WO2006087479A2

WO2006087479A2 - Video and/or image data processing with determination of representation functions

Info

Publication number: WO2006087479A2
Application number: PCT/FR2006/000367
Authority: WO
Inventors: Michel Fliess; Cedric Join; Mamadou Mboup
Original assignee: Ecole Polytechnique; Universite Rene Descartes; Universite Henri Poincare
Priority date: 2005-02-18
Filing date: 2006-02-17
Publication date: 2006-08-24
Also published as: WO2006087479A3

Abstract

The invention relates to video and/or image data processing with determination of representation functions. According to the invention, one device (DT) is dedicated to processing n-dimensional input data that are representative of an image or a sequence of images with n dimensions corresponding to n directions, whereby n = 2. The invention comprises processing means (MT) which, upon receipt of input data, perform the following steps consisting in: (i) defining one or more connecting n-dimensional domains having complementary shapes which are defined by position data and containing input data; (ii) subsequently, for each n-dimensional domain, determining a representation function that is representative of the set of input data contained therein and constructed from basic functions selected from at least one database of basic functions and respectively associated with selected coefficients, whereby said representation functions can be connected in pairs at least; and (iii) transforming at least some of the coefficients of at least one of the representation functions into representation data which, together with the associated position data, define output data that are representative of the input data.

Description

PROCESSING IMAGE AND / OR VIDEO DATA BY DETERMINING REPRESENTATION FUNCTIONS

The invention relates to the processing of data representative of an image or a sequence of images constituting, for example, apart from sound, all or part of a video.

By "image" is meant here any type of data that can be displayed on a screen or reproduced on any type of medium, such as for example a sheet of paper in the case of a photocopier or a fax machine.

Moreover, here "data processing" is understood to mean any type of processing likely to modify the number of data representative of at least part of an image (or sequence of images) and / or to modify the representation of at least a part of an image (or sequence of images) and / or to determine information relating to at least part of an image (or sequence of images). It may especially be a compression, a spatial and / or temporal transformation, a tattoo or a contour detection or movement. As known to those skilled in the art, many techniques have been proposed to process data, generally digital, representative of an image or a sequence of images, so as to facilitate their analysis, and / or their manipulation, and / or storage on a medium (or medium) of computer storage, such as a compact disc (CD), a video disc (DVD or VCD), or a digital photographic equipment memory or terminal (computer, mobile phone, fax machine, photocopier, game console, or television), and / or their transmission within a telephony network and / or a computer network, such as the Internet. Thus, in the field of compression / decompression several techniques have been subject to standardization. This is particularly the case of the JPEG technique for still images and the relative MPEG technique animated images (or sequences of images). These techniques, which have become standards, have been the subject of many changes that make it possible to achieve very high compression ratios. But, obtaining these high compression rates is always to the detriment of the quality of restitution. In other words, when it is desired to obtain uncompressed quality images, it is necessary to choose relatively low compression ratios which are currently incompatible with the storage and / or processing capacities of certain terminals and / or the capacities transmission of current networks. Moreover, these methods of compression, like the corresponding decompression methods, are time-consuming, so that they are hardly compatible with uses in real time.

In addition, the implementation of known compression and decompression methods generally requires components such as DCT (Discrete Cosine Transform) type filters or coders and Run-Length-Coding (RLC) type encoders, which contribute to degrade the quality of the images.

In addition, these methods of compression and decompression do not facilitate other treatments intended either to transform all or part of an image, such as for example a rotation or a close-up, or to detect an outline or a movement, or to slow down the frame rate of a video

The invention therefore aims to remedy at least one of the aforementioned drawbacks. It proposes for this purpose a device dedicated to the processing of n-dimensional input data representative of an image or sequence of n-dimensional images corresponding to n directions, with n ≥ 2.

Here we mean "n directions", ie n directions of space related to a n-dimensional spatial coordinate system (in the case of a fixed image), or n-1 directions of space combined with a temporal direction and reported together at a spatio-temporal landmark with n dimensions (in the case of a sequence of images or animated images).

This treatment device is characterized by the fact that it comprises loaded processing means (MT), in case of receipt of input data:

defining one or more contiguous n-dimensional domains, of complementary forms defined by position data, and containing the input data, and then determining for each n-dimensional domain a representational function representative of all the input data that it contains and constructed from basic functions, selected from at least one base function base, and respectively associated with coefficients, said representation functions being connectable at least two by two, and

transforming at least some of the coefficients of at least one of the representation functions into representation data defining, with the associated position data, output data representative of the input data. The treatment device according to the invention may comprise other characteristics that can be taken separately or in combination, and in particular:

its processing means can be responsible for defining n-dimensional domains of identical general shapes chosen and of identical dimensions;

the processing means can then be instructed to transform into representation data only the coefficients which are associated with so-called crossed products of different basic functions (the other coefficients corresponding to the values of the representation functions on the boundaries of the contiguous domains, which can be deduced step by step);

as a variant, its processing means may be responsible for defining the n-dimensional domains i) by determining the partial and approximate k-th derivative in the n directions of the input data, with k≥1, then ii) by comparing certain at least partial k-th derivatives at a variation threshold chosen so as to determine each input datum, then called border datum, at which the threshold of variation occurs, and iii) constituting the n-domain domains. dimensional from pairs of neighboring boundary data in the n directions, at least some of the positions of the boundary data constituting the position data of the n-dimensional domains;

the processing means can then be instructed to transform all the coefficients into representation data;

the order k of the partial derivatives is for example equal to two (2);

the processing means may be responsible for determining each partial and approximate k-th derivative in one of the n directions from a Taylor development truncated on a selected number of segments in the direction in question, the set of segments following the direction in question covering the range of input data in that direction;

when the number of directions n is equal to two, the domains are preferably rectangular in shape; when the number of directions n is equal to three, the domains are preferably volumes of parallelepipedal shape;

its processing means can be responsible for constructing all the representation functions from the same basic function base;

its processing means may be responsible for constructing each representation function from a linear combination of the basic functions of the chosen base, a combination in which each basic function is weighted by a chosen coefficient;

the basic functions can be polynomials of orders between zero and M, with M≥1; the processing means may be responsible for constructing the representation functions (and therefore determining the coefficients that represent them) by means of an interpolation technique, for example of spline or least squares type;

in the presence of input data sets representative of complementary components of color image (s) (for example red, green and blue) its processing means may be responsible for processing the games independently of one another; its processing means may be responsible for constituting a file comprising at least the output data representative of the input data. The file is then arranged as a sequence or array of output data groups representative of the input data contained in each of the n-dimensional domains, and each group includes representation data and eventual data. position data;

its processing means may be responsible for applying a spatial and / or temporal transformation to at least one of the representation functions before transforming at least the coefficients of each transformed representation function into representation data defining with the data of associated position of the output data; spatial transformations are for example chosen from rotations and close-ups (or "zoom"); its processing means may be responsible for applying a chosen mathematical operator to the input data so as to perform contour or motion detection. The mathematical operator may for example be a simple derivative with respect to one of the directions or a Laplacian or a divergence; its processing means may be responsible for hiding information at selected points (or pixels) of an image, corresponding to at least one of the representation functions (tattoo effect), before transforming at least some of the coefficients representation data representation functions defining with the associated position data output data.

