Meridian tire x-ray image defect automatic detection method based on improved YOLO-v5 model

Abstract

The application discloses a kind of meridian tire X-ray image disease flaw automatic detection method based on improved YOLO-v5 model, belong to meridian tire detection technical field, including following steps: step 1, meridian tire X-ray image is collected and is segmented processing, uniform resolution, model training sample data are made;Step 2, for the image strip, block missing caused by X-ray machine instability, thereby affecting the detection effect Problem, image restoration processing is carried out;Step 3, design improved YOLO-v5 model, including increasing the fourth detection layer, increase attention module, and improve loss function;Step 4, using meridian tire disease flaw data carries out model training;Step 5, using the model of training completion carries out actual application scene tire disease flaw detection.The application is a kind of automatic detection method, based on improved model can automatically identify multiple disease flaws, and different disease flaws are detected and classified, and detection efficiency and accuracy are higher.

Description

Technical Field

[0001] This invention belongs to the field of radial tire inspection technology, specifically relating to an automatic defect detection method for radial tire X-ray images based on an improved YOLO-v5 model. Background Technology

[0002] Radial tires, as a major component of automobiles, are closely related to people's lives and industrial production. Radial tires have a thicker tread, making them more pressure-resistant and wear-resistant. They also feature a belt layer with high elongation and high hardness, which acts as a buffer and tightens the tire body. This results in a larger contact patch area and greater adhesion; lower contact pressure and more even load distribution, thus extending the tire's mileage. Due to the superior performance of radial tires, they have achieved a dominant market position, leading to increasingly higher requirements for tire manufacturing quality and the performance of finished products. Therefore, final product inspection is a crucial step in tire production quality control, ensuring that the tires leaving the factory meet excellent quality and performance standards.

[0003] Currently, my country's corresponding automatic tire defect detection algorithms are not yet perfect. Generally, tire X-ray images are observed manually to determine whether the tire has defects and to classify the types of defects, resulting in low inspection efficiency. Secondly, tire X-ray images obtained by X-ray machines are subject to noise interference and have low brightness. Furthermore, the image shapes of various defects are different, making it difficult to extract different defect features using a single method. Summary of the Invention

[0004] Given the variety of radial tire models, the diversity of tire tread patterns, and the numerous types of defects present in radial tires, and the potential differences in shape, scale, and grayscale in X-ray images of the same defect across different radial tire models, this invention proposes an automatic defect detection method for radial tire X-ray images based on an improved YOLO-v5 model, capable of identifying a wide range of defects.

[0005] The technical solution of the present invention is as follows:

[0006] An automatic defect detection method for radial tire X-ray images based on an improved YOLO-v5 model includes the following steps:

[0007] Step 1: Collect radial tire X-ray images, perform segmentation processing, unify resolution, and create model training sample data;

[0008] Step 2: Perform image restoration processing;

[0009] Step 3: Design and improve the YOLO-v5 model;

[0010] Step 4: Use radial tire defect data for model training;

[0011] Step 5: Use the trained model to detect tire defects in real-world application scenarios.

[0012] Furthermore, the specific steps of step 1 are as follows:

[0013] Step 1.1 First, for an original radial tire X-ray image, reduce its width to 1900 pixels and its height proportionally. The corresponding actual annotation box in the defect description file is also reduced accordingly.

[0014] Step 1.2: Use a 1900*1900 window to slide along the height direction in the original image, and set the window to slide 10 times;

[0015] Step 1.3: If the window does not contain any defects while sliding, continue sliding forward; otherwise, correct the location information of the defects and generate sub-images and corresponding defect description files.

[0016] Furthermore, the specific steps of step 2 are as follows:

[0017] Let the radial tire X-ray image be a third-order tensor. n1, n2, and n3 represent the dimensions of each order, and the constructed restoration model is as follows:

[0018]

[0019] in, For a clean image to be solved, N represents the tensor. The order of , FCTN is a fully connected tensor network. This represents the N factor tensors obtained from the fully connected tensor decomposition. It is an implicit regularizer based on deep learning, where λ is the regularization weight parameter, σ is the denoising parameter, and Ω represents... The set of pixel indices observed in the image. This represents mapping by Ω;

[0020] The model is solved by alternating iterative updates using the following algorithm:

[0021]

[0022] in, This represents a two-dimensional denoising convolutional neural network trained on a natural image dataset, where q represents the number of iterations, i = 1, ..., N represents the tensor of each factor, ρ is the proximal parameter, and Ω is the denoted Ω. C It represents the complement of set Ω.

