FR2993690A1 - Method for determining movement vectors or field model movement vectors between successive digital images for video compression, involves calculating vector or vectors field according to calculation that is replaced by another calculation - Google Patents

Abstract

The method involves calculating (C1) vector or estimated vectors field according to a first calculation by a sinusoidal regression, where the first calculation is reiterated by successive sinusoidal regressions until reaching a sufficient precision of the vectors or vectors field. The first calculation is replaced or supplemented by a second calculation (C2) when a condition is reached between brightness gradients according to a multiplication of average of the brightness gradients and a difference in brightness between corresponding images pixels is divided by a square of a gradient norm. An independent claim is also included for a video encoding and decoding system.

Description

The invention relates to a method for determining a motion vector or a motion model between two digital images according to the luminance data of the pixels of the images, as well as an encoding and decoding system. of motion or a movement model between two digital images according to the luminance data of the pixels of the images and an encoding and decoding system which comprises calculation means for the determination of motion vectors operating according to such determination method. More particularly, the method is performed by applying a sinusoidal regression supplemented or replaced by the projection of the motion vector on the orthogonal direction of the luminance gradients under certain conditions.

The search for a motion vector or a movement model is a subject of great importance in the image processing fields, especially for video compression, this also for medical imaging, geographical imagery, etc. Currently, the motion search methods are essentially based either on an exhaustive comparison pixel by pixel, or by minima searches of the Newton type. The exhaustive comparison has a very high calculation complexity, especially in the case where the distance of the movement is important, because this complexity increases with the square of the search distance. The exhaustive comparison therefore presents a calculation cost that is often prohibitive. Multiple partial search patterns have been developed to speed up the calculation by reducing the number of tests to be performed. Such partial searches are, for example, described in US-A2007 / 0206677 and US-A-2006/0018554.

The common problem with these accelerated search patterns is that by reducing the number of tests, a reduction in the accuracy of the search is also inevitable. These acceleration schemes are therefore only compromises, more or less effective, between the performance and the cost of calculation. The methods of the second type by minima searches of the Newton type give a mediocre result because of false minima which distort the determination of motion vectors or motion fields. Another motion search method is gradient descent. This method uses the partial derivatives of luminance to predict the motion vector. This is, for example, described in an article by Lurng-Kuo Liu et al. with the title "A block-based gradient descent search algorithm for block motion estimation in video coding, Circuits and Systems for Video Technology," IEEE Transactions on, 6/4, 419-422 (1996). This method of searching based on gradient descent is rapid, but the method relies on a Newton-Ralphson type estimate, so the result is often trapped by local minima. In addition, the formulas are relatively sensitive to noise in the images. As a result, the research results are not very accurate. EP-A-1 050 850 relates to the estimation of the predominant motion direction between two images and proposes to make this estimate to combine a peak detection algorithm and a multi-resolution regression detection algorithm in order to increase the accuracy of the estimate. At the kth iteration, a vector of motion Ok is calculated according to the vector Ok-1 according to the formula: Ok = Ok-1 + dOk. The iteration is repeated until the correction amplitude of the parameters of the overall motion becomes smaller than a threshold value. The same considerations as mentioned above apply to the teaching of this document. The problem of the present invention is to design a method for searching a motion vector or a motion model based on the luminance data of pixels in an image segment that gives the most accurate estimate. possible motion vector or motion pattern while keeping a reasonable cost. For this purpose, the subject of the invention is a method for determining motion vectors or a motion vector field model between two successive digital images according to the luminance data of the pixels of the images, the second image corresponding to the first image after a motion according to a vector or a motion vector field to be determined, the first image having its own vector or initial vector field, in which process after segmentation and matching of each segment of the two images, it is carried out calculating a vector or an estimated vector field of said vector or said vector field to be determined, pixel by pixel for all or part of the pixels of the segment, the vector or the estimated vector field being added to the vector or initial vector field, which method is characterized in that the vector or the estimated vector field is calculated according to a first calculation by regr sinusoidal pressure, the sinusoidal regression formula being the sum of the difference in luminance between pixels in correspondence of the first and second images multiplied by a first factor dependent on the partial derivative of the luminance and divided by a second factor depending on the square of the luminance gradient standard, divided by the sum of the squares of the first factor dependent on the partial derivative of the luminance and divided by a second factor dependent on the square of the norm of the luminance gradient, said first calculation being replaced or completed, when a condition is reached between the luminance gradients, by a second calculation according to the average of the luminance gradients multiplied by the luminance difference between corresponding pixels of the two images and divided by the square of the norm of the gradient. The technical effect obtained is an accurate and rapid estimation of a motion vector or a motion vector field in a sequence of successive digital images, which reduces the durations of the encoding and decoding processes. such a suite and also gives a high quality of final display images. The images obtained according to the method of the invention are displayed by means of a screen, for example, a liquid crystal screen or light-emitting diodes, the management of the display of the images on this screen being managed by an electronic card associated with the screen. The reduction of the duration of the encoding and decoding processes in a series of digital images can easily be verified by comparing the display images obtained according to the determination according to the present invention with those obtained according to a known determination of the state. of the technique.

