WO2015090682A1

WO2015090682A1 - Method of estimating a coding bitrate of an image of a sequence of images, method of coding, device and computer program corresponding thereto

Info

Publication number: WO2015090682A1
Application number: PCT/EP2014/072604
Authority: WO
Inventors: Thibaud BIATEK; Mickaël RAULET; Olivier Deforges
Original assignee: Tdf
Priority date: 2013-12-17
Filing date: 2014-10-22
Publication date: 2015-06-25
Also published as: FR3015163B1; FR3015163A1; EP3085087A1

Abstract

The invention relates to a method of estimating a coding bitrate of an image (100) of a sequence of images implementing a bitrate model according to a so-called ρ-domain approach such that R(ρ) = θ(1 - ρ), where: - R denotes said bitrate, - ρ denotes a percentage of zero coefficients in at least one coding residual for said image, after transformation (107a) and quantification (107b) of said residual, said residual being obtained by differencing between a block of said image and a prediction of said block, - θ denotes a slope parameter of said bitrate model. According to the invention, such a method of estimating comprises a step of calculating said slope parameter θ taking account of a surface area of said coding residual(s) coded for said image.

Description

A method of estimating an encoding rate of an image of an image sequence, encoding method, device and corresponding computer program.

1. Field of the invention

The field of the invention is that of the coding of image sequences and the compression of video sequences, in particular, but not exclusively, in the context of H EVC / H.265 standards (for "High Efficiency Video Coding") and SHVC (for Scalable-High Efficiency Video Coding) established by standardization bodies such as the ITU-T Video Coding Expert Group (VCEG) and the ISO / IEC Motion Picture Expert Group (MPEG).

More precisely, the invention relates to the estimation and the regulation of the data rate at the output of a video encoder, implemented in the context of such a compression.

2. Prior art and its disadvantages

The compression of a video sequence consists in reducing the amount of data necessary to represent it, while seeking to minimize the phenomena of loss of information, and therefore of quality. During such a compression, it is necessary to adapt the data rate at the output of the encoding system, to take account, for example, of the constraints of the network on which they must be transmitted, or the size of the data carrier on which they must be stored. This rate adaptation must also take into account the quality of reproduction of the video sequence that one wishes to achieve for the decoded video stream.

To do this, flow control strategies, aimed at modulating the amount of information at the output of the encoder to obtain an optimal bit rate and quality, are conventionally implemented. They allow to control certain parameters of the encoding to obtain a flow which respects a set of imposed constraints. One of the main challenges of such a flow control is therefore to find a model connecting the parameters of the encoding to the data rate at the output of the encoder.

More particularly, in the known predictive coding techniques, such flow control strategies most often act on the parameter of quantification (or QP, for "quantization parameter"), implemented during a step of quantifying the residuals of the prediction.

It is recalled that the video coding, as proposed for example in the standard MPEG-4 AVC (for "Advanced Video Coding") or in the standard HEVC, implements the following steps:

The images are first divided into blocks of variable size, depending on the complexity of the movement to be represented;

These blocks are then predicted with respect to each other, by spatial and / or temporal prediction. The information relating to the prediction modes used and the motion vectors are retained to be transmitted in the stream, or bitstream, final;

The residues are then formed, by difference between the original blocks and their prediction: Block _Residue = Block _0riginal - Block _Predicts ;

The residual signal of the prediction then undergoes a set of treatments aimed at removing as many redundancies as possible, namely a transformation, to make it pass in the frequency domain, followed by a quantization;

The final stream, or bitstream, is then formed respecting the normative structure, an entropy encoding for compressing the information in the bitstream.

It is therefore by acting on the value of the quantization parameter (QP), used during the quantization step of the prediction residuals, that it is possible to vary the amount of information lost during coding, and therefore to vary the flow and quality obtained at the output of the encoder.

It is therefore particularly interesting to identify a model linking the quantification parameter QP and the flow rate.

Different models of the relationship between QP and flow have been studied in the literature. Of these, Z. He et al., In "Optimum Bit Allocation and Accurate Rate Control for Video Coding via p-Domain Source Modeling," IEEE Transactions on Circuits and Systems for Video Technology, Vol. 12, No. 10, Oct. 2002, an interesting approach for the regulation of flow both from the point of view of the precision of its representation as of its low complexity in algorithmic terms. This approach, called p-domain, links a parameter p, defined as the percentage of zero coefficients in a macroblock, or in an image, after transformation and quantization, with the flow R.

According to this p-domain approach, the relation between the parameter p and the flow rate has a very linear aspect, and can be approximated by the formula:

R (p) = 6 (1 - p).

The line represented by this equation passes through the point (1; 0). Indeed, when p is 1, all the coefficients of the macroblock or of the image are zero, and the rate R is equal to 0. To characterize this relation, it is therefore sufficient to determine the slope Θ of the line.

To date, the known techniques providing an estimate of this slope Θ all rely on the implementation of measures to better approach the value of Θ, which varies from one image to another, depending on its content.

In "Encoding Strategies for Scalable Video Encoders" (Thesis), INSA Rennes, 2009, Y. Pitrey proposes two approaches, in the context of the MPEG-4-AVC and SVC standards, to determine the value of the slope Θ.

According to a first approach, each image is encoded twice: the first pass is used to initialize the R-p rate model, by establishing a slope parameter Θ; the optimal value of the quantization parameter QP is then chosen thanks to the flow model; a second encoding pass permanently encodes the image using the QP quantization parameter chosen to achieve the target rate.