[0023] Furthermore, in step 3,

[0024] The improved YOLO-v5 model structure includes an input end, a backbone network, a bottleneck network, and a prediction end. The input end represents the input image, where image preprocessing is performed, scaling the input image to the network's input size and normalizing it. The backbone network is used to extract some general feature representations. The bottleneck network is located between the backbone network and the prediction end, and is used to further improve the diversity and robustness of features. The prediction end is used to output the object detection results.

[0025] The YOLO-v5 model improvement process is as follows:

[0026] Step 3.1: Based on the original YOLO-v5 model, add a fourth detection layer to upsample the feature map with the largest output scale, further expanding the feature map. At the same time, cascade and fuse the obtained feature map with the same size in the backbone network.

[0027] Step 3.2: Add an attention module to each detection layer feature processing stage; the attention module works as follows:

[0028] The intermediate feature map T is subjected to MaxPool max pooling and AvePool average pooling in the spatial dimension, respectively, to obtain two pooling vectors. These vectors are then input into a shared multilayer mapping neural network (MLP) for nonlinear mapping, resulting in two new vectors. These two vectors are then summed bit-by-bit and further nonlinearly mapped using the Sigmoid activation function to obtain the channel attention module. The transformation formula is as follows:

[0029] A1(T)=σ1{MLP[AvgPool(T)]+MLP[MaxPool(T)]} (3)

[0030] Where A1 represents the channel attention module and σ1 represents the Sigmoid function;

[0031] The intermediate feature maps T are subjected to MaxPool max pooling and AvePool average pooling operations in the channel dimension, then stacked, and then mapped to a single-band, same-size model through convolution operation to obtain the spatial attention module A2.

[0032]

[0033] in, This indicates that a 1×1 convolutional neural network is used for feature transformation, where σ1 represents the Sigmoid function, and its function equation is:

[0034]

[0035] Where x represents the input tensor;

[0036] Step 3.3: Improve the loss function; add an impact factor, measure it on the aspect ratio of the detection box, and further use DIoU_P, which is more consistent with the regression mechanism. The formula is:

[0037]

[0038] Where IoU represents the overlap ratio between the predicted bounding box and the ground truth bounding box, and b and b gt ρ1 and c represent the center points of the detection box and the target box, respectively; ρ1 represents the Euclidean distance between the two center points; c represents the diagonal distance of the smallest closure region that can simultaneously contain the predicted box and the ground truth box; μ is a parameter used to balance the scale; m is used to measure the proportional consistency between the width and height of the predicted box and the ground truth box.

[0039] Furthermore, the specific steps of step 4 are as follows:

[0040] Step 4.1: Process the model training sample data created in Step 1 to establish a dataset including defects in X-ray images of tires with different meridians. Label the image information and classify it to obtain the training dataset.

[0041] Step 4.2: Build the improved YOLO-v5 network model, set the training path, and read the training parameters;

[0042] Step 4.3: Load the pre-trained model and obtain the initial network model parameters;

[0043] Step 4.4: Select the training optimizer;

[0044] Step 4.5: Load the training set;

[0045] Step 4.6: Start training, update network model parameters until the iteration ends and model training is complete.

[0046] Furthermore, the specific steps of step 5 are as follows:

[0047] Step 5.1: Acquire the current original radial tire X-ray image in real time and perform image preprocessing, including image segmentation, resolution unification, and image restoration;

[0048] Step 5.2: After obtaining the preprocessed image, input it into the trained improved YOLO-v5 model, and then perform image partitioning;

[0049] Step 5.3: Based on the different locations of the defects, they are divided into local defects and global defects; by further extracting detailed features in different areas of the tire, corresponding defect detection is performed, and the model provides detection information, including the location information of the defects in the image and the predicted confidence levels.

[0050] Step 5.4: Then, perform confidence filtering. Based on a pre-set threshold, only retain candidate boxes with a confidence level greater than the threshold and display them on the current detection image to obtain the current tire defect detection status.

[0051] The beneficial technical effects of this invention are as follows:

[0052] The X-ray images of tires are processed by segmentation to unify the resolution. The segmented images are then used to train the model, which solves the problems of large tire X-ray images and inconsistencies in imaging equipment.

