FR3135342A1 - Method for identifying an O-ring seal on a technical level and associated program product. - Google Patents

Abstract

Title: Method for identifying an O-ring seal on a technical level and associated program product. The invention relates to a method for identifying an O-ring seal on a technical plan, said method comprising, based on an image of said O-ring of given size from the technical plan, the following computer-implemented steps: processing said image in a convolutional neural network comprising only a single input channel and a single output channel; processing the image obtained from the output channel of the convolutional neural network to distinguish, by thresholding, a pixel from the image associated with an O-ring seal of a pixel not associated with an O-ring seal; implement a data partitioning algorithm of the image obtained at the end of step B) to identify the O-ring seal. Figure for abstract: Figure 2.

Description

Method for identifying an O-ring seal on a technical level and associated program product. Technical field of the invention

The invention relates to a method for identifying an O-ring seal on a technical level, as well as an associated program product.

It can find application in different fields, in particular in the field of transport, notably in the automobile sector.

Technical background

In the automotive field, many pipes are used to transport fluids and fittings to connect these pipes to the components of a vehicle. A key component of a fitting is the O-ring which provides the sealing function between vehicle components connected by a pipe.

When a connection is designed, the technical plan associated with it is stored in a database. This storage is carried out with different classification criteria in the database in order to allow easier subsequent research.

An important criterion is in particular the presence of an O-ring, and their number.

Today, the identification of an O-ring, and their number, on a technical level is manual. This leads to a classification of the different technical plans which is a source of error. This classification depends on the technical competence and experience of the person entering the information into the database. Furthermore, it is also time-consuming because it requires opening and analyzing the plan and all its views to be sure not to forget anything.

An objective of the invention is to improve the identification of an O-ring seal on technical levels, in particular with the objective of making a better classification.

To this end, the invention proposes a method for identifying an O-ring seal on a technical level, said method comprising, from an image of a given size of the technical plan, the steps implemented by computer following:
A) processing said image in a convolutional neural network comprising only a single input channel and a single output channel;
B) processing the image obtained from the output channel of the convolutional neural network to distinguish, by thresholding, a pixel of the image associated with an O-ring seal from a pixel not associated with an O-ring seal;
C) implement a data partitioning algorithm of the image obtained at the end of step B) to identify the O-ring seal.