A disadvantage of this first two-pass encoding approach is that the pre-encoding phase necessary to initialize the bitrate model induces a sharp increase in encoding time: it therefore strongly penalizes the speed of the mechanism, and introduces a complexity in terms of calculations.

According to a second approach, which exploits the correlations between successive images, the encoding is performed in a single pass, and the flow control on the current image is based on the information obtained after the encoding of the images. preceding. More specifically, for the rate regulation of a current image, the slope Θ of the rate model Rp is initialized to a value calculated on a previous image chosen as a reference for the rate model, namely the image which has the behavior closest to that of the image being controlled. This value of Θ is then corrected, as the encoding progresses, thanks to the feedback of the values of Θ measured on the previously coded images.

If this second approach has the advantage of inducing a reduced encoding time compared to the first approach above, thanks to the removal of the first encoding pass, and a very low cost in terms of complexity calculations, it induces a decrease in the accuracy of flow control. Indeed, the changes of content between successive images, even minimal, introduce a bias in the calculation of the optimal quantization parameter. In addition, if there is a significant change in activity between successive images (large variations in movement and detail), the results obtained according to this second approach are less efficient than in the first approach.

There is therefore a need for a flow control technique based on a p-domain approach that does not have these disadvantages of the prior art.

In particular, there is a need for such a technique that makes it possible to estimate the value of the slope Θ of the R-p flow model accurately and deterministically.

There is also a need for such a technique that allows, compared to known methods, to perform the encoding of images in a single encoding pass, while allowing a better accuracy in the flow control.

More generally, there is a need for a precise and deterministic estimation technique of the coding complexity of an image to be encoded, based on such an approach called p-domain.

3. Presentation of the invention

The invention responds to this need by proposing a method for estimating a coding bit rate of an image of a sequence of images implementing a bit rate model according to a so-called p-domain approach such that R (p ) = θ (1 - p), where: R denotes the flow,

p denotes a percentage of zero coefficients in at least one coding residue of the image, after transformation and quantization of the residue, the residue being obtained by difference between a block of the image and a prediction of the block, Θ denotes a parameter of slope of said flow model.

According to the invention, such an estimation method comprises a step of calculating the slope parameter Θ taking into account a multiple of a surface of the coded coding residue (s) for the image, the surface being the sum of the resolutions of each of the components of the image.

Thus, the invention is based on a completely new and inventive approach to the estimation of the coding complexity of an image to be encoded, which may notably but not exclusively be applied to the regulation of output flow of a video encoder, for example according to H EVC or SHVC standard. Indeed, while the known flow estimation mechanisms based on a so-called p-domain approach were all based on an empirical evaluation of the slope parameter Θ of the rate-p model R (p) = θ (1-p), putting As a set of measurements, the rate estimation method of the invention is based on a precise and deterministic calculation of this slope parameter Θ.

Indeed, as explained in more detail later in this document, the inventors of the present patent application have established the demonstration, in the context of HEVC type video coding, of a direct link between the slope parameter. Θ and the surface of a coding residue, or more generally, the image to be encoded.

In particular, such a link between the slope parameter Θ and the surface of a coding residue, or more generally, the image to be encoded may be a proportionality link. In this case, the slope parameter Θ is calculated as a multiple of this area.

It will be noted that, by surface area of the image, here and in the entirety of this document is meant the sum of the resolutions of each of the components of the image in the chosen decomposition space. So, in the YUV space for example, the surface of the image corresponds to the sum of the resolutions of each of the chrominance and luminance components of the image.

Such a rate estimation method, implemented in an H EVC / H.265 or SHVC type coder, advantageously makes it possible to achieve optimal compression as soon as the first coding pass, and to substantially reduce the algorithmic complexity of such an encoder.

According to a first characteristic of the invention, p denoting a percentage of zero coefficients in a residue of size NxN of the image after transformation and quantization of the residue, the slope parameter Θ is determined, during the calculation step, under the form 0 = 4 * N ² .

Thus, the inventors of the present patent application have established a formulation of the Rp model at a coding residue of the image, in which the slope parameter Θ depends on the size of the residue. In the case of video coding of HEVC or SHVC type, in which the residues are macroblocks of NxN pixels where N is 4, 8, 16 or 32, the slope parameter Θ can thus be very simply calculated in the form Θ = Α * Ν ² . The flow estimation method of the invention thus offers a precise and deterministic formulation of the flow rate model Rp, at the level of a residue, and thus allows a simplification of the algorithmic complexity with respect to the flow control methods of the invention. prior art.

According to another characteristic of the invention, such an estimation method further comprises a step of estimating the coding rate at the residue such that: R (p) = 4 * N ² * (1- p) + Δ, where Δ denotes an adjustment parameter dependent on a distribution of coefficients in the residue.

It is thus possible to refine the R-p rate model, to obtain a more accurate estimate of the flow rate, by introducing such an adjustment parameter.

In one embodiment of the invention, the adjustment parameter Δ depends on the statistical distribution of the coefficients and their location in the residue. According to another embodiment of the invention, where p denotes a percentage of zero coefficients in the surface image S after transformation and quantization, the slope parameter Θ is determined, during the calculation step, in the form Θ = 4 * S.

