System and method for automatically identifying the interface of a tire product profile

Abstract

The present invention relates to a computer-implemented method for training a model for automatically identifying positions in a tire product profile, whereby an error term is calculated relative to a label assigned in a reference of an interface to be searched in an image of the captured tire product profile, and the error term is to be used as an input to a neural network to be trained. The present invention also relates to a system (100) for automatically identifying an interface change captured in a sample image according to the disclosed method.

Description

Technical Field

[0001] The present invention relates to a method for training a model for automatically identifying material changes in a tire product profile (including the production of products for incorporation into tires). Background Art

[0002] In the field of tires, tires are required to have various performance qualities (e.g., reduced rolling resistance, increased wear resistance, comparable grip in wet and dry conditions, sufficient mileage, etc.). A tire is an object with a known geometry and typically includes several superimposed rubber layers (or "plies"), as well as a metallic or fabric fiber structure that forms the carcass to reinforce the tire structure. The properties of the rubber and the properties of the reinforcing materials are selected according to the desired final characteristics.

[0003] The tire also includes a tread that is added to the outer surface of the tire. Referring Figure 1 to, representative tire 10 has a tread 12 designed to contact the ground through tread surface 12a. Tire 10 further includes a crown reinforcement having a working reinforcement 14 and a hoop reinforcement 16, and the working reinforcement 14 has two working layers 14a and 14b. Tire 10 also includes two sidewalls ( Figure 1 one sidewall 18 is shown in), and two beads 20 reinforced with bead cores 22. The radial carcass ply 24 extends in a known manner from one bead to the other bead around the bead core. The tread 12 has a reinforcement formed by, for example, superimposed layers having known reinforcing patterns. In an embodiment of tire 10, the tire may include rubber 26 that dissipates static electricity generated during driving.

[0004] During the process of tire production, tire samples are studied to develop different tire performance qualities. These samples include tire cross-sections that show the overall cross-sectional profile of the tire product (or "product profile") and the cross-sectional profile of each layer used to construct the final tire product ("layer profile"). These cross-sections also show the boundaries between adjacent superimposed layers (the "boundaries" or "limits" of the layers) and whether these boundaries are correctly aligned with each other to ensure the integrity of the final tire.

[0005] There are already some solutions to measure the parameters for the correct incorporation of the control layer into the product profile. For example, U.S. Patent No. 5,865,059 discloses an apparatus that uses a hydraulic backscatter sensor to measure the thickness of non-metallic materials in the form of films and their equivalents. U.S. Patent No. 7,909,078 discloses a system that combines a laser for scanning the surface of the product profile and means for generating real-time incremental measurement data (which is a system for measuring the exterior of the product profile without the ability to measure the interior of complex profiles). European Patent EP2404136 discloses an apparatus for measuring a tire tread that includes a plurality of light-emitting means, reflecting means, and imaging means for measuring the depth of the tread.

[0006] In the field of artificial intelligence, neural networks (also known as "neural nets") are well known. They are based on being "trained" in a large number of scenarios. By adjusting the weighting coefficients during the learning phase, the performance of neural networks can be described for new scenarios presented to them. For example, neural networks such as GoogleNet, Inception, or Resnet are well known for object recognition and classification. In a specific example, the "profound AI" (or "Deep AI") algorithm is dedicated to the search for cancer tumors (see the publication "Artificial Intelligence for Digital Breast Tomosynthesis - Reader Study Results", https: / / www.icadmed.com / assets / dmm253-reader-studies-results-rev-a.pdf).

[0007] Reference Figure 2 gives, as an example, a representative architecture of an artificial neural network (also known as an "ANN"). Figure 2 The neural network NN in E includes an input layer E having one or more neurons N X (where X varies from 1 to N depending on the number of hidden layers in the network employed), and an output layer S having one or more neurons N S By an algorithm designed to change the weights of the connections in the network, the neural network allows learning from new data (e.g., performing a specific task by analyzing examples used for training).

[0008] It will be understood that the neural network NN is given as an example, and it is known that several types of neural networks can adopt different learning methods, including supervised learning (where the network NN trained on a set of labeled data self-modifies until it can obtain the desired result), unsupervised learning (where the data is not labeled so that the network can adapt to increase the accuracy of the algorithm), reinforcement learning (where the neural network reinforces positive results and sanctions negative results), and active learning (where the network requests examples and labels to improve its predictions) (see https: / / www.lebigdata.fr / reseau-de-neurones-artificiels-definition). Examples of neural networks are presented in the prior art (see the example “The NeuralNetwork Zoo”, Fjodor Van Veen, https: / / www.asimovinstitute.org / neural-network-zoo / ) (September 14, 2016).

