WO2015052064A1

WO2015052064A1 - Method for coding and decoding floating data of an image block and associated devices

Info

Publication number: WO2015052064A1
Application number: PCT/EP2014/071112
Authority: WO
Inventors: Dominique Thoreau; Mikael LE PENDU; Yannick Olivier; Christine Guillemot
Original assignee: Thomson Licensing SAS
Current assignee: Thomson Licensing SAS
Priority date: 2013-10-07
Filing date: 2014-10-02
Publication date: 2015-04-16
Anticipated expiration: 2016-04-07

Abstract

The invention relates to a method and device for coding, to a coded data signal, an image block whose image data are represented by floating values. Each floating value being expressed by a mantissa and an exponent, the method is characterised in that it comprises the following steps: - determining (10) a prediction block (B_pred) for the image block to be coded; - determining (20) a prediction residual error block (B_res) by extracting said prediction block from the image block to be coded; - determining (30) a unique exponent value (Exp_res) for the floating values representing the image data of the prediction residual error block (B_res), - coding (40) said exponent value thus determined (Exp_res) for the prediction residual error block (B_res), and - coding (50) the mantissas (M_res) of the floating values representing the image data of the prediction residual error block (B_res) by spatial and/or temporal prediction. The invention also relates to a method and device for decoding the coded data signal.

Description

METHOD FOR CODING AND DECODING FLOATING DATA OF AN IMAGE BLOCK AND ASSOCIATED DEVICES

1. Domain of the invention

The invention relates to the general domain of image coding. More specifically, the invention relates to a method for coding to a coded data signal an image block whose data are represented by floating values and a method for decoding a stream of coded data representative of an image block whose data are represented by floating values. The invention also relates to methods for coding and decoding a sequence of images and devices implementing these methods.

2. Prior art

This section is intended to introduce various aspects of the prior art which may be linked to various aspects of the invention which are described and/or claimed below. This section is useful since it provides readers with basic information which will facilitate a better understanding of the various aspects of the invention. However, this information is to be used for this purpose only, and must in no way be considered as a summary of the prior art recognised by the inventor.

Traditional methods for coding sequences of images are designed to code integral data.

These methods use temporal prediction or inter-image prediction and/or spatial prediction or intra-image prediction. Intra-image prediction (or inter-image prediction) is used to improve the compression of a sequence of images. It comprises, for a current image block, the generation of a prediction image block and the coding of a difference, also called residual image block, between the current image block and the prediction image block. The more the prediction image block is correlated with the current image block, the lower the number of bits required to code the current image block and therefore the more efficient the compression.

Examples of such methods include those defined by the MPEG-4 AVC/H.264 standard described in the ISO/I EC 14496-10 document, or those defined by the HEVC (High Efficiency Video Coding) standard described in the document (B. Bross, W.J. Han, G. J. Sullivan, J. R. Ohm, T. Wiegand JCTVC-K1003, "High Efficiency Video Coding (HEVC) text specification draft 9," Oct 201 2.).

Historically, standards such as AVC and HEVC are used to compress images and videos based on pixels whose components are initially coded on 8 bits. Each of the components in the set of pixels in these images is therefore represented with the same dynamic range and by definition these components are thus said to have the same exponent in a virtually floating representation. The same applies to prediction residual error which has an additional sign bit.

However, these methods are not suitable for encoding or decoding sequences of images whose data are represented by floating values as defined in general by the following equation :

C_p = M. 2^E (1 )

where C_p is one of the R, G or B components of a pixel, with M and E representing mantissa and exponent respectively.

Representations such as RGBE (with E meaning exponent) or OpenEXR RGB half float (cf K. Myszkowski, R Mantiuk, G Krawczyk "High Dynamic Range video," Synthesis Lectures on Computer Graphics and Animation, Morgan & Claypool Publishers 2008) comprise components which are coded with 5 bits of mantissa and 1 0 bits of exponent for the OpenEXR format and 8 bits of mantissa and 1 0 bits of exponent for the RGBE format.

