MXPA99008424A - Image compression method and device to implement this method - Google Patents

Abstract

The invention relates to an image compression method, especially of the MPEG2 type, in which the images are encoded according to the groups (GOP) each of which comprises N images with an I image encoded in intra mode, P predicted images as a function of the intra image 1 or of the preceding image P, each image P preceded and followed by n bi-directionally predicted images B, n is possibly zero. The number M = n + 1 represents the structure of the group. At least one meter which characterizes the source images which are to be coded according to a group is determined using a test coding (70), and the numbers N and M are made dependent on this pair or these parameters. In the course of the test coding, defined values are conferred to N, M and the quantization interval

Description

IMAGE COMPRESSION METHOD AND DEVICE TO IMPLEMENT THIS METHOD DESCRIPTION OF THE INVENTION The invention relates to an image compression method in which images are coded according to groups of varying lengths. It relates more particularly to a method of the MPEG type, particularly of the MPEG2 type. Although the invention is not limited to this standard, it will be referred to mainly in the remainder of the description. The principle of such compression is reiterated below. In the MPEG2 video standard, compression of the digital video signals is obtained by taking advantage of spatial redundancy and temporal redundancy of the encoded images. Spatial redundancy is evaluated mainly by virtue of a succession of three operations: an operation commonly called a discrete cosine transform and indicated as DCT ("discrete cosine transform"), a quantification operation of the coefficients that arise from DCT and a variable length coding operation to describe the quantized coefficients that arise from DCT.

Temporal redundancy is analyzed by a movement compensation operation which consists, by translating each block of the current image, in the search for the most similar block located in the reference image, when searching for the most similar block located in the reference image. In analysis of the temporal redundancy leads to a field of certain translation vectors, commonly called motion vectors, as well as a prediction error which is the difference between the signal of the current image and the signal of the image predicted by the movement compensation. The prediction error is then analyzed according to the principle of spatial redundancy. MPEG encoding is predictive. It is found that the decoding which is associated with it regularly must be reinitialized regularly in order to protect the signal against any transmission error or any interruption in the signal due to the decoder being switched from one program to another. For this purpose, the MPEG2 standard establishes that, periodically, the images must be encoded in spatial mode, that is, in accordance with a mode of use only of spatial redundancy. The images encoded in the spatial mode are called INTRA images or I images. The images encoded using the temporal redundancy are of two types: on the one hand, the images constructed by reference to a previous temporal image based on a frontal prediction and, on the other hand, the images constructed by reference to two temporal images, previous and Subsequent, based on a frontal prediction and a rear prediction. Coded images constructed based on a frontal prediction are called predicted images or images P, and the coded images constructed on the basis of a frontal and a back prediction are called bidirectional images or B images. An I image is decoded without reference being made to images other than itself. An image P is decoded with reference to the image P or I which precedes it. An image B is decoded by relying on the I or P image which precedes it and an I or P image which follows it. The periodicity of images I defines a group of images broadly referred to as GOP ("group of illustrations"). Within a single GOP, the amount of data contained in an I image is generally greater than the amount of data contained in a P image and the amount of data contained in the P image is generally greater than the amount of data contained in an image. P.

