Method and device for processing multidimensional microscopy data for localizing dislocations in a sample

Abstract

Method and device for processing multidimensional microscopy data for locating dislocations in a sample. This method is applied to an N-dimensional tile of input data representative of a part of the observed sample and comprises, for each homogeneous component, obtaining a super-resolved droplet tile, comprising droplets representative of an atomic structure of a part of the observed sample, followed by at least one of a first treatment (46) of applying spectral filtering to said super-resolved droplet tile, and / or a second treatment (48) comprising calculating a super-resolved diffraction pattern of the super-resolved droplet tile, and / or a third treatment (50) comprising calculating a structure tensor associated with at least a subset of points of the super-resolved droplet tile.An enriched signature tile is formed from the output data of at least one of the first, second, and third processing steps, followed by the detection and localization (54) of at least one dislocation, by applying a machine learning classification method to said enriched signature tile. Figure for the abbreviation: Figure 2.

Description

Title of the invention: Method and device for processing multidimensional microscopy data for the localization of dislocations in a sample

[0001] The present invention relates to a method and a device for processing multidimensional microscopy data for locating dislocations in a sample of crystalline material, comprising an acquisition of at least one microscopy image of the sample, forming an input data block, the or each image of said block being representative of a part of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two.

[0002] The invention lies in the field of processing multidimensional microscopy data, obtained by observing samples composed of one or more materials, for the analysis of physical properties of their structures.

[0003] More particularly, the invention finds applications in the localization of dislocation-type defects observed in crystalline materials, for example the invention applies in the quality inspection in a production line of devices based on crystalline materials.

[0004] As is known, crystals are composed of a crystal lattice, defined by a Bravais lattice. More precisely, a crystal is defined as a solid whose diffraction pattern is essentially discrete. Indeed, the IUCR (International Union of Crystallography) defines crystal as follows: “By crystal we mean any solid having an essentially discrete diffraction pattern”. In this document, the definitions rigorously established by the IUCR are chosen to define crystal and amorphous material.

[0005] In practice, observed materials comprising one or more crystalline and / or amorphous species may exhibit deformations or irregularities which may induce stresses or malfunctions in devices using such materials, for example leakage currents in transistors.

[0006] In particular, "dislocation" type deformations correspond to a linear defect of a crystal obtained by sliding of a part of the crystal along a plane. A dislocation is characterized by a sliding vector, also called the Burgers vector of the dislocation. Known dislocations are "screw" dislocations, "corner" dislocations or mixed dislocations.

[0007] More generally, considering a data block in an N-dimensional space, the dislocation is defined by a sliding of a part of the data along a hyperplane.

[0008] It is useful to have high-performance tools for carrying out the inspection of materials of the aforementioned type, and detecting the possible presence of dislocations, in the calibration phase or in the technological optimization phase of various devices using the materials.

[0009] The analysis of physical properties of materials by processing spectral data (1D) and images (2D) obtained by microscopy has been developed for this purpose.

[0010] The images and spectra are for example obtained by high-resolution transmission electron microscopy (HRTEM and HRSTEM and 4D-STEM), X-ray fluorescence microscopy (HRSTEM-EDX), electron energy loss microscopy (HRSTEM-EELS and HRSTEM-VEELS), energy-filtered transmission electron microscopy (EF-HRTEM, EF-HRSTEM), atomic force microscopy (AFM), scanning tunneling microscopy (STM), atom probe tomography or electron tomography for example. These techniques are all considered as microscopies allowing the analysis of the structure of the sample, that is to say the distribution of chemical species. The microscopy images obtained are atomic resolution images, that is to say with a spatial resolution of less than or equal to 2 nanometers.

[0011] Spectral and image data sets, each representative of at least a portion of the observed sample, are obtained by microscopy, forming an N-dimensional data block (or multidimensional data), also called a "datacube", N being an integer greater than or equal to 2. This data block comprises drops, also called spots, or "spikes" or "blobs" in English, which stand out against a homogeneous background (for example, light drops or spots on a dark homogeneous background). These drops are representative of structural characteristics of the material of the observed sample, for example, an alignment of atoms along the observation direction. The data block acquired by microscopy may also contain noise.

[0012] For example, when observing crystals, drops arranged in a regular pattern represent the crystal lattice. In two dimensions, drops are also called spots. In one dimension, drops are points.

[0013] Mathematically, a drop is a simply connected component of a discrete topological space, in the sense of general topology. This means that any loop traced in a drop can be reduced by homotopy to a point. Physically, a drop represents, for example, the electrical signal produced by the pixels of a matrix detector following the impact of a particle (electron, photon, ion, fermions, boson, etc.). The particle is always much smaller than the pixel, so the drop can always be reduced to a point by homotopy, in accordance with the mathematical definition of the drop. There is therefore a precise agreement between the physical definition and the mathematical definition. gout theme.

[0014] In the state of the art, methods are known for measuring dislocations in materials observed by microscopy, using the Radon transformation. One of the difficulties is to locate dislocation-type defects with good precision.

[0015] There is therefore a need to improve the accuracy of localization of dislocations in observed samples of crystalline material, while preserving the observed sample.

[0016] To this end, the invention proposes, according to one aspect, a method for processing multidimensional microscopy data for the localization of dislocations in a sample of crystalline material, comprising an acquisition of at least one microscopy image of said sample forming an input data block, the or each image of said input data block being representative of a part of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two, each data item of said block corresponding to a point in the N-dimensional space. This method comprises steps of: - for at least one data block extracted from said input data block, called a component block, comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample, applying a transformation to the component block to obtain a block, called a super-resolved drop block, comprising drops representative of an atomic structure of a part of the observed sample, followed by at least one of the following three treatments: - applying a first processing to said super-resolved droplet block comprising an application of spectral filtering to said super-resolved droplet block, and / or - applying a second processing to said super-resolved droplet tile comprising a calculation of a super-resolved diffraction pattern of the super-resolved droplet tile, and / or - applying a third processing to said super-resolved droplet tile comprising a calculation of a structure tensor associated with at least a subset of points of the super-resolved droplet tile; - formation of an enriched signature block associated with the superresolved drop block, said enriched signature block combining, for each point of said drop block, output data from at least one of said first, second and third processing operations; - detection and localization of at least one dislocation, by applying a machine learning classification method to the enriched signature block.

[0017] Advantageously, the combination of a plurality of treatments makes it possible to obtain a block of signatures rich in information, which makes it possible to detect and locate precisely and unequivocally any type of dislocation.

[0018] This method of processing multidimensional microscopy data for the localization of dislocations according to the invention may also have one or more of the characteristics below, taken independently or in all technically conceivable combinations.

[0019] The machine learning classification method is either supervised, semi-supervised, or autonomous after prior learning on signature tiles calculated on simulated dislocations, making it possible to obtain a map of the dislocations corresponding to said drop tile.

[0020] The method further comprises a step of dividing the input data block into component blocks, each component block comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample.

[0021] Applying a transformation to obtain a super-resolved droplet tile comprising drops representative of an atomic structure of a portion of the observed sample implements a segmentation step using a machine learning algorithm, trained to classify the data of the super-resolved droplet tile into two classes, respectively a “background” class and a “droplet” class.

[0022] The method further comprises refining to obtain a refined superresolved droplet tile, by applying a refining spectral filtering mask, in which the refining spectral filtering mask is obtained from the maximum of the machine learning segmentation of the Fourier transform of the component tile normalized between 0 and 1 and an adjusted theoretical diffractogram normalized between 0 and 1.

