Method and system for checking tyres

Abstract

Method for checking tyres, comprising: providing an initial image (100) representative of a finished tyre, said initial image (100) comprising first pixels (11), representative of plain or decorated first background areas (21), and second pixels (12), representative of second areas (22) containing writings and / or logos; applying a segmentation algorithm to said initial image (100), thereby obtaining a corresponding segmented image, wherein said segmentation algorithm is based on a first neural network. The first neural network is trained to: associate said first pixels (11) of said initial image (100) with a first category; associate said second pixels (12) of said initial image (100) with a second category; associate pixels other than said first pixels (11) and said second pixels (12) with said first category. The method further comprises: selecting, starting from said segmented image, the pixels belonging to said first category, thereby obtaining a processed image; applying an inpainting algorithm to said processed image, thereby obtaining a resulting image. In said resulting image, those parts corresponding to the pixels not belonging to the first category are filled as a function of the pixels belonging to the first category and adjacent to said parts corresponding to the pixels not belonging to the first category. The method further comprises applying a first anomaly detection algorithm to said resulting image, said first anomaly detection algorithm being based on a second neural network trained to recognize any anomalies with reference to images representative of said first areas (22) with substantially no defects; generating a first notification signal if said first anomaly detection algorithm identifies an anomaly in said resulting image.

Description

[0001] " METHOD AH D SYSTEM FOR CHECKING TYRES"

[0002] DESCRIPTION

[0003] The present invention relates to a method for checking tyres.

[0004] The present invention further relates to a system for checking tyres. A tyre for vehicle wheels generally comprises a carcass structure including at least one carcass ply having respectively opposite end flaps in engagement with respective annular anchoring structures, generally referred to as "bead cores", integrated into the regions usually identified as "beads", the inside diameter of which substantially matches a so-called "fitting diameter" of the tyre for fitting it onto a respective rim, The tyre also comprises a crown structure including at least one belt strip located in a radially external position relative to the carcass ply, and a tread band which is radially external to the belt strip. Between the tread band and the belt strip(s) a so-called "underlayer" of elastomeric material may be interposed, the properties of which are suitable for providing a stable union between the belt strip(s) and the tread band, In addition, respective sidewalls of elastomeric material are applied to the side surfaces of the carcass structure, each extending from one of the side edges of the tread band to the respective annular bead anchoring structure. In tyres of the "tubeless" type, the carcass ply is internally coated with a layer of elastomeric material, preferably a butyl-based one, commonly referred to as "liner", which has air tightness properties and extends from one bead to the other.

[0005] A "decorated area" is a tyre area where one or more graphic elements are printed (embossed and / or debossed), which do not have a definite semantic meaning. Some examples of decoration are knurling and stippling.

[0006] " Inpainting operation" or "inpainting algorithm" refer to an operation or an algorithm whereby, given an initial image, a portion of said initial Image is filled with filling pixels, defined on the basis of pixelsof said initial image adjacent to that position. For example, based on said adjacent pixels, a prediction algorithm is applied in order to define said filling pixels.

[0007] " Adjacent pixels" that are adjacent to a portion of the image are pixels which, within said image, are distant from said portion by an order of magnitude corresponding to a maximum dimension of said portion. Preferably, said distance is shorter than or equal to said maximum dimension.

[0008] A "mould drawing" is a graphics file representative of the shape and dimensions of the mould. The mould drawing may be a graphics file containing the mould design in its final version, which is used for producing the mould, e.g. a CAD (Computer Aided Design) graphics file.

[0009] The expression "images having the same dimensions" means that, given a first image and a second image, the first image has dimensions, measured as a number of pixels, which are equal to those of the second image. For example, when rectangular images are involved, the first image has a first side of a length defined by a first number of pixels, and a second side of a length defined by a second number of pixels; the second image has a first side of a length defined by the first number of pixels, and a second side of a length defined by the second number of pixels.

[0010] The tyre moulding and curing process has a dual purpose of giving the tyre its final mechanical / structural characteristics and forming a series of symbols on the tyre sidewalls.

[0011] Such symbols may represent technical and / or commercial information, such as, for example: manufacturer's name / logo, tyre size, tyre type, intended season, possible certifications, etc.

[0012] Typically, the outcome of the process of moulding and curing the sidewalls of a tyre is verified by a human operator who, through a visual inspection, checks whether there are any anomalies / defects.Document EP 2500686 describes a method and a device for inspecting the shape of a tyre. An original sample image is provided, which is a two-dimensional image of the sidewall of a sample tyre. Boundary lines, which are contours of uneven marks, are detected in the original sample image. A mask Image showing the positions of said boundary lines is then generated. Thereupon, a height offset image is generated, which shows the height of the uneven marks. This is done by classifying the height of regions in the sample original image which remain after excluding regions corresponding to the positions of the boundary lines shown in the mask image, using discrete height threshold values. An inspection image is acquired, which is a two-dimensional image of the sidewall of a tyre to be inspected. By removing the uneven marks from the inspection image, an image with no uneven marks is created. The latter is generated by subtracting the height offset image from the inspection image. Finally, based on the image with no uneven marks, a shape defect on the sidewall of the tyre being inspected is detected.