More generally, however, floating data are expressed by:

• a mantissa comprising Nm bits which enables the mantissa values to extend between 0 and 2^Nm-1 , and

· an exponent of Ne bits which enables the exponent values to extend between 0 et 2^Ne-1

Thus, the values expressed with a floating representation composed of 1 sign bit, Nm mantissa bits and Ne exponent bits can vary from 0 to:

(2^A,m - l) x 2⁽²"ⁱ-¹⁾ .

It can be seen that the higher the floating value, the greater the quantisation of this value. Take as an example the case of a floating value expressed in the RGBE format (Ne = 8, Nm = 8). An exponent of value equal to 1 08 can be used to represent component values comprised between 0 and 0.000243187 with a quantisation step of 2.^£~128 = 9.53674E-07. Similarly, an exponent of value equal to 148 can be used to represent component values comprised between 0 and 267386880 with a quantisation step of 1048576.

This representation of image data by floating value shows that the quantisation of a component of an image changes with the amplitude of this component. Since this change follows a logarithmic scale, it is possible to represent a significant dynamic range of an image component with a restricted number of bits, admittedly at the cost of a variation of the quantisation in the linear domain.

3. Summary of the invention

It is sought to improve the coding of images composed of pixels coded by floating values through a predictive-type coding method.

For this purpose, the invention relates to a method for coding, to a coded data signal, an image block whose image data are represented by floating values. Each floating value being expressed by a mantissa and an exponent, the method is characterised in that it comprises the following steps:

- determining a unique exponent value for the floating values representing the image data of a prediction residual error block determined by extracting a prediction block from the image block;

- coding said exponent value thus determined for the prediction residual error block, and

- coding the mantissas of the floating values representing the image data of the prediction residual error block by spatial and/or temporal prediction.

According to an embodiment, said exponent value thus determined is coded without loss.

According to an embodiment, said exponent value is predicted by a prediction value and the difference between the prediction value and the exponent value is coded.

According to an embodiment, the prediction value is defined from a causal neighbouring area formed of at least one image block previously processed by said method. According to an embodiment, the causal neighbourhood is formed of a block belonging to the image to which the block to be coded belongs and/or of at least one block belonging to at least one other image different from the image to which the image block to be coded belongs.

According to an embodiment, the method comprises a step of selecting a coding mode for the image block to be coded from a set of coding modes during which step the coding mode selected is that which minimises a compromise between a minimum reconstruction error for this image block coded and then decoded according to this coding mode and a minimum coding cost taking into account the coding cost according to this mode of coding the exponent's unique value and the cost of coding the mantissas of the floating values representing the image data of the prediction residual error block.

The invention also relates to a method for decoding a coded data signal representative of an image block whose data are represented by floating values. Each floating value being expressed by a mantissa and an exponent, the method is characterised in that it comprises the following steps:

- decoding from the coded data signal an exponent value relative to a prediction residual error block relative to the image block;

- decoding from the coded data signal the mantissas of the floating values representing the data of the prediction residual error block by spatial and/or temporal prediction;

- reconstructing the floating values representing the image data of the prediction residual error block from said exponent value and said mantissas thus decoded;

- determining a prediction block for the image block; and

- reconstructing the floating values representing the image data of the image block from the prediction residual error block and the prediction block.

According to another of its aspects, the invention relates to a coded data signal representative of an image block whose data are represented by floating values. Each item of floating data being expressed by a mantissa and an exponent, the signal is characterised in that it comprises at least one item of coded data describing a single exponent value for floating values and at least one other item of coded data describing the mantissas of floating values representing the image data of the prediction error block of the image block.