[0023] The first treatment comprises steps of: - application of a Fourier transform on the super-resolved droplet block to obtain a frequency block, - spectral filtering by applying a masking window to the frequency block, said masking window corresponding to at least two crystal planes corresponding to two drops of the frequency space symmetrical with respect to the center of the Fourier transform of the super-resolved drops of the real space; - application of an inverse Fourier transform to obtain corresponding signature data.

[0024] The method further comprises a step of correcting artifacts introduced by the Former transform.

[0025] Spectral filtering is applied for a plurality of masking windows corresponding to distinct crystal planes.

[0026] The second treatment comprises steps, applied to said super-resolved droplet block, of:

[0027] - addition of a circumference on the periphery, according to each of the dimensions, of at least one zero value point, allowing to obtain an increased block,

[0028] - application of a Fourier transform on the augmented block,

[0029] - application of a refining spectral filtering mask to the result of the step previous to obtain a refined super-resolved diffractogram of the super-resolved droplet tile, forming a second signature tile.

[0030] The method further comprises, before the addition step, a drop erosion step, each drop of the super-resolved drop block having an initial size, so as to reduce said initial size of each drop, to obtain an eroded drop block.

[0031] The method further comprises, after the step of applying a filtering mask, a step of applying a theoretical convolution to the super-resolved diffractogram to obtain a convolved super-resolved diffractogram of each super-resolved component tile.

[0032] The third processing comprises, for at least a subset of neighboring points of said block of drops, steps of:

[0033] -calculation of a structure tensor as a function of the values ​​of directional gradients, each direction corresponding to a dimension of said space, in a neighborhood of said point,

[0034] -calculation of one or more characteristic values ​​of the structure tensor from among: gradient tensors in each direction, eigenvalue tensors of the structure tensor, an energy tensor, a coherence tensor, an orientation tensor, a Harris index tensor, a Harris-Laplace index tensor,

[0035] - formation of a signature tensor associated with said point, the signature tensor having as components the said characteristic values ​​of the structure tensor;

[0036] -obtaining a third signature tile corresponding to said super-resolved drop tile, said third signature tile bringing together the signature tensors associated with the points of said super-resolved drop tile.

[0037] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a method for processing multidimensional microscopy data for the localization of dislocations as defined above.

[0038] The invention also relates to a device for processing multidimensional microscopy data for locating dislocations in a sample of crystalline material, configured to acquire at least one microscopy image of said sample forming an input data block, the or each image of said input data block being representative of a part of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two, each data item of said block corresponding to a point in the N-dimensional space. This device is configured to implement: - for at least one data block extracted from said input data block, called a component block, comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample, a module for applying a transformation to the component block to obtain a block, called a super-resolved drop block, comprising drops representative of an atomic structure of a part of the observed sample, and for implementing at least one of the following three modules: - a module for applying a first processing to said super-resolved droplet block comprising an application of spectral filtering to said super-resolved droplet block, and / or - a module applying a second processing to said super-resolved droplet block comprising a calculation of a super-resolved diffraction pattern of the super-resolved droplet block, and / or - a module for applying a third processing to said super-resolved droplet tile comprising a calculation of a structure tensor associated with each point of at least one subset of points of the super-resolved droplet tile

[0039] as well as:

[0040] - a module for forming an enriched signature block associated with the drop block super-resolved, said enriched signature tile bringing together, for each point of said super-resolved drop tile, output data from at least one of said first, second and third processing;

[0041] - a module for detecting and locating at least one dislocation, by application from a machine learning classification method to an enriched signature block.

[0042] The invention will appear more clearly on reading the description which follows, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0043] [Fig-1] [Fig.l] is a block diagram of a multi-channel data inspection system microscopy data comprising a device for processing multidimensional microscopy data according to one embodiment;

[0044] [Fig.2] [Fig.2] is a block diagram of a multi-data processing method microscopy measurements for locating dislocations according to one embodiment;

[0045] [Fig.3] [Fig.3] is a first example of division into homogeneous components;

[0046] [Fig.4] [Fig.4] is a second example of division into components homogeneous;

[0047] [Fig.5] [Fig.5] illustrates a processing to obtain an image comprising simply connected components for a crystal;

[0048] [Fig.6] [Fig.6] is a synopsis of the main steps of a first treatment implemented according to one embodiment;

[0049] [Fig.7] [Fig.7] illustrates by an example the results of intermediate stages of the first treatment;

[0050] [Fig.8] [Fig.8] illustrates the presence of artifacts due to the Fourier transform and the correction of artifacts;

[0051] [Fig.9] [Fig.9] is a synopsis of the main stages of a second treatment implemented according to an embodiment;

[0052] [Fig. 10] [Fig. 10] illustrates by a crystal image and the associated diffractogram obtained by the second treatment;

[0053] [Fig. 11] [Fig. 11] is a synopsis of the main steps of a third treatment implemented according to one embodiment;

[0054] [Fig. 12] [Fig. 12] illustrates examples of the signature values ​​obtained by the third processing.

[0055] [Fig.l] schematically illustrates a system for detecting dislocations 2 in a sample of crystalline material, comprising one or more types of crystals, from images and spectra acquired by an electron microscope 4.

[0056] In one embodiment, the microscope 4 is a transmission electron microscope (TEM) which makes it possible to acquire images of a sample composed of one or more materials, for example comprising crystalline materials.

[0057] The electron microscope 4 makes it possible to simultaneously acquire electron diffraction images, EELS spectra (for “Electron Energy Loss Spectroscopy”), EDX spectra (for “Energy Dispersive X-Ray Analysis”) of the signals from different sensors (BF for “Bright field”, DF for “Dark field”, DPC for “Differential phase contrast” or any other suitable or customized sensor), for each point of the sample in scanning microscopy mode called STEM mode (for “Scanning Transmission Electron Microscope”). Another possible acquisition mode is the TEM mode which makes it possible to obtain a global image in an instant without having to scan the electron beam sequentially.

[0058] In variants, another type of microscope 4 is used, for example an atomic force microscope or a scanning tunneling microscope.

[0059] Each acquired spectrum or image is represented in the form of a digital image, comprising points or pixels, each image being representative of at least part of the observed sample.

[0060] All the acquired spectra and images form an N-dimensional data block, N being a natural integer greater than or equal to 2. Such a data block is also called a “datacube”.

[0061] In one embodiment, multiple spectra or images are acquired over time, showing an evolution of the sample during analysis. In this embodiment, time is an additional dimension of the data block.

[0062] Additionally, another dimension of the data block is the microscope focus, electron energy, an angle (of the sample, electron collection, electron convergence), or any other microscope adjustment parameter that may vary in a controlled manner during the measurement.

[0063] Incomplete data can possibly be extrapolated from neighboring values ​​if necessary, for example to correct any imperfections encountered during data acquisition. As a general rule, measurements at the atomic scale are very often marred by defects because a simple acoustic or electronic noise can sometimes disturb them. Conventional reconstruction techniques published in the literature therefore make it possible to correct any measurement artifacts.

[0064] In a multidimensional data block of a crystal structure sample, the crystal lattice forms a homogeneous network of observable drops arranged in a regular pattern that represents the regular crystal lattice.

[0065] Structural defects or irregularities of such a structure also form drops, but which are for example irregularly arranged.

