WO2016037988A1

WO2016037988A1 - Method and device for restoring a digital image comprising compression artefacts

Info

Publication number: WO2016037988A1
Application number: PCT/EP2015/070408
Authority: WO
Inventors: Carole THIEBAUT; Roberto CAMARERO; Christophe Latry
Original assignee: Centre National D'etudes Spatiales
Priority date: 2014-09-09
Filing date: 2015-09-07
Publication date: 2016-03-17
Also published as: FR3025640A1; FR3025640B1

Abstract

The invention relates to a method for restoring a digital image comprising compression artefacts, such a digital image being represented by at least one two-dimensional matrix of image samples in a spatial domain. Said method comprises the following steps, applied to a digital input image: applying (40) a transformation to said digital input image in order to obtain a transformed digital image, represented in a domain transformed by a plurality of blocks of coefficients, each block of coefficients corresponding to a block of image samples, and, for at least one processed coefficient of a block of coefficients: calculating (46) a noise threshold in accordance with a value representing a model of image acquisition noise associated with said processed coefficient or with said block of coefficients; comparing (50) the absolute value of said processed coefficient with the noise threshold, and, when said absolute value of the processed coefficient is lower than said noise threshold, replacing (52) said processed coefficient with a noise value that is dependent on said value representing a model of image acquisition noise.

Description

Method and apparatus for digital image restoration with compression artifacts

The present invention relates to a digital image restoration method comprising compression artifacts, wherein said digital image is represented by at least one three-dimensional array of image samples. It also relates to an associated restoration device.

The invention lies in the field of improving the quality of digital images including artifacts, in particular artifacts due to compression.

Image compression is widely used, whether it is still images or digital video.

The invention deals more particularly with the case of still images, acquired by any type of acquisition device.

The digital images are generally represented by arrays of image samples or pixels, each image sample having a given dynamic intensity.

In general, the compression methods include the application of three treatments to a source digital image, which are: decorrelation, quantization, coding. In several digital image compression standards, the decorrelation is obtained by applying a transformation, applied to the pixels to obtain transformed coefficients.

The JPEG ("Joint Photography Expert Group") compression standard uses a discrete cosine transform (DCT) applied to 8x8-sized image blocks, followed by block quantization and entropy encoding.

The JPEG2000 compression standard was developed after JPEG. This standard uses discrete wavelet transform (DWT) of the source digital image. Quantization and coding are performed in a single step, thanks to quality layer coding, obtained by bitmap coding of the transformed coefficients.

The various compression standards introduce, at a high compression rate, visual artefacts, also called compression artifacts, which are structured, being related to the basic functions of the transformation applied for the decorrelation. Such compression artifacts are, for example, block effects for JPEG and solid colors and butterfly patterns for compression methods using wavelet transformation.

In general, compression artifacts are visually annoying to users. The subject of the invention is a digital image restoration method comprising compression artifacts, making it possible to improve the subjective quality of the digital image.

To this end, the invention proposes, in one aspect, a digital image restoration method comprising compression artifacts, a said digital image being represented by at least one two-dimensional array of image samples in a spatial domain.

This method comprises the following steps, applied to an input digital image:

applying a transformation to said input digital image to obtain a transformed digital image, represented in a transformed domain by a plurality of coefficient blocks, each coefficient block corresponding to a block of image samples, and for at least one processed coefficient of a block of coefficients:

calculating a noise threshold as a function of a value representative of an image acquisition noise model associated with said processed coefficient or said coefficient block, comparing the absolute value of the coefficient treated with the noise threshold and,

• when the absolute value of the treated coefficient is greater than or equal to said noise threshold, said coefficient is left unchanged,

When the absolute value of the processed coefficient is lower than said noise threshold, replacement of the processed coefficient value by a noise value dependent on said representative value of an image acquisition noise model.

Advantageously, the digital image restoration method comprising compression artifacts according to the invention makes it possible to restore a granular noise similar to image acquisition noise, and thus to re-introduce a granularity similar to the natural granularity of the images. , which improves the subjective quality of the images and hides the compression artifacts.

The digital image restoration method with compression artifacts according to the invention may have one or more of the features below.

The representative value of an image acquisition noise model is a variance value of the acquisition noise, and the noise value is a function of the standard deviation of the acquisition noise.