At 50 Hertz, the GOP is presented as an I image followed by a sequence of images B and P which, most of the time, show the following sequence I, B, B, P, B, B, P, B, B, P, B, B * However, the standard does not require that N = 12 images be provided in a GOP, as is the general case, nor that the distances M between two images P always be equal to 3. More precisely, the distance M is the number n of B images that precede or follow an image P, increased by one unit, that is, M = N + 1. The number N represents the size or length of the GOP, while the number M represents its structure. The invention results from the observation that it is possible to act on parameters M and N to improve the level of compression and / or improve the quality of the coding. The coding method according to the invention is characterized in that a parameter is determined which characterizes the source of images which are to be coded according to a group and where the length and structure of the group are made to depend on this parameter or of these parameters. In one embodiment, the parameter or parameters that characterize the source images is or are determined with the aid of a test encoding in the course of which defined values N, M and the quantization interval Q are assigned. The test coding the open circuit is carried out, for example. In a particularly simple embodiment, a parameter (Pcost) characterizing the images P obtained during the test coding and a parameter (Bcost) characterizing the B images obtained during the test coding are determined separately, these parameters characterize the images P and B and, preferably, are the average coding costs of images P and B. The cost of encoding an image is the number of bits (including headers) which are necessary for coding. In this case, it is possible to make the number N depend on the characterizing parameter of the images P and the number M depends on the characterizing parameter of the images B. During the tests carried out in the context of the invention, in sequences of images of various types, it was noted that, for each type of sequence, there is an optimal N number that is provided as a minimum coding cost (or performance) for the P images and an optimal M number that provides a minimum coding cost (or performance) for the B images, these costs are obtained during the test coding. These sequences are differentiated from others by the movement of variable amplitudes, different objects, different spatial definitions and different contents. Furthermore, it is noted that there is a practically linear relationship between the optimum N number and the performance of the P images. Similarly, there is a practically linear relationship between the number m and the performance of the B images. Therefore, when knowing the returns, from the P and B images it is easy to calculate the numbers N and M which provides the best results. In an example corresponding to the MPEG2 standard, 50 Hz, the test coding is carried out with N = 12, M = 13 and Q = 15, the relation between N and the performance of the P images is approximately as follows: (i) N = INT 389000 - Pcost + 1 10000 with 12 < N < 30 and the relationship between M and the yield, or cost, Bcost of the R images is as follows: (2) M < 7. N = INT 179000 - Bcost with 1 < + 1 20000 It is also possible to limit M to 5. In these formulas, INT means the whole part. The limitation of N between 12 and 30, and the limitation of M to a maximum value of 7 makes it possible to have a simple mode of the encoders and limit the time of program change. With the same objective it is also possible to impose other limitations or restrictions, particularly that M is constant in the GOP and / or that it is a submultiple of N. In a modality, if the values of N and M taken individually and together they are not compatible with the restrictions, the values of M and N that are closer to the calculated values and which satisfy the stipulated compatibility will be chosen. In this case, the value of M will be favored, that is, if it is going to be a choice between several pairs M, N, the pair will be chosen for which the M value is closer to the one of the results from the calculation . The formula (2) above applies only if Bcost does not exceed 179800. In the opposite case, that is without Bcost >; 179000, the experiment has shown that it is necessary, in this example, for M to be chosen as follows: (3) with 1 < M < 7, M = 5. INT Pcost Bcost If the cost of an image B is greater than the cost of an image P it is preferable for the GOP not to contain an image B, that is, M = 1. This is because the images P, show a better quality of prediction than B images and, supposedly, are of lower cost, and the presence of such B images would be an inconvenience in this case. The costs, in bits of each image P and of each image B, are determined, for example, in the manner and at the moment in which these images appear. In one embodiment, the values of M and N are selected by averaging over all P and B images of the test encoding, and the appropriate coding is carried out only after the N-source image encoding of the test, N is determined by the cost of coding the P images. In this case, the M parameter may remain constant in the GOP. In another modality, which allows a more rapid adaptation to the variations in the content of the scenes as well as a reduction in the delay between the arrival of the source images and the appropriate coding (and which therefore allows a smaller capacity). buffer), the appropriate encoding is established as soon as the test encoding supplies data that allows this start. Therefore, the first image B of the test encoding provides an M number that allows the encoding to be started and the number N is supplied by the first image P of the test encoding. It is also possible to have the start of coding only after the test coding of the first image P; in this case, the coding starts when the value of N and the value of M are known. With this type of coding "in the process", the M number, ie the structure, can vary within a GOP, which allows faster adaptation to variations in the content of the scene. In the coding carried out progressively, the GOP is interrupted when the number of images already encoded in the current GOP is at least equal to the number N measured (measured by Pcost in the previous example) or before a scene change. In order to avoid significant variations in the parameters between the groups which follow each other, it can be shown that it is convenient to deviate from the calculated values. For example, if the calculation shows that, for a large part of the length of the GOP, for example at least 80%, M = 1 would be necessary, while for the rest of the GOP the calculation shows that M would be greater than 1. , the value of 1 will be adopted for M, despite everything, even if the calculation shows that a different value is necessary. Similarly, if for the preceding GOP, M = 1 and if, for the current GOP, the calculation shows that an M = 1 value would be necessary for a significant part of the current GOP, for example at least 60%, it will also be adopted the value of 1 for M, even if the result of the calculation, which is obtained from formula (2) above, implies a different value. It is known that when a scene change occurs, that is, when a discontinuity appears in the sequence of video images, it is necessary to adapt the GOP image groups on both sides of the discontinuity so that the new group, which starts with an I image, correspond with the new scene. In one modality, if a scene change occurs in a group, the new scene constitutes the I image of a new group, and the affected group is shortened so that it stops before this new scene if the scene change occurs in the affected group, at a distance from the beginning at least equal to the minimum number allowed for N. The start of the affected group is used to lengthen the group which precedes and when the sum of the number of images preceding the change of scene in the group affected and the number of images of the group which precedes, does not exceed the maximum manageable for N. In this previous group modified in this way (shortened or lengthened), it may be necessary to modify the M number previously calculated for this GOP. In a variant, which is preferably used in the case in which the length of the affected group is less than the minimum allowable for N when a scene change occurs in a group, the new scene constitutes the I image for a group new, this new group has a length equal to the average of the length of the group before it was affected and the length of the group that precedes it. With this variant, it may be necessary to modify the M number previously calibrated for the GOPs. When two modifications are possible, for example when the length of the affected group is less than the minimum allowable for N, a choice can be made between these two modifications by carrying out a calculation, for each modification, of the distance of the pair (M , N) obtained or the pair M, N before modification, and select the pair for which the distance is smaller. In order to determine the parameters M and N, the measurement resource of parameters different from the measurement of the yields must be taken. For example, it can be used, in order to determine N of the energy of the Intra I images. It is also possible to determine the amplitude of the movements or other motion compensation error, known as DFD (displaced frame difference) to determine M and N. Other characteristics and advantages of the invention will emerge with the description of some of its modalities, this description is provided with reference to the accompanying drawings in which: Figure 1 illustrates a macroblock for standard 4.2.0, Figure 2 is a diagram illustrating the DCT transform, Figure 3 shows a group of GOP images, In accordance with the MPEG standard or similar standard, Figures 4 to 7 are diagrams that illustrate the method according to the invention, and Figure 8 is a diagram of a distribution for implementing the method according to the invention. Reference will first be made to figures 1 to 3 which attempt to reiterate certain principles used in MPEG2 coding. In the MPEG2 standard, a point of view can be an image that includes, in progressive mode, 576 lines each of 720 points. In interlaced mode, this image consists of two frames, each of which comprises 288 lines, each also of 720 points. Each image is broken into microblocks, each of which is formed by a square of 16 x 16 luminance points. In this way each microblock is formed of four square blocks of 8 x 8 luminance points. With each of these four luminance blocks, two chrominance blocks, each of which shows 8 x 8 dots, are associated (in the 4.2.0 format), one of the blocks represents the color difference Cr signals or red chrominance and the other block represents the Cb signals of color difference or blue chrominance. In the 4.2.2 format, four chrominance blocks 8 x 8 are associated with each luminance macroblock, two blocks for the blue chrominance and two blocks for the red chrominance. There is also a 4.4.4 format. for which each of the luminance and chrominance components includes four blocks 8 x 8. In figure 1 four luminance blocks 8 8 denoted in general with the number 10, and blocks 12 and 14 of chrominance 8 x 8 are represented. for the blue and red chrominance respectively, the whole illustrates a macroblock for the 4.2.0 standard. Each block is coded using a denoted DCT transform which is a discrete cosine transform which makes it possible to transform a luminance block (for example) into a block of coefficients representing spatial frequencies. As can be seen in figure 2, the source block 16 becomes a block 18 of coefficients 8 x 8. The upper left corner of block 18 corresponds to the spatial frequencies zero (average value of the block) and, starting from from this origin 20, the horizontal frequencies increase to the right, represented by the arrow 22, while the vertical spatial frequencies increase starting from the top down, represented by the arrow 24.

