METHOD AND APPARATUS FOR RECONSTRUCTING DIGITAL IMAGES

Description

[0001] The present invention relates to a method for restoring a digital image from an initial image acquired by an image acquisition device having associated instrumental acquisition noise, then compressed by a predetermined compression method. It also relates to an associated restoration device.

[0002] The invention lies in the field of improving the quality of digital images.

[0003] In particular, the invention relates to improving the quality of digital images exhibiting acquisition and compression noise. These images may be monospectral, multispectral, or hyperspectral.

[0004] Image compression is widely used, whether for still images or digital videos.

[0005] The invention deals more particularly with the case of still images, acquired by any type of acquisition device.

[0006] Digital images are generally represented by matrices of image samples or pixels, each image sample having a radiometric value of given dynamic range. A color image (multi or hyperspectral) is then represented by as many matrices as there are color components, each sample having a radiometric value for each color called a spectral band.

[0007] Generally speaking, compression processes involve applying three processes to a source digital image: decorrelation, quantization, and coding. In several digital image compression standards, decorrelation is achieved by applying a transformation to the pixels to obtain transformed coefficients.

[0008] The Joint Photographic Expert Group (JPEG) compression standard uses a discrete cosine transform (DCT) applied to 8x8 image blocks, followed by block quantization and entropy coding.

[0009] The JPEG2000 compression standard was developed after JPEG. This standard uses a discrete wavelet transform (DWT) of the source digital image. Quantization and coding are performed in a single step, using quality layer coding, obtained by bit-plane coding of the transformed coefficients.

[0010] Various compression standards introduce, at high compression rates, visual artifacts, also called compression artifacts, which are structured, being related to the basic functions of the applied transformation. Such compression artifacts are, for example, blocky effects for JPEG and flat areas and butterfly patterns for compression processes using a wavelet transform.

[0011] In various applications, for example in satellite image acquisition, images are acquired by an acquisition device, compressed by a given compression method and stored or transmitted to a remote device for further processing.

[0012] In the case of satellite imagery, image acquisition and compression are carried out on board a satellite, and processing is carried out on the ground. In this case, the image is usually composed of a broadband image called panchromatic with high spatial resolution and n images acquired in much narrower spectral bands representing the multispectral color information, the acquisition being carried out with a sampling step k times larger than the sampling step of the panchromatic band, k generally being 4. An image fusion process makes it possible to mix the panchromatic image and the n multispectral images in order to obtain n color images at the spatial resolution of the panchromatic. The following description applies to both panchromatic and multispectral images.

[0013] Digital images, whether panchromatic or multispectral, contain compression artifacts, blur, and noise. Typically, images are decompressed, restored, and merged, with restoration typically involving deconvolution followed by denoising.

[0014] As is known, the blur introduced by an acquisition device can be modeled by the convolution of the observed landscape by a characteristic function of the acquisition device, called impulse response, which represents the image of a point object. Equivalently, this blur is translated in the Fourier domain by the multiplication of the Fourier transform of the landscape by the Fourier transform of the impulse response, called Modulation Transfer Function (MTF). The MTF characterizes the attenuation of spatial frequencies by the instrument. The deconvolution of an image aims to compensate for this blur by convolving the image with a so-called deconvolution filter, whose Fourier transform is close to the inverse of the MTF.

[0015] Artifacts may remain, especially when the digital image to be restored includes both noise and compression artifacts. Such artifacts are visually distracting to users and may impair further exploitation of the images, especially on the merged product.

[0016] Methods for reducing noise in images are known, for example in publication US2005 / 0259889 A1.

[0017] There is a need to improve the quality of image restoration in this context, and more generally, the restoration of noisy images.

[0018] For this purpose, the invention proposes a digital image restoration method according to claim 1.

[0019] Advantageously, the method of the invention has the effect of improving the quality of the images obtained after restoration, the restitution of instrumental acquisition noise making it possible to modify the noise distribution in the digital images before restoration (denoise removal and deconvolution). In the case of images acquired by satellite, denoising is followed by deconvolution and fusion.

[0020] The method according to the invention may have one or more of the characteristics below, taken in all their technically acceptable combinations.