• l’étape A) comprend les sous-étapes mises en œuvre par ordinateur suivantes :
  a) traiter ladite image, depuis l’unique canal d’entrée, dans une première couche du réseau neuronal convolutif, selon les sous-étapes suivantes :
  - effectuer une convolution avec comme hyperparamètres : un nombre de noyaux, définissant le nombre de canaux de sortie de la première couche, compris entre deux et quatre-vingt, une taille de noyau d’au moins treize, un pas valant l’unité ou deux et une marge choisie pour conserver la taille d’image à la valeur du pas près,
  - effectuer une correction avec une fonction d’activation,
  - effectuer une mise en commun avec une taille de noyau donnée, avantageusement entre cinq et vingt-et-un, un pas compris entre deux et quatre, une marge choisie pour conserver la taille d’image à la valeur du pas près;
  chaque image présentant alors une taille réduite ;
  b) traiter les images issues des canaux de sortie de la convolution de la sous-étape précédente dans une deuxième couche du réseau neuronal convolutif, selon les sous-étapes suivantes :
  - effectuer une convolution avec comme hyperparamètres : un nombre de noyaux, définissant un nombre de canaux de sortie pour la deuxième couche, d’au moins quatre-vingt-un, une taille de noyau comprise entre trois et vingt-et-un, un pas valant l’unité ou deux et une marge choisie pour conserver la taille d’image à la valeur du pas près,
  - effectuer une correction avec une fonction d’activation,
  - effectuer une mise en commun avec une taille de noyau donnée, avantageusement entre cinq et vingt-et-un, un pas compris entre deux et quatre, et une marge choisie pour conserver la taille d’image à la valeur du pas près,
  chaque image présentant alors une taille encore réduite;
  c) traiter les images issues des canaux de sortie de la convolution de la sous-étape précédente dans une autre couche du réseau neuronal convolutif avec l’étape suivante :
  - effectuer une convolution avec comme hyperparamètres: un nombre de noyaux de valeur unité définissant un canal de sortie unique pour cette autre couche, une taille de noyau de valeur unité, un pas de valeur unité et une marge nulle pour conserver la taille d’image.
• le nombre de noyaux pris en considération pour la convolution de la sous-étape a) est compris entre 10 et 64, avantageusement entre 18 et 64 ;
• la taille de noyau prise en considération pour la convolution de la sous-étape a) est d’au moins 21, avantageusement d’au moins 31 ;
• le nombre de noyaux pris en considération pour la convolution de la sous-étape b) est d’au moins 100, avantageusement d’au moins 120 ;
• la taille du noyau prise en considération pour la convolution de la sous-étape b) est d’au plus 15, avantageusement d’au plus 9 ;
• le pas pris en considération pour effectuer la convolution de la sous-étape a) est de valeur unité si bien que la marge prise en considération pour cette même convolution conserve la taille d’image ; et/ou le pas pris en considération pour effectuer la convolution de la sous-étape b) est de valeur unité si bien que la marge prise en considération pour cette même convolution conserve la taille d’image ;
• l’étape A) comporte une sous-étape supplémentaire mise en œuvre par ordinateur, entre la sous-étape b) et la sous-étape c), consistant à traiter les images issues des canaux de sortie de la convolution de la sous-étape b) dans une troisième couche du réseau neuronal convolutif, selon les sous-étapes suivantes :
  - effectuer une convolution avec comme hyperparamètres: un nombre de noyaux compris entre 2 et 80, une taille de noyau d’au plus 15, un pas de valeur unité ou deux et une marge choisie pour conserver la taille d’image à la valeur du pas près,
  - effectuer une correction avec une fonction d’activation,
  - effectuer une mise en commun avec un nombre de noyaux donné, avantageusement entre trois et onze, un pas compris entre deux et quatre, et une marge choisie pour conserver la taille d’image à la valeur du pas près ;
• le nombre de noyaux pris en considération pour effectuer la convolution de la sous-étape supplémentaire est compris entre 10 et 64, avantageusement entre 18 et 64 ;
• la taille du noyau prise en considération pour effectuer la convolution de la sous-étape supplémentaire est d’au plus 9 et avantageusement d’au plus 5 ;
• le pas pris en considération pour effectuer la convolution de la sous-étape supplémentaire est de valeur unité si bien que la marge prise en considération pour cette même convolution conserve la taille d’image ;
• la fonction d’activation choisie pour effectuer la correction est une fonction d’activation non saturante, aussi nommée Unité Linéaire Rectifiée (ReLU) ;
• la mise en commun est réalisée au moyen d’une fonction max (MaxPool2D) ;
• les sous-étapes mises en œuvre par ordinateur suivantes, avant de mettre en œuvre l’étape a) : mettre en œuvre un algorithme de partitionnement de données sur l’image du plan technique ; définir une image du plan technique, de taille restreinte et comportant des zones d’intérêt susceptibles de contenir un joint d’étanchéité torique ;
• l’algorithme de partitionnement de données est l’algorithme DBSCAN.

The method according to the invention may comprise at least one of the following characteristics, taken alone or in combination:

