METHOD FOR ESTIMATING PARAMETERS OF AN OBJECT TO BE ESTIMATED IN A DIGITAL IMAGE AND METHOD FOR REMOVING THE OBJECT FROM THE DIGITAL IMAGE

Description

[0001] The invention relates to a method for parametrically estimating an object to be estimated in a digital image, and to a method for removing the object from the digital image. It is particularly applicable to digital images obtained by X-ray imaging, but can be extended to any type of digital imager, for example, infrared or visible spectrum imaging.

[0002] In this application, three components may be present in the digital image: 1. The clinical background or content, corresponding to the actual object being imaged, which we seek to visualize. This could be, for example, a patient, or a specific area of ​​a patient, in X-ray imaging. 2. Acquisition noise, corresponding to random disturbances in image formation. This can include electronic noise, to which, in X-ray imaging, we must add photon noise, due to the small number of photons present in the image and which follows Poisson statistics. 3. Objects that we seek to estimate, which may interfere with the visualization of the content.

[0003] The digital image can then be written in two different ways depending on whether the noise is treated independently of the other two components: Image = Object + Contenu + bruit

[0004] Or considered as belonging to the content: Image = Object + Contenu

[0005] There figure 1This represents an example of an image containing an object to be assessed. In this case, the figure represents a radiological image whose content is an area of ​​a patient. To improve image contrast, and more generally the quality of the radiological image, it is common practice for the practitioner to use an anti-scatter grid. Positioned between the patient and the detector, it reduces the contribution of scattered radiation caused by the patient. On the figure 1 The horizontal variations are due to the presence of the anti-scatter grid; these variations can obscure important information and potentially lead to misdiagnosis. Image processing must therefore be applied to remove these horizontal variations from the digital image without altering its content.

[0006] However, such processing presents two main difficulties. First, it is necessary to be able to remove only the object, knowing that the image also contains noise, and that the content can be a very strong and unpredictable signal. Indeed, the content can vary significantly depending on the area of ​​the patient being imaged, and even within the same area, the content can vary from one patient to another. The other difficulty arises from the fact that the object to be estimated is not perfectly defined; that is, several parameters of the object may be unknown. figure 2 This represents another example of an object to be estimated. It is a CR square superimposed on the IM image, which interferes with the correct visualization of the content. Neither the object's amplitude (the pixel value, for example, the gray level for a radiology image) nor its size is known.

[0007] Currently, two types of solutions allow a signal to be processed in order to remove unwanted components.

[0008] The first group includes classic filters, notably the matched filter, the Wiener filter, and the Kalman filter. These solutions are based on an estimation and statistical modeling of the content and the signal that interferes with the content estimation (generally random noise, uncorrelated with the content). In our case, the content (the patient) cannot be modeled, yet it represents the majority of the signal present in the image. Such solutions can therefore be implemented in so-called "flat" areas of the image where the patient is not present, but remain suboptimal when the patient is present. Furthermore, these solutions have the drawback of modifying the image content to varying degrees when the object is removed, for example, if the image includes a component that resembles the object to be removed. With reference to the figure 1, if there is, in the content acquired by the detector, an element resembling the anti-scatter grid, this element will potentially be removed in the event of appropriate filtering.

[0009] The second type of solution includes spectral estimation methods, notably the IAA (Iterative Adaptive Approach), described in particular in the article « Source localization and sensing: A nonparametric lterative Adaptive Approach based on weighted least squares” (Yardibi T. et al., Aerospace and Electronic Systems, IEEE Transactions on 46.1 2010), and the SLIM method (Sparse Learning via Iterative Minimization), described in particular in the article «Sparse Learning via Iterative Minimization with application to MIMO Radar Imaging” (Tan X. et al., IEEE Transactions on Signal Processing, vol. 59, no. 3, March 2011). These methods are suitable for radar and telecommunications applications, where the signal is one-dimensional, but cannot currently be transposed to the case of a two-dimensional signal and thus used for image processing. Furthermore, in the SLIM method, the noise is assumed to be white. This assumption cannot be used for an image acquired by a detector, where the noise is specific to each imaging technique. In particular, in X-ray imaging, photon noise is not white.

