FR2930059A1 - Digital video sequence processing method for e.g. computer, involves selecting location for writing pixel values of given memory block among set of locations of cache memory using labels of memory blocks stored at locations - Google Patents

Abstract

The method involves receiving a field of movement vectors, where each vector associates a pixel macroblock of one image to a memory block assembly of another image. Pixel values of memory blocks of the assembly are read using a cache memory. A count of number of macroblocks is stored in a label corresponding to the memory block. A location for writing pixel values of the given memory block is selected among a set of locations (38-1p-38-4p) of the cache memory using the labels of the memory blocks stored at the locations, during reading of the pixel values of the given memory block. Independent claims are also included for the following: (1) a computer program comprising instructions for implementing a video sequence processing method (2) a device for processing a video sequence of images sub-divided into pixel macroblocks (3) a memory comprising a set of memory zones.

Description

PROCESSING OF A VIDEO SEQUENCE IN A CACHE MEMORY

The invention relates to the processing of a digital video sequence using a cache memory, for example for encoding purposes. The image processing methods can use a segmentation of the images into macroblocks of, for example, 16x16 pixels. Typically, it is known to search for motion vectors associating a macroblock of a first image with a set of pixels of a second image. The values of these pixels can be read for a variety of purposes. For example, within the framework of an encoding method, it is possible to seek to determine motion vectors obtained from sub-sampled images, called lower level images, in order to orient the choice of motion vectors for the higher-level image macroblocks, as illustrated in figure 1. For a given lower-level macroblock B1 N, we can choose as candidate vector for the higher-level macroblock B11 N-1 the motion vector V1, N_1 starting from the center of macroblock B11 and of length determined by multiplying the coefficients of motion vector V1, N corresponding to macroblock B1 by respective constants. These constants correspond to the sub-sampling carried out. In the example shown in Figure 1, the level N image is obtained by low-pass filtering the video signal in the horizontal and vertical directions, then removing every other pixel in the horizontal direction, and one pixel in the horizontal direction. every second pixel in the vertical direction, so the constants are 2 horizontally and 2 vertically. The candidate vector V1, N_1 defines a set of pixels, the values of which are read for comparison purposes with the pixels of the higher level macroblock B11. A definitive motion vector is estimated from the results of these comparisons. Similarly, the vector V1, N makes it possible to determine the candidate vectors and the sets of pixels for the blocks B12, B13 and B14.

According to another example, one can seek to determine motion vectors for decision-making purposes as to the types of the macroblocks, for example type I (like Intra in English) or type P (like Predicted in English).

The pixel values to be read are stored in a block memory, structured in memory areas. Each memory area is able to store the pixel values of a memory block of the second image. To read a value of a given pixel, it may be necessary to read all the pixel values of the pixels of the memory block to which the given pixel belongs. Also, for a given macroblock, the set of pixels associated with this macroblock by the corresponding motion vector corresponds in fact to a set of memory blocks. When reading the pixel values associated with a given macroblock, it is therefore necessary to read all the pixel values of the set of memory blocks corresponding to these pixel values. Reading the pixel values of memory blocks from block memory can take a relatively long time. It is known to use a cache memory in order to speed up read requests. When the pixel values of a given memory block are read for the first time from the block memory, these values can be copied back to the cache memory, in order to allow a later reading of these values more quickly. However, since cache memories are relatively expensive, the size of these memories is generally limited. It is therefore desirable to optimize the choice of memory blocks to be kept in the cache memory, in order to allow overall faster processing. According to a first aspect, the subject of the invention is a method of processing a video sequence of images subdivided into macroblocks of pixels, in which: a field of motion vectors is received, each motion vector associating by addressing a macroblock of pixels from a first image to a set of memory blocks of a second image, this set comprising at least one memory block, 3 - for at least one macroblock, the pixel values of the memory blocks of the set of Associated memory blocks are read using a cache memory. The method comprises the steps of: - for at least one memory block of the second image, storing in a tag corresponding to this memory block a count of the number of macroblocks whose sets of associated memory blocks include this memory block, and - when reading the pixel values of a given memory block in a block memory separate from the cache memory, deciding on the writing location of these pixel values of this given memory block, from among a set of locations of the cache memory corresponding to this given memory block, from the labels of the memory blocks stored at these locations.

