Tire defect detection method

Abstract

The present disclosure provides a reliable and fast tire defect detection method in which the accuracy of tire defect detection is improved to minimize manual inspection and ensure that industrial quality requirements are met, and the image of the tire is used as an input to a trained artificial intelligence algorithm, detecting whether there is at least one tire defect in the image by a trained artificial intelligence algorithm, marking the at least one tire defect in the image by the trained artificial intelligence algorithm, and outputting a marked image having the at least one tire defect by the trained artificial intelligence algorithm, wherein the output marking image with the at least one tire defect is used as an input for an image processing algorithm which determines and marks at least one geometric feature of the detected at least one tire defect in the marking image, wherein the image processing algorithm outputs a marking image with the determined and marked at least one geometric feature of the detected at least one tire defect, and wherein the marking image is displayed on a display.

Description

Technical Field

[0001] The present invention relates to a method for detecting tire defects, wherein an image of a tire is used as input to a trained artificial intelligence algorithm, wherein the trained artificial intelligence algorithm detects whether at least one tire defect is present in the image, the trained artificial intelligence algorithm marks the at least one tire defect in the image, and the trained artificial intelligence algorithm outputs a marked image with the at least one tire defect. Background Art

[0002] The number of vehicles is increasing worldwide, and therefore the demand for tires is also increasing. Tire quality is very important as it is the basis for safety in operating a vehicle. In addition, tire quality has a strong influence on the tire life itself. A tire basically consists of several plies, such as carcass ply, crown ply, steel belt, radial ply, etc. During the tire manufacturing process, a variety of different tire defects may occur in different tire plies. Tire defects usually appear randomly at different locations on the outside of the tire or on the inside of the tire. For example, misalignment of the ply, impurities, scratches, etc. are common types of tire defects. Due to tire defects, the quality of the tire is affected.

[0003] In order to ensure a certain tire quality or meet certain industrial quality requirements for tires, it is a mandatory process in tire manufacturing that each individual tire needs to be thoroughly optically inspected before delivery. Currently, this inspection is done manually by tire manufacturing personnel, which requires a lot of experience and can still be inaccurate, for example due to personnel lack of concentration or subjectivity. In addition, manual inspection is time-consuming and therefore expensive.

[0004] In order to overcome the inaccuracies and costs of manual inspection, different image processing methods and imaging systems are known to optically inspect the exterior of the tire, such as the outer surface or the inner surface. By using X-rays, the tire can also be optically inspected internally. Due to the random nature of tire defects, conventional image processing algorithms, such as filtering, edge or contour extraction, pixel scanning, etc., are technically difficult to implement and are usually not accurate enough, especially for small defects.

[0005] A method is described in DE102019217179A1 that combines manual tire inspection by personnel and automatic inspection using an imaging system. The imaging system includes a camera that takes images of the outside of the tire. These images are compared with reference images by an algorithm (e.g., a deep learning algorithm). Depending on the condition of the tire, the imaging system generates a positive or negative signal to indicate whether the tire passes the specified quality requirements. If the image from the camera does not match the reference image, a neutral signal is generated and instructs the personnel to manually check the image or the tire itself. After the personnel decides whether the tire passes the specified quality requirements, the decision and the corresponding images are used to train the deep learning algorithm of the imaging system. The more the deep learning algorithm is trained, the better its performance and accuracy are for grading the quality of tires.

[0006] Typically, artificial intelligence (AI) algorithms, such as machine learning algorithms or deep learning algorithms, are based on neural networks. A neural network in principle consists of an input layer, a number of intermediate (hidden) layers, and an output layer. In each layer a number of nodes or neurons are arranged, which are connected to nodes in other layers of the neural network to form a certain topology. The topology of a neural network essentially depends on the type of neural network and its intended application. Node connections are weighted, where each weight represents the strength of the connection between the two nodes it connects. A single node calculates its output based on its weighted input. The values ​​of weights and other parameters (such as biases) can be initially set and can be adjusted by training the artificial intelligence algorithm.