[0010] Finally, the aforementioned methods only consider two signals: the content and the noise. Therefore, they do not allow for a parametric estimation of an object that interferes with the visualization of the content and is distinct from the acquisition noise.

[0011] The document "Artifact Suppressed Dictionary Learning for Low-Dose CT Image Processing" (Chen Yang et al.) describes a method for suppressing artifacts present in a CT image.

[0012] The invention therefore aims to provide a method for parametrically estimating an object in a digital image, distinct from the acquisition noise, and requiring no prior precise knowledge of the content, the object to be estimated, or the acquisition noise. The invention also aims to provide a method for removing the object from the digital image, which does not alter the content of the image corresponding to the actual imaged object.

[0013] An object of the invention is therefore a parametric estimation method as defined in claims 1 to 12.

[0014] The invention also relates to a suppression method as defined in claim 13.

[0015] Other features, details and advantages of the invention will become apparent from the description provided with reference to the accompanying drawings given by way of example, which represent, respectively: there figure 1 As already described, an example of oscillation present in an X-ray image, due to the presence of an anti-scattering grid; the figure 2 , an example of an object to estimate in a digital image; the figure 3 a functional diagram of the process according to the invention; the figure 4A , an example of a dictionary of content components, obtained from Fourier components; the figure 4B , another example of a content component dictionary, obtained from the detector's impulse response; the figure 5 , an example of an object component dictionary, for which the parameter to be estimated is the size of a square; the figure 6, a schematic illustration of the different stages of the algorithm for jointly determining the amplitudes of object and content components, according to a first embodiment; the figure 7 , a schematic illustration of the different stages of the algorithm for jointly determining the amplitudes of object and content components, according to a second embodiment; the figure 8 a graph illustrating the magnitude of each object component; the figure 9 , the image of the figure 2 , in which the object was removed.

[0016] There figure 2 illustrates a digital image IM, including an object (a CR square) that whitens a corresponding part of the digital image. On the figure 2The dotted lines surrounding the CR square are shown only for clarity. We therefore seek to estimate the value of at least one parameter characterizing the object, and the object's amplitude, to estimate how much the CR square brightens the image.

[0017] There figure 3 illustrates a functional diagram of the process according to the invention.

[0018] The initial first step a) is to create a dictionary of content components and to create a dictionary of object components.

[0019] A first sub-step therefore consists of creating a dictionary of content components.

[0020] According to a first embodiment, the content component dictionary can be created without prior assumptions, by generating a set of images of basic two-dimensional sinusoidal signals. figure 4Aillustrates such a dictionary. Each of the content components (COMP_CONT_1, COMP_CONT_2, ..., COMP_CONT_K) (K greater than or equal to 1) represents a two-dimensional sinusoidal signal, that is, at a particular frequency. The represented content components are called Fourier components. The number of content components represented on the figure 4 is not exhaustive.

[0021] According to another embodiment, the content component dictionary can be created by generating a set of images representative of the modulation transfer function of the sensor forming the digital image. figure 4BThis illustrates such a dictionary. This dictionary is generated from the detector's impulse response. Each component (COMP_CONT_1, COMP_CONT_2, ..., COMP_CONT_K) (K greater than or equal to 1) represents the detector's impulse response at a pixel. The content dictionary thus consists of all these images, for each pixel of the detector. This dictionary advantageously closely reflects the physical characteristics of the detector, particularly in X-ray imaging.

[0022] In another embodiment, the content component dictionary can be created using a machine learning algorithm, specifically the K-SVD algorithm. K-SVD is particularly well-suited for creating and learning dictionary components, making it easily implementable. A content component dictionary using such an algorithm could, for example, generate components from a database of X-ray images. Other machine learning algorithms, such as neural networks or the Support Vector Machine (SVM), can also be used.

[0023] The dimensions of each content component can be the same as the dimensions of the digital image containing the object. Parametric estimation can also be performed on a portion of the digital image, for example, if the object is highly localized. In this case, the dimensions of each content component are the same as the dimensions of that portion of the digital image.