Thus, the information provided by the vector field is used in order to better manage the writings of the cache memory, and this in a manner that is relatively simple to implement. Taking into account the number of macroblocks for which a memory block is requested allows a relative optimization of the management of the cache memory. The process is relatively quick to implement. Macroblocks and memory blocks can be 16x16 pixels, 8x8 pixels, 16x32 pixels, 32x32 pixels, or the like. Macroblocks may or may not have the same dimensions as memory blocks.

The reading of the pixel values of the blocks associated with a macroblock can be performed for each macroblock, or alternatively for only part of the macroblocks. The storage of the counting of the number of macroblocks corresponding to a memory block can be carried out for each memory block, or alternatively for only part of the memory blocks, for example those for which this value exceeds or is below a certain threshold.

The cache memory can for example comprise an MRAM (standing for Magnetic Random Access Memory), or the like. The block memory can include SDRAM (Synchronous Dynamic Random Access Memory), or the like.

The pixel values can for example include RGB component values (Red Green Blue), YUV component values, or the like. Each motion vector can directly associate a macroblock with a set of memory blocks.

Alternatively, each motion vector can associate a set of pixels with a macroblock, the motion vectors thus being relatively precise. Each set of pixels can correspond to a set of memory blocks, the memory storing the pixel values being structured in memory blocks.

The set of memory blocks may include one or more memory block (s). Advantageously, for at least one memory block, the count value is decremented on each read request for the pixel values of this memory block. Thus, when choosing the location of the cache memory, account is taken of the number of macroblocks to be processed, for which a reading of the pixel values of the memory block will be necessary, and not of the macroblocks for which this reading has already been carried out. This decrementation can be carried out for all the memory blocks, or for only a part of them.

Alternatively, the count value is not decremented, so that this value remains constant during the processing of an image. Advantageously, for each memory block, an identifier of a macroblock of which the associated set of memory blocks comprises this memory block is stored in the corresponding label. Thus, the location of the macroblock in the first image is taken into account in order to manage the cache memory. For example, during the processing of macroblocks at the top left of the first image, it is possible to choose to overwrite in the cache memory the pixel values of memory blocks associated with a macroblock at the bottom right of the first image. Thus, the most relevant macroblock among the macroblocks associated with a given memory block would be the next macroblock, according to a macroblock processing order, to request the given memory block. When processing a macroblock at the top left of the first image, we can choose to keep in the cache memory the pixel values of memory blocks which we know will soon be requested by macroblocks also at the top left of The first picture.

Advantageously, for each memory block, the macroblock whose identifier is stored is the last macroblock, according to a processing order of the macroblocks, of which the associated set of memory blocks comprises this memory block. The choice of the last macroblock offers a compromise between simplicity of implementation of the method and relevance of the macroblock whose identifier is stored. The invention is in no way limited by these improvements. Provision can be made to be satisfied with counting, for each memory block, the number of macroblocks requesting this memory block. Advantageously, the method further comprises the steps of estimating a motion coherence parameter from the received motion vector field, choosing a degree of associativity of the cache memory from the estimated value of the motion coherence parameter , and assigning to each memory block a number of locations of the cache memory equal to the chosen degree of associativity.

Thus, the degree of associativity of the cache memory is chosen as a function of the consistency of the vector field. For example, if the displacements between two images are relatively inconsistent, one can choose a degree of associativity equal to 1 or 2. When writing pixel values of a given memory block, the number of locations of the set of locations corresponding to said given memory block is relatively small. Indeed, the displacements being relatively little correlated with each other, it is more efficient to have a relatively low associativity. If the 6 movements are not correlated at all with each other, a degree of associativity of 1 may be the most effective. The coherence parameter can be a variance of the vector field for example, or else another parameter. The variance can be obtained from the orientations of the vectors and / or the moduli of the vectors. The method according to one aspect of the invention finds various applications. For example, the step of reading the pixel values of one or more memory block (s) associated with a given macroblock can be followed by a step of finding a definitive motion vector for this given macroblock . In this case, the motion vector received for the given macroblock may be a candidate vector indicating a set of pixels from which to determine the final motion vector. In particular, provision can be made to determine, prior to the step of receiving the field of vectors, a field of preliminary vectors each associating a macroblock of pixels of an image obtained by sub-sampling of the first image to pixels of an image obtained by sub-sampling the second image. The first and second images may themselves be obtained by sub-sampling (s) of two images of the video sequence, or else they may be two images of the video sequence. The field of vectors is estimated from the field of preliminary vectors, for example by multiplying the coefficients of the preliminary vectors by respective constants, the values of these constants depending on the subsampling carried out. The vector field is thus obtained in a hierarchical manner. The pixel values of the memory blocks read can in particular be used to search for a definitive motion vector, the vectors of the vector field being considered as candidate vectors. The method according to one aspect of the invention can thus be used in a method of hierarchical search for motion vectors.

