FR2790853A1 - METHOD FOR COMPRESSING DIGITAL IMAGES IN WHICH A CODING QUALITY FACTOR IS ATTACHED TO EACH AREA - Google Patents

Abstract

The invention relates to a method for coding digitized images with a view to their compression in which the image is separated into zones, such as macroblocks, individually coded. This method is characterized in that a filtering (10) is applied to the source image and a filtering to the same image resulting from the decoding of the coded image, so as to determine a quality factor for each zone, representing l 'difference (20) between the decoded picture and the source picture, and in that this quality factor is used to modify an encoding parameter of each area so as to minimize the differences between quality factors of the various areas. Preferably, the filtering is psycho-visual. Application to MPEG coding.

Description

METHOD FOR COMPRESSING DIGITAL IMAGES IN WHICH A

  CODING QUALITY FACTOR IS ATTACHED TO EACH AREA

  The invention relates to a compression method

of scanned images.

  A scanned image corresponds to a large amount of information. To store and / or transmit image information, in particular moving images, it is therefore necessary to

  use compression methods to record

  sort or transmit a reduced number of digital data

  without affecting the quality of the reproduced image.

  Compression involves using various techniques

  to avoid the transmission or recording of information

  redundant mations and Huffman coding assigning a reduced number of digits (bits) to the information which appears most frequently. Compression also uses the properties of human vision, particularly the fact that the human eye

  is less sensitive to fine details (high frequencies spa-

  coarse detail (low spatial frequencies) and the visibility of fine detail decreases as movement increases. Thus, compression processes use a

  transformation which assigns a greater weight to the

  low spatial frequencies than high spatial frequencies.

  In the current compression process which is the subject of the MPEG standard, the discrete cosine transform (DCT,

"Discrete Cosine Transform").

  In what follows, for simplicity, reference will be made above all to the MPEG2 standard although the invention is not limited to

to this standard.

  In this method, the image is divided into macroblocks of 16 × 16 picture elements (pixels or pels), each macroblock being

  formed of 4 blocks, each of which has 8 x 8 pixels.

  According to the MPEG2 standard, each macroblock of 16x16 pixels is transformed into a macroblock of coefficients of a matrix

  whose lines correspond to the horizontal space frequencies

  tales and columns correspond to vertical spatial frequencies. These coefficients are coded according to a precision which is defined by a quantization step. The higher the pitch, the lower the accuracy. According to the MPEG2 standard, the quantization step can be chosen for each macroblock. Thus, for each macroblock, a quantization step can be selected which is the result of a compromise between the need to obtain good image quality (which therefore implies a low quantization step) and to limit the number of bits. to transmit (which implies a high quantification step).

  Despite this possibility of quantification modulation,

  tion for each macroblock, in practice, the same quantization step is used for each image or a quantization step which can vary from one horizontal row of macroblocks to another.

  In addition, transmission or stock standards-

  Age often impose a determined bit rate or a coding cost determined by image. This coding cost is generally imposed by

a regulation module.

  We have already proposed a coding in which each macro-

  block, or each block, has a quantization step chosen so as to obtain good quality on each macroblock, or block, while respecting the constraint imposed on the overall coding cost of the image. In other words, we proposed to choose the pitch

  quantification of each macroblock, or block, so as to opti-

  focus on the quality of each of the macroblocks, or blocks, while respecting the condition imposed, namely that the total number of

  image bits (coding cost) is equal to a specific value

  undermined. However, this coding method is not used in

  practical because experience shows that the images obtained present

try flaws.

  The invention results from the observation that the human eye is sensitive to the diversity of the reproduction qualities of the various areas of the image, that is to say, that the areas reproduced with different qualities are perceptible to

  the eye and give the impression that the image has defects.

  Thus, the method according to the invention is characterized in that each image zone, in particular each block or macroblock, is coded, so that all the zones have substantially the same quality factor, preferably a quality factor

psycho-visual type.