For each macroblock, the type of coding must be chosen: either "intra" or "inter". The intra-coding consists of applying the DCT transform to a source block of the image, while the inter-coding consists of applying the DCT transform to a block that represents the difference between a source block and a predicted block or prediction block, of a preceding or subsequent image. The choice depends partly on the type of images to which the macroblock belongs. These images are of three types: the first type is the type known as I or intra, for which the coding is intra for all of the macroblocks. The second type is of P or type of prediction; In this type of images, the coding of each macroblock can be intra or inter. In the case of an inter-coding on a P-type image, the DCT transform is applied to the difference between the current macroblock of this P-image and a prediction macroblock arising from the preceding I or P image. The third type of images is called B or bidirectional. Each macroblock of such an image type is encoded in either intra mode, or is encoded in inter mode. The inter-coding also consists of applying the transform to the difference between the current macroblock of this B-image and a prediction macroblock. This prediction macroblock can arise either from the preceding image or from the next image or both at the same time (bidirectional prediction), the prediction images named precedent or later can only be type I or P. In figure 3 has presented a set of images, which form a group called the GOP ("group of illustrations") which comprises 12 images, specifically an image I followed by 11 images B and P, according to the following sequence: B, B , P, B, B, P, B, .B, P, B, B. A GOP is characterized by a length, that is, a number of images N which, in an example it can only be found between 12 and 30, and by a structural parameter M that represents the distance between two P images, that is, the number of B images between two successive P images, increased by one unit. In this example, this parameter M is equal to 3. In addition, by way of example, this number M can be found between 1 (without image B) and 7. It also stipulates that this number M can be a submultiple of the number N with In order to simplify the encoder. To date, images have been encoded while keeping N and M constant in the encoder. The invention results from the observation that there are optimal values of M and N which differ according to the sequences of the encoded images. This is because, based on whether the image sequences have more or less definition and significant or minor movement, the optimal values of M and N can differ significantly. Optimal values are those which, for the same quality, require a mim number of bits. First of all, the experimental studies carried out in the context of the invention have shown that the optimal Nopt size of the GOP for a defined sequence of images corresponds to the mim value Pcost on this sequence, of the number of bits which is necessary to use to encode the P images (including headers). This property is illustrated by the diagram of figure 4 in which the number N has been plotted on the abscissa and, on the ordinate, the value Pcost for a sequence denoted as I. This Pcost value is the number of bits that will be used to encode an image P at an average value on the sequence i. Therefore it is observed that the value Pcost (i) is represented by a curve 32 that shows a mim 34 for which the value of N is optimal (Nopt). In a similar way, it has been observed that the optimum value of the number M corresponds to the mim Bcost (i) of the number of bits that will be used on average to encode the B images on a defined sequence, indicated as i. Therefore, in the diagram of figure 5, the number M has been presented again on the abscissa and the number Bcost (i) has been plotted on the ordinate. In this diagram it is observed that curve 36 shows a mim 38 corresponding to the optimum value of M (Mopt). Measurements have been taken, particularly in test sequences which are conventional in the MPEG coding called "horses", "flower garden" and "Mobcal". The sequence "horse" corresponds to fast movements with good definition, the sequence "flower garden" also corresponds to a good definition and average movements, while the sequence "Mobcal" corresponds to little movement and high definition. Other sequences have been tried such as a sequence in ayak with fast movements and little definition, a sequence of basketbol and a sequence with average uniform movements, and images with good definition. It has also been noted that, if the group is subjected to test coding with defined values of M, N and the quantization interval Q, these values do not necessarily correspond to the optimal values of the sequence i in question, the average cost of the The coding of the P, Pcost images, and the average cost of the coding of the B, Bcost images, represent N and M, respectively. In addition, as shown in Figure 6, there is a simple relationship between the numbers Nopt for each sequence i and the cost of coding Pcost given M, N and Q. This relationship is linear or substantially linear, and is represented by the straight line 40 ( figure 6) on which the different points 42, 44, etc., represent different sequences. Figure 7 is a diagram in which the Nopt groups are plotted on the abscissa and the cost of Pcost coding (with M, N and Q defined) are plotted on the ordinate; * each point 52, 54, 56, etc., corresponds to a given sequence. It is noted that these points are on line 60 straight. Therefore, you have a linear relationship between Nopt and the cost of proof coding. When the values of M, N and Q used in the course of the test coding are as follows: M = 12, N = 3, and Q = 15, the values of M and N satisfy the following relationships: (1) N = INT 389000 - Pcost + 1 10000 with 12 < N < 30 (2) with 1 < M < 7 N = INT 179000 - Bcost + 1 20000 Although, for formula (2) above, it has been indicated that M must be between 1 and 7, it is observed in the diagram of figure 7 that in fact M can be limited to 5 The layout designed to implement the invention is depicted in Figure 8. It comprises a "first MPEG2 encoder 70 designed to carry out the" first pass "coding or test coding.This prμeba coding is adjusted with the fixed parameters indicated before, specifically, in this example: M = 12, N = 3 and Q = 15. This test encoder operates, in this example, in an open circuit, ie without regulation, The encoder 70 supplies the Bcost and Pcost values which are applied to a converter 72 which carries out the conversions of Pcost in Nopt and Bcost in Mopt, represented in figure 6 and 7, according to the relations (1) and (2) above. are calculated for a group of images, as described in the above and then applied to a control input 76 of an MPEG2 encoder 74. The data at the input of the encoder 74 is the same as that of the input of the test encoder 70. Therefore, an intermediate memory 78 is provided to take into account the processing time in the test coder 70 and in the converter 72, this memory 78 retains the data during processing.