Indeed, under the assumption that all the residues of an image are coded, the relation obtained at the residue level can advantageously be extended to an image, or frame, and allows to calculate simply the slope parameter Θ . Thus, in the context of H EVC or SHVC type video coding, the inventors of the present patent application have established that this slope parameter is expressed as quadruple of the surface of the image.

According to an alternative embodiment of the invention, where p denotes a percentage of zero coefficients in the surface image S after transformation and quantization, the slope parameter Θ is determined, during the calculation step, in the form 0 = 4 * (l- r _skip ) * S, where r _skip denotes a skip rate in the image.

By reference to the H EVC or SHVC standards, the skip _{skip rate} designates here, and in all of this document, the total area covered by the coding units CU (for "Coding Units") whose flag cu_skip_flag is worth Ί ', divided by the total area of the image.

Indeed, the inventors have demonstrated that, during the coding of an image, all the coding residues are not coded: when the prediction is sufficiently good, the residue can be zero. In this case, when coding the image, the associated macroblock may not be compressed: it is said that it is "skipped". The invention therefore advantageously proposes to refine the R-p flow model, by expressing the slope parameter Θ, not from the total surface of the image, but from the unskipped surface of the image. Such a formulation makes it possible to obtain an increased accuracy of the flow model.

According to an additional characteristic of the two aforementioned embodiments, such an estimation method further comprises a step of estimating the coding rate at the level of the image such that: R (p) = Θ * (1-p) + _F , where A _F denotes an adjustment parameter dependent on a distribution of coding residues for the image. Again, such an adjustment parameter further improves the rate estimate. In one embodiment of the invention, the adjustment parameter Δ depends on the statistical distribution of the coefficients and their location in the residue.

The invention also relates to a coding method of an image sequence comprising a step of cutting each of the images into at least one image block, and an encoding phase comprising steps of:

predicting each of the blocks, delivering a predicted block;

for each of the blocks, forming a coding residue by difference between the picture block and the predicted block;

transformation and quantification of coding residues;

the encoding method implementing a mechanism for estimating an encoding rate of the images from a rate model according to a so-called p-domain approach such that R (p) = θ (1-p), where R denotes the rate, p denotes a percentage of zero coefficients in at least one coding residue of the image, after transformation and quantization of the residue, and Θ denotes a slope parameter of the rate model.

According to the invention, such a coding method comprises a step of calculating the slope parameter Θ taking into account a surface of the coded coding residue (s) for the image, the surface being the sum of the resolutions of each component of the image.

According to one aspect of the invention, such a slope parameter Θ is in particular calculated as a multiple of a surface of the coded encoding residue (s) for the image.

According to another aspect of the invention, for each of these blocks, such a coding method implements a single iteration of the encoding phase.

Thus, the video coding method of the invention makes it possible, with respect to prior video coding methods based on a rate regulation according to a so-called p-domain approach, to realize only one encoding pass, the slope parameter Θ can be calculated directly, without having to evaluate it during a first initialization pass of the encoding. Deleting a password encoding, the encoding method of the invention can significantly improve latency latency encoders of the prior art.

In addition, compared to the approaches of the prior art which proposed a flow control already based on a single encoding pass, the method of the invention allows a more accurate estimation of the slope parameter Θ.

The invention also relates to a coding device for an image sequence comprising a module for cutting each of the images into at least one image block, and an encoding module comprising:

a prediction module of each of the blocks, delivering a predicted block;

a module for transforming a coding residue formed by difference between the image block and the predicted block;

a quantization module of the coding residue.

Such a coding device comprises means for regulating a coding rate of the images from a rate model according to a so-called p-domain approach such that R (p) = θ (1-p), where R denotes the rate, p denotes a percentage of zero coefficients in at least one coding residue of the image, after transformation and quantization of the residue, and Θ denotes a slope parameter of the rate model.

According to the invention, the regulation means comprise a module for calculating the slope parameter Θ taking into account a surface of the coded coding residue (s) for the image, the surface being the sum of the resolutions each of the components of the image, and each of the blocks is encoded by a single passage in the encoding module.

More generally, the invention relates to a coding device having in combination all or some of the features set forth throughout this document.

The invention also relates to computer programs comprising instructions for implementing a control method or coding method as described above when the program is executed by a processor. Finally, the invention relates to computer readable recording media on which is recorded a computer program comprising instructions for executing the steps of the control method or coding method as described above.

4. List of figures

Other objects, features and advantages of the invention will appear more clearly on reading the following description, given as a simple illustrative and non-limiting example, in relation to the figures, among which:

FIG. 1 illustrates the architecture of a video coder of the prior art making it possible to deliver a video stream conforming to the H EVC standard;

FIGS. 2A and 2B show a comparison between the R-p model of the invention and simulation results for various residues derived from reference bitstreams;

FIG. 3 illustrates a comparison between the R-p model according to a first embodiment of the invention and simulation results for different class A to E images;

FIG. 4 illustrates a comparison between the R-p model according to a second embodiment of the invention and simulation results for different class A to E images;

FIG. 5 illustrates a first variant embodiment of the encoder of FIG. 1, in which a regulation of flow rate at the residue level is implemented in accordance with one embodiment of the invention;

FIG. 6 illustrates a frame structure in Random Access configuration, with encoding order, in accordance with the H EVC standard;

FIG. 7 illustrates a second variant embodiment of the encoder of FIG. 1, in which frame level flow regulation is implemented in accordance with one embodiment of the invention.