The invention also proposes an image or video data encoder equipped with a processing device of the type described above.

This encoder may also include a post-compression module responsible for applying auxiliary compression to the output data that is output from the processing device to output a post-compressed data file.

The invention also proposes a device dedicated to the reconstitution of image or video data from input data resulting from a coding (or treatment) by means of a processing device or an encoder of the type of those presented above.

This reconstitution device is characterized in that it comprises reconstitution means having the base of basic functions and the construction technique which are used during the coding (or processing) and responsible for reconstituting each representation function which is associated with each domain from the representation data that is contained in each input data group and from the basic function database and construction technique (eg a linear combination), and then reconstruct (or reconstruct ) a main function from the reconstructed (or reconstructed) representation functions and position data contained in each corresponding input data group.

The means for reconstituting such a reconstitution device may be responsible for transforming the main function (and more precisely the representation functions that constitute it) into output data.

The means for reconstituting such a reconstitution device may also be responsible for decoding input data sets that correspond to complementary color image components (for example, R, G, B), in order to deliver output datasets. Finally, the invention also proposes an input data decoder representative of an image or video and resulting from an encoding by means of a processing device or an encoder of the type of those presented above, equipped with a reconstitution device of the type shown above. This decoder may also include a pre-decompression module loaded, when receiving post-compressed input data, to apply auxiliary decompression to the input data received, so as to supply the decompression device with input data. Other features and advantages of the invention will emerge on examining the following detailed description, and the accompanying drawings, in which: FIG. 1 very schematically and functionally illustrates an example of embodiment of an encoder equipped with an exemplary embodiment of a treatment device according to the invention,

FIG. 2 is an example of an algorithm detailing the main steps of a processing method that can be implemented by a processing device according to the invention,

FIG. 3 is an example of an evolution diagram along a direction X of a part of a partial derivative of order k (d ^k S / dX _n ^k ) of an input signal (S),

FIG. 4 very schematically illustrates a part of the decomposition of an image (I) into domains Di (here rectangular), this decomposition resulting in particular from the evolution diagram of FIG. 3,

FIG. 5 very schematically and functionally illustrates an exemplary embodiment of a decoder equipped with an exemplary embodiment of a reconstitution device according to the invention; FIG. 6 is an example of an algorithm detailing the main steps of a reconstitution process (or decoding) that can be implemented by a reconstitution device according to the invention.

The attached drawings may not only serve to complete the invention, but also contribute to its definition, if any. Referring first to Figures 1 to 4 to describe a processing device DT according to the invention and the processing method that it implements.

As illustrated in FIG. 1, the processing device DT according to the invention comprises a processing module MT which is responsible for performing at least one processing on input data representative of an image I or a sequence of images constituting for example all or part of a video.

The input data, which feeds the processing module DT, are for example arranged in the form of a binary (or digital) file, also called source file. Such a file may come from any type of source, including digital photographic equipment, fax machine, photocopier, camera phone, recorder a compact disc (CD) or videodisc (DVD or VCD), a game console, or a television, or a medium (or medium) for computer storage, such as a compact disc or a videodisk .

The source file is for example supplied to the processing device DT by a coder CR in which it is implanted and which is coupled directly or indirectly to the source or which is part of it. This is particularly the case of the embodiment, non-limiting, illustrated in Figure 1.

The invention relates to the processing of n-dimensional data, that is to say representative of an image or sequence of n-dimensional images, with n ≥ 2.

These n dimensions correspond either to n directions of space related to a spatial coordinate system defined by n vectors, as in the case of a fixed image, or to n - 1 directions of space combined with a temporal direction and reported together to a spatio-temporal landmark defined by n vectors, as in the case of a sequence of images (or moving images).

In the case of a fixed image n is generally equal to 2. The two dimensions then correspond to two directions of space, perpendicular to each other and referenced X and Y.

In the case of a sequence of images, n is generally equal to 3. The three dimensions then correspond to two directions of space, perpendicular to each other and referenced X and Y, as well as to the temporal direction referenced t.

But, one can envisage situations in which the images have more than two spatial dimensions (n> 2), as for example the so-called 3D images (three-dimensional) or the images belonging to curved spaces.

Furthermore, the invention also relates to monochromatic images, each pixel of which is defined by a (data) position and a (data) of chromatic value (for example between 0 and 255 in the case of a gray component ), that the polychromatic images, each pixel is defined by a (data) position and by three (data) chromatic values (corresponding for example to the colors red, green and blue, and for example between 0 and 255). In the case of an image polychromatic, the source file is for example in the form of a concatenation of three subfiles (or sets of input data) each corresponding to one of the three color components.

We consider in the following the case of monochromatic images. The processing of polychromatic images may, for example, consist of individually processing each set of input data corresponding to one of the three color components, and then concatenating at the end the three (processed) output data sets. But other methods can be used.

According to the invention, when the processing module MT receives a source file containing input data, it starts by defining one or more contiguous n-dimensional domains, of complementary forms defined by position data, and containing the data of entry to be processed.

The case of a single domain corresponds to situations in which the image is for example uniform monochrome (typically all black). In other cases, we define several domains (at least two).

Here, joined domains are domains that have a common side in the two-dimensional case (n = 2) or a common face in the other cases (n> 2). Moreover, here "complementary shape domains" are understood to mean domains which, once united, cover the entire surface of an image (I) or the entire volume of a sequence of images.

For example, in the two-dimensional case (n = 2), the domains can be rectangles and in the three-dimensional case (n = 3) the domains can be volumes such as rectangular parallelepipeds.

We shall return later on the modes of obtaining the joined domains Di.

As illustrated in FIG. 1, the processing module MT may comprise an SMD cutting sub-module dedicated to the determination of the domains Di and to the input data which they respectively contain. This SMD cutting sub-module is for example coupled to the input of the processing module MT and it delivers on its output the position data defining each domain Di. Once it has determined the join domains Di (for example by means of its SMD sub-module), the processing module MT determines for each of them a representation function FRi representative of the set of data. input that it contains. More precisely, each representation function FRi is constructed from one or more basic functions chosen within at least one basis of basic functions that are respectively associated with selected coefficients.