[0007] In tire manufacturing, it is necessary to inspect not only the outside of the tire, but also the inside. Most rubber tires have a complex inner ply structure embedded in the tire rubber. In order to inspect the inside of the tire, X-ray scanning is widely used. Methods and devices are known that are able to take X-ray images of tires. Usually, the tire is mounted on a rotating device to rotate the tire 360° around its rotation axis when the X-ray image is taken. Defects inside the tire as well as outside the tire can then be detected by manually inspecting the X-ray image, for example by displaying the X-ray image to a person on a screen.

[0008] In order to minimize the expensive manual inspection of X-ray images, artificial intelligence algorithms such as deep learning algorithms can also be used. For example, in CN108564563A or CN110660049A, a method for detecting tire defects in X-ray images using a deep learning algorithm based on a convolutional neural network (CNN) is proposed. Due to its performance, deep learning algorithms based on CNN are widely used in the field of object detection in images. Currently, deep learning algorithms based on CNNs like region-based convolutional neural network (R-CNN), fast R-CNN, You Only Look Once (YOLO) or fast R-CNN are commonly used to detect tire defects in X-ray images. In CN110335242A, an algorithm based on improved Faster R-CNN and YOLO is proposed.

[0009] In principle, if an AI algorithm is trained to a certain level of accuracy, it can take an image as input, assign weights and biases to various aspects / objects in the image (the middle layer), and interpret the objects in the image (the output). In other words, a trained AI algorithm can determine whether an X-ray image of a tire contains a tire defect.

[0010] Basically, the artificial intelligence algorithm is trained in certain steps to enable it to detect tire defects in X-ray images with a certain accuracy. In a first step, a large number of X-ray images are collected, for example in the field, in which different tire defects are present to form a representative database for training the algorithm. The collected images can be pre-processed for use in the algorithm, for example by dividing and / or cropping the images to a certain common image size. The images can also be compressed, sharpened, denoised, etc. to reduce the necessary computing power. In addition, the pre-processed images are manually marked by marking the tire defects (e.g. type and location) with appropriate marking software.

[0011] The labeled images are randomly grouped into training sets, validation sets, and test sets in appropriate proportions. For example, in CN110120036A, 70% of the labeled images are placed in the training set, 15% are placed in the validation set, and 15% are placed in the test set. In the training phase, the images in the training set are used to train the artificial intelligence algorithm and adjust the parameter values ​​of the algorithm, such as weights and biases. During the training phase, it is important to prevent the algorithm from overfitting, which means that the algorithm is well performed on the images of the training set but not on unseen images. Therefore, in the validation phase, the (unseen) images of the validation group are used to check the accuracy of the algorithm in detecting tire defects, where the value parameters of the algorithm are further adjusted according to the determined accuracy. If the determined accuracy is not sufficient compared to the predetermined accuracy, the algorithm can be trained again by the images in the training set and checked again by the images in the validation set. For example, the artificial intelligence algorithm is trained to converge, for example, until a certain accuracy according to industry quality requirements is met.

[0012] After the artificial intelligence algorithm is trained to a certain accuracy, in the testing phase, the images in the test set (also unseen) can be used to check the performance of the trained algorithm, for example, whether the tire defect detection works as expected. If the performance is sufficient, the trained algorithm can be used to detect tire defects during the tire manufacturing process. However, although these algorithms can detect the type and / or location of defects, industrial quality requirements also require the determination of the geometric characteristics of tire defects, such as size, ply position, orientation, etc., to grade tire quality. Summary of the invention

[0013] Therefore, an object of the present invention is to provide a reliable and fast tire defect detection method, in which the accuracy of detecting tire defects is improved to minimize manual inspection and ensure that industry quality requirements are met.

[0014] This object is achieved by the features of the independent claims.The invention proposes a tire defect detection method based on a combination of a trained artificial intelligence algorithm and a traditional image processing algorithm.

[0015] The output marked image with the at least one tire defect is used as input to an image processing algorithm, wherein the image processing algorithm determines and marks at least one geometric feature of the at least one tire defect detected in the marked image, the image processing algorithm outputs a marked image with the determined and marked at least one geometric feature of the at least one tire defect detected, and the marked image is displayed on a display. In addition to detecting at least one tire defect with a trained artificial intelligence algorithm, geometric features of the tire defect, such as size, ply position, orientation, etc., can also be determined in combination with an image processing algorithm. By providing detailed information on the tire defect, the tire defect can be detected more accurately, which can further improve tire quality grading and reduce manual inspection. In this way, industry quality requirements are met.

