FR2891965A1 - METHOD, DEVICE AND PROGRAM FOR SCHEDULING DATA UNITS OF A SCALABLE VIDEO STREAM, AND CORRESPONDING TERMINAL THEREFOR. - Google Patents

Abstract

The invention relates to a method for scheduling data units (51) of a scalable video stream at least in time and quantization dimensions, obtained by hierarchical coding, comprising a scheduling step delivering an ordered list of said data units comprising, for a set of data units to be ordered, the following steps: - introducing into said ordered list of data units of the quantization base layer; - association (53), to each data unit of the quantization base layer; a quantization refinement layer, an efficiency information representative of an increase in quality and / or a rate increase provided by said data unit; scheduling (54) said data layer data units; quantization refinement according to said effectiveness information and a presence criterion, indicating that said reference data units of the data unit es considered are already present in said ordered list.

Description

Method, device and program for scheduling data units of a

  scalable video stream, and corresponding terminal. FIELD OF THE DISCLOSURE The field of the invention is that of coding and decoding of data streams, and more particularly of scalable (or scalable) data streams having a hierarchical structure in layers, or in levels. These layers can be defined in several dimensions, and classically the temporal dimensions, in quality and / or spatial. In general, it is desirable to order the data units constituting the data stream according to their efficiency, that is to say according to the improvement in quality that they provide, with respect to the required bit rate. This allows, in the case where all the data units are not used in the terminal, to provide an optimized playback quality. Priority levels are thus defined for ordering the data units of a data stream. In other words, we must define an ordered list of data units. The invention relates more precisely to this aspect. It therefore aims to propose a scheduling technique such that the most relevant data are extracted first during a rate adaptation and / or spatio-temporal resolution.

  The invention applies in particular, but not exclusively, to Scalable Video Coding (SVC), also known as scalable video encoders, which operate in a closed loop. It is clear, however, that other scalable closed-loop coders can implement the invention. 2. Techniques of the Prior Art 2.1. Scalable Video Encoding Most video encoders generate a single compressed stream corresponding to the entire encoded sequence. If several clients wish to use the compressed file for decoding and visualization, they will have to download (or stream) the complete compressed file. More recently, so-called scalable video coding algorithms (adaptive quality and variable spatio-temporal resolution) have appeared for which the coder generates a compressed stream in several layers, each of its layers being nested in the layer of higher level. Such coders are very useful for all applications for which the generation of a single compressed stream, organized into several scalability layers, can serve several clients of different characteristics, and for example: - video-on-demand services (in English video on demand, or VOD), for terminals of the UMTS type, PC ADSL, TV ADSL ... session mobility (recovery on a PDA (electronic personal assistant) of a video session started on a television, or on a mobile UMTS of a session started on the GPRS network); - continuity of session (bandwidth sharing with a new application); High definition television (single encoding to serve standard definition (SD) or high definition (HD) clients; videoconferencing (single encoding for UMTS clients and Internet clients) Some of these scalable video coding algorithms are now is now being adopted by the MPEG ("Moving Picture Expert Group" for "Image Coding Expert Group") as part of the MPEG4 Working Group.

  Notably, the model recently adopted by the MPEG-4 AVC (Scalable Video Coding) Working Group is called SVM (Scalable Video Model for Scalable Video Model), and is based on a scalable encoder based on AVC solutions (Advanced Video Coding for Advanced Video Coding). This model is described in more detail in N6716 ISO / IEC JTC 1 / SC 29 / WG 11, Scalable Video Model 3.0, October 2004, Palma de Mallorca, Spain. The aim of the MPEG 4 working group is to propose a standard for the supply of scalable medium grain streams in time, space and quality. 2.2 The SVM MPEG-4 SVC Encoder 2.2.1 Key Characteristics of the Encoder Figure 1 illustrates the structure of such an encoder, which has a pyramidal structure. The video input components 10 undergo dyadic subsampling (2D decimation by 2 referenced 11, 2D decimation by 4 referenced 12). Each of the subsampled streams is then subjected to a temporal decomposition 13 of the MCTF (Motion Compensated Temporal Filtering) type. A low resolution version of the video sequence is encoded 19 up to a given bit rate which corresponds to the maximum decodable bit rate for the low spatial resolution (this base level is AVC compatible).