[0066] A dislocation-type defect corresponds to a sliding of a part of the crystal along a plane. In N dimensions, this corresponds to a sliding of a part of the data along a hyperplane.

[0067] As already indicated above, mathematically, a drop is a simply connected component of a discrete topological space, in the sense of general topology.

[0068] The values ​​of the pixels belonging to a drop are distinguished from an image background.

[0069] A block of multidimensional data of a sample of an observed physical structure, e.g. a crystalline material, is transmitted to a device 6 for processing multidimensional microscopy data for the localization of dislocations in the observed sample.

[0070] For example, the transmission is carried out by a wired connection or by a wireless connection (optical, radio, or other).

[0071] The processing device 6 is, in one embodiment, a programmable electronic device, e.g. a computer.

[0072] The device 6 comprises a processor 8 (CPU or GPU) associated with an electronic memory 10. Optionally, the device 6 comprises a human-machine interface (12), comprising in particular a data display device (not shown). In addition, the device 6 comprises or is connected to a storage memory 14. The elements 8, 10, 12, 14 of the device 6 are adapted to communicate via a communication bus 16.

[0073] The processor 8 is configured to execute modules 18, 20, 22, 24, 26, 28, 30, 32 stored in the electronic memory 10, to implement a method for processing multidimensional microscopy data for the localization of dislocations, as described below.

[0074] Module 18 is a module for obtaining a block of multidimensional data to be processed, configured to obtain a block of data comprising at least one input image from an image obtained by microscopy, called a microscopy image.

[0075] For example, the obtaining module 18 is configured to obtain the data block to be processed (or input data block) from an electronic memory, where this data has been stored after acquisition.

[0076] Module 20 is a module for dividing the input data block into component blocks, each component comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample.

[0077] The component tiles may have different sizes than the input data tile, or may be the same size as the input data tile.

[0078] Size here refers to the number of points per dimension, according to each of the dimensions.

[0079] For example, the module 20 cuts the data block into one or more blocks of either crystalline materials, each crystalline material being characterized by a crystalline mesh, or amorphous materials or a mixture of phases of such materials which have a homogeneous appearance. Mathematically, the structural homogeneity criterion can be precisely quantified by appropriate image processing described in the literature.

[0080] Module 22 is a module for applying a transformation making it possible to obtain, from a component block, a super-resolved block of drops, comprising drops representative of an atomic structure of a part of the observed sample.

[0081] For example, the transformation module 22 implements filtering, normalization and then either thresholding or segmentation by machine learning as described in more detail below.

[0082] Modules 24, 26 and 28 implement first, second and third processing operations applied to the super-resolved droplet block, making it possible to obtain at output feature tiles, also called signature tiles, corresponding to the input data tile.

[0083] According to variants, only one of the modules 24, 26, 28 is implemented, or a combination of two of the three modules, or all three modules are implemented.

[0084] The module 24 for applying a first processing to the droplet block performs a Fourier transform on said super-resolved droplet block, and spectral filtering as detailed below. At the output of the first processing, a first block of signatures is obtained.

[0085] The module 26 for applying a second processing to the super-resolved droplet block calculates a super-resolved diffraction diagram. At the output of the second processing, a second block of signatures is obtained.

[0086] The module 28 for applying a third processing to the super-resolved droplet block implements a calculation of a structure tensor associated with each point of at least one subset of points of said super-resolved droplet block. At the output of the third processing, a third signature block is obtained.

[0087] The device also implements a module 30 for forming an enriched signature block associated with the super-resolved drop block, said enriched signature block combining, for each point of said drop block, at least part of the data of the first, and / or second and / or third signature blocks, obtained at the output of said first, second and third processing operations respectively.

[0088] Module 32 is a module for detecting and locating a dislocation or several dislocations by applying a machine learning classification method to the enriched signature block, previously trained, for example on signature blocks calculated on simulated dislocations, making it possible to obtain a map of the dislocations corresponding to said superresolved drop block.

[0089] The map 35 obtained is stored and, for example, is then displayed on the human-machine interface 12 for an operator or transmitted to a remote device.

[0090] The parameters 34 of the classification method, obtained by machine learning, are stored, for example in the electronic storage memory 14.

[0091] Machine learning is implemented in a preliminary learning step, by an electronic computing device of the same type, on simulated data blocks, the dislocations of which are also simulated and therefore the theoretical characteristics of which are known.

[0092] The enriched signature block is calculated for each simulated data block of a learning base, so as to train the classification method, and in particular particular obtain by machine learning the values ​​of the parameters of the classification method to be applied.

[0093] For example, machine learning is performed with a fast random forest type algorithm, described in the article by Léo Breiman (2001), “Random Forests”, published in Machine Learning. 45(1):5-32).

[0094] Alternatively, other machine learning classification algorithms may be implemented, for example Bayesian type, or function-based, or rule-based, or a combination of such algorithms, or based on the entire range of tools used in machine learning classification.

[0095] According to another variant, the chosen classification method implements a neural network, trained by supervised learning, for example a convolutional neural network CNN (for “Convolutional Neural Networks”). Another variant uses “deep learning” type learning.

[0096] In one embodiment, the modules 18, 20, 22, 24, 26, 28, 30, 32 are each produced in the form of software, and form a computer program, comprising software instructions which, when executed by a computer, implement a method for processing multidimensional microscopy data for the localization of dislocations as described in more detail below.

[0097] In a variant not shown, the modules 18, 20, 22, 24, 26, 28, 30, 32 are each produced in the form of a programmable logic component, such as an FPGA (Field Programmable Gate Array), a GPU (graphics processor) or a GPGPU (General-purpose processing on graphics processing), or in the form of a dedicated integrated circuit, such as an ASIC (Application Specific Integrated Circuit).

[0098] The software for processing multidimensional microscopy data for locating dislocations is further capable of being recorded, in the form of a computer program comprising software instructions, on a computer-readable medium, not shown. The computer-readable medium is, for example, a medium capable of storing the electronic instructions and of being coupled to a bus of a computer system. By way of example, the readable medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, SD, NVRAM, RRAM, PCRAM, 2D NAND, 3D NAND, SLC NAND, MLC NAND, TLC NAND, V-NAND, QLC), a magnetic card or an optical card, an SSD disk.

[0099] [Fig.2] is a synopsis of the main steps of a method for processing multidimensional microscopy data for the localization of dislocations in a sample observed by microscopy according to one embodiment.

[0100] The method comprises a step 40 of acquiring a block of data (of images and of spectra) of N-dimensional microscopy, the data being representative of a sample of crystalline material observed, N being an integer greater than or equal to 2.

[0101] Each image data has a digital value, or intensity, represented on a chosen number of bits, for example 8, 16, 32 or 64.

[0102] In the special case where N is equal to two, each multidimensional data tile is a two-dimensional (2D) image.

[0103] The term image will be used hereinafter to designate 2D data blocks; it being understood that the methods described apply analogously for N dimensions, with N greater than or equal to 2.

[0104] The illustrations given as examples are illustrations of 2-dimensional digital images, from simulated HRSTEM-HAADF (High Resolution Scanning Transmission Electron Microscope - High-angle annular dark-field) images, the intensity of each point of the image having a value corresponding to the signal measured by a HAADF detector.