According to another example, the first and the second image being two images of the video sequence, provision can be made for each macroblock, following the reading step, the pixel values of the macroblock to be compared with at least part of the values. of pixels read. From the result of the comparison, a decision can be made as to the encoding type for this macroblock, for example an I encoding type such as Intra or P such as Predicted in English. The invention is of course not limited by these applications. According to another aspect, the subject of the invention is a computer program intended to be stored in a memory of a video sequence processing device, and / or stored on a memory medium intended to cooperate with a reader of said device and / or downloaded via a telecommunications network. This program includes instructions for implementing the method according to one aspect of the invention, when these instructions are executed by a processor of the video sequence processing device. According to yet another aspect, the subject of the invention is a device for processing a video sequence of images subdivided into macroblocks of pixels, comprising means for estimating a field of motion vectors, each motion vector associating by addressing a macroblock of pixels from a first image to a set of memory blocks of a second image, this set comprising at least one memory block. Access means make it possible, for at least one macroblock, to read the pixel values of the memory blocks of the set of memory blocks associated with this macroblock, these access means comprising a cache memory. For at least one memory block of the second image, a label corresponding to this memory block is provided. This tag is intended to store a count of the number of macroblocks whose sets of associated memory blocks include this memory block. Calculation means make it possible, during a reading of the pixel values of a given memory block in a block memory distinct from the cache memory, to decide on the location among a set of locations in the cache memory corresponding to this given memory block, for writing these pixel values of this given memory block, these calculation means using the contents of the labels of the memory blocks stored at these locations. This device can be, for example, a computer, a server, a terminal, a video encoding device, or the like.

According to yet another aspect, the invention relates to a memory comprising a plurality of memory areas, each memory area comprising a data field, able to store the pixel values of a memory block of an image, and a label . This label makes it possible to store a number of macroblocks of another image associated by respective motion vectors with the memory block whose pixel values are stored in the data field. This memory can be a block memory or else a cache memory. In the latter case, the memory areas can be called cache lines.

Other features and advantages of the present invention will appear in the description below. FIG. 1, already commented on, illustrates an example of a method for estimating hierarchical motion vectors. FIG. 2 illustrates an example of a motion vector associating a macroblock with a set of memory blocks. FIG. 3 shows an example of a video sequence processing device according to one embodiment of the invention. FIGS. 4A and 4B show an example of a flowchart according to an embodiment of the invention. Fig. 5 shows an example of a flowchart according to another embodiment of the invention. Identical references are used to designate identical or similar objects. Represented in FIG. 2 is an image portion 22. The values of the pixels of image 22 are stored in a block memory (not shown in FIG. 2), structured in zones. Memory blocks of the image are represented under the reference B i, the values 9 of the pixels of the same memory block being stored in the same zone of the block memory. A motion vector V makes it possible to associate a macroblock of another image MB ,,,, shown here in dotted lines, with a set of pixels 20.

To read the pixel values of this set of pixels, it is necessary to read the pixel values of the set 21, due to the block access to the block memory. For example, all the pixels of the block B; o, io must be read, even if only a part of these pixels are used subsequently. The motion vector V therefore associates a set of blocks 21 of one image with a macroblock MB ,,, of another image. FIG. 3 shows an example of device 34 according to one embodiment of the invention. This device can for example comprise or form part of a video encoding server. This video sequence processing device 34 comprises a block memory 32 structured in memory areas 351, 352, ..., 35p, ... This memory 32 may for example be a RAM memory (standing for Random Access Memory). , SDRAM, or other. Each zone 351, 352, ..., 35p, ... is able to store the pixel values of a memory block, and for this purpose comprises a data field 371, 372, ..., 37p ,, ... Each data field 371, 372, ..., 37p ,, ... makes it possible to store the values of the pixels data_1_1, data_1_2, ..., data_io_jo, ... corresponding respectively to memory blocks. For example, the data field 37p is intended to store the pixel values data_io_jo of the memory block B; o, io in FIG. 2.