[0024] A second substep of the initial first step (a) consists of creating a dictionary of object components. The object component dictionary can be created by generating a set of images, each with a different parameter value. The parameter can be, but is not limited to, the object's frequency, shape, size, and location.

[0025] There figure 5 An example illustrates a dictionary of object components, for which the parameter to be estimated is the size of the square to be estimated in the figure 2The number of object components (COMP_OBJ_1, COMP_OBJ_2, COMP_OBJ_3, COMP_OBJ_4) present in the figure 5 The number of possible values ​​that the parameter to be estimated can take is not limited. In particular, it can include as many components as there are different sizes of squares in the digital image (for example 1x1 pixel, 2x2 pixels, 3x3 pixels, etc.).

[0026] The most frequent case is where there is a parameter to estimate (for example, the size of the square) and the corresponding object's amplitude. Thus, the object component dictionary contains as many object components as there are possible values ​​for the parameter. The object component with the largest amplitude corresponds to the correct parameter value. It is also possible to consider the case where there are several parameters to estimate (for example, the location, the size of the square), in addition to the corresponding object's amplitude. If the object is a sum of known components with unknown amplitudes, then the amplitudes of each component must be estimated.

[0027] We thus have a dictionary of content components, and a dictionary of object components. Each component is a matrix of I rows and J columns which we will write in the form of a vector of size M, M being the size of the image, or of the part of the image in which the object is located (i.e. the set of corresponding pixels).

[0028] The joint determination of the amplitude of each of the content components of the content component dictionary and the object components of the object component dictionary present in the digital image (step b) in the figure 3 ) can be carried out in accordance with the two embodiments described below. 1 er< method of implementation

[0029] Noise present in the digital image is processed separately from the content and the object. This separate processing can be justified by the different nature of the data being processed: the data of the content and the object are not random, whereas that of the noise is.

[0030] Let D be an MxL matrix containing the different components to be estimated (object and content). The matrix D is therefore a concatenation of the dictionary of content components and the dictionary of object components. L is the number of components in the dictionary of content components and the dictionary of object components. D i< , which is a vector of size M, is the i-th component of the matrix D.

[0031] The correlation information for the noise present in the digital image is stored in a matrix N, distinct from the matrix D. The noise in the digital image is therefore treated separately from the content and the object. The noise correlation matrix N can be determined by prior knowledge of the digital image acquisition detector. Noise can indeed be specific to each imaging technique (X-ray, infrared, visible). In particular, in X-ray imaging, photon noise is generally not white noise. The noise correlation matrix N can be estimated using the Wiener-Khintchine theorem, by taking the inverse Fourier transform of the power spectral density of a "flat" image, that is, one without content.

[0032] Alternatively, assuming that there is no correlation between pixels, N can be an identity matrix of size MxM.

[0033] Either ya vector representing the digital image, namely the value (also called brightness) of each pixel.

[0034] Either β an estimation vector of size L containing the estimated amplitudes of each component D i< . So, β = [β 1 , β 2 , ..., β L ]

[0035] Let a, b, and q be parameters of the algorithm; a and b depend on the nature and knowledge of the noise. If there is no particular knowledge about the noise, then a=0 and b=0.

[0036] The algorithm for jointly determining the amplitude of each of the content components of the content component dictionary and the object components of the object component dictionary present in the digital image comprises an initialization step and six iteration steps, as illustrated in the figure 6 . Initialization step

[0037] The value of the estimation vector βcan be initialized in any way (for example with random values), or by performing the dot product between each component D i< and the vector y representing the digital image: β i = D i . y , pour 1 ≤ i ≤ L

[0038] Furthermore, an estimate of the noise variance (σ²< ) is initialized, either in any way (for example with random values), or with the following relationship: σ 2 = y − Dβ H N − 1 y − Dβ + a M + b β being at its initial value. Iteration steps