[0016] Preferably, the image of the tire is an X-ray image of the tire or an image of the tire surface. Depending on the type of image used in the method, different parts of the tire and the plies can be inspected. For example, with an image of the outer or inner surface of the tire, defects on the outer side of the tire can be determined. However, with an X-ray image, the interior of the tire can be inspected, which will be described in more detail below.

[0017] The trained artificial intelligence algorithm is advantageously an artificial intelligence algorithm based on a neural network, wherein images of tires marked with at least one tire defect, preferably preprocessed, are used to train the artificial intelligence algorithm, wherein the images are grouped into a training set and a validation set, wherein the images in the training set are used to train the artificial intelligence algorithm, wherein parameter values ​​of the artificial intelligence algorithm are adjusted and the images in the validation set are used to determine the accuracy of the artificial intelligence algorithm in detecting tire defects, and the artificial intelligence algorithm is trained when the determined accuracy meets a predetermined accuracy of the artificial intelligence algorithm in detecting tire defects.

[0018] Advantageously, the trained artificial intelligence algorithm is tested with a test set to determine the performance of the trained artificial intelligence algorithm, the testing with the test set comprising testing with tire images (preferably pre-processed) marked with at least one tire defect.

[0019] The trained artificial intelligence algorithm preferably marks the position (preferably relative to a reference coordinate system of the tire) and the type of at least one detected tire defect in the image. In this way, for example in the case where a tire has to be inspected manually, the user can determine where the at least one detected tire defect is actually located on the tire. Advantageously, the user can also know which type of tire defect is present in the tire. Thus, the marked image can also be displayed on the display.

[0020] Advantageously, the trained artificial intelligence algorithm marks the location of at least one detected tire defect in the image as a border around at least part of the detected tire defect, and the image processing algorithm crops the marked image to the border around the detected tire defect to determine and mark at least one geometric feature of the detected tire defect in the cropped marked image. By cropping the marked image, the image processing algorithm processes a smaller image. Thus, the geometric features of the tire defect can be determined in more detail and the necessary processing power can be reduced.

[0021] At least one geometrical feature of the detected tire defect describes at least a size or a ply position or orientation, preferably relative to a reference coordinate system of the tire. By providing detailed information of the tire defect, the detection of tire defects and the tire quality grading can be improved. Furthermore, the location of the detected tire defect can be described more accurately by determining the at least one geometrical feature.

[0022] The marked image with the determined at least one geometrical feature of the at least one detected tire defect is preferably used to grade the tire quality by comparing the at least one determined geometrical feature of the at least one detected tire defect with a predetermined geometrical feature. By determining the geometrical feature of the tire defect, the accuracy of the tire quality grading is improved, which reduces manual inspection.

[0023] If the trained artificial intelligence algorithm detects that no tire defects are present in the image of the tire, then advantageously the trained artificial intelligence algorithm outputs a signal to initiate an action, wherein preferably a message is shown on a display. In this way, the user can check the image of the tire to confirm that indeed no tire defects are present in the image. If the user determines that at least one tire defect is present, the user can manually mark at least one tire defect in the image of the tire. This information can be used to improve the trained artificial intelligence algorithm.

[0024] Preferably, the trained artificial intelligence algorithm is implemented on an AI acceleration chip. In this way, the performance of the trained artificial intelligence algorithm can be improved. Thus, for example, the speed of the tire defect detection method can be increased. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Refer to the following Figures 1 to 6 Describing the present invention in more detail, Figures 1 to 6 Exemplary, schematic and non-limiting advantageous embodiments of the present invention are shown. In the drawings:

[0026] Figure 1 shows a portion of an X-ray image of a tire,

[0027] Figure 2 Shows several common types of tire defects.

[0028] Figure 3 shows the results from a trained artificial intelligence algorithm,

[0029] Figure 4 shows the steps and results of the image processing algorithm,

[0030] Figure 5 shows another result of the image processing algorithm, and

[0031] Figure 6 The general steps of the proposed detection method are shown. DETAILED DESCRIPTION

[0032] Figure 1 Schematically an X-ray image is shown as an image 1 of a tire 2. Of course, in real life, the X-ray image will look different, especially in the form of a grayscale image. However, for the purpose of describing the proposed detection method, a black and white image is used for illustration purposes.