[0009] In addition, the introduction of deep learning algorithms has improved the performance of object detection, localization, recognition, and classification. In several fields, most of the performance of imaging systems has come from their ability to process multiple images and “overlay” their interpretations. For example, the publication WO2018 / 009405 discloses an instance segmentation system that is connected to an acquisition system to perform the automatic detection and localization of anatomical objects (internal objects) in the images generated by the imaging system. This system associates very specific hardware with artificial intelligence in order to perform segmentation. The goal is not to measure anything, but to detect and identify the elements in the image. In addition, this system consists of several neural networks for localization and classification and for segmentation.

[0010] US20190287237 discloses a system for analyzing the car body (outer surface) passed by a photographic acquisition system. This system is connected to an artificial intelligence-based algorithm to analyze the reflections of the car body using histogram normalization in this system and infer the presence of defects (e.g., dents).

[0011] Therefore, in the tire field, lighting and perspective can be increased to observe the cross-section of the product from different angles in order to more precisely identify its internal contour. The detection and segmentation of products during the production process allows for the search, finding, and identification of product cross-sections and the boundaries between them. Therefore, it is important to detect inconsistencies with product cross-section parameters and signal the need for intervention before the post-execution stage of the production process.

[0012] Thus, the disclosed invention meets the need for quality measurement of plies / carcasses to ensure the performance of pneumatic products. To avoid expensive hardware solutions that may not be suitable for each factory, the proposed invention uses neural network and image-based training. This training can be adapted to each factory and automate the quality measurement of the outer and inner profiles. Summary of the Invention

[0013] The present invention relates to a computer-implemented method for training a model for automatically identifying positions in a tire product profile, characterized in that the method comprises the following steps:

[0014] - providing a system for automatically identifying an interface captured in an image of a tire product profile;

[0015] - creating a reference of the interface searched for in the captured image of the tire product profile, the interface reference including superposed layers in the product profile and the interfaces between the superposed layers indicated by the captured image;

[0016] For each in a set of tire samples obtained from one or more tire products:

[0017] - capturing an image of the tire product profile, which is performed by the system;

[0018] - analyzing the captured image;

[0019] - a neural network training step, wherein the captured image is cut into thumbnails and labeled, wherein the neural network takes thumbnails of the same size as input, and wherein the neural network outputs a corresponding image of the same size, the corresponding image representing its prediction of the position of the interface of the product profile; and

[0020] - a comparison step, during which the prediction of the position of the interface of the product profile is used to construct at least one model representing the true interface in the captured image relative to the interface predicted in the interface reference;

[0021] Thus, the images output from the neural network are compared by calculating an error term relative to the labels assigned in the interface reference, and the offset between the true interface and the predicted interface is represented by the residual between the prediction of the interface position of the product profile and the model established during the training step, and this error indicates the variation in the samples.

[0022] In some embodiments of the method, the method further comprises a cutting step of obtaining at least one sample from one or more tire products.

[0023] In some embodiments of the method, the step of capturing an image of the tire product profile comprises capturing the image under various coded illuminations.

[0024] In some embodiments of the method, the comparing step includes the step of measuring the accuracy of a neural network, during which the neural network is assigned a metric value representing the measurement of its accuracy.

[0025] In some embodiments of the method, the step of measuring the accuracy of the neural network is performed iteratively until a constant metric value greater than 0.5 is achieved.

[0026] In some embodiments of the method, the method further includes the step of feeding a cut sample into the system.

[0027] In some embodiments of the method, the step of creating the interface reference includes a neural network training step, during which the neural network takes the true position of the interface as input.

[0028] The present invention also relates to a system for automatically identifying interface changes captured in an image of a sample according to the disclosed method, characterized in that the system includes an imaging device that performs image capture, training, and comparison steps, wherein the imaging device includes a digital profile projector configured to identify a selected product profile from a sample corresponding to an automatic selection of a corresponding control program of the projector, and the projector includes at least one processor.

[0029] In some embodiments of the system, the projector includes an image capture device for capturing an image of an obtained tire sample.