Each memory area 351, 352, ..., 35p, ... also includes a label 361, 362, ..., 36p, ... for storing an address & 1,1, & 1,2, ..., of the corresponding block, a count value C1,1, C122, ..., and an identifier MBdm (1'1), MBdrr, (12) MBdm (io, io) The address & 1,1, & 1,2, ..., can for example comprise the values of two indices i, j making it possible to identify the pixels of the block corresponding to the respective memory zone, or any other address making it possible to identify the pixels of the block. For a given memory area 35p, the count value Cio, io is a number of macroblocks associated with sets of memory blocks comprising the memory block B; o, io corresponding to this area 35p. These count values can be determined by going through the macroblocks, for example from left to right and from bottom to top: for each macroblock, the count values of the memory blocks associated with this macroblock are incremented by 1 by the corresponding motion vector. During processing, when readings of the pixel values of the memory blocks are taken, the count value corresponding to a given memory block is decremented on each access to the pixel values of that given memory block, regardless of whether the reading is carried out on the cache memory or block memory. The count value Cio, io therefore gives an indication of the number of readings of the memory area 35p to be expected. Thus, when all the necessary accesses to a given memory block are performed, the count value of this macroblock is zero. For a given memory zone 35p, the identifier MBdmr '' °) comprises an identifier, for example an address, of the last macroblock, according to the order of processing of the macroblocks, associated with the memory block B; o, io corresponding to this zone 35p. The order of processing of the macroblocks can for example be from bottom to top, and from left to right. The macroblock address can comprise two indices of macroblocks for example. The device 34 further comprises a cache memory 31. In this example, the cache memory 31 is K times associative, K being in this example equal to 4. The cache memory 31 therefore comprises K = 4 sub-caches 301, 302, 303, 304, and to each memory area 35p of the block memory 32 correspond K = 4 areas or locations 381, p, 382, p, 383, p, 384, p of the cache memory 31. Each of these locations 381, p, 382, p, 383, p, 384, p, also called cache lines, belongs to a respective sub-cache 301, 302, 303, 304.

The content of a given memory area 35p can be written in 4 corresponding cache lines 381, p, 382, p, 383, p, 384, p. Furthermore, each cache line corresponds to more than one memory area. Typically, each cache line 381, p, 382, p, 383, p, 384, p corresponds to memory areas 35p_4, 35p, 35p + 4, 35p + 8, etc. that is to say to the 35p mod (K) memory areas. Each cache line 381, p, 382, p, 383, p, 384, p has a data field 391, p, 392, p, 393, p, 394, p and a tag 401, p, 402, p, 403, p, 404, p.

Each data field 391, p, 392, p, 393, p, 394, p is intended to store pixel values data_ii ji, data_i2j2, data_i3j3, data_i4j4 of a memory block also stored in block memory 32. Each tag 401, p, 402, p, 403, p, 404, p is intended to store: - an identifier or an address & il, n, & j2p, & j3,] 3, & j4j4 of the memory block whose pixel values are stored in the data field 391, p, 392, p, 393, p, 394, p corresponding, - a number of macroblocks C; l, il, C; 2, i2, C; 3p, Ci4, i4 associated with this memory block , and - an identifier of the last macroblock MBdm ('' '' '), MBdm02' 2>, MBdm (13''3), MBdm ('4''4) according to the macroblock processing order, associated with this memory block. During processing, when a reading of the pixel values of a memory block is performed on the cache memory, the number of macroblocks stored in the cache memory and associated with this given memory block is decremented. If the reading is performed on the block memory, the value of the number of macroblocks copied into the cache memory is also decremented so as to take this reading into account. The device further comprises processing means 33, for example a processor, a DSP (standing for Digital Signal Processor) or the like. The processor 33 comprises means for estimating vector fields making it possible to estimate a field of motion vectors. Each vector associates a macroblock of a first image with a set of memory blocks of a second image, as illustrated by FIG. 2.

For a given macroblock MBm, the processor 33 reads the pixel values of the memory blocks B; o, io, Bi0 + 4, i0 + 3 in the block memory 32, using access means, here two buses 41, 42 and cache 31.

When the cache memory 31 receives via the bus 41 a request to read the pixels of a memory area 35p, the cache memory checks whether these pixel values are stored in the cache memory. Where appropriate, the pixel values are supplied to the processor 33 (cache hit or cache hit in English). If, on the contrary, the cache memory 31 does not have these pixel values (cache default or cache miss in English), the cache memory 31 reads the area 35p of the block memory 32, via the bus 42. The content of zone 35p is then copied into cache memory 31, and also transmitted to processor 33.

