 can be obtained as follows:

Ordered Hadamard matrix:

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1-1-1 1 1 -1 -1 1 1 ^■ -1 1 1 -1 -1

I 1 -1-1-1-1 1 1 1 -1 -1 -1-1 1 1-1 1-1-1 1 -1 1 1 ^• 1 -1-1 1 -1

II 1 1 -1-1 -1-1-1 ■ -1 -1 1 1 1 1-1-1 1 -1 1 1 -1-1 1 -1 1 -1 -1 1 1 -1-1 1 1 -1-1-1 1 1 -1-1 1 1-1 1-1 1 -1 1 -1-1 -1 1 -1 1 -1

H] _t 1 1 1 1 1 1 1 1 -1 -1 -1 -1-1-1 1-1-1 1 1 -1 -1 1 -1 1 -1 -1 1 1 1 1-1-1 -1-1 1 1 -1 1 1 1 1 -1 1-1 1-1-1 1 -1 1 -1 -1 1 1 -1 1 1 1 1 1 -1-1 -1-1 1 1 1 -1 -1 -1 1-1-1 1 -1 1 1 -1 1 -1 1-1 1 1

1 1—1—1 1 1-1 -1 1 -1 -1 1 1 -1 1-1 1-1 1 -1 1 -1 1 1 -1 1 -1 1

Construction of the orthogonal conversion matrix A:

m Matlab notation:

A = [zeros (i serous (8); serous (8), C _R ]

Tn A H16

The values of the coefficients in the lines of the C8 atrix: 2, 2, 5 and 16 each result in squared and, in turn, add up the square of the equal part 17.

The odd basic functions of T16 result in (again Matlab notation):

T16 (2: 2: 16, :) = Tn ([9 13 11 15 10 14 12 16],:) -

That means the 9th line of Tn comes in the 2nd line of T16, the -13. Row from Tn to the 4th of T16, and so on.

The advantage of this matrix T ^ g is besides a higher one

The construction is transformed using the previously mentioned highly symmetrical matrices. This enables an efficient implementation in a coder or a corresponding decoder for the transformation coding of pixel information within blocks.

The transformation gain of an orthogonal transformation with N base vectors is given by the ratio of the variance of the input signal after quantization to the variance of the signal transformed and then quantized with the transformation matrix T _N. This transformation gain is usually given in dB. Work in literature real image signals are often modeled by a first-order autoregressive process. This model is fully described by the signal variance and the correlation coefficient. Assuming optimal quantization and coding, the transformation gain for this signal model can be specified directly. For the correlation coefficient of 0.95, which is typical for image signals, there is a transformation gain of 9.46 dB for the DCT, 8.17 dB for the 16x16 matrix presented in the first example and 8.86 dB for the latter 16x16.

With the method of the invention, encoders and decoders for frequency domain coding / decoding of moving picture sequences can be constructed. Your

For this purpose, transformation devices must be such that transformation coefficients are processed in accordance with the method steps presented, or the original moving image sequence can be reconstructed from these coefficients.

Fast algorithms can be developed for the ICT matrices of sizes 8x8 and 16x16, which minimize the number of additions and multiplications required.

An example of such an algorithm is shown in FIG. 1, which schematically represents the two transformations in a flow graph. The ICT matrices of size 8x8 and 16x16 are marked in Figure 1 by boxes with the designations T8 and T16. Since the straight

Basic functions of the T16 correspond exactly to the T8, the T8 block is fully integrated in the T18 block. The coefficients of the input signal of length 16 are designated xO, xl, x2, ..., xF and are on the left Side of the graph. The coefficients of the output signal are Oy, ly, 2y, ..., yF.

Nodes in the graph represent additions. Multiplications with constants are indicated by corresponding numbers on the

Called edges. A minus sign on one edge means

Subtraction instead of addition. in order to keep the presentation clearer, the

The graph is separated at the nodes marked by boxes and part of the further course is shown next to the main graph. The corresponding nodes have the same names.

The number of additions and multiplications required for the 8x8 transformation and the 16x16 transformation is shown in the table below. For comparison, the table contains information on the number of additions and multiplications for the fast discrete cosine transformation according to [7].

The flow graph for the rapid implementation of the T16 and T8 according to FIG. 1 contains some elements that occur several times. This regularity reflects the symmetries within the transformation. The graph is divided into 4 levels, which are characterized by a vertical plane of nodes. These are referred to below as levels 1 to 4.