  The higher levels are then encoded by subtracting the previous reconstructed and over-sampled level and encoding the residuals as: a base level; optionally one or more enhancement levels obtained by multi-pass encoding of bit planes (hereinafter referred to as FGS for Fine Grain Scalability, fine grained scalability). According to this approach, the motion information 17 and the texture information 18 are distinguished. In order to achieve a rate adaptation, the texture information 18 is coded using a progressive scheme: coding of a first level of minimum quality (called Base Layer, or Base Layer); coding of progressive refinement levels (called Enhancement Layer, or Layer of Enhancement). With reference to FIG. 1, the texture information 18 feeds a coding module of the quantization base layer 19. The coded data, at the output of the module 19, are used to feed a block 21 for spatial transformation and entropic coding, which is working on the levels of signal refinement. The output data of the module 21 feeds an interpolation 20 from the base level. A multiplexing module 22 orders the different sub-streams generated in a global compressed data stream 23.

  The compressed stream 23, at the output of the coder, is therefore structured in unitary data units called Network Abstraction Layer Units (NALUs). Each NALU is associated with an image resulting from spatio-temporal decomposition, a spatial resolution level, and a quantization level. This structuring in elementary units makes it possible to adapt the bit rate and / or space-time resolution by eliminating the NALUs with too great spatial resolution, or with too long a time frequency, or even with too much encoding quality. 2.2.2 Priority level mechanism The principle of the priority level mechanism has been developed by the inventors of this patent application, and is described in patent application FR-05 00331, not yet published. It consists in associating elements or portions of elements of an SVC (Scalable Video Coding) stream with so-called priority level information. Thus, this priority level information allows an extractor (also called quality adaptation module) to efficiently extract the most relevant information during a rate adaptation and / or spatio-temporal resolution. FIG. 2 illustrates the results obtained by means of the scalable coder of FIG. 1, in the form of flow-distortion curves, represented for operations with and without a priority level mechanism and for a QCIF (Quarter Common Interface Format) resolution. QCIF corresponds to a quarter of TV format) at 15 Hz. The ordinate shows the PSNR (Peak Signal to Noise Ratio for peak signal-to-noise ratio) and in abscissa, the rate expressed in kbits / s. Thus, the referenced curve 24 corresponds to an operation without a priority level mechanism and the curve referenced to operation with a priority level mechanism. The priority level mechanism therefore makes it possible for the same quality to gain on the order of 10 to 15% of flow. The estimation of the priority levels to be used is based on a rate-distortion formalism, this type of estimation comprises four steps: a first step during which it is estimated, for each sub-part of the flow considered, the variations in deltaR flow rate and in deltaD quality generated; in the next step, from the estimated variations, the convex envelope of the set of quality-rate pairs for each image coded in the stream is calculated; then in a third step, the points are recovered which are defined by quality-flow couples positioned on the calculated convex envelope and in which it is possible to associate a slope value; finally, in a fourth step, each retrieved point is associated with a priority level corresponding to a quantized slope value. 2.2.3 Operation of the MPEG-4 SVC coder in open loop and closed loop The MPEG-4 SVC encoder can implement two types of time decorrelation: MCTF time decorrelation, which includes a prediction phase consisting of a calculation of residue on images resulting from a prediction by motion compensation, and an update phase consisting of a feedback by compensation of the movement of the previously calculated residuals in the reference images used for the prediction previous ; the temporal decorrelation of the Unconstrained Motion Compensated Temporal Filtering (UMCTF) type for temporal filtering compensated in unconstrained motion, which comprises only the aforementioned prediction phase. It should be noted that this unique prediction phase also applies for the case of a temporal transformation of hierarchical B-frames (hierarchical B-frames). This encoder can also operate in two distinct modes: in open loop mode, in which the motion compensations to the encoder are performed using the original values of the images; in closed loop mode, in which the motion compensations to the encoder are performed using the coded-decoded values of these images. The closed-loop coding technique is a technique commonly used in prediction coders (for example in the MPEG1, MPEG2, ... coding schemes) to avoid drift phenomena. Indeed, an encoder can make a prediction from original images, while the corresponding decoder has only coded-decoded versions of these images. This closed-loop coding technique therefore makes it possible to provide the coder and the decoder with the same reference images.