[0053] A model-data jointly driven image denoising method is proposed. Based on the tensor fully connected decomposition model, a pre-trained deep denoiser is inserted as a plug-and-play module. This denoising method has the good transferability of tensor models and the superior denoising performance of deep learning, and solves the problem of noise interference in tire X-ray images.

[0054] The YOLO-v5 model is improved by adding an extra detection layer and attention module, and the original loss function is improved in the same way. The improved model can capture small-scale feature information of radial tire X-ray images, classify and extract defect features of radial tire X-ray images more comprehensively, and detect the location of defects more comprehensively. It solves the problems in tire defect detection, such as the defect target may be relatively small, lack sufficient appearance feature information, and be difficult to distinguish them from the background or similar targets, as well as the differences in X-ray imaging in real scenes, target occlusion and target scale changes.

[0055] This invention is an automatic detection method that, based on an improved model, can automatically identify various defects and classify them according to their different characteristics, resulting in high detection efficiency and accuracy. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the segmentation process of a meridional X-ray image according to the present invention.

[0057] Figure 2 This is a diagram showing the overall structure of the improved YOLO-v5 model of this invention.

[0058] Figure 3 This is a structural diagram of the basic components of the improved YOLO-v5 model of this invention.

[0059] Figure 4 This is a structural diagram of the first method for implementing the channel attention module and spatial attention module of the present invention;

[0060] Figure 5 This is a structural diagram of the second method for implementing the channel attention module and spatial attention module of the present invention;

[0061] Figure 6 This is a flowchart illustrating the defect detection process based on the trained model of this invention.

[0062] Figure 7 This is a diagram illustrating the tire area division in an embodiment of the present invention;

[0063] Figure 8 This is a defect detection diagram of a tire image in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0065] This invention proposes an automatic defect detection method for radial tire X-ray images based on an improved YOLO-v5 model, which mainly includes the following steps:

[0066] Step 1: Collect radial tire X-ray images, segment them, unify the resolution, and create model training sample data.

[0067] To address the issue of large X-ray images of radial tires, the images are segmented and their resolution standardized. These segmented images are then used as training samples for the model. Considering that the original tire X-ray image size is approximately 3000 pixels * 10000 pixels, such a large image cannot be directly input into the detection model. Furthermore, a tire X-ray image often only shows defects in certain localized areas, with most areas displaying normal textures. Therefore, it is necessary to crop the original image based on a defect description file, effectively "extracting" the defects to construct a sample set that can be used as input to the model. The specific cropping process is as follows: Figure 1 As shown:

[0068] Step 1.1 First, for an original radial tire X-ray image, reduce its width to 1900 pixels and its height proportionally. The corresponding actual annotation box in the defect description file also needs to be reduced accordingly.

[0069] Step 1.2: Use a 1900*1900 window to slide along the height direction in the original image. In this invention, the window is set to slide 10 times. Therefore, for a scaled-down tire X-ray image, assuming its scaled-down height is 9000, each adjacent window overlaps by at least half, and basically no defects will be missed.

[0070] Step 1.3: Through the defect description file, the actual position and size of the defect's bounding box can be determined. Therefore, the relative position of the defect to the window can be easily determined when the window is slid. If the window does not contain any defects when slid, continue sliding forward; otherwise, correct the defect's position information and generate a sub-image and its corresponding defect description file.

[0071] Step 2: To address the issue of striped or blocky image defects caused by X-ray machine instability, which affect the detection results, image restoration processing is performed.

[0072] Due to differences in the performance of radial tire X-ray machines and the influence of image acquisition equipment, tire X-ray images may have drawbacks such as noise, missing images, and low resolution, which affect the detection results.

[0073] To address the issue of missing or noisy tire X-ray images affecting detection results, image restoration processing was implemented. The meridional tire X-ray image was assumed to be a third-order tensor. n1, n2, and n3 represent the dimensions of each order, and the constructed restoration model is as follows:

[0074]

[0075] in For a clean image to be solved, N represents the tensor. The order of , FCTN is a fully connected tensor network. This represents the N factor tensors obtained from the fully connected tensor decomposition. It is an implicit regularizer based on deep learning, where λ is the regularization weight parameter, σ is the denoising parameter, and Ω represents... The set of pixel indices observed in the image. This indicates mapping by Ω.