According to another of its aspects, the invention relates to a device for coding, to a coded data signal, an image block whose image data are represented by floating values. Each floating value being expressed by a mantissa and an exponent, the device is characterised in that it comprises the following means for:

According to another of its aspects, the invention relates to a device for decoding a coded data signal representative of an image block whose data are represented by floating values. Each floating value being expressed by a mantissa and an exponent, the device is characterised in that it comprises the following means for:

- determining a prediction block for the image block; and

4. List of figures The invention will be better understood and illustrated by means of non-restrictive embodiments and advantageous implementations, with reference to the accompanying drawings, wherein:

- Fig. 1 depicts an example of the definition of an image block according to the HEVC standard;

- Fig. 2 depicts a diagram showing the steps of a method for coding an image block whose image data are represented by floating values;

- Fig. 3 depicts a diagram showing the steps of a method for coding an exponent value;

- Fig. 4 depicts an example of the causal neighbourhood of an image block;

- Fig. 5 depicts a diagram showing the steps of an embodiment of the method for decoding a coded data signal; and

- Fig. 6 depicts an example of the internal architecture of a device. 5. Detailed description of the invention

The invention relates to a method for coding an image block B_c {whose data are represented by floating values to a coded data signal F, a method for decoding a coded data signal F representative of an image block B_c whose data are represented by floating values and a coded data signal F representative of a prediction error block relative to an image block B_c whose data are represented by floating values.

An image block is from an image which itself may be from a sequence of images. An image block comprises pixels or image points with each of which at least one item of image data represented by one (or several) floating value(s) is associated. An item of image data is, for example, an item of luminance data or an item of chrominance data.

The term "image block" hereafter includes any set of image data to which a transform is possibly applied. Though it is usual to use square- or rectangular-shaped blocks, the invention is in no way limited to these shapes and indeed any shape may be used.

Furthermore, an image block may be a macroblock to which a DCT (Discrete Cosinus Transform) or any other transform such as wavelet decomposition is applied. In the case of the HEVC standard, an image block, in the sense of the invention, corresponds to a Transform Unit (TU). In fact, as illustrated in Fig. 1 , the HEVC standard defines a recursive partitioning of an image into prediction units (PU) and each PU can in turn be decomposed into several transform units TU. Each unit TU is then encoded by applying a suitable transform. A unit PU can thus be encoded using various transforms depending on the partitioning of this PU into TUs. In this case, each TU is an image block in the sense of the invention. According to the example in Fig. 1 , a Coding Unit CU is subdivided into 4 CUs and one of these CUs is subdivided into 4 PUs. In this example, a PU can then be encoded either by applying a single TU transform to this PU, for example an 8x8 DCT, or by applying a TU transform to each half of this PU, for example by applying a TU transform twice to each half of this PU, for example by applying a 4x8 or 8x4 DCT twice, or by applying a TU transform to each quarter of this PU, for example by applying a TU transform four times to each quarter of this PU, for example by applying a 4x4 DCT four times.

Since the floating values representing the image data of the image block to be coded are defined in a space of at least one colour component, the steps for determining and coding the exponent and the coding of the mantissas of the floating values representing the image data of the image block to be coded are executed by the colour component of this space.

Hereafter, the term "motion data" is to be understood in the widest sense. It comprises the motion vectors and possibly the reference image indexes enabling a reference image to be identified in the reconstructed image sequence. It can also comprise an item of information indicating the type of interpolation that must be applied to a reference image block to derive a prediction block.

The term "prediction block" includes any set of image data used to predict an image block. A prediction block is obtained from an image block of the same image as the image to which the image block that it predicts belongs (spatial prediction or intra-image prediction) or from one or several image blocks belonging to one (mono-directional prediction) or several (bi-directional prediction) different images (temporal prediction or inter-image prediction) of the image to which the image block that it predicts belongs. The term "residual error block" includes any set of image data obtained after extraction of other data. This term is synonymous with the term "residues".

The term "prediction residual error block" therefore includes any set of image data obtained by subtracting a prediction block from an image block that it predicts.