  In the document JVT-P059, entitled Comparison of MCTF and closed-loop hierarchical B peaks, July 2005, Poznan, it is shown that the addition of a closed-loop coding technique and a temporal transformation of B-images Hierarchical type provides better performance (in terms of rate-distortion) than an open-loop MCTF coding technique. 2.2.4 Hierarchical closed-loop B-frame encoding structure Throughout the rest of this document, we will consider the following notation: for an image B, y the first index x corresponds to the temporal hierarchical level (or enhancement layer). temporal) of this image and the second index y to the index of this image in its temporal hierarchical level. By hierarchical level is meant here the level of temporal resolution. The maximum time hierarchical level corresponds to the highest resolution level, ie the highest temporal frequency. We will also speak in this case of the deepest hierarchical level. A closed-loop hierarchical B-frames encoding structure gives rise to a dyadic decomposition: at each new hierarchical level, a new image is inserted between each image pair of the previous hierarchical levels. Of course, more generally, it is possible to have for each new hierarchical level, the insertion of any, one or more images between each image pair of the previous hierarchical levels. A hierarchical type B (hierarchical B-frames) coding structure is now described with reference to FIG. 3: the image at t + 2 (B ,,,) is coded using a prediction obtained. from the images at t (Bo ,,) and at t + 4 (Bo, 2); the image at t + 1 (B2,1) is encoded using a prediction obtained from the images at t (Bo ,,) and at t + 2 (B ,,,); the image at t + 3 (B2.2) is coded using a prediction obtained from the images at t + 2 (B 1,) and at t + 4 (80.2). In closed-loop operation, the predictions of the different images are obtained using the coded-decoded versions of the reference images. In the context of scalable (i.e., multi-level) coding, closed-loop coding is performed at each quantization level. The prediction residuals at each quantization level are differentially encoded with respect to the coded-decoded prediction residue of the previous quantization level. The information thus coded forms a quality increment (also called quantization increment).

  With such a closed-loop coding structure, the increments of coded residues for an image are therefore conditional on the fact that the reference images are previously coded-decoded for the quality level of coding considered. 3. Disadvantages of the Prior Art The technique for calculating the levels of priority currently used is a technique which is based on the concept of separability of the images to be encoded. That is, it is assumed that the influence of the coding level of an image does not affect the performance expected by encoding another image. This property is ensured by the use of an open-loop MCTF coding technique as a temporal decorrelation tool. In the context of a closed-loop hierarchical B-frames (or UMCTF) encoding technique, this property is no longer valid, since the coding of the images depends for each hierarchical level of time on at least two reference images , called parent images, of a higher temporal hierarchical level. As a result, the current technique for calculating priority levels is not well suited. 4. OBJECTIVES OF THE INVENTION The object of the invention is notably to overcome these disadvantages of the prior art. More specifically, an object of the invention is to provide a technique for scheduling the data units of a scalable video stream, in particular coded in a closed loop, which is more efficient, in terms of quality and / or bitrate, than the known techniques.

  In other words, an object of the invention is to provide such a technique, which maximizes the efficiency of the reconstruction of the flow, especially when it is operated only partially. Another object of the invention is to provide such a technique, which does not introduce complexity or important treatments. In particular, an objective of the invention is to allow a simple and efficient extraction of the most important data of an SVC stream during a rate adaptation and / or spatio-temporal resolution. It is another object of the invention to provide such a technique which, in a particular embodiment, is adapted to an MPEG4-SVC coder and decoder implementing a hierarchical B-frames coding technique and operating in a closed loop. 5. Main features of the invention These objectives, as well as others which will appear more clearly later, are achieved by means of a method of scheduling data units of a scalable video stream at least in the temporal and quantization dimensions, obtained by hierarchical coding, said data units being organized in successive layers comprising a temporal base layer and at least one temporal enhancement layer in said temporal dimension and a quantization base layer, and least one quantization refinement layer in said quantization dimension, each data unit of a given temporal enhancement layer being encoded taking into account at least one reference data unit of a lower temporal layer. According to the invention, the method comprises a scheduling step 20 delivering an ordered list of said data units comprising, for a set of data units to be ordered, the following steps: input into said ordered list of data units of the quantization base layer; associating, with each data unit of a quantization refinement layer, information of efficiency indicative of an increase in quality and / or an increase in bit rate provided by said data unit; scheduling said quantization refinement layer data units based on said effectiveness information and a presence criterion, indicating that said reference data units of the relevant data unit are already present in said ordered list. Thus, the data units are organized in an optimized manner and adapted in particular to a closed-loop coding. In particular, an efficiency superior to that of the prior art is obtained. Advantageously, the method comprises a step of defining dependency relationships between said data units, so as to control said presence criterion. It is indeed desirable to know in advance these dependency relationships, in particular for the implementation of closed-loop coding. Preferably, said step of defining dependency relationships comprises a step of identifying, for each data unit of a temporal enhancement layer, a single reference data unit, called the parent reference data unit, chosen from so that the presence of said parent reference data unit in said ordered list ensures the presence therein of all of said reference data units. Advantageously, said reference data unit father of a data unit of a n-level temporal enhancement layer is the reference data unit having the highest or the lowest index among set of reference data units of the n-1 level temporal enhancement layer. According to a preferred embodiment of the invention, said data units are coded using a hierarchical coding implementing a temporal transformation delivering images encoded in bidirectional closed-loop mode. In this case, said data units may comprise: Intra images, called I images; predicted images in monodirectional mode, called P-pictures; predicted images in bidirectional mode, called B images.