[0030] In some embodiments of the system, the image capture device includes:

[0031] - A substantially flat tray having a predefined capture area;

[0032] - A camera capable of capturing an image of a sample placed in the predefined capture area; and

[0033] - A light source including one or more illuminators that serve as a light source on the sample during capture of an image of the sample.

[0034] In some embodiments of the system, the system further includes a cutting device having a cutting system for cutting a rubber product and obtaining at least one sample from one or more tire products from the cutting system.

[0035] Other aspects of the present invention will become apparent from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The nature and various advantages of the present invention will become more apparent from the following detailed description in conjunction with the accompanying drawings, in which like reference numerals throughout indicate like parts, and in which:

[0037] Figure 1 Schematic cross-sectional view showing an embodiment of a known tire.

[0038] Figure 2 Schematic showing an embodiment of a known artificial neural network.

[0039] Figure 3 Schematic diagram showing an embodiment of the automatic identification method of the present invention.

[0040] Figure 4 Schematic diagram showing an embodiment of the imaging device of the automatic identification system of the present invention.

[0041] Figure 5 and Figure 6 representative image obtained by the imaging device shown in Figure 4

[0042] Figure 7 Side cross-sectional view of a tread sample processed during the method of training the automatic identification model of the present invention. DETAILED DESCRIPTION

[0043] Reference is now made to Figure 3 and 4 in which like numerals indicate like elements, Figure 3 which depicts an automatic interface identification system (or "system") 100 of the present invention. System 100 performs a method for automatically identifying changes between interfaces captured in an image of a tire product cross-section. In the method performed, a deep learning algorithm implemented by a commercially available imaging system is utilized to obtain an image of a tire product cross-section (or "sample"), such as overlapping layers in a tire structure.

[0044] Samples processed by system 100 include Figure 1 cross-sections of tire products of the type depicted. Tires are complex rubber profile elements, which are profile elements composed of different profile elements made of different elastomeric compounds and assembled with each other. Considering that tires need to have various performances (e.g., reduced rolling resistance, improved wear resistance, comparable grip in wet and dry conditions, sufficient mileage, etc.), tires may include superimposed rubber layers, including reinforced rubber layers, which have different rubber compounds selected according to desired final characteristics.

[0045] The construction of a tire is typically described by representing its components in the meridian plane (i.e., the plane containing the tire's axis of rotation). The radial, axial, and circumferential directions refer respectively to directions perpendicular to the tire's axis of rotation, parallel to the tire's axis of rotation, and perpendicular to any meridian plane. The terms "radially," "axially," and "circumferentially" denote respectively "radially on," "axially on," and "circumferentially on" the tire. The terms "radially inward" and "radially outward" denote respectively "towards" or "away from" the axis of rotation of the tire in the radial direction.

[0046] Referring again to Figure 3 , system 100 includes a cutting device 200 from which samples are obtained from one or more tire products. The cutting device 200 includes a cutting system having a cutting device known for cutting rubber products, the cutting device including but not limited to saws, blades, cutters, and / or their equivalents. At least one rolling device (e.g., one or more conveyor shafts) conveys the tire products along a predefined rolling path to the cutting device. The rolling device may include a lateral holding device that provides a substantially linear and aligned positioning of the tire product relative to the cutting device. The rolling device conveys the samples obtained from the cutting device to a device downstream of the cutting device 200, and an automatic identification method is performed at the cutting device 200. In one embodiment of the cutting device 200, each cutting device may be coupled to a linear actuator that adjusts the position of the cut sample (e.g., ensuring relative alignment between the tire product and the cutting device to achieve optimized cutting). Each linear actuator operates in a known manner (e.g., pneumatically or hydraulically). As an example, the cutting system may include a tread profile cutting device of the type disclosed in applicant's patent FR3052388.

[0047] Referring again to Figure 3 , and further referring to Figure 4 , system 100 further includes an imaging device 300 in which an image of the sample interface is obtained. The imaging device 300 may be integrated into an already installed tire production line.