Indeed, during an encoding of images, for each given source image, a prediction image is calculated by adding a calculated motion vector field to a previous reference image obtained from the previous source image, as a residual image is calculated from said given source image to which said prediction image is subtracted. Similarly, during an image decoding, display images are delivered from the data collected during the encoding, in particular from the residual image for each source image, the calculated motion vector field serving obtaining a prediction image, said prediction image being added to the residual image to provide a reference image for developing the corresponding display image. The calculation of a motion vector or a motion vector field thus has a direct consequence on the encoding time of a series of source images. In addition, the calculation of a motion vector or a motion vector field has an effect on the quality of the display image delivered after decoding, on the one hand during the encoding for the calculation the residual image that can be distorted by a not-precise estimate of the vector or the vector field and on the other hand during the decoding for the calculation of the reference image giving the display image.

This is easily verifiable by comparison between display images obtained with different vector determination methods or motion vector fields, as detailed in the example given in the description of the present patent application.

Advantageously, the estimation by the first calculation is repeated by successive sinusoidal regressions until a sufficient accuracy of the estimated vector or vector field is reached.

According to an advantageous variant of the method, the coordinates of the estimated vector are calculated as follows: - for which coordinates: f (pi) is the function of the luminance of the first image, g (p,) is the function of the luminance of the second image, F1 gx (pi) is the first factor dependent on the partial derivative in x of the luminance, F1 gy (pi) is the first factor dependent on the partial derivative in y of the F2 V,) 112 luminance and H ( p is the second factor dependent on the square of the norm of the luminance gradient.

According to an advantageous variant of the method, the first sinusoidal regression calculation is applied to the orthogonal basic functions of the motion model, in order to find the coefficients of the motion vector field according to this motion model. According to another variant of the method for a motion model, the motion model is a set of functions k (x, y) which make it possible to define a motion vector field at M (x, y) = (U (x , y), V (x, y)) using the coefficients ci and di according to the formulas: d For which: where for which coordinates: f (pi) is the function of the luminance of the first image, g (pi) is the function of the luminance of the second image, F1 gx (pi) is the first factor dependent on the partial derivative in x of the luminance, F1 gy (pi) is the first factor dependent on the partial derivative The luminance is the second factor dependent on the square of the norm of the luminance gradient. In another variant of the method, the first factors F1 gx (pi) and F1 gy (pi) are the derivatives, respectively. partial x and y of the luminance of the second image. 11 (p;) 112 In another variant of the method, is the norm of the gradient of the second image. In another variant of the method, the first factors F1 gx (pi) and F1 gy (pi) are respectively the average of the x and y partial derivatives of the luminance for the associated pixels of the first and second images. In another variant of the method, the average of the luminance gradients for the associated pixels of the first and second images. In another variant of the method, the relevance of the result of the first regression calculation is verified by calculating the sum of the squares or absolute values of the difference between the results and the sinusoidal curve, the estimation of the vector or the motion model. being successful if this sum is less than a predefined value. In another variant of the method, the images are filtered by suppressing too rapid changes in luminance, prior to the sinusoidal regression steps. In another variant of the method, the series of vectors or motion vector fields undergoing sinusoidal regression is weighted, said weighting being a function of either the gradient norm, the pixel position, or the gradient norm. and the position of the pixel. In another variation of the method, the first sinusoidal regression calculation is first applied to the reduced size images to obtain a vector or field of coarse preliminary motion vectors, which vector or preliminary vector field is then transferred to the original images for use as an initial vector or vector field, the first sinusoidal regression calculation then being applied to the original size images with said initial vector or vector field for obtaining the vector or vector. estimated vector field.