  Said scheduling step then advantageously consists in selecting, quantization refinement layer by quantization refinement layer, starting with said base layer, all of said images I, then all of said images P and finally all said images B ordered by temporal enhancement layers. According to an advantageous aspect of the invention, said scheduling step simultaneously processes groups of at least two data units, so as to take into account the sum of said efficiency information and privilege one group with respect to another, if it is overall more efficient.

  Preferably, said scheduling step can select said groups of data units so that a flow rate / quality representative of said path corresponds to a convex envelope of all of the different eligible hierarchical paths. Advantageously, said scheduling step comprises, for each temporal enhancement layer n, a step of merging ordered sub-lists of the dependent data units, for each data unit of said temporal enhancement layer n, and then a step of merging ordered lists of the set of quantization refinement layers. According to another advantageous characteristic of the invention, the method comprises a step of inserting, in a signal comprising said data units, at least one piece of information representative of said ordered list. The invention also relates to a device for transmitting and / or storing a scalable video stream implementing the method as described above.

  It should be noted that the method can be implemented in many devices, at different locations of a data transmission or storage chain, as needed. The invention also relates to a computer program product downloadable from a communication network and / or stored on a computer readable medium and / or executable by a microprocessor, comprising program code instructions for the implementation of the method of scheduling described above. The invention also applies to terminals able to receive a stream whose data units are ordered according to the method described above. Such a terminal comprises means for selecting a part of said data units within said stream, from said ordered list and as a function of at least one criterion of bit rate, spatial resolution and / or frequency of data. display, and means for reconstructing a video stream adapted to said criterion, using said selected data units. 6. List of Figures Other features and advantages of the invention will appear more clearly on reading the following description of a preferred embodiment of the invention, given as a simple illustrative and non-limiting example, and annexed drawings among which: FIG. 1, already described in relation with the prior art, presents a block diagram of the MPEG-4 SVC encoder described in the aforementioned document N6716; FIG. 2, also described in relation with the prior art, illustrates the flow-distortion curves obtained from the coder of FIG. 1 with and without a priority level mechanism; Figure 3, commented in relation to the prior art, illustrates a coding structure using hierarchical B-frames; Figure 4 is a general flowchart of the method of the invention; Fig. 5 is an example of scheduling increments of data items; FIG. 6 illustrates the evolution of the distortion as a function of the rate of two scheduling of envisaged data elements; Figure 7 shows an embodiment of the scheduling algorithm implemented in the invention; Figures 8a to 8d illustrate the result of the algorithm presented in connection with Figure 7; Fig. 9 is also an illustration of the scheduling algorithm of the data elements of the invention; FIGS. 10A and 10B schematically illustrate a coding device and a terminal implementing the invention. 7. DESCRIPTION OF A PARTICULAR EMBODIMENT The invention thus proposes an optimized and efficient technique for scheduling data elements that make up images, which can be applied to the case of images encoded according to a closed-loop structure. It will be recalled that for this type of coding, and as illustrated in FIG. 3, the elements of images of a temporal hierarchical level (or temporal enhancement layer) n each depend on at least one reference picture element, at a hierarchical level of lower resolution n-1 (the temporal frequency is lower). In addition, as detailed above, this closed-loop coding is performed for each quantization level in order to obtain a quality / rate variation. This quality increment is an information of efficiency representative of the increase in quality as a function of the increase in flow rate. The main problem to be solved for the scheduling of the data units composing an image flow is thus to define a path, or ordered list, which represents a hierarchy of the data according to their importance, so as to obtain the best efficiency, when we consider the evolution of quality as a function of flow. 7.1 General Principle The main steps of the invention are presented in the form of an algorithm in FIG. 4. The coded images 51 are used for a first step 52 of defining the dependency relationships between them. These dependency relationships will subsequently allow constraints on the scheduling of the elements of each quality increment.