5. Detailed Description of Embodiments of the Invention

The general principle of the invention is based on a precise and deterministic calculation of the slope parameter Θ in a rate model Rp of a so-called p-domain approach. implemented as part of a mechanism for estimating the complexity of coding an image. Such a principle applies in particular advantageously, but not exclusively, to the regulation of the data rate at the output of a video encoder. Indeed, the inventors of the present patent application have established that, in the context of a video coding of type H EVC or SHVC for example, this slope parameter Θ could simply be calculated from a surface of the coded residues of the image, and in particular, in the embodiment described below, as the quadruple of this surface. More generally, the invention makes it possible to estimate the coding complexity of an image to be encoded or of a coding residue, in the context of a predictive video coding.

An embodiment of the invention is now presented, in connection with the figures, in the particular context of the video coding described in the HEVC standard as defined by the ITU-T Video Coding Experts Group and the ISO / IEC Moving Picture. Experts Group. The invention is of course not limited to this type of video coding, and may in particular be extended to scalable, scalable video coding, of the SHVC type, and more generally to any other type of predictive video coding in which the parameter of slope Θ can be directly associated with the size of coding residues of an image.

The HEVC standard has been designed to meet all existing H. 264 / MPEG-4 AVC standard applications, allowing for increased video resolution and increased use of parallel processing architectures.

FIG. 1 illustrates the architecture of a video encoder making it possible to produce a video stream having a structure and syntax in accordance with the H EVC standard, and capable of implementing the rate control method of the invention. Such a video coder is known from the prior art, but its operating principle is recalled to better illustrate the principle of the invention.

For a detailed description of the mode of operation of such a video encoder, reference may be made to the "Overview of the High Efficiency Video Coding (HEVC) Standard" document by Gary J. Sullivan et al., IEEE Transactions on Circuits and Systems for Video Technology, Vol. 22, No. 12, December 2012. In summary, such a video encoder proceeds as follows. Each image 100 is cut into block-shaped regions, called CTU for Coding Tree Units composed of CTB for Coding Tree Blocks, one per color component. The exact division into blocks is transmitted to the decoder. Each CTU is itself subdivided into coding units called CUs (for "Coding Units", composed of CBs for "Coding Blocks", namely one per color component), themselves divided into prediction units called PUs (for " Prediction Units ", composed of PB for" Prediction Blocks ", one per color component) and in transformation units called TUs (for" Transform Units ", composed of TB for" Transform Blocks ", one per colorimetric component).

The first image of a video sequence (and each image used as a random access point to the video sequence) is encoded using the intra prediction mode (ie a spatial prediction from other regions of the same image, but without using other images of the sequence). It is the block referenced 101 which estimates which intra prediction mode is optimal, while the block referenced 102 realizes the prediction from this optimal mode. For all the other images of the sequence, or between the different random access points to the sequence, for most blocks, an inter-image temporal prediction is used. Thus, the block referenced 103 estimates the optimal motion vector or vectors (from the buffer 110 containing the previously encoded images); the block referenced 104 then applies these movements by making motion compensation for the prediction. For both types of intra and inter prediction, the important prediction information is stored and transmitted (via data signals referenced 105 and 106) to be encoded and placed in the bitstream. The encoder and decoder generate identical inter prediction signals, applying motion compensation from the motion vectors, and information about the prediction modes used, which are transmitted in the bitstream.

The residual of the inter or intra prediction (applied to each PB of a PU block) is formed by the difference between the original block and its prediction. The more accurate the prediction, the less energy the residue contains. It is then transformed by a spatial linear transform (applied to each TB of a TU block). The coefficients of the transform are then scaled, quantized (block referenced 107). The thus transformed and quantized residue 108 is retained to be encoded and transmitted in the bitstream. In order for the decoder to reproduce the same reference images used for the temporal prediction, the encoder integrates a quantization and inverse transformation block 109. Block partitioning can create visible artifacts and level differences during the transformation phases. quantization, which may negatively impact the accuracy of the prediction. To overcome this problem, the steps of filtering the reconstructed blocks intended for the buffer of the reference images are implemented. A first pre-analysis step 111 determines the optimal filters to be applied, then these filters are applied in the block referenced 112. The referenced information 113 relating to the choice of filters are retained to be encoded and transmitted in the bitstream. During encoding, a control block 114 manages the CTUs and the prediction modes to be applied. The information relating to this management is also stored in the stream referenced 115 for encoding and insertion in the final bitstream. Finally, the bitstream 116 is formed respecting the normative structure of the H EVC stream. An entropic coder CABAC (Arithmetic Binary Encoding with Context Adaptation) is used to compress all stored data (i.e. data referenced 115, 108, 105, 113 and 106).

The quantization of the coefficients resulting from the transform, implemented in the block referenced 107 in FIG. 1, constitutes a very important step in the encoding process with respect to the problematic of the invention, since it makes it possible to vary the the rate needed to encode the signal by adjusting the amount of data lost.

In practice, it is on the quantization parameter QP (Quantization Parameter) that one acts to vary the amount of information lost. Indeed, the quantization operation leads to the cancellation of the coefficients whose values are not significant in the transformed signal.

One of the main issues of flow control is therefore to find a model linking the parameters of the encoding to the data rate at the output of the video encoder of FIG. 1. The present invention proposes a regulation of flow rate based on a so-called p-domain approach, which is interesting from the point of view of both the precision of its representation and its low complexity in algorithmic terms.