In another variant of the method, the condition reached between the luminance gradients for the replacement or complement of the first sinusoidal regression calculation by said second calculation is that the correlation of luminance gradients in the segment exceeds a predetermined correlation coefficient, said coefficient varying between 0.65 and 0.9.

The invention also relates to a video encoding and decoding system comprising means for encoding video images comprising calculation means for each given source image from the second source image of a prediction image from a motion vector field and then a reference image and a residual image from said prediction image, at least temporary storage means and residual image and field data transmission means motion vectors and decoding means for obtaining each display image, said system comprising calculation means for reconstructing the series of reference images for obtaining display images from the transmitted data relating to the vector fields and the residual images, which system is characterized in that the decoding means comprise calculation means for the determination of n of motion vectors or of a motion vector field model between two successive digital images for the implementation of such a method.

The invention will now be described in more detail but in a nonlimiting manner with reference to the appended figures, in which: FIG. 1 is a diagrammatic view of an encoding and decoding system according to the present invention, FIG. 2 is a schematic view of the various steps of the method for determining motion vectors or a motion vector field model between two successive digital images according to the luminance data of the pixels of the images.

A nonlimiting example of an embodiment of a video encoding and decoding system according to the present invention is given with reference to FIG. 1. In this figure, a video stream of source images and a vector field of movement relative to these images reach encoding means 1.

The encoding means 1 of video images comprise calculation means for each given source image of a prediction image, a reference image and a residual image. The calculation of the prediction image is done so that the prediction image is either identical to the source image or that obtained after application of the motion vector field to the previous or subsequent reference images. According to the present invention, the motion vector field is then determined by calculation means present in the encoding means 1 operating in accordance with the determination method according to the invention which will be described later. For the residual image, this can be obtained by the difference between the source image and the prediction image, while the reference image can be deduced from the prediction image by adding the residual image.

Then, the data relating to the residual images and the motion vector fields are stored by storage means which are referenced 2 in FIG. 1. These data can be transmitted to transmission means 3 which, via the intermediary of FIG. any communication network 4, in particular the Internet, sends the data to reception means 5.

The video encoding and decoding system according to the embodiment shown in FIG. 5 comprises decoding means 6 making it possible to decode the transmitted data. These decoding means 6 make it possible to obtain each display image and comprise calculation means for reconstructing the series of reference images used to obtain display images from the data transmitted relative to the fields of the display. vectors and residual images. Finally, the decoding system comprises display means 7 of the television screen type, computer screen or projector for viewing the display images. In this embodiment, the encoding means are separate from the decoding means but this is not necessarily always the case.

FIG. 2 illustrates the steps of the method for determining motion vectors or a motion vector field model between two successive digital images according to the luminance data of the pixels of the images, the second image corresponding to the first image after a motion according to a vector or a motion vector field to be determined, the first image having its own vector or initial vector field. In the method according to the invention, it is proceeded to the segmentation SI of the first source image and to the segmentation S2 of the second source image.

At the step referenced M in FIG. 2, each segment of the two images is matched. The next step is the calculation according to a first calculation C1 and / or a second calculation C2 of a vector or an estimated vector field of said vector or said vector field to be determined, pixel by pixel for all or part of the pixels of the segment. The calculation can be done in two different ways either by the first calculation C1 or the second calculation C2 or this calculation can be the complementary association of the first C1 and second C2 calculations. After one of these calculations C1, C2 or the combination of these calculations C1, C2, the vector or the estimated vector field is added to the initial vector or vector field.

During the first calculation C1, the vector or the estimated vector field is calculated by sinusoidal regression, the sinusoidal regression formula being the sum of the difference in luminance between pixels in correspondence of the first and second images multiplied by a first factor dependent on the partial derivative of the luminance and divided by the standard square of a second luminance gradient dependent factor, divided by the sum of the squares of the first partial luminance derivative dependent factor and divided by a second luminance dependent factor square of the norm of the luminance gradient.

As previously mentioned, the first calculation C1 can be replaced or completed, when a condition is reached between the luminance gradients, by a second calculation C2. This second calculation C2 is based on the average of the luminance gradients multiplied by the difference in luminance between corresponding pixels of the two images and divided by the square of the norm of the gradient. An example of a motion model is the case where the basic functions are the linear functions on the spatial coordinates of the images, and the sinusoidal regression gives the coefficients of a vector field of an affine transformation. For a digital image, luminance forms a two-variable function, f (x, y) for a first image.