[0076] The model is solved by alternating iterative updates using the following algorithm:

[0077]

[0078] in, This represents a two-dimensional denoising convolutional neural network trained on a natural image dataset, where q represents the number of iterations, i = 1, ..., N represents the tensor of each factor, ρ is the proximal parameter, and Ω is the denoted Ω. C It represents the complement of set Ω.

[0079] Step 3: Design and improve the YOLO-v5 model.

[0080] The ultimate goal of tire defect detection is to replace traditional manual inspection with advanced computer vision technology. Therefore, tire defect detection algorithms must be deployed in business application scenarios to achieve intelligent tire defect identification. This requires the defect identification model to meet the requirements of real-time recognition and detection accuracy. The YOLO series algorithms offer fast recognition speed, strong real-time performance, and high target detection accuracy. This aligns well with the intelligent tire defect identification task; therefore, the YOLO-v5 model can be used as the basic model for tire defect identification. However, due to certain drawbacks of the original YOLO-v5 model, such as insufficient recognition accuracy, this invention improves the YOLO-v5 model. The improved architecture is as follows: Figure 2 As shown.

[0081] The improved YOLO-v5 model structure includes an input layer, a backbone network, a bottleneck network, and a prediction layer. The input layer represents the input image, typically involving image preprocessing such as scaling the input image to the network's input size and performing normalization. The backbone network extracts general feature representations. The bottleneck network, usually located between the backbone and prediction layers, further enhances feature diversity and robustness. The prediction layer outputs the object detection results.

[0082] The improved YOLO-v5 model basic components are described below:

[0083] (1) The CBL module consists of ConV+BN+Leaky_Relu activation functions, such as Figure 3 Module (a) is shown in the diagram.

[0084] (2) The Res Unit borrows the residual structure from the ResNet network to construct deep networks, such as... Figure 3 Module (b) is shown in the diagram.

[0085] (3) CSP1_X divides the input into two branches. One branch first passes through a CBL, then a Res Unit, and then performs a convolution; the other branch performs a direct convolution. The two branches are then concatenated, passed through a BN layer, and then through a Leaky ReLU activation function. For example... Figure 3 Module (c) is shown in the diagram.

[0086] (4) The only difference between CSP2_X and CSP1_X is that CSP2_X replaces the Res Unit with CBL, and it is mainly used in bottleneck networks, such as... Figure 3 Module (d) is shown in the diagram.

[0087] (5) SPP employs max pooling methods of 1×1, 5×5, 9×9, and 13×13 for multi-scale feature fusion, such as... Figure 3 Module (e) in the middle.

[0088] (6) The Focus structure first concatenates multiple slice results, and then sends them into the CBL module. For example... Figure 3 Module (f) in the middle.

[0089] The various layers within the basic components are explained below:

[0090] (1) ConV layer, also known as convolutional layer, is used to extract feature information of input image.

[0091] (2) BN layer, also known as batch normalization layer, is used to standardize the input, thereby greatly suppressing the overfitting and gradient vanishing problems of neural network models.

[0092] (3) The LeakyReLU activation function mainly introduces nonlinear factors into the neural network model, enabling the neural network model to perform nonlinear tasks.

[0093] (4) Concat layer, also known as splicing layer, mainly splices two or more feature maps in the channel dimension to achieve feature information fusion at different levels.

[0094] (5) Slice operation: slices the input image so that the Focus module can achieve double downsampling without losing information.

[0095] (6) Maxpool layer, also known as pooling layer, is used to improve the training speed of neural networks and effectively prevent overfitting.

[0096] The main improvements to the YOLO-v5 model are as follows:

[0097] (1) Add a fourth detection layer.

[0098] To address issues in tire defect detection, such as the small size of some defects, insufficient appearance information, difficulty in distinguishing them from background or similar targets, differences in X-ray imaging in real-world scenarios, target occlusion, and changes in target scale, the YOLO-v5 network structure was improved. The original YOLO-v5 network outputs three detection layers. Taking an input size of 608×608×3 as an example, the output scales are 76×76×25 (corresponding to the first detection layer), 38×38×25 (corresponding to the second detection layer), and 19×19×25 (corresponding to the third detection layer). Considering that some defect points are small, the original 76×76 output feature map cannot effectively display the defect features. To obtain larger feature maps for small target detection and improve the detection accuracy of small target defects, a fourth detection layer was added to the improved model. The 76×76 feature map output by the YOLO-v5 model was upsampled to further expand the feature map. At the same time, the obtained 152×152 feature map was cascaded and fused with the 152×152 feature map in the backbone network.