  To achieve uniformity in the quality of

  various zones, we preferably act on the quanti-

  fication. The quality factor used is, in one example, the signal to image noise ratio. This factor is measured, in a manner known per se, by performing, for each pixel, the difference between the luminance of the pixel of the source image and the luminance of the corresponding pixel of the decoded image, the signal to image noise ratio. in each zone being, for example, a function of the sum of the squares of the differences between luminances for all

the pixels of the area.

  In an example, the signal-to-noise ratio has the value PSNR or SNR defined by the following first and second formulas: 2,552 PSNR = 10log 2 - IN diff2 max SNR = 10log 1 YN diff2 diff being the difference in luminance between pixels of the source image and the decoded image, max the maximum value of a pixel in the image considered, and N the total number of pixels. While this quality factor provides results

  satisfactory, the best results are obtained with a fac-

  high quality that takes into account the psycho-visual properties of human vision. The quality factor depends, in this case, for each zone, on the luminance, the spatial frequencies, the

colors and movement.

  We can define a quality factor of this type as follows: We apply psycho-visual filtering to the source image and the decoded image and we make a difference, pixel by

  pixel, which will represent, for each point of the image, the fac-

sought-after quality.

  Psycho-visual filtering consists in applying thresholds translating a non-linear relationship between, on the one hand, the luminance and possibly the color, and, on the other hand, its visibility by the human eye. Psycho-visual filtering can also take into account the so-called masking phenomenon which consists of a reduction in the visibility of an image part due to a non-uniformity of the environment surrounding this image part. The non-uniformity can be of spatial type; she can

  also correspond to movement or time.

  Spatial masking can be defined by a parameter which expresses the visibility of the part of the image concerned as a function of the contrast of the spatial frequencies of this part in relation to the environment, the fine details may not be perceptible, whereas defects with a very spectrum

  different from the surrounding spectrum are much more visible.

  Such spatial masking filters are described in particular in the following articles: 1. F.X.J. Lukas & Z.L Budrikis, "Picture Quality Prediction Based on a Visual Model"

  image based on a visual model) IEEE transaction on communi-

  cation (vol.COM-30, n 7), July 1982.

  2. N.B. Hill, "A Visual Model Weighted Cosine Transform for Image Compression and Quality Assessment" (Un

  weighted cosine transform visual model for compression

  image sion and quality determination), IEEE transaction

  on communications (vol. COM-33, n 6), June 1985.

  3. K.N.Ngan, K.S.Leong and H.Singh, "Adaptive Cosine Transform Coding of Images in Perceptual Domain"

  in the perceptual domain by cosine transformation adapta-

  tive) IEEE transaction on acoustics, speech and signal proces-

  sing (vol. 37, no 11), November 1989.

  4. L. Carrioli and M.G. Albanesi "Multigray Image

  Compression based on the Human Eye Transfer Function "(compres-

  Zion of grayscale images based on the transfer function of the human eye) Alta Frequenza (vol. 57, n 5)

pages 273-284, June 1988.

  5. S.J. Daly, "A Noise Adaptive Visual Contrast Sensibility Function for Application to Image Data Compression" (A noise-adaptive visual contrast sensitivity function for application to image data compression)

Eastman Kodak company.

  Motion masking is a filtering that translates

  the fact that sensitivity to fine detail decreases with the

  space frequencies for medium speeds

  bass are more visible than high spatial frequencies.

  For more details on motion masking, see the following publications: 6. D.H. Kelly "Motion and Vision. II. Stabilized

  spatio-temporal threshold surface "(Movement and vision. II.

  stabilized space-time threshold surface) J. Opt. Soc. Am.

(vol. 69, no 10), October 1979.

  7. S. Daly, "Engineering observations from spatio-

  velocity and spatiotemporal visual models "(technical observations

  questions related to the visual model space / speed and space / time) Human vision and Electronic Imaging III, Proceedings of SPIE

  (vol. 3299), pages 180-191, January 1998.

