WO2007135626A2

WO2007135626A2 - Method of video coding

Info

Publication number: WO2007135626A2
Application number: PCT/IB2007/051856
Authority: WO
Inventors: Detlef Goetting
Original assignee: Nxp B.V.
Priority date: 2006-05-18
Filing date: 2007-05-15
Publication date: 2007-11-29
Also published as: WO2007135626A3

Abstract

The invention relates to a method of video coding in which a signal is quantized, there being a method of rounding that operates with an adaptive dynamic dividing up of the quantization interval, the dividing-up being effected by means of a parameter M.

Description

Method of video coding

The invention relates to a method of video coding.

Such methods are known from, for example, US 2004/0234150 Al. In a method of video coding of this kind, use is of made of a quantization matrix. Underflow of a video buffer verifier is advantageously avoided in this way. US 2006/0034522 Al discloses a method for equalizing video quality using selective decoding.

Overall, the prior art methods are very complicated and costly and are not greatly optimized as far as the requirements of specific signals are concerned. It is therefore an object of the invention to provide a method of video coding that can be applied in a more specific way to the specific characteristics or requirements of individual signals or variables.

The object of the invention is achieved by a method of video coding in which a signal is quantized, there being a rounding method that operates with an adaptive dynamic dividing-up of the quantization interval, the dividing-up being effected by means of a parameter M.

It is useful if the standard rounding method is performed in the case where M = N, where N is the number of steps that are possible in the dividing-up. In this way, an adaptive method can advantageously be converted into a standard non-adaptive method, or vice versa, where required in a specific case. It is particularly advantageous if the adaptive rounding method is performed in bit-rate control.

It is also advantageous if the inverse quantization is performed with a standard method of rounding. Drifts or changes between the coding and decoding can be avoided in this way. It is particularly advantageous if the adaptive method of rounding is performed with different divisions of the quantization interval for the luminance component (Y component) and the chrominance component(s) (Cb and/or Cr component). This is advantageously possible because an additional disruption in the chrominance is less markedly perceptible than in the luminance. Advantageous embodiments are described in the dependent claims. These and other aspects of the invention are apparent from and will be elucidated with reference to the embodiments described hereinafter.

In the drawings:

Fig. 1 shows a schematic block circuit diagram relating to the video coding. Fig. 2 is a graph to illustrate the forward quantization. Fig. 3 is a graph to illustrate the forward quantization. Fig. 4 is a graph to illustrate the run length coding.

Fig. 5 is a graph to illustrate the forward quantization, and Fig. 6 is a graph to illustrate the forward quantization.

Fig. 1 shows a block circuit diagram 1 for a method of video coding. An input signal, i.e. input data 2 to be coded is converted by means of individual blocks into a coded data stream forming an output signal 3.

As well as lossless spatial and temporal prediction, all the relevant methods of video coding, such as MPEG- 1/2/4, H.261/3/4 and VC-l/WMV, also support a lossy process that is referred to as quantization. An encoder supports a forward quantization process 11 and an inverse quantization process 12 in this case, as can be seen from Fig. 1. What are used as an input to the forward quantization 11 are the temporally and/or spatially decorrelated transformed coefficients C. These coefficients are processed by the forward quantization process into transformed coefficients C_Q that serve as input variables to the inverse quantization 12. The inverse quantization processes the reconstructed coefficient C_Q into the reconstructed coefficient C_R.

Even though the particular implementations may differ for the various video standards, the general, fundamental, mechanism is nevertheless comparable. The equations that apply are

C_Q = Round ( C/Q ) (1)

for the scalar quantization operation and

C_Qx,_y = Round ( C_x,_y/Q_x,_y ) (2) for the matrix-based representation of the quantization, these two equations being considered equivalent to one another for a particular coefficient.