A transform may be applied to a residual error block. A DCT (Discrete Cosine Transform) is an example of such a transform described in chapter 3.4.2.2 of the book by I. E. Richardson entitled "H.264 and MPEG-4 video compression" published by J. Wiley & Sons in September 2003. The wavelet transform described in chapter 3.4.2.3 of the book by I. E. Richardson and the Hadamard transform are other examples. Such transforms "transform" data of a residual error block, for example luminance and/or chrominance residual data, into a "block of transformed data" also called "block of frequency data" or "block of coefficients". The block of coefficients generally comprises a low- frequency coefficient known under the name of direct current coefficient or DC coefficient and high-frequency coefficients known under the name of AC coefficients.

According to the invention, illustrated in Fig. 2, the method for coding an image block B_c comprises the following steps:

- determining (10) a prediction block B_pred for the image block to be coded;

- determining (20) a prediction residual error block B_res by extracting said prediction block from the image block to be coded;

- determining (30) a unique exponent value Exp_res for the floating values representing the image data of the prediction residual error block B_res,

- coding (40) said exponent value thus determined Exp_res for the prediction residual error block B_res, and

- coding (50) the mantissas M_res of the floating values representing the image data of the prediction residual error block B_res by spatial and/or temporal prediction.

At the end of the method, a coded data signal F, illustrated in Fig. 2, is obtained. This stream comprises an item of coded data Ec describing a unique exponent value relative to the floating values which represent the image data of this prediction residual error block and an item of coded data Ec describing the mantissas of the floating values of the image data of this prediction residual error block. This item of coded data Ec is composed of several items of information, each describing the mantissa of a floating value of an item of data of this block.

According to a preferred embodiment of step 30, the exponent value Exp_res represents the maximum value of pixel exponents of the prediction residual error block B_res.

In mathematical terms, the unique exponent value Exp_res is given by the following equation:

Exp _ns ^{= ar}gmax{^exP_res ' )} where exp_res(i, j) is the exponent value for a pixel of indices i and j of the prediction residual error block B_res.

The invention is not limited to a particular determination of this exponent but extends to any possible approach such as the exponent mean or median value, to give just a few of examples.

According to an embodiment of step 40, the exponent value Exp_res is coded without loss.

For example, the exponent value is coded using Fixed Length Coding

(FLC).

This embodiment is particularly advantageous due to its simplicity of implementation into an image sequence coding scheme, since it is sufficient to transmit the Fixed Length Code to the decoder so that this decoder can find the exponent value Exp_res associated with the prediction residual error block

B_res.

As a variant, the exponent value Exp_res is coded by Variable Length Coding (VLC), using a Huffman-type coder or any other entropic coder well known to those skilled in the art, including, notably, those used in standards such as H264/AVC.

According to an embodiment of step 40, shown in Fig. 3, the exponent value Exp_res is predicted (step 41 ) by a prediction value Exp_pred, a so-called difference value denoted Exp_diff is then calculated (step 42) by extracting the prediction value Exp_pred from the exponent value Exp_res, and this difference ExV_diff ^is coded (step 43) giving rise to an item of coded data Ec added to the coded data signal F.

According to an embodiment of step 43, the difference Exp_diff is coded without loss.

For example, this difference is coded using Fixed Length Coding (FLC).

Preferably, this difference Exp_diff is coded via the intermediary of a Variable Length Coder (VLC), a Huffman-type coder or any other entropic coder well known to those skilled in the art, including, notably, those used in standards such as H264/AVC.

It is particularly advantageous to combine a prediction with a variable length coder to code the calculated difference between exponent values Exp_res and Exp_pred because the length of the code used, which corresponds to a coding cost denoted Exp£^c _ff , depends on the dynamic range of this difference, which is statistically always lower than the dynamic range of the unique exponent value Exp_res. Moreover, coding will be all the more efficient since the predictor is efficient.