• step A) includes the following computer-implemented substeps:
  a) processing said image, from the single input channel, in a first layer of the convolutional neural network, according to the following substeps:
  - perform a convolution with hyperparameters: a number of kernels, defining the number of output channels of the first layer, between two and eighty, a kernel size of at least thirteen, a step value of unity or two and a margin chosen to keep the image size to the nearest step value,
  - perform a correction with an activation function,
  - perform a pooling with a given kernel size, advantageously between five and twenty-one, a step of between two and four, a margin chosen to keep the image size to the nearest value of the step;
  each image then having a reduced size;
  b) processing the images from the output channels of the convolution of the previous sub-step in a second layer of the convolutional neural network, according to the following sub-steps:
  - perform a convolution with as hyperparameters: a number of kernels, defining a number of output channels for the second layer, of at least eighty-one, a kernel size between three and twenty-one, one not worth the unit or two and a margin chosen to keep the image size to the nearest step value,
  - perform a correction with an activation function,
  - perform a pooling with a given kernel size, advantageously between five and twenty-one, a step of between two and four, and a margin chosen to keep the image size to the nearest step value,
  each image then having an even smaller size;
  c) process the images from the output channels of the convolution of the previous sub-step in another layer of the convolutional neural network with the following step:
  - perform a convolution with as hyperparameters: a number of kernels of unity value defining a single output channel for this other layer, a kernel size of unity value, a step of unity value and a zero margin to preserve the image size .
• the number of kernels taken into consideration for the convolution of sub-step a) is between 10 and 64, advantageously between 18 and 64;
• the kernel size taken into consideration for the convolution of sub-step a) is at least 21, advantageously at least 31;
• the number of kernels taken into consideration for the convolution of sub-step b) is at least 100, advantageously at least 120;
• the size of the kernel taken into consideration for the convolution of sub-step b) is at most 15, advantageously at most 9;
• the step taken into consideration to carry out the convolution of sub-step a) is of unity value so that the margin taken into consideration for this same convolution preserves the image size; and/or the step taken into consideration to carry out the convolution of sub-step b) is of unity value so that the margin taken into consideration for this same convolution preserves the image size;
• step A) comprises an additional sub-step implemented by computer, between sub-step b) and sub-step c), consisting of processing the images from the output channels of the convolution of the sub-step b) in a third layer of the convolutional neural network, according to the following sub-steps:
  - perform a convolution with hyperparameters: a number of kernels between 2 and 80, a kernel size of at most 15, a step of unit value or two and a margin chosen to keep the image size at the value of not near,
  - perform a correction with an activation function,
  - perform a pooling with a given number of cores, advantageously between three and eleven, a step of between two and four, and a margin chosen to maintain the image size to the nearest value of the step;
• the number of kernels taken into consideration to carry out the convolution of the additional sub-step is between 10 and 64, advantageously between 18 and 64;
• the size of the kernel taken into consideration to carry out the convolution of the additional sub-step is at most 9 and advantageously at most 5;
• the step taken into consideration to carry out the convolution of the additional sub-step is of unity value so that the margin taken into consideration for this same convolution preserves the image size;
• the activation function chosen to carry out the correction is a non-saturating activation function, also called Rectified Linear Unit (ReLU);
• pooling is carried out using a max function (MaxPool2D);
• the following computer-implemented sub-steps, before implementing step a): implementing a data partitioning algorithm on the technical plane image; define an image of the technical plan, of restricted size and including areas of interest likely to contain an O-ring seal;
• The data partitioning algorithm is the DBSCAN algorithm.

The invention also relates to a computer program product downloadable from a communications network and/or recorded on a computer-readable medium, comprising instructions for executing a method according to the invention when the computer program product is executed by a computer system.

Brief description of the figures

The invention will be better described with the help of the figures:

There is a cross-sectional diagram of an O-ring.

There is a diagram representative of an exemplary embodiment of a method according to the invention.

Detailed description of the invention

The invention relates to a method for identifying an O-ring seal on a technical level, said method comprising, from an image of a given size of said technical plan, the following steps implemented by computer:
A) processing said image in a convolutional neural network, comprising only a single input channel and a single output channel;
B) processing the image obtained from the output channel of the convolutional neural network to distinguish, by thresholding, a pixel of the image associated with an O-ring seal from a pixel not associated with an O-ring seal;
C) implement a data partitioning algorithm of the image obtained at the end of step B) to identify the O-ring seal.

Generally speaking, a convolutional neural network has interesting properties for processing images. Indeed, it presents in particular properties of spatial invariance (the network is capable of classifying the image whatever the position of the search object in the image) and invariance to geometric distortion (the network is capable to classify whatever the orientation or scale of the object sought in an image).