  Finally, the temporal masking corresponds to the temporal summation of Bloch's law for a part of image of short duration, less than 20 ms. According to Bloch's law: the perception of a difference (or error) in luminance dL1 of duration t1 ms in the image is the same as the perception of a difference in luminance dL2 of duration t2 ms, if these values verify the following relationships: dLl.t1 = dL2.t2 (t1, t2) <20ms Thus, from these filterings and these differences between the filtered source image and the filtered decoded image, we establish, for each pixel of the image, or for a part of the pixels of the image, and, in any case, for each zone (block or macroblock in particular) a quality factor and we choose at

  minus a coding parameter, in particular the quantization step

  tion, so as to minimize disparities or disper-

let us know about the quality in the image.

  In general, the relationship between the various values of the quality factor and the coding parameter considered is not known, because it depends on the content of each image. Under these conditions, to obtain good results, it is preferable to use an iterative process which consists, for example, during a first step of assigning a determined value, the same for all the zones, to the coding parameter and to establish the quality map and, from this quality map, an average value of the quality factor is determined and the coding parameter is modified in such a way that it

  increases the quality at least for part of the quality zones

  below average and decreases quality at least for some of the above average quality areas. For example, we increase the quantization step to decrease the quality and we decrease the quantization step for

increase the quality.

  During this first iteration step, it is preferable that the increase and decrease in the value of the

  coding parameters are minimal, this variation corresponds

  in, for example, an increment.

  In the second step, we apply the

  coding parameters that were determined during the pre-

  first step. This second step may be sufficient to obtain

  to establish a correct homogenization of quality.

  However, better results are obtained by

  following the iteration as follows: we determine the para-

  coding meter for the next step on the assumption that, from one step to another, the variation of the quality factor

  is proportional to the variation of the coding parameter.

  So the proportionality coefficient is the ratio between,

  on the one hand, the difference between the values of the factor of

  lity at the stages of rows n and n-l and, on the other hand, the difference

  between the values of the coding parameter in steps of rows n and n-1. This proportionality coefficient can be used to determine the new value of the parameter of

  coding that would result from redetermining the varia-

quality factor.

  In the case where the cost of coding the image must pre-

  feel a determined value, in order to satisfy this condition

  tion, in one embodiment, some of the zones are selected for which the coding parameter is varied. The

  number and nature of the areas selected, as well as the

  study of the variation of the coding parameter are chosen with a view to the result to be achieved, that is to say that the coding cost has a determined value. For example if, during a step, the coding cost is lower than the desired coding cost, for the K% (for example 10%) zones having the lowest quality

  the quantification step is reduced, which, on the one hand, increases

  quality lies (in the sense of homogenization) and, on the other hand, increases the cost of coding. Conversely, if the coding cost is too high, we select K '%, for example 5%, of the zones having the best quality and, to homogenize the quality, we increase their quantization step, which

  reduces the cost of coding.

  Preferably, during a step, x% of the lowest quality areas and y% of the best areas are chosen.

  quality to modify the coding parameter.

  If, at the end of this operation performed by calculation, the coding cost does not vary correctly, x and / or y are modified, for example by a determined factor chosen so that the coding cost tends towards the correct value. For example, if the coding cost is too high, the number of zones whose quantization step is decreased can be reduced and / or the number of zones for which the quantization step is increased can be increased. These various adjustment operations are carried out iteratively by calculation. This secondary iteration should not be confused with the main iteration during which the quantization steps are actually modified. During this secondary iteration, the quantization step is modified each time, preferably by a minimum value, by

example of an increment.

  Since, in general, we will not be able to achieve complete standardization of the quality factor over the entire image, we interrupt the main iteration when, from one stage of this iteration to another, there is no had decrease

  of disparities between zones.

  Thus the invention relates, in general, to a method of coding digitized images with a view to their compression in which the image is separated into zones, such as macroblocks, individually coded, which is characterized in that filtering is applied to the source image and filtering to the same image resulting from the decoding of the coded image, so as to determine a quality factor for each zone, representing the difference between the decoded image and the image source, and in that this quality factor is used to modify a coding parameter of each zone so as to minimize the differences between

  quality factors of the various areas.