According to an embodiment of step 41 , the prediction value Exp_pred is defined from a block that belongs to an image, called a reference image, other than the image to which image block B_c belongs.

According to a variant of this embodiment, this block is co-located with the image block B_c, that is to say located in the same spatial position.

According to a variant of this embodiment, the reference image block is designated by a motion vector which, in a coding/decoding scheme, is transmitted to the decoder.

According to an embodiment of step 41 , the prediction value Exp_pred is defined from a causal neighbourhood V_res formed of at least one block B?_es previously processed by the method. Each block B?_es corresponds to a prediction residual error block of an image block and is thus associated with a unique exponent value Exp^_x determined, for example, according to step 30. A neighbourhood is qualified as causal when it is formed of available data blocks. Thus, in a coding/decoding scheme, a neighbourhood is causal if the decoder is able to form this neighbourhood from data that it has previously decoded. In this case, the decoder is capable of obtaining the block(s) B?_es as well as its/their associated values which are required to form the neighbourhood V_res . The use of a neighbourhood is particularly advantageous in a context of transmitting the signal F between a coder and a decoder, because the decoder is then able to obtain a prediction value of the exponent value without any data being transmitted by the coder, as soon as the coder and decoder have the same rule for obtaining this prediction value.

According to a variant of step 41 , the causal neighbourhood V_res is formed of at least one block belonging to the image to which the block to be processed belongs.

Fig. 4 illustrates an example of a causal neighbourhood. In this example, the neighbourhood is formed by three blocks denoted B _es, 5r_es, 5_r ³ _es with which the exponent values Exp _es, Exp?_es, Exp?_es are associated respectively.

According to an embodiment of step 41 , a causal neighbouring area V_res is formed of at least one block belonging to at least one other image different to the image to which the image block to be coded belongs. This other image may be a reference image used for the estimation of motion data in a coding scheme or an image pointed to by a motion vector.

According to a variant of this embodiment, at least one of the blocks which belongs to the reference image is co-located with the image block Be, that is to say located in the same spatial position.

According to another variant of this embodiment, at least one of the blocks which belongs to the reference image is designated by a motion vector which, in a coding/decoding scheme, is transmitted to the decoder.

According to an embodiment of step 41 , the prediction value Exp_pred is defined from the prediction residual error.

According to a variant of this embodiment, the prediction value Exp_pred is defined from a median-type predictor defined by the following equation:

Exp_Pred = median(Exp?_es) pour n E {1, ... , N] where N is a prediction residual error block number.

According to a variant of this embodiment, the prediction value Exp_pred is defined from a mean-type predictor defined by the following equation:

N n=l

According to a variant of this embodiment, the prediction value Exp_pred is defined from a maximum-type predictor defined by the following equation:

ExPpred = ^ar§ i^EXP

According to an embodiment of step 41 , the prediction value Exp_pred is defined from a set of reconstructed data. Thus, in a coding/decoding scheme, these data are reconstructed by a decoder and form what is commonly called a causal zone.

Exp_pred = median (^ Expⁿ ^ pou r n E {1, ... , N] where Expⁿ = median xp" (i, j)} i e {l ,..., L}ei j e {Ι,. ,., Μ } and where L and M are dimensions of a block of index n and exp" (i, j) is the exponent value for a pixel (i,j) belonging to a block n of the neighbourhood used to predict the exponent. This block n was previously reconstructed.

Exppred = arg maX ^P") V ur fi £ {l, ...,JV} where Expⁿ = arg _max {exp"( , j)} i e {l,..., h}et j≡ {l,..., } and where L and M are dimensions of a block of index n and exp" (i, j) is the exponent value for a pixel (i,j) belonging to a block n of the neighbourhood used to predict the exponent. This block n was previously reconstructed.