Numerous bases can be envisaged, and in particular bases of polynomial functions or sinusoidal functions. In the case of a fixed image related to the two spatial directions X and Y, it is possible, for example, to choose as base of basic functions the polynomial base whose components (or basic functions) are defined by the relation x ^α .y ^β , where α and β are integers bound by the relation α + β ≤ M. This polynomial base can comprise any polynomial of degree m between zero and M, with M≥1.

In the presence of such a base, a representation function FRi, associated with a domain Di, can for example be constructed by linear combination of the basic functions weighted by coefficients. In this case and when M is equal to 3, the representation function FRi associated with a domain Di is for example in the form:

FRi = ao _O + aox + ao-iy + a-nxy + 820X ^{2 +} 8

02 ^ + + 812Xy ^{2 +} 830X ^{3 +} a _O 3y ³ where the a _pq (p, q = 0 to 3) are the coefficients associated with the basic (polynomial) functions. These coefficients are for example real numbers.

In the case of an image comprising 512 pixels along X and 512 pixels along Y, then 1 ≤ x, y ≤ 512, for example. In practice, for a digital processing problem, the 512 pixel positions are coded between 0 and 1

In the case of an image sequence related to the two spatial directions X and Y as well as to the temporal direction t, one can for example choose the polynomial base whose components are defined by the relation x ^α .y ^β .t ^v , where α, β and y are integers bound by the relation α + β + y ≤ M, for example M =

3.

In the presence of such a base, a representation function FRi, associated with a domain Di, can for example be constructed by linear combination of the basic functions weighted by coefficients. In this case and when M is equal to 3, the mapping function associated with a FRi domain Di is for example in the form: FRi = a o + _OO _OO have x + a ⁺ _O ioy a∞it + an _O xy + aioixt + a _O nyt + amxyt + a ₂ od ^ + 8020 / + aocet ² + 8210X ² Y + a-i2oxy ² + aio2Xt ² + ao ^ yt ² + a ₀ i2yt ² + a3oox ³ +

where the _pqr (p, q, r = 0 to 3) are the coefficients associated with the basic (polynomial) functions. These coefficients are for example real numbers. In the case of a sequence of T images comprising 512 pixels along X and 512 pixels along Y ₁ then 1 ≤ x, y ≤ 512 and 1 ≤ t ≤ T, for example. In practice, for a digital processing issue, the 512 pixel positions are coded between 0 and 1.

The polynomial functions are advantageous because of the simplicity of their implementation. In the presence of such basic functions, it is indeed possible to construct representation functions FRi, that is to say, to determine the coefficients associated with the basic functions, by means of a spline interpolation or lower interpolation technique. square.

The higher the order M of the polynomials, the more the domains Di can be large and therefore the fewer they are to cover an image (or sequence of images). However, the higher the order M of polynomials, the greater the number of coefficients to be determined (and stored). When processing is about data compression, a compromise must be found. For example you can choose an order M equal to 3. Although we can choose different groups of basic functions from a basic functions basic to the construction of FRi representative functions associated with different areas Su ₁ is however simpler to construct all FRi representation functions with the same group of (or with all) basic functions chosen in a database.

Moreover, it can be envisaged that the processing module MT has several different bases and that it chooses one of them according to the needs corresponding to the input data to be processed. Each representation function FRi thus determined approaches, as far as possible, the portion of the source signal S which is defined by the input data contained in the corresponding domain Di.

It is important to note that the FRi representation functions must satisfy boundary conditions at the sides or faces that their respective Di domains share with their neighbors. Consequently, there is a continuity, in the mathematical sense of the term, of the representation functions FRM and FRi, which correspond to adjoining neighboring domains.

Di-1 and Di ₁ at their side or common face (e). These representation functions FRi, which are connectable at least two by two and which correspond to all the domains Di covering an image (or a sequence of images), is once assembled a main function FP.

In the presence of adjacent neighbor domains Di-1 and Di identical (in shape and dimensions), the continuity at their common border is for example ensured by means of equalities which link the coefficients of their representation functions FRM and FRi.

In the presence of adjacent neighbor domains DM and Di different, the continuity at their common border is for example obtained by applying a weighted least squares technique. Border errors are artificially increased by weighting to be better minimized later.

Once the processing module MT has constructed all the representation functions FRi, it transforms at least some of the coefficients, which define, in the base used, at least one of these representation functions, in representation data which constitute with the position data (which defines the respective shapes and locations of the associated Di domains) output data representative of the input data of the source file.

It is important to note that the number of coefficients that must be transformed into representation data depends on the type of processing performed and the type of Di domain chosen.

In fact, a compression-type process aims to reduce the number of output data as much as possible, given the desired rendering quality. A contour or motion detection or close-up (or "zoom") type processing may possibly require greater precision and therefore more coefficients to represent each representation function. On the other hand, as will be seen below, if all the domains Di are of general shapes and of identical dimensions and the representation functions FRi are constructed by linear combination of basic functions, then we need not take into account (For example, to transform them into representation data) the coefficients that are associated with products, called cross-products, of different basic functions (for example, and in a nonlimiting manner, xy or xy ² or xyt or X ² Q. Joining Di domains sharing one of their sides or one of their faces (so-called boundary), the other coefficients that correspond to the values of the representation functions on the different boundaries of the associated domain can be deduced. step by step from the values of the coefficients associated with the crossed products .- - This-allows: to reduce the number of coefficients to be transformed into data of n and therefore significantly increase the compression ratio when this is the goal. The compression ratio therefore depends on the compromise found between the size of the domains (and therefore their number) and the number of coefficients defining each representation function.

In the presence of domains of identical general shapes, but of different dimensions, the joined Di domains generally share only part of their common border (side or face). Therefore, unlike in the previous case (which is a special case of the present case), the coefficients, which are not associated with the crossed products, do not correspond to the border level and therefore can not be deduced step by step. In this general case, all the coefficients that define each representation function FRi must be taken into account (for example to transform them into representation data).

As illustrated in FIG. 1, the processing module MT may comprise a construction sub-module SMC dedicated to the construction of the representation functions FRi which correspond to the domains Di. This sub- SMC construction module is for example coupled to the input of the processing module MT and the output of the SMD sub-module, and it delivers on its output the coefficients (or representation data) of each representation function. corresponding to each domain Di. Once the representation data has been determined (for example by means of the SMC construction sub-module), several situations can be envisaged.