  According to one embodiment, an average quality factor is determined for the entire image and the modification of the coding parameter for each zone is carried out so as to minimize for each zone the difference between the quality factor

  corresponding and the average quality factor.

  In one embodiment, during a first step, a coding parameter of determined value is assigned to all the zones, and the coding parameters for the next step are determined by modifying this parameter in such a way that it tends to bring the quality factor of each area closer to the

average quality factor.

  Preferably, during the first step, the coding parameter of each zone is varied by a minimum increment. Advantageously, during a second step, or a subsequent step, the coding parameter which will be used for the next step is determined from a predictive model according to which the variations of the coding parameter are, for each area, proportional to variations in

quality factor.

  According to one embodiment, the cost of coding the image

is predetermined or controlled.

  In this embodiment, preferably, during the second step, or a subsequent step, the coding cost for the next step is determined from a predictive model according to which the variation in the coding cost is, for each

  zone, proportional to the variation of the coding parameter.

  Advantageously in the latter case, the coding parameters for each zone are determined so that in the next step the variation in the cost of coding is minimized.

compared to the current step.

  Preferably, during the second step, or a subsequent step: the coding parameter for each zone is determined by iteration by modifying by an increment the coding parameter of a first determined fraction of zones having the best factor quality so as to decrease this quality and / or by changing the coding parameter of an

  second determined fraction of areas with a lower qua-

  In order to improve this quality, the variation of the coding cost which results from this (or these) variation (s) of the coding parameter is determined, and,

  we modify the number of zones for which we modify

  the coding parameter, so as to minimize the difference by

  relative to the desired coding cost of the image.

  Advantageously in this case, during each iteration step modifying by one increment the coding parameter of the zones having the best and / or the least quality factor, the average of the deviations of the quality factor for all the zones is determined. compared to the average quality factor and the iteration is interrupted when this average has

more likely to decrease.

  According to one embodiment, the coding parameter is the quantization step of the value assigned to the coefficients

  attached to each pixel in an area.

  According to one embodiment, the filtering is of the psycho-

  visual. This psycho-visual filtering takes account for example of

  the effect of spatial and / or movement and / or temporal masking.

  As a variant, the filtering determines the signal to

image noise.

  In one embodiment, the coding is of the object oriented type and the quality factor is linked to the degree of interest of the regions. When the coding is of the object oriented type, it is possible to optimize the homogeneity of the quality factor on each of the

objects taken individually.

  The method according to the invention applies in particular to

  storage and / or transmission of images.

  Other characteristics and advantages of the invention

  will appear with the description of some of its modes of

  embodiment, this one being carried out with reference to the attached drawings in which: FIG. 1 is a block diagram of the method and the device in accordance with the invention, FIG. 2 is a diagram used to explain a step in the method in accordance with l invention, and Figures 3 and 4 are diagrams illustrating

  results obtained with the process according to the invention.

  The method which we will describe in relation to the figures consists in proceeding by iteration to homogenize the subjective quality of MPEG2 macroblocks by acting on the quantization step of each of the macroblocks while keeping constant

  and / or controlled the cost of coding each image.

  As shown in FIG. 1, the source image is subjected to a psycho-visual filter 10 and to an encoder 12 containing

  at least one variable parameter which is, in the example,

  killed by the quantification step. This parameter is communicated to the

  encoder 12 by a calculation unit 14 of this quantification step

  tion. The coded image 12 is decoded (block 16) and this image

  decoded is also subjected to the psycho-visual filter 10.

  The filter 10 includes a means which determines, for each pixel, the difference between the quality factor of the source image and the quality factor of the decoded image. In this way, a map of the dispersions of the qualities of the image is established which is supplemented, for each macroblock, by the difference

  between the average quality factor of the source image and the fac-

  average quality of the decoded image.

  For the determination of the quantification step, the

  block 14 uses the data constituted by the disper-

  qualities 20 and the cost of coding provided by a block 22.

  In the example, there is no simple relationship between the quality dispersions and the quantification steps.

  for each macroblock. To reduce quality dispersions

  it is therefore necessary to perform an iteration in use

  reading a prediction correction model, as explained below.