It is recalled that a process of flow regulation generally takes place in four stages:

firstly, the constraints to be applied to the flow are determined according to the context envisaged;

the available resources are then calculated and distributed among the different levels of the flow;

the value of the quantization parameter is chosen in accordance with the resources to be respected;

the data is encoded using the chosen QP and the available resources are updated.

The p-domain rate regulation approach, implemented in the invention, is based on an expression of the relationship between the value of the quantization parameter and the output rate of the encoder.

Starting from the H EVC standard, the inventors of the present patent application have established a formulation of the R-p model, first of all at the level of the residue. This relationship depends on the size NxN of the residue, and an adjustment parameter that depends on the distribution of the coefficients in the residue. We then have:

R = 4 * N ² * (1 - p) + Δ (4) This first relation can be advantageously used in the context of a flow control mechanism at the level of the residue.

The adjustment parameter Δ can be neglected at first approach.

However, for reasons of algorithmic complexity, it is often reasonable to consider flow control, not at the level of the residue, but at the level of an image. The inventors have therefore extended the relation (4) above to an entire frame, considering, according to a first variant embodiment, that all the residues of an image are coded.

Since the video signal is defined by colorimetric components, a frame is conventionally composed of several components, for example the components YUV. If we consider more generally the case of a component C of the image (C can, in a particular case, be one of the components Y, U or V), the relation (4) can then be summed on this component C of the image, denoting by RZ the set of residues in this component C:

^R c = Σ ^R i (5) ieRZ

By replacing in equation (5) the formulation of the flow established in equation (4), we obtain the following relation:

Pi representing the percentage of zero coefficients in the residue i after transformation and quantization, it can be replaced by the number of zero coefficients (Zi) in the residue i divided by the total number of coefficients (N, ² ) in the same residue i . We get then: ieRZ ieRZ ieRZ

By replacing this formula (7) in equation (6), we obtain:

R _c = 4 * (Σ N - z) + A _i (8)

GRZ i £ RZ 'i £ RZ

In this equation (8), we therefore find the expression of the total area covered by the residuals S _c for the component C (Σi _£ Rz N) as well as the total quantity of zero coefficients Z _c for the component C (Σ i _£ Rz Zi), and an overall adjustment coefficient A _FC for the component C (ZigRz ^ i) - We then have the following relation:

R _c = 4 * (5 _C - Z _c ) + A _FC (9) In order to obtain the relation on the total frame, we sum on all the components C of the set of components of the image:

R Frame ⁼ ^ (10)

CGComposantes

If we take the example of a decomposed image in the space of the YUV components, we obtain, by replacing in the equation (10) each of the components Rc = R _Y , Rc = Ru, Rc = Rv, by the expression obtained in equation (9):

+ S _u + S _v ) - (Z _Y + Z _U + Zy)) + (Δ ^ + Δ _ρυ + Α ") (11)

^ Frame = * (S _{To ta} the ~ ^ Totale) + ^ FTotale {^ -) By re-expressing the percentage of null coefficients in the image, after the transformation and quantization steps, p = Z _Total / S _Total we obtain the final relationship below.

^ Frame- * S _Totale * (1- p) + pTotal (13)

This relation depends on the surface of the image S _Tota i _e (namely the sum of the surfaces of each component, more simply designated by S in the remainder of this document), and of a global adjustment parameter which depends on the distribution of all the residues on each component _FTota i _e (more simply designated by A _F in the remainder of this document). In fact, during the coding mechanism of the transformed coefficients, several capital information are signaled, such as the absolute value (si> 1, si> 2), the sign, the position, and so on. In the HEVC standard, this information is reported through the following syntax elements:

• transform_skip_flag

• last_sig_coeff_x_prefix

• last_sig_coeff_y_prefix

• last_sig_coeff_x_suffix

• last_sig_coeff_y_suffix

• coded_sub_block_flag

• sig_coeff_flag

• coeff_abs_level_greaterl_flag

• coeff_abs_level_greater2_flag

• coeff_sign_flag

• coeff_abs_level_remaining

These syntax elements have signaling conditions that depend on the statistical distribution of the coefficients and their location in the residue. For example, the following two residues presented in Tables 1 and 2 below will not have the same cost in terms of bits, even though they contain the same coefficients:

TABLE 1 TABLE 2

The simulations carried out made it possible to observe the reality of the model established by equation (13).

FIGS. 2A and 2B show the curves of the rate R as a function of the parameter p, after entropic coding of the CABAC type, respectively for class A (FIG. 2A) and class B (FIG. 2B) images. It is recalled that the reference sequences are divided into classes. These classes are composed of sequences of specific resolutions and specific content (in some cases). Class A contains two sequences of 2500x1600 resolution, the two sequences containing many movements. Class B contains 6 sequences of 1920x1080 resolution with varied content (tracking, fixed shots, with or without plane changes). For more information on this notion of class, reference may be made to JCTVC-O0022 prepared by the Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG 16 WP 3 and ISO / IEC JTC 1 / SC 29 / WG 11 "at the 15th meeting held in Geneva, Switzerland, from 23 October to 1 November 2013.