[0039] 1. We calculate P = | β | q< Z, where q is a user-defined parameter; it is a real number between 1 and 2. The user can change the value of this parameter depending on the final result, namely the quality of the object removal in the image. Z is a diagonal matrix used to prioritize the estimation of certain components; in the general case, Z will be the identity matrix. Z is user-configurable. 2. We calculate a matrix R estimating the covariance matrix of the signal y digital image: R = Ddiag P D H + σ 2 N 3. We calculate a normalization term: Nor = β i − q 4. We then calculate each estimated amplitude with component β i using the formula: β i = D i H R − 1 y Nor 5. The estimate of the noise variance σ2< is calculated using the following formula: σ 2 = y − Dβ H N − 1 y − Dβ + a M + b 6. Optionally, the dictionary components can be modified to minimize a projection error ε defined by: ε = y − Dβ 2 Each component D i<The dictionary can, in particular, be updated using an update parameter. λ between 0 and 1: D new i = 1 − λ D old i + λ y − Dβ + D old i β i β i Or D old i is the component D i< dictionary to be updated, and D new i , the component D i< dictionary update. The update parameter λ can be set to 0; in this case, there is no update to the dictionary components. Updating the components D i< Dictionary updates can be performed using other methods, including binary search or gradient descent. This update step can be applied to the content components of the content component dictionary and / or the object components of the object component dictionary.

[0040] These six steps are repeated until the algorithm converges, that is, when the estimation vector βand the estimate of the noise variance σ²< convergent. The noise variance σ²< being directly dependent on β The convergence of one implies the convergence of the other.

[0041] The convergence of the estimation vector β and the convergence of the noise variance estimate σ² can be considered reached respectively when the norm of the estimation vector β evolves below a predetermined threshold between two successive iterations, and when the value of the noise variance estimate σ 2< evolves below a predetermined threshold between two successive iterations.

[0042] Alternatively, convergence can be considered to be reached after a predetermined number of iterations. It is then possible to determine the value(s) of the object's parameter(s) (step c) in the figure 3 Step c) includes the following substeps: determination of a subset of object components of greater amplitude; linear combination of the object components of said subset.

[0043] The subset of object components with the highest amplitude may include, in particular, a single component.

[0044] From the algorithm described above, it is also possible to reconstruct the content, corresponding to the actual imaged content (step d) in the figure 3 ), by performing a linear combination of all content components. This reconstruction makes it possible to obtain a noise-free image. 2nd< method of implementation

[0045] According to this second unclaimed embodiment, the noise present in the digital image is treated with the content and the object, namely in one and the same matrix.

[0046] Compared to the first embodiment, the matrix D is therefore the result of the concatenation of the object component dictionary, the content component dictionary, and the noise correlation information present in the digital image (which constitutes the matrix N in the first embodiment).

[0047] y is the vector representing the digital image, namely the value of each pixel.

[0048] β is the L-sized estimation vector containing the estimated amplitudes of each component D i< . So, β = [β 1 , β 2 , ..., β L ]

[0049] q is a parameter of the algorithm. Compared to the first embodiment, there is no calculation of the estimate of the noise variance σ 2< ; the parameters a and b are therefore not used.

[0050] The joint determination algorithm also includes an initialization step and five iteration steps, as illustrated in the figure 7 . Initialization step

[0051] The value of the estimation vector β can be initialized in any way (for example with random values), or by performing the dot product between each component D i< and the vector y representing the digital image: β i = D i . y , pour 1 ≤ i ≤ L Iteration steps

[0052] 1. We calculate P = | β | q< Z, where q is a user-defined parameter; it is a real number between 1 and 2. The user can change the value of this parameter depending on the final result, namely the quality of the object removal in the image. Z is a diagonal matrix used to prioritize the estimation of certain components; in the general case, Z will be the identity matrix. Z is user-configurable. 2. We calculate a matrix R estimating the covariance matrix of the signal y digital image: R = Ddiag P D H 3. We calculate a normalization term: Nor = D i H R − 1 D i 4. We then calculate each estimated amplitude with component β i using the formula: β i = D i H R − 1 y Nor 5. Optionally, the dictionary components can be modified to minimize a projection error ε, in the same way as for the first embodiment.