[0048] Figure 4An embodiment of a digital profile projector (or "projector") 302 installed at an imaging device 300 and including at least one processor (not shown), the processor configured to detect, locate, and segment one or more sample images captured by the projector. The term "processor" refers to a device capable of processing and analyzing data and including software for processing the data (e.g., one or more integrated circuits included in a computer, one or more controllers, one or more microcontrollers, one or more microcomputers, one or more programmable logic controllers (or "PLCs"), one or more application-specific integrated circuits, one or more field-programmable gate arrays (or "FPGAs") and / or one or more other known equivalent programmable circuits known to those skilled in the art). The processor includes software for processing the images (and the corresponding data obtained) captured by the imaging device, as well as software for identifying and classifying the product profile and the interface between the layers overlying it.

[0049] The projector 302 can be selected from commercially available optical vision systems. In Figure 4 the embodiment shown, the projector 302 is a two-dimensional (or "2D") measurement system that includes the identification of a sample of the product profile, which is selected corresponding to the automatic selection of a corresponding control program. It will be understood that the projector 302 can be selected from commercially available projectors for measuring objects (including samples obtained at the cutting device 200) and for recording the measurement results obtained (e.g., the projector 302 can be of the type provided by Ayonis Corporation, but it will be understood that other equivalent projectors can be used).

[0050] The projector 302 includes an image capture device (or "device") 304 that captures an image of a sample obtained at the cutting device 200. The projector 302 further includes a substantially flat sample tray (or "tray") 306. The tray 306 includes a predefined capture area (or "area") in which the sample is placed during the capture of the image of the sample. The image captured by the device 304 is transmitted to an image processing device (e.g., a processor), which can process, identify, and classify the product profile included in the sample. It will be understood that the position of the tray 306 can be changed in a rotatable manner relative to the device 304 to capture sample images at several different angles.

[0051] The image capture device 304 includes a camera 304a, such as a video and / or photographic camera, an infrared camera, an ultraviolet camera, a set of electromagnetic sensors capable of capturing images, and their equivalents. The image capture device 304 also includes an illumination source 304b having one or more illuminators (e.g., known programmable LEDs) to serve as a light source at the sample. The illuminators can be encoded in the device 304, or they can be pre-encoded during the neural network training method (e.g., using a computer program containing data representing illumination and images of known tire product profile interfaces). Changes in the illumination source 304b captured in the sample images obtained by the camera 304a are represented by pixels of different color intensities. For example, the obtained sample images can contain indications of reflections due to illumination changes. Thus, the images captured by the image capture device 304 indicate the overlapping layers in the product profile and the interfaces between them. The camera 304a and the illumination source 304b can be moved in an alternating or random manner to respectively adjust the objective lens and the illumination based on the parameters of the sample (e.g., based on its length) placed in the capture area of the tray 306 (see Figure 4 for the arrows A and B).

[0052] Referring again to Figure 3 and 4 , and further referring to Figure 5 and 6 , Figure 5 and Figure 6 depict representative images obtained by the projector 302.

[0053] Figure 5 and Figure 6 each depict a meridian cross-section of a sample tire product profile in the form of a tread. It will be understood that the sample for which the imaging device 300 generates an image can take the shape of another product profile.

[0054] Figure 5 depicts an image of a tire product profile in the form of a tread 400. The tread 400 is intended to contact the ground via the tread surface 400a. In the sampled tire architecture, a plurality of circumferential grooves (or "grooves") 402 are arranged in the tread 400, each circumferential groove 402 having a predefined width 402a and depth 402b (it will be understood that the width and depth of the grooves 402 can vary between one groove and another). The depth 402b of the groove 402 represents the maximum radial distance between the rolling surface 400a and the bottom surface 402c of the groove.

[0055] In Figure 5, the tread 400 is substantially continuous with the axial outer surface of the tire, the tangent of the tread is plotted in the meridian section of the tire, and at any point on such tread in the transition region of the sidewall. The first axial edge 404 is such a point where the angle β is defined as between the tangent and the axis Y - Y' (i.e., the radially outermost point). The same procedure will be followed to determine the second axial edge 406 of the tread 400. The image of the cross-section of the tread 400 captured by the projector 302 is represented by the outer contour E.