In FIGS. 2A and 2B, the lines referenced 20, 21, 22 and 23 represent the theoretical model of the rate R as a function of the parameter p obtained according to the calculation method of the invention, considering the slope parameter Θ equal to 4 * N ² , for coding residuals of size 4x4 (right referenced 20), 8X8 (right referenced 21), 16x16 (right referenced 22), 32x32 (right referenced 23), and neglecting the adjustment coefficient Δ. The scatter plots represent the flow values measured on residues from reference bitstreams, respectively of size 4x4, 8x8, 16x16 and 32x32. These reference bitstreams correspond to the sequences from the classes mentioned previously encoded using the reference software, in Random-Access configuration and for four fixed QP values (22, 27, 32 and 37). We recalls that the HM reference software consists of an encoder and a decoder used for compliance and interoperability testing. The reference bitstreams are streams, encoded with the reference software, and made available by the BBC (registered trademark) to be able for example to test the decoders. Bitstreams are available and cover for each sequence the eight configurations (All-Intra, All-Intra 10, Low-Delay P, Low-Delay P 10, Low-Delay B, Low-Delay B 10, Random-Access and Random Access 10), for four different QP's (22, 27, 32 and 37). In the context of the present work, the inventors have used the HM11 version of the abovementioned reference software, as defined at the 13 ^th meeting of the Joint Collaborative Team on Video Coding (JCT-VC) of the ITU-TSG16 WP3 and ISO / IEC JTC1 / SC29 / WG11 ", held in Incheon, South Korea, from 18 to 26 April 2013, and described in JCTVC-M1010.

Note also that a measurement of the average value of the adjustment coefficient Δ for the curves of Figures 2A and 2B gives the mean values of deviations listed in Table 3 below, for the different sizes of residues:

Size Δ Medium

4x4 -0.11 bits

8x8 -1.15 bits

16x16 -2.75 bit

32x32 -6.35 bits

Average -2.59 bits

TABLE 3

FIG. 3 illustrates the evolution of the rate R as a function of the parameter p, no longer at the level of a residual, but at the level of one frame, for different class A to E images. Indeed, to achieve a saving in In terms of calculations, it is sometimes preferable to perform rate control at the image level and to encode each macroblock with the same value of the quantization parameter QP.

Thus, the straight lines referenced 30 to 34 illustrate the theoretical model of the flow rate R as a function of the parameter p, obtained according to the calculation method of the invention, according to which the slope parameter θ = 4S, where S is the total area of the image, and neglecting the adjustment parameter Δ _Ρ , and this, respectively, for class A, B, C, D and E images. Indeed, it is recalled that the test sequences are divided into 5 classes of resolutions and varied contents (A : 2560x1600, B: 1920x1080, C: 832x480, D: 416x240 and E: 1280x720). The clouds of points concentrated around the lines referenced 30 to 34 represent, for their part, the values measured on frames from the aforementioned reference bitstreams for these different classes A to E.

It may be noted that the flow rate model Rp illustrated by the curves referenced 30 to 34 is therefore fairly close to reality, but does not prove to be as accurate as the flow rate model at the residue level illustrated in FIGS. 2A and 2B. Indeed, when encoding a frame, a non-negligible number of residues are not coded: the prediction being perfect, the residues are indeed null, which causes the "skip" (ie the non-compression ) of some elementary coding units in the image, or CL ⁾ for "Coding Unit".

In order to evaluate the slope of the p-domain more precisely, an alternative embodiment of the invention relies on an approach making it possible to estimate it more precisely, and thus to move the lines referenced 30 to 34 of FIG. 3 to the associated point clouds, depending on the properties of the frame.

In this variant embodiment, the total area of the frame is not considered to be the coded surface, but the non-skippered surface in the image is taken into account. This approach results in the following equation, where r _skip represents the skip rate in the image:

S '= (l - r _skip ) * S (14) Using this approach in equation (13), we obtain by replacing S with S':

R = 4 * (1 - r _skip ) * S * (1 - p) + A _F (15) This formulation of the flow rate model Rp makes it possible to obtain a significant precision, and substantially increased compared to prior techniques.

Figure 4 presents a comparison of theoretical values obtained by equation (15) and measured values for the slope Θ of the rate model Rp, as a function of the skip rate _skip . More precisely, the lines referenced 40 to 44 represent the evolution of the slope θ = 4 * (l-r _skip ) * S as a function of the skip rate ^r skip for class A to E images respectively. The clouds of points concentrated around these lines represent the corresponding values measured for the slope Θ, for the same images. In practice, the rate of skip _skip r, the parameter p and the flow rate outputted from the encoder were measured when decoding each frame, for all sequences. For each skip rate interval of 1%, the slope of the line is estimated by linear regression. These estimated values are then compared to the slope values determined using equation (15).

Table 4 below makes it possible to observe the average accuracy obtained in the slope estimation using the method of the invention. This precision is increased on the range corresponding to the most frequent cases in practice (namely + 50% of skip).

Class After CABAC

92.49%

B 92.93%

C 91.60%

90.49%

E 90.14%

Average 91.53%

TABLE 4

The flow control method of the invention thus makes it possible, compared with the known methods of the prior art implementing flow control in two encoding passes, to eliminate a first encoding pass. It also makes it possible to estimate the slope of the p-domain more precisely in the context of one-pass regulation.

FIG. 5 illustrates a modified architecture of the video coder of FIG. 1 for implementing flow control at the CTU level, in accordance with the principle of the invention described above.