[0021] The digital image to be restored was obtained by applying compression to the initial acquired image to obtain an initial compressed image, then decompressing the initial compressed image.

[0022] The restitution of instrumental acquisition noise is carried out in connection with the decompression of the initial compressed image.

[0023] The digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands and the steps of obtaining an intermediate digital image and of denoising the intermediate digital image are applied to each of the images acquired in different spectral bands, and the deconvolution step is applied to at least one of the intermediate digital images.

[0024] The method further comprises an image fusion step applied after denoising or after deconvolution, to obtain a final restored image.

[0025] The instrumental acquisition noise is modeled by a model parameterized by two coefficients, the coefficients having values used during the restitution of instrumental noise, and the denoising step involves the application of a variance stabilization transformation parameterized by the two coefficients characterizing said parameterized model of the instrumental acquisition noise.

[0026] The digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands, and values of the coefficients of the instrumental acquisition noise model dependent on the acquisition spectral band are determined for each acquisition spectral band.

[0027] The denoising step consists of the following substeps: applying the variance stabilization transformation to obtain a stabilized intermediate image, applying a stationary noise denoising method to the stabilized intermediate image, applying an inverse variance stabilization transformation to the digital image resulting from the step of applying a denoising method to obtain a denoised intermediate digital image.

[0028] In one embodiment, the variance stabilizing transformation is an Anscombe transformation.

[0029] The step of obtaining an intermediate digital image includes sub-steps of: obtaining, by applying a transformation called a compression transformation, a representation of the digital image to be restored by a plurality of blocks of coefficients, each block of coefficients corresponding to a block of digital image pixels to be restored, for at least one processed coefficient of a block of coefficients: calculating a noise threshold as a function of a value representative of the instrumental image acquisition noise model, comparing the absolute value of said processed coefficient to the noise threshold, and, when the absolute value of the processed coefficient is greater than or equal to said noise threshold, said coefficient is left unchanged, when the absolute value of the processed coefficient is less than said noise threshold, replacing the processed coefficient value with a noise value depending on said value representative of an image acquisition noise model.

[0030] The instrumental acquisition noise is modeled by a parameterized model defining the noise standard deviation σ by σ = a 2 + b ⋅ s , where a and b are coefficients of said model and s is a coefficient value representative of an average of the values associated with the pixels of a block of digital image pixels considered.

[0031] According to another aspect, the invention relates to a digital image restoration device according to claim 12.

[0032] According to one embodiment, in which the digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands, the device further comprises an image fusion module adapted to apply a fusion after denoising or after deconvolution, to obtain a final restored image.

[0033] According to another aspect, the invention relates to a computer program according to claim 11.

[0034] Other characteristics and advantages of the invention will emerge from the description given below, for information purposes only and in no way limiting, with reference to the appended figures, among which: there figure 1 schematically illustrates the main modules of an image processing system in which the invention is implemented; the figure 2 is a synopsis of the main modules of a restoration device according to one embodiment; the figure 3 is a diagram representing the functional blocks of a programmable device capable of implementing the invention; the figure 4 is a synopsis of the main steps of a restoration method according to a first embodiment of the invention; the figure 5 is a synopsis of the main steps of a restoration method according to a second embodiment of the invention.

[0035] The invention applies in particular in the context of satellite imaging, but is not limited to this field. It can also be applied to the restoration of any type of digital images presenting acquisition and compression noise.

[0036] The invention applies to the processing of digital images. In the following description, the term “image” is used to designate a digital image.

[0037] There figure 1 schematically illustrates an image processing system 1, comprising a first part 10 for acquiring images and a second part 20 for exploiting the acquired images.

[0038] In the case of satellite acquisition, the first part 10 is implemented on board a satellite, while the second part 20 is for example implemented in a ground-based image processing center.

[0039] The first acquisition part 10 comprises an image acquisition device 12, having an associated instrumental acquisition noise.

[0040] For example, the image acquisition device 12 comprises a plurality of image detector arrays fixed to a satellite. This device acquires images as the satellite passes over a landscape, by pushbroom scanning. In one embodiment, the image acquisition device 12 is of the multispectral type and allows the acquisition of multispectral images in several different narrow spectral bands, as well as a panchromatic image of wide spectral band, having a higher spatial resolution than multispectral images.