[0099] (2) Add an attention module.

[0100] An attention module is added to each detection layer feature processing stage to enhance attention and achieve adaptive refinement of image target detection points.

[0101] This invention proposes two methods for designing attention modules.

[0102] Method 1: Given an intermediate feature map, perform convolution operations on the intermediate feature map in the spatial dimension to obtain a channel attention module, and perform convolution operations in the channel dimension to obtain a spatial attention module. Both convolution operations are performed under supervised training. For example... Figure 4 As shown, firstly, the intermediate feature map undergoes multiple convolution operations in the spatial dimension, keeping the number of channels constant, ultimately resulting in a p×1×1 one-dimensional vector, the same dimension as the vector dimension of a single channel in T. Similarly, the intermediate feature map undergoes multiple convolution operations in the channel dimension, keeping the size of each band constant, ultimately resulting in a 1×r×r two-dimensional vector, the same dimension as the dimension of a single band in T.

[0103] Method 2: For example Figure 5 As shown, the intermediate feature map T undergoes MaxPool max pooling and AvePool average pooling in the spatial dimension, respectively, to obtain two pooled vectors. These vectors are then input into a shared multilayer mapping neural network (MLP) for nonlinear mapping, resulting in two new vectors. These two vectors are then summed bit-by-bit and further nonlinearly mapped using the Sigmoid activation function to finally obtain the channel attention module. The transformation formula is as follows:

[0104] A1(T)=σ1{MLP[AvgPool(T)]+MLP[MaxPool(T)]} (3)

[0105] Where A1 represents the channel attention module and σ1 represents the Sigmoid function.

[0106] The intermediate feature maps T are subjected to MaxPool max pooling and AvePool average pooling operations in the channel dimension, then stacked, and then mapped to a single-band model of the same size through convolution operation, and then the spatial attention module A2 is obtained.

[0107]

[0108] in, This indicates that a 1×1 convolutional neural network is used for feature transformation, where σ1 represents the Sigmoid function, and its function equation is:

[0109]

[0110] Where x represents the input tensor.

[0111] (3) Improve the loss function.

[0112] In the original YOLO-v5 model, the Cross-Union Ratio (GIoU) ​​loss function is used. GIoU considers both overlapping and non-overlapping regions and reacts well to the degree of overlap. It penalizes candidate boxes with larger errors more severely. Different proportions of candidate boxes can achieve good detection results under this method. The formula is:

[0113]

[0114] Where IoU represents the overlap ratio between the predicted and ground truth boxes, i.e., the ratio of their intersection to their union; C includes the area of ​​the smallest bounding box of both the predicted and ground truth boxes; and B represents the area of ​​the predicted box. gt This represents the area of ​​the actual bounding box.

[0115] However, GIoU does not consider factors such as the distance, overlap rate, and scale between the target and anchor, and divergence issues can occur during training. This can cause the GIoU regression strategy to degenerate into an IoU regression approach. This phenomenon occurs because when IoU is 0, GIoU rapidly overlaps the detection and target boxes, and the penalty mechanism of the loss function becomes ineffective iteratively. This results in the inclusion of non-overlapping detection and target boxes being considered acceptable. To improve this method, this invention introduces a new loss function and incorporates an influence factor, measuring the aspect ratio of the detection boxes. Furthermore, it uses DIoU_P, which is more consistent with the regression mechanism, as shown in the formula:

[0116]

[0117] Among them, b and b gt ρ1 and c represent the center points of the detection box and the target box, respectively; ρ1 represents the Euclidean distance between the two center points; c represents the diagonal distance of the smallest closure region that can simultaneously contain the predicted box and the ground truth box; μ is a parameter used to balance the scale; m is used to measure the proportional consistency between the width and height of the predicted box and the ground truth box.

[0118] Step 4: Use radial tire defect data for model training; the specific training process for improving the YOLO-v5 model is as follows:

[0119] Step 4.1: Process the model training sample data created in Step 1 to establish a dataset including defects in X-ray images of tires with different meridians. Label the image information and classify it to obtain the training dataset.

[0120] Step 4.2: Build the improved YOLO-v5 network model, set the training path, and read the training parameters;

[0121] Step 4.3: Load the pre-trained model and obtain the initial network model parameters;

[0122] Step 4.4: Select the training optimizer;

[0123] Step 4.5: Load the training set;

[0124] Step 4.6: Start training, update network model parameters until the iteration ends and model training is complete.