[0013] Document WO 2017 / 017371 describes a method for inspecting a surface of a tyre, wherein the following steps are implemented: an image of the texture of the sidewall of a reference tyre is captured using an acquisition system, and the obtained data are transmitted to a processor; an operator, via interaction with said processor, parameterises the main characteristics and obtains a map of the sidewall of the reference tyre by cutting the image of the surface of the sidewall of the reference tyre into a plurality of separate zones of interest; a specific control and registration algorithm is assigned to each of the zones of interest; then, an image of a tyre to be inspected is captured automatically; after image pre-processing has been carried out, the map is superposed on the image of said tyre to be inspected, and the control and registration algorithms specific to each of the zones of interest of the tyre are implemented in order to determine the conformity thereof.The Applicant has verified that manual inspections may not prove fully effective and require non-negligible execution times, Such execution times may also be incompatible with an in-line process designed to operate with substantially no interruption.

[0014] The Applicant has also observed that the systems described in documents EP 2500686 and WO 2017 / 017371, while they are based on techniques that are more advanced than visual inspections, are not, however, fully satisfactory.

[0015] In particular, the technique described in EP 2500686 does not appear to be highly reliable, since it is based on an image subtraction operation, which is typically affected by criticalities and cannot always give precise and accurate results. Moreover, EP 2500686 teaches to analyze the whole sidewall of the tyre by using a single anomaly detection technique (or, anyway, no information is given as to the use of different techniques). This poses an additional limitation in terms of accuracy and effectiveness, because areas with different morphological characteristics (e.g. writings and a substantially plain background) typically react differently to the same analysis algorithm, so that, depending on the specific algorithm selected, defects in some areas will be detected with more difficulty, and hence with less reliability, than defects in other areas.

[0016] The system described in WO 2017 / 017371 envisages to divide the image to be analyzed (representing a tyre under examination) into smaller areas, and to apply different algorithms to the different areas. However, such division is, along with other operations, made manually by an operator, resulting in some of the limitations and drawbacks highlighted above for visual inspections. Furthermore, WO 2017 / 017371 provides no teachings as to the detection of any defects in the substantially plain background of the tyre (zone Z20 in Figure 5). This is clearly a significant limitation, because it may result in defects in that part of the tyre not being detected, which may in turn result in a defectivetyre, which may even have some potentially dangerous flaws, being marketed and used.

[0017] In this frame, the Applicant has felt the need for a system which is capable of automatically and effectively detecting any defects / anomalies on a tyre, while considering at least all parts of a tyre sidewall.

[0018] As perceived by the Applicant, this solution should ensure reliable and repeatable control over the outcome of the process of moulding and curing the sidewall of a tyre. This solution should also allow integration into an in-line building process, without introducing any pauses or delays in the production cycle,

[0019] The Applicant has realized that the use of different algorithms, each one dedicated to different pixel zones / categories of the tyre (e.g. background, writings / logos, etc.) can ensure a particularly effective defect / anomaly detection.

[0020] The Applicant has also realized that, in said algorithms, the use of one or more neural networks, suitably trained, can be useful to achieve such goals. More specifically, the Applicant has explored the possibility of applying an anomaly detection neural network to selected portions of images representing a tyre to be inspected.

[0021] In this regard, the Applicant has verified that neural networks are not very effective on images from which some parts have been removed, because the missing parts (or any monochromatic fillings not correlated with the context) are then interpreted as image features and may be misleading when training and using the neural network.

[0022] Therefore, the Applicant has perceived that, in order to obtain satisfactory results from a neural network, it is necessary to have the neural network generate outputs that are not affected by the removal of one or more parts.

[0023] The Applicant has thus realized that, by replacing the parts removed from an image with fillings that are similar to the context (e.g. similar tothe background) from which such parts have been removed, it is possible to have the neural network operate transparently with respect to such zones, so that it can identify any defects in the remaining zones.

[0024] Lastly, the Applicant has found that, performing a segmentation operation, wherein some parts are selected and removed (to be then processed separately), followed by an inpainting operation (for filling the removed pixels coherently with the context), the neural network dedicated to analyzing the resulting image will be able to identify any defects / anomalies correctly without being adversely affected by the segmentation operation. This is also due to the fact that, for example, the pixels of defects in writings / logos, which are not recognized as belonging to that category, are associated with the background and processed by a dedicated anomaly detection algorithm, so that the latter will be able to detect such defects.

[0025] In accordance with a first aspect, the invention relates to a method for checking tyres.

[0026] Preferably, it is envisaged to provide an initial Image.