In the particular context of the invention, the convolutional neural network only has a single input channel (so it only takes a single image) due to the fact that the image of the technical plan to be processed during the step A) is a black and white image.

Furthermore, also in the particular context of the invention, the convolutional neural network only has a single output channel. At the end of step A), we therefore only recover one image.

The implementation of steps B) and C), applied after step A) from which only a single channel (therefore a single image) emerges, allows rapid processing. This processing is in particular faster than a classic solution consisting of implementing a “fully connected” layer (“fully connected” according to Anglo-Saxon terminology) at the output of the convolutional neural network, which generally includes numerous output channels .

The thresholding carried out in step B) makes it possible to obtain an image where a pixel associated with an O-ring seal is represented in black and a pixel not associated with an O-ring seal is represented in white (classification).

The object of step C) is then to recognize the O-ring seal from the pixels thus classified in step B).

The data partitioning algorithm implemented in step C) is advantageously the DBSCAN algorithm (acronym for “Density-Based Spatial Clustering of Applications with Noise” according to Anglo-Saxon terminology). However, we could use another algorithm, for example the OPTICS algorithm (acronym for “Ordering Points To Identify the Clustering Structure” according to Anglo-Saxon terminology) or the agglomerative pooling algorithm (better known by its Anglo-Saxon terminology). -Saxon “Agglomerative clustering”). For the application considered in the context of the invention, which should be associated with an image size that can typically be encountered, it is possible in particular to use DBSCAN software with the following hyperparameters: a distance of 15 (counted in pixels) and a minimum number of points of 20.

With the implementation of steps A), B) and C), the identification of an O-ring is particularly quick and efficient (with no or extremely low probability of error).

The image processing carried out during step A) by the convolutional neural network can be carried out in different ways.

However, a particularly well-suited design of the convolutional neural network for the identification of an O-ring seal, ensuring both accuracy and reasonable processing time, considers a step A) carried out according to the following sub-steps:
a) process the image, from the single input channel, in a first layer of the convolutional neural network, according to the following sub-steps:
- perform a convolution with hyperparameters: a number of kernels, defining a number of output channels for the first layer, between two (2) and eighty (80), a kernel size of at least thirteen (13 ), a step worth one (1) or two (2) and a margin chosen to keep the image size to the nearest step value,
- perform a correction with an activation function,
- perform a pooling with a given kernel size, advantageously between five (5) and twenty-one (21), a step between two (2) and four (4), and a margin chosen to keep the image size to the nearest step value;
each image then having a reduced size;
b) processing the images from the output channels of the convolution of the previous sub-step in a second layer of the convolutional neural network, according to the following sub-steps:
- perform a convolution with as hyperparameters: a number of kernels, defining a number of output channels for the second layer, of at least eighty-one (81), a kernel size between three (3) and twenty -and-one (21), a step worth one (1) or two (2) and a margin chosen to keep the image size to the nearest step value,
- perform a correction with an activation function,
- perform a pooling with a given kernel size, advantageously between five (5) and twenty-one (21), a step between two (2) and four (4), and a margin chosen to keep the image size to the nearest step value,
each image then having an even smaller size;
c) process the images from the output channels of the convolution of the previous sub-step in another layer of the convolutional neural network with the following step:
- perform a convolution with as hyperparameters: a kernel number of unity value defining a single output channel for this other layer, a kernel size of unity value, a step of unity value and a zero margin to preserve the image size .

Substeps a), b) and c) are implemented by computer.

With these sub-steps, the number of layers remains very limited and as a result, the parameters to learn are fewer.

Convolution with a neural network implements several hyperparameters such as the number of kernels (called “kernel” according to Anglo-Saxon terminology), the size of the kernel (called “kernel size” according to Anglo-Saxon terminology), the step (named “stride” according to Anglo-Saxon terminology), the margin (called “padding” or sometimes “zero padding” according to Anglo-Saxon terminology). This convolution is sometimes called “Conv2D”.