According to a variant of this embodiment, the prediction value Exp_pred is defined from a mean-type predictor defined by the following equation: N n=l

1 ^{L M}

where £φ" = ∑ ∑ exp and where L and M are dimensions of a block of index n and exp" (i, j) is the exponent value for a pixel (i,j) belonging to a block n of the neighbourhood used to predict the exponent. This block n was previously reconstructed.

According to a variant, this set of reconstructed data may be formed of pixels used to construct a prediction block. The invention is not limited to a particular set of reconstructed data.

It may be noted that these various embodiments and variants of step 41 , used to define the prediction value Exp_pred , can be combined with each other to obtain a spatial prediction or a temporal prediction or a spatio- temporal prediction.

The invention is not limited to a particular type of method for coding the mantissas M_res of the floating values representing the image data of the prediction residual error block B_res, nor is it limited to any known coding method using, for example, a transform such as a DCT, DST (Discrete Sinus Transform) or wavelet, nor those using DPCM-type coding (Differential pulse- code modulation) (" Differential pulse-code modulation from ΗΗ available at: httpi /qps-tscupces/GTAV/Torres/Teaching/IVC-Notes/dpcm enqlish.pdf) nor that used in the H264 standard ("Advanced video coding for generic audiovisual services" available at: http://www.itu.int/rec/T-REC-H.264-201304- I) known as I PCM (Intra Pulse Code Modulation), to give just a few of examples. It is also well known that these various coders use quantisers and entropic coders.

Fig. 5 represents a method for decoding a coded data signal F representative of an image block whose data are represented by floating values.

The method comprises the following steps: - decoding (51 ) from the coded data signal F an exponent value relative to a prediction residual error block relative to the image block;

- decoding (52) from the coded data signal F the mantissas of the floating values representing the data of the prediction residual error block by spatial and/or temporal prediction;

- reconstructing (53) the floating values representing the image data of the prediction residual error block from said exponent value and said mantissas thus decoded;

- determining (54) a prediction block for the image block; and

- reconstructing (55) the floating values representing the image data of the image block from the prediction residual error block and the prediction block.

The floating values representing the data of the prediction residual error blocks to be decoded are obtained, for example, using equation (1 ).

According to an embodiment of step 51 , the exponent value of the prediction residual error block is obtained by decoding an item of coded data Ec from the coded data signal F. The Ec data are decoded according to either a fixed- or variable-length code.

According to an embodiment of step 51 , the Ec data represent the exponent value of the prediction residual error block.

According to another embodiment of step 51 , the Ec data represent a value called the difference value which corresponds to the difference Exp_diff between the exponent value Exp_res of the image block which was coded and a prediction value Exp_pred of this exponent value. The prediction value Exp_pred is then obtained from a causal neighbourhood in a manner identical to step 41 described above. The exponent value Exp_res is then decoded (reconstructed) by summing this prediction value Exp_pred with the decoded item of Ec data

(difference Exp_diff).

The mantissas of floating values representing the image data of the prediction residual error block are a set of integer values which is coded, according to the invention, by a coding method which uses, for example, a transform. A decoding method which is the inverse of this coding method is then used to decode the coded data of signal F in order to find the mantissas of these floating values. The invention is not limited to a particular method for decoding integral data.

The previously described methods for coding and decoding an image block can be used to code and decode an image which may belong to a sequence of images. In this case, each image is divided into image blocks, each image block is coded according to one of the methods described in relation to Fig. 2 to 4 and the coded data signal comprises as many items of data (Ec, Mc) as it does image blocks to be coded. The coded data signal F can then be decoded according to one of the methods described in relation to Fig. 5, with the aim of decoding (reconstructing) an image and possibly a sequence of images.

According to a variant of the invention, the method for coding an image block B_c also comprises a step (step 60) of selecting a coding mode for the residual error block from a set of coding modes.

The set of coding modes may include several methods to determine a unique exponent value for the floating values representing the image data of the prediction residual error block (step 30) and/or several predictors of this unique exponent value (step 41 ) and/or several coders to code the difference between this unique exponent value and a prediction value (step 43) and/or several coders to code the mantissas of the floating values representing the image data of the prediction residual error block (step 50). A coding mode in the set may also be a combination of one of these specific coders and predictors.