The noise threshold is calculated for a block of coefficients and depends on an image acquisition model associated with said coefficient block, the comparison being applied to all the coefficients of the coefficient block. Each coefficient block comprises a low frequency coefficient and a plurality of frequency coefficients, and the representative value of the noise model is the variance value of the acquisition noise in the calculated transformed domain, for a coefficient block, by a parameterized model. according to the value of the low frequency coefficient of said block of coefficients.

When the transformation applied to the input digital image is a wavelet transformation on several decomposition levels, for each decomposition level, the representative value of the acquisition noise pattern is the variance value of the acquisition noise in the calculated transformed domain, for each frequency coefficient processed, as a function of the value of a low frequency coefficient of the same level of decomposition and spatially corresponding to said frequency coefficient processed.

The noise threshold is equal to the standard deviation of the acquisition noise weighted by a predetermined noise proportionality factor.

The digital input image is obtained by compression of a source digital image acquired by an acquisition device, and the acquisition noise model is determined according to said acquisition device.

The input digital image is obtained by compressing a source digital image according to a compression method including the application of a compression transformation, and at the transforming step, said compression transformation is applied.

In the step of replacing said processed coefficient, said noise value is calculated from a value obtained by drawing according to a predetermined distribution law multiplied by a reconstruction proportionality factor.

The predetermined distribution law is a Gaussian law, of zero mean and of variance equal to unity.

Said treated coefficient belongs to a frequency subband, and in the step of replacing said processed coefficient, said noise value is weighted by a deconvolution weighting coefficient dependent on said frequency subband.

Said deconvolution weighting coefficient is calculated as a function of said transformation, and a predetermined deconvolution filter.

According to another aspect, the invention relates to a digital image restoration device comprising compression artifacts, a said digital image being represented by at least one two-dimensional matrix of image samples in a spatial domain, characterized in that it includes units applied to an input digital image, adapted to: applying a transformation to said input digital image to obtain a transformed digital image, represented in a transformed domain by a plurality of coefficient blocks, each coefficient block corresponding to a block of image samples,

for at least one processed coefficient of a block of coefficients:

calculating a noise threshold as a function of a value representative of an image acquisition noise model associated with said processed coefficient or said coefficient block,

comparing the absolute value of the coefficient treated with the noise threshold, and, when the absolute value of a treated coefficient is lower than said noise threshold, replacing said processed coefficient with a noise value dependent on said representative value of a noise model. image acquisition noise.

The digital image restoration device is adapted to implement the steps of the digital image restoration method as briefly introduced above.

According to a particular characteristic, the image restoration device is able to calculate a noise threshold for a block of coefficients according to the image acquisition noise model associated with said coefficient block, and to compare the absolute value of each coefficient of the coefficient block at the calculated noise threshold.

According to one particular characteristic, the digital image restoration device is able to process blocks of coefficients comprising a low frequency coefficient and a plurality of frequency coefficients, the value representative of the acquisition noise model being the noise variance value. method of acquisition in the transformed domain calculated for a block of coefficients by a model parameterized as a function of the value of the low frequency coefficient of said block of coefficients.

According to another aspect, the invention relates to a computer program comprising instructions for implementing the steps of a digital image restoration method including compression artifacts as briefly described above during execution. of the program by a processor of a programmable device.

According to another aspect, the invention relates to an information recording medium, characterized in that it comprises instructions for the execution of a digital image restoration method including compression artifacts as presented herein. above, when these instructions are executed by a programmable device. Other features and advantages of the invention will emerge from the description given below, by way of indication and in no way limiting, with reference to the appended figures, among which:

FIG. 1 is a block diagram of an image restoration device implementing a restoration method according to the invention;

FIG. 2 diagrammatically represents a set of sub-bands obtained by applying a discrete wavelet transformation to a digital image;

FIG. 3 schematically represents a block of transformed coefficients obtained by a discrete wavelet transformation;

- Figure 4 is a diagram showing the functional blocks of a programmable device adapted to implement the invention;

FIG. 5 is a block diagram of an image restoration method according to a first embodiment of the invention;

FIG. 6 is a block diagram of an image restoration method according to a second embodiment of the invention;

FIG. 7 describes an embodiment of the calculation of the deconvolution weighting coefficients.

Figure 1 illustrates a device 10 for image restoration.

The device 10 receives as input an input digital input image, composed of one or more image sample matrices or pixels, each image sample having an associated intensity value. This domain of representation is called spatial domain.