[0105] Of course, the method of the invention is not limited to this type of image acquisition by microscopy, and applies with any variant of data acquisition by a microscope 4, as indicated above, provided that the signal is calculably linked to the position of the atoms in the material, that is to say when the adjustment and magnification of the microscope are correct.

[0106] The image and spectra data block obtained in the acquisition step is called the input data block.

[0107] The method then optionally comprises a step 42 of cutting (or partitioning or segmenting by regions) the input data block into component blocks (also called homogeneous components), each component comprising homogeneous data according to a homogeneity criterion relating to at least one subset of the input data.

[0108] The homogeneity criterion is for example a homogeneity of structure, with reference to repetitive patterns, the pattern corresponding to a type of structure for a given material.

[0109] For example, observing a crystal of given thickness will make it possible to obtain an image consisting of a periodic pattern that is easy to identify (pattern 43 of [Fig.3]). When the diffractogram of this image is not made up of drops but of annular shapes, then the component is classically described as amorphous (pattern 45 of [Fig.3]), in accordance with the usual meaning given by the IUCR. Indeed, a ring is not a drop because it contains a hole, so it is not simply connected. An amorphous material classically gives diffraction rings.

[0110] There are a wide variety of methods for classification and segmentation into regions, known to those skilled in the art. The aim is to group the pixels into regions, which constitute a tiling or partition of the image, or more generally of the data block, according to a chosen criterion. For example, the peak-pair distribution function or radial distribution function and the distribution of polar angles of the segments that connect the nearest neighbors make it possible to quantify the distribution of drops in the data blocks. To segment the image, a division into regions that have a distribution of distances and angles between near neighbors that is homogeneous, that is to say in a chosen interval, is carried out. These regions are the homogeneous components. More generally, the similarity between 2 images or 2 regions of an image is classically obtained when the distance that separates them is less than a chosen numerical threshold.

[0111] In other words, each distinct component tile is a part of the tile of given which corresponds to a distinct structure, characterized by a given distribution of distances and polar angles between neighboring drops. Within a component block, all diffractograms averaged over a region containing at least G neighboring drops are similar, for example G=16. A component is an area which exhibits structural homogeneity in the sense of crystallography. The homogeneity threshold is chosen according to each particular case.

[0112] In general, preferably, at least 3 to the power of N neighboring drops are used to define the size of a representative domain of the structure, where N represents the dimension of the droplet tile.

[0113] For an image, at least G=9 neighboring drops are therefore required to define a domain representative of the structure. A component is said to be homogeneous when the Fourier transforms of all its representative domains are all similar. As a reminder, there are many classical methods for mathematically quantifying the similarity between two images. The Fourier transform gives an image of the structural order, similar to the crystalline order revealed by an electron diffraction image.

[0114] In the case of a multilayer material, a component corresponds to a homogeneous layer of the stack of layers. In the case of an inclusion in a matrix, one component can be chosen for the inclusion and another component for the matrix.

[0115] In an example application, the cutting step 42 makes it possible to obtain a block of crystalline material and a block of amorphous material, for example a separation between pure silicon and an amorphous layer.

[0116] In one embodiment, the cutting of step 42 implements automated segmentation (or classification) by machine learning.

[0117] Alternatively, the cutting is carried out manually by an operator.

[0118] The cutting step 42 is optional, the input data block being able to represent a homogeneous material. In this case, in the rest of the method, the block of input data is the component block mentioned later.

[0119] Figures 3 and 4 respectively illustrate divisions of an input image (two-dimensional input data block) into sub-images comprising homogeneous components. These sub-images have a size less than or equal to that of the input data block. The simplest approach is to keep the same size and replace the removed areas with a black background, i.e. zero signal intensity.

[0120] In [Fig.3], an input image 41 is cut into an image 43 of crystalline material (eg silicon) and an image 45 of amorphous material (eg a-SiwCxOyHz).

[0121] In [Fig.4], an input image 47 is cut into three images 49, 51 and 53 each corresponding to a visibly differentiated structure.

[0122] According to variants, the homogeneity criterion is a criterion of homogeneity of data intensity values ​​and / or dielectric permittivity and / or electromagnetic properties, for example measured in situ during the acquisition of the data block. The VEELS measurements make it possible in particular to obtain the local dielectric permittivity, as shown in the literature.

[0123] Returning to [Fig.2], the cutting step 42 is followed by a step 44 of applying a numerical transformation to obtain a superresolved droplet block comprising drops representative of an atomic structure of at least part of the observed sample.

[0124] Advantageously, the use of a super-resolved droplet block makes the result more robust and less erratic.

[0125] By definition, for the drops to be considered super-resolved, the input data must first be denoised, preferably using an HBSG filter, then possibly the number of points must be increased, and then the data must be renormalized.

[0126] The HBSG filter is described in patent FR3118256 “IMAGE PROCESSING FILTER FOR LOCALIZING DROPS IN A MULTIDIMENSIONAL IMAGE AND ASSOCIATED IMAGE PROCESSING METHOD”. It is represented by image 57 of [Fig.5]. This filtering, for example, uses here a square HBSG kernel whose size (in this particular case 21 pixels x 21 pixels) is close to the median distance between neighboring drops, with smoothing of order 2 or 3.

[0127] To increase the number of points in the input tile in order to increase the resolution, the super-resolved image (and more generally the super-resolved tile) is generated either by interpolation or using a reconstruction technique or a digital super-resolution technique such as the total variation (TV) method. This method consists of regulating the creation of the new points on the total variation of the contrasts of the original image. This approach makes it possible to preserve the overall variations in contrast while preserving the edges of the drops of the image.

[0128] The total variation method is notably described in the publication LI Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms”, Physica D, 60 (1), 259-268 (1992). By applying this method with 100 iterations, we obtain image 59 of [Fig.5].

[0129] After this resolution increase step, the previous HBSG filtering (57) is possibly reapplied to erase any digital resolution increase artifacts.

[0130] Finally, the renormalization step consists of dividing the image by a local variance, obtained by calculating the square root of the convolution by a Gaussian filter of the square of the image. Generally, the size of the Gaussian filter used to calculate the local variance is ideally close to the median of the distances between neighboring drops. In [Fig.5], this normalization step makes it possible to transform image 59 into image 61.

[0131] The last transformation of step 44 consists of segmenting superresolved drops, in order to replace the neighborhood of the drops with a uniformly black background.

[0132] To carry out this segmentation, in a first embodiment, the method implements a machine learning algorithm to distinguish two classes, respectively the “drop” class (or “spots” in English) and the “background” class. The learning is carried out in parallel and in a differentiated manner on data blocks corresponding to theoretically known dislocations, for each crystallographic orientation and for each microscope setting. The algorithm for dividing into 2 differentiated classes is trained by supervised, semi-supervised or autonomous learning after learning on a very large number of particular cases. Another variant uses “deep learning”.

[0133] For example, machine learning is performed with a fast random forest algorithm, described in the article by Léo Breiman (2001), “Random Forests,” published in Machine Learning. 45(1):5-32).

[0134] Alternatively, other machine learning classification algorithms may be implemented, for example of the Bayesian type, or based on functions, or based on rules, or a combination of such algorithms, or based on the entire range of tools used in machine learning classification, in particular “deep learning”.

[0135] Optionally, the two-class classification is followed by watershed segmentation, which is a classical morphological segmentation method. J.B.T.M. Roerdink and A. Meijster. “The watershed transform: definitions, algorithms, and parallelization strategies,” in Mathematical Morphology, vol. 41 (2000), pp. 187-228. This image segmentation method from mathematical morphology considers the grayscale image as a relief topographical area whose flooding is simulated.

