FR3058816A1 - METHOD FOR NON-DESTRUCTIVE CONTROL OF METAL PIECE - Google Patents

Abstract

The invention relates to a method for non-destructive testing of a metal part for aeronautics comprising the following steps: - generation of a 2D image of the metal part using an X-ray generator, - processing of the 2D image adapted to attenuate artifacts and / or variabilities of the 2D image, - processing of the part pixels in order to tessellate the part in the 2D image into a plurality of images, - for at least one image , automatic extraction of a signature of said thumbnail, determining a plurality of descriptors p, from descriptors of a statistical model, - automatic classification of the signature of the thumbnail in a p-dimensional space for attributing to said imagette, a label representative of the presence of anomalies in the room, said reference database comprising reference images to which are assigned a label relating to the presence of anomalies.

Description

GENERAL TECHNICAL AREA

The invention relates to the field of non-destructive testing (CND) on industrial parts made of metallic material, in particular in the aeronautical field, using images acquired by means of an X-ray generator.

More specifically, the invention relates to the automation of NDT methods.

STATE OF THE ART

The CND is essential to control the material health of materials. For example, high pressure turbine blades are critical parts that need to be fully checked.

Each part from the production chain is individually checked to ensure that it complies with specifications. This material health check can be carried out by means of an X-ray generator which bombards the part to be inspected. The attenuation of the X-rays is observed thanks to a detector placed behind the part forming a 2D image (cf. Figure 1). The grayscale of this projected image is proportional to the attenuation of the X-rays passing through the room.

The high pressure turbine blades have a complex, hollow internal structure, allowing air to pass through to cool them. Additional thicknesses and sub-thicknesses may be present inside the dawn. In addition, other defects may appear during the part's manufacturing process: more or less dense foreign bodies, shrinkage, inclusions, residues, oxides ...

Traditionally, after acquiring a projection of the part (2D image), an operator comes to manually check this image in search of indications of critical faults.

The image is coded in 16 bits (i.e. 65536 grayscale), but the control screens only allow the display of 8 bit images. The operator must then constantly adjust the contrast of the image to distinguish small irregularities in gray levels in the image. This phenomenon is accentuated for parts of variable thickness (such as blade parts), the gray levels being correlated with this variation in thickness. The operator must therefore scan the entire image while adjusting the contrast of the image. It is a relatively long and tedious operation, and a source of error.

After having manually detected all the indications of the part, the part is either accepted and mounted on the engine, or is rejected according to the derogation process.

There is therefore a need to automate this detection process, in order to reduce the analysis time compared to a manual analysis, and to reduce the inspection costs. An automated process also makes it possible to overcome the subjectivity introduced by the operators (visual fatigue, intra-operator variability, inter-operator variability). The process being more repetitive and reliable, it allows the improvement of the yield.

There are known non-destructive testing methods implementing automatic methods for detecting workpiece defects.

These can use descriptors, which define analyzes of the gray levels of the image, to search for characteristics of the image corresponding to defects.

Thus, these methods conventionally use generic descriptors (ex: US20110182495A1).

There are also known methods using manually designed descriptors (eg US 20090066939A1).

However, there is a need to design more efficient descriptors making it possible to automatically obtain descriptors, optimal, specific to a part to be examined and to the types of associated faults. In order to improve the detection and characterization of faults present in a metal part.

PRESENTATION OF THE INVENTION

The invention implements non-destructive testing, by X-rays generating 2D images of the parts to be inspected, and by an automatic analysis method using statistical methods of supervised deep learning.

The invention provides a method of non-destructive testing of a metal part for aeronautics comprising the following steps:

E0: generation of a 2D image of the metal part using an X-ray generator, said 2D image corresponding to a projection onto a plane of the part,

- El: 2D image processing adapted to attenuate artifacts and / or variability of the 2D image,

- E3: processing of the room pixels in order to produce a tiling of the room in the 2D image in a plurality of thumbnails,

- E4: for at least one thumbnail, automatic extraction of a signature from said thumbnail, determining a plurality of p descriptors, a descriptor being the result of an image analysis, using descriptors of a previously determined statistical model by deep learning supervised from a reference database,

- E5: automatic classification of the signature of the thumbnail in a p-dimensional space allowing to attribute to said thumbnail, a label representative of the presence of anomalies in the part, the dimensional space and a partition of said space being determined automatically beforehand by supervised deep learning from said reference database, said reference database comprising reference thumbnails to which are assigned a label relating to the presence of anomalies.