Subsequently, it is considered to simplify the explanation that a digital image to be processed is composed of a matrix of pixels of MxN dimension, each pixel having an intensity initially encoded on P bits. For example, P = 12.

The input digital image input is derived from a source digital image, acquired by an acquisition device, having been compressed and having compression artifacts.

The input digital image is processed by a transformation unit 12, able to apply a transformation to the input image, to output a plurality of transformed coefficient blocks of predetermined size, each block of transformed coefficients corresponding to a corresponding block of pixels of the input image, of the same size. This representation of the representation input image is called in the transformed domain.

For example, blocks of size 8x8 are formed. The transformation applied by the transformation unit 12 is preferably the compression transformation applied by the compression algorithm to the source digital image, when this compression transformation is known.

For example, if the source digital image was compressed by applying the JPEG2000 standard, a wavelet transform (or DWT) with a filter and a given number of decomposition levels was applied. In this case, the same wavelet transformation with the same filter and the same number of decomposition levels is applied. In addition, a spatial coherence restoration of the transformed coefficients is performed, as explained below, with reference to FIGS. 2 and 3.

If the source digital image has been compressed by applying the JPEG standard, an 8x8 block decomposition and a discrete cosine transform (DCT) is applied for each block.

If the compression algorithm applied beforehand is not known, the choice is given to the user as to the transformation to be applied by the transformation unit 12.

Following transformation, a representation of the image blocks B _k of transform coefficients _C, including a low-frequency coefficient C _0,0, also called DC coefficient, and frequency coefficients G _j, is obtained.

Then, a replacement unit of the transformed coefficients 14 implements the calculation of a noise threshold according to the local noise estimated via a model, as a function of the low frequency coefficient C _0.0 and model parameters stored in a memory 16 .

Transformed coefficients whose absolute value is below the noise threshold are replaced by a noise value, calculated by the replacement unit of the transformed coefficients 14.

An inverse transformation unit 18 applies an inverse transformation, making it possible to obtain a restored digital image, denoted as output, whose visual quality is improved with respect to the input digital image input.

The inverse transformation unit 18 applies the transformation inverse to that applied by the transformation unit 12.

According to one embodiment, a discrete wavelet transformation (DWT) is applied by the transformation unit 12, on 3 levels of decomposition, as schematically illustrated in FIG. 2.

In a known manner, the application of a DWT is first carried out on the complete image matrix, making it possible to obtain four sub-bands, each subband including transformed coefficients corresponding to the application of a pass filtering low L and a high-pass filtering H, first in a first direction of the matrix, then in a second direction, perpendicular to the first direction.

The application of a first level of decomposition to the matrix I makes it possible to obtain the sub-bands of coefficients LL _1; LH ^ HL-ι and Hh ^. The sub-band LL-ι comprises a low-pass version of the initial image I at a lower resolution level than the resolution level of the initial image I. If the initial image is composed of M rows and N columns, each LL ₁ sub-band _; The first level of decomposition has M / 2 lines and N / 2 columns.

The application of the respective filters L and H is iterated on the sub-band LL _1; to obtain the sub-bands LL ₂ , LH ₂ , HL ₂ and HH ₂ , second decomposition level and size M / 4 rows and N / 4 columns.

The LL ₂ sub-band is filtered again to obtain the third decomposition sub-bands LL ₃ , LH ₃ , HL ₃ and HH ₃ .

Thus, after applying a wavelet transformation on three decomposition levels, as illustrated in FIG. 2, the digital image is represented by a set of transformed coefficient subbands. At a decomposition level p, the size of the subbands is M 12 ^P x N 12 ^P.

The transformed coefficients of the subbands are grouped in blocks of spatially located coefficients of size 2 ^P x 2 ^p for p decomposition levels, thus blocks of size 8x8 for a decomposition with three levels of decomposition.

Thus, as illustrated in FIGS. 2 and 3, a block B _k of transformed coefficients is obtained by spatially grouping coefficients of the DWT transformation corresponding to the same spatial position of the input digital image input. This operation is also called spatial coherence.

Each block of coefficients B _k comprises a low-frequency coefficient DC, derived from the sub-band LL ₃ in the example, and several AC frequency coefficients derived from the sub-bands LH ₃ , HL ₃ , HH ₃ , LH ₂ , HL ₂ , HH ₂ , LH _1; HL _1; Hh ^ as illustrated.