[0041] Alternatively, the image acquisition device is of the matrix type comprising a matrix of CCD or CMOS detectors.

[0042] Thus, the image acquisition device 12 makes it possible to acquire at least one digital image.

[0043] A digital image composed of one or more matrices of image samples or pixels, each image sample having an associated radiometry value. This representation domain is called the spatial domain. The spatial resolution of a digital image is defined by the number of pixels per row and column of the representation matrix.

[0044] Preferably, the image acquisition device 12 is adapted to acquire images in a plurality of spectral bands, or color images, formed from as many pixel matrices as there are acquisition spectral bands.

[0045] The initial digital image acquired then undergoes compression according to a given compression method, carried out by the compression module 14, and the digital data relating to the compressed digital image are stored, for example in a file, and / or transmitted for later use.

[0046] There are several known and standardized compression methods, for example JPEG or JPEG2000 for still images. Each compression standard implements a transformation of the digital image, hereinafter called a compression transformation. For example, JPEG implements a block-based discrete cosine transform or DCT, JPEG2000 implements a sub-band wavelet transform or DWT.

[0047] The first acquisition part 10 is connected to a transmission module 16, capable of transmitting digital data to the second part 20 for receiving and exploiting the acquired digital images. The transmission module 16 carries out, for example, a radio transmission.

[0048] The second part 20 of the system is connected to a reception module 22, capable of receiving digital data from the transmission module 16.

[0049] The reception module 22, for example a satellite reception antenna, is connected to a decompression module 24, adapted to carry out a corresponding decompression to obtain a usable digital image.

[0050] According to one embodiment of the invention, the module 24 performs decompression with restitution of the instrumental acquisition noise, making it possible to obtain an intermediate digital image.

[0051] For example, the restitution of instrumental noise for images having undergone compression described in patent application FR 3 025 640 is applied.

[0052] In one embodiment, the decompression module 24 performs standard decompression, and the instrumental noise restitution is subsequently performed by the image restoration module 26, with or without knowledge of the compression / decompression method applied. The instrumental noise restitution uses a noise model that is also used by the image restoration module 26.

[0053] The restoration module 26 implements a method for restoring digital images according to the invention.

[0054] According to a variant, the compression 14 and decompression 24 modules are optional, the image restoration module 26 nevertheless performing instrumental noise restitution on the digital image to be restored.

[0055] The intermediate digital image is therefore a digital image in the spatial domain, obtained by restitution of instrumental noise on the digital image to be restored.

[0056] The digital image to be restored is obtained either by applying compression to the initial acquired image to obtain an initial compressed image, then decompressing the initial compressed image, or it is simply the initial digital image acquired.

[0057] There figure 2 is a synopsis of the modules implemented by the restoration module 26 according to one embodiment.

[0058] The restoration module 26 is adapted to implement the obtaining 30 of an intermediate digital image by decompression with restitution of the instrumental acquisition noise, the denoising 32 on the intermediate digital image to obtain a denoised intermediate digital image and the deconvolution 34, optionally followed by a fusion 36 in the case of the processing of multispectral images, of the denoised digital image to obtain a final restored digital image.

[0059] Detailed implementations of decompression with restitution of instrumental acquisition noise, denoising and deconvolution / fusion will be described below.

[0060] Such an image restoration module 26 is for example implemented by a set of computer program instructions, executable by a programmable device.

[0061] There figure 3 schematically illustrates the main functional blocks of a programmable device, for example a computer, a workstation, adapted to implement a restoration method according to the invention.

[0062] A programmable device 40 suitable for implementing the method of the invention comprises a central processing unit 42, for example a processor (CPU), capable of executing pre-programmed operations or computer program instructions when the device 40 is powered up.

[0063] In one embodiment, a multi-processor central processing unit is used, allowing parallel calculations to be performed. The device 40 also comprises information storage means 44, for example registers, capable of storing executable code instructions allowing the implementation of programs comprising code instructions capable of implementing the method according to the invention.

[0064] The device 40 comprises control means 46 for updating parameters and receiving commands from an operator. When the programmable device 40 is an on-board device, the control means 46 comprise a telecommunications device for receiving commands and parameter values remotely.