[0136] Typically, drops represent less than 20% of the image, and their size is less than 1 nm, after calibration to the actual sizes in the sample. Simulations of the theoretical images make it possible to determine the expected distribution of drops in the image in order to verify that the segmentation is correct. A segmentation is correct if the number and position of drops approach the expected theoretical values ​​beyond a given uncertainty threshold. Optionally, convolution filtering to fill in any artifacts within the resulting drops is also implemented.

[0137] For example, an HBSG filter as described in patent application FR3118256 “IMAGE PROCESSING FILTER FOR LOCALIZING DROPS IN A MULTIDIMENSIONAL IMAGE AND ASSOCIATED IMAGE PROCESSING METHOD” is implemented, and / or a “fill holes” type algorithm (K.-J. Oh, S. Yea, and Y.-S. Ho, “Hole filling method using depth based inpainting for view synthesis in free viewpoint television and 3-D video,” in Proc. Picture Coding Symp., Chicago, IL, May 2009.). Again, the objective is to approach the ideal theoretical shape, which is free of holes.

[0138] A second segmentation mode consists, for example, of applying an adaptive Bemsen thresholding described in "Dynamic Thresholding of Grey-Level Images", Proc. of the 8th Int. Conf. on Pattern Recognition, 1986, with a radius close to the half-median distance between neighboring drops (here 10 pixels). It is the atomistic simulation of the "perfect" theoretical images which makes it possible to verify that the image processing parameters make it possible to converge towards the perfect theoretical form, for each crystal, each orientation and each set of instrumental parameters.

[0139] Then, to get even closer to the theory, additional refining steps can be applied. These refining steps consist of applying a spectral windowing mask, called a refining spectral filtering mask, comprising real numerical values ​​between 0 and 1. This mask has the same dimension as the Fourier transform of the component block and by multiplying each term of the mask by its counterpart in the Fourier transform of the component block, only the points of the block with non-zero frequencies of the mask are retained, the other points being set to zero. The mask therefore has the function of removing the frequency contributions of zero value, for each point of the Fourier transform of the component block.

[0140] The adjusted theoretical diffractogram can be used as a spectral filtering mask to converge more precisely to the theoretical droplet distribution. The adjustment is obtained by varying the crystallographic parameters (a,b,c, alpha, beta, gamma) of the theoretical diffractogram so that it best superimposes on the Fourier transform of the component. This method allows to correct any distortion artifacts that may occur during data acquisition. The spectral filtering mask can also be obtained by taking the maximum between (A) the adjusted theoretical diffractogram normalized between 0 and 1 and (B) the mask obtained by machine learning segmentation of the Fourier transform of the component after normalization between 0 and 1. This last method allows not to delete the frequency information coming from the defects. By definition, this set of steps allows to obtain the refined super-resolved droplet tile, which allows to greatly improve the quality of the processing during the following steps. This refinement provides the maximum resolution and precision.

[0141] [Fig.5] illustrates, on examples of 2D digital images, the result of step 44 in its first embodiment described above: the processing is applied to a homogeneous component image 55 (experimental image of 193x193 pixels), to which the HBSG kernel referenced 57 is applied; image 59 is obtained by convolution filtering with an HBSG kernel, image 61 is the image obtained by local normalization from image 59 and image 63 (super-resolved drop image) is the result of the segmentation of image 61 which shows the drops (represented by light spots) which stand out from the uniform background of zero light intensity.

[0142] Finally, step 44 makes it possible to obtain a super-resolved droplet tile for each component tile. The refined super-resolved component gets even closer to the theory if necessary.

[0143] Step 44 is applied to at least one of the component tiles obtained at the end of the cutting step 42.

[0144] Preferably, step 44 and the following steps are repeated, in a similar manner, for each component tile obtained in step 42.

[0145] In the following, the treatments applied to a super-resolved, and possibly refined, droplet block are described, it being understood that these treatments are applied in an analogous manner to each super-resolved, and possibly refined, droplet block obtained by the treatments 42, 44.

[0146] The super-resolved droplet tile is then processed by applying at least one of a plurality of processes, respectively a first process 46, and / or a second process 48 and / or a third process 50, to obtain first, and / or second and / or third signature tiles, comprising characteristics of the processed super-resolved droplet tile.

[0147] The first processing 46 implements at least one spectral filtering applied to the super-resolved (possibly refined) droplet tile obtained in step 44, as described in more detail, according to one embodiment, with reference to [Fig.6].

[0148] The first processing 46 comprises a step 60 of applying a transform of Fourier, on the droplet block to obtain a frequency block. The domain of representation of the result of the Fourier transform is called the spectral domain.

[0149] The expression "Fourier transform" means the Fourier transform and all its variants and generalizations to hyperspaces, in particular the Hartley transform, the discrete Hartley transform (DHT), the fast Fourier transform, the discrete Fourier transform, the generalized discrete Fourier transform, the short-time Fourier transform (STFT), the fractional Fourier transform (FRFT), the Chirplet transform, the Hankel transform, the Fourier-Bros-Lagolnitzer transform, the canonical linear transform, the discrete-time Fourier transform (DTFT), the discrete-space Fourier transform (DSFT), the Z transform, the modified discrete cosine transform (MDCT),and the Fourier transform of finite groups. ,

[0150] Optionally, a black peripheral frame, i.e. zero-value points (zero-padding) is added around the edge of the droplet pad to increase the spatial resolution after Fourier transform.

[0151] The first processing 46 then comprises a spectral filtering 62, comprising the application of a predetermined windowing mask, in the spectral domain, corresponding to a sought spectral signature.

[0152] The filtering 62 comprises the application of a windowing mask comprising real digital values ​​between 0 and 1. This mask has the same dimension as the Fourier transform of the block of the component and by multiplying each term of the mask by its counterpart in the Fourier transform of the component only the points of the block of the non-zero frequencies of the mask are retained, the other points being set to zero. The mask therefore has the function of removing the frequency contributions of zero value, for each point of the Fourier transform of the component.

[0153] The determination of the mask corresponding to a given crystal plane is determined on blocks representing theoretical networks.

[0154] For example, the mask to be applied, corresponding to a given material, is obtained, in one embodiment, by simulation of the diffraction diagram corresponding to the observed crystal, for example

[110] oriented silicon, and adjustment of the crystalline parameters of lengths (a,b,c) and angles (a,[3,y) defining the crystal mesh to best correspond to the spots obtained in the frequency block obtained by step 60, which conventionally represents a good approximation of the diffraction diagram of the sample. The spots observed in the frequency block are hence called digital diffraction spots, just like the physical diffraction spots classically observed in electron diffraction patterns. In higher-dimensional hyperspaces, diffraction balls are numerically obtained from the data block representative of the sample.

[0155] For example, by performing the statistics of all the half-height widths of the diffraction spots, we choose a diameter equal to 4 times the largest half-height width obtained to define the diameter of the spots of the windowing mask. In a space of dimension N strictly greater than 2 (hyperspace), the largest diffraction ball of the frequency block is selected, then represented in hyperspherical coordinates from its center. The chosen diameter is equal to four times the half-height width of the intensity averaged over all the angles of the hyperspace. More generally, atomistic simulation makes it possible to choose the best diameter for each particular case, in order to retain only the useful spectral information and reject the frequencies which correspond to artifacts such as noise.