[0033] Depending on which components and plies of the tire 2 should be inspected, different types of images 1 of the tire 2 can be used. For example, using an X-ray image, the inner ply structure of the tire 2 can be inspected. Figure 1 , represented as a shadow pattern in . Instead of an X-ray image, an image 1 of the surface (e.g., the outer surface or the inner surface) of the tire 2 can also be captured, which can be used to inspect the outside of the tire 2. For the method, one type of image 1 is preferably used, although different types of images can also be used. In the following and without limiting the generality, the proposed detection method is described with the help of X-ray images.

[0034] It is also noted that the (X-ray) image 1 shown is only used to illustrate the invention, but it is not necessary to display the image 1 on a display in order to implement the invention.

[0035] Depending on which type of image 1 of the tire 2 should be taken, different imaging units for capturing the image 1 may be used. To capture the image 1 of the surface of the tire 2, for example, a suitable digital camera with an illumination unit (e.g., an LED array) for adequate illumination may be used. To capture the X-ray image 1, an X-ray scanning device (e.g., a backscatter or transmission X-ray scanner) may be used as an imaging unit.

[0036] In tire manufacturing, the entire circumference of a tire 2 is usually captured to detect tire defects 3. One common approach is to mount the tire 2 on a rotating device to rotate the tire 2 360° around its rotation axis while taking an image 1 of the tire 2, thereby producing an elongated image 1 showing the entire circumferential surface of the tire 2 or a circumferential inner area. Figure 1In , only a segment of an elongated image 1 (in case of an X-ray image) in the circumferential direction is shown. Another way to capture the entire circumference of the tire 2 is by imaging the tire 2 from different angles with multiple imaging units and combining the images 1 from the multiple imaging units.

[0037] Figure 1 The X-ray image 1 of a tire 2 in FIG. 1 shows different parts of the tire 2, such as the tread, the sidewalls and the inner ply structure of the tire 2. Figure 1 In the case of the exemplary X-ray image 1 in FIG. 1 , the outer segment represents the sidewall of the tire 2. The inner segment is a portion of the tread, the outer surface of which is in contact with the road when the tire 2 is in use. Depending on the type of tire 2, different types of materials are used, and the tire 2 includes a certain number of plies (e.g., carcass plies, crown plies, steel belts, radial plies, etc.), which form an inner ply structure of the tire 2. These plies contain cords, such as steel cords, which are arranged in different orientations or patterns within a certain ply. In addition, depending on the type of material and its thickness, the absorption capacity of X-rays varies on different parts of the tire 2. The sidewall is generally thinner than the area of ​​the tread and includes fewer plies. As a result, different contrast levels may appear in the X-ray image 1 of the tire 2, resulting in a grayscale image.

[0038] As mentioned at the beginning of this article, during tire manufacturing, various tire defects 3 may appear in different parts of the tire 2. The tire defects 3 appear randomly at different locations on the outside or inside of the tire 1. Figure 2 , a non-limiting number of common types of tire defects 3 are shown, which can usually be detected in an X-ray image 1 of a tire 2. For illustration purposes only, Figure 2 The types of tire defects 3 in are each marked by a line or a rectangle. When the method of the invention is implemented, such markings will of course not appear in image 1. In the first (upper) row starting from the left image, cord misalignment, impurities, upward tilt of the cord, bubbles and off-center cord plies are shown. Off-center cord plies are represented only by black and white dotted lines. In the second (lower) row starting from the left image, cord ply misalignment (step off), cord twisting, overlap, cord puncture and missing cord are shown. The list of types of tire defects 3 is not exhaustive. In addition Figure 2 Besides the tire defects 3 shown, other tire defects 3 may also occur, such as scratches on the outside or inside of the tire 2 .

[0039] For the proposed tire defect detection method, an image 1 of a tire 2 is used as input to a trained artificial intelligence algorithm, wherein the trained artificial intelligence algorithm detects whether at least one tire defect 3 is present in the captured image 1 of the tire 2. The trained artificial intelligence algorithm is preferably a machine learning algorithm or a deep learning algorithm based on a neural network (e.g., FasterR-CNN or similar), which is trained as described below.