Now, the cache being K = 4 times associative, the content of a given memory area 35p can be written in K = 4 corresponding cache lines 381, p, 382, p, 383, p, 384, p. Computing means, here integrated into the processor 33, decide which cache line to choose from among these K = 4 cache lines. The processor uses for this purpose the contents of the labels of the cache lines 381, p, 382, p, 383, p, 384, p. For each cache line 38k, p, k running from 1 to K, a local usage density is calculated: Clk Jk A + diSt (MBm, MBd / ') Where A is a strictly positive constant, for example A = 1 .

The estimated distance dist (MB, MBdm'k) between the current macroblock MB ,,, and the macroblock whose identifier MBdm'k is stored in the label 40k, p can be any distance, for example a distance Euclidean. We will preferably choose a relatively quick distance to estimate, such as a distance from Manhattan.

The distance can be estimated from block coordinates, or pixel coordinates for example. Thus, the local usage density Dk for a cache line 40k, p increases with the number Clk,% k of macroblocks requesting the corresponding memory block, and decreases with the distance between the current macroblock MB ,,, and the macroblock whose MBdm'k identifier is stored in the 40k tag, p. Dk = 13 For a 40k cache line, p, the local usage density therefore gives an indication of the probability that the content of this cache line will soon be read by the processor 33. The processor 33 chooses to write the content of the given memory area 35p in the cache line 40k, p for which the use density Dk is the lowest. This process makes it possible to preserve the useful data in the cache memory in the near future, and thus to significantly reduce the number of cache defects compared to an LRU (standing for Last Recently Used) cache management, for the same sequence. video. FIGS. 4A and 4B illustrate an example of a video sequence processing method according to one embodiment of the invention. The algorithm represented is suitable for a pair of images, but it is of course adaptable to a video sequence comprising any number of images. Two images Im_1, Im_2 of a video sequence are received during a step 400. These two images can follow one another immediately, or else be separated from each other by one or more images in the video sequence.

These two images are stored as zero level images Im_1 (°), Im_2 (°), during a step 401. Then these images are subjected to successive sub-samplings. A loop indexed by the variable I is executed, with the classic steps of initialization, test and incrementation. During each loop, two level I images Im_1 (1), Im_2WWW are obtained by sub-sampling of two respective images Im_1 (1-1), Im_20-1) themselves obtained in the previous loop. After obtaining a predetermined number L of pairs of sub-sampled images, a search is carried out for fields of motion vectors {Vm, Lfinall between the two most sub-sampled images Im_1 (L), Im_2 (L) , during a step 403. The variable m indexes the macroblocks of an image Im_1 (1).

This search can be carried out using a conventional method of searching for motion vectors. For example, for each macroblock MB ,,, of the image Im_1 (L), the pixels of this macroblock are compared with the pixels of a search zone of the image Im_2 (L). The search zone can for example be chosen around the location corresponding to the macroblock in the image Im_2 (L). These comparisons lead to selecting a group of pixels of the image Im_2 (L) as having values relatively close to the values of the pixels of the MB ,,, of the image Im_1 (L) We choose as the motion vector Vm, Lfinal the vector associating the MBm macroblock with this group of pixels. This set of motion vectors between the two most under-sampled images Im_1 (L), Im_2 (L) is then used to estimate sets of motion vectors between two images Im_10), Im_20), the index I varying between L-1 and O. Also a loop on this index I is set up, with classical steps of initialization, test and decrementation. For two images Im_10), Im_20) of level I given, one estimates during a step 404 a field of motion vectors {Vm, i} from the set of definitive motion vectors for the images of level 1 + 1 Im_10 +1) and Im_20 + 1). These final motion vectors Vm, 1 + 1fina 'for level 1 + 1 images are said to be preliminary during the processing of level 1 images. As illustrated in Figure 1, a vector Vm, i can be obtained by multiplying by a constant a vector obtained for the images Im_1 0+ '), Im_20 + 1). If, for example, each sub-sampling of step 402 is a sub-sampling by a factor of 2, this constant is 2. It will be noted that the notation f0, used several times in FIGS. 4A, 4B and 5, does not correspond to a particular function, but serves only to indicate that certain data is obtained from other data.

During a step 450, a set of definitive motion vectors {Vm, i final} is determined from the set of motion vectors {Vm, i}. This step is detailed in Figure 4B.

The search for motion vectors is thus carried out in a hierarchical manner, by loops indexed by I. During the execution of the last loop, I = 0, and the images for which the field of final vectors is estimated are the images of the video sequence Im_1, Im_2.