The invention frees up time for the operator, who can then perform tasks with higher added value. In addition, the subjectivity introduced by the operators, linked to intra- and inter-operator variability is eliminated.

In addition, the financial cost of the inspection is reduced.

The invention can also include the following features:

the processing step comprises a segmentation of the image into room pixels and empty pixels,

- the signatures are vectors with p dimensions, p being the number of descriptors,

the vectors having the same label define a partition of the p-dimensional space into domains with which said label and said label value are associated, said domains being stored,

- The classification step associates a label with the vector of one of the thumbnails when said vector is in a domain associated with said label.

- said reference database is generated beforehand from at least one learning piece, on which the following steps are implemented:

• FO: generation of the 2D image of said test part using an X-ray generator, • F1: processing of the 2D image adapted to attenuate artifacts and / or variabilities of the 2D image, • F3: processing of the room pixels in order to produce a tiling of the room in the 2D image in a plurality of thumbnails, • F4: for each thumbnail of said plurality of thumbnails, allocation to said thumbnail, of a label and of a label value relating to the presence of anomalies as a function of a visual inspection of the thumbnail, said visual inspection having the aim of detecting anomalies within said thumbnail, and storage of said allocation in the database reference data • F5: for a plurality of thumbnails of the database, automatic evaluation of a signature of each thumbnail, by deep learning supervised from said reference database, determining a plurality of descriptors, • F6 : det automatic ermination of a p-dimensional space and a partition of said space by supervised deep learning from said reference database, and classification of the signature of the thumbnail in said p-dimensional space allowing automatic attribution to said thumbnail, a label representative of the presence of anomalies in the room, • F7: Storage of said statistical model comprising said plurality of descriptors and the p-dimensional space and a partition of said space in a learning database.

- deep learning is implemented by a convolutional neural network architecture comprising a set of convolution and subsampling layers for the automatic signature evaluation step, and at least one layer of neurons allowing the automatic classification of said signature,

- each thumbnail of the learning piece is modeled in the form of a vector with p components, the vector being positioned in the p-dimensional space, and each thumbnail has a label and a label value assigned using visual analysis,

- the paving step is performed for the entire image,

- at least one label has only one class, said domain of dimension p being a closed domain,

- at least one label with at least two classes, said classes corresponding to a region without defect and at least to a region with defect, corresponding to at least two domains of dimension p, which are two domains and in which a border between these two areas is established.

- the paving treatment involves overlapping areas between thumbnails, in order to minimize the leakage rate.

The invention also relates to a calculation unit, comprising processing means and a memory, said unit being configured to implement a method according to one of the preceding characteristics.

The invention also relates to a computer program product, configured to be executed by a calculation unit as described above and comprising code instructions for the execution of a method according to one of the preceding characteristics.

PRESENTATION OF THE FIGURES

Other characteristics, aims and advantages of the present invention will appear on reading the detailed description which follows and with reference to the appended drawings, given by way of examples and not limiting and in which:

FIG. 1, already presented, illustrates a 2D image acquisition system by an X-ray generator,

FIG. 2 illustrates certain steps of an embodiment in accordance with the invention, as well as the links between a method of

CND according to the invention and a database generation method for implementing a CND,

FIGS. 3, 4 and 5A, 5B illustrate the various devices and products involved in the context of the invention,

FIG. 6 illustrates a network architecture of convolutional neurons,

FIG. 7 illustrates a p-dimensional space in which points forming domains are positioned, and

- Figure 8 illustrates the performance comparison between different types of descriptors.

DETAILED DESCRIPTION

The CND can be applied at different stages of manufacturing a part 10 to be inspected. The part 10 is made of metallic material.

The parts on which the methods described can be implemented are preferably intended for aeronautics. It is a high pressure turbine blade for example, more specifically an M53 engine blade (Mirage 2000 engine).