It is indeed possible to directly constitute an output file from the representation data and the associated position data. This is especially what we do when we simply want to compress or denoise input data.

The coefficients of at least one of the representation functions can also be processed so as to apply to them a spatial and / or temporal transformation, and then possibly constitute an output file to

From the unprocessed and / or transformed representation data and associated position data.

___ ta ^{^} constitution of un'fichier-output-is-for example to form UNE-sequence (or array) of G groups of output data representative of the input data contained in each selected domain Di . Each

Group Gi is associated with a domain Di and comprises the representation data defining (at least partially) the corresponding representation function FRi (possibly transformed) and any position data at least partially defining said domain Di.

As will be seen below, the position data is only necessary when the domains are of identical general shapes but of different dimensions.

Indeed, in the presence of identical domains Di, the fact of placing the groups Gi of representation data in a chosen order, which

30 corresponds to the arrangement of said domains Di relative to the source image (or sequence of images), enough to indicate the Di domains to which they correspond. It is also possible to associate with each group Gi of representation data a position data representing the number order of the domain Di corresponding within the division.

When the domains D 1 have identical general shapes but different dimensions, it is essential that each group G 1 of representation data, corresponding to each domain D _1, is accompanied by position data defining at least its dimensions.

For example, in the case of Di domains of general rectangular shape must be provided for each Di area positions within the image of two of its corners which are opposite in a diagonal.

It will be appreciated that because of the type of input data processing that results in the determination of coefficients representing functions, the number of representation data is very small relative to the number of input data. Therefore, the main processing of the invention almost systematically induces data compression.

As illustrated in FIG. 1, the processing module MT may include an SMG file generation sub-module dedicated to the constitution of the output files.

£ ___ ^. _ |, _ | ___ _j_ treatment of ^~ ^~ data - of - representation - (complementary determining data representation), the sub-SMG file generation module is for example coupled to the outputs of the construction sub-module SMC and SMD cutting sub-module.

When providing one or more complementary processing (to the determination of representation data), it is possible to insert an SMT transformation sub-module between the SMG file generation sub-module and the SMC construction sub-modules and - SMD cutting module.

Since each image is represented by representation functions FRi, defined by coefficients relative to a base function base, it is therefore particularly easy to apply to at least one of them one (or more) processing (s). ) complementary. As mentioned above, these complementary treatments may for example consist of modifying the representation of at least part of an image (or sequence of images) and / or of determining information relating to at least part of an image ( or sequence of images). Thus, the invention makes it possible to transform spatially and / or temporally all or part of an image or a sequence of images, for example to perform a rotation or a close-up ("or zoom"). Geometric transformations, such as orthogonal geometry isometries, and time transformations for image sequences are easily calculated by performing the corresponding variable changes on the basic functions used to construct the representation functions FRi.

For example, a digital zoom can be obtained without decreasing the resolution by the numerical values of the FRi representation functions taken in more points.

The invention also makes it possible to insert hidden (but not visually perceptible) information in an image or a sequence of images by a so-called "tattoo" technique. It is recalled here that the

Tattooing is to hide information at certain points (or pixels) of an image. Any type of information can be inserted, including information

- Representatives of the sign language (for the hearing impaired) or sound information.

This insertion of hidden information can for example be done in

20 imposing a tangent of slope equal to ± 1 to the representation functions FRi representing the points concerned, so as to set to 0 or 1 the corresponding bit. This imposition of tangent values constitutes an additional constraint when determining the coefficients that represent the representation functions FRi.

By thus transforming the representation functions FRi, an image containing hidden information can be formed. To recover the hidden information, it is necessary to perform tests on the reconstructed FRi representation functions to determine the points represented by bits at 0 and 1 where the tangents are equal to ± 1.

Other types of complementary processing may be envisaged, such as contour detection or motion detection, or the slowing down of the frame rate of images of a video (slow motion effect). The contour detection is performed by means of a chosen mathematical operator that is directly applied to at least a portion of the input data, but which could also be applied to one or more FRi representation functions. The mathematical operator can for example be a simple derivative with respect to one of the directions (d / dX) or a Laplacian (d ² FRi / dX ² + d ² FRi / dY ² , case n = 2) or else a divergence (dFRi / dX + dFRi / dY, case n = 2). When the domains Di are determined by means of the derivatives, it is possible, for example, to use the matrix of the derivatives to perform the contour detection. In addition, we can perform a thresholding. The Laplacian operator is the one who provides the best results.

Motion detection can for example be obtained with a technique similar to that implemented for contour detection, but applied to the temporal direction. Alternatively, it can also be done following the path of a pixel, represented by a portion of a representation function, from one image to another. For example, in the case of polychromatic images, the motion detection may operate on the basis of at least one color change detection on a chosen pixel.

An idle without decreasing the number of frames per second can also be obtained.

When the images are of the polychromatic type, the SMG file generation sub-module may for example be responsible for concatenating the three sets of output data that correspond to the three color components to form the output files. When the three data sets contain common data, for example position data, some of them can be deleted in order to avoid redundancies. Of course, the number of output sets associated with a polychromatic image is a function of the color coding mode.

The sequence of the main steps of the treatment method described above is given in the simplified algorithm example illustrated in FIG. 2, and summarized below.

In a step 10, the processing module MT receives the source file to be processed. Then, in a step 20 it determines the domains Di which cut

(or recover) the entire image (or sequence of images) into domains

Di, as well as the input data of the source file which are contained in these areas Di ₁ for example by means of its submodule SMD.

Then, in a step 30, the processing module MT constructs each representation function FRi that approaches the input data contained in the corresponding domain Di, and transforms into representation data at least some of the coefficients that define, in the base used, at least one of the representation functions. For this purpose, he uses his SMC construction sub-module, for example.

At this stage, a complementary step 40 may be provided in which the processing module MT applies at least one complementary processing to the representation data of at least one of the representation functions so as to deliver transformed representation data (at least for some of them). He uses his-sub-module-of-transformation-SMT, -for example

Then, in a step 50, the processing module MT constitutes the output file from representation data (possibly transformed) and any associated position data, for example by means of its SMG file generating sub-module.

A complementary step 60 may be added after step 50 to subject the output data of the output file to auxiliary compression. Several auxiliary compression techniques, known to those skilled in the art, can be used, for example compression type

"Zip" or vector quantization applied to the vectors that represent the coefficients of the representation functions FRi. In this case, an auxiliary compression module MCA is provided at the output of the processing module MT.