This vector field can be transmitted to a video encoding module for example, and / or to a block type decision module, during a step 411. FIG. 4B shows in more detail the method making it possible to determine set of final motion vectors {Vm, I final} of step 450.

From the set of motion vectors {Vm ,,}, a set of count values {C;, i} is determined during a step 452. Each count value Cz, j corresponds to a memory block Bq and indicates a number of gg

macroblocks whose sets of associated memory blocks include this memory block Bq. The count values C i, i are stored in the block memory. Also, for each memory block Bq, an identifier of the last macroblock MBdm'j ') is stored during a step 453, according to a processing order of the macroblocks, of which the associated set of memory blocks includes this memory block. 20 There follows a loop going through the different macroblocks of the image Im_1 (1). This loop is indexed by m varying between 1 and M (I), the number M (') of macroblocks of an image being variable according to the level of this image. This loop comprises the classic steps of initialization, testing and incrementation. Provision can be made for the variable m to make it possible to browse all the macroblocks of the image Im_10), for example from bottom to top and from left to right. For a given macroblock MBm, during a step 412, a set of memory blocks {Bq} associated with this given macroblock MBm is determined from the corresponding motion vector Vm. The variable q indexes the memory blocks of this set. This index varies between 1 and a value e1) this value Q (mI) being liable to vary according to the associated set.

For each memory block Bq, a reading of the values of the pixels of this memory block Bq is performed. A loop 405 is therefore executed going through the various memory blocks of the set {Bq}, with the conventional steps of initialization, test, and incrementation.

A test 406 is first carried out as regards the storage of the values of the pixels of the memory block Bq in a cache memory. If these values are already cached, the read is fast. Otherwise, these values are read in a corresponding area of a RAM memory. Then the read values are copied into the cache memory, in anticipation of a next reading of these values. In the case of a K times associative cache memory, these values are likely to be copied into K cache locations. To determine which cache location to choose, a local use density value Dk is calculated for each cache location, during a step 407. A loop is performed to browse the K cache locations corresponding to the memory block Bq. This loop is indexed by k varying between 1 and K and comprises the conventional steps of initialization, incrementation and test.

Then, during a step 408, the location, indexed ko, whose local use density value Dk is the lowest is selected. The pixel values of the memory block Bq are copied into this cache location, from the block memory, during a step 409. An address of this memory block Bq, the macroblock identifier MBdm, is also copied onto the cache memory. 'j') and the count value Cz corresponding to this memory block Bq. gg

Whatever the results of test 406, that is to say that the reading of the pixel values of the memory block Bq has been carried out on the cache memory or on the block memory, the count value CZ 1 corresponding to gg

to the memory block Bq is decremented during a step 451. This decrementation affects both the count value stored in the block memory and the count value stored in the cache memory. After reading all the pixel values of the set of blocks {Bq}, comparisons are made between the pixel values of the macroblock MB ,,, and pixel values corresponding to parts of the set of blocks {Bq }. These comparisons make it possible to determine a definitive motion vector Vm'lfinal for this macroblock MBm of the image Im_1 WWW, during a step 410. The final motion vectors of the other macroblocks of the image Im_1 are determined in the same way. (1). As explained above with reference to FIG. 4A, this set of definitive motion vectors {Vm, lfinal} for the images Im_1 0W, Im_2 ~ 1 ~ makes it possible to determine during step 404 a set of motion vectors {N / m ,, _ 1} for the images of level I-1 Im_1 0- ') During the execution of the last loop, I = 0, and the images for which the final vector field is estimated are the images of the video footage Im_1, Im_2. This definitive vector field {Vm, ofinal} can be transmitted to a video encoding module for example, and / or to a block type decision module, during step 411. The type decision module block can also implement a method according to an embodiment of the invention. FIG. 5 shows an example of a flowchart capable of being executed by such a block type decision module.

This module, the calculation means, the vector field estimation means, etc. can be integrated in the same processor or else in separate processors. A step 500 of receiving a field of motion vectors {Vm, ofinal} is implemented.

In this example, this vector field is received from the means for estimating vectors implementing a method according to one embodiment of the invention, but it can of course be otherwise. For example, the field of motion vectors may have been determined in a manner known to those skilled in the art, for example by testing for each macroblock the vectors in all directions, or else the vectors around a vector determined for one. previous macroblock of the image. Also received and stored is a set of count values {Ci,} during a step 501 and a set of last macroblock identifiers {MBdm ')} during a step 502. Each count value C1 corresponds to a memory block Bq and gg indicates a number of macroblocks whose sets of associated memory blocks include this memory block Bq.