FIG. 2 represents certain steps of an embodiment of the invention.

In order to be able to carry out a non-destructive check inside the room 10, a preliminary acquisition step using an imaging system 20, such as an X-ray generator, and calculation means ( see figure 3) is performed. This step provides a 2D image of part 10.

This image is called a projection. A set of projections is acquired without moving. For example, in one embodiment, the generator 21 and the detector 22 being fixed, an articulated arm (example:

robot type from the manufacturer Kuka ®) automatically grasps the part and places it between the generator 21 and the detector 22 at a position defined in advance in a range of acquisitions.

A computing unit 30 is provided for carrying out the computer processing. It includes processing means 31 such as a microprocessor and a memory 32.

With reference to FIGS. 4 and 5A, 5B, the method implemented for automatically detecting anomalies requires databases (BDD) 40 and 41. BDD 40 and 41 are stored on at least one server 40, and a calculation unit 50, which can be the same calculation unit as for the acquisition of the 2D image, is used. The calculation unit 50 also includes data processing means 51 and a memory 52. The server 40 can be this calculation unit 50 and the BDDs 40 and 41 can be stored in said memory 52.

Calculation of a statistical model

In what follows, the part 10 is a so-called learning part and will be referenced 11. This learning part 11 will make it possible to generate reference images in a reference BDD 40 which will help to calculate a statistical model of the images of said learning piece 11.

Image preprocessing (Fl)

Before implementing the methods which will be described later, preprocessing steps F1 are preferably carried out to improve the quality of the analyzes, following step F0 of acquiring the 2D image of the learning part 11 .

Indeed, during this acquisition, artefacts and variability may appear (defective pixels, noise generated by the diffused flux, diffraction ...). Preprocessings can therefore be applied to the 2D image to remove and / or attenuate these artefacts / variabilities.

Thus, it is possible to implement on the 2D image filtering of the noise of the image by using, for example, a median filter, a Gaussian filter (class of low pass filters).

The 2D image can also advantageously undergo a normalization step to standardize the 2D images so that they are all comparable. Several standardization methods can be used independently or coupled together:

- a registration of the images applied to the images in order to put them all in a common frame of reference in order to absorb the variability of the acquisition.

- A calculation of the median value and the MAD (Median Absolute Difference) on the pixels corresponding to the part for each image and fixed at identical predetermined values for all the images.

- A local contrast calculation: the part inspected does not always have a constant thickness (this is never the case for a blade). Also, the local contrast is calculated for all the pixels of the image (subtraction of the low frequencies from the standardized image) (bilateral filter [1], anisotropic scattering [2], total variation filter [3], Edge-avoiding wavelets [4]).

In addition, a mask of the learning part 11 is generated from the 2D image, that is to say that a segmentation between two classes of pixels, air and material, is carried out (step F2). This segmentation can be achieved by automatically calculating the optimal threshold allowing the two classes to be separated.

An example of a method consists in calculating the histogram of the gray levels of the image and in maximizing the interclass variance of this histogram in order to define the threshold which separates the pixels of air and those of matter (see reference [5 ]). Other automatic thresholding methods are possible (see reference [6]).

This step makes it possible to perform the analysis of the part only on regions of interest, relating to the material zones of the part 11.

After processing, in a step F2, the 2D image of the training part 11 itself is recovered by the processing unit for the following topological analysis.

Paving step F3

The 2D image is paved by fixed size boxes (which can overlap). The size of the boxes as well as the percentage of overlap (generally between 50 and 75%) are parameters which are optimized according to the applications.

The image is therefore broken down into a set of thumbnails (patch). For example, the size of the patch could be equal to 32x32 pixels (representing a few mm ² on part 11).

At the end of the tiling step, a plurality of thumbnails I'I, ..., I'n are thus obtained, the recombination of which makes it possible to reconstruct the image. In the following description, it is assumed that the tiling of the image is done using n thumbnails I'I, ..., I'n, n being a natural integer.

The superimposition makes it possible to reduce the leakage rate, thus to avoid missing a defect, the paving of the part 11 involves overlapping zones between two thumbnails.