Fig. 2 shows a graph 13 to illustrate the forward quantization, in which graph the range index C_Q of the transformed coefficient is shown on the y axis, plotted against the coefficient C on the x axis. In the forward quantization, the continuous numeral interval covered by the coefficient C is, in essence, converted, by dividing (by Q) and rounding, into separate, discrete points C_Q (also referred to as indices). An index C_Q always represents a sub-interval of C in this case. After the inverse quantization, the quantized value, or what is also called the index C_Q, has been mapped, by multiplication by Q, onto only one value, the reconstructed coefficient C_R.

This is shown by way of example in Fig. 2 as follows: the interval 2 (see 17b) is mapped onto the index 2 (C_Q) (see 18b) on the y axis in the forward quantization and in the inverse quantization is mapped in turn onto the value C_R, and thus represents the interval 3 (see 16).

In the quantizing method, the exact mapping of the coefficient C onto an index C_Q is determined by the rounding rules. The standard rounding rule that is used in the prior art methods is "rounding to the nearest neighbor". In this case a mapping error (C -> C_Q -> C_R) is minimized in such a way that the result of the forward quantization is always rounded up or down to the nearest whole number. In Fig. 2, the interval 1 for example (see 17a) is mapped downwards to the index 18a that is likewise identified as 1 and the interval 2 (see 17b) is mapped upwards to the index 18b that is identified as 2. As shown in Fig. 2, what this produces is two dividing-up intervals 1 and 2 of equal sizes.

The characteristic 14 is shown as a straight line in this case and the step function 15 represents the quantization. The step function has a step width Q and a step height of 1. The forward quantization separates the entire range of data covered by the coefficient C into separate, individual range sets or intervals 16, which means that only the range index C_Q of the transformed coefficient is coded and after the inverse quantization the entire set is represented by one of its elements. This element is also called the reconstruction point C_R.

The exact definition of the coefficient C for an index C_Q is determined by the application of a rounding method, using equation 1.

The assignment of the coefficients C to the indices C_Q and the rounding errors that arise in the process are indicated in Fig. 3 by the triangles marked 3 and 4 that carry the reference numerals 20, 21, 22 and 23. Whereas prior art methods still employ the "rounding to the nearest neighbor" rounding rule and regulate the quantization by varying the divider Q (also known as the quantizer), the very point to which the present invention applies is, precisely, the rounding rule that has been described. One of the disadvantages of the prior art methods is also the fact that a change in the divider Q has to be indicated in the coded data stream and that this is permitted only at certain points, which depend on the video coding standard that is being used.

If equation 1 is converted as follows,

C_Q = Roun

what is obtained as equation 3 is the starting point for the adaptive method of rounding employing the two parameters M and N.

As shown in Fig. 1, the quantized coefficients or indices C_Q are fed not only to the inverse quantization process at 12 in Fig. 1, but also to a lossless entropy coding process at 4 in Fig. 1. A particular aspect of the entropy coding 4 is, in the present case, the variable run length encoding (RLE) 5 that precedes it. By following a special path (called a zigzag scan), this run length encoding 5 converts the sequenced indices C_QX,_Y into the symbol

Sign^Run^Level^ (see for example Fig. 4).

In this symbol, the value Sig^ denotes the plus or minus sign and the value

LeVeI₁ denotes the amplitude of the coefficient C_Q1 = C_QX,_Y which is other than zero. The value RUn₁ gives the number of values, preceding the coefficient C_Q1, whose amplitude was zero.

This also means that coefficients of value zero are not themselves represented by a symbol. In the entropy coding 4 (such as Huffman coding for example), the symbols are typically mapped onto a codeword of variable length, in accordance with the probability of their occurrence (a process known as variable length coding VLC). It should be noted in this case that the mapping of the LeVeI₁ value increases monotonically. This means that, in the entropy coding, a higher value, Level, > Level_1? always generates a codeword of equal, or else greater, length than LeVeI₁. The length of the codeword that is produced in this way, measured in bits, is also referred to as the cost of the non-zero index C_Q when mapped onto Sig^ and LeVeI₁. A method of regulation in video coding, also referred to as bit-rate control, always attempts to find a compromise between the mapping error and the coding cost that is incurred.