[0065] Alternatively and optionally, the control means 46 are means for entering commands from an operator, for example a keyboard.

[0066] Optionally, the programmable device 40 comprises a screen 48 and an additional pointing means 50, such as a mouse.

[0067] The various functional blocks of the device 40 described above are connected via a communication bus 52.

[0068] Alternatively, the methods of the invention are implemented by graphics processors or GPUs, with a parallelized hardware architecture.

[0069] Alternatively, the methods of the invention are implemented by electronic devices of the programmable logic circuit type, such as electronic cards based on FPGA or ASIC, or chips that can be integrated into electronic devices such as mobile phones or cameras.

[0070] There figure 4 is a synopsis of the main steps of a digital image restoration method in a first embodiment of the invention.

[0071] As input, the process receives an input digital image which corresponds to the digital image to be restored after compression.

[0072] The first decompression step 30 with restitution of instrumental acquisition noise is broken down into three sub-steps 60, 62, 64.

[0073] In the first step 60, a representation of the input digital image is obtained in the compression transformation domain, also called the transformed domain.

[0074] It should be noted that the input digital image may be a digital image represented in the spatial domain, or be provided in an encoded representation after compression.

[0075] When the input digital image is represented in the spatial domain, the corresponding compression transformation, for example DCT or DWT, is applied to obtain a representation of the image by blocks of coefficients in the transformed domain.

[0076] When the input digital image is represented by compressed data, decoding is applied according to the compression standard used, to obtain a representation of the image by blocks of coefficients, called representation in the transformed domain.

[0077] When the input digital image is represented in the spatial domain, the input digital image may result from compression followed by decompression, but the compression / decompression method is not known from the input digital image. In this case, a chosen compression transformation is applied in step 60, for example a block DCT or a wavelet transformation on a chosen number of decomposition levels. In this case, the applied compression transformation is chosen in step 60, and not as a function of the compression / decompression method previously applied to the initial image which is not known at this stage.

[0078] Step 60 is followed by a step 62 of modifying the coefficients according to a threshold value calculated according to a parameterized model of instrumental acquisition noise.

[0079] Preferably, the parameterized model of instrumental noise defines the noise standard deviation σ by: σ = a 2 + b ⋅ s where a and b are coefficients of said instrumental noise model and s is a coefficient value in the transformed domain representative of the average of the sample values of a processed block of the digital image.

[0080] For example, s is the DC coefficient of the processed block when the compression transform is a block-wise DCT, and s is the low-frequency wavelet coefficient spatially corresponding to the processed block for a discrete wavelet transform DWT.

[0081] According to another variant and in the case of a compressor based on the discrete wavelet transformation DWT, the renoising is refined by applying it to each step of the recomposition of the wavelet coefficients, for each recomposition level, the value s then representing the low-frequency coefficient of the recomposition level considered and the renoising concerning the wavelet coefficients belonging to the three high-frequency sub-bands corresponding to the low-frequency sub-band of the recomposition level considered. When the acquisition device is known, the values of the coefficients a and b are known and stored prior to the implementation of the method.

[0082] For example, for the PLEIADES-1A satellite with a configuration of 13 TDI stages, the a and b values in the panchromatic band, in the spatial domain, are equal to 2.267 and 0.0393 respectively for a signal s expressed in encoder "steps" (in English "digital count"). These values are adapted, if necessary, according to the compression transformation for the use of the noise model in the transformed domain, as explained above.

[0083] In the case of multispectral image acquisition, a pair of values (a,b) is estimated and stored for each spectral acquisition band.

[0084] Alternatively, the values of the coefficients a and b are provided by an operator, or estimated from one or more images acquired by a given acquisition device or from one or more arbitrary images.

[0085] We define a noise threshold S bruit depending on a noise proportionality factor K d , and a reconstruction proportionality factor K r

[0086] The noise threshold S bruit is obtained from the standard deviation σ of the acquisition noise: s bruit = K d ⋅ σ

[0087] Where K d is the user-supplied constant noise proportionality factor.

[0088] K d can typically be 0.5 or 1, and more generally be between 0 and 3.