In the first stage, two input coefficients are added and subtracted, so that there are again 16 coefficients. The sums of xO and xF, xl and x: E, x2 and xD, etc., form the input coefficients of the T8 block, which represents the even basic functions. The differences are at nodes 0-7 shown in the figure. The odd proportions result from them. The star structure resulting from these additions and

Subtractions results, repeated with 8 instead of 16 coefficients at the input of the block labeled T8 at level 2; then with the 4 upper coefficients (even basic functions of T8) at level 3, and with two coefficients before the outputs yO and yF at level 4. This structure is equivalent to fast implementations of the discrete cosine transformation.

Stars that do not represent pure additions and subtractions, but contain weighting coefficients also occur several times. Thus, from the nodes 3-4, 2-5, 1-6 and 0-7 designated in FIG. 1, there are two weighted stars with 4 coefficients, which differ only in the arrangement of the weights. This structure is repeated before the outputs y4 and yC with different weights and two coefficients.

Variations in this weighted structure result for the odd basic functions of the T16. At level 3 ^• there are four structures, each with two outputs with weighted coefficients and two outputs with pure additions / summations. These differ in the distribution of the weights. With three of these structures, only two nodes form the entrance. So distorted stars arise.

literature

[1] Telecom. Standardization Sector of ITU, "New integer transforms for H.26L", in Study Group 16, Question 15, Meeting J, (Osaka, Japan), ITU, May 2000. [2] Telecom. Standardization Sector of ITU, "Addition of 8x8 transform to H.26L", in Study Group 16, Question 15, Meeting I, (Red Bank, New Jersey), ITU, Oct. 1999th

[3] W. Cham, "Development of integer cosine transforms by the principle of dyadic symmetry", IEE Proc, vol. 136, pp. 276-282, Aug. 1999.

[4]. Cham and Y. Chan, "An order-16 integer cosine transform," IEEE Trans. Signal Processing, vol.39, pp.1205-1208, May 1991.

[5] R. Srinivasan and K. Rao, "An approximation to the discrete cosine transform for n = 16", Signal Processing 5, pp. 81-85, 1983.

[6] A.K. Jain, Fundamentals of digital i age processing. Englewood Cliffs, NJ: Prentice Hall, 1989.

[7] W.H. Chen, C.H. Smith, and S.C. Fralick, "A Fast Computatial Algorithm for the Discrete Cosine Transform", IEEE Trans. Comm., Vol. COM-25, No. 9, Sep. 1977, pp. 1004-1009,

Claims

Expectations

1. Method for obtaining integer approximated cosine transformation coefficients, in particular for the coding of pixel blocks, the transformation coefficients being selected as follows:

- a limited range of values is specified for the transformation coefficients,

- The coefficients of the base vectors for the sub-matrices are taken into account

Orthogonality condition chosen so that the sum of their squares gives the square of the DC component,

- The coefficients of the alternating components are derived from these coefficients.

2. The method according to claim 1, characterized in that a uniform normalization and / or quantization factor is used for all coefficients.

Method according to claim 1 or 2, characterized in that block sizes coupled to the motion compensation are used for the transformation.

4. The method according to any one of claims 1 to 3, characterized in that the integer-cosine transformation matrices with the aid of a conversion matrix from the Hadamard transformation of the same matrix size, e.g. 16x16.

5. The method according to any one of claims 1 to 4, characterized in that the value 17 is selected for the coefficients of the base vectors for the DC component.

6. The method according to claim 4 and 5, characterized in that the coefficients of the base vectors for the lowest alternating component are chosen to be 15 and 8 in terms of amount.

7. The method according to claim 4 and 5 or 6, characterized in that the coefficients of the base vectors for the second lowest alternating component in terms of amount

12, 9 and 8 can be selected.

8. The method according to claim 4 and 5 or 6 or 7, characterized in that the coefficients of the base vectors for the third lowest alternating component in terms of amount

13, 10, 4 and 2 can be selected

9. The method according to claim 4 and 5 or 6 or 7, characterized in that the coefficients of the base vectors for the third lowest alternating component in terms of amount

16, 5, 2 and 2 can be selected.

10. Encoder for frequency domain coding of moving picture sequences with a Transformation device, which is designed to create transformation coefficients for a moving image sequence, which were prepared with the method steps according to claims 1 to 9.

11. Decoder for frequency range decoding of moving picture sequences with a

Transformation device, which is designed to reconstruct a moving image sequence from the transformation coefficients that were created with the method steps according to claims 1 to 9.

12. The method according to any one of claims 1 to 9, characterized in that essentially symmetrical algorithms are used for the approximated transformations, the input-side transformation coefficients being led step by step to an addition or subraction node and weighted accordingly for necessary multiplications.