[0027] Preferably, said initial image is representative of a finished tyre. Preferably, said initial image comprises first pixels.

[0028] Preferably, said first pixels are representative of plain or decorated first background areas.

[0029] Preferably, said initial image comprises second pixels.

[0030] Preferably, said second pixels are representative of second areas containing writings and / or logos.

[0031] Preferably, a segmentation algorithm is applied to said initial image, Preferably, through said segmentation algorithm, a corresponding segmented image is obtained from said initial image.

[0032] Preferably, said segmentation algorithm is based on a first neural network.

[0033] Preferably, said first neural network is trained to associate said firstpixels of said initial image with a first category.

[0034] Preferably, said first neural network is trained to associate said second pixels of said initial image with a second category,

[0035] Preferably, said first neural network is trained to associate pixels other than said first pixels and said second pixels with said first category.

[0036] Preferably, it is envisaged to select, starting from said segmented image, the pixels belonging to said first category, thereby obtaining a processed image.

[0037] Preferably, it is envisaged to apply an inpainting algorithm to said processed image,

[0038] Preferably, the application of said inpainting algorithm to said processed image gives a resulting image.

[0039] Preferably, in said resulting image, those parts corresponding to the pixels not belonging to the first category are filled as a function of the pixels belonging to the first category and adjacent to said parts corresponding to the pixels not belonging to the first category, Preferably, it is envisaged to apply a first anomaly detection algorithm to said resulting image.

[0040] Preferably, said first anomaly detection algorithm is based on a second neural network.

[0041] Preferably, said second neural network is trained to recognize any anomalies with reference to images representative of said first areas with substantially no defects.

[0042] Preferably, it is envisaged to generate a first notification signal if said first anomaly detection algorithm identifies an anomaly in said resulting image.

[0043] In accordance with a second aspect, the invention relates to a system for checking tyres.

[0044] Preferably, an input module is provided.

[0045] Preferably, said input module is configured for receiving or acquiringan initial image.

[0046] Preferably, said initial image is representative of a finished tyre. Preferably, said initial image comprises first pixels.

[0047] Preferably, said first pixels are representative of plain or decorated first background areas.

[0048] Preferably, said initial image comprises second pixels.

[0049] Preferably, said second pixels are representative of second areas containing writings and / or logos.

[0050] Preferably, a processing module is provided.

[0051] Preferably, said processing module is associated with said input module.

[0052] Preferably, said processing module is configured for applying a segmentation algorithm to said initial image.

[0053] Preferably, by applying said segmentation algorithm to said initial image, a corresponding segmented image is obtained.

[0054] Preferably, said segmentation algorithm is based on a first neural network.

[0055] Preferably, said first neural network is trained to associate said first pixels of said initial image with a first category.

[0056] Preferably, said first neural network is trained to associate said second pixels of said initial image with a second category.

[0057] Preferably, said first neural network is trained to associate pixels other than said first pixels and said second pixels with said first category.

[0058] Preferably, said processing module is configured for selecting, starting from said segmented image, the pixels belonging to said first category, thereby obtaining a processed image.

[0059] Preferably, said processing module is configured for applying an inpainting algorithm to said processed image.

[0060] Preferably, by applying said inpainting algorithm to said processed image, a resulting Image is obtained.Preferably, in said resulting image, those parts corresponding to the pixels not belonging to the first category are filled as a function of the pixels belonging to the first category and adjacent to said parts corresponding to the pixels not belonging to the first category.

[0061] Preferably, said processing module is configured for applying a first anomaly detection algorithm to said resulting image.

[0062] Preferably, said first anomaly detection algorithm is based on a second neural network.

[0063] Preferably, said second neural network is trained to recognize any anomalies with reference to images representative of said first areas with substantially no defects.

[0064] Preferably, said processing module is configured for generating a first notification signal if said first anomaly detection algorithm identifies an anomaly in said resulting image.

[0065] The Applicant believes that the proposed solution makes it possible to reliably, effectively, and efficiently in terms of processing times, detect any defects / anomalies that may be present on a finished tyre, particularly on a sidewall of a finished tyre.

[0066] In one or more of the above aspects, the invention may comprise one or more of the following preferred features.

[0067] Preferably, it is envisaged to make a first comparison between said initial image, or said segmented image, and a drawing of a mould used for manufacturing said finished tyre.

[0068] Preferably, through said first comparison it is envisaged to identify which second areas are present in said initial image or said segmented image.

[0069] Preferably, it is envisaged to apply a second anomaly detection algorithm to said identified second areas.

[0070] Preferably, said anomaly detection algorithm is configured to recognize any anomalies with reference to images representative of saidsecond areas with substantially no defects.

[0071] Preferably, it is envisaged to generate a second notification signal if the second anomaly detection algorithm identifies an anomaly.