  INITIALIZATION STEP OF THE CALCULATION OF THE STEP

QUANTIFICATION.

  During the initialization step, either the coder determines a reference coding cost or this coding cost

  is determined by regulation (not shown).

  In the example, the cost of coding is determined by

  a regulation which determines the data rate.

  The cost of coding each image must be constant or controlled, i.e. determined for each image but not

  necessarily the same for all images.

  Furthermore, it is assumed that the quantization step is

  represented by an integer between 1 and 31.

  The quantization step is, during the initialization step, established at a constant value and is supplied

also by regulation.

  With this data, the coder 12 provides, on the one hand, a coding cost (block 22) and, on the other hand, an image decoded by the block 16 which is subjected to psycho-visual filtering 10 at the same

  time the source image to provide the map 20 of the dis-

  persions of qualities. Block 20 provides not only the dispersion map but also the average dispersion, or error

for the whole image.

  During this first step, block 14, prepares the second step by calculating, for each macroblock, the steps of

  quantification that will be used during the second step.

  This calculation is carried out as follows: For each macroblock, it is determined whether the quality dispersion or error is greater or less than the average error for the entire image, and we decrease by one (increment) the quantization step for macroblocks whose error is greater than the average and the quantization step for macroblocks whose

  the error is below average.

  In other words, we slightly decrease the quality of macroblocks with errors below the average and we

  increases the quality of macroblocks with higher errors

laughing at the average.

  If one or more macroblocks have an error equal to the average, it is modified, for example by increasing it by an increment, to allow an initialization of the Ak values

and Bk defined below.

SECONIDE ETAPF

  For this second step, the coder uses the quantization steps determined, for each macroblock, at the end of the first step (no increased original quantization or

decreased by one).

  To then determine the quantification steps

  QSCnk for each macroblock, k representing the number of the macro-

  block and n the step number, we use a pre-

  diction relating the quantification step to the quality factor and the coding cost, the quantification step determining the

  Emapnk quality factor and Costnk coding cost.

  This prediction model is a simple linear function as defined by the formulas below Enapnk = AkxQscnk + cte, Costnk = BkxQscnk + cte '(2) In these formulas cte and cte' represent constant values. Thus, at each step the coefficients Ak and Bk for each macroblock can be determined or updated according to the formulas below: Ak Costk-Cost (, - I, ACostk k Qsc ,, k-Qsc (i ,, - = AQsck Bk = Enmap ,, k-Emap (,, - i AEmnapk (3) Qsck-Qsc (t ,, - k - AQsck If the quantization step does not vary from step to step

  the next one, i.e. AQsck = 0, then Ak and Bk are main-

  held constant. However, during the second stage, the step of

  quantification necessarily varied from the pre-

  due to the variation imposed on it during

from the first step.

  Thanks to this prediction model, we can, from the second step, predict the quality factor and the coding cost of each macroblock (the quantization step having been

  determined in the previous step).

  The predicted variation in APictCostPred coding cost for the entire image is expressed by the following relationship: APictCostPred = _BkxAQsck (4) k Given that we want to obtain a determined coding cost of RefPictCost value, we are looking for therefore to minimize the quantity: ACostErr = APictCostPred- (PictCost, - RejfPictCost) (5) In this formula, PictCostn represents the cost of coding the image at the step of rank n. This ACostErr quantity thus represents the difference between the predicted variation of

  coding cost and the excess cost determined previously.

  In other words, the quantification steps or their variation AQscK are sought which make it possible to keep the overall cost of coding the image as close as possible to the cost of reference coding of this image while seeking, of course, to homogenize the factors quality, ie to reduce errors on quality factors (deviations from the average value). To this end, we divide the macroblocks into three subsets (Figure 2):

  - The first subset, noted [AQsck] +, is

  killed by all macroblocks for which AQsck will be positive

  (increase in the quantization step) and the error on the fac-

  quality is less than the average error.