The convolution of sub-step a) (first layer) uses a number of kernels, which corresponds to the number of output channels of the first layer, between 2 and 80. This choice makes it possible to limit the volume of data to be processed by the following. Advantageously, a number of cores could be provided between 10 and 64, more advantageously between 18 and 64, and in particular between 18 and 48 on the one hand to further limit the volume of data to be processed (upper limit) and on the other hand to maintain a reasonable number of cores (lower limit) and not lose processing quality.

This convolution is also carried out with a minimum or even relatively large kernel size, in this case at least 13, advantageously at least 15, preferably at least 21 and even more preferably at least 31. The advantage of a large kernel size is to be able to detect a wide range of image characteristics (e.g. O-ring seal or other geometrically close element such as a circle, a circle associated with straight lines, etc.) from the first layer. We can therefore also consider implementing a limited number of layers to achieve the desired detection in a reasonable processing time.

This convolution is also carried out with a step (non-zero) of at most two, and advantageously of unity value, so as not to lose spatial precision.

The choice of the margin, designated P, depends on the size of the core, designated K, and the value of the step, designated S.

Generally speaking, the hyperparameters P, K and S are linked to the size of the image at the input of the first layer, designated W _I , and to the size of the image (for a given output channel) at the output of the first layer W _O , by the following relation:

W _O = 1 + (W _I – K + 2P)/S (R1)

Saying then that the margin is chosen to preserve the image size to the nearest step value translates mathematically into the condition (C1): W _O = W _I /S. In this case, the relation R1 is reduced to the following relation R2:

S – K + 2P = 0 (R2)

In the advantageous case where the step S is fixed at the unity value (S = 1), then the size of the output image of the first layer is identical to the size of the input image of the first layer, and the relation R2 can be rewritten in the form of the following relation R3:

P = (K -1)/2 (R3)

Thus, and as an application of the relation R3, if the kernel size is chosen at K = 25, then the margin is fixed at P = 12 and if the kernel size is chosen at K = 31, then the margin is fixed at P = 15.

The activation function can be diverse.

As non-limiting examples, we can choose a hyperbolic tangent function, a sigmoid function or even a non-saturating function. The non-saturating activation function, also known as the Rectified Linear Unit (also referred to as “ReLU” for “Rectified Linear Unit” according to Anglo-Saxon terminology), and expressed in the form [Maths 1] is however advantageous because it allows particularly rapid and precise processing, in particular because it avoids convergence problems.

Concerning the pooling sub-step (called “pooling” according to Anglo-Saxon terminology), a step of at least two ensures a reduction in the image size (measured size in pixels*pixels), in accordance with the application of the relationship R1 with the condition C1. However, at the same time, it is advisable not to reduce this image size too much as this would result in a loss of resolution for the detection of the shape to be identified (arc of a circle, etc., representing the desired O-ring). This is why the step is at most four.

This pooling can be carried out according to different schemes.

By way of non-limiting examples, it is possible to carry out average pooling (“average pooling” according to Anglo-Saxon terminology) for which the output is the average value of the input values (the number of input values is defined by the number of cores chosen), or a pooling implementing a max function for which the output is the maximum value of the input values. It is advantageous to implement a max function (this function is better known under the designation “MaxPool2D” in English terminology) because it favors strong activations. Its use is also particularly advantageous when the correction by activation function is made with the ReLU function.

The convolution of sub-step b) (second layer) is carried out with a higher number of kernels than for sub-step a) (first layer), in this case at least 81, advantageously at least least 100 and preferably at least 120. Indeed, due to the pooling operation carried out during substep a), the image size has been reduced and we can therefore increase the number of cores during substep b), that is to say increase the number of output channels of the second layer, without this being done at the expense of calculation time. However, it is the number of output channels which makes it possible to carry out the search for characteristic areas of the image which will make it possible to solve the classification task.

In parallel, we use a limited maximum kernel size, in this case at most 21 (which corresponds to a kernel whose size is at most 21 pixels * 21 pixels), but which is advantageously at most 15 , and preferably at most 9.