Because of this overlap, an anomaly that is not completely included in a given thumbnail will be completely included in another thumbnail positioned at a nearby location.

Constitution of a reference database 40 (F4)

It will be recalled that a plurality of learning pieces 11 is necessary to generate a reference database 40.

This step involves a visual inspection of the 2D image representing a projection of the learning part 11. This visual inspection is carried out by an operator, according to methods known from the prior art. However, any method which makes it possible to identify a fault with certainty can be used, whether carried out manually or semi-automatically by an operator.

Following the visual inspection, each thumbnail ΙΊ, ..., is assigned a label. As the part, with its possible defects, has been visually inspected, the signature will serve as a reference.

There are two ways to mark a thumbnail.

A single class Ll label, allowing the detection of faults, is assigned if it is not possible to find thumbnails having the opposite class. For example, if it is not possible to identify a thumbnail image with anomalies, the single-class Ll label has the only possible value "without defect" (or "healthy").

A multi-class L2 label, allowing the detection and characterization of anomalies, is awarded if the thumbnail can be identified according to M sub-categories: oxide, residue, excess thickness, etc. The most elementary L2 multiclass label consists in assigning a value "without defect" ("healthy") or a value "defect" ("anomaly").

A vector Y of dimension n is thus obtained, corresponding to the number of thumbnails, each component of which corresponds to the value of a label.

The value of the label can advantageously be defined numerically, that is to say that a flawless ("healthy") thumbnail will have a label Ll or

L2 equal to 0, and the defects will be numbered from 1 to M for example, M being a natural integer.

Thus, in p-dimensional space E, each vector of the vector cloud comprises a label L1 or L2 with an assigned value. As an illustration, a vector Y can have the following form: [Ll = 0, Ll = 0, L2 = 2, Ll = 0, L2 = M, L2 = 0, Ll = 0, ..., L2 = l], with n components.

This correspondence between thumbnail and label is stored in the reference database 40.

To obtain a reliably usable reference database 40, it must be constructed using a large number of images representing the variability of the images of the part 11 and the variability of the production. Several parts 11 are advantageously used.

The number of thumbnails to be inserted in the BDD 40 must be large to have a robust description of the signatures of a healthy region and an abnormal region.

A balancing phase of the database can possibly be carried out. Indeed, it is generally much more difficult to obtain defect samples compared to healthy samples. Also, the number of samples is very unbalanced which is detrimental during statistical learning. To fill this imbalance and to improve the robustness of learning, transformations can be added to the patches superimposing a defect to increase the representativeness of the anomalies. Thus, each “defect” sample will be boosted (application of small realistic geometric transformations), so that for each “defect” sample, a hundred other defect samples are generated. This boost of BDD 40 makes it possible to increase variability and thus limit overfitting.

Image signature (step F5)

From each thumbnail I'I, I'n is then automatically calculated a signature from the reference BDD 40. A signature is a set of numerical values, which we conveniently define in the form of a vector, to which we associates a point P. These numerical values are obtained using descriptors, which are for example convolution operations applied to the thumbnail. The descriptors constitute analyzes of the gray levels N DG of the thumbnail.

For each thumbnail, there is preferably a plurality of descriptors. We define p as the number of descriptors, which is a natural integer.

Each thumbnail is therefore representable as a point P in a p-dimensional space E. Otherwise formulated, the vector associated with the thumbnail is projected into this p-dimensional space E (see Figure 7 detailed later in the description).

In step F5, the signature of the thumbnail is obtained automatically by deep statistical methods. These methods make it possible to automatically extract the optimal signature of the thumbnail and to obtain descriptors specific to a given application.

Deep statistical methods can be carried out by supervised learning.

Figure 6 illustrates an example of deep learning architecture through a convolutional neural network.

The input data of the neural network is a thumbnail of a dimension of 32 × 32 pixels for example.

The convolutional neural networks extract the signature of said thumbnail by a succession of processing layers comprising convolution operations (Cl, C2, ..., C6) and optionally pooling operations (P2, P4, P6).

A convolution is a mathematical operation which makes it possible to modify the initial image by a convolution kernel by filtering it.