  - The second subset, noted [AQsck] -, is

  killed by all macroblocks for which AQsck will be negative (reduction of the quantization step) and for which the error

  on the quality factor is greater than the average error.

  - The third subset, noted [AQsck] 0, is

  made up of all macroblocks for which the quant

tification will not change.

  Thus, the coding costs of the macroblocks of the first subset correspond to a decrease while the

  coding costs of the macroblocks of the second subset correspond to growth.

  From this division into subsets, during the second step (or a step of rank n), we proceed to another iteration (or secondary iteration) so as to determine

  the quantification steps which provide, for the next step

  vant, the greater homogeneity of the quality factor and the minimization of the cost error compared to the reference cost. To this end, in the first subset, we choose the K +% macroblocks (for example 10%) which have the most

  low coding error. In the second subset, we choose

  locate the K-% macroblocks (for example 15%) with the largest

coding error.

  For all K +% macroblocks of the subset [AQsck] +, the quantization step is increased by a determined value dQ +, for example an increment, which decreases the cost

  coding and reconciles the quality error with the average error.

  Similarly for all the K-% macroblocks of the subset [AQsck] -, we decrease by a determined quantity dQ-, for example an increment, the quantization step so

  to compare the error on the factor of

  and increase the cost of coding.

* From these modifications of the quantification step

  tion, we can, thanks to the linear prediction model described above

  above, predict the quality factor of each macroblock. We can therefore calculate the new average quality factor that results. Similarly, we can calculate the predicted cost variation

  coding from formula (5) above.

  If these calculations show that the variation in the coding cost increases, K-% is reduced by a determined factor, for example by a factor of 2, so as to reduce the influence of the K-% macroblocks of the subset [AQsck ] - and reduce the cost of coding. We can also, for the same purpose, increase in parallel the influence of K +% macroblocks of the subset [AQsck] l + of

  determined factor, for example a quarter.

  Conversely, if the ACostErr quantity decreases, i.e.

  say if this parameter becomes negative and increases in absolute value

  read, we make a reverse correction, by decreasing the influence of

  K + and possibly increasing the influence of K-.

  However, it may happen that, despite this adjustment,

  the error on the cost of coding always varies in the same direction.

  In this case, we increase the quantification step of K-% macro-

  blocks and the quantification step of the K +% macroblocks is reduced.

  In other words, if it is found that the iteration gives the opposite result to that sought, that is to say that this iteration does not allow convergence towards the cost of reference coding, but, on the contrary, an increase in the divergence, so we

  varies the quantization steps in an inverse manner.

  This iteration, performed by calculation, during the step of rank n, is interrupted when the quality error

no longer decreases.

  This secondary iteration thus provides the quantification steps for the next step.

  The various main iteration steps are inter-

  broken when the errors on the quality factors which are obtained no longer decrease. This main iteration therefore provides, for each macroblock, a quantification step which, in principle, is optimal for the homogeneity of the quality factor.

  on the image with a desired coding cost.

  An example of a varia-

  coding error as a function of the step number. On this diagram, the abscissa corresponds to the step number, that is to say to the rank of the main iteration, and the ordinate represents

the average coding error.

  The abscissa of the diagram in Figure 4 also indicates

  the rows of the main iteration, while the ordinate rep-

  the relative cost of coding an image, that is to say the ratio between, on the one hand, the difference between the cost of coding and the

  reference cost, and, on the other hand, the reference cost.

  The method according to the invention applies to other coding parameters than the quantization step. The only limitation on the coding parameter to choose is that it has a monotonous influence on the cost of coding the macroblock (or,

  more generally, the area considered.

  In addition, this process is independent of the type of yarn.

  trage, psycho-visual or not, used.

  Because of its iterative nature, the method according to the invention is well suited to the compression of images to be stored, for example for the production of DVD discs, or of storage on tape. But the method can also be used for real-time storage or real-time transmission, if the

  processing speed is high enough.

  The coding method according to the invention is not limited to

the MPEG process.

  It also applies to other compression processes

  such as JPEG (still image compression) or ETSI.