It should be noted that in an advantageous case, the convolution of sub-step a) provides a number of kernels that is neither too small nor too large, for example between 18 and 64, with a relatively high kernel size, for example d at least 21 and at the same time, the convolution of sub-step b) provides an advantageously smaller kernel size, for example at most 15 with the consequence of an increase in the number of kernels of at least 81 or more as previously stated.

This convolution nevertheless remains carried out with a step (not zero) of at most two, and preferably of unity value, so as not to lose spatial precision.

Finally, the margin also remains chosen to maintain the size of the image to the nearest step value, in accordance with the relation (R1) and the condition (C1) provided previously.

For this substep b), the comments relating to correction and pooling made for the convolution of substep a) are applicable.

Finally, it should be noted that step A) may optionally include an additional sub-step, between sub-step b) and sub-step c), consisting of processing the images from the output channels of the convolution of sub-step b) (second layer) in a third layer of the convolutional neural network, according to the following sub-steps:
- perform a convolution with hyperparameters: a number of kernels between 2 and 80, a kernel size of at most 15, a step of unit value or two and a margin chosen to keep the image size at the value of not near,
- perform a correction with an activation function,
- perform a pooling with a given number of cores, advantageously between three (3) and eleven (11), a step of between two and four, and a margin chosen to maintain the image size to the nearest value of the step .

This additional sub-step is also implemented by computer.

As we can see, this additional sub-step (third layer) limits, compared to sub-step b) (second layer) the number of cores, that is to say the number of output channels, in l occurrence between 2 and 80. Advantageously, we can provide a number of cores between 10 and 64, more advantageously between 18 and 64, and in particular between 18 and 48. This is also done without requiring a size of important kernel since this is clearly limited to 15. Advantageously, the kernel size will also be at most 9, or even at most 5. In all cases, the kernel size is non-zero.

This additional convolution nevertheless remains carried out with a step (non-zero) of at most two, and preferably of unity value, so as not to lose spatial precision.

Finally, the margin also remains chosen to ensure that the third layer, when implemented, maintains the image size at the nearest step value, in accordance with the relationship (R1) with condition (C1) previously provided.

For this additional convolution, the comments relating to the correction and pooling made for the convolution of substep a) are applicable.

Optionally but advantageously, step A) can provide, before implementing sub-step a), to provide an image of the technical plan coming from processing by an image data partitioning algorithm. This makes it possible to select an image forming only part of the technical plan including areas of interest, i.e. areas likely to contain an O-ring seal (this amounts to segmenting the initial image to finally work only on the relevant views of this image). For this purpose, the different examples of partitioning algorithms cited above can be used, in particular the DBSCAN algorithm. For the application considered in the context of the invention, which should be associated with an image size that can typically be encountered, one can for example, and to establish ideas, implement a DBSCAN algorithm with a distance of 25 (counted in pixels) and a minimum number of points of 10. The values of the hyperparameters must of course be adapted according to the effective size of the image to be processed.

Thus, prior to step A) of the method according to the invention, the following steps implemented by computer (preprocessing) can be provided:
- implement a data partitioning algorithm on the image of the technical plan;
- define an image of the technical plan, of limited size and including at least one area of interest likely to contain an O-ring seal.

In order for the convolutional neural network to function correctly, it must first be trained to recognize an O-ring seal on a technical level. However, on a technical level, an O-ring seal, generally in section, is in the form of a combination of a hatched circle and straight lines extending from said lines.

We have thus represented on the the typical shape of an O-ring seal on such a technical level.

Advantageously, the convolutional neural network used in step A) is a convolutional neural network having been trained with at least 10,000 images comprising a representation of at least one O-ring seal as represented on there , possibly without hatching, the latter having no impact on the definition of the contour of the O-ring.

These images can be images from real technical plans and/or from technical plans (or more generally images) constructed ad hoc for the purposes of learning.

Finally, it should be noted that the invention also relates to a computer program product downloadable from a communications network and/or recorded on a computer-readable medium, comprising instructions for executing a method according to the invention when the computer program product is executed by a computer system.