[0040] For training the artificial intelligence algorithm, a plurality of images 1 of tires 2 containing at least one tire defect 3 are used. The number of images 1 of tires 2 can be collected on site. As described above, preferably one type of image 1 is used. For describing the training of the artificial intelligence algorithm, X-ray images 1 are used below. Preferably, the images 1 are preprocessed. For example, by cropping the images 1 to equal length and width. Depending on the artificial intelligence algorithm and the type of images 1, certain known preprocessing steps of image processing (e.g. compression, sharpening, noise reduction, etc.) can help improve the training.

[0041] As a next step, at least one tire defect 3 in the image 1 is manually marked by the user. For example, the location of the at least one tire defect 3 is marked as a rectangle (e.g., similar to Figure 2 In addition, the type of at least one tire defect 3 is marked so that the artificial intelligence algorithm can learn different types of tire defects 3 and can distinguish between them. To mark the image 1, appropriate marking software can be used.

[0042] After labeling, the images 1 are preferably randomly grouped into a training set, a validation set, and a test set according to appropriate proportions. For example, 70% of the images 1 are placed in the training set, 15% are placed in the validation set, and 15% are placed in the test set.

[0043] The images 1 in the training set are used to train the artificial intelligence algorithm. Parameter values ​​of weights and / or other parameters of the artificial intelligence algorithm, such as bias, are initially set. Those parameter values ​​are adjusted by training the artificial intelligence algorithm. In this way, the artificial intelligence algorithm learns to recognize different types of tire defects 3 and their locations in the images 1. It should be noted that there are known learning methods (e.g., supervised learning, unsupervised learning, or similar methods) for training artificial intelligence algorithms that can be used.

[0044] Images 1 in the validation set are used to determine the accuracy of detecting tire defects 3 by the artificial intelligence algorithm. Since only images 1 in the training set are used to train the artificial intelligence algorithm, images 1 in the validation set that have not been used are "unseen" by the artificial intelligence algorithm. Therefore, images 1 in the validation set can be used to prevent the artificial intelligence algorithm from overfitting (as described at the beginning). For example, if the artificial intelligence algorithm is trained with a certain number of images 1 in the training set, images 1 in the validation set can be used to check the accuracy of the artificial intelligence algorithm. If the determined accuracy meets the predetermined accuracy of the artificial intelligence algorithm in detecting tire defects 3 (e.g., the percentage of identified tire defects of different defect types), the learning of the artificial intelligence algorithm is completed. If the determined accuracy is not sufficient compared to the predetermined accuracy, the artificial intelligence algorithm can be trained again with images 1 in the training set, and the artificial intelligence algorithm can be checked again with images 1 in the validation set. For example, the artificial intelligence algorithm is trained to converge, for example, until a certain accuracy according to industry quality requirements is met.

[0045] The achievable accuracy of the AI ​​algorithm depends on the number of images1 used for training. According to the test results, 2000 images1 were used to train the AI ​​algorithm, achieving Figure 2 An accuracy of more than 99% is shown for detecting different types of tire defects 3. However, this is only an example, since in addition to the number of images 1, the quality of the images 1 (e.g. quality of labeling, pre-processing, etc.) is also relevant for training the artificial intelligence algorithm.

[0046] Training of AI algorithms can be performed on appropriate processing units, e.g., GPUs with high computing power, to ensure fast training with large numbers of images1.

[0047] In the testing phase, the (also unseen) images 1 in the test set can be used to determine the performance of the trained artificial intelligence algorithm, for example, whether the tire defect detection works as expected. The trained artificial intelligence algorithm is preferably implemented on an AI (artificial intelligence) acceleration chip (also called an AI accelerator) for use in the tire manufacturing process to detect tire defects 3. The AI ​​acceleration chip is a dedicated hardware component (e.g., based on a GPU) specifically designed for efficient and fast processing of AI workloads (such as the calculation of neural networks).

[0048] exist Figure 3 , where an X-ray image 1 of a tire 2 is used as input to the trained artificial intelligence algorithm. As mentioned above, typically the image 1 of the tire 2 is an elongated image 1 of the entire circumference of the tire 2. Before using the image 1 as input to the trained artificial intelligence algorithm, the elongated image 1 (e.g., Figure 1 shown).