In the converter 72, it is also verified that the pairs N, M resulting from the formulas (1) and (2) are compatible with the restrictions imposed in the modality, in particular that M is a submultiple of N. If the values that results of the calculation are not compatible, values of N * and M are adopted which are the closest to those that were calculated, favoring, however, the values of M. The converter 72 also takes into consideration supplementary conditions. First, make a comparison between Bcost and Pcost and, if Bcost is greater than Pcost, a value of 1 is assigned to the number M, the GOP that does not contain image B. This is because, with this assumption, the B images imply a coding cost which is greater than that of the B images; it is preferable to keep only P images which show a higher prediction quality. Secondly, the converter compares Bcost with the value 179000 and, if Bcost exceeds 179000, the previous relation (2) is replaced by the following heuristic relationship: (3) with 1 < M < 7, M = 5. INT Pcost - 1 Bcost The converter 72 also makes it possible to take into account the two special cases for which it is necessary to deviate from the relation (2) in order to obtain uniformity in the image quality. The first case is as follows: the test coding shows that M must show a value at least equal to 2 but, in addition, it is proof coding also shows that the intermediate values obtained by M are equal to 1 over a large part of the group, for example at least 80%. In this case, the converter 72 stipulates that M equals l. The second case is similar to the first: the test coding shows that M must be at least equal to 2, but the intermediate values obtained for M are equal to 1 for at least a part of the group length, for example 60% (this limit is below the anticipated limit in the first case), and the preceding group is such that M = 1. In this case, the value 1 is also conferred on the number M. These two special cases, for which the value 1 is established for M, result from tests carried out in the context of the invention which have shown that these conditions allow a good quality uniformity for the same type of sequence, on successive groups. Finally, the converter 72 takes into account the scene changes or "cuts" which are usually detected in the encoders. When such a scene change occurs, the GOP starts with a new scene, that is, when a new scene appears, an image I is attributed. In addition, with the method of the invention, when a scene change is detected, the preceding GOP and the GOP Current settings are based on the following considerations: if the scene change appears in a GOP after the eleventh image, the new GOP starts with the scene change, and the preceding GOP is limited or shortened, in contrast, if the change of scene appears before the twelfth image, it is not possible to limit the preceding GOP, so that it is deleted just before the scene change since, in this case, its number of images would be less than the stipulated minimum number. The preceding GOP and the current GOP are then modified as follows, distinguishing two cases. In the first case, the change of scene appears at the moment in which the sum of the number of images of the preceding GOP and the number of current GOP images, just before the change of scene, is at least equal to 30. In this In case, the preceding GOP is lengthened. In the second case, the sum of the number of images of the Previous GOP and the number of images of the current GOP just before the scene change is greater than 30. The preceding GOP and the current GOP are then rearranged by calculating an average that corresponds to these two GOPs.