[0053] These five steps are repeated until the algorithm converges, that is, when the estimation vector β converge.

[0054] The convergence of the estimation vector β can be considered reached when the norm of the estimation vector β evolves below a predetermined threshold between two successive iterations.

[0055] Alternatively, convergence can be considered to be reached after a predetermined number of iterations. Steps c) and d) are the same as for the first embodiment. 3rd< method of implementation

[0056] According to this third, unclaimed embodiment, noise, content, and object are treated separately. This separate treatment can be justified by the different nature of the data being processed: noise is random, content is modeled by a dictionary of content components, and the object is modeled by a dictionary of parametric object components.

[0057] Let Dc be a matrix of size MxLc containing the different components to be estimated from the content. Let Ds( µ ) a parametric matrix of size MxLs containing the different components of the object to be estimated. The vector µ is a parameter selected from a set including, but not limited to, the frequency, shape, size, and location of the object. Lc is the number of components in the content component dictionary, and Ls is the number of components in the object component dictionary. Dc i<, which is a vector of size M, is the i-th component of the matrix Dc, and Ds i< (µ) , which is a vector of size M, is the i-th component of the matrix Ds( µ ).

[0058] Let D be the matrix resulting from the concatenation of the dictionary of object components Ds( µ ) and the Dc content component dictionary.

[0059] The correlation information for the noise present in the digital image is found in a matrix N, distinct from the matrices Dc and Ds( µThe noise present in the digital image is therefore treated separately from the content and the object. The noise correlation matrix N can be determined by prior knowledge of the digital image acquisition detector. Noise can indeed be specific to each imaging technique (X-ray, infrared, visible). In particular, in X-ray imaging, photon noise is generally not white noise. The noise correlation matrix N can be estimated using the Wiener-Khintchine theorem, by taking the inverse Fourier transform of the power spectral density of a "flat" image, that is, one without content.

[0060] Alternatively, assuming that there is no correlation between pixels, N can be an identity matrix of size MxM.

[0061] Either y a vector representing the digital image, namely the value (also called brightness) of each pixel.

[0062] Either βca content estimation vector, of size Lc, containing the estimated amplitudes of each content component Dc i< . So, βc = [βc 1, βc 2, ..., βc Lc]

[0063] Either βs an object estimation vector, of size Lc, containing the estimated amplitudes of each object component Ds i< (µ) . So, βs = [βs 1, βs 2, ..., βs Ls]

[0064] Either β the vector resulting from the concatenation of the vector containing the amplitudes of the object components βs and the vector containing the amplitudes of the content components βc .

[0065] Let a, b, and q be parameters of the algorithm; a and b depend on the nature and knowledge of the noise. If there is no particular knowledge about the noise, then a=0 and b=0.

[0066] The algorithm for jointly determining the amplitude of each of the content components of the content component dictionary and the object components of the object component dictionary present in the digital image comprises an initialization step and six iteration steps, as illustrated in the figure 6 . Initialization step

[0067] The values ​​of the content estimation vectors βc can be initialized in any way (for example with random values), or by performing the dot product between each component Dc i< and the vector y representing the digital image: βc i = D c i . y , pour 1 ≤ i ≤ Lc

[0068] The values ​​of the object estimation vectors βs can be initialized in any way (for example with random values), or by performing the dot product between each component Ds i< and the vector yrepresenting the digital image: βs i = D s i μ . y , pour 1 ≤ i ≤ Ls

[0069] Furthermore, an estimate of the noise variance (σ²< ) is initialized, either in any way (for example with random values), or with the following relationship: σ 2 = y − Dcβc − Ds μ βs H N − 1 y − Dcβc − Ds μ βs + a M + b