[0125] Step 5: Use the trained model to detect tire defects in real-world application scenarios.

[0126] like Figure 6 As shown, the specific process includes the following:

[0127] Step 5.1: Acquire the current original radial tire X-ray image in real time and perform image preprocessing, including image segmentation, resolution unification, and image restoration.

[0128] Step 5.2: After obtaining the preprocessed image, it is input into the trained improved YOLO-v5 model, and then image partitioning is performed. Based on the differences in imaging features, the main components of a tire can be divided into the tread, shoulder, sidewall, and bead, such as... Figure 7 As shown.

[0129] (1) Tire crown

[0130] The tire crown, located in the center of the tire, mainly consists of the tread, belt layers, and ply layers. It is the primary component that directly contacts the ground during vehicle operation. The tread often features different tread patterns to adjust performance indicators such as traction, depending on the vehicle's driving requirements. The belt layers, also known as the buffer layers, play a role in cushioning impacts. Their structure relies on nearly parallel cords as reinforcement and is the tire's main load-bearing component. The ply layers form the tire's skeleton, supporting the belt layers and bearing the load.

[0131] (2) Fetal shoulder

[0132] This is the transition zone between the tire crown and the tire sidewall. Because the cord angles of the tread and carcass differ significantly, the edges of the tread must be filled with rubber to connect the two parts, forming a shoulder structure different from other tire structures. The shoulder thickness should not be too thick; otherwise, due to the limited cord distribution and the fact that it is merely filled with rubber, it may be prone to cracking. Similarly, this area is usually also patterned with tread grooves, connecting to the tread grooves, to facilitate heat dissipation and improve grip.

[0133] (3) Fetal side

[0134] The section between the tire shoulder and the bead refers to the rubber layer that covers the outer sidewall of the tire, perpendicular to the tire tread. It protects the internal cords of the tire from damage. At the same time, because this part is perpendicular to the ground when the car is moving, it must have sufficient elasticity to allow the tire tread to move up and down when encountering uneven road surfaces.

[0135] (4) Tire bead

[0136] The rim is the part of the tire that mounts to the wheel. It needs to be able to withstand the vertical stress caused by internal pressure, and also ensure that the rim can withstand the lateral force of the tire when cornering so that the tire does not come off the rim. Therefore, a dense and sturdy material that is not easily deformed is usually chosen.

[0137] Step 5.3: Based on the location of the defects, they are divided into local defects and global defects. For example, sparse lines, bends, and scattered lines all appear in the cord and belt layer areas, belonging to local defects. Impurities and bubbles can appear anywhere on the X-ray image, belonging to global defects. By further extracting detailed features from different areas of the tire, corresponding defect detection is performed. The model provides detection information, including the location information of the defect in the image and the predicted confidence levels.

[0138] Step 5.4: Then, perform confidence level filtering. Based on a pre-set threshold, only candidate boxes with a confidence level greater than the threshold are retained and displayed on the current detection image, thereby obtaining the current tire defect detection status. For example... Figure 8 The image shows a defect detection result, with the values ​​0.69, 0.80, and 0.42 representing the confidence levels. It can be seen that there is a belt layer open-root defect in the tread area, an impurity defect in the shoulder area, and cross-defects throughout the tire body.