For each memory block Bq, the corresponding macroblock identifier MBdm'j ') indicates the last macroblock, according to a macroblock processing order, of which the associated set of memory blocks includes this memory block. The set of initial count values and / or the set of identifiers of the last macroblock {MBdm ')} may have been determined previously, for example in a step similar to step 452 and / or 453 of FIG. 4B, and this (these) same set (s) is or are used by the block type decision module. In this case, provision can be made for copies of these count values to be decremented on each reading of the corresponding block 20, during a step similar to step 451 of FIG. 4B. Alternatively, the set of count values {Ci, i} and / or the set of last macroblock identifiers {MBdm ')} are determined by the block type decision module, from the motion vector field { Vm, final o}.

There follows a step 509 of determining a degree of associativity K ′ of a cache memory possibly distinct from the cache memory described with reference to FIGS. 4A and 4B. A variance a of the field of motion vectors {vm, Ofinal} is first estimated during a step 503. This variance a is compared with thresholds S1, S2, during test steps 504, 507, and the value assigned to the variable K 'depends on the result of these tests (see steps 505, 506, 508).

19 Thus, if the variance a is small, which would be the case if the vectors of the field {Vm, ofinal} are relatively similar to each other, the degree of associativity of the cache K 'is relatively high. When reading pixel values from a memory block into block memory, the calculations to determine which cache line to write these new pixel values to are relatively long, due to the high value of the number K 'of. cache lines likely to accommodate these new values. Nevertheless, the variance being low, one can hope to read in the cache memory a good number of pixel values of memory blocks, so that these calculations would remain relatively rare. There follows a loop relating to macroblocks, and indexed by the variable m, with the classic steps of initialization, test and incrementation. For a given macroblock MBm, during a step 512, a set of memory blocks {Bq,} associated with this given macroblock MBm is determined from the corresponding motion vector Vm, o final. The variable q 'indexes the memory blocks of this set. This set may be different from the set {Bq} of FIG. 4B. The index q 'varies between 1 and a maximum value liable to vary according to the associated set. Then the pixel values of the set {Bq,} are read during a step 513. This reading uses the cache memory for which the degree of associativity K 'was determined during step 509. L step 513 can be performed relatively similarly to step 405 of Figure 4B. This reading of the pixel values makes it possible to perform various calculations, for example subpixellic interpolations, comparisons of at least part of the pixel values read with the pixel values of the macroblock MBm, or the like. These calculations, represented by step 514, make it possible to determine the type of coding to be used for this macroblock.

Claims (12)