[0072] The pixels pertaining to defects in writings / logos are also subtracted from the writings / logos category (because, as aforesaid, they are associated with the background), and their absence is also detected by the anomaly detection algorithm dedicated to writings / logos, thus providing a further confirmation of the presence of a defect. In the case of defects that are present only in the background or only in writings / logos, they will be detected by the anomaly detection algorithm dedicated to the respective category.

[0073] Preferably, said second anomaly detection algorithm is selected on the basis of said first comparison.

[0074] Preferably, it is envisaged to execute an Optical Character Recognition, OCR, operation on one or more writings contained in said identified second areas, thereby obtaining an alphanumerical result.

[0075] Preferably, it is envisaged to acquire one or more text portions corresponding to said one or more writings from said drawing.

[0076] Preferably, it is envisaged to make a second comparison between said alphanumerical result and said one or more text portions.

[0077] Preferably, said second notification signal is generated also when said second comparison identifies at least one difference between said alphanumerical result and said one or more text portions.

[0078] Preferably, said resulting image has the same dimensions as said initial image.

[0079] Preferably, it Is envisaged to generate a map of defects of said finished tyre.

[0080] Preferably, said map of defects is generated on the basis of said first notification signal.

[0081] Preferably, said map of defects is generated on the basis of saidsecond notification signal.

[0082] Preferably, the anomalies reported by said first and second notification signals are shown in said map of defects in their respective positions.

[0083] Preferably, said processing module is configured for making a first comparison between said initial image, or said segmented image, and a drawing of a mould used for manufacturing said finished tyre.

[0084] Preferably, through said first comparison it is possible to identify which second areas are present in said initial image or said segmented image.

[0085] Preferably, said processing module is configured for applying a second anomaly detection algorithm to said identified second areas.

[0086] Preferably, said second anomaly detection algorithm is configured to recognize any anomalies with reference to images representative of said second areas with substantially no defects.

[0087] Preferably, said processing module is configured for generating a second notification signal if the second anomaly detection algorithm identifies an anomaly.

[0088] Preferably, said processing module is configured for selecting said second anomaly detection algorithm on the basis of said first comparison.

[0089] Preferably, said processing module is configured for executing an Optical Character Recognition, OCR, operation on one or more writings contained in said identified second areas, thereby obtaining an alphanumerical result.

[0090] Preferably, said processing module is configured for acquiring, from said drawing, one or more text portions corresponding to said one or more writings.

[0091] Preferably, said processing module is configured for making a second comparison between said alphanumerical result and said one or more text portions.Preferably, said processing module is configured for generating said second notification signal also when said second comparison identifies at least one difference between said alphanumerical result and said one or more text portions.

[0092] Preferably, said resulting image has the same dimensions as said initial image.

[0093] Further features and advantages will become more apparent in the light of the following detailed description of a preferred, but non-limiting, embodiment of the invention. Such description is provided herein with reference to the annexed drawings, which are also supplied by way of non-limiting example, wherein:

[0094] - Figure 1 shows a block diagram of a system in accordance with the present invention;

[0095] - Figures 2-6 show some images used in the present invention; - Figure 7 shows a flow chart representative of the processing steps carried out in the present invention.

[0096] With reference to the accompanying figures, numeral 1 designates as a whole a system for checking tyres.

[0097] The system 1 (Figure 1) comprises an input module 2, configured for receiving or acquiring an initial image 100 representative of a finished tyre FT.

[0098] In particular, the initial image 100 is representative of a sidewall of the tyre FT.

[0099] The initial image 100 can be obtained, for example, through an acquisition operation executed by means of one or more suitable acquisition devices 3, such as one or more cameras, included in or associated with said input module 2.

[0100] The initial image 100 can be obtained from multiple acquisitions forming, when combined together, the initial image 100. Each acquisition may show a respective sub-part of the finished tyre FT.One or more lighting devices 4, appropriately positioned and adjusted to maximize the quality of the acquired images, may be associated with the acquisition devices 3.

[0101] In particular, the acquisition devices 3 and the associated lighting devices 4 may be synchronized to progressively acquire different portions of the finished tyre FT (in particular, of at least one sidewall of the finished tyre FT), so as to make up the initial image 100.

[0102] The system 1 further comprises a processing module 5, associated with the input module 2,

[0103] The processing module 5 may consist of one or more processors associated with one or more respective memory areas and configured to execute the operations described below.

[0104] For example, in order to obtain the initial image 100, the processing module 5 may apply one or more compensation algorithms so as to compensate for any deformations (whether small or big) of the finished tyre FT. It is in fact necessary to consider that, during the acquisition process, the finished tyre FT is not typically mounted on a rim and, therefore, is not inflated. For example, the finished tyre FT may be lying on a support table (not shown). As a consequence, the finished tyre FT may not have a perfectly regular shape, and this may lead to obtaining an initial image not showing its exact appearance (e.g. writings, logos, etc.). By way of example, one may use the compensation technique described in international patent application WO 2015 / 097635 A1 in the name of the present Applicant.