[0056] Figure 6 An image of a cross-section of a second tire product in the form of a tread 500 with a tread surface 500a is depicted, where Figure 6 The axial width of the tread surface LSR and the axial width of the off-shoulder tread surface LSR' are also depicted. In Figure 6 the sampled tire architecture shown, a plurality of grooves 502 are arranged in the tread 500, each groove 502 having a predefined width 502a and depth 502b (it will be understood that the width and depth of the grooves 502 can vary from one groove to another). In Figure 6 it, the axial edge 504 is determined by the projector 302. The radial distance is defined as between the radially outer surface of the upper covering corrugated layer (e.g., as shown by the point F in Figure 6 ) and the radially outermost point of the rib pattern 505, and this distance is perpendicular to the center of the bottom surface 502c of the groove 502. The radial distance is defined as between the radially outer surface of the radially outermost covering layer and the radially outermost point of the rib pattern 505, at the corrugation level of the covering layer (e.g., as shown by the point F in Figure 6 ). In an embodiment of a tire including a wear indicator, the radial distance can be defined as between the tread surface 500a and the radially outermost point of the wear indicator. The image of the tread cross-section 500 captured by the projector 302 is represented by the outer contour E' and the inner contour I.

[0057] The projector 302 utilizes known tools (e.g., optical, mathematical, geometric, and / or statistical tools) and programmable illumination, as well as software that allows for sample inspection, measurement result evaluation, and monitoring and use of control devices. The projector 302 includes one or more programming modes, including learning, feeding, modifying, and training neural networks. The selection of semantic segmentation over instance segmentation enables the present invention to be generalized to any product profile. As used herein, "semantic segmentation" refers to a method of assigning labels to each pixel in an image to treat multiple objects of the same category as a single entity. A typical semantic segmentation architecture can be considered an encoder network (e.g., a pre-trained classification network such as GoogleNet, Inception, Resnet, or one or more equivalent forms of such a network), followed by a decoder network. The processor can utilize ground truth data to train and / or expand the neural network to automatically detect the space where an object (e.g., a product profile) should be located and / or the rubber space surrounding it. The ground truth data is represented in the interface reference created during the automatic recognition method (described below) of the present invention.

[0058] Light projected from the illumination source 304b is reflected onto the sample, and the resulting image is captured by the projector 302. The image consists of a matrix of pixels, each pixel having a different color and brightness indicating its position in the layer profile of the sample. Software for identifying changes along the layer profile can convert the captured image into a set of 2D images, each image being the same as the image within one illumination change (or change in the layer profile) of the product profile boundary. The resulting image changes indicating one or more positions of the product profile boundary train the neural network to identify all positions of the product profile boundary. These image changes are used as inputs to the neural network, and the output of the neural network is a classification of whether the position is a profile boundary. The purpose of this algorithm is to automatically identify and indicate the outer profile of the sample and the internal interface between the profiles of the layers, and then give exact measurements of the outer profile and / or the inner profile. It will be understood that any neural network can be implemented, including but not limited to convolutional neural networks and their equivalent forms.

[0059] Fragments classified as containing changes (and potential anomalies) are overlaid on the original image. In an embodiment of the present invention, this step includes capturing multiple images, each image including a change displaced from the previous image such that all points showing the change are visible relative to the rest of the sample. Different changes can have different parameters and / or "weights" and will thus be treated differently when fed into the classification algorithm.

[0060] In some embodiments of the system 100, multiple sets of dimensional scales can be analyzed independently, thereby creating a multi-scale classifier. In these embodiments, the multi-scale classifier can cross-reference information from all classifiers and can display all fragments showing changes.

[0061] The projector processor 302 continuously trains a neural network based on the data of newly captured sample images obtained by the projector. To automatically detect the boundary between the layers of the product and / or the surrounding rubber, the projector 302 captures images (which may include videos) and acquires an image dataset of the product profile and / or the surrounding rubber from multiple product samples. Before being recorded, the image dataset can be annotated based on user input to create ground truth data. For example, in some embodiments, to assist the neural network in detecting and identifying the profile boundary and / or the surrounding rubber, the image dataset is annotated based on the knowledge of tire professionals and known variations are manually identified. Thus, the ground truth data described herein generally refers to information provided by direct observation of professionals in the field, rather than information provided by inference. They can have data from multiple sources to develop the neural network, including multiple professionals at remote locations. The feedback loop of the annotated images can be updated over time with additional ground truth data to improve the accuracy of the system 100.

[0062] The dataset of images can be divided into multiple groups (e.g., the ground truth data can be divided into groups including at least a training dataset and a validation dataset). In some embodiments, the processor can be configured to implement supervised learning to minimize the error between the output of the neural network and the ground truth data (referred to as "supervised learning"). It will be understood that other deep learning techniques can be utilized, including but not limited to reinforcement learning, unsupervised learning, etc. In some embodiments, the development of the training data can be based on a reward / punishment function such that no labeled data needs to be specified.