As illustrated in FIG. 1, this auxiliary compression module MCA can be part of the CR encoder. But, it could also be part of the DT treatment device.

As indicated above, the joined domains Di can be obtained in at least two different ways. A first way, particularly simple to implement and economical in computing time, is to impose the same shape and dimensions identical to all areas Di covering an image (or a sequence of images). For example, one can choose domains Di having sides extending on 3 or 4 pixels.

In this case, the processing module MT has only to cut, in the above-mentioned step 20, the image (or the image sequence) in identical D1 domains.

A second, particularly effective way is to let the processing module MT determine, in the aforementioned step 20, the dimensions of each domain Di covering an image (or a sequence of images). In this case, it is preferable to impose a constraint on the general form of the domains Di, so that they all have the same general shape (for example a rectangle or a rectangular parallelepiped) but different dimensions.

In order to be able to determine the dimensions of the domains D1, the processing module MT, and for example its submodule SMD, must be specifically arranged.

The invention proposes that the determination of the dimensions of the domains Di takes place in three main stages.

In a first step, the processing module MT determines the partial and approximate k-th derivatives in the n directions X _n (d ^k S / dX _n ^k ) of the input data S, k being chosen greater than or equal to 1.

For example, in the two-dimensional case (n = 2), corresponding to the spatial directions X and Y, the processing module MT determines the partial and approximate ^kth derivative in the direction X _1, ie d ^k S / dX ^k , and the derivative k-th partial and approximate in the Y direction, ie d ^ / dY *.

The higher the order of the derivative, the more sensitive one is to the residual noise that is superimposed on the input data. However, the higher the order of the derivative, the greater the number of variations observed, and therefore the larger the number of domains Di will be important. Now, the more the number of domains Di is important, the more the number of coefficients to transform into output data is high (and therefore the smaller the compression ratio). A Compromise must therefore be found when the compression ratio has to be high. For example we can choose an order k equal to 2 or 3.

It is important to stress that a classical partial derivative calculation, that is to say approached by variation rates, is not adapted to the treatment mode of the invention, in particular because of the noise (structured and / or unstructured) which is superimposed on the representative data of the image and which together with these constitute the input data of the source file. In addition, exact calculation of the derivatives is not possible because the input data is discrete. Preferably, each partial and approximate k-th derivative in one of the n directions X _n (d ^k S / dX _n ^k ) is determined from a truncated Taylor expansion on a selected number J _n of segments SGj _n ( J _n = 1 to J _n ) in the direction X _n concerned.

The set of segments SGj _n in the direction X _n covers the extent of the input data in this direction X _n .

All segments SGj _n of a direction X _n have the same length. It is important to note that the length of a segment SGj _n in the direction X _n is decorrelated from the lengths of the domains Di in the same direction. This length is indeed fixed initially and the determination of the domains is done after the computation (approached) of the partial derivatives k-th.

On the other hand, the length of the segments SGj _n in the direction X _n is adjustable according to a compromise between the desired denoising and the reactivity with respect to the dynamics of the signal (input data). The operators used for calculating partial and approximate k-th derivatives are equivalent to low-pass filters, so that the larger the sliding window (mentioned above), the smaller the cutoff frequency.

In addition, some segments SGj _n of a direction X _n can overlap partially in order to use only convergent estimates of the derivatives resulting from the truncated Taylor development. In this case, one can for example use a slippery window type technique.

The mathematical formalism on which the calculation of the partial and approximate derivatives presented above is based is given in the Appendix. With regard to the definition of windows (the segments above) and the calculation of partial and approximate derivatives which follows, more detailed explanations may be drawn from patent application PCT / FR 2005/002165 filed on August 30 2005, and entitled "Signal compression method". For all purposes, and in particular to supplement the above explanations if necessary, the descriptive content of this earlier application is to be considered incorporated by reference to the present description, in particular for the end of its description (pages 20 to 35). and the corresponding drawings. If necessary, these elements will be physically incorporated in the present patent application.

Once the processing module MT has determined the partial and approximate ^kth derivatives in the n directions X _n (d ^k S / dX _n ^k ) of the source signal (input data) S to be compressed, it passes to the second main step. In this second step, the processing module MT begins by comparing the variations of at least some of the k-th partial derivatives d ^k S / dX _n ^k with a chosen threshold th (n), as illustrated schematically in FIG. figure shows that, starting from a given point, one monitors the amplitude of variation (maximum - minimum) of the derivative concerned, until the point where it exceeds (or reaches) the threshold.

The thresholds th (n) associated with each of the directions X _n can be equal or different.

The higher the values of th (n) thresholds, the smaller the number of coefficients representing the representation functions (and hence the higher the compression ratio). Indeed, when the thresholds are high, the output signal (output data) is equivalent to the definition of the domains Di (position data) and the associated coefficients. As a result, higher threshold values tend to induce a smaller number of Di domains. Moreover, the smaller the th (n) threshold values, the smaller the dimensions of the Di domains, and therefore the better the approximation of the source signal (input data) S by the main function FP consisting of the set of FRi representation functions. As a result of these observations, especially when the processing essentially concerns the compression of the input data, the respective values of th (n) thresholds must be chosen as a function of a compromise between the compression ratio and the coding noise. . The comparison made by the processing module MT is intended to determine each input data (or point), called the border data, at which a threshold crossing of variation th (n) (by higher value) of the threshold is observed. at least one of its n partial k-derivatives d ^k S / dX _n ^k . In the example illustrated in FIG. 3, for a line in a direction, the threshold th (n) is a threshold representative of a partial k-th derivative variation amplitude d ^k S / dX _π ^k . The comparison carried out by the processing module MT is intended to determine for each of the n directions X _n groups or blocks of input data (points) for which the difference between the two maximum and minimum values of their derivatives k- Part i d ^k S / dX _n ^k is less than or equal to the threshold of variation th (n). Each input data (point) at which a crossing of the th (n) threshold is observed (by higher value) is a raw boundary datum of a block that will be used to build the interface between two neighboring contiguous domains. , for the direction X _n considered. Each border datum (point) where we stopped is the last of a domain in the direction X _n considered. As already indicated, this is done for the variations of at least some of the partial k-th derivatives d ^k S / dX _n ^k at a chosen threshold th (n). All lines in each direction, or a portion thereof, can be processed according to a compromise between the compression ratio and the image quality after compression. In the example, domains are constrained to be rectangles. The limit point of the domain can therefore be the first encountered, on all lines, for which the threshold is crossed. A more advanced method can be used, for example statistical (as "10% of the lines have crossed the threshold", this example being purely indicative). We will come back to it later.