[0105] Preferably, the initial image 100 is a grey-scale image,

[0106] The initial image 100 (Figure 2) comprises first pixels 11 and second pixels 12.

[0107] The first pixels 11 belong to plain or decorated first background areas 21.

[0108] The second pixels 12 belong to second areas 22 containing writingsand / or logos.

[0109] The writings and the iogos represent technical and / or commercial information, such as, for example: manufacturer's name / iogo, tyre size, tyre type, intended season, possible certifications, etc.

[0110] Once the initial image 100 has been obtained, a segmentation algorithm 1000 is applied thereto.

[0111] The segmentation algorithm is based on a first neural network performing the task of associating each pixel of the initial image 100 with a category.

[0112] The categories to be used and the associations between the pixels and such categories are defined while training the first neural network.

[0113] In more detail, an untrained neural network is first provided, which is generally configured to execute segmentation operations.

[0114] Said untrained neural network is then trained with a training dataset comprising: (i) a plurality of training images, similar to said initial image 100; (ii) for each one of such training images, a corresponding mask, i.e. a corresponding annotated image, in which each pixel of the training image is associated with a category.

[0115] By inputting this training dataset to the untrained neural network, the latter is progressively made to learn which categories have to be taken into account, and establishes a relationship between the pixels and the corresponding categories.

[0116] Segmentation neural networks of this known type are described, for example, in scientific papers accessible through the following ULRs:

[0117]

[0118] Segmentation

[0119]

[0120] and [1703.06870] Mask R-CNN (last access: December 10, 2024).

[0121] In order to obtain the first neural network, at least a first category Cl and a second category C2 are defined during the training phase.

[0122] The pixels belonging to plain or decorated background areas areassociated with the first category Cl.

[0123] The pixels belonging to areas containing writings and / or logos are associated with the second category C2.

[0124] In one embodiment, a more elaborate segmentation is provided: for example, the first category Cl may be divided into two or more sub-categories (e.g. plain zones and decorated zones) and / or the second category C2 may be divided Into two or more sub-categories (e.g. plain writings and decorated writings). At any rate, even with more complex segmentations, the general principles described and claimed herein will still apply.

[0125] At the end of the training said first neural network is obtained, which, when it receives the initial image 100 as input, will output a corresponding segmented image 200 (Figure 3).

[0126] The segmented image 200 substantially consists of the initial image 100 with the addition of metadata representative of the association between each pixel and the respective category Cl, C2.

[0127] Thus, the first pixels 11 of the initial image 100 are associated with the first category Cl, and the second pixels 12 of the initial image 100 are associated with the second category C2.

[0128] Advantageously, also those pixels which are not recognized as either first pixels or second pixels are associated with the first category Cl.

[0129] In more detail, it may happen that the initial image 100 has parts representative of quite big defects, so that the first neural network may not be able to associate the corresponding pixels to either the first category Cl or the second category C2. It is however important to handle such pixels adequately to obtain a reliable anomaly / defect detection.

[0130] For this reason, the first neural network is trained to associate any pixels of the initial image 100 other than the first pixels 11 and the second pixels 12 with the first category Cl.In practical terms, this means that the first category Cl comprises all pixels belonging to plain background zones, decorated background zones, and unrecognized zones (typically major defects).

[0131] For example, as schematized in Figures 3-4, the pixels forming the anomaly X depicted In Figure 5 are also associated with the first category Cl.

[0132] In other words, the segmentation algorithm 1000 uses a first neural network trained to discern between different pixel categories Cl, C2, which must then be processed differently to detect any defects / anomalies; when any pixels are not recognized as belonging to either one of the categories Cl, C2, they will be assigned to a predefined category, in particular to the background pixel category, i.e. category Cl.

[0133] As will become apparent below, this expedient ensures a particularly accurate and effective detection of any anomalies / defects that may be present on the finished tyre FT.

[0134] In one embodiment, to help divide the pixels of the initial image 100 into the first category Cl and the second category C2 (and, optionally, into sub-categories), it is envisaged to compare the initial image 100 with the drawing 30 of a mould used for manufacturing said finished tyre FT (i.e. the tyre represented in the initial image 100). In this way, it is possible to corroborate and / or confirm the assignment of the various pixels to the different categories / sub-categories.

[0135] Figure 6 schematically shows the drawing 30 of the mould used for manufacturing the finished tyre FT.

[0136] Subsequently, the invention includes a selection step 2000, wherein, in the segmented image 200, the pixels belonging to the first category Cl are selected, thereby obtaining a processed image 300 (Figure 4).

[0137] The processed image 300 is essentially the initial image 100 from which the pixels belonging to writings and / or logos have been removed.

[0138] An inpainting algorithm 3000 is then applied to the processed image300, thereby obtaining a resulting image 400 (Figure 5). The inpainting algorithm 3000 is based on a respective inpainting neural network.