  Instead of the discrete cosine function DCT, we can, as a variant, call on a Fourier transformation

  fast or a separation of the spectrum in sub-bands.

  The compression method according to the invention also applies to object-oriented coding, MPEG4 for example. In this case, the quality factor consists of a parameter giving

  the interest of the image regions; alternatively (or in combination

  sound), the various objects of the image (multimedia) are treated individually so as to homogenize the quality factors

of each of its objects.

  Whatever its application, the process and the dis-

  positive according to the invention improve the sub-

  jective of coded video sequences.

Claims (18)

  1. Method for coding digitized images with a view to their compression in which the image is separated into zones, such as macroblocks, individually coded, characterized in that filtering (10) is applied to the source image and filtering at the same image resulting from the decoding of the coded image, so as to determine a quality factor for each zone, representing the difference (20) between the decoded image and the source image, and in that this quality factor is used to modify a coding parameter for each zone so as to minimize the differences between   quality factors of the various areas.   2. Method according to claim 1, characterized in that an average quality factor for the whole of the image is determined and in that the modification of the coding parameter for each zone is carried out so as to minimize for each zone the difference between the corresponding quality factor and the average quality factor.   3. Method according to claim 1 or 2, characterized in that during a first step, a coding parameter of determined value is assigned to all the zones, and in that the coding parameters are determined for the 'next step   by modifying this parameter in such a way that it tends to approximate   expensive the quality factor of each area of the average quality factor. 4. Method according to claim 3, characterized in   which during the first stage, we vary by an incré   The minimum coding parameter for each zone.   5. Method according to claim 3 or 4, characterized   in that during a second stage, or a subsequent stage   lower, the coding parameter which will be used for the next step is determined from a predictive model according to which the variations of the coding parameter are, for each   area, proportional to variations in the quality factor.   6. Method according to any one of the claims   previous, characterized in that the cost of coding the image is predetermined or controlled.   7. Method according to claims 5 and 6, character-   The fact that during the second step, or a later step, the coding cost for the next step is determined from a predictive model according to which the variation in the coding cost is, for each zone, proportional to the variation the coding parameter.   8. Method according to claim 7, characterized in that the coding parameters are determined for each zone so that in the next step the variation of the coding cost   is minimized compared to the current stage.   9. Method according to claim 8, characterized in that during the second step, or of a subsequent step: the coding parameter for each zone is determined by iteration by modifying the coding parameter of a first determined fraction of zones having the best quality factor so as to reduce this quality and / or by modifying the coding parameter of an   second determined fraction of areas with a lower qua-   In order to improve this quality, the variation of the coding cost which results from this (or these) variation (s) of the coding parameter is determined, and,   we modify the number of zones for which we modify   the coding parameter, so as to minimize the difference by   relative to the desired coding cost of the image.   10. Method according to claim 9, characterized in   that during each iteration step modifying by an increment   the coding parameter of the zones with the best and / or the least quality factor, the average of the quality factor deviations for all the zones is determined with respect to   to the average quality factor and in that the iteration is inter-   broken when this average no longer tends to decrease.   11. Method according to any one of the claims   previous, characterized in that the coding parameter is the quantization step of the value assigned to the coefficients   attached to each pixel in an area.   12. Method according to any one of claims   previous, characterized in that the filtering is of the psycho-visual type. 13. Method according to claim 12, characterized in   what psycho-visual filtering takes into account the effect of mas-   spatial and / or movement and / or temporal quage.   14. Method according to any one of the claims   1 to 11, characterized in that the filtering determines the ratio image noise signal.   15. Application of the method according to any one of   previous claims to an MPEG2 type coding.   16. Method according to any one of the claims   1 to 14, characterized in that the coding being of the object oriented type, the quality factor is linked to the degree of interest of the regions.   17. Method according to any one of the claims   1 to 14, characterized in that the coding being of the object oriented type, the quality factor homogeneity is optimized on   each of the objects taken individually.   18. Application of the method according to any one of   previous claims for storage and / or transmission- images.