CLAIMS 1. A method of processing a video sequence of images subdivided into macroblocks of pixels, comprising the steps of: receiving (404) a field of motion vectors ({Vm ,,}), each motion vector associating by addressing a macroblock of pixels (MBm) of a first image (Im_1 W) to a set ({Bq}) of memory blocks of a second image (Im_2W), said set comprising at least one memory block, and for at least one macroblock, reading (405) the pixel values of the memory blocks of the associated set of memory blocks, using a cache memory, characterized in that the method further comprises the steps of for at least one memory block of the second image , storing (452) in a tag corresponding to said memory block a count 15 (C;, i) of the number of macroblocks whose sets of associated memory blocks include said memory block, and when reading the pixel values of a memory block given in a block memory separate from the cache memory, deciding (408) the location for writing said pixel values of said given memory block, from among a set of cache memory locations, said set of locations corresponding to said given memory block, from the labels of memory blocks stored at said locations. 2. A method of processing a video sequence according to claim 1, the method comprising, for at least one memory block (Bq), a step of decrementing (451) the count (CZg, 1) of the number of macroblocks of which the. sets of associated memory blocks comprise said memory block, said step being performed on each read request (405) of the pixel values of said memory block. 3. A method of processing a video sequence according to one of claims 1 or 2, the method comprising a step consisting in for each memory block (B; o, jo), storing in the corresponding tag (36p) an identifier. (MBdm (10'j °)) of a macroblock whose associated set of memory blocks includes said memory block. 4. A method of processing a video sequence according to claim 3 wherein for each memory block (B; o, jo), the macroblock whose identifier (MBdm ('o'o)) is stored is the last macroblock, according to a processing order of the macroblocks, of which the associated set of memory blocks comprises said memory block. 5. A method of processing a video sequence according to one of claims 3 or 4, further comprising, during the reading of the pixel values of a given memory block (B; o, jo), the steps consisting in for each location of the set of locations (381, p, 382, p, 383, p, 384, p) corresponding to said given memory block, calculate a local use density (Dk) from the count (Cil Vil, C212 , C313, Ci414) of the number of macroblocks stored in the tag (401, p, 402, p, 403, p, 404, p) corresponding to the location, of the macroblock identifier (MBdm (i "), MBdm ('2, i2), MBdm03, j3> MBdm (i4, j4)) stored in the label, and an identifier of the macroblock (MBm) associated with the given memory block, and for which the pixel values of the block given memory are read, choose the location corresponding to the lowest usage density to write the pixel values of the given memory block. 6. A method of processing a video sequence according to claim 5, wherein, when reading the pixel values of a given memory block (Bio, io), for each location of the set of locations (381, p , 382, p, 383, p, 384, p) corresponding to said given memory block, the local usage density checks: C. '% k A + dist (MB, MBdm'k) Clk,.% K is the count of the number of macroblocks stored in the label (40k, p) corresponding to said location (38k, p), 10 A is strictly positive, and dist (MB, MBdm'k) is an estimated distance between the macroblock for which the values of pixels of the given memory block are read, and the macroblock whose identifier (MBdm'k) is stored in the label (40k, p) corresponding to said location. 15 7. A method of processing a video sequence according to any one of claims 1 to 6, further comprising the steps of estimating (503) a motion coherence parameter from the received motion vector field, choosing. (505, 506, 508) a degree of associativity (K ') of the cache memory from the estimated value of the motion consistency parameter, in which the number of cache memory locations allocated to each memory block is equal to the degree of associativity chosen. 25 8. A method of processing a video sequence according to any one of claims 1 to 7, comprising the steps of, prior to the step of receiving the vector field, determining (403, 450) a preliminary vector field ( {Vm, 1 + 1final}) 30 each preliminary vector associating a macroblock of pixels of an image Dk = where 23 obtained by subsampling the first image to pixels of an image obtained by subsampling the second image , and estimate (404) the field of vectors ({Vm, l}) from the field of preliminary vectors, and following the read step, for each macroblock, use (410) the pixel values of the memory blocks read to find a definitive motion vector (Vm, lfinal) 9. A method of processing a video sequence according to any one of claims 1 to 7, wherein the first image and the second image are two images (Im_1, Im2) of the video sequence, the method further comprising steps comprising, for each macroblock, following the reading step, comparing (514) the pixel values of the macroblock to at least a portion of the read pixel values, and making (514) a decision as to the type of encoding for said macroblock from the result of the comparison. 10. Computer program intended to be stored in a memory of a video sequence processing device, and / or stored on a memory medium intended to cooperate with a reader of said device and / or downloaded via a telecommunications network, characterized in that it comprises instructions for implementing the method according to one of the preceding claims, when said instructions are executed by a processor of said video sequence processing device. 11. A device for processing a video sequence of images subdivided into macroblocks of pixels, comprising means (33) for estimating a field of motion vectors, each motion vector associating by addressing a macroblock of pixels of a first one. image to a set of memory blocks of a second image, said set comprising at least one memory block, and access means (41, 31, 42) capable of, for at least one macroblock, reading the pixel values of the memory blocks of the set of memory blocks associated with said macroblock, said access means comprising a cache memory (31), characterized in that the device further comprises for at least one memory block of the second image, a label (361 , 362, 36p, 401, p, 402, p, 403, p, 404, p) corresponding to said memory block, said label being intended to store a count of the number of macroblocks (Cil, C1,2, C; oe), C; l, jl, Ci2, j2, C; 3p, C; 4j4) whose sets of associated memory blocks include said memory block oire, and calculation means (33) capable of, when reading the pixel values of a given memory block (35p) in a block memory separate from the cache memory, deciding the write location said pixel values of said given memory block from among a set of cache memory locations (381, p, 382, p, 383, p, 384, p) corresponding to said given memory block, said computing means using the contents of labels (401, p, 402, p, 403, p, 404, p) of the memory blocks stored at said locations. 12. Memory (31; 32) comprising a plurality of memory areas (351, 352, 35p; 381, p, 382, p, 383, p, 384, p) each memory area comprising a data field (371, 372, 37p; 391, p, 392, p, 393, p, 394, p) capable of storing the pixel values of a memory block of an image, and a label (361, 362, 36p; 401, p, 402 , p, 403, p, 404, p), said label making it possible to store a number of macroblocks of another image associated by respective motion vectors with the memory block whose pixel values are stored in the data field.