[0139] Through the inpainting algorithm 3000, those parts which correspond to pixels not belonging to the first category Cl are filled, so that the image will have no missing parts or "holes". In other words, the resulting image 400 will have the same number of pixels as the initial image 100. Therefore, the resulting image 400 will also have the same dimensions as the initial image 100.

[0140] The filling is applied as a function of the pixels belonging to the first category Cl and adjacent to the parts that need to be filled.

[0141] Merely by way of example, if a part to be filled is surrounded by a plain background zone, then it will be filled with pixels simulating a plain background; conversely, if the part to be filled is surrounded by a decorated (e.g. knurled) zone, then it will be filled with pixels simulating a decoration that Is coherent with the neighbouring zone.

[0142] In general, the goal of the inpainting operation is to fill the zones left empty by the pixels of the second category C2, so as to avoid creating zones that might be Interpreted as anomalies / defects by the anomaly detection algorithm, which will be described below.

[0143] The Applicant observes that the inpainting neural network may be advantageously trained with images containing only textures (e.g. decorations) and plain parts, without any writing.

[0144] In more detail, the training of the inpainting neural network may be carried out by inputting, as a training dataset, images in which some parts have been removed / blackened, and by Imposing the task of filling such parts on the basis of the starting images (i.e. the images from which some parts have been removed / blackened).

[0145] Therefore, the training phase simulates the context in which the inpainting algorithm will have to work in operation, i.e. a context in which fillings will have to be applied to images without any writing / logo (aspreviously explained, after the segmentation step), which substantially only contain plain or decorated parts.

[0146] Some known inpainting neural networks are described, for example, in scientific papers accessible through the following ULRs: [1604.07379] Context Encoders Feature Learning by Inpainting and [1801.07892] Generative image Inpainting with Contextual Attention (last access: December 10, 2024).

[0147] A first anomaly detection algorithm 4000 is then applied to the resulting image 400.

[0148] The task of the anomaly detection algorithm 4000 is to detect any anomalies / defects represented In the resulting image 400.

[0149] In particular, the first anomaly detection algorithm 4000 is based on a second neural network trained to recognize any anomalies / defects with reference to images representative of the first areas 21 (i.e. the plain or decorated background areas) with no defects.

[0150] Some known anomaly detection neural networks are described in the scientific paper accessible through the following URL: MVTec AD — A

[0151]

[0152] (last access: December 10, 2024).

[0153] Preferably, the second neural network is trained by inputting substantially only images or image portions without any defect. In this way, when an image is received in operation, e.g. the resulting image 400, which contains some irregularities (or, more generally, a portion significantly differing from what has been learned to be an "image without defects"), the second neural network will generate a first notification signal NS1.

[0154] For example, the first notification signal may be representative of the presence of the anomaly X (Figure 5).

[0155] The first notification signal NS1 may be provided in the form of a visual message highlighting the zone which the second neural networkwas not able to recognize as belonging to an image without defects, and which may, as such, represent a defect.

[0156] It should be noted that, preferably, the training of the second neural network occurs through image portions (also known as "crops") containing no anomalies; such crops represent totally plain background portions, totally decorated background portions, or plain and decorated background portions combined in many different ways.

[0157] Advantageously, the second areas 22 containing writings and / or logos are treated differently than described above for the pixels belonging to the first category Cl, thus making the general anomaly / defect detection activity more effective.

[0158] In particular, a first comparison 5000 is made between the initial image 100, or the segmented image 200, and the drawing 30 of a mould used for manufacturing said finished tyre FT (i.e. the tyre represented in the initial image 100).

[0159] By way of example, the first comparison 5000 between the initial image 100 (or the segmented image 200) and the mould drawing 30 may be made as described in international patent application WO 2023 / 119086 in the name of the present Applicant.

[0160] Through said comparison it is possible to identify the second areas 22 (areas containing writings and / or logos) that are present in the initial image 100 or in the segmented image 200.

[0161] In other words, the segmentation operation substantially only separates the first background areas 21 (plain or decorated) from the second areas 22 containing writings / logos, without determining which writings / logos are actually present in the initial image 100, while the first comparison 5000 with the mould drawing 30 allows identifying which writings and / or logos are present, so that the next verification can be carried out.

[0162] A second anomaly detection algorithm 6000 is then applied to thesecond areas 22 thus identified

[0163] The second anomaly detection algorithm 6000 is configured to recognize any anomalies that may be present in said identified second areas 22.

[0164] In particular, the second anomaly detection algorithm 6000 is based on a third neural network trained to recognize any anomalies with reference to images representative of the second areas 22 with substantially no defects.

[0165] The third neural network is preferably trained only with images (or image portions) not containing any anomaiy / defect. Thus, in operation, when the third neural network is inputted an image, e.g. the initial image 100 or the segmented image 200, that contains some irregularities (or, more generally, that contains a portion significantly differing from what has been learned to be an "image without defects"), it will generate a second notification signal NS2.