Alternatively, a plurality of low frequency coefficients can be considered by decomposition levels: a low frequency coefficient DC, derived from the sub-band LL ₃ for the third decomposition level, a low frequency coefficient matrix LL ₂ for the second level of decomposition. decomposition, a matrix of low frequency coefficients for the first level of decomposition. In general, a matrix of low frequency coefficients per level of decomposition can be obtained, regardless of the number of levels of decomposition considered.

Alternatively, according to an alternative embodiment, the input image is divided into 8x8 pixel blocks and a discrete cosine transform (DCT) is applied on each block of pixels, making it possible to obtain blocks 8x8 of spatially localized transformed coefficients.

In general, any decorrelation transformation, making it possible to obtain blocks of coefficients of size mxn, is applicable by the transformation unit 12.

The digital image restoration method of the invention is implemented by an electronic device, for example a computer-type, workstation-type programmable device whose main functional blocks are illustrated in FIG. 4.

A programmable device 20 capable of implementing the method of the invention comprises a central processing unit 22, for example a processor (CPU), capable of executing preprogrammed operations or computer program instructions when the device 20 is put under control. voltage.

In one embodiment, a multiprocessor CPU is used to perform parallel computations. The device 20 also comprises information storage means 24, for example registers, capable of storing executable code instructions allowing the implementation of programs comprising code instructions able to implement the method according to the invention. .

The device 20 comprises control means 26 for updating parameters and receiving commands from an operator. When the programmable device 20 is an onboard device, the control means 26 comprise a telecommunication device for receiving remote commands and parameter values.

Alternatively and optionally, the control means 26 are means for inputting commands from an operator, for example a keyboard.

Optionally, the programmable device 20 comprises a screen 28 and additional pointing means 30, such as a mouse.

The various functional blocks of the device 20 described above are connected via a communication bus 32.

In a variant, the methods of the invention are implemented by electronic devices of the programmable logic circuit type, such as electronic cards based on FPGAs or ASICs.

FIG. 5 illustrates the main steps of a digital image restoration method according to the invention, according to a first embodiment.

These steps are implemented by processor modules 22 of a device 20 adapted to implement the invention or by an electronic card developed for the implementation of the invention. An input digital image, the input, which is the digital image to be processed, is inputted to the process.

In a first step 40, a predetermined transformation T is applied to the input digital image, as explained above, the transformation T being for example a DCT or a DWT.

The transformation step 40 comprises a spatial re-organization of the coefficients transformed into blocks of coefficients B _k , which is implemented when the applied transformation T is a DWT. The size mxn of the blocks B _k depends on the number of decomposition levels applied in the DWT transformation, as explained above.

When T is a block DCT, this reorganization step is omitted.

Step 40 is followed by a step 42 of replacing transformed coefficients, comprising steps 44 to 56.

A current block to be processed is extracted during a step 44 of current block extraction B _kc .

Next, an acquisition noise threshold for the current block is calculated at the step of calculating a noise threshold 46, as a function of parameters 48 which comprise: an acquisition noise model, a proportionality factor of noise threshold K _d , a reconstruction proportionality factor K _r

The acquisition noise model defines, in the preferred embodiment, the variance of the acquisition noise of the source digital image, acquired by an image acquisition device, before application of a compression algorithm.

In a preferred embodiment, the acquisition device of the source digital image is known, and an acquisition noise model is also known. For example, several acquisition devices are listed, and for each of these acquisition devices, a noise model is stored in a table.

According to a variant, a predetermined acquisition noise model is applied, without knowledge of the acquisition device of the source digital image. For example, a standard noise model can be applied in which the noise variance is an affine function of the local DC coefficient.

According to another variant, prior to the implementation of the digital image restoration method, an acquisition noise estimation method is applied from the digital input image or from a plurality of digital images acquired by the acquisition device.

For example, in the context of Earth observation images, a noise estimation method can be applied as described in the article "Geometric / radiometric calibration from ordinary images for high resolution satellite systems "by C. Latry, published in Proc. SPIE 8153, Earth Observing Systems XVI, 81531 K, September 201 1.

In the step of calculating a noise threshold 46, a value representative of the acquisition noise model, associated with the current block B _kc , is calculated.

In the preferred embodiment, a parameterized noise pattern is used.