In this FIG. 5, all the blocks and signals bearing the same reference numerals as in FIG. 1 are identical to the corresponding blocks and signals of FIG. this figure, and we can refer to the description of Figure 1 to understand the operation.

The architecture of the encoder of FIG. 5 differs from that of FIG. 1 in that the transform-quantization block 107 is split into a first block referenced 107a operating the transformation of the residues, and a second block referenced 107b operating their quantization.

The block referenced 118 receives the budget allocated for the coding, as well as the size of the residue, from the control block referenced 114. From the equation (4) proposed above, it calculates the value of the parameter p, it transmits to the block referenced 119.

The block referenced 119 then calculates, from the value of p received from the block referenced 118, the appropriate quantization parameter QP.

The block referenced 107b then proceeds to the quantization of the transformed residue by the block referenced 107a from the quantization parameter QP determined by the block referenced 119.

We now show, in connection with Figure 6, an example of calculation of skip rate _sfcip .

In this figure, we present a sequence of twenty-one frames, or frames, in Random-Access configuration. It is recalled that the Random-Access configuration is the case of use for broadcasting: the images are placed in the decoding order in the bitstream (which is different from the display order). Ramdom-Access Point pictures (RAPs) appear regularly in the bitstream so that you can resume decoding the current program. The numerical value associated with each of these frames indicates the order of encoding (and thus decoding) of these different images. These images are grouped into groups of images (or GOP, for Groups of Pictures): thus, the images numbered from 1 to 8 constitute a first GOP; the images referenced from 9 to 16 constitute a second GOP; etc.

In this case, it is a hierarchical GOP, each including several Inter images, except during the refresh where the first image of the GOP in the encoding order is necessarily an Intra (RAP) image. We recall that an Intra image uses only spatial prediction, and in particular allows to initialize the encoding of a sequence (RAP), because it is coded independently of the other images; an Inter image uses intra- and inter-image prediction and therefore exploits spatial redundancies in the image, and temporal in the image sequences.

In the hierarchical GOP example of FIG. 6, the images are encoded in successive time levels according to a pyramidal structure based on Inter images.

Thus, in the first GOP of FIG. 6, the frame 1 is a unidirectional Inter picture, and the frames numbered 2 to 8 are bidirectional Inter pictures.

In the context of the invention, different methods for determining the skip rate can be implemented.

For example, one may use a moving average prediction model for predicting the skip rate r i _sk _P frame n from the skip rate r i _sk _P measured on the previously encoded frames. The initialization for the first GOP may be as described in Table 5 below, with the POC (for "Picture Order Count") which numbers the position of the frame in the GOP:

TABLE 5

Thus, in the configuration of FIG. 6, using a moving average, the skip rate _ski p of the numbered frame 15 can be predicted with respect to the frames previously encoded according to the equation:

[13] + 2 [12] + 1 [8]

RsfcipL ¹⁵ J- 7 (16)

Similarly, for the numbered frame 14, the skip rate can be predicted according to the equation: According to another embodiment, in a scalable video coding scheme such as SHVC for example, the skip rate _sk i _P of the enhancement layer can be determined from the skip rate _Sk i _P measured on the layer. basic.

According to yet another approach, the skip rate _sk i _P can be determined based on the correlation between an image and its nearest neighbors, which can give an indication of the proportion of the image that is motionless, and who could be skippered.

Thus, the measurement of r skip _skip rate during encoding can be carried out simply, using a counter: for a given frame, when a coding unit CU is skippered (ie uncompressed because the residue is zero) this counter is incremented with the surface of this skippered CU coding unit, taking into account the 3 characteristic YUV components of the color space used. When the encoding of the frame is completed, the value S _skip stored in the counter, corresponding to the total area skipped for this frame, is divided by the total area of the image S _{Frame in} order to obtain the r _Sk i _P - The counter is then reset for the next frame. The formula for calculating the r _skip is then:

_ ^S Skip

rskip- ç U °) ^ Frame

For example, for a configuration with a 1920x1080 resolution that is progressive with a 4: 2: 0 pixel ratio (representing the sampling ratio of the colorimetric components, 4: 2: 0 equals a ratio of 1.5) , and a skippered surface S _skip =

1500000 samples measured on a frame, we obtain:

1.5 million

r _S kiv = = 48.22% (19) s ^klv 1920 * 1080 * 1.5 '

Other methods of measuring r skip _skip levels can of course be implemented in the flow control method of the invention.

FIG. 7 illustrates a variant of the architecture of the encoder of FIG. 1, making it possible to implement a flow rate regulation at the frame level, according to the principle of the invention.

All the elements bearing the same numerical references as in FIG. 1 are identical in principle and in their operation to what has been described. previously in connection with Figure 1, and will not be described here again in more detail.

The architecture of the video encoder of FIG. 7 differs from that of the encoder of FIG. 1 by the elements described hereinafter.

As the CTUs are encoded, a counter of the skippered surface is incremented. When cu_skip_flag = 1 is reported, the corresponding area is added to the counter, which is updated in block referenced 118.

Once the encoding of a frame has been completed, the block referenced 119 receives the value of the counter coming from the block referenced 118 and divides it by the total surface of the image (ie the sum of the surfaces of the components) in order to determine the value of the r _skip parameter.

The value of the ratio r _skip is transmitted to a database 120 of r _skips per frame (according to the hierarchical level of the frame in the GOP).