For example, if the preceding GOP is such that N = 25 and M = 2, and if a scene change occurs after the eighth image of the current GOP for which the calculation indicates N = 20 and N = 3, the GOP precedent, lengthened by the current shortened GOP includes 33 images. Since this value * exceeds the maximum allowable (30), an "average" corresponding to the two GOPs is looked for, the total number of images of which is 33, each, of the GOPs must comply with the imposed restrictions. In this case, it is observed that a choice can be made between N = 18 and M = 2 for the preceding GOP, and N = 15 and M = 3 for the GOP just before the scene change. The lengths 18 and 15 are close to the average (16, 5) of the length of the preceding group (25) and the length (8) of the current group affected. Tests have been carried out in twelve different sequences with scene changes, flashes of light and relatively long durations, and the results obtained with a conventional method of coding, corresponding to fixed values of M and N have been compared with the results obtained with the method according to the invention which adapts the values of M and N to the sequences. The tests were carried out with various performances. An increase in quality is observed, measured by the parameter PSNR (maximum signal-to-noise ratio) of 0.2 dB and 1.14 dB. This increase in PSNR corresponds to a saving in terms of bits, which is between approximately 2 and 22%.

The method according to the invention can be used for any type of video image compression method in which I, P and B images are provided. It is applied both for recording, in real time or off line as well as for transmission. The method is not limited to the case in which the size of the GOP is determined before coding. Also, when the parameters M and N are calculated for each image, the appropriate encoding is carried out in process. In this case, the M number may vary with a GOP, when a new GOP is started, for example, when the image number is coded in the current GOP it is at least equal to the calculated N number. The M number may vary as a function of the complexity of the images within a GOP. In this case, it is not necessary to store the entire GOP in the intermediate memory 78 (whose capacity can be reduced), the restrictions for the values of M and N are reduced, restrictions which are defined only by the MPEG2 standard; the restrictions imposed on scene changes are also less severe.