[0166] The second notification signal NS2 may be provided in the form of a visual message highlighting the zone which the third neural network was not able to recognize as belonging to an image without defects, and which may, as such, represent a defect.

[0167] In one embodiment, a plurality of second anomaly detection algorithms 6000.1-6000. n are used, each one based on a respective third neural network. Each third neural network is specifically trained to identify anomalies in specific writings / logos.

[0168] In more detail, a given untrained anomaly detection neural network is subjected to a specific training for detecting anomalies in specific writings / logos.

[0169] The same untrained neural network is then also subjected to a different training, which is specific for other writings / logos.

[0170] A plurality of third neural networks are thus obtained, each one focused, as aforesaid, on specific writings and / or specific logos.Each third neural network constitutes the base for a respective second anomaly detection algorithm.

[0171] In summary, starting from one untrained neural network, different third neural networks are generated by means of different training datasets, and a respective second anomaly detection algorithm is based on each third neural network.

[0172] In one variant embodiment, the third neural networks (and hence the second anomaly detection algorithms) may be obtained from different untrained neural networks, e.g. should this be preferable to attain a higher degree of specialization for specific writings or logos.

[0173] Advantageously, after said first comparison 5000, the writings and logos that should be present on the finished tyre FT are identified; based on such identification, it will then be possible to select one or more third neural networks. In particular, the selected third neural network (or group of third neural networks) will be the one specifically trained for the writings and logos that are present in the initial image 100 (or in the segmented image 200).

[0174] Preferably, in addition to the above, an Optical Character Recognition, OCR, operation 7000 is executed on one or more writings that are present in the identified second areas 22. From a practical viewpoint, the image portion that contains a writing is subjected to OCR for the purpose of recognizing the text contained therein. This operation can be carried out for different image portions that contain writings, preferably on all image portions that contain writings. An alphanumerical result 23 is thus obtained.

[0175] A second comparison 8000 is then made, wherein the alphanumerical result 23 is compared with one or more corresponding text portions 31 acquired from the mould drawing 30.

[0176] If said second comparison 8000 identifies at least one difference between the alphanumerical result 23 (i.e. the result of the OCRoperation) and the corresponding text 31 in the mould drawing 30 (which represents the ideal result of the moulding and curing process), then the second notification signal NS2 will be generated.

[0177] In this case, the second notification signal NS2 may be a visual indication of the discrepancy detected by the second comparison 8000.

[0178] In one embodiment, a step of combining the first notification signal NS1 with the second notification signal NS2 is carried out to allow for a comprehensive evaluation of the anomalies detected.

[0179] This operation may be useful when, for example, a defect is partially present on the plain background and partially present on a writing: assume that the two anomaly portions are individually too small to be considered a defect but, when combined together, they actually constitute a defect that should be taken into account; the two anomaly portions would be individually detected by the respective anomaly detection algorithms, but the defect would probably not be considered in its real importance, and the tyre quality check would not give a sufficiently accurate result.

[0180] To this end, the processing module 5 may be configured for generating a map 500 of defects of the finished tyre FT, by combining the information contained in the first notification signal NS1 and in the second notification signal NS2.

[0181] The map 500 of defects includes all detected anomalies in their respective positions. In practice, an image similar to the initial image 100 is generated which represents, instead of writings, logos, decorations, etc., the anomalies indicated in the first and second notification signals NS1, NS2, in the positions where they were detected.

[0182] The processing module 5 can then execute a further step of checking the anomalies against the map 500, so as to confirm the actual presence and importance of potential defects.

[0183] The Applicant observes that in the main image 100 there may besome substantially circular marks, commonly referred to as "vents", which are caused by vent inserts included in the mould used for manufacturing the tyre. In one embodiment, such marks may be associated with a dedicated category during the segmentation step, and may then be subjected to a dedicated anomaly detection operation.