A convolution kernel will analyze a characteristic of the input image. A non-linear operation is applied to each step (at each convolution): sigmoid, re-read, tanh, ... allowing the extraction of visual characteristics (side, angle, end detection ...).

The convolution kernel can be for example a matrix of size 3x3.

To analyze several characteristics, we stack strata of independent convolution kernels, each stratum analyzing a characteristic of the image. All of the strata thus stacked form the convolutional treatment layer. The number of treatment layers defines the depth of the convolution layer.

All the outputs of a processing layer make it possible to reconstruct an intermediate image, which will serve as the basis for the next processing layer.

As illustrated in FIG. 6, the first Layerl layer comprises a convolution layer comprising a depth of size 20. Said first layer is followed by a second Layer2 treatment layer comprising a second convolution layer C2, for example of the same depth and core size . The convolution in said second layer is followed by a pooling layer.

A pooling layer allows a step of sub-sampling the image, thus reducing the quantity of parameters and calculation, allowing large gains in computing power. The pooling layer also allows a multi-scale approach, the next convolution layer processing a larger part of the initial image.

It is therefore frequent to periodically insert a pooling layer between two successive convolutional layers of an architecture of deep statistical methods.

Pooling reduces the spatial size of an intermediate image by taking for example the maximum response between the different pixels (max pooling) for example by a 2x2 Max-Pool method (compression by a factor of 4).

We can also use, average pooling, L2norm pooling or a stochastic pooling function.

After applying Max-Pool 2x2 pooling in the P2 pooling layer, the size of the images is therefore 16x16 pixels.

A third treatment layer (layer 3) comprises a convolution layer C3 of depth 50.

A fourth processing layer Iayer4, includes a convolution layer C4 and a pooling layer P4 which increases the size of the image to be analyzed, for example to 8x8 pixels.

The architecture can also include a fifth and a sixth treatment layer, Iayer5 and layerô, of a depth for example of size 100.

For each thumbnail I'I, I'n of the learning piece 11, a signature is obtained in the form of a vector with p dimensions. For a set of thumbnails I'I, ... I'n, we obtain a point cloud P in p-dimensional space, by projection of the associated vectors.

Consequently, one obtains for the learning piece 11 concerned a matrix X of dimension n by p, the coefficients of which correspond to the numerical values of the descriptors.

Signature classification (F6)

Following the sequences of convolutional and pooling layers, the architecture includes at least one fully connected layer

FC, which is a perceptron-like layer. The neurons in a fully connected layer have connections to all of the outputs of the previous layer (as seen regularly in regular neural networks).

Thus, in a step F6, this layer makes it possible to classify the signature of a thumbnail and therefore to assign a label to the thumbnail of said signature.

To allow the classification of the thumbnails, the thumbnails of the reference database 40 having the same labels are brought together in order to determine ranges of values of the descriptors making it possible to characterize an thumbnail.

It is a question of determining using the thumbnails having the same labels (that is to say a cloud of points P), of the subdomains Dl, D2, \ D2 of the p-dimensional space E ( see figure 7 again). These subdomains D1, D2, \ D2 can be of dimension p if the set of descriptors is taken into account, or of smaller dimension, if at least one descriptor is ignored. At the end of step F6, a partition of space E is generated.

The generation of these domains D1, D2, \ D2 makes it possible to determine which label and which label value to grant to a point placed in the p-dimensional space E.

If it is a single class Ll label, the flawless ("healthy") thumbnails therefore form a point cloud P in space E with p dimensions. A closed domain is thus defined (see Dl in FIG. 7) inside which are all the points or a large majority of the points comprising the same value of a label of type L1. In fact, it is possible to ignore certain points (values that are too far apart, "errors", etc.).

When a new vector is inserted (new thumbnail), the opposite value of the single-class label will be assigned to it if it is outside the defined area.

If it is a multi-class L2 label, the classification step amounts to defining a decision border between the signatures of healthy thumbnails and that of abnormal thumbnails. At least two open domains are thus defined (see D2, \ D2 in FIG. 7), all (or almost all) the points comprising the same value of a label of type L2 being in the same domain. As an example, the boundary can be a hyperplane of p-dimensional space (i.e. a plane of dimension p-1), which splits space into two domains.