  It follows, for each image (or increment, or unit of data) of each hierarchical level of time, a step 53 of calculating the variations of flow as a function of the level of quantification (according to the approach described above): we obtain and efficiency information associated with each data element. The higher the slope of the flow variation depending on the quality, the greater the efficiency.

  From these first two steps 52 and 53, a step 54 of scheduling this efficiency information is performed. By this new scheduling step, we will see later that we obtain a convex envelope of quality evolution as a function of the flow, thus generating the best possible performances in terms of image quality as a function of the flow rate. and therefore in terms of efficiency. The scheduling obtained in step 54 makes it possible to obtain an ordered list, that is to say, to order the various efficiency information associated with each data element. A rising priority level with their order of appearance is then assigned to these elements for example, using a quantization of their order of appearance. This follows a step 55 of quantizing the priority levels, which are transmitted to the decoder in a step 56 of transmission of the priority levels. The number k of quantization levels available is different from the size of the order of appearance, up to a value n. It is then possible to quantify an order of appearance between 1 and n over a range of quantization values between 1 and k. We thus achieve a hierarchy of data, in order of importance, or appearance. We now detail some of the steps of the method of the invention. 7.2 Preliminary Notations One considers the case of a hierarchical transformation of type hierarchical B-frame. An image noted B ~ 1 thus corresponds to the i-th image located at the nth level of decomposition, that is to say the n-th hierarchical level of temporal refinement (or layer of temporal enhancement).

  It is further recalled that the predictions of each image Bn, 1 are obtained from two reference images, of lower hierarchical level (lower frequency resolution), by temporal refinement. In addition, for each image of each hierarchical time level, quantization levels of index q are defined. Note then Bn, (q) the q-th encoding quantization increment, associated with the image Bn,. On the other hand, Bn ,, r, and 13n2.2 will be noted as the two reference images used for the prediction of Bn,. In a more general context, it is of course possible to take into account a number of reference images greater than two. 7.3 Definition of dependency constraints In this section, step 52 of the algorithm of the invention is detailed. Due to the dependence of data elements (or units, or NALUs) on data with other higher hierarchical level data elements (lower resolution), a dependency relationship must be established between each of them. Indeed, it is necessary that two reference data elements are known, and therefore present for the decoding of the lower hierarchical level data element (higher refinement). For this, the technique of the invention associates with each piece of information of effectiveness of the data element Bn, (q) the two information of effectiveness of the reference data elements used in the prediction of the image Bn ;. Of these two efficiency information of the reference data elements, and because of the hierarchical B-frame coding structure, there is necessarily one of the two reference elements used in the prediction of the Bn element; a reference element located at the hierarchical level of directly higher temporal decomposition n-1. Among these reference elements located at level n-1, then a reference datum element Bnp, p. This parent data item is obtained by selecting the highest index reference data item at hierarchical level n-1.

  In FIG. 3, the parent data item of B21 1 will be for example B111. This construction of relationship between the data elements then makes it possible to take into account the dependence constraints between the enhancement information, stating that the quality information of the data element Bn,; (q) will only appear after the raising information of the data element Bnn,; n (q). It is noted that the images of first hierarchical level, that is to say of lower temporal resolution, do not have a parent reference image. In a particular embodiment of the invention, it is possible to group all these images of the first hierarchical level into a single quantization information value. Finally, at each quality increment BX, y (q) (q-th quantization level of the y-th image of the hierarchical level x), we associate a list of dependent references LDX, y (q). This list of references LDX, y (q) then contains the set of elements that depend on BX, y (q). In other words, LDX, y (q) contains the efficiency information of the child data elements with parent reference data element BX, y (quality increment BX, y (q)). referencing can be built recursively: - LD ,,, y (q) is empty if no parent element Bn ,; has parent reference element BX, y. LDX, y (q) contains the Bn, (q) and LDn, (q), such that BX, y is the parent reference element of Bn,; So we see that a parent reference element can have multiple child data elements. 7.4 Calculation of Rate Variations and Distortion With step 52, it is known that the parent reference data items must appear before the data items that depend on them. It remains to define the order in which the child data elements appear (for the same parent reference image).