[0049] If at least one tire defect 3 is present in the image 1, the trained artificial intelligence algorithm marks at least one tire defect 3 in the image 1. Preferably, the trained artificial intelligence algorithm marks the approximate position relative to the reference coordinate system of the tire 2 and the type of at least one detected tire defect 3 in the image 1. For example, Figure 3 As shown, the trained artificial intelligence algorithm marks the approximate location of at least one detected tire defect 3 in the image 1 as a boundary of a given geometric shape (such as a rectangle) surrounding at least a portion of the detected tire defect 3, as well as the type of the detected tire defect 3.

[0050] If the trained artificial intelligence algorithm detects that no tire defect 3 is present in the image 1 of the tire 2, the trained artificial intelligence algorithm outputs a signal to initiate an action, wherein preferably a message is displayed on a display, such as a display of a workstation or a mobile device, to inform the user. In this way, a user (e.g., tire manufacturing personnel) can manually check the image 1 of the tire 2, at least on a sample basis, to confirm that there is indeed no tire defect 3 in the image 1. If the user determines that at least one tire defect 3 is present in the image 1, the user can manually mark at least one tire defect 3 in the image 1. This information can also be used to further improve the performance of the trained artificial intelligence algorithm by retraining the artificial intelligence algorithm using this information. If the trained artificial intelligence algorithm does not detect a tire defect 3 due to poor image quality, the user can also interrupt the detection method and restart the method with a new image 1 of the tire 2 with better image quality.

[0051] In the case where the user has to manually check the detected tire defect 3 on the actual tire 2, the position relative to the reference coordinate system is preferably used to locate the detected tire defect 3 on the actual tire 2. As the reference coordinate system of the tire 2, a spherical coordinate system can be used, whose source point is located at a predetermined reference point of the tire 2, for example, the center of the tire 2.

[0052] The trained artificial intelligence algorithm outputs a marked image 4 with at least one tire defect 3. Preferably, the marked image 4 is displayed on a display. In this way, a user can manually check the at least one tire defect 3 detected by the trained artificial intelligence algorithm. It is also possible to display only a fragment of the marked image 4 on the display (e.g. Figure 3 ), for example, a fragment in which a tire defect 3 is present.

[0053] The marked image 4 with at least one tire defect 3 is then used as input to an image processing algorithm. An image processing algorithm is basically an algorithm that processes a digital image represented by pixels at a given resolution. The image processing algorithm determines and marks at least one geometric feature 5 of at least one detected tire defect 3 in the marked image 4. The at least one geometric feature 5 is preferably a size, a ply position or an orientation of the at least one tire defect 3, preferably relative to a reference coordinate system of the tire 2. Furthermore, by determining the at least one geometric feature 5, the position of the detected tire defect 3 can be described more accurately than the approximate position marked by the trained artificial intelligence algorithm.

[0054] Preferably, the image processing algorithm crops the marker image 4 (eg, Figure 3 ). The image processing algorithm determines and marks at least one geometric feature 5 of the tire defect 3 detected in the cropped marked image 4. Due to the cropped image 4, the image processing algorithm processes a smaller image 4. Therefore, more details of the geometric features 5 of the tire defect 3 can be determined and the necessary processing power can be reduced.

[0055] Depending on the type of tire defect 3 and / or geometric feature 5, filtering, edge or contour extraction, pixel scanning, blob recognition, etc. can be used as image processing algorithms. Generally, those known methods are used in image processing to detect points and / or areas in the image (as well as their size, position, orientation, etc.) that differ in properties such as brightness or color compared to the surrounding or background in order to identify at least one geometric feature 5. For example, in the case of blob identification, a blob can be defined as a group of connected pixels in the image that share the property (e.g. the same or approximately the same brightness value that is different from the surrounding pixels). A blob can be a bright object in a dark background, or vice versa, with the borders of the object shifting abruptly from dark to light or from light to dark. Advantageously, a suitable image processing algorithm is implemented as image processing algorithm for each type of tire defect 3.