[0063] The neural network can be used to provide predictions for the validation data and newly captured data almost in real time. The neural network can be trained to locate and demarcate the product profile data (defined as product profile boundary data). The neural network can also be trained to locate and demarcate the layer profile data (defined as layer profile boundary data that may include surrounding rubber data). The difference between training the neural network to locate product profile objects and training the neural network to demarcate profile objects includes how to label the data and architectural details.

[0064] It will be understood that the neural network can be trained using the data of images obtained by the projector 302 from multiple samples. It will also be understood that the neural network can be trained by multiple imaging devices (including those of the type represented by the imaging device 300) that acquire profile data from multiple samples over time. For all embodiments, there may be differences in image size, intensity, contrast, and / or texture.

[0065] Therefore, the system 100 utilizes innovative methods to construct a large amount of labeled training and thus form a large enough network to effectively utilize all the acquired data.

[0066] Referring again to Figures 3 to 6 , a detailed description is given as an example of the automatic recognition method (or "process") of the present invention performed by system 100.

[0067] When starting the automatic recognition method of the present invention, the automatic recognition method includes the step of creating a reference of the interface to be searched in the images to be captured by projector 302. The interface reference created during this step includes: the expected image corresponding to the overlay in the product profile and the interface between the overlays indicated by the image captured by the projector (including the outer and inner contours of the product profile). This step can be performed before other steps in the method to feed the true position of the expected interface to the neural network by analyzing the captured images. This step includes creating a negative image of the captured image, and each negative image is a "label" or is "tagged". In an embodiment of the method, at least a part of the interface reference is created by one or more technicians in the art.

[0068] The automatic recognition method further includes a cutting step to obtain one or more tire product profile samples. This step performed by cutting device 200 of system 100 can be performed iteratively depending on the number of samples to be provided to feed the neural network. In an embodiment of the method, this step includes the step of cutting the tire in a meridian cut, which is used to determine the centers of different radial distances, grooves, and the bottom surfaces of the furrows. If it is necessary to change the cutting medium (during the cutting step or between method cycles), system 100 allows rapid retraining of the neural network.

[0069] The automatic recognition method further includes the step of capturing an image of the cut sample, which is performed by projector 302, and more specifically, by image capture device 304. This step includes the step of placing the cut sample in the capture area of tray 306 of device 304. This step further includes capturing the image under various illuminations performed by illumination source 304b of device 304. The illumination can be carried out according to the number of images to be captured with automatic focusing. In one embodiment of system 100, the system can transfer the cut sample to tray 306. In this embodiment, the step of capturing the image further includes the step of conveying the cut sample to tray 306, which includes the step of placing the sample in the capture area of the tray.

[0070] It will be understood that each captured image may include traces reflected from the darker areas of the rubber material where there is no direct incidence of light. In this case, the image consists of a matrix of pixels with different colors and brightnesses.

[0071] In one embodiment, during the image capture step, the projector 302 can automatically generate multiple high-resolution images by changing the illumination incident configuration and intensity. The intensity is intentionally changed to increase the robustness of the neural network. Instead of training the neural network on the basis of the same intensity, so that the neural network is too sensitive to possible variations such as those caused by brightness loss (due to diode aging, screen contamination, or environmental changes), but to ensure the accuracy of the neural network.

[0072] The automatic identification method of the present invention further includes the step of analyzing the images captured by the projector 302. For example, each curve represents a stripe (the curve represented by the internal profile I of the image captured in Figure 6 ). Each stripe is divided into segments. The criterion for dividing the stripe into segments can simply divide each stripe into segments of the same pixel size. Features are extracted from the geometry and curvature to provide information for the anomaly classification algorithm. The features mentioned herein may include, but are not limited to, the geometric features of the segment and / or the curvature statistics calculated for each point in the segment (including but not limited to the sum of curvatures, average value, variance, kurtosis, or a combination of these statistics). In addition, a tolerance value can be derived from the curvature, and curvatures below this tolerance value will not be considered when calculating the boundary profile position.