  For this, a step 53 is performed for calculating the variations in flow and distortion. In other words, for each quality increment B ,,,; (q) what is its deltaRn ,, (q) impact on the overall flow, as well as its impact deltaD, u (q) on the quality of global reconstruction. The impact on the overall bit rate is obtained for example by using the size of the bits of the coded information relating to this quality increment: one then obtains information of effectiveness relating to each data element. The impact on the distortion of the reconstructed images is obtained for example by measuring the difference in distortions between the case where the quality increment is present and that where it is not present. 7.5 Scheduling of Quality Increments The purpose of step 54 of scheduling is therefore to find the most effective combination of quality increments among the different combinations that satisfy the constraint that B ,,,; (q) appears before the included elements. in their corresponding list LDn,; (q). An example of scheduling is illustrated in FIG. 5. We consider the simple case where two data elements B1.2 and B ,,, have the same reference datum element Bo, l. In this example, the set 61 is formed of the data element B ,,, associated with its two flow variations deltaR = 2 and quality deltaD = 1. Similarly, the assembly 62 is formed of the element of data B12 2 having a deltaR = 2 flow variation and a deltaD = 2 quality variation. Finally, the set 63 is formed of the data element Bo, i and its deltaR = 1 and deltaD quality flow variations. = 1.

  We thus have a hierarchical structure: the two sets 61 and 62 are at the same hierarchical level N = 1 and dependent (611, 621) of the set 63 at the directly lower hierarchical level N = 0, of lower temporal frequency. The dependency relationships make it possible to know that the quality increment of Bo ,, (q) must appear before the quality increments B1.2 (q) and B1,1 (q) of the two child data elements.

  It remains to determine which of these two son data elements should appear first, so as to optimize the bit rate and the quality of the restitution of the images. In relation to FIG. 6, the evolution of the flow rate as a function of the distortion for each data element 61, 62, 63 is studied. The abscissa represents the evolution of the flow and the ordinate axis. evolution of quality. The two possible strategies concerning the order of appearance of the efficiency information are illustrated by the envelopes 71 and 72.

  We first show the quality increment of Bo, l. Strategy 1 consists of then displaying B1,2 and then Strategy 2 consists conversely of first showing B22 2 then B1,2. The envelope 71 of the strategy 1 is convex. This strategy is then adapted for scheduling.

  The optimal scheduling is the scheduling that corresponds to the upper convex envelope of the different compatible scheduling strategies. 7.6 General algorithm for scheduling quality increments The scheduling method developed in the previous section concerns a single hierarchical time level, for a single quantization level q. We now present the algorithm of the invention which makes it possible to find the general scheduling and to find the upper convex envelope, as a function of the efficiency information obtained for each data element, generalized to all the levels. hierarchical and the set of quantification levels. In relation to FIG. 7, two initialization steps 811 and 812 start at the first quantization level q = 1, and at the highest temporal hierarchical level of time n = N (we speak of the maximum level for the hierarchical level time of maximum frequency).

  There follows a step 82 of initialization of the lists of this level. By level list is meant the process detailed in the preceding paragraph which orders the appearance of the elements contained in the lists LOX, y (q) of each parent reference element. The following step 83 retrieves, for each increment element Bn, i of level n, the ordered reference lists LOfn, fi (q) associated with the elements Bf ,,, f1, whose Bpi is the data element of parent reference. In this same step, the set of ordered referencing lists is then merged so as to nest them in each other and thus optimize the general scheduling for a given hierarchical level. The scheduling list merge algorithm is detailed in the following section. A test 84 follows step 83 to check whether the last hierarchical time level is reached. If not, the hierarchical level index n is decremented 85 and steps 82 and 83 are repeated. If the test 84 is positive, a second test 86 is performed to check whether the current processing is performed on the last quantization level. Otherwise, the quantization level is incremented 87 and all of the steps 812, 82, 83, 84, 85 and 86 are reiterated. If so, a final step 88 of melting the lists obtained for each quantization level is performed. At the end of steps 82 to 86 repeated for each quantization level, the algorithm generated for each quantization level q an ordered list LO (q) of the quality increments relative to the coding level q. Step 88 then concatenates the lists in order of coding level in order to obtain a final list LO regrouping the different levels of coding quantization. 7.7 Algorithm for Merging Ordered Lists In step 83 of FIG. 7, a merge operation of the ordered lists of each element dependent on the parent reference elements is performed. Here we detail the algorithm implemented, always in order to obtain a convex envelope. Consider the quality increment element Bn, i (q). It comprises an ordered indexing list, composed of quality elements depending on Bn, (q), and each of which is associated with an ordered list of references LOfn, fi (q). FIGS. 8a, 8b, 8c and 8d illustrate by an example the evolution of the merger of reference lists.