During step 60, a prediction residual error block is coded according to one of the coding modes i of the set and a coding cost Csti is then determined.

According to an embodiment relative to the case where the exponent value is predicted, the coding cost Csti takes into account the cost of coding MB^sj the mantissas of the floating values representing the image data of the prediction residual error block, the cost of coding Exp_d ^c _{ff i} the difference ExV_diff calculated by extracting the prediction value Exp_pred from the exponent value Exp_res and the cost of coding Hdr_{cst i} the syntax elements which enable the decoder to decode the image block B_c, such as an index to designate a specific prediction mode or a motion vector.

In mathematical terms, the coding cost Cst _j relative to a coding mode i of the image block B_c is given by:

where sse[^ec is a metric quantising a reconstruction error for image block Be and Bf^st = MB_r ^c _{s i} + Exp_d ^c _{ff i} + Hdr_{cst i}, with value A being adjusted to have the best compromise between bitrate and distortion. The adjustment of parameter A, called Lagrangian, is widely known. In principle, the best choice of coding mode, of index (i_opt), is made by selecting the mode which leads to the best compromise between a minimum coding cost for the exponent value and the mantissas of floating values representing the image data of the prediction residual error image, and a minimum reconstruction error of the image block.

In mathematical terms, the optimal coding cost is defined by:

(i_opt) = argmin (Csti)

(0

According to an embodiment, the metric sse_{rec i} is a square error calculated between the image block B_c and the coded/decoded block according to a coding mode (i).

In Fig. 2, 3 and 5, the modules shown are functional units that may or may not correspond to physically distinguishable units. For example, these modules or some of them can be grouped together in a single component or circuit, or constitute functions of the same software. On the contrary, some modules may be composed of separate physical entities. The prediction and/or coding devices compatible with the invention can be implemented according to a purely hardware embodiment, for example in the form of a dedicated component (for example in an ASIC (Application Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array) or VLSI (Very Large Scale Integration)) or of several electronic components integrated into a device or even in the form of a mixture of hardware elements and software elements.

Fig. 6 describes an example of the internal architecture of a device configured to implement at least one of the methods for coding an image block which are described in relation to Fig. 2, 3 and 4, and at least one of the methods for decoding a coded data signal which are described in relation to Fig. 5.

Device 600 comprises the following elements, interconnected by a digital address and data bus 601 :

- A calculation unit 603 (also called a central processing unit);

- A memory 605;

- A network interface 604, for interconnections between device 600 and other remote devices connected via a connection 602;

The calculation unit 603 can be implemented by a (possibly dedicated) microprocessor, a (possibly also dedicated) microcontroller, etc. The memory 605 can be implemented in a volatile and/or non-volatile form such as a RAM (random access memory), a hard disc, an EPROM (erasable programmable ROM), etc.

According to an embodiment, means 603, 604 and possibly 605 cooperate with each other in order to determine a unique exponent value for the floating values representing the image data of a residual error block which is from an image block to be coded, in order to code said exponent value thus determined and in order to code the mantissas of the floating values representing the image data of the residual error block by spatial and/or temporal prediction.

According to an embodiment, means 603, 604 and possibly 605 then cooperate with each other in order to decode an exponent value from the coded data signal for a residual error block relative to an image block to be decoded, in order to decode the mantissas of the floating values representing the image data of the residual error block from the coded data signal by spatial and/or temporal prediction and in order to reconstruct the floating values representing the image data of the residual error block to be decoded from said exponent value and said mantissas thus decoded.

The two embodiments above may also be combined so that the device 600 is designed both to code and decode a residual error block relative to an image block to be coded.