[0070] βc And βs being at their initial value. Iteration steps

[0071] 1. We calculate P = |β| q< Z , where q is a user-defined parameter; it is a real number between 1 and 2. The user can change the value of this parameter depending on the final result, namely the quality of the object removal in the image. Z is a diagonal matrix used to prioritize the estimation of certain components; in the general case, Z will be the identity matrix. Z is user-configurable. 2. We calculate an estimation matrix R for the covariance matrix of the signal y of the digital image: R = Ddiag P D H + σ 2 N 3. We calculate a normalization term: Nor = β i − q 4. We then calculate each estimated amplitude with component β i using the formula: β i = D i H R − 1 y Nor 5. The estimate of the noise variance σ2< can be calculated, for example, using the following formula: σ 2 = y − Dcβ c − Ds μ βs H N − 1 y − Dcβ c − Ds μ β s + a M + b 6. Optionally, the object dictionary components can be modified to minimize a projection error ε defined by: ε = y − Dcβc − Ds μ βs 2 Each component D i< The dictionary can notably be updated from a new value of the parameter µ by a gradient descent process. For this, we call J the Jacobian of the function | y - Dc βc - Ds(µ) βs | compared to µ: J = ∂ y − Dcβ c − Ds μ βs ∂ μ The new value of the parameter µ is then given by μ new = μ old + J T × J − 1 × J T × y − Dcβc − Ds μ β s

[0072] These six steps are repeated until the algorithm converges, that is, when the estimation vector βand the estimate of the noise variance σ²< convergent. The noise variance σ²< being directly dependent on β The convergence of one implies the convergence of the other.

[0073] The convergence of the estimation vector β and the convergence of the noise variance estimate σ² can be considered reached respectively when the norm of the estimation vector β evolves below a predetermined threshold between two successive iterations, and when the value of the noise variance estimate σ 2< evolves below a predetermined threshold between two successive iterations.

[0074] Alternatively, convergence can be considered to be reached after a predetermined number of iterations. It is then possible to determine the value(s) of the object's parameter(s) (step c) in the figure 3 Step c) includes the following substeps: Determining the value of µ

[0075] From the algorithm described above, it is also possible to reconstruct the content, corresponding to the actual imaged content (step d) in the figure 3 ), by performing a linear combination of all content components. This reconstruction makes it possible to obtain a noise-free image.

[0076] The invention also relates to a method for removing an object present in a digital image representing real-world image content, comprising the steps of: estimation of the value of at least one parameter of the object and its amplitude in accordance with the aforementioned parametric estimation method; determination of the object from the parameter of the object and its amplitude; pixel-by-pixel subtraction of the object in the digital image.

[0077] Using the first implementation of the algorithm, it is possible to obtain an estimate of the size and amplitude of the square. figure 8represents the result of the algorithm's execution. The parameter value of each component of the object dictionary is on the x-axis, and the estimated amplitude of each of these components is on the y-axis (on a logarithmic scale). This figure shows that the algorithm detected a square of size 7 with a large amplitude. Thus, the square located at the figure 2 is 7 inches wide and has an amplitude of 52. Using this information, it is then possible to correct the image to obtain a perfect correction, as illustrated by the figure 9 .

[0078] The method has been described for the entire image formed by the detector. It can also be applied to a part of the image formed by the detector, by adapting the dimensions of the matrices and vectors necessary for the parametric estimation.

[0079] The method was described in the case of an additive object: Image = Objet + Contenu It can be applied to the case of a multiplicative object: Image = Objet × Contenu For example, the anti-scatter grid is a multiplicative object; the frequency value and its amplitude then depend on the patient. To move from the additive case to the multiplicative case, simply calculate the logarithm of the image (i.e., of the y-vector): Image = log Objet + log Contenu Using the additive method described previously to obtain the estimation vector β , it is possible to estimate log( Object ) and to deduce from this: Objet = 10 log Objet

[0080] The present invention can be used in any process requiring improved image quality. It applies, among other things, to the field of medical imaging (X-ray radiography, CT scans, MRI, etc.), as well as to all other types of imagers (X-ray, visible, infrared).