If the L2 label has three classes, we obtain three sub-domains.

Determination of the best statistical model and storage (F7)

In a step called "feed forward" from the first layer to the last processing layer (layerl to FC). We extract from the thumbnails presented to the neural network a signature, as well as a predicted label.

The “feed forward” stage makes it possible to predict a label by learning the optimal signatures and the partition of the dimensional space.

For each thumbnail, the label predicted in the "feed forward" step is compared to the label determined by the operator of said thumbnail of the annotated database. If these differ, an error is determined. Statistical learning is carried out by backpropagation of the error made.

We calculate the derivative of this error to determine how to change the weights of the convolution kernels of the different layers so as to reduce the error.

The error is thus back-propagated in the network layer by layer (FC to layerl), from the last to the first layer to update the weights of each layer. The learning algorithm implemented for backpropagation is for example the stochastic gradient descent method.

It is therefore these convolution kernels that are learned during statistical learning when updating weights. Statistical learning then makes it possible to determine:

- learning the best signature (descriptors) of the images making it possible to discriminate healthy images from faulty images,

- learning the best separation (decision boundary) between the signatures of sound images and signatures of anomalies in a part, by generating a representation space in which the data can be classified with controlled performance.

In a step F7, the parameters of the statistical model obtained by supervised learning from the reference BDD 40 are stored in a training database 41. The stored parameters are for example, the best signature (descriptor (s)) of learning room 11, and the best decision boundary between the signatures of the healthy images and the signatures of the anomalies of a room.

Of course, other architectures and other deep statistical methods can be used, for example multi-layer neural networks, multi-layer Boltzman machines (Deep Belief Network), stacked auto-encoders.

FIG. 8 illustrates the performance gain of a convolutional neural network compared to conventional methods using generic descriptors which can be used for different applications. A non-exhaustive list of such possible methods for calculating said generic descriptors is described below:

Histogram of the image, where statistical analysis of it through the mean, the median, the mode, the standard deviation, the kurtosis (which each constitutes a descriptor),

Haralick descriptors which characterize the texture through the spatial organization of bi-points,

“Gray Level Run Length Matrix (GLRLM)” descriptors which characterize the length of the homogeneous areas of the image texture,

“Gray Levei Size Zone Matrix (GLSZM)” descriptors which characterize the size of the homogeneous zones of the texture of an image,

Granulometry: using a sieving process, a granulometry makes it possible to count the elements constituting a determined characteristic by classifying them by size,

Gabor filters: a bank of Gabor filters allows a directional wavelet analysis of the texture by looking at the energy of the response of each filter,

Autocorrelation: Use of 3D volume autocorrelation in certain preferred directions,

Non-directional frequency analysis: from the Fourier transform of the volume, calculation of the energy in different frequency bands, "Histogram of Oriented Gradient 3D (HOG)": after calculating the orientation of the gradient for each pixel of l 'image, we count the number of pixels having a similar direction. The unitary sphere is then paved in different zones which correspond to the different boxes of the histogram.

Wavelet / curvelette analysis: time / frequency analysis of the tomographic volume, we use for example Haar, Daubechies, Morlet wavelets ...

Binary descriptors: Local Binary Pattern, DAISY, FREAK,

BRIEF, BRISK ...

The abscissa line represents the percentage of false positives, the ordinate line represents the percentage of true positives. The line defined by the points (x = 0, y = 0) and (x = l, y = l) represents a decision which would be taken randomly.

The perfect point that we aim at in Figure 4 is the point (x = 0, y = l) corresponding to 0% false positives for 100% of detected defects. The curve (CNN) corresponding to a network model of convolutional neurons is very close to the target point and offers much better performance than that of conventional methods.

Example of implementation of a non-destructive testing automation process

The process is similar to the generation of the BDD database, except that the database is used to assign a label L1, L2 to the thumbnails II, ... In of the image of the part 10 We no longer speak of a reference signature for part 10, since we are trying to characterize whether the part is healthy or abnormal using these signatures.