[0073] The automatic identification method of the present invention further includes a training step, during which the captured images are cut and labeled into even smaller thumbnails. Specifically, a desired number of thumbnails are defined, and these thumbnails are regarded as a single set. Therefore, one or more predetermined thumbnails combined with their labels (and thus "labeled") are considered training examples. Each training example consists of a set of thumbnails and their labels. Thumbnails allow a significant increase in the number of available examples, which not only provide the category (i.e., the label), but also provide additional information about the spatial position of these categories relative to the rubber area.

[0074] During the training step, the neural network takes thumbnails of the same dimension as input and outputs corresponding images of the same dimension, which are its predictions of the product profile interface position (the "output image" of the neural network). The prediction of the product profile interface is used to construct one or more models, and the one or more models represent the true interface in the captured image relative to the predicted interface in the interface reference. The deviation between the true interface and the predicted interface is represented by the calculated error, which indicates the variation in the sample (and can be a potential anomaly). The calculated error can be input into the interface reference to expand the reference (see Figure 3 ).

[0075] The automatic recognition method of the present invention further includes a comparison step, during which the images output from the neural network are compared by calculating an error term relative to the labels assigned in the reference at the interface. Then, in an optimization method that is part of this comparison step, the residual between the predicted interface position of the product profile and the model (established during the training step) is utilized. By performing stochastic gradient descent (refining more or less according to the magnitude of the error), the parameters of the neural network are modified to reduce this residual.

[0076] The comparison step includes a step of measuring the accuracy of the neural network. During this step, the neural network is assigned a value called "metric" (for example, a value ranging from 0 to 1), which represents the measurement of its accuracy. The higher the metric value, the more accurate the model (in the sense of the metric). This step can be iteratively executed until a constant metric value greater than at least 0.5 is reached.

[0077] At the end of the comparison step, the resulting model after all iterations is called the saved "trained model". The trained model allows for fine-grained inference by performing dense prediction by deriving the labels of each pixel, so each pixel is labeled with the category of the corresponding object (variation) and the surrounding area (rubber). Now, in order to predict the position of the interface in the image of the captured sample, the number of images defined during the training step can be reduced without the need to label them. Therefore, the system 100 is very flexible because it only requires the captured images to execute the method of the present invention.

[0078] Therefore, the system 100 of the present invention is based on a neural network, the basis of which will be trained according to a large number of situations (for example, images of samples) so as to be able to describe new situations that will be presented to it. On the one hand, the neural network must be told what it should recognize and then be trained ("annotated"). On the other hand, it is necessary to evaluate the performance of the neural network and present relevant samples to avoid falling into specific biases (such as overtraining), which will reduce the performance of the network and even eliminate its practicality.

[0079] Compared with known quality measurement devices (for example, through a profiler system), the automatic recognition method of the present invention is executed by the system 100 in a reduced time. For example, the disclosed method currently takes a few seconds for more than 10,000 pixels, instead of several minutes for 10 to 15 manual points (the number of which will increase as the complexity of the tire product profile interface increases).

[0080] Using the system 100, the internal and external interfaces can also be obtained immediately, and their images can be accessed by the same system using the same software. Therefore, the method of the present invention can be easily used by the internal team for all tire architectures. As an example,Figure 7 Figure 1 shows a side cross-sectional view of a sample of a tread 600 that can be processed during the method of the present invention. The depicted tread 600 incorporates a wear-resistant and road grip layer 602, a profile lining (or "sub-layer") 604, and a layer 606 of rubber for eliminating static electricity generated during driving. Also incorporated is a profile of a seam cover 608 that serves as side impact protection. In the captured image, the outer profile 600E is well shown together with the inner profiles (see the profiles 602I, profile 604I, profile 606I, and profile 606I in Figure 7 ) to enable the automation of point measurements. It will be understood that samples with other product profiles will feed the images captured of these samples into a neural network to build a model.

[0081] To achieve industrial productivity, solutions incorporating artificial intelligence can be used, which can replace the "algorithmic" description of cases that people attempt to detect through learning and instance work. Once the system gives a result, it is simpler to implement but more difficult to interpret. Through localization and detection, the disclosed invention thus eliminates the need for image preprocessing. The solution is not only highly effective but can also be flexibly adjusted according to specific circumstances if requirements or operating conditions change.

[0082] System 100 is applicable to tires made of various rubber compounds without reducing industrial productivity. Thus, the present invention takes into account the quality of the parameters measured and analyzed to ensure the quality of the tires.