  The quality variation with respect to the rate variation of a parent reference element Bn, i is considered in relation to FIG. This element is a parent of two elements Bfn ,,,;, and Bfn2 2 each of which is associated with a scheduling list (which means that they themselves are each parent reference element of other data elements), illustrated respectively in Figures 8b and 8c. These listing lists are determined according to the procedure detailed in section 7.5. Figure 8d shows the result of the scheduling of the two merged lists. The parent reference element of slope a is first inserted. We then choose among the first elements of each list on which depends (Figures 8b and 8c), the one with the strongest slope. This element corresponds to the slope element b1 in this particular example. It is then the slope element and has the largest slope. The algorithm continues its progression by juxtaposing the elements of slope less and less strong. The final list is illustrated in Figure 8d and has a convex hull. The elements of the two referencing lists (of FIGS. 8b and 8c) have been progressively introduced by taking into account the slopes of the associated convex envelope. This algorithm for merging the ordered lists of dependent increments is now generalized, in relation to FIG. 9. A quality increment element Bn,; (q) is associated with a series of lists LOfn, i (q) of list scheduling, each associated with the quality increment elements dependent on Bn, (q). The progress of the algorithm is as follows: Step 1: the slopes of the convex envelope for each point of the different ordered lists LOf, fn (q) are calculated and obtained; Step 2: we effectively merge the lists; Step 2a: initialize the output list LOn, (q) by taking the element B, u (q), which is the parent reference element; Step 2b: As long as all the lists LOfn, f, (q) are not empty: 2b i: We look for the element B of quality increment at the top of one of the lists LOfn, f; (q) having the largest slope, named B; 2b ii: we then remove this element B from the list where it was located. The first element of this list then becomes the element that followed B previously in this list; 2b iii: Add B at the end of the output list LO ,,,; (q) (corresponding to the merged list of the element B ,,,; (q)). An illustration of the result of this fusion algorithm is presented in Figure 9. At point PO is associated the slope PEO. At points P1 and P2 is associated slope PT1. Points PT2 and M are respectively associated with points 3 and 4. Note that only one slope PT1 is associated with all the points P1 and P2: we are in a situation where all the first elements of the lists concerned have a Slope lower than the next element. To avoid losing the convexity of the envelope, we merge the two slopes into one so as to maintain the convexity. In a variant of this algorithm, it is possible, for each quantization level q, to order the increments of qualities according to their hierarchical level value. 7.8 Implementation Devices The method of the invention can be implemented in numerous devices, such as stream servers, intermediate nodes of a network, transmitters, data storage devices, etc. The simplified general structure of such a device is illustrated schematically in FIG. 10A. It comprises a memory M 100, a processing unit 101, equipped for example with a microprocessor, and controlled by the computer program Pg 102. At initialization, the code instructions of the computer program 102 are for example loaded into a RAM memory before being executed by the processor of the processing unit 101. The processing unit 101 receives as input a data stream to be ordered 103. The microprocessor P of the processing unit 101 implements the process described above according to the instructions of the program Pg 102. The processing unit 101 outputs an ordered data stream and / or an ordered list 104.

  The invention also relates to the receivers of an ordered stream according to the invention, whose simplified overall structure is illustrated schematically in FIG. It comprises a memory M 110, a processing unit 111, equipped for example with a microprocessor, and controlled by the computer program Pg 112. At initialization, the code instructions of the computer program 112 are for example loaded into a RAM before being executed by the processor of the processing unit 111. The processing unit 111 receives as input an ordered data stream and / or an ordered list 113. The microprocessor tP of the unit processing 111 implements the method according to the instructions of the program Pg 112, to deliver a video of optimized quality 114, depending on the capabilities of the receiver, and if necessary the transmission network.