[0089] More generally, the noise threshold S bruit is obtained by applying a function f(), which is not necessarily linear: s bruit = f K d σ

[0090] For each block of the transformed domain considered, the absolute value of each coefficient noted C i,j of the processed block is compared to the noise threshold S bruit . When the absolute value of the coefficient C i,j is greater than the noise threshold, it is left unchanged. When the absolute value of the coefficient C i,j is less than the noise threshold S bruit , the coefficient C i,j is replaced by a local noise value which is a function of the representative value of the acquisition noise model associated with the processed block. In general, we can write C ij ′ = g K r σ .

[0091] For example, the function g() is given by: C ij ′ = sign C ij ⋅ K r ⋅ σ ⋅ rand Or : sign ( C ij ) is the sign of the coefficient C i,j; K r is the reconstruction proportionality factor, which is typically 1 but can more generally be between 0 and 3, rand is a value obtained by pseudo-random drawing according to a predetermined distribution law, for example Gaussian, with variance equal to 1. The absolute value of the rand value, noted | rand |, is used in the formula of equation (EQ 4).

[0092] Preferably, all coefficient blocks of the transformed digital image are processed.

[0093] The inverse compression transformation is then applied during an inverse transformation application step 64 to obtain an intermediate digital image.

[0094] Advantageously, this intermediate digital image no longer presents compression artifacts, but presents instrumental acquisition noise artifacts similar to those introduced by the image acquisition device. In other words, the image at the processing output has noise having the same characteristics as the instrumental noise present in the image before compression.

[0095] The first step 30 of decompression with restitution of instrumental noise is followed by a step 32 of denoising applied to the intermediate image.

[0096] Denoising step 32 includes substeps 66 to 68.

[0097] Substep 66 consists of applying a variance stabilizing transformation, which is preferably the Anscombe transformation, to the previously obtained intermediate image. A stabilized intermediate image is then obtained.

[0098] Depending on possible variants, other variance-stabilizing transformations, e.g., the Freeman-Tukey transformation, are applicable.

[0099] The Anscombe transformation is parameterized by two coefficients a, b. Advantageously, the value of the coefficients (a,b) defining the applied Anscombe transformation is the same as that of the coefficients of the observed instrumental noise model.

[0100] The Anscombe transform, applied in the spatial domain, is defined by: A p = 2 b ⋅ bp + 3 8 b 2 + a 2

[0101] Where a and b are coefficients, of the same value as the coefficients of the instrumental acquisition noise model defined above, in the spatial domain, and p is the radiometry value of a processed pixel.

[0102] As mentioned above, in the case of multispectral acquisition, each spectral band has an associated pair of coefficient values (a,b).

[0103] Step 66 of applying the stabilization transformation is followed by a step 68 of applying a stationary noise denoising method, itself followed by a step 70 of applying the inverse variance stabilization transformation, to obtain a denoised intermediate digital image.

[0104] Stationary noise is noise whose standard deviation is substantially constant over the image.

[0105] The inverse Anscombe transform is given by: A − 1 p = 1 4 bp 2 − 3 8 b − a 2 b

[0106] Denoising step 32 is followed by deconvolution step 34 to obtain the restored digital image. Deconvolution improves the sharpness of the resulting image.

[0107] Any known deconvolution method is applicable. For example, a Fourier transform, deconvolution filtering in the Fourier domain, and an inverse Fourier transform are applied.

[0108] The deconvolution filtering used is related to the characteristic optical transfer function of the image acquisition device, which is measured and known, in the same way as the instrumental noise parameters a and b.

[0109] As a non-limiting example, we can apply the deconvolution method described in the article “Restoration technique for Pleiades-HR panchromatic images” by C. Latry et al, published in International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B1, 2012.

[0110] Preferably, the stationary noise denoising method applied in step 68 is a non-local Bayesian method, such as the method described in the article by Marc Lebrun, Antoni Buades, and Jean-Michel Morel, “A Nonlocal Bayesian Image Denoising Algorithm” published in SIAM Journal on Imaging Sciences, vol. 6, no. 3, pages 1665-1668. The principle of the NL-Bayes method is to replace the value p of a pixel p(x,y) by a weighted average of the values of other pixels p' of the processed image, chosen on a criterion of distance between neighborhoods of given size.