[0139] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. An automatic defect detection method for radial tire X-ray images based on an improved YOLO-v5 model, characterized in that, Includes the following steps: Step 1: Collect radial tire X-ray images, perform segmentation processing, unify resolution, and create model training sample data; Step 2: Perform image restoration processing; the specific steps are as follows: Let the radial tire X-ray image be a third-order tensor. Let n1, n2, and n3 represent the dimensions of each order, and the constructed restoration model is as follows: (1) in, For a clean image to be solved, N represents the tensor. The order of FCTN is given by the fact that FCTN is a fully connected tensor network. This represents the N factor tensors obtained from the fully connected tensor decomposition. It is an implicit regularizer based on deep learning. For regularization weight parameters, For noise reduction parameters, represent The set of pixel indices observed in the image. Representatives Perform mapping; The model is solved by alternating iterative updates using the following algorithm: (2) in, This represents a two-dimensional denoising convolutional neural network trained on a natural image dataset. Represents the number of iterations. Represents each factor tensor. These are proximal parameters. Represents a set The complement; Step 3: Design and improve the YOLO-v5 model; The improved YOLO-v5 model structure includes an input end, a backbone network, a bottleneck network, and a prediction end. The input end performs image preprocessing after receiving the input image, scaling the input image to the network's input size and performing normalization. The backbone network is used to extract general feature representations. The bottleneck network is located between the backbone network and the prediction end, and is used to further improve the diversity and robustness of features. The prediction end is used to output the object detection results. The YOLO-v5 model improvement process is as follows: Step 3.1: Based on the original YOLO-v5 model, add a fourth detection layer to upsample the feature map with the largest output scale, further expanding the feature map. At the same time, cascade and fuse the obtained feature map with the same size in the backbone network. Step 3.2: Add an attention module to each detection layer feature processing stage; the attention module works as follows: The intermediate feature map T is subjected to MaxPool max pooling and AvePool average pooling in the spatial dimension, respectively, to obtain two pooling vectors. These vectors are then input into a shared multilayer mapping neural network (MLP) for nonlinear mapping, resulting in two new vectors. These two vectors are then summed bit-by-bit and further nonlinearly mapped using the Sigmoid activation function to obtain the channel attention module. The transformation formula is as follows: (3) Where A1 represents the channel attention module, Represents the Sigmoid function; The intermediate feature maps T are subjected to MaxPool max pooling and AvePool average pooling operations in the channel dimension, then stacked, and then mapped to a single-band, same-size model through convolution operation to obtain the spatial attention module A2. (4) in, This indicates that a 1×1 convolutional neural network is used for feature transformation. This represents the Sigmoid function, and its equation is: (5) Where x represents the input tensor; Step 3.3: Improve the loss function; add an impact factor, measure it on the aspect ratio of the detection box, and further use DIoU_P, which is more consistent with the regression mechanism. The formula is: (7) Where IoU represents the overlap ratio between the predicted bounding box and the ground truth bounding box, and b and These represent the center points of the detection box and the target box, respectively; represents the Euclidean distance between two center points; c represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted and ground truth boxes. is a parameter used to balance the scale; m is used to measure the proportional consistency between the width and height of the predicted box and the ground truth box. Step 4: Use radial tire defect data for model training; Step 5: Use the trained model to detect tire defects in real-world application scenarios.

2. The automatic defect detection method for radial tire X-ray images based on the improved YOLO-v5 model according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 1.1 First, for an original X-ray image of a radial tire, reduce its width to 1900 pixels and its height proportionally. The corresponding actual annotation box in the defect description file is also reduced accordingly. Step 1.2: Use a 1900*1900 window to slide along the height direction in the original image, and set the window to slide 10 times; Step 1.3: If the window does not contain any defects while sliding, continue sliding forward; otherwise, correct the location information of the defects and generate sub-images and corresponding defect description files.

3. The automatic defect detection method for radial tire X-ray images based on the improved YOLO-v5 model according to claim 1, characterized in that, The specific steps of step 4 are as follows: Step 4.1: Process the model training sample data created in Step 1 to establish a dataset including defects in X-ray images of tires with different meridians. Label the image information and classify it to obtain the training dataset. Step 4.2: Build the improved YOLO-v5 network model, set the training path, and read the training parameters; Step 4.3: Load the pre-trained model and obtain the initial network model parameters; Step 4.4: Select the training optimizer; Step 4.5: Load the training set; Step 4.6: Start training, update network model parameters until the iteration ends, and model training ends.

4. The automatic defect detection method for radial tire X-ray images based on the improved YOLO-v5 model according to claim 1, characterized in that, The specific steps of step 5 are as follows: Step 5.1: Acquire the current original radial tire X-ray image in real time and perform image preprocessing, including image segmentation, resolution unification, and image restoration; Step 5.2: After obtaining the preprocessed image, input it into the trained improved YOLO-v5 model, and then perform image partitioning; Step 5.3: Based on the different locations of the defects, they are divided into local defects and global defects; by further extracting detailed features in different areas of the tire, corresponding defect detection is performed, and the model provides detection information, including the location information of the defects in the image and the predicted confidence levels. Step 5.4: Then, perform confidence filtering. Based on a pre-set threshold, only retain candidate boxes with a confidence level greater than the threshold and display them on the current detection image to obtain the current tire defect detection status.

Images