Once all the frontier data has been obtained in a first direction, for example X _n = X, the above operation is performed. (Partial derivation and threshold comparison) for another of the directions, for example X _n = Y, taking as starting input data each boundary datum determined for the previous direction (here X). Each first input data (point) at which is observed a crossing of the th (n) threshold of variation (by higher value) is then a border data for the Y direction. One stops then to derive in this direction. Y. The border datum (point) where we stopped is then the last of the domain in the Y direction. The first and last frontier data in each direction determine the shape and position of each domain. Once the processing module MT has determined all the boundary data, it proceeds to the third main step.

In this third step, the processing module MT constitutes domains Di from neighboring pairs of data in the n directions X _n . In the example illustrated in FIG. 3, it can be observed that over the distance between 0 and Xmax (which defines the extent of the image in the direction X _n ) six lxι lengths (Ixi to Ixβ) have been determined. These six lengths define, for example, the respective dimensions of the first six rectangular domains D1 to D6 schematically illustrated in FIG. 4 and contributing to the covering (or cutting) of the image I.

As indicated previously, the positions of the input data defining each domain Di need not all necessarily be stored. Only those from which the positions of the other input data can be deduced can be stored. For example, in the case of rectangular domains it is sufficient to store for each domain Di the positions within the image of the input data which are located at two of its opposite corners along a diagonal.

As seen in the example of FIG. 3, the comparison of the variations of the partial derivatives with the threshold (s) provides boundary data in each dimension of the image. More generally, these boundary data define multidimensional blocks or contours that are a priori of any shape. Taking the two-dimensional case, these outlines are lines. In the example of the constraint by rectangles, it is about adjusting rectangles without holes on these contours. Figure 4 shows a way to do it, as a simple example, in no way limiting. Here, the constraint by rectangles is interpreted as "the right edge of the domain D7 is fixed by the left edge of the domain D2"; alternatively, the constraint by rectangles could be interpreted as "the bottom edge of the domain D2 is at the same level as D1" and the right edge of D7 is determined in another way involving the outlines and / or the constraint by rectangles . In general, this is the algorithmic of contour adjustment on polygons (or other forms simply definable). As already noted, the rectangles are definable by two points; it is also the same for triangles, whose form is imposed to a near homothety. These algorithmic choices for applying the constraint by rectangles are also a compromise between the compression ratio and the image quality after compression. Moreover, at decoding, the rectangles (or other shapes) are traversed in a determined order. They can therefore overlap, the image portion reconstructed on a given rectangle is then partially "overwritten" by the image portion reconstructed on a next rectangle. Thus, "joined rectangles" here means that there are no significant holes (relative to the content of the image), but does not exclude the recovery. The output files that are delivered by the processing device DT

(or by the CR encoder) can be of low weight while allowing a good quality of restitution, which facilitates their storage and transmission by wave or by wire. In addition, as will be seen below, these files can be decoded very quickly and without requiring the provision of significant computing resources.

The processing device DT, and in particular its processing module MT, and / or the post-compression module MPC of the encoder CR, can be realized in the form of electronic circuits (or "hardware"), software modules (or computer or "software"), or a combination of circuits and software.

Reference is now made to FIGS. 5 and 6 to describe a reconstitution device DD according to the invention and the method of reconstitution DR that it implements. As illustrated in FIG. 5, the reconstitution device DD ₁ according to the invention comprises a reconstitution module MR which is responsible for decoding input data resulting from a coding (or processing) by means of the method of processing described above and representative of an image I or a sequence of images constituting for example all or part of a video.

It is important to note that the MR reconstruction module must include the basic functions of the database that is used during the processing as well as the technique used to build the FRi representation functions from these basic functions.

The input data, which feeds the reconstitution module MR, are arranged in the form of a binary (or digital) file, also called source file to be decoded.

Such a file can come from any type of source, and in particular from a CR encoder, possibly implanted in a digital photographic equipment, a fax machine, a photocopier, a telephone, a compact disc (CD) or videodisc recorder (DVD or VCD), a game console, or a television, or a medium (or medium) for computer storage, such as a compact disc or a videodisk. The source file to be decoded is for example supplied to the reconstitution device DD by a DR decoder in which it is implanted and which is coupled directly or indirectly to the source or which is part of it. This is particularly the case of the exemplary embodiment, not limiting, illustrated in FIG. 5. According to the invention, when the reconstitution module MR receives a source file containing input data to be decoded, it begins by determining each group of representation data Gi associated with any position data defining each domain Di.

As illustrated in FIG. 5, the reconstitution module MR may comprise a SMDG unbundling sub-module dedicated to this task. This SMDG unbundling sub-module is for example coupled to the input of the reconstitution module MR and it delivers on its output the different groups of representation data Gi and the domain definitions Di correspondents.

Then, the reconstitution module MR reconstructs (or reconstructs) each representation function FRi which is associated with each domain Di from the representation data groups Gi. This operation is particularly simple and fast thanks to the invention.

It suffices to transform the representation data of each group Gi into coefficients, then to multiply each coefficient by the corresponding basic function, and finally to combine the results of these multiplications according to the construction technique used during the processing (or coding). For example, the construction technique is a linear combination. The main function FP (initial) is thus reconstituted by putting end to end all the reconstructed FRi representation functions.

The reconstitution module MR then transforms the main function FP (or representation functions FRi) into output data whose respective positions are those of the pixels of the decoded or reconstructed image (or picture sequence).

As illustrated in FIG. 5, the reconstitution module MR may comprise a reconstruction sub-module SMR dedicated to the reconstruction of the representation functions FRi and to their transformation into output data. This SMR reconstruction sub-module is for example coupled to the output of the SMDG unbundling sub-module and it delivers on its output the different output data resulting from the transformation of the representation functions FRi (or of the main function FP).

Once the reconstitution module MR has determined all the output data (for example by means of the reconstruction sub-module

SMR), it can form an output file. To do this, it may for example constitute a sequence (or a table) of output data representative of the pixels of the decoded or reconstructed image (or picture sequence).

As illustrated in FIG. 5, the reconstitution module MR can comprise an SMGF file generation sub-module dedicated to the constitution of the output files. This sub-module for generating files

SMGF is for example coupled to the output of the reconstruction sub-module

SMR and has an output constituting the output of the reconstitution device DD.