[0156] The windowing mask obtained then comprises patterns (balls) which have a first chosen diameter. An example of mask 71 is illustrated in [Fig.7], described below. This mask consisting of 2 diffraction spots symmetrical with respect to the center makes it possible to select the 2 crystalline planes corresponding to the spots and which are represented in figure 73 after the inverse Fourier transform. Each pair of spots symmetrical with respect to the center, therefore each family of selected crystalline planes, can be used to generate a spectral signature of the defect, for example of the dislocation, to be mapped.

[0157] The spectral filtering step 62 is followed by a step 64 of applying an inverse Fourier transform to the filtered frequency block, to obtain corresponding signature data.

[0158] The first processing 46 further comprises a subsequent step 66 of correcting artifacts introduced by the Fourier transform. Indeed, it is known that the Fourier transform is likely to add artifacts, for example a periodization or a spreading of the signal.

[0159] The artifact correction 66 implements the application of spatial filtering, which makes it possible to remove the artifacts, thus correcting the signature data obtained by the first processing. An automatic processing consists of segmenting the image of drops (eg image 63 in [Fig.5]) for example with the Bernsen method or by machine learning, then replacing the diameter of each drop by the median distance between neighboring drops. This masking makes it possible to retain only the information centered in the vicinity of the initial drops, within a radius equal for example to half the median distance between neighboring drops. The spreading effects beyond this limit are eliminated by this processing. The radius makes it possible to define the level of spreading retained in the vicinity of each drop.

[0160] In embodiments, the steps of the first processing are repeated for a plurality of windowing masks corresponding to the chosen crystal planes. It is useful to choose at least two diffraction balls symmetrical with respect to the center of the diffractogram to construct the windowing mask of the crystal planes. The more different crystal planes (and therefore diffraction balls) are chosen, the more a range of dislocation signatures is available, in order to identify them as precisely as possible.

[0161] Theoretical simulations make it possible to reveal which crystal planes are most representative of each dislocation. Indeed, certain dislocations sometimes have almost no impact on certain families of planes but deform other planes in a very obvious way. Theoretical simulation therefore guides the choice of crystal planes to be selected in order to better decide on their identification.

[0162] The data resulting from the first processing 46, for each windowing mask, are stored and grouped, so as to form a first block of signatures corresponding to the processed super-resolved block of drops.

[0163] As an example, [Fig.7] illustrates an input image 65, corresponding to a 2D image of a sample of Silicon Si

[110] having a central “screw” type dislocation, and image 67 (super-resolved droplet paving) obtained at the end of step 44 from the input image 65. The result of the application of the Fourier transform is illustrated by image 69 (representation in the spectral domain). The spectral windowing mask 71 corresponds to the crystal planes (-1-11) and (11-1) of the

[110] oriented Silicon. The result obtained after filtering the image 71 by the spectral windowing mask 71 and inverse Fourier transform is image 73 which makes it possible to visualize the planes (-1-11) and (11-1) of the sample. It should be noted that in the example in [Fig.7], the contrast and brightness are adjusted to the maximum, in order to better visualize the crystalline planes.

[0164] The artifacts due to the Fourier transform and their correction are illustrated in an example in [Fig.8].

[0165] In [Fig.8] are illustrated an input image 75, the corresponding image 77 obtained after applying spectral filtering and inverse Fourier transform on the corresponding super-resolved drop image.

[0166] As can be seen in [Fig.8], image 77 includes a border of artifacts. The corresponding spatial mask 79 is applied in correction step 66, which makes it possible to obtain the first corrected signature tile 81, also represented as an image in this example.

[0167] More generally, the size of the first signature block may differ from the size of the processed input data block, provided that the information is kept at a minimum useful, that is to say representative of the component block treated, and therefore of the material component observed.

[0168] The second processing 48 implements a calculation of a super-resolved diffraction diagram, also simply called a super-resolved diffractogram, of the droplet block obtained in step 44, as described in more detail, according to one embodiment, with reference to [Fig.9].

[0169] Preferably, the second processing implements the method for obtaining a super-resolved diffractogram described in patent FR3118254 “IMAGE PROCESSING METHOD FOR OBTAINING A SUPER-RESOLVED IMAGE AND A SUPER-RESOLVED DIFFRACTOGRAM.”

[0170] In one embodiment, starting from a super-resolved droplet tile, the second processing 48 comprises a step 70 of erosion of the drops, a step 72 of addition of a peripheral outline of zeros (or “zero-padding” in English), a step 74 of application of a Fourier transform and a step 76 of application of a segmentation.

[0171] The optional drop erosion step 70 involves reducing the initial size of each drop of the super-resolved drop tile. Each drop may be reduced by a predetermined ratio or so as to be formed of a predetermined number of points. After erosion, an eroded tile is obtained. In the extreme, each drop of the super-resolved drop tile may be reduced to one or a few points.

[0172] Step 72 of adding a peripheral contour consists of adding to the eroded tile a set of points of zero value at the periphery, in each of the N dimensions of the representation space, which makes it possible to obtain an augmented tile. The peripheral contour may in particular comprise at least a single point in each dimension. The addition of the peripheral contour has the effect of increasing the number of points of the eroded tile and therefore of increasing the spatial resolution of the diffractogram. It should be noted that in the absence of the erosion step 70, step 72 is applied directly to the treated super-resolved droplet tile.

[0173] Step 74 of applying a Fourier transform makes it possible to move from the spatial domain to the spectral domain. The implementation of step 74 is analogous to the implementation of step 60 described above.

[0174] The second processing 48 then comprises a segmentation step 76, during which the values ​​of the points of the diffractogram below a predetermined threshold are set to zero in order to retain only the brightest diffraction drops which stand out against a black background, in order to allow a direct and rapid comparison with the simulations of diffraction images. Alternatively, the segmentation 76 is carried out by a segmentation method trained by machine learning with the variants described previously.

[0175] The refinement method for approaching the theoretical diffractogram more precisely consists of multiplying the diffractogram of the droplet block by the theoretical diffractogram previously adjusted on the diffractogram of the droplet block. This step 78 is called "theoretical convolution". To carry out this adjustment, the 6 crystallographic parameters of length (a,b,c) and angles (a,[3,y) of the crystal lattice are explored to minimize the mean square deviation between the theoretical diffractogram and the numerical diffractogram of the superresolved droplet block. Ideally, the theory should overlap with the experiment, but in practice, the crystallographic parameters must be varied to obtain the best possible agreement. This optimized theoretical diffractogram is normalized in the range of real numbers between 0 and 1 to finally obtain a mask of the adjusted theoretical diffractogram of the droplet block.Masking allows to keep in the experimental result only the spectral information close to the expected theoretical values. The spectral filtering mask can also be obtained by taking the maximum between the adjusted theoretical diffractogram and the mask obtained by machine learning segmentation of the Fourier transform of the component block. This last method allows not to delete the frequency information coming from the possible defects. By definition, this set of steps allows to obtain an extremely precise refined super-resolved diffractogram which constitutes a new and distinct signature of each type of dislocation. The mask includes the theoretical information and the presence of all the defects, therefore the level of accuracy and precision of this method is maximal.

[0176] [Fig. 10] illustrates an image 83 corresponding to a data block comprising a “screw” type dislocation 85, and the corresponding super-resolved diffractogram 87, obtained after the second processing 48.