Claims

CLAIMS1. Method for checking tyres, comprising:- providing an initial image (100) representative of a finished tyre (FT), said initial image (100) comprising first pixels (11), representative of plain or decorated first background areas (21), and second pixels (12), representative of second areas (22) containing writings and / or logos;- applying a segmentation algorithm (1000) to said initial image (100), thereby obtaining a corresponding segmented image (200), wherein said segmentation algorithm (1000) is based on a first neural network, wherein said first neural network is trained to:associate said first pixels (11) of said initial image (100) with a first category (Cl);associate said second pixels (12) of said initial image (100) with a second category (C2);associate pixels other than said first pixels (11) and said second pixels (12) with said first category (Cl);- selecting (2000), starting from said segmented image (200), the pixels belonging to said first category (Cl), thereby obtaining a processed image (300);- applying an inpainting algorithm (3000) to said processed image (300), thereby obtaining a resulting image (400); wherein, in said resulting image (400), those parts corresponding to the pixels not belonging to the first category (Cl) are filled as a function of the pixels belonging to the first category (Cl) and adjacent to said parts corresponding to the pixels not belonging to the first category (Cl);- applying a first anomaly detection algorithm (4000) to said resulting image (400), said first anomaly detection algorithm (4000) being based on a second neural network trained to recognize any anomalies with reference to images representative of said first areas (22)with substantially no defects;- generating a first notification signal (NS1) if said first anomaly detection algorithm (4000) identifies an anomaly (X) in said resulting image (400),2. Method according to claim 1, further comprising:making a first comparison (5000) between said initial image (100) or said segmented image (200) and a drawing (30) of a mould used for manufacturing said finished tyre (FT), in order to identify which second areas (22) are present in said initial image (100) or said segmented image (200);applying to said identified second areas (22) a second anomaly detection algorithm (6000) configured to recognize any anomalies with reference to images representative of said second areas (22) with substantially no defects;generating a second notification signal (NS2) if the second anomaly detection algorithm (6000) identifies an anomaly,3. Method according to claim 2, wherein said second anomaly detection algorithm (6000) is selected on the basis of said first comparison (5000),4. Method according to claim 2 or 3, further comprising: executing an Optical Character Recognition, OCR, operation (7000) on one or more writings contained in said identified second areas (22), thereby obtaining an alphanumerical result (23);acquiring from said drawing (30) one or more text portions (31) corresponding to said one or more writings;making a second comparison (8000) between said alphanumerical result (23) and said one or more text portions (31);wherein said second notification signal (NS2) is generated also when said second comparison (8000) identifies at least one difference between said alphanumerical result (23) and said one or more text portions (31).

5. Method according to any one of the preceding claims, wherein said resulting image (300) has the same dimensions as said initial image (100).

6. Method according to claim 2 or any one of claims 3-5 when dependent on claim 2, comprising generating a map (500) of defects of said finished tyre (FT) on the basis of said first notification signal (NS1) and said second notification signal (NS2).

7. Method according to claim 6, wherein, in said map (500) of defects, the anomalies reported by said first and second notification signals (NS1, NS2) are shown in their respective positions.

8. System for checking tyres, comprising:an input module (2) for receiving or acquiring an initial image (100) representative of a finished tyre (FT), said initial image (100) comprising first pixels (11), representative of plain or decorated first background areas (21), and second pixels (12), representative of second areas (22) containing writings and / or logos;a processing module (5), associated with said input module (2) and configured for:- applying a segmentation algorithm (1000) to said initial image (100), thereby obtaining a corresponding segmented image (200), wherein said segmentation algorithm (1000) is based on a first neural network, wherein said first neural network is trained to:associate said first pixels (11) of said initial image (100) with afirst category (Cl);associate said second pixels (12) of said initial image (100) with a second category (C2);associate pixels other than said first pixels (11) and said second pixels (12) with said first category (Cl);- selecting, starting from said segmented image (200), the pixels belonging to said first category (Cl), thereby obtaining a processed image (300);- applying an inpainting algorithm (3000) to said processed image (300), thereby obtaining a resulting image (400); wherein, in said resulting image (400), those parts corresponding to the pixels not belonging to the first category (Cl) are filled as a function of the pixels belonging to the first category (Cl) and adjacent to said parts corresponding to the pixels not belonging to the first category (Cl);- applying a first anomaly detection algorithm (4000) to said resulting image (400), said first anomaly detection algorithm (4000) being based on a second neural network trained to recognize any anomalies with reference to images representative of said first areas (21) with substantially no defects;- generating a first notification signal (NS1) if said first anomaly detection algorithm (4000) identifies an anomaly in said resulting image (400).

9. System according to claim 8, wherein said processing module is configured for:making a first comparison (5000) between said initial image (100) or said segmented image (200) and a drawing (30) of a mould used for manufacturing said finished tyre (FT), in order to identify which second areas (22) are present in said initial image (100) or said segmented image (200);applying to said identified second areas (22) a second anomaly detection algorithm (6000) configured to recognize any anomalies with reference to images representative of said second areas (22) with substantially no defects;generating a second notification signal (22) if the second anomaly detection algorithm (6000) identifies an anomaly.

10. System according to claim 9, wherein said processing module (5) is configured for selecting said second anomaly detection algorithm (6000) on the basis of said first comparison (5000).

11. System according to claim 9 or 10, wherein said processing module (5) is configured for:executing an Optical Character Recognition, OCR, operation (7000) on one or more writings contained in said identified second areas (22), thereby obtaining an alphanumerical result (23);acquiring from said drawing (30) one or more text portions (31) corresponding to said one or more writings;making a second comparison (8000) between said alphanumerical result (23) and said one or more text portions (31);generating said second notification signal (NS2) also when said second comparison (8000) identifies at least one difference between said alphanumerical result (23) and said one or more text portions (31).

12. System according to any one of claims 8-11, wherein said resulting image (300) has the same dimensions as said initial image (100).

Images