When the images are of the polychromatic type, the file generation sub-module SMGF may for example be responsible for concatenating the three sets of output data, which correspond to the three color components, to form the output files. When the three data sets contain common data, for example position data, some of them can be deleted in order to avoid redundancies. Of course, the number of output sets associated with a polychromatic image is a function of the color coding mode. It is important to note that the reconstitution module MR may optionally comprise an SMT transformation sub-module coupled to its reconstruction sub-module SMR. This SMT transformation sub-module is substantially identical to the SMT transformation sub-module of the processing module MT ₁ described above. It is for example responsible for processing the coefficients of at least one of the representation functions so as to apply to them a spatial and / or temporal transformation.

This transformation sub-module SMT 'can also reconstruct representation functions FRi which represent a wider range of colors. For example, you can double or triple the number of colors. To do this, it is necessary, as in the case of a zoom, to reconstruct each FRi representation function with a greater number of points in the color space.

The SMT transformation sub-module can also reconstruct representation functions FRi which represent a larger number of pixels, so as to increase the definition of the image. For example, you can double or triple the definition of the image. We can thus go from a definition of type 512x512 to a definition of type 1024x1024. To do this, it is necessary, as in the case of a zoom, to reconstruct each FRi representation function with a larger number of points. The sequence of the main steps of the reconstitution process

(or decoding) described above is given in the example of simplified algorithm illustrated in Figure 6, and summarized below.

In a step 100, the reconstitution module MR receives the file source to decode.

At this stage, when the data of the source file has been subjected to a treatment (or coding) according to the invention followed by an auxiliary compression, a pre-decompression step 110 is provided for applying to said data a decompression "Inverse" of said auxiliary compression.

In this case, a pre-decompression module MPD is provided upstream of the input of the decompression module MD. As illustrated in FIG. 5, this MPD pre-decompression module can be part of the DR decoder. But, it could also be part of the DD reconstruction device. In a step 120, the reconstitution module MR determines the groups of input (or representation) data Gi which are associated with the definitions of the domains Di which completely cut (or overlap) the image (or the sequence of images). .

Then, in a step 130, the reconstitution module MR reconstructs each representation function FRi of each domain Di.

At this stage, a possible step 140 may be provided in which the reconstitution module MR performs at least one treatment on the reconstituted representation functions FRi.

Then, in a step 150, the reconstitution module MR transforms the reconstituted and possibly processed representation functions into output data.

Finally, in a step 160 the reconstitution module MR constitutes the output file from the output data.

The reconstitution device DD, and in particular its reconstitution module MR, and / or the pre-decompression module MPD of the decoder

DR, can be realized in the form of electronic circuits (or

"Hardware"), software modules (or "software"), or a combination of circuits and software.

It is important to note that the invention provides a particularly elegant and effective means for attenuating structured and / or unstructured noises that are superimposed on the representative data of an image. This results from the fact that the transformation of input data into representation functions uses, among other things, iterated integrations which act as a filter lowpass. The noises concerned are not necessarily additive. In the case of multiplicative noise we can indeed use the trivial decomposition wl = I + (w-1) l, where w is the noise, which brings back the case of additive noise.

In the described embodiments, the data are considered monochromatic. For black-and-white (or black-and-white) images, these monochromatic data can be referred to as luminance data. As already indicated, in the case of color images, the format used for the data may comprise three chromatic components according to three complementary colors suitably chosen, such as the classic triplet {red, green, blue}, called RGB. It is also possible to use different video data format variants, such as combinations between a luminance component and at least two color components (chrominance) according to two complementary colors. The third component can be recovered by difference. Another example is the CCR601 digital television standard, in that it advocates a luminance and chrominance file approach.

Chroma and luminance are linked in the color space concerned. Accordingly, a format containing chrominance data, with or without luminance, can be considered to include data representative of complementary color image components.

The invention is not limited to the embodiments of treatment device, encoder, reconstitution device and decoder described above, only by way of example, but it encompasses all variants that may be considered by humans of the art in the context of the claims below. 30

Annex 1

1 Mathematical Framework

1.1 Images and videos

A video, or an animated image, is a vector distribution / = (I ₁ , ..., I _m ) on M ^m × R, with values in W. The coordinates xι,. . . , x _m of R ^m (resp. y _\ , ..., y _p of MP) are the variables of space (respectively the chromatic components), that t of E is the time. In general, m = 2 or 3, p = 1 or 3. The video is called monochromatic if p = 1. If § £ ≡ 0, 1 is a still or inanimate image. Let's make the following assumptions, which seem reasonable, in practice:

the projections of the support of I on M, ^m and R are respectively compact and closed in M ₊ ,

- in almost every point of M x x E, i is locally equal to an analytic function of the variables X _\,. . . , x _m , t.

1.2 Estimation of the derivatives 1.2.1 Case of a single variable Let a function of the time x (t), with real values, supposed analytic over the interval tι <t <* ₂ - To simplify, one will suppose x (t) analytic around t = 0. Introduce his truncated Taylor development

Approximate x (t) over the interval (0, e), ε> 0, by the polynomial xw (t) = of degree N. The usual rules of symbolic computation in the theory of

from Schwartz provide

≈ χ (0) δW + x (0) δ ^(N -V + ... +

where δ is the measure of Dirac in 0. The well-known equalities tδ = 0, tδ ^ = -aδ ^ ⁰¹ ^, a> 1, lead to the triangular system of linear equations to determine the estimated values [α; ^ (0 )] _e derivatives: _t ^a _χ ^(N + ⁱ⁾ ^ _{= t} ^a ([ _x (p)] _eδ W + [χ (0)] _e δ ^(N - ^ + • • • + [x ^ (Q) )] _e δ), a = 0, ..., N (l)

The derivatives of x (t), the Dirac measure and its derivatives are eliminated by integrating at least N times with respect to time the two members of (1):

+ ^{• • •} + [ ^{χ {N)} (o)] _e δ) (2) 31

for v> N, a = 0,. . . , N and where J ^ = J ₀ * / _o ^r "* ... / _Q ^T1 A very suitable estimate is obtained over a short time window [0, t].

Let us add that these integrals are low-pass filters, which make it possible to attenuate the noises.

1.2.2 Multi-variable generalization

Let A _n + it 'Weyl algebra of the linear differential operators

with polynomial coefficients, α _Q , ₁₎ ... _{) Q} , _kingJ g GM [xχ,. . . , x _m , t]. Let us introduce the Z) -rnodule 3 = (Ji, ..., 3 _P ) = sγ> a, n _{Am + 1} (I), which is the module on the left on A _7n ^ x generated by J.

Lemma 1.1 Suppose that in Xi = ζi,% = 1,. . . , m, t = r, £ α component i of I is locally equal to a polynomial function P ₁ GM [xι,. . . , x _m , t]. So, every coefficient c of the polynomial is given by

finite where vo GR [xi,. . . , x _m , t], i GJ ₁ . The iterated integral

/

is chosen such that / i = Σ _nn i _e / 'TtP ₁ , where J' is another iterated integral and π GR [xι,. . . , x _m , t].