The devices described above are also designed to code and decode an image and/or a sequence of images. Information regarding image data, image sequence data and/or image block data are possibly received via the interface 604 and connection 602 or these data can be obtained from the memory 605.

The data signal F may be saved in the memory 605 and/or transmitted to an item of remote equipment via the interface 604 and connection 602.

Naturally, the invention is not limited to the embodiment examples mentioned above. In particular, those skilled in the art may apply any variant to the stated embodiments and combine them to benefit from their various advantages. The invention is in no way limited by the mode of coding the mantissas. Notably, the invention can be used with a DCT, a DST, a Hadarmard transform or a wavelet transform.

Claims

1 . Method for coding, to a coded data signal, an image block whose image data are represented by floating values, each item of floating data being expressed by a mantissa and an exponent, characterised in that it comprises the following steps:

- determining (30) a unique exponent value {Exp_res) for the floating values representing the image data of a prediction residual error block (B_res) determined by extracting a prediction block from the image block; - coding (40) said exponent value thus determined {Exp_res) for the prediction residual error block (B_res), and

- coding (50) the mantissas (M_res) of the floating values representing the image data of the prediction residual error block (B_res) by spatial and/or temporal prediction.

2. Method according to claim 1 , wherein said exponent value thus determined is coded without loss.

3. Method according to claim 1 or 2, wherein said exponent value is predicted by a prediction value and the difference between the prediction value and the exponent value is coded.

4. Method according to claim 3, wherein the prediction value is defined from a causal neighbourhood formed of at least one image block previously processed by said method.

5. Method according to claim 4, wherein the causal neighbourhood is formed of a block belonging to the image to which the block to be coded belongs and/or of at least one block belonging to at least one other image different to the image to which the image block to be coded belongs.

6. Method according to one of the previous claims, which comprises a step of selecting a coding mode for the image block to be coded from a set of coding modes during which the coding mode selected is that which minimises a compromise between a minimum reconstruction error of this image block which is coded and then decoded according to this coding mode and a minimum coding cost taking into account the coding cost according to this mode of coding the unique value of the exponent and the cost of coding the mantissas of the floating values representing the image data of the prediction residual error block.

7. Method for decoding a coded data signal representative of an image block whose data are represented by floating values, each floating value being expressed by a mantissa and an exponent, characterised in that it comprises the following steps:

- decoding (51 ) from the coded data signal an exponent value relative to a prediction residual error block relative to the image block;

- determining (54) a prediction block for the image block; and

8. Coded data signal representative of an image block whose data are represented by floating values, each item of floating data being expressed by a mantissa and an exponent, characterised in that it comprises at least one item of coded data describing a single exponent value for floating values and at least one other item of coded data describing the mantissas of floating values representing the image data of a prediction error block of the image block.

9. Device for coding, to a coded data signal, an image block whose image data are represented by floating values, each floating value being expressed by a mantissa and an exponent, characterised in that it comprises a processor configured to:

- determine (30) a unique exponent value {Exp_res) for the floating values representing the image data of a prediction residual error block (B_res) determined by extracting a prediction block from the image block;

- code (40) said exponent value thus determined {Exp_res) for the prediction residual error block (B_res), and

- code (50) the mantissas (M_res) of the floating values representing the image data of the prediction residual error block (B_res) by spatial and/or temporal prediction.

10. Device for decoding a coded data signal representative of an image block whose data are represented by floating values, each floating value being expressed by a mantissa and an exponent, characterised in that it comprises a processor configured to:

- decode (51 ) from the coded data signal an exponent value relative to a prediction residual error block relative to the image block;

- decode (52) from the coded data signal F the mantissas of the floating values representing the data of the prediction residual error block by spatial and/or temporal prediction;

- reconstruct (53) the floating values representing the image data of the prediction residual error block from said exponent value and said mantissas thus decoded;

- determine (54) a prediction block for the image block; and

- reconstruct (55) the floating values representing the image data of the image block from the prediction residual error block and the prediction block.