The representative value of the acquisition noise model, denoted V, is a function of this acquisition noise model and of the value of the low frequency coefficient C _0.0 , representative of the average of the intensity of the pixels of the current block. B _kc , according to the following formula:

V {B _kc ) = a ² + P- C _Qfi (Eq 1)

The value V (B _kc ) is the variance of the acquisition noise associated with the block B _kc , and the parameters a and β are provided in the set of parameters 48. These parameters a and β are predetermined constants.

According to another variant, the noise threshold can be calculated independently for each coefficient in the image by applying the noise model defined by the equation (Eq 1) on each image pixel (input), the value C _0.0 of the equation (Eq 1) being replaced by the value of the pixel, in order to establish a noise image with the standard deviation values given by the model.

Then the same transform T is applied to this noise image in order to obtain a noise threshold (S _noiselj ) for each coefficient in the transformed domain.

For example, the constants a and β are representative of the acquisition sensor of the source digital image of the input image input, and are determined by calibration of the sensor and stored.

For the PLEIADES-1A satellite with a 13-stage TDI configuration, the a and β values in the panchromatic band are equal to 2.267 and 0.0393, respectively.

Alternatively, for a non-white detection noise, the noise calculation may take into account different models for each sub-band or transform coefficient.

The variance of the acquisition noise is obtained with the following formula:

V (B _kc ) = Σ <*, + # _. , · <\ (Eq 2)

i

(U) ≠ (0,0)

The noise threshold s _noise is obtained from the standard deviation - V (B _kc ) of the acquisition noise:

¾ _n * = K _d - V {B _kc ) (Eq 3) Where K _d is a constant noise proportionality factor provided by the user. K _d may be typically 0.5 or 1 and more generally between 0 and

3.

More generally, the noise threshold s _brU it is obtained by applying a function f (), which is not necessarily linear:

s _noise = f {K _d ^ V {B _kc )) (Eq 4)

Then, each coefficient _C, of the current block of transformed coefficients _kc B is compared to the noise s _noise threshold, in absolute value, when the comparison step 50.

If the absolute value of the coefficient C _ti is less than the noise s _noise threshold, then during a step of replacing the coefficient 52, the coefficient _C, is replaced by a local noise value based on the representative value of acquisition noise model associated with the current block. In general, one can write Cy ^' = giK ^ ViB ^)).

In this embodiment, the function g () is given by:

Cy ^' = sign (Cy) - K _r - ^ V (B _kc ) ^■ \ rand \ (Eq 5)

Or :

sign (Cy) is the sign of the coefficient _C,,,;

K _r is a reconstruction proportionality factor, which is typically 1 but may be more generally between 0 and 3,

rand is a value obtained by pseudo-random drawing according to a predetermined distribution law and of variance equal to 1. The absolute value of the rand value, noted \ rand \, is used in the equation formula (Eq 5).

For example, it is a Gaussian law of zero mean and variance equal to 1.

Alternatively, another distribution law, for example a uniform distribution or a Poisson distribution is used.

If the absolute value of the coefficient Qj is greater than or equal to the noise threshold s _brU it, the converted coefficient is unchanged (step 54), so the replacement coefficient C'ij is equal to the original coefficient Cij.

According to another variant, the value of the standard deviation to be used for calculating the replacement coefficient may vary for each coefficient. This value is obtained by applying the noise model on each image pixel (input) to establish a noise image with the standard deviation values given by the acquisition noise model. Then the same transform T is applied to this noise image in order to obtain the standard deviation value to be applied for each C'y. These standard deviation values replace the value ^ V {B _kc ) in the equation (Eq 5).

The block of replacement coefficients C '^ is stored, in the storage step 56, in the current block B _kc , instead of the coefficients Q _j .

If necessary, step 56 is followed by step 44 previously described, for the processing of another block of coefficients transformed as a current block.

Preferably, all the transformed coefficient blocks of the transformed digital image are processed.

When there are no longer any blocks of transformed coefficients to be processed, step 56 is followed by a step 58 for applying the inverse transformation T ¹ , which makes it possible to obtain the output digital image that is output, whose quality visual is improved.

When the transformation T applied in step 40 is a DWT transformation, the step 58 of applying the inverse transformation T ¹ comprises a reorganization of the coefficients transformed into subbands.