The result is checked if m = 0. If m> 1, we go from m - 1 to m in a similar way. The effective expression of (3) follows from the demonstration.

Suppose now / _{t is} locally equal to an analytic function F _L. Let us note [c] _ejv (/ ii, • • • ih _m , k), the estimate of the coefficient c of a monomial of degree d of the development of Tayor F, in Xi = ξi, t = r, computed,

using the formula (3) obtained from the Taylor Taylor development of F in Xi = ξi, t = r, truncated to the order N, and replacing F ^ r with F;

- by integrating iteratively, according to (4), on the intervals] ξi, ξi + hi [, i = 1,. . . , m, and] r, T + k [.

The following proposition justifies the practical trade-offs between degrees of polynomials and expanses of windows: Proposition 1.2 If N> d, hm [c] _eN {hi _,. . . , h _m , k) ≈ Jim [c) _eN (h ₁ , ..., h _m , k) = c (5)

Claims

1. Device (DT) for processing n-dimensional input data, representative of an image or sequence of n-dimensional images corresponding to n directions, with n ≥ 2, characterized in that it comprises processing means (MT) arranged, upon receipt of input data, i) for defining one or more contiguous n-dimensional domains, of complementary shapes defined by position data and containing said input data, and then ii) for determining for each n-dimensional domain a representational function representative of the set of input data that it contains and constructed from basic functions, chosen in at least one base function base, and respectively associated with chosen coefficients, said representation functions being connectable at least two by two, and iii) for transforming at least some of the coefficients of at least one of said representation functions into representation data defining with the associated position data output data representative of the input data.

2. Device according to claim 1, characterized in that said processing means (MT) are arranged to define n-dimensional domains of identical selected general shapes and identical dimensions.

3. Device according to claim 2, characterized in that said processing means (MT) are arranged to transform into representation data the coefficients which are associated with products, said cross, of different basic functions.

4. Device according to claim 1, characterized in that said processing means (MT) are arranged to define said n-dimensional domains i) by determining the partial and approximate k-th derivative in said n directions of said input data, with k≥1, then ii) comparing at least some of said partial k-th derivatives to a variation threshold chosen so as to determine each input datum, then called border datum, at which a crossing of said datum threshold occurs. variation, and iii) to form said n-dimensional domains from neighboring pairs of boundary data in said n directions, at least some said positions of said boundary data constituting said position data of said n-dimensional domains.

5. Device according to claim 4, characterized in that said processing means (MT) are arranged to transform into representation data all the coefficients.

6. Device according to one of claims 4 and 5, characterized in that said order k is equal to two.

7. Device according to one of claims 4 to 6, characterized in that said processing means (MT) are arranged to determine each derivative k-th partial and approximated in one of said n directions from a development of Taylor truncated on a selected number of segments in the direction concerned, the set of segments in said direction covering the range of the input data in that direction.

8. Device according to one of claims 1 to 7, characterized in that n is equal to two, and in that said domains are rectangular in shape.

9. Device according to one of claims 1 to 7, characterized in that n is equal to three, and in that said domains are parallelepiped-shaped volumes.

10. Device according to one of claims 1 to 9, characterized in that said processing means (MT) are arranged to construct all said representation functions from the same base of basic functions.

11. Device according to one of claims 1 to 10, characterized in that said processing means (MT) are arranged to construct each representation function from a linear combination of the basic functions of the chosen base, in which each basic function is weighted by a chosen coefficient.

12. Device according to one of claims 1 to 11, characterized in that said basic functions are polynomials of orders between zero and M, with M≥L

13. Device according to claim 12, characterized in that said processing means (MT) are arranged to construct said representation functions by means of a spline or least squares interpolation technique.

14. Device according to one of claims 1 to 13, characterized in that in the presence of input data sets representative of color image complementary components (s) said processing means (MT) are arranged to process said games independently of one another.

15. Device according to one of claims 1 to 14, characterized in that said processing means (MT) are arranged to form a file comprising at least the output data representative of said input data, said file being arranged under the as a sequence or array of output data groups representative of the input data contained in each of said n-dimensional domains, and each group including representation data and any position data.

16. Device according to one of claims 1 to 15, characterized in that said processing means (MT) are arranged to apply a spatial and / or temporal transformation to at least one of the representation functions before transforming at least the coefficients of said representation functions transformed into representation data defining with the associated position data output data.

17. Device according to claim 16, characterized in that said spatial transformations are chosen from a group comprising at least rotations and close-ups.

18. Device according to one of claims 1 to 17, characterized in that said processing means (MT) are arranged to apply a selected mathematical operator to said input data so as to perform contour detection or movement.

19. Device according to one of claims 1 to 18, characterized in that said processing means (MT) are arranged to hide information at selected points (or pixels) of an image, corresponding to at least one representation functions, before transforming at least some of the coefficients of said representation functions into representation data defining with the associated position data output data.

20. Encoder (CR) image data or video, characterized in that it comprises a processing device (DT) according to one of the preceding claims.

21. Encoder according to claim 20, characterized in that it comprises a post-compression module (MPC) adapted to apply an auxiliary compression to said output data delivered by said processing device (DT), so as to deliver a file post-compressed data.

22. Device (DD) for reconstructing image or video data from input data resulting from encoding by means of a processing device (DT) or an encoder (CR) according to the one of the preceding claims, characterized in that it comprises reconstitution means (MR) having the base of basic functions and the construction technique used during the coding and arranged to reconstitute each representation function associated with each domain to from the representation data contained in each input data group and said basic function database and building technique, and then reconstructing said main function from said reconstituted representation functions and position data contained in each data group corresponding entry.

23. Device according to claim 22, characterized in that said reconstitution means (MR) are arranged to transform said main function into output data.

24. Device according to one of claims 22 and 23, characterized in that said reconstitution means (MR) are arranged to decode input data sets corresponding to complementary color image components, of to deliver output datasets.

25. Decoder (DR) of image or video representative input data obtained by means of a processing device (DT) or an encoder

(CR) according to one of claims 1 to 21, characterized in that it comprises a data reconstruction device (DD) according to one of claims 22 to 24.

26. Decoder according to claim 25, characterized in that it comprises a clean pre-decompression module (MPD), in the case of reception of post-compressed input data, to apply an auxiliary decompression to said received input data. , so as to supply said data reconstitution device (DD) with input data.