[0177] The output data of the second processing, i.e. the super-resolved diffractogram (possibly refined) obtained are stored, and form a second signature block corresponding to the super-resolved drop block.

[0178] The third processing 50 implements a calculation of characteristics from the structure tensor at each point of at least a subset of points of the super-resolved droplet tile obtained in step 44, as described in more detail, according to one embodiment, with reference to [Fig. 11].

[0179] The third processing comprises the calculation 80, at each point of the super-resolved droplet block, of a corresponding structure tensor, a function of the directional gradient values, each direction corresponding to a dimension of the N-dimensional representation space.

[0180] At a given point P, with coordinates Xp=(xl_p,x2_p,.. .,xN_p), the structure tensor is defined by the following formula, for N=2: [Math.l]

[0181] Where fx is the partial derivative along the first dimension, f is the partial derivative along the second dimension, and w() is an isotropic observation window, centered at P. The symbol denotes the convolution operation.

[0182] The observation window makes it possible to apply smoothing to neighboring points, in other words to take into account a neighborhood of the point P considered.

[0183] In 3 dimensions, the formula [MATH 1] becomes: [Math.2]

[0184] The generalization of the formula [MATH 2] to N dimensions, with N>3, is analogous, by calculating the field of local covariance matrices of the partial first derivatives in the corresponding hyperspaces.

[0185] Subsequently, the formulas are given for N=2, the generalization to N>2 being within the reach of those skilled in the art.

[0186] The terms of the matrix expressed by [MATH 1] are also noted: [Math.3] JW=\ 7 / \ T / \

[0187] With Ji2(p)=J2i(p) as it appears from [MATH 1]

[0188] For example, in one embodiment, the partial derivatives are obtained by calculating directional gradients, according to each of the dimensions of the space.

[0189] The structure tensor is defined by the matrix of [MATH 1], applied with the values ​​of the directional gradients in each direction.

[0190] The neighborhood of points taken into consideration is for example defined by a number No of neighboring points around the point P considered in each direction.

[0191] For example, No is chosen equal to the median distance between neighboring drops for a given crystal lattice structure.

[0192] The third processing 50 also comprises a step 82 of calculating one or more characteristic values ​​of the structure tensor.

[0193] In N dimensions, the characteristic values ​​of the structure tensor form tensors, among: tensors of the gradients in each direction, tensors of the eigenvalues ​​of the structure tensor, an energy tensor, a coherence tensor, an orientation tensor, a Harris index tensor, a Harris-Laplace index tensor.

[0194] In one embodiment, this step 82 of calculating characteristic values ​​of the structure tensor comprises several sub-steps.

[0195] Thus, the calculation step 82 comprises a sub-step 84 of calculating the eigenvalues ​​of the structure tensor matrix associated with the point P considered.

[0196] In the case where N=2, the structure tensor at each point P considered is a 2x2 matrix, which includes two associated eigenvalues, which are calculated analytically.

[0197] The eigenvalues ​​of the structure tensor associated with the point P considered are noted respectively Xi(p) and X2(p).

[0198] The eigenvalues ​​make it possible to obtain information on the distribution of the gradient in the window w, and to discriminate between regions with directional gradient, uniform regions or regions with rotational symmetry.

[0199] The eigenvalues ​​are stored in substep 84.

[0200] In addition, the method optionally comprises a sub-step 86 of calculating an energy value associated with the structure tensor, by applying the following formula:

[0201] [Math.4] e p = Pi( / ? )l + 14(p)I

[0202] The energy of a structure tensor, defined as the sum of its eigenvalues, characterizes the dynamics. The energy of the tensor at a point expresses the local contrast in the vicinity of this point. (J. Angulo. Structure Tensor Image Filtering using Riemannian L1 and Loo Center-of-mass. Image Analysis and Stereology, vol. 33, no. 2, pp. 95-105,2014.)

[0203] If the energy is close to 0, the point P belongs to a homogeneous region.

[0204] The energy value associated with the point considered is stored.

[0205] The calculation step 82 also optionally comprises a sub-step 88 of calculating the associated coherence, the coherence being defined by the difference between the maximum eigenvalue and the minimum eigenvalue divided by the sum of the maximum eigenvalue and the minimum eigenvalue.

[0206] Coherence is also called confidence factor or dispersion indicator. Coherence represents a measure of local anisotropy and characterizes the dispersion of the gradient orientation, i.e. the local variability of the image geometry.

[0207] Let Xmax(p)=max(Xi(p), X2(p)) be the maximum eigenvalue and Xmin(p)=min(Xi(p),X2(p)) the minimum eigenvalue, the formula for calculating the coherence is:

[0208] [Math.5]

[0209] The coherence value associated with the point considered is stored.

[0210] The calculation step 82 further comprises, optionally, a sub-step 90 of calculating an orientation associated with the point considered, from the components of the structure tensor.

[0211] The orientation value is for example calculated by the formula:

[0212] [Math.6] 6p= jarctanl ) ' 2 \ 22^-P)~-E nP) /

[0213] The orientation value associated with the point considered is stored.

[0214] More generally, the method for calculating the orientation is for example described in the following publication: IL Dryden, A. Koloydenko and D. Zhou. Non-Euclidean Statistics for Covariance Matrices, with Applications to Diffusion Tensor Imaging. The Annals of Applied Statistics, vol. 3, no. 3, pp. 1102-1123, 2009.

[0215] The calculation step 82 further comprises, optionally, a sub-step 92 of calculating a Harris index value at point P and / or a calculation of a Harris-Laplace index value at point P. The Harris index, known in the field of image processing, is used for detecting corners in digital images. Similarly, the calculation of Harris-Laplace index values ​​is known in image processing.

[0216] The Harris and / or Harris-Laplace index values ​​are stored.

[0217] Thus, at the end of step 82 of calculating characteristic values ​​of the structure tensor, a signature tensor associated with the point considered is formed (step 94), the components of which are characteristic values ​​previously calculated.

[0218] In one embodiment, the characteristics of the structure tensor retained in the signature tensor include: tensors of the gradients in each direction, tensors of the eigenvalues ​​of the structure tensor, an energy tensor, a coherence tensor, an orientation tensor, a Harris index tensor, a Harris-Laplace index tensor

[0219] According to variants, the characteristics of the structure tensor retained in the signature tensor comprise a subset of one or more characteristics mentioned above.

[0220] It should be noted that in such a variant, certain calculation sub-steps are not carried out, for example if the energy value is not retained among the components of the signature tensor, the sub-step of calculating and storing the energy value is not carried out.

[0221] Steps 80 to 94 are repeated for each point considered.

[0222] Thus, for each point considered, an associated signature tensor is obtained and stored.

[0223] After processing all the points considered, the third processing 50 comprises a step 96 of obtaining a third signature tile, associated with the super-resolved drop tile, bringing together the tensors of the signatures associated with the dislocations considered. Each signature constitutes a tensor, therefore the final signature tile is an assembly of tensors which precisely characterizes each dislocation.

[0224] As an example, [Fig. 12] illustrates, for N=2, block 91 of coherence values ​​corresponding to data block 83 of [Fig. 10]; block 93 of corresponding orientation values; block 95 of corresponding energy values ​​and block 97 of corresponding Harris index values.

[0225] Returning to [Fig.2], the first, second and third signature tiles, obtained respectively at the output of the first processing 46, second processing 48 and third processing 50, are combined in step 52 to form an enriched signature tile associated with the processed super-resolved drop tile.