According to an alternative embodiment, when the transformation T applied in step 40 is a DWT transformation, the current block of coefficients B _kc is processed by decomposition level. For each frequency coefficient Cjj of one of the frequency sub-bands HL _P , LH _P or HH _P of the decomposition level p, the value C _0.0 in the equations (Eq 1) and (Eq 2) is the value of the coefficient low frequency of the sub-band LL _P , spatially corresponding to the frequency coefficient Cjj.

Thus, in this embodiment, a local low frequency coefficient value is used.

In this embodiment, the frequency coefficients are processed by descending decomposition level index. For example, when p = 3 decomposition levels are applied, as illustrated in FIG. 3, the noise threshold for the decomposition level frequency coefficients 3 belonging to the sub-bands LH ₃ is calculated during step 46. HL ₃ , HH ₃ as a function of the value of the coefficient LL ₃ .

The steps 50 to 54 described above are applied, and the replacement values of the decomposition level frequency coefficients 3 are stored.

An additional step of applying an inverse DWT transformation for a decomposition level is applied, to obtain, from memorized LL ₃ , LH ₃ , HL ₃ , HH ₃ coefficients, a low frequency subband of level 2, denoted LL ₂ , and comprising, in this example, a matrix of 2x2 coefficients. Noise thresholds are calculated (step 46) for the decomposition level frequency coefficients 2 belonging to the sub-bands LH ₂ , HL ₂ , HH ₂ as a function of the values of the spatially corresponding coefficients of the sub-band LL ₂ . Since all the sub-bands of the same level of decomposition consist of matrices of coefficients of the same size, the low frequency coefficient spatially corresponding to a frequency coefficient is the coefficient of the low frequency sub-band LL ₂ of the same position. in the matrix.

The steps 50 to 54 described above are applied, and the replacement values of the decomposition level frequency coefficients 2 are stored.

An additional step of applying an inverse DWT transformation for a decomposition level is applied, to obtain, from memorized LL ₂ , LH ₂ , HL ₂ , HH ₂ coefficients, a low frequency subband of level 1, denoted LL _1; and comprising, in this example, a matrix of 4x4 coefficients.

Noise thresholds for the decomposition level frequency coefficients 1 belonging to the sub-bands LH _{1 are} calculated (step 46) _; HL _1; HHi as a function of the values of the coefficients corresponding spatially of the subband. Since all the sub-bands of the same level of decomposition consist of matrices of coefficients of the same size, the low frequency coefficient spatially corresponding to a frequency coefficient is the coefficient of the low frequency sub-band L-ι of the same position in the matrix.

The steps 50 to 54 described above are applied, and the replacement values of the decomposition level frequency coefficients 1 are stored.

Obviously, this alternative embodiment is applicable regardless of the maximum number of decomposition levels.

Advantageously, this alternative embodiment makes it possible to perform a treatment adapted locally to the characteristics of the image and to refine the restoration obtained.

According to a second embodiment, the image restoration according to the invention anticipates a deconvolution of the output image 1 output, by application of a predetermined deconvolution filter known in advance, denoted F _deconv .

FIG. 6 illustrates a second embodiment of the invention, anticipating such a deconvolution of the image. These steps schematically illustrated in FIG. 6 are implemented by modules of the processor 22 of a device 20 adapted to implement the invention or by an electronic card developed for the implementation of the invention.

Steps 40 to 46 are similar to the steps described with reference to Figure 5 for the first embodiment, so they are not re-described here.

In this embodiment, the method further comprises a preliminary step 60 for calculating and storing deconvolution weighting coefficients, a function of the deconvolution filter F _dec0 nv. A deconvolution weighting coefficient is calculated per frequency subband.

When the applied transformation is a wavelet transformation, the frequency subbands are illustrated in Figure 2.

When the applied transformation is a DCT, each transformed coefficient differs from the low frequency coefficient C _0.0 corresponding to a frequency sub-band. Thus, 63 deconvolution weighting coefficients P are calculated and stored for the application of a DCT.

Step 50 for comparing a transformed coefficient _C, at the sound s _BRU threshold it is followed, when the absolute value of the coefficient is below the noise threshold, a step 62 of obtaining weighting coefficient deconvolution of P j for the coefficient _C,, ,, for example by reading in a memory space where the deconvolution weighting coefficients have been stored.

Step 62 is followed by a step 64 of replacing the coefficient Qj with a local noise value dependent on the value representative of the acquisition noise model associated with the current block and the deconvolution weighting coefficient Pi j. In general, the local noise value is calculated by a function h (): C _v = h (K _r , P _v , V (B _k ).