Claims (14)

  A method of scheduling data units (51) of a scalable video stream at least in time and quantization dimensions, obtained by hierarchical coding, said data units being organized in successive layers comprising a base layer time and at least one temporal enhancement layer in said time dimension and a quantization base layer and at least one quantization refinement layer in said quantization dimension, each data unit of a given temporal enhancement layer being encoded taking into account at least one reference data unit of a lower temporal layer, characterized in that it comprises a scheduling step delivering an ordered list of said data units comprising, for a set of data units to order, the following steps: introduction into said ordered list of data units of the couc he basic quantization; associating (53), with each data unit of a quantization refinement layer, information of efficiency indicative of an increase in quality and / or an increase in bit rate provided by said data unit; scheduling (54) said quantization refinement layer data units based on said effectiveness information and a presence criterion, indicating that said reference data units of said data unit are already present in said quantization refinement layer data units; ordered list.   2. scheduling method according to claim 1, characterized in that it comprises a step (52) for defining dependency relationships between said data units, so as to control said presence criterion.   3. A scheduling method according to claim 2, characterized in that said dependency relationship definition step (52) comprises an identification step, for each data unit of a time enhancement layer, of a unit single reference data unit, referred to as the parent reference data unit, selected such that the presence of said parent reference data unit in said ordered list ensures that all of said reference data units are present therein.   4. A scheduling method according to claim 3, characterized in that said reference data unit father of a data unit of an n-level temporal enhancement layer is the reference data unit presenting the index. the highest, respectively the lowest, of the set of reference data units of the n-1 level temporal enhancement layer.   5. scheduling method according to any one of claims 1 to 4, characterized in that said data units (51) are encoded using a hierarchical coding implementing a temporal transformation delivering images coded in bidirectional mode in closed loop.   The scheduling method according to claim 5, characterized in that said data units (51) comprise: Intra images, referred to as I frames; predicted images in monodirectional mode, called P-pictures; predicted images in bidirectional mode, called B images.   7. A scheduling method according to claim 6, characterized in that said scheduling step consists in selecting, quantization refinement layer by quantization refinement layer, starting with said base layer, all of said images I , then all of said images P and finally all of said images B ordered by temporal enhancement layers.   8. scheduling method according to any one of claims 1 to 7, characterized in that said scheduling step simultaneously processes groups of at least two data units, so as to take into account the sum of said information efficiency and favor one group over another if it is more effective overall.   The scheduling method according to claim 8, characterized in that said scheduling step selects said data unit groups of S such that a representative bit rate / quality curve of said path corresponds to a convex hull of the set. different qualifying hierarchical paths.   10. The scheduling method as claimed in claim 9, characterized in that said scheduling step comprises, for each temporal uplift layer n, a step of merging ordered sub-lists of the dependent data units, for each unit of data of said temporal enhancement layer n, then a step of merging the ordered lists of the set of quantization refinement layers. 15   11. A scheduling method according to any one of claims 1 to 10, characterized in that it comprises a step of insertion, in a signal comprising said data units, of at least one piece of information representative of said ordered list. .   A device for transmitting and / or storing a scalable video stream at least in time and quantization dimensions, obtained by hierarchical coding, formed of data units organized in successive layers comprising a time base layer and at least one temporal enhancement layer in said temporal dimension and a quantization base layer and at least one quantization refinement layer in said quantization dimension, each data unit of a given temporal enhancement layer being encoded taking into account at least one reference data unit of a lower temporal layer, characterized in that it comprises scheduling means defining an ordered list of said data units and comprising, for a set of data units to order: means for introducing into said ordered list of data units of the base layer of qua ntification; means for associating each data unit of a quantization refinement layer with information of efficiency representative of an increase in quality and / or a rate increase provided by said data unit; means for scheduling said quantization refinement layer data units based on said effectiveness information and a presence criterion, indicating that said reference data units of the given data unit are already present in said quantization refinement layer data units; ordered list.   13. Computer program product downloadable from a communication network and / or stored on a computer readable medium and / or executable by a microprocessor, characterized in that it comprises program code instructions for the implementation of the scheduling method according to at least one of claims 1 to 11.   14. Terminal for receiving data units representative of a video stream ordered according to the method of at least one of claims 1 to 11, characterized in that it comprises means for selecting a part of said data units. within said stream, from said ordered list and according to at least one criterion of bit rate, spatial resolution and / or display frequency, and means for reconstructing a video stream adapted to said criterion with the aid of said selected data units.