The same steps E0, E1, E2, and E3 corresponding to steps F0, F1, F2, and F3 are carried out.

In a step E4, by applying the statistical model and more particularly its descriptors stored in the learning database 41, a signature is obtained.

Once the signature obtained, in a step E5, instead of assigning it a label by visual inspection, and storing it (step F4), this signature is projected into the p-dimensional space E of the statistical model of the BDD d ' learning 41. In other words, each vector associated with an image II, ..., In is projected into the p-dimensional space, so as to position said thumbnail in said space E, and to classify said thumbnail. Then, due to the application of the statistical model and the partition of the space E, the placement of the point P corresponding to the vector makes it possible to automatically assign to it the value of a label L1, L2.

More precisely, the point P corresponding to one of the thumbnails II, ..., In is positioned in the p-dimensional space E. From this positioning, it is determined in which domain D1, D2, \ D2 predetermined of the learning BDD 41 this vector belongs, which makes it possible to attribute to the thumbnail the value of the label which characterizes said domain.

By positioning all the vectors of the thumbnails II, ..., In constituting the image of the piece 10, the set of signatures determined by the supervised statistical learning method made it possible to generate a partition of the space E, which in turn allows to generate a vector Y, called prediction, as defined above, grouping all the values of the labels. It suffices that a single coordinate of this vector Y, that is to say a value of a label, corresponds to an image with a defect so that the part 10 is not considered to be without defect. Tolerance criteria can nevertheless be set according to specifications.

If the domain is defined by a border, we calculate the algebraic distance to this border to determine the label. If it can happen that the point falls too close to a border, it is also possible to artificially move the border, in particular in order to reduce the rate of leaks: it is in fact better to falsely detect an anomaly than to to lack.

For example, if the thumbnail is far from a label L1 corresponding to a flawless piece, it can be deduced that the thumbnail corresponds to an unhealthy piece.

In order to allow compatibility between the reference BDD 40 and the automatic CND process, the size of the paving boxes of the paving step E3 has dimensions identical to that of the step F3. In this way, all the thumbnails used have the same dimensions.

In an optional step, a map of the defects present in the image of the part 10 is generated. If necessary, a visual inspection, called consolidation, is carried out by an operator to validate the detection.

In addition, following this visual inspection, the data generated by the process can complete the database, so that the

Reference BDD 40 is enriched as the parts are analyzed. Preferably, these data only complete the database if a difference between the result obtained and the consolidation is observed. Thus, the part 10 becomes a learning part 11.

References [1] Carlo Tomasi and Roberto Manduchi, Bilateral filtering for gray and color images, in Computer Vision, 1998. Sixth International Conférence on. IEEE, 1998, pp. 839-846.

[2] Pietro Perona and Jitendra Malik (November 1987). Scale-space and edge detection using anisotropic diffusion. Proceedings of IEEE Computer Society Workshop on Computer Vision ,, pp. 16-22.

[3] Rudin, L. I .; Osher, S .; Fatemi, E. (1992). Nonlinear total variation based noise removal algorithms. Physica D 60: 259-268.

[4] Raanan Fattal Edge-avoiding wavelets and their applications, ACM Trans. Graph, 28, 3, 2009.

[5] K Dabov, A Foi, V Katkovnik Image and video denoising by sparse 3D transform-domain collaborative filtering.

[6] Nobuyuki Otsu, "A threshold selection method from gray-level histograms", IEEE Trans. Sys., Man., Cyber., Vol. 9, 1979, p. 62-66 (DOI 10.1109 / TSMC.1979.4310076).