[0056] exist Figure 4 , exemplary steps and results of an image processing algorithm are exemplarily and schematically shown, wherein at least one geometrical feature 5 of a tire defect 3 (upward inclination of a cord) detected in a marked image 4 is determined and marked. Figure 4The upward inclination of the cords in is schematically shown with hatching. The image processing algorithm has cropped the marked image 4 to a certain width and length, which is shown in the left figure. In this example, the width and length of the cropped image 4 are represented by a number of pixels given on the ordinate and abscissa. As shown in the middle image, the (bright) background of the cropped image 4 can be darkened and the tire defect 3 can be brightened (e.g., increase the contrast) to distinguish it from the dark background. By counting the number of pixels, the maximum height (distance) of the upward inclination of the cords as geometric feature 5 can be determined. In this example, the height is 36 pixels. The height can be marked in the image 4, for example by an arrow, and as shown in Figure 4 As shown in the right figure in , the determined height is marked in pixels. Using the knowledge of the pixel size in the marked image 4 and the known characteristics of the optical system used to capture the image (such as focal length, distance to the imaged object, etc.), the upward tilt height can also be given in units of length in the real world.

[0057] Figure 5 A further exemplary result of the image processing is shown in , in which at least one geometric feature 5 of a tire defect 3 (impurity) detected in the marked image 4 is determined and marked. Figure 5 The impurities in the figure are schematically shown with hatching. In this example, the size (width and length) of the tire defect 3 is determined as a geometric feature 5, for example by spot recognition. Figure 5 As shown, the size of an impurity is characterized by a sudden shift of pixels from dark to light or from light to dark. The size of an impurity can be determined by these shifts and can be marked by a rectangle and given, for example, in pixels or as mentioned in the real world in units of length. It is also possible to outline the boundaries of an impurity and calculate the size and / or area of ​​the impurity more accurately.

[0058] The image processing algorithm outputs a marked image 6 having at least one determined and marked geometric feature 5 of at least one detected tire defect 3, such as Figure 4 and Figure 5 As shown. The marked image 6 is displayed on the display. In addition, the marked image 6 with the determined at least one geometric feature 5 of the at least one detected tire defect 3 can be used to grade the tire quality by comparing the at least one determined geometric feature 5 of the at least one detected tire defect 3 with a predetermined geometric feature. For example, the predetermined geometric feature can be the maximum allowed size of the impurity or the height of the upward inclination. The predetermined geometric feature can be predetermined according to the industry quality requirements. If the tire quality is sufficient, the tire 2 passes the quality requirements and can enter delivery. If the tire quality is not enough, the tire 2 can be discarded.

[0059] The displayed marking image 6 can also be checked by the user. In this way, the user can manually confirm whether the tire quality is sufficient. For example, a user's check may be necessary if the comparison of the determined at least one geometrical feature 5 of the at least one tire defect 3 detected with the predetermined geometrical feature is unclear. The user's decision whether the tire quality is sufficient can be used for further improvement of the detection method.

[0060] exist Figure 6 , the general steps of the proposed detection method are shown. In a first step S10, an image 1 of a tire 2 is used as input to a trained artificial intelligence algorithm. The trained artificial intelligence algorithm detects in step S20 whether at least one tire defect 3 is present in the image 1. The trained artificial intelligence algorithm marks the at least one tire defect 3 in the image 1. If the trained artificial intelligence algorithm detects that no tire defect 3 is present in the image 1 of the tire 2, the trained artificial intelligence algorithm outputs a signal in step S21 to inform the user. If the user determines that at least one tire defect 3 is present in the image 1, the user can manually mark the at least one tire defect 3 or interrupt the detection method, for example due to poor image quality, and restart the method with a new image 1 of the tire 2.

[0061] In step S30, the trained artificial intelligence algorithm outputs a marked image 4 with at least one tire defect 3. The marked image 4 with at least one tire defect 3 is used as input to an image processing algorithm in step S40, wherein the image processing algorithm determines and marks at least one geometric feature 5 of the detected at least one tire defect 3 in the marked image 4. In step S50, the image processing algorithm outputs a marked image 6 with the determined at least one geometric feature 5 of the detected at least one tire defect 3, and the marked image 6 is displayed on a display.

[0062] Preferably, the marked image 6 can be used in step S60 to grade the tire quality by comparing the determined at least one geometric feature 5 of the detected at least one tire defect 3 with a predetermined geometric feature. Optionally, in step S61, the marked image 6 with the determined at least one geometric feature 5 of the detected at least one tire defect 3 is displayed on a display for inspection and confirmation by a user. The user can then determine whether the tire quality is sufficient.