[0226] This enriched signature block is provided as input to a step 54 of dislocation detection and localization.

[0227] Step 54 implements a method for classifying the signature tile into one or more dislocation categories. The classification method is previously trained by machine learning, in a preliminary learning phase, carried out from examples of simulated dislocations on simulated data tiles, for theoretical structures.

[0228] For example, machine learning is performed with a fast random forest algorithm, described in the article by Léo Breiman (2001). "Random Forests", published in Machine Learning. 45(1):5-32).

[0229] Alternatively, other machine learning classification algorithms may be implemented, for example Bayesian type, or function-based, or rule-based, or a combination of such algorithms, or based on the entire range of tools used in machine learning classification.

[0230] According to another variant, the chosen classification method implements a neural network, trained by supervised learning, for example a convolutional neural network (CNN). Another variant is obtained by deep learning.

[0231] At the output of step 54, a map 35 of the dislocations is obtained, including the location and type of each dislocation detected.

[0232] Advantageously, the accumulation of several treatments and the exceptional sensitivity of these methods allows precise characterization of dislocations of any type (screw; corner or mixed), unlike all current state-of-the-art methods. 27

Claims

1. Claims Method for processing multidimensional microscopy data for the localization of dislocations in a sample of crystalline material, comprising an acquisition of at least one microscopy image of said sample forming an input data block, the or each image of said input data block being representative of a part of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two, each data item of said block corresponding to a point in the N-dimensional space, the method being characterized in that it comprises steps of: - for at least one data block extracted from said input data block, called component block, comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample, application (44) of a transformation to the component block to obtain a block, called super-resolved drop block, comprising drops representative of an atomic structure of a part of the observed sample, followed by at least one of the following three treatments: - applying a first processing (46) to said super-resolved droplet block comprising an application of spectral filtering to said super-resolved droplet block, and / or - applying a second processing (48) to said super-resolved droplet tile comprising a calculation of a super-resolved diffraction pattern of the super-resolved droplet tile, and / or - applying a third processing (50) to said super-resolved droplet tile comprising a calculation of a structure tensor associated with at least one subset of points of the super-resolved droplet tile; - formation (52) of an enriched signature block associated with the super-resolved drop block, said enriched signature block bringing together, for each point of said drop block, output data from at least one of said first, second and third processing operations; - detection and localization (54) of at least one dislocation, by application of a learning classification method audit machine enriched signature pad.

2. Method according to claim 1, in which the machine learning classification method is either supervised, semi-supervised, or autonomous after prior learning on signature tiles calculated on simulated dislocations, making it possible to obtain a mapping of the dislocations corresponding to said drop tile.

3. Method according to one of claims 1 or 2, comprising a step of cutting (42) the input data block into component blocks, each component block comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample.

4. A method according to any one of claims 1 to 3, wherein applying a transformation (44) to obtain a super-resolved droplet tile comprising drops representative of an atomic structure of a portion of the observed sample implements a segmentation step using a machine learning algorithm, trained to classify the data of the super-resolved droplet tile into two classes, respectively a “background” class and a “drop” class.

5. A method according to any one of claims 1 to 3, further comprising refining to obtain a refined super-resolved droplet tile, by applying a refining spectral filtering mask, wherein the refining spectral filtering mask is obtained from the maximum of the machine learning segmentation of the Fourier transform of the component tile normalized between 0 and 1 and an adjusted theoretical diffractogram normalized between 0 and 1.

6. Method according to one of claims 1 to 5, in which the first processing (46) comprises steps of: - applying (60) a Fourier transform to the super-resolved droplet block to obtain a frequency block, - spectral filtering (62) by applying a masking window to the frequency block, said masking window corresponding to at least two crystal planes corresponding to two drops of the frequency space symmetrical with respect to the center of the Fourier transform of the super-resolved drops of the real space; - application (64) of an inverse Fourier transform to obtain corresponding signature data.

7. Method according to claim 6, further comprising a step of correcting artifacts (66) introduced by the Fourier transform.

8. A method according to any one of claims 6 or 7, wherein the spectral filtering (62) is applied for a plurality of masking windows corresponding to distinct crystal planes.

9. Method according to any one of claims 1 to 8, in which the second processing (48) comprises steps, applied to said super-resolved droplet block, of: - adding (72) a periphery, along each of the dimensions, of at least one point of zero value, making it possible to obtain an augmented block, - applying (74) a Fourier transform to the augmented block, - applying (76) a refining spectral filtering mask to the result of the previous step to obtain a refined super-resolved diffractogram of the super-resolved droplet block, forming a second block of signatures.

10. The method of claim 9, further comprising, before the adding step (72), a step of eroding the drops, each drop of the super-resolved droplet tile having an initial size, so as to reduce said initial size of each drop, to obtain an eroded droplet tile.

11. A method according to claim 9 or 10, further comprising, after the step of applying (76) a filtering mask, a step of applying (78) a theoretical convolution to the super-resolved diffractogram to obtain a convolved super-resolved diffractogram of each super-resolved component tile.

12. Method according to any one of claims 1 to 11, in which said third processing (50) comprises, for at least a subset of neighboring points of said block of drops, steps of: -calculation (80) of a structure tensor as a function of the values ​​of directional gradients, each direction corresponding to a dimension of said space, in a neighborhood of said point, -calculation (82) of one or more characteristic values ​​of the structure tensor among: tensors of the gradients in each direction, tensors of the eigenvalues ​​of the structure tensor, a tensor energy, a coherence tensor, an orientation tensor, a Harris index tensor, a Harris-Laplace index tensor, - formation (94) of a signature tensor associated with said point, the signature tensor having as components said characteristic values ​​of the structure tensor; - obtaining (96) a third signature tile corresponding to said super-resolved drop tile, said third signature tile bringing together the signature tensors associated with the points of said super-resolved drop tile.

13. A computer program comprising software instructions which, when executed by a programmable electronic device, implement a method of processing multidimensional microscopy data for the localization of dislocations in a sample according to claims 1 to 12.

14. Device for processing multidimensional microscopy data for locating dislocations in a sample of crystalline material, configured to acquire at least one microscopy image of said sample forming an input data block, the or each image of said input data block being representative of a part of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two, each data item of said block corresponding to a point in the N-dimensional space, the device being characterized in that it is configured to implement: - for at least one data block extracted from said input data block, called a component block, comprising homogeneous data according to a homogeneity criterion relating to at least one material of the observed sample, a module (22) for applying a transformation to the component block to obtain a block,said super-resolved droplet block, comprising drops representative of an atomic structure of a part of the observed sample, and to implement at least one of the following three modules: - a module (24) for applying a first processing to said super-resolved droplet block comprising an application of spectral filtering to said super-resolved droplet block, and / or - a module (26) for applying a second processing to said, super-resolved droplet tile comprising a super-resolved diffraction pattern calculation of the super-resolved droplet tile, and / or - a module (28) for applying a third processing to said super-resolved droplet block comprising a calculation of a structure tensor associated with each point of at least one subset of points of the super-resolved droplet block; as well as : - a module (30) for forming an enriched signature block associated with the super-resolved drop block, said enriched signature block combining, for each point of said super-resolved drop block, output data from at least one of said first, second and third processing operations; - a module (32) for detecting and locating at least one dislocation, by applying a machine learning classification method to said enriched signature block.

Images