Let the function g (x, y) corresponding to a second image coming from the first image after a uniform movement with a motion vector RT = (U, V), there then exists a relation between the functions f and g which is: f (x, y) = g (x + U, y + V). Let be the gradient on a point (xo, yo), cp = arctan (gy / gx) the angle = (h, 'hy) the normalized vector in the directional orientation of v, where / 11711 = eastcp) and h .- = gy 71i = sin (q)) Around the point (xo, yo), the limited expansion of the function gives: f (xo, yo) = g (xo-FU, yo + V) = g (xo , yo) -FUgx + Vgy + 02 where 02 is a term of order greater than or equal to 2 for u and y. So for an approximation of order 1, we can omit the term 02, to have: 1-igaVg) 41ff II II> II cost, where a is the angle between the motion vector M and the gradient. This means that the orthogonal projection of M over 7 is equal to the vector: 11, 11e), in the case where T is non-zero. The pixels n, P2, --- Pn of the segment of the image give a series of vectors: In such a case, the sinusoidal regression makes it possible to find a vector _). m = (u,) whose orthogonal projections on vectors 1/1/7 = (ai, bi) are as close to 11-17 as possible. The formulas of the sinusoidal regression are: V - In the case currently described, it is taken as first factors F1 gx (pi) and F1 gy (pi) respectively the partial derivatives in x and y of the luminance of the second image. It should be borne in mind that this is not limiting and that it can be taken, in the general formulas of the calculation of u and v given above according to the factors F1 gx (pi) and F1 gy (pi ), these factors F1 gx (pi) and F1 gy (pi) comprising said partial derivatives in x and y of the luminance of the second image while not being limited to said derivatives and may be other.

For example, without being limiting, it can be taken from the first factors F1 gx (pi) and F1 gy (pi) being respectively the average of the partial derivatives in x and y of the luminance for the associated pixels of the first and second images .

The same applies to the standard II F2 V (0112 which may be other than the norm of as a gradient of the second image, for example being the norm of the average of the luminance gradients for the associated pixels of the first and second images.

This regression always gives a result, except in the case where all the terms in the sum of a denominator are zero. In the latter case, a value corresponding to O is given as a value.

In the case where the two sets of partial derivatives (gx (pi), gx (Pi), gx (pi,)) and (gy (pi), gy (pi), gy (pn)) are not highly correlated and where the vector of motion M is not very important compared to 02, the vector obtained in "after the regression offers an excellent approximation of M. So for a recursive application of the regression, the convergence is fast.

In the case where the two sets of partial derivatives are strongly correlated, the situation is ambiguous and the restricted study in the segment can not determine the vector Fi unambiguously. In this case, it is possible, in accordance with the present invention, when a condition is reached between the luminance gradients to replace or supplement the regression by the orthogonal projection of M on the correlated gradient direction. The projection of ersur the orthogonal direction of the gradients is indeterminate, but it can be determined by a study of the situations in the neighboring segments.

Sinusoidal regression also applies to a motion model. Such a model of motion is a finite set of basic functions ki (x, y) (j = 1, ..., i, ..., t), preferably orthogonal, functions that make it possible to define a vector field of motion Ie (x, y) = (U (x, y), V (x, y)) using the constant coefficients ci and di (j = 1, ..., i, ..., t ): U (.r Cr. Y The following regression formulas are used to determine the coefficients c and di (p,) - g (pi) k (p,) / 117 (Pi) IF (Pi) (pi) / 11 (P;) Here again, the general formulas for calculating cj and dj are made from the factors F1 gx (pi) and F1 gy (pi) which are not limited to the partial derivatives in x and y of the luminance of the second image, being for example but not necessarily a function of the average of the partial derivatives in x and y of the luminance for the associated pixels of the first and second images.

The same applies to the standard II F2 V (pall2 which may be other than the norm of as a gradient of the second image, for example being the norm of the average of the luminance gradients for the associated pixels of the first and second images.

A special case of motion model is the case and ki (x, y) = x-x0, k2 (x, y) = y-yo where (x-xo) is the center of the segment and. In this case, the coefficients obtained by the sinusoidal regressions determine an affine transformation matrix.

Tests of the method according to the present invention have been performed on a few hundred pairs of images from video sequences of different types and levels of difficulty. The segments are square blocks whose sides are between 8 and 16 pixels.

In most cases, the starting motion distance is below 2 to 3 pixels and the reiterations with sinusoidal regression converge to a position with a residual motion vector below 0.2 pixel after 1 to 3 iterations.