In one embodiment the function h () is given by:

C _y = sign (C _y ) ^■ P ^■ K _r ^■ V (B _kc ) ^■ \ rand \ (Eq 6)

Or :

sign (Cy) is the sign of the coefficient _C,,,;

rand is a value obtained by pseudo-random drawing according to a predetermined distribution law and of variance equal to 1. The absolute value of the rand value, noted \ rand \, is used in the equation formula (Eq 5). For example, it is a Gaussian law of zero mean and variance equal to 1.

When the absolute value of the transformed coefficient _C,, is greater than or equal to the noise threshold mit _b s, it remains unchanged, step 50 is followed by step 66, analogous to the step 54 previously described.

The transformed transformed coefficients C '^ are stored in step 68, analogous to step 56 previously described.

Step 68 is followed by the reverse transformation step 58 previously described.

One embodiment of calculating the deconvolution weighting coefficients is illustrated in FIG. 7.

A gaussian white noise image of zero mean and standard deviation, denoted l_noise, is input.

The image I noise must be of sufficient size to correctly estimate variances on the transformed coefficients. A typical size of 1024x1024 is sufficient for a DCT transform or wavelet transform on 3 levels.

The transformation T applied during step 40 is applied to the noise image l_noise during step 70. At the output, a transformed noise image, denoted l_bruit_transf, composed of a set of transformed coefficient blocks, is obtained.

For each different frequency sub-band denoted (i, j) (step 72), a standard deviation of the transformed coefficients of the image l_noise_transf is calculated in step 74.

This standard deviation is noted o _ht (i, j).

As explained above, if the transformation T is a block DCT, 63 frequency sub-bands corresponding to all other frequency coefficients than the low frequency coefficient are considered, and thus 63 standard deviation values are calculated.

If the transform T is a dyadic wavelet transform on n levels, 3 ^{n 1} standard deviation values are created, corresponding to the 3 ^{n 1} frequency subbands other than the low frequency subband LL _n .

Substantially in parallel or sequentially, a convolution is applied by the inverse of the deconvolution filter F _dec0 nv to the noise image I noise in step 76, then the transformation T is applied to this image convolved at step 78 , to obtain a transformed convolved image I conv transf. For each different frequency sub-band, noted (i, j) (step 80), a standard deviation of the transformed coefficients of the convective image Ic is calculated in step 82.

This standard deviation is noted a _ct (i, j).

Then, at step 84, for each frequency subband considered, the deconvolution weighting coefficient is calculated by the following formula:

These deconvolution weighting coefficients are stored for later use during the step 64 previously described.

Advantageously, the application of the deconvolution weighting coefficients makes it possible to obtain a white noise during the subsequent deconvolution of the image obtained by the deconvolution filter F _dec .

The method of the invention has been described above in particular for restoring digital images with compression artifacts due to JPEG or JPEG2000 compression standards.

However, it is clear that the method according to the invention generally applies with any other compression method, in particular with the standard CCSDS IDC ("Image Data Compression") defined by the CCSDS (Consultative Committee for Space Data Systems). used for the compression of digital images obtained by satellite.

Advantageously, the application of the digital image restoration method allows a zeroing by the compressor itself of the transformed coefficients whose amplitude is lower than the noise threshold by a posteriori correcting the artifacts, which allows to increase significantly the performance of the compression.

Claims

1. Digital image restoration method comprising compression artifacts, a said digital image being represented by at least one three-dimensional matrix of image samples in a spatial domain, characterized in that it comprises the following steps, applied to a digital input image:

- applying (40) a transformation to said input digital image to obtain a transformed digital image, represented in a transformed domain by a plurality of coefficient blocks, each coefficient block corresponding to a block of image samples ,

for at least one processed coefficient of a block of coefficients:

calculating (46) a noise threshold as a function of a value representative of an image acquisition noise model associated with said processed coefficient or said coefficient block,

-comparison (50) of the absolute value of said coefficient treated at the noise threshold, and,

When the absolute value of the processed coefficient is lower than said noise threshold, replacement (52, 64) of the coefficient value processed by a noise value dependent on said representative value of an image acquisition noise model.

2. The digital image restoration method as claimed in claim 1, characterized in that said value representative of an image acquisition noise model is a variance value of the acquisition noise, and that the noise value is a function of the standard deviation of the acquisition noise.