[0111] Alternatively, other known stationary noise denoising methods, such as the method known by the acronym BM3D for "block matching and 3D filtering" or "deep learning" type denoising, are also applicable.

[0112] There figure 5 is a synopsis of the main steps of a restoration method according to a second embodiment of the invention, applied in particular to images acquired by satellite.

[0113] In this embodiment the input images are on the one hand the image l 0 which is the panchromatic image and n images I 1 to I n , each of the images I 1 to I n corresponding to a given acquisition spectral band.

[0114] Images I 1 to I n are hereinafter called multispectral images, it being understood that each of these images corresponds to a given narrow spectral acquisition band.

[0115] The images I 1 to I n have a lower spatial resolution than the spatial resolution of the panchromatic image I 0 . Typically the resolution ratio is 4 for satellite imagery.

[0116] These images are acquired during acquisition steps 80, 82.

[0117] The image acquisition steps are followed by compression steps 84, 86, according to a predetermined compression method, using a compression transformation as explained above.

[0118] For example, in one embodiment the acquisition steps 80, 82 and compression steps 84, 86 are performed on board a satellite.

[0119] These steps 84, 86 are followed by decompression steps with instrumental noise restitution 88, 90 for the panchromatic image I 0 , and 92, 94 for each of the multispectral images I 1 to I n . The decompression with instrumental noise restitution is carried out in a manner analogous to that described with reference to the figure 4 for the implementation of step 30.

[0120] At the end of the decompression steps with instrumental noise restitution 88, 90, 92, 94 we obtain an intermediate panchromatic image I 0_interm and intermediate multispectral images I 1_interm to I n_interm .

[0121] Decompression with instrumental noise restitution is followed by a denoising step 96 for the intermediate panchromatic image I 0_interm and a denoising step 98 for each intermediate multispectral image I 1_interm to I n_interm .

[0122] The respective denoising steps 96, 98 are analogous to those described with reference to the figure 4 .

[0123] The denoising step 98 comprises, in one embodiment, for each intermediate multispectral image I k_interm , the application of a variance stabilization transformation, for example the Anscombe transformation parameterized with the coefficients a,b of the instrumental noise model for the spectral band associated with I k , then the application of a stationary noise denoising method, and finally the application of the inverse variance stabilization transformation.

[0124] At the end of the denoising step 98, restored multispectral images I' 1 to I' n are obtained.

[0125] Similarly, the denoising step 96 applied to the intermediate panchromatic image I 0_interm comprises, in one embodiment, the application of a variance stabilization transformation, for example the Anscombe transformation parameterized with the coefficients a, b of the instrumental noise model for the panchromatic band, then the application of a stationary noise denoising method, and finally the application of the inverse variance stabilization transformation.

[0126] The denoising step 96 is followed by a deconvolution step 100, which makes it possible to obtain a restored panchromatic image I' 0 .

[0127] Finally, steps 98 and 100 are followed by a step 102 of merging all or part of the restored images I' 0 , I' 1 to I' n .

[0128] Preferably, the multi-resolution image fusion method described in the French patent published under number FR2994007 by S. Fourest and C. Latry is applied.

[0129] Alternatively, another multi-resolution image fusion method known to those skilled in the art may be applied.

[0130] We obtain a final restored color image, I final, whose visual quality is improved compared to an image obtained by a classic compression / decompression, restoration and fusion process.

[0131] According to a variant of the embodiment of the figure 5 , the deconvolution step 100 is also applied to each of the multispectral images resulting from the denoising step 98, in a manner analogous to the application of the deconvolution after the denoising step 96, so as to improve the sharpness of the multispectral images.

[0132] Advantageously, the application of a variance stabilization transformation makes it possible to transform the instrumental noise restored on the images into a noise whose variance is independent of the image signal, which allows the use of a stationary noise denoising algorithm.