To measure the positioning accuracy of the segments, pairs of good quality graphic images having regular movements as a difference were selected for a test. After process applications, segment by segment, the irregular variations in positioning between neighboring segments are measured. The average of these irregular variations is about 0.1 pixel, despite the fact that aggressive reiteration parameterization can reduce the residual motion vectors below 0.03 pixel. The irregular variations are thus probably due to noises caused by pixelation and previous compressions in the images. The tests show that the sinusoidal regression is able to recover the movement offsets up to twenty pixels depending on the situation. With a hierarchical search applying the sinusoidal regressions first on 3-fold reduced images and with neighborhood extensions, motions of more than 100 pixels are found without difficulty.

The complexity of the search for movements is essentially a function of the complexity of mutual movements and deformations of the scene. This complexity can be measured with the number of calculations of 11711 Il VII per pixel on average.

For a pair of images where the movements are mainly regular, without many objects in mutual and complicated movements, even if the scene contains movements of an affine or projective nature, the number of computations per pixel is generally between 2 and 4 , for motion determination with a maximum accuracy of 0.1 pixel. By reducing the accuracy to 0.3-0.5 pixels, this calculation number can be reduced to less than 1 per pixel.

For a scene with fast and highly complex movements such as a sports scene, the implementation of the process will put up to 8 to 10 calculations per pixel for maximum accuracy, because there are repetitive trials in regions with occlusions or deformities.

The capacity and accuracy of the process according to the present invention is illustrated by the following example. This example concerns images extracted from a video sequence, with a time difference of 8 images (1/3 second). This video consists of an aerial shot with a rotation of about 15 degrees, displacements up to 150 pixels, and mutual movements with occlusion of more than 50 pixels. The size of the images is 960x600 pixels, images that are deduced from a high-definition video sequence with 1920x1080 pixels.

This size reduction removes most of the artifacts created by the previous compression, to prevent these artifacts from distorting the result of the experiment. The calculation is based entirely on the two images, ignoring information in other images in the sequence that may help to increase accuracy.

After an average of 7.5 calculations per pixel, the method according to the present invention produces a vector field with a resolution of 3 pixels. Except for areas obscured by mutually moving objects, this vector field is correct and accurate in the vast majority of locations. But there are also small irregularities in the dark areas, indicating that the limit of the search capacity is almost reached for the parameters chosen in this test.

This precision of the vector field makes it possible to recreate the entire video sequence between the two images, with residual images reduced to a minimum. For a comparison, the H.264 process was tested on this video sequence, as implemented by the x264 software.

It is found that the search for motion in x264 is already very approximate with an interval of two images in the sequence, since in this case, the residual image (of type P) has a size equivalent to 25% of the totality of the image. image (type I image), and between two P-type images, the B-type image always requires a non-negligible residual image equivalent to 6% of the entire image. It is thus estimated that for this video sequence, a motion search based on the method according to the invention makes it possible to divide the sizes of the residual images of type P and B by at least 4, compared to the search method in the software. x264 for a probably shorter search time.20

Claims (15)