Example of production

There schematizes the process according to the invention in a particular embodiment.

The process was implemented with the PYTHON language.

We took into consideration an image of any technical plan, of size 9000 pixels*6000 pixels.

Preprocessing of this image was carried out by computer to obtain an image of limited size and comprising at least one area of interest likely to contain an O-ring seal. This processing was carried out by the DBSCAN algorithm with hyperparameters: 25 (named “eps” under PYTHON) for the distance and 10 for the minimum number of points (named “min_samples” under PYTHON). The image thus obtained is the one shown on the left on the , of size 1600pixels*1400pixels. In this image, we can note the presence of two zones each likely to be assimilated to an O-ring seal. In other words, this image appears to have two O-ring seals.

This image was then introduced into the convolutional neural network to implement step A) of the method according to the invention. The convolutional neural network shown on the was trained with over 10,000 images featuring a representation of at least one O-ring.

Step A) of the method according to the invention was then implemented on a computer with this last image and with the “PyTorch” library of the PYTHON software, according to the following algorithm:

In this algorithm, there are 4 layers in the convolutional neural network, the last being used to define a single output channel.

In each layer of the convolutional neural network, a convolution is implemented, processing by the “ReLU” activation function then, pooling with the “MaxPool2D” function.

If we take for example the first layer of the convolutional neural network (sub step a) of the method according to the invention), and we are interested in the convolution itself, we note the following points.

The first number (i.e. 1) defines the number of input channels. We can also observe on the only one image (in black and white) is taken into consideration.

The second number (i.e. 64) defines the number of output channels (number of cores) of the first layer.

The third number (i.e. 31) defines the size of the kernel (i.e. 31 pixels*31 pixels).

The fourth number defines the step (i.e. the unit value)

The fifth number defines the margin (namely 15, note that 15*2+1 = 31 thus verifying the relationship (R3) provided previously).

Continuing with the first layer of the convolutional neural network, and focusing on the pooling sub-step, we note that the 3 numbers concerning the hyperparameters, respectively the kernel size (namely 15 , defining a kernel of size 15pixels*15pixels), the step (namely 2, thus dividing the size of the image by 4, which can be verified on the where the image having undergone a first convolution, in this example from 500 pixels*500 pixels passes to 250 pixels*250 pixels at the level of the second layer).

Taking into account the above explanations, we can thus understand which parameters are implemented for the other layers of the convolutional neural network ( ^2nd layer, ^3rd layer and last layer).

At the output of the convolutional neural network, therefore at the end of step A), we obtain an image expressing a probability that a given pixel belongs to the “O-ring seal” class.

Step B) then makes it possible to make this information binary, with thresholding.

Step C) finalizes the identification of the O-rings, in this example embodiment, with the DBSCAN algorithm and the following hyperparameters: 15 for the distance and 20 for the minimum number of points.

The data is saved in a table reproduced below:

From this table, we can see the identification of two O-rings, with their coordinates (x_origin; y_origin) in the image and their respective dimensions (“width” for width and “height” for height, in pixels).

Claims (17)

Method for identifying an O-ring seal on a technical plan, said method comprising, from an image of a given size of the technical plan, the following computer-implemented steps:
A) processing said image in a convolutional neural network comprising only a single input channel and a single output channel, step A) comprising the following computer-implemented sub-steps:
a) processing said image, from the single input channel, in a first layer of the convolutional neural network, according to the following sub-steps:

• perform a convolution with hyperparameters: a number of kernels, defining the number of output channels of the first layer, between two and eighty, a kernel size of at least thirteen, a step worth unity or two and a margin chosen to keep the image size to the nearest step value,
• perform a correction with an activation function,
• perform a pooling with a given kernel size, advantageously between five and twenty-one, a step of between two and four, a margin chosen to maintain the image size to the nearest step value;

each image then having a reduced size;
b) processing the images from the output channels of the convolution of the previous sub-step in a second layer of the convolutional neural network, according to the following sub-steps:

• perform a convolution with as hyperparameters: a number of kernels, defining a number of output channels for the second layer, of at least eighty-one, a kernel size between three and twenty-one, a step worth the unit or two and a margin chosen to keep the image size to the nearest step value,
• perform a correction with an activation function,
• perform a pooling with a given kernel size, advantageously between five and twenty-one, a step of between two and four, and a margin chosen to maintain the image size to the nearest step value,

each image then having an even smaller size;
c) process the images from the output channels of the convolution of the previous sub-step in another layer of the convolutional neural network with the following step:
perform a convolution with as hyperparameters: a number of kernels of unity value defining a single output channel for this other layer, a kernel size of unity value, a step of unity value and a zero margin to preserve the image size;
B) processing the image obtained from the single output channel of the convolutional neural network to distinguish, by thresholding, a pixel of the image associated with an O-ring from a pixel not associated with an O-ring;
C) implement a data partitioning algorithm of the image obtained at the end of step B) to identify the O-ring seal. Method according to claim 1, in which the number of kernels taken into consideration for the convolution of substep a) is between 10 and 64, advantageously between 18 and 64. Method according to one of the preceding claims, in which the kernel size taken into consideration for the convolution of substep a) is at least 21, advantageously at least 31. Method according to one of the preceding claims, in which the number of kernels taken into consideration for the convolution of substep b) is at least 100, advantageously at least 120. Method according to one of the preceding claims, in which the size of the kernel taken into consideration for the convolution of substep b) is at most 15, advantageously at most 9. Method according to one of the preceding claims, in which:

• the step taken into consideration to carry out the convolution of sub-step a) is of unity value so that the margin taken into consideration for this same convolution preserves the image size; and or
• the step taken into consideration to carry out the convolution of sub-step b) is of unity value so that the margin taken into consideration for this same convolution preserves the image size.

Method according to one of the preceding claims, in which the activation function chosen to carry out the correction in sub-step a) and/or in sub-step b) is a non-saturating activation function, also called Unit Rectified Linear (ReLU). Method according to one of the preceding claims, in which the pooling of substep a) or substep b) is carried out by means of a max function (MaxPool2D). Method according to one of the preceding claims, in which the data partitioning algorithm implemented in step C) is the DBSCAN algorithm. Method according to one of the preceding claims, in which step A) comprises an additional sub-step implemented by computer, between sub-step b) and sub-step c), consisting of processing the images resulting from the output channels of the convolution of sub-step b) in a third layer of the convolutional neural network, according to the following sub-steps:

• perform a convolution with hyperparameters: a number of kernels between 2 and 80, a kernel size of at most 15, a step of unit value or two and a margin chosen to keep the image size at the step value close,
• perform a correction with an activation function,
• perform a pooling with a given number of cores, advantageously between three and eleven, a step of between two and four, and a margin chosen to maintain the image size to the nearest value of the step.

Method according to the preceding claim, in which the number of kernels taken into consideration to carry out the convolution of the additional sub-step is between 10 and 64, advantageously between 18 and 64. Method according to one of claims 10 or 11, in which the size of the kernel taken into consideration to carry out the convolution of the additional sub-step is at most 9 and advantageously at most 5. Method according to one of claims 10 to 12, in which the step taken into consideration to carry out the convolution of the additional sub-step is of unity value so that the margin taken into consideration for this same convolution preserves the image size . Method according to one of claims 10 to 13, in which the activation function of the additional sub-step chosen to carry out the correction is a non-saturating activation function, also called Rectified Linear Unit (ReLU). Method according to one of Claims 10 to 14, in which the pooling of the additional substep is carried out by means of a max function (MaxPool2D). Method according to one of the preceding claims, in which step A) comprises the following computer-implemented sub-steps, before implementing step a):

• implement a data partitioning algorithm on the technical plan image;
• define an image of the technical plan, of restricted size and including areas of interest likely to contain an O-ring seal.

Computer program product downloadable from a communications network and/or recorded on a computer-readable medium, comprising instructions for executing a method according to one of the preceding claims when the computer program product is executed by a computer system.