[0083] The automatic identification method of the present invention can be accomplished through PLC control and can include pre-programming of management information. For example, the method settings can be associated with tire parameters, rubber material characteristics, and operating parameters of the interconnected layers.

[0084] In an embodiment of the present invention, system 100 (and / or a device incorporating system 100) can receive voice instructions or other audio data indicating, for example, to start or stop capturing an image of a sample placed on tray 306. The instructions can include a request for the current state of the automatic identification method cycle. The generated response can be presented in an auditory, visual, tactile (e.g., using a haptic interface), and / or virtual and / or augmented manner.

[0085] In all embodiments of system 100, a monitoring system can be implemented. At least a portion of the monitoring system can be provided in a wearable device, such as a mobile network device (e.g., a mobile phone, a laptop computer, a network-connected wearable device (including "augmented reality" and / or "virtual reality" devices, network-connected wearable devices, and / or any combination and / or equivalents). It will be understood that the detection and comparison steps can be performed iteratively.

[0086] The terms "at least one" and "one or more" are used interchangeably. A range expressed as "between a and b" includes the values of "a" and "b".

[0087] Although specific embodiments of the disclosed devices have been illustrated and described, it should be understood that various changes, additions, and modifications can be practiced without departing from the spirit and scope of the disclosure. Accordingly, no limitation should be imposed on the scope of the described invention other than those set forth in the appended claims.

Claims

1. A computer - implemented method for training a model for automatically identifying positions in a tire product profile, characterized in that, the method comprises the following steps: - a step of providing a system (100) for automatically identifying an interface captured in an image of a tire product profile; - a step of creating a reference of the interface searched for in the captured image of the tire product profile, the interface reference including an overlay in the product profile and the interface between the overlays indicated by the captured image; For each of a set of tire samples obtained from one or more tire products: - a step of capturing an image of the tire product profile by the system (100); - a step of analyzing the captured image; - a neural network training step, wherein the captured image is cut into thumbnails and labeled, wherein the neural network takes thumbnails of the same size as input, and wherein the neural network outputs corresponding images of the same size, the corresponding images representing its prediction of the position of the interface of the product profile; and - a comparison step, during which the prediction of the position of the interface of the product profile is used to construct at least one model representing the true interface in the captured image relative to the interface predicted in the interface reference; Thus, the images output from the neural network are compared by calculating an error term relative to the labels assigned in the interface reference, and the offset between the true interface and the predicted interface is represented by the residual between the prediction of the interface position of the product profile and the model established during the training step, and this error indicates the variation in the sample.

2. The method according to claim 1, further comprising a cutting step of obtaining at least one sample from one or more tire products.

3. The method according to claim 1 or claim 2, wherein, the step of capturing an image of the tire product profile includes the step of capturing the image under various coded illuminations.

4. The method according to claim 1 or claim 2, wherein, the comparison step includes the step of measuring the accuracy of the neural network, during which the neural network is assigned a metric value representing the measurement of its accuracy.

5. The method according to claim 4, wherein, the step of measuring the accuracy of the neural network is performed iteratively until a constant metric value greater than 0.5 is expected.

6. The method according to claim 1 or claim 2, further comprising the step of conveying the cut sample to the system.

7. The method according to claim 1 or claim 2, wherein, the step of creating the interface reference includes the step of training a neural network, during which the neural network takes the true position of the interface as input.

8. A system (100) for automatically identifying interface variations captured in an image of a sample, according to the method of any one of claims 1 to 7, characterized in that, The system includes an imaging device (300), and the imaging device (300) performs steps of capturing an image, training, and comparison. Among them, the imaging device includes a digital profile projector (302), and the digital profile projector (302) is configured to identify a selected product profile from a sample corresponding to an automatic selection of a corresponding control program of the projector. The projector includes at least one processor.

9. The system (100) according to claim 8, wherein, the projector (302) includes an image capture device (304), and the image capture device (304) captures an image of the obtained tire sample.

10. The system (100) according to claim 9, wherein, the image capture device (304) includes: - a substantially flat tray (306) having a predefined capture area; - a camera (304a) capable of capturing an image of a sample placed in the predefined capture area; and - an illumination source (304b) including one or more illuminators for serving as a light source on the sample during capturing an image of the sample.

11. The system (100) according to any one of claims 8 to 10 further includes a cutting device (200), the cutting device (200) having a cutting system for cutting a rubber product and obtaining at least one sample from one or more tire products from the cutting system.

Images