If, with this mind, use is made of equation 3 and the parameters M and N that have been introduced are generalized, the result is a new method of control, which is referred to here as "adaptive rounding" and which can be advantageously used for efficient video coding.

The parameters M and N that have been introduced can be interpreted as follows in this case. The parameter N in equation 4 defines the granularity that is possible for the controlling parameter M and will be used here to define the numbers of escalation steps that are possible. The parameter M, on the other hand, defines the adaptive choice and the application of a given "escalation step", with M advantageously covering the range from 0 to N ( M = { 0, ... , N } ). It is useful for standard-compliant quantization if M is set as equal to the value N, because this converts equation 4 back into equation 3.

Dynamic variation of the parameter M from N (for "rounding to the nearest neighbor") to 0 (for "rounding down") allows the bit-rate to be regulated at no additional cost in terms of signaling and without any other standard- specific restriction. The method described is shown by way of example in Fig. 5. As is shown by way of example in Fig. 5, what takes place when the parameter

M varies from 0 to N is a weighted dividing-up of the quantization interval. Whereas two unequal intervals 1 and 2 arise when M is not equal to N (see Fig. 5), when M is equal to N there are, in turn, two intervals of equal sizes. It is precisely the independence from particular requirements preset by the coding standard that allows a plurality of advantageous applications. It is particularly advantageous if the adaptive rounding method is performed with different values of M for the luminance data (Y component) and chrominance data (Cb and Cr components) of a video picture. It is useful in this case, if required, for M to be taken as being of a lower value for the quantization of the chrominance coefficients than for the quantization of the luminance coefficients. This is advantageously possible because an additional error, such as is indicated at (4) in Fig. 6, is less markedly perceptible to a human viewer in the chrominance coefficient or even in the color difference coefficient (Cb and Cr).

The additional rounding error compared with the standard method of rounding (see the shape marked 4 in Fig. 6), is always small when M is selected to be close to N. It is also useful if the adaptive method of rounding is used in bit-rate control to achieve a uniform distribution of quality without any standard-specific restrictions at the level of the individual coefficients. The differing significance that the coefficients C_Q1 have for the human perceptive faculty can be exploited in this case in that coefficients in higher frequency bands are, if required, processed with a lower parameter M than the coefficients in lower frequency bands that are significant for human perception.

The inverse quantization in accordance with equation 5 is not intended to be affected by the method described and takes place unaltered in accordance with the relevant standard

C_x = C₀ - Q (5).

LIST OF REFERENCE NUMERALS

1 Block circuit diagram

2 Input data

3 Output data

4 Entropy coding

5 Run length encoding

10 Block

11 Forward quantization

12 Inverse quantization

13 Graph

14 Characteristic

15 Step function

16 Interval

17a Interval

17b Interval

18a Index

18b Index

20 Rounding error indicated by triangle

21 Rounding error indicated by triangle

22 Rounding error indicated by triangle

23 Rounding error indicated by triangle

Claims

1. A method of video coding in which a signal is quantized, characterized in that a method of rounding operates with an adaptive dynamic dividing-up of the quantization interval, the dividing-up being effected by means of a parameter M.

2. A method as claimed in Claim 1, characterized in that the standard method of rounding is performed in the case where M = N, where N is the number of steps that are possible in the dividing-up.

3. A method as claimed in Claim 1 or 2, characterized in that the adaptive method of rounding is performed in bit-rate control.

4. A method as claimed in Claims 1 to 3, characterized in that the inverse quantization is carried out with a standard method of rounding.

5. A method as claimed in any one of the preceding claims, characterized in that the adaptive method of rounding is performed with different divisions of the quantization interval for the luminance component and the chrominance component(s).