REVENDICATIONS1. Method for determining motion vectors or a motion vector field model between two successive digital images according to the luminance data of the pixels of the images, the second image corresponding to the first image after a movement according to a vector or a field of motion vectors to be determined, the first image having its own vector or initial vector field, in which process after segmentation (51, S2) and mapping (M) of each segment of the two images, it is computed ( C1, C2) of a vector or an estimated vector field of said vector or said vector field to be determined, pixel by pixel for all or part of the pixels of the segment, the vector or the estimated vector field being added to the vector or initial vector field, which method is characterized in that the vector or the estimated vector field is calculated according to a first calculation (C1) by sinuso regression dale, the sinusoidal regression formula being the sum of the difference in luminance between pixels in correspondence of the first and second images multiplied by a first factor dependent on the partial derivative of the luminance and divided by the square of the norm of a second luminance gradient dependent factor, divided by the sum of the squares of the first partial luminance derivative dependent factor and divided by a second square dependence factor of the luminance gradient standard, said first calculation (C1) being replaced or completed, when a condition is reached between the luminance gradients, by a second calculation (C2) according to the average of the luminance gradients multiplied by the luminance difference between corresponding pixels of the two images and divided by the square of the gradient norm . 2. The method of claim 1, wherein the estimate by the first calculation (C1) is repeated by successive sinusoidal regressions until a sufficient accuracy of the vector or the estimated vector field. 3. The method according to claim 1, wherein the coordinates of the estimated vector are calculated as follows: / r (P,) f II F2 (i) If In (Pi :) - g (Pi)) FLgy (Pi) / II F.71 (Pi) ir v i-1 11,2e, -1 for whose coordinates: f (pi) is the function of the luminance of the first image, g (pi) is the function of the luminance of the second image, F1 gx (pi) is the first factor dependent on the partial derivative in x of the luminance, F1 gy (pi) is the first factor dependent on the partial derivative in y of the luminance, Il F2 V (pi) 112 is the second factor depending on the square of the norm of the luminance gradient. 4. Method according to one of claims 1, 2 or 3, wherein the first calculation (C1) by sinusoidal regression is applied to the orthogonal basic functions of the motion model, in order to find the coefficients of the motion vector field. according to this model of movement. 5. Method according to claim 4, wherein the motion model is a set of functions k (x, y) which make it possible to define a motion vector field at M (x, y) = (U (x, y) , V (x, y)) using the coefficients ci and d; according to the formulas: ## EQU3 ## where (r, y) = Eci "1 / (27.y, E (4ki (1, y) for which: (pi) - g (pi) ) k, (pi) F gr (pi) / 1177e (pq k '(p1) F1g2r (p1) / 11 r77.- (p1) 1 (Pi) g (Pi) k, Fig, (p1) / 11 F2 (P;) I I2 = r-1 k2 (Pi) 11: g2e (pi) III F2 (Pi) where for which coordinates: f (pi) is the function of the luminance of the first image, g (pi) is the function of the luminance of the second image, F1 gx (pi) is the first factor dependent on the partial derivative in x of the luminance, F1 gy (pi) is the first factor dependent on the partial derivative in y of the luminance F 2 V (pi) II is the second square-dependent factor of the luminance gradient standard. 6. Method according to any one of claims 3, 4 or 5, wherein the first factors F1 gx (pi) and F1 gy (pi) are respectively the partial derivatives in x and y of the luminance of the second image. 20 2 V 7. The method of claim 6, wherein Il F (p) 112 is the gradient standard of the second image. The method according to claim 3, wherein the first factors F1 gx (pi) and F1 gy (pi) are respectively the mean of the partial derivatives in x and y of the luminance for the associated pixels of the first and second images. 9. A method according to any of claims 3 or 5, for which F 2 V (W 112) is the norm of the average of the luminance gradients for the associated pixels of the first and second images. The method according to any one of the preceding claims, wherein the relevance of the result of the first computation (1) by regression is verified by calculating the sum of squares or absolute values of the difference between the results and the sinusoidal curve. estimation of the motion vector or model being successful if this sum is less than a predefined value. The method of any of the preceding claims, wherein the images are filtered by suppressing too rapid changes in luminance prior to the sinusoidal regression steps. A method according to any one of the preceding claims, wherein the series of vectors or motion vector fields undergoing sinusoidal regression is weighted, said weighting being a function of either the gradient standard or the pixel position. , the gradient standard and the position of the pixel. A method according to any one of the preceding claims, wherein the sinusoidal regression calculation (C1) is first applied to the reduced size images to obtain a rough preliminary motion vector or vector field, said preliminary vector or vector field is then transferred to the original images to serve as an initial vector or vector field, the first sinusoidal regression calculation (C1) then being applied to the original size images with said vector or field initial vectors for obtaining the vector or the estimated vector field. A method according to any one of the preceding claims, wherein the condition reached between the luminance gradients for replacing or complementing the calculation (C1) by sinusoidal regression by said second calculation (C2) is that the correlation of the gradients of luminance in the segment exceeds a predetermined correlation coefficient, said coefficient varying between 0.65 and 0.9. A video encoding and decoding system comprising: video image encoding means (1) comprising calculation means for each given source image from the second source image of a prediction image from a motion vector field, and then a reference image and a residual image from said prediction image; - at least one temporary storage and transmission means (3 to 5); ) data relating to the residual images and to the motion vector fields, and decoding means (6) for obtaining each display image, comprising calculation means for reconstituting the series of reference images serving obtaining display images from the transmitted data relating to the vector fields and the residual images, which system is characterized in that the decoding means comprise calculation means for the determination of vector vectors. motion or motion vector field pattern between two successive digital images for carrying out a method according to any one of the preceding claims.