3. - digital image restoration method according to one of claims 1 or 2, characterized in that said noise threshold is calculated for the coefficient block and depends on the image acquisition noise model associated with said block of coefficients, and in that the comparison (50) is applied to all the coefficients of said coefficient block.

4. The method of restoring digital image according to claim 3, wherein each block of coefficients comprises a low frequency coefficient and a plurality of frequency coefficients, characterized in that the representative value of the Acquisition noise model is the variance value of the acquisition noise in the transformed domain calculated, for a block of coefficients, by a model parameterized as a function of the value of the low frequency coefficient of said block of coefficients.

5. A digital image restoration method according to one of claims 1 or 2, characterized in that, when the transformation applied to the input digital image is a wavelet transformation on several levels of decomposition, for each level of decomposition,

the representative value of the acquisition noise model is the variance value of the acquisition noise in the calculated transformed domain, for each frequency coefficient processed, as a function of the value of a low frequency coefficient of the same decomposition level and corresponding spatially to said frequency coefficient treated.

6. The digital image restoration method as claimed in claim 2, wherein said noise threshold is equal to the standard deviation of the noise of acquisition weighted by a noise proportionality factor. predetermined.

The method of digital image restoration according to any one of claims 1 to 6, wherein said digital input image is obtained by compression of a source digital image acquired by an acquisition device, characterized in that that said acquisition noise pattern is determined according to said acquisition device.

The digital image restoration method according to any one of claims 1 to 7, wherein said digital input image is obtained by compressing a source digital image according to a compression method comprising the application of a compression transformation, characterized in that, at the transforming step (40), said compression transformation is applied.

9. A digital image restoration method according to any one of claims 1 to 8, characterized in that, in the step of replacing (52, 64) said processed coefficient, said noise value is calculated from a value obtained by drawing according to a predetermined distribution law multiplied by a reconstruction proportionality factor (K _r ).

10. A method of digital image restoration according to claim 9, characterized in that the predetermined distribution law is a Gaussian law, of zero mean and variance equal to unity.

1 1 .- A digital image restoration method according to any one of claims 1 to 10, characterized in that said processed coefficient belongs to a frequency subband, and in that in the step of replacing said coefficient processed, said noise value is weighted by a deconvolution weighting coefficient (P) dependent on said frequency subband.

12. A digital image restoration method according to claim 1 1, characterized in that said deconvolution weighting coefficient is calculated according to said transformation, and a predetermined deconvolution filter.

13. A digital image restoration device comprising compression artifacts, a said digital image being represented by at least one two-dimensional matrix of image samples in a spatial domain, characterized in that it comprises units applied to a digital input image, adapted to:

applying a transformation to said input digital image to obtain a transformed digital image, represented in a transformed domain by a plurality of coefficient blocks, each coefficient block corresponding to a block of image samples,

for at least one processed coefficient of a block of coefficients:

-comparate the absolute value of the coefficient treated at the noise threshold, and,

• when the absolute value of the treated coefficient is greater than or equal to said noise threshold, leave said coefficient unchanged,

When the absolute value of a processed coefficient is lower than said noise threshold, replacing said processed coefficient with a noise value dependent on said representative value of an image acquisition noise model.

14. A digital image restoration device according to claim 13, characterized in that it is able to calculate a noise threshold for a block of coefficients according to the image acquisition noise model associated with said block of noise. coefficients, and compare the absolute value of each coefficient of the coefficient block with the calculated noise threshold.

15. - digital image restoration device according to claim 14, adapted to process blocks of coefficients comprising a low frequency coefficient and a plurality of frequency coefficients, characterized in that the representative value of the acquisition noise model is the variance value of the acquisition noise in the calculated transformed domain, for a block of coefficients, by a model parameterized as a function of the value of the low frequency coefficient of said coefficient block.

16. - digital image restoration device according to claim 13, characterized in that, when the transformation applied to the input digital image is a wavelet transformation on several levels of decomposition, for each level of decomposition, the representative value of the acquisition noise model is the variance value of the acquisition noise in the calculated transformed domain, for each frequency coefficient processed, as a function of the value of a low frequency coefficient of the same level of decomposition and corresponding spatially frequency coefficient processed.

17. Computer program comprising instructions for implementing the steps of a digital image restoration process including compression artifacts according to one of claims 1 to 12 during the execution of the program by a processor. a programmable device.