The block referenced 121 determines the r _skip for the next frame, for example by moving average (as explained above in relation to FIG. 6), using the database 120. It transmits the value thus determined to the block referenced 122. , which also receives from block referenced 114 the budget allocated to the coding and the total surface of the image. The block referenced 122 can then deduce the value of the parameter p from the equation (15) at the frame level, and transmits it to the block referenced 123.

The block referenced 123 then determines the value of the quantization parameter QP, and transmits it to the controller 114, which, in turn, communicates it to the quantization block referenced 107.

The rate control method of the invention makes it possible to substantially reduce the complexity of the video encoder compared to the methods of the prior art. In particular, it makes it possible to halve the latency of a double-pass encoder. It also makes it possible to substantially increase the accuracy, the quality or the bit rate, according to the fixed constraints, of a simple pass encoder.

The set of principles and characteristics described above in relation to the H EVC standard can of course be extended to other types of video coding, and more particularly to the SHVC scalable video coding standard.

Claims

A method for estimating a coding rate of an image (100) of a sequence of images implementing a rate model according to a p-domain approach such that R (p) = θ (1) - p), where:

R denotes said flow,

p denotes a percentage of zero coefficients in at least one coding residue of said image, after transformation (107a) and quantization (107b) of said residue, said residue being obtained by difference between a block of said image and a prediction of said block,

Θ denotes a slope parameter of said flow model,

characterized in that it comprises a step of calculating said slope parameter Θ taking into account a surface of said encoded coding residue (s) for said image, said surface being the sum of the resolutions of each of the components of said image.

2. Estimation method according to claim 1, characterized in that during said calculation step, said slope parameter Θ is calculated as a multiple of said surface of said coded coding residue (s) (s). ) for said image.

3. Estimation method according to any one of claims 1 and 2, characterized in that, p denoting a percentage of zero coefficients in a residue of size NxN of said image after transformation and quantification of said residue, the slope parameter Θ is determined, in said calculation step, in the form 0 = 4 * N ² .

4. An estimation method according to claim 3, characterized in that it further comprises a step of estimating said coding rate at said residue such that: G ⁰ ) ⁼ 4 * N ² * (1 p ) + Δ, where Δ denotes an adjustment parameter dependent on a coefficient distribution in said residue.

5. Estimation method according to any one of claims 1 and 2, characterized in that, p designating a percentage of zero coefficients in said image of surface S after transformation and quantization, the slope parameter Θ is determined, during the said calculation step, in the form Θ = 4 * S.

6. Estimation method according to any one of claims 1 and 2, characterized in that, p denoting a percentage of zero coefficients in said surface image S after transformation and quantization, the slope parameter Θ is determined, when of said calculating step, in the form Θ = 4 * (1-r _skip ) * S, where r _skip denotes a skip rate in said image.

7. Estimation method according to any one of claims 5 and 6, characterized in that it further comprises a step of estimating said coding rate at said image such that: R (p) = Θ * (1-p) + A _F , where A _F denotes an adjustment parameter dependent on a distribution of said coding residues for said image.

A method of coding an image sequence comprising a step of detecting each of said images in at least one image block, and an encoding phase comprising steps of:

predicting (102) each of said blocks, delivering a predicted block;

for each of said blocks, forming a difference coding residue between said picture block and said predicted block;

transforming and quantizing (107; 107a, 107b) said coding residues; said encoding method implementing a mechanism for estimating a coding rate of said images from a rate model according to a so-called p-domain approach such that R (p) = θ (1-p), where R denotes said rate, p denotes a percentage of zero coefficients in at least one coding residue of said image, after transformation and quantification of said residue, and Θ denotes a slope parameter of said flow model,

characterized in that it comprises a step of calculating said slope parameter Θ taking into account a surface area of said code-encoded residue (s) for said image, said area being the sum of the resolutions of each components of the image.

9. Coding method according to claim 8, characterized in that, during said calculation step, said slope parameter Θ is calculated as a multiple of said surface of said coded coding residue (s) (s). ) for said image.

10. Coding method according to any one of claims 8 and 9, characterized in that, for each of said blocks, it implements a single iteration of said encoding phase.

Apparatus for coding an image sequence comprising a module for cutting each of said images into at least one image block, and an encoding module comprising:

a prediction module (102) of each of said blocks, delivering a predicted block; a transformation module (107a) of a coding residue formed by difference between said image block and said predicted block;

a quantization module (107b) of said coding residue;

said coding device comprising means for estimating a coding rate of said images from a rate model according to a p-domain approach such that R (p) = θ (1-p), where R denotes said rate, p denotes a percentage of zero coefficients in at least one coding residue of said image, after transformation and quantification of said residue, and Θ denotes a slope parameter of said flow model, characterized in that said regulating means comprise a module for calculating said slope parameter Θ taking into account a surface of said coded coding residue (s) for said image, said surface being the sum of the resolutions of each of the components of said image,

and in that each of said blocks is encoded by a single pass in said encoding module.

12. Computer program comprising instructions for implementing an estimation method according to any one of claims 1 to 7 when said program is executed by a processor.

Computer program comprising instructions for implementing an encoding method according to any of claims 8 to 10 when said program is executed by a processor.

A computer-readable recording medium on which is recorded a computer program comprising instructions for performing the steps of the estimation method according to any one of claims 1 to 7.

A computer-readable recording medium on which a computer program is recorded including instructions for performing the steps of the encoding method according to any of claims 8 to 10.