Claims (12)

Claims 1. Method for non-destructive testing of a metal part (10) for aeronautics, comprising the following steps: - (E0) generation of a 2D image of the metal part (10) using an X-ray generator (20), said 2D image corresponding to a projection on a plane of the part (10), - (E1) 2D image processing adapted to attenuate artifacts and / or variabilities of the 2D image, - (E3) processing of the room pixels in order to produce a tiling of the room in the 2D image in a plurality of thumbnails (II, In), - (E4) for at least one thumbnail, automatic extraction of a signature from said thumbnail, determining a plurality of p descriptors, a descriptor being the result of an image analysis, from descriptors of a determined statistical model previously by deep learning supervised from a reference database (40), - (E5) automatic classification of the signature of the thumbnail (II, ..., In) in a space (E) p-dimensional allowing to assign to said thumbnail, a label representative of the presence of anomalies in the piece (L1, L2), the p-dimensional space (E) and a partition of said space being determined beforehand automatically by supervised deep learning from said reference database (40), said reference database ( 40) comprising reference thumbnails to which a label (L1, L2) is assigned relating to the presence of anomalies. 2. A method of non-destructive testing of a metal part (10) according to the preceding claim, in which the processing step comprises a segmentation of the image into room pixels and empty pixels. 3. The method of claim 1 or 2, wherein; - the signatures are vectors with p dimensions, p being the number of descriptors, the vectors having the same label (L1, L2) define a partition of the p-dimensional space (E) into domains (D) with which said label and said label value are associated, said domains being stored, - the classification step (E5) associates a label with the vector of one of the thumbnails (II, ..., In) when said vector is in a domain associated with said label. 4. Method according to one of the preceding claims, in which said reference database (40) is generated beforehand from at least one learning part (11), on which the following steps are implemented: - (F0) generation of the 2D image of said test piece (11) using an X-ray generator (20), - (Fl) 2D image processing adapted to attenuate artifacts and / or variabilities of the 2D image, - (F3) processing of the room pixels in order to produce a tiling of the room in the 2D image in a plurality of thumbnails (II, ..., In), - (F4) for each thumbnail of said plurality of thumbnails (II, ..., In), allocation to said thumbnail of a label (L1, L2) and of a label value relating to the presence of anomalies based on a visual inspection of the thumbnail (Γ1, .... Γη), said visual inspection having the aim of detecting anomalies within said thumbnail, and storage of said allocation in the reference database ( 40). 5. Method according to the preceding claim, in which said descriptors of said statistical model are generated beforehand from the reference database (40), on which the following steps are implemented: - (F5) for a plurality of thumbnails of the database (40), automatic evaluation of a signature of each thumbnail (Γ1, .... Γη), by supervised deep learning from said database reference (40), determining a plurality (p) of descriptors, - (F6) automatic determination of a p-dimensional space (E) and of a partition of said space by supervised deep learning from said reference database (40), and classification of the signature of the thumbnail ( II, ..., In) in said p-dimensional space (E) making it possible to automatically assign to said thumbnail, a label representative of the presence of anomalies in the part (L1, L2), - (F7) Storage of said statistical model comprising said plurality of descriptors and the p-dimensional space (E) and a partition of said space in a training database (41). 6. Method according to the preceding claim, in which: each thumbnail (Γ1, ..., Γη) of the learning piece (11) is modeled as a vector with p components, the vector being positioned in p-dimensional space, and each thumbnail (Γ1 , ..., I'n) has a label (L1, L2) and a label value assigned using visual analysis. 7. Method according to one of the preceding claims, in which the deep learning is implemented by a convolutional neural network architecture comprising a set of convolution and subsampling layers for the step of automatic evaluation of the signature, and at least one layer of neurons allowing the automatic classification of said signature. 8. Method according to any one of the preceding claims, in which the tiling step (E3, F3) is carried out for the entire image. 9. Method according to any one of the preceding claims, in which at least one label (L1) has only one class, said domain of dimension p being a closed domain (Dl). 10. Method according to any one of the preceding claims, in which at least one label (L2) has at least two classes (M), said classes (M) corresponding to a faultless region and at least to a faulty region, corresponding to at least two open domains (D2, \ D2) of dimension p, which are two domains (D2, \ D2) and in which a border between these two domains is established. 11. Method according to any one of the preceding claims, in which the paving treatment (E3, F3) involves overlapping zones between thumbnails (II, ..., In, I'I, ... I'n ), to minimize the leak rate. 12. Computer program product, intended to be executed by processing means of a calculation unit, said calculation unit further comprising a memory, and comprising code instructions for the execution of a method according to the 'one of the preceding claims. 2/7