Claims (20)

1. Image compression method in which the images are coded according to groups, each of which comprises a number N of images, * N represents the length of the group, which includes an I image encoded in intra mode, P images predicted on the basis of the intra I image or a preceding P image, each P image is preceded and followed by n bi-directionally predicted B images, n possibly being zero, the number M, which is equal to the number n increased by one unit, represents the structure of the group, the method is characterized in that at least one parameter that characterizes the image source is determined which will be coded according to a group and wherein the length N and the structure M of the group are made dependent of this parameter or these parameters.

2. The method according to claim 1, characterized in that the parameter or parameters that characterize the source images is or are determined with the aid of a test coding in the course of which defined values are assigned to N, M and the interval of quantification Q.

3. The method according to claim 2, characterized in that the test coding is carried out in open circuit.

. The method according to claim 2 or 3, characterized in that, in order to characterize the source images, a parameter is characterized which characterizes the P images obtained during the test coding, and a parameter that characterizes the B images obtained during the the test coding.

5. The method according to claim 4, characterized in that the number N is determined based on the parameter that characterizes at least one image P, and wherein the number M is determined based on the parameter that characterizes at least one image B .

6. The method according to claim 4 or 5, characterized in that the parameters that characterize the images P and B are the costs of coding the images P and B, for example the average costs.

7. The method according to claims 5 and 6, characterized in that, when, in the course of the test coding, the average coding cost of each B-image is greater than the average coding cost of each P-image, a value of 1 is given to the M-number, and this way the group does not contain image B.

8. The method according to claim 6 or 7, characterized in that, in the course of the test coding, the coding cost of each B-image and the corresponding M-number are determined in the stage with the arrival of the source images.

9. The method according to claim 8, characterized in that, when the numbers M determined before the end of the test coding are equal to 1 for a significant fraction of the group, the value of 1 is conferred to the number M of the group.

10. The method according to claim 8 or 9, characterized in that when in the numbers M determined before the end of the test coding are equal to 1, during at least a defined fraction of the group, and when the number M equals 1 for the preceding group, the value of 1 is given to the number M for the group.

11. The method according to any one of the preceding claims, characterized in that, in the event that a scene change occurs in a group, the new scene constitutes the image I of a new group, the affected group is shortened so as to be stop before this new scene if the change of scene occurs in an affected group, at a distance from the beginning at least equal to the number, minimum allowable for N, the sign of the affected group is used to lengthen the group which precedes when the sum of the number of images that precede the change of scene is the affected group, and the number of images of the group which precedes it does not exceed the maximum number admissible for N.

12. The method according to any of claims 1 to 9, characterized in that, in the event of a scene change in a group, the new scene constitutes the image I of a new group, the affected group and the group the which precedes it is rearranged so that each one shows a length close to the average of the length of the group after the alteration and of the group which precedes it.

13. The method according to claim 6, characterized in that the test coding is carried out according to a standard MPEG at 50 Hertz with N = 12, M = 3 and Q = 15, the numbers N and M are a function , respectively, of the average coding costs of images B and P, according to the following ratios: (1) 389000 - Pcost N = INT + 1 10000 with 12 < N < 30 (2) with 1 < , < . 7, N = INT 179000 - Bcost + 1 20000 INT means whole part.

14. The method according to claim 13, characterized in that 1 < . M < . 7

15. The method according to claim 13 or 14, characterized in that, when the cost of coding is greater than 179000, the number M is determined by the following relation: (3) with 1 < M < 7. M = 5. INT Pcost Bcost

16. The method according to any of claims 1 to 6, characterized in that the number M is produced to vary within a group.

17. The method according to any of claims 2 to 6, characterized in that the compression is carried out after the test coding.

18. The method according to any of claims 4 to 6, characterized in that the compression begins after the parameter characterizing the first image B or the first image P has been determined.

19. The method according to claim 18, characterized in that the formation of a coded group is interrupted when its number of encoded images is at least equal to the number N determined based on the current P-image.

20. A coding device for implementing the method according to claim 4, 5 or 6, characterized in that it includes a channel for carrying out the test coding and determining the parameters M and N, and a coding channel that receives the information of the first channel in a way that carries out the appropriate encoding.