[0063] The proposed tire defect detection method can be used in different stages of tire manufacturing to detect tire defects 3, such as in printing, hot stamping, die cutting, etc. In addition, performance tests of the proposed method were conducted on site. By determining the geometrical features 5 of the tire defects 3, an accuracy of more than 99% of correctly grading tire quality was achieved, leaving only less than 1% of (uncertain) cases for which manual inspection by the user was required. In addition to minimizing manual inspection, a detection time of less than 3 s was measured to detect a tire defect 3 of one tire 2 and to grade its tire quality using the determined geometrical features 5 of the tire defects 3.

Claims

1. A method for detecting tire defects, wherein an image (1) of a tire (2) is used as input to a trained artificial intelligence algorithm, wherein the trained artificial intelligence algorithm detects whether at least one tire defect (3) is present in the image (1), the trained artificial intelligence algorithm marks at least one tire defect (3) in the image (1), and the trained artificial intelligence algorithm outputs a marked image (4) with at least one tire defect (3), characterized in that: the output marked image (4) with the at least one tire defect (3) is used as input into an image processing algorithm, wherein the image processing algorithm determines and marks at least one geometric feature (5) of the at least one tire defect (3) detected in the marked image (4), The image processing algorithm outputs a marked image (6) having the at least one determined and marked geometric feature (5) of the at least one tire defect (3) detected, and The marked image (6) is displayed on a display.

2. The method according to claim 1, characterized in that The image (1) of the tire (2) is an X-ray image of the tire (2) or a surface image of the tire (2).

3. The method according to claim 1 or 2, characterized in that: The trained artificial intelligence algorithm is an artificial intelligence algorithm based on a neural network, wherein preferably preprocessed images (1) of tires (2) marked with at least one tire defect (3) are used to train the artificial intelligence algorithm, The images (1) are grouped into a training set and a validation set, wherein the images (1) in the training set are used to train the artificial intelligence algorithm, wherein parameter values ​​of the artificial intelligence algorithm are adjusted, and the images (1) in the validation set are used to determine the accuracy of tire defects (3) detected by the artificial intelligence algorithm, and When the determined accuracy satisfies a predetermined accuracy for detecting tire defects (3) by the artificial intelligence algorithm, the artificial intelligence algorithm is trained.

4. The method according to claim 3, characterized in that The trained artificial intelligence algorithm is tested using a test set comprising images (1) of tires (2) marked with at least one tire defect (3), preferably pre-processed, to determine the performance of the trained artificial intelligence algorithm.

5. The method according to any one of claims 1 to 4, characterized in that The trained artificial intelligence algorithm marks the position and type of at least one of the tire defects (3) detected in the image (1), the position preferably being relative to a reference coordinate system of the tire (2).

6. The method according to any one of claims 1 to 5, characterized in that The trained artificial intelligence algorithm marks the location of at least one of the detected tire defects (3) in the image (1) as a boundary around at least part of the detected tire defect (3), and The image processing algorithm crops the marked image (4) to the boundary around the detected tire defect (3) to determine and mark the at least one geometric feature (5) of the detected tire defect (3) in the cropped marked image (4).

7. The method according to any one of claims 1 to 6, characterized in that The at least one geometrical characteristic (5) of the detected tire defect (3) describes at least a size or a ply position or orientation, preferably relative to the reference coordinate system of the tire (2).

8. The method according to any one of claims 1 to 7, characterized in that The marked image (6) having the determined at least one geometrical feature (5) of at least one detected tire defect (3) is used for grading the tire quality by comparing the determined at least one geometrical feature (5) of at least one detected tire defect (3) with predetermined geometrical features.

9. The method according to any one of claims 1 to 8, characterized in that If the trained artificial intelligence algorithm detects that no tire defect (3) is present in the image (1) of the tire (2), the trained artificial intelligence algorithm outputs a signal to initiate an action, wherein preferably a message is shown on the display to inform a user.

10. The method according to any one of claims 1 to 9, characterized in that The trained artificial intelligence algorithm is implemented on an AI acceleration chip.

Images