Claims

1. A method implemented by a programmable device, for restoring a digital image, the digital image to be restored being derived from an initial image acquired by an image acquisition device having associated instrumental acquisition noise, in which the digital image to be restored was obtained by applying a compression to the initial acquired image to obtain a compressed initial image, then a decompression of the compressed initial image, characterized in that it includes the following steps: - obtaining (30, 60-64, 80-94) an intermediate digital image, the intermediate digital image being a digital image obtained by restoring the instrumental acquisition noise to the digital image to be restored, - denoising (32, 66-70, 96, 98) of the intermediate digital image to obtain a denoised intermediate digital image, - deconvoluting (34, 100) the denoised intermediate digital image to obtain a restored digital image.

2. The method for restoring a digital image according to claim 1, wherein said restitution of instrumental acquisition noise is performed on the basis of a noise model comprising a parameter is defined as a function of a decompression coefficient of the initial compressed image.

3. The method for restoring a digital image according to any of claims 1 to 2, wherein the digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands and wherein the steps of obtaining (80-94) an intermediate digital image and denoising (96, 98) the intermediate digital image are applied to each of the images acquired in different spectral bands, and the deconvolution step (100) is applied to at least one of the intermediate digital images.

4. The method for restoring a digital image according to claim 3, further including a step of merging (102) images applied after denoising (96, 98) or after deconvolution (100), to obtain a final restored image.

5. The method for restoring a digital image according to any of claims 1 to 4, wherein the instrumental acquisition noise is modeled by a model parameterized by two coefficients, the coefficients having values used during the restitution of instrumental noise, and wherein the denoising step includes the application of a variance stabilization transformation parameterized by the two coefficients characterizing said parameterized model of the instrumental acquisition noise.

6. The method for restoring a digital image according to claim 5, wherein the digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands, and wherein values of the coefficients of the instrumental acquisition noise model depend on the acquisition spectral band and are determined for each acquisition spectral band.

7. The method for restoring a digital image according to any one of claims 5 or 6, wherein the denoising step includes the following substeps: - applying (66) the variance stabilizing transformation to obtain a stabilized intermediate image, - applying (68) a stationary noise denoising method to the stabilized intermediate image, - applying (70) an inverse transformation of said variance stabilizing transformation to the digital image resulting from the step of applying a denoising method to obtain a denoised intermediate digital image.

8. The method for restoring a digital image according to claim 7, wherein said variance stabilization transformation is an Anscombe transformation,9. The method for restoring a digital image according to any of claims 1 to 8, wherein the step of obtaining an intermediate digital image comprises the following sub-steps: - obtaining (60), by applying a transformation known as a compression transformation, a representation of the digital image to be restored by a plurality of coefficient blocks, each coefficient block corresponding to a block of digital image pixels to be restored, - for at least one processed coefficient of a coefficient block: - calculating a noise threshold based on a value representative of the instrumental image acquisition noise model, - comparing the absolute value of said processed coefficient with the noise threshold, and, • when the absolute value of the processed 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 (62) of the processed coefficient value by a noise value dependent on said value representative of an image acquisition noise model.

10. The method for restoring a digital image according to any one of claims 1 to 9, wherein the instrumental acquisition noise is modeled by a parameterized model defining the standard deviation of noise σ by σ = a 2 + b . s , where a and b are coefficients of said model and s is a coefficient value representative of an average of the values associated with the pixels of a block of digital image pixels in question.

11. A computer program including instructions for implementing the steps of a method for restoring a digital image according to any one of claims 1 to 10 during execution of the program by a processor of a programmable device.

12. A device for restoring a digital image, the digital image to be restored being derived from an initial image acquired by an image acquisition device having associated instrumental acquisition noise, wherein the digital image to be restored was obtained by applying a compression to the initial acquired image to obtain an initial compressed image, followed by a decompression of the initial compressed image, characterized in that it includes modules, implemented by a processor of a programmable device, adapted to: - obtain an intermediate digital image with restitution of the instrumental acquisition noise, - denoise the intermediate digital image to obtain a denoised intermediate digital image, - deconvolve the denoised intermediate digital image to obtain a restored digital image.

13. The device for restoring a digital image according to claim 12, wherein the digital image to be restored is a multispectral image composed of a plurality of images acquired in different spectral bands, further including an image fusion module adapted to apply a fusion after denoising or after deconvolution, to obtain a final restored image.