[0081] Practical applications include, for example: The elimination of a disturbance, distinct from noise, that degrades visibility and can lead to misdiagnosis in medical imaging, as illustrated by the figure 1 The invention advantageously allows noise to be left in the digital image, which may be desirable for some medical imaging practitioners; the characterization of objects for technical or medical purposes (for example, the size of a lump in the patient's body); and noise reduction: if the first embodiment is used, it is possible to obtain an estimate of the object, its content, and the noise (knowing the estimated noise variance σ²). It is then possible to reconstruct the signal using only the object and content components. The same operation can be performed without an object, estimating only the content and the noise.

[0082] The method according to the invention is thus capable of correctly estimating each component of dictionaries, even though these dictionaries contain numerous components. The use of an iterative process makes it possible to estimate the components one by one, filtering each time to best refine the previously estimated components.

Claims

1. Method for estimating parameters of an object to be estimated in a digital image representing at least one real imaged content, and the object to be estimated able to interfere with the visualisation of the real imaged content, comprising at least: a) one initial step comprising the production of a dictionary of content components and the production of a dictionary of object components, the content components (D) and the object components having the same dimensions as the digital image (y); b) one step of jointly establishing the magnitude of each of the content components of the dictionary of component contents and of object components of the dictionary of object components present in the digital image; c) one step of establishing the value of at least one parameter characterising the object to be estimated from the magnitude of each of the object components in an estimation vector containing the magnitudes of each component, the parameter being selected from a set comprising the frequency, shape, size and location of the object, and comprising the substeps of: - establishing a subset of object components of greater magnitude; - establishing the value of said parameter as a function of a linear combination of the object components of said subset; the joint establishment step b) being carried out by iterative establishment of an estimation vector (β) comprising the different magnitude values of each of the components of the content dictionary and the object dictionary present in the image, until convergence of the estimation vector (β); the joint establishment step b) comprising the iterative establishment of an estimation of the noise variance (σ2), established from the estimation vector (β) and a correlation matrix N of the noise present in the digital image, until convergence of the estimation of the noise variance (σ2); the noise variance (σ2) being defined by: σ 2 = y − Dβ H N − 1 y − Dβ + a M + b where N represents the noise correlation matrix, M represents the number of pixels of the image, and a and b are parameters established as a function of the nature of the noise.

2. Parametric estimation method according to any one of the preceding claims, the production of the content dictionary comprising the generation of a set of two-dimensional sinusoidal signal images.

3. Parametric estimation method according to claim 1, the production of the content dictionary comprising the generation of a set of images representative of the modulation transfer function of the sensor forming the digital image.

4. Parametric estimation method according to claim 1, the production of the content dictionary being carried out from a machine learning algorithm.

5. Parametric estimation method according to claim 4, the machine learning algorithm being a K-SVD algorithm.

6. Parametric estimation method according to any one of the preceding claims, the production of a dictionary of object components comprising the generation of a set of images, each having a different value of the parameter.

7. Parametric estimation method according to any one of the preceding claims, the convergence of the estimation vector (β) being considered reached when the norm of the estimation vector (β) evolves below a pre-established threshold between two successive iterations.

8. Parametric estimation method according to any one of claims 1 to 6, the convergence of the estimation vector (β) being considered to be reached after a pre-established number of iterations.

9. Parametric estimation method according to any one of the preceding claims, comprising, at the end of each iteration, a step of modifying the components of the dictionary of content components and / or dictionary of object components to minimise a projection error (ε) defined by: ε = y − Dβ 2 , where y is the vector representing the digital image, and D is the matrix representing the components of the dictionary to be modified.

10. Parametric estimation method according to any one of the preceding claims, the noise correlation matrix N being established by prior knowledge of the digital image acquisition detector.

11. Parametric estimation method according to any one of the preceding claims, comprising a step d) of reconstructing the content image by linear combination of all the content components.

12. Parametric estimation method according to any one of the preceding claims, the digital image being acquired by X-ray imaging.

13. Method for removing an object present in a digital image representing real-world image content, comprising the steps of: - estimating the value of at least one parameter of the object and the amplitudes for multiple components established according to the parametric estimation method according to any one of the preceding claims; - estimating the object from the value of the parameter characterising the object and its magnitude; pixel-by-pixel subtraction of the estimated object in the digital image.