A damage detection method for smart rail roads

Abstract

A method for detecting damage to smart rail roads belongs to the field of damage detection technology, and includes: (1) identifying magnetic nail damage, marking line damage, cracks, and potholes in smart rails as damage types, extracting features of each damage type, and constructing a data set for these damage types; (2) constructing an improved smart rail pavement damage detection network YOLOv8-CMDA; (3) training and accuracy verification of the network YOLOv8-CMDA; and (4) constructing a graphical user interface (GUI). The model provided by the present invention has excellent detection performance; in actual use, images can be collected through devices such as mobile phones, high-definition cameras, and drones, and input into the GUI interface developed by the present invention for automatic damage detection. The method can output detection anchor frames based on the damage, and can also output detection losses based on the detection results, providing data support for subsequent staff to carry out reinforcement and repair.

Description

Technical Field

[0001] The present invention belongs to the field of damage detection technology, and specifically relates to a damage detection method for a smart rail road. Background Art

[0002] Damage detection methods for smart rail tracks can be categorized as non-destructive testing and destructive testing. Currently, damage detection methods used for smart rail tracks include core drilling, ultrasonic testing, and infrared thermal imaging. The core drilling method involves drilling samples from the road for laboratory analysis, including compressive strength and crack distribution. While accurate and suitable for material strength assessment, this method damages the smart rail pavement, requiring repair and resulting in high testing costs. Non-destructive testing primarily utilizes ultrasonic testing technology. Ultrasonic testing utilizes ultrasonic probes to inspect the internal structure of the track, analyzing the echo signal for defects such as cracks and cavities. This testing method is suitable for structural health monitoring. While this method offers high accuracy, it suffers from low efficiency, complex equipment, and high manual operation costs.

[0003] Existing deep learning algorithms still face bottlenecks in detection accuracy. For example, while object detection algorithms have achieved some success in road damage detection, these methods rely on large amounts of labeled data. Real-world data on road damage and road stud conditions under different environments is difficult to obtain, limiting the model's generalization capabilities. Furthermore, road damage types are complex (e.g., cracks, potholes, and wear) and vary in form. Environmental factors (e.g., varying lighting, pollutant coverage, and rain and snow) further reduce the stability and accuracy of detection systems. Summary of the Invention

[0004] In order to solve the problems existing in the prior art, the present invention proposes a damage detection method for smart rail roads, aiming to realize the automatic detection of smart rail pavement damage by using the YOLO target detection technology in deep learning. Based on the image target detection method in deep learning, the present invention constructs a smart rail road pavement image dataset for detection network training and verification; based on the YOLOv8n framework, the CMUNeXt module is integrated, and the dual attention mechanism DANet is added to the C2f-CSP module in the YOLOv8 model, and the feature pyramid structure is improved, and dynamic convolution is added to the FPN structure. In order to show the experimental results more intuitively, three loss functions, cls-loss, box-loss, and dfl-loss, are used to measure the detection results. In order to verify the robustness of the model, the model is verified using a public dataset. The results show that: whether it is a constructed dataset or a public dataset, the model has good detection performance. At the same time, the present invention constructs a GUI interface to reduce the difficulty of damage detection and improve work efficiency.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] A damage detection method for a smart track road, comprising:

[0007] (1) Define magnetic nail damage, marking line damage, cracks, and potholes in smart tracks as damage types, extract the features of each damage type, and construct a dataset for these damage types for training and verification of the YOLO network;

[0008] (2) Constructing an improved intelligent rail pavement damage detection network YOLOv8-CMDA based on multiple YOLO models;

[0009] (3) Training and accuracy verification for the network YOLOv8-CMDA;

[0010] (4) Build a graphical user interface (GUI).

[0011] Preferably, in the step (1), the data set is expanded using image data augmentation technology; and the data set is divided into a training set: validation set ratio of 7:3, and data augmentation is performed using four methods: rotation transformation, flip transformation, color transformation, and random cropping.

[0012] Preferably, in step (2), the multiple YOLO models refer to: YOLOv3, YOLOv5 and YOLOv8.

[0013] Preferably, in the step (2), YOLOv8n is used as the backbone network, the backbone network is added with a CMUNeXt module, the CSP-C2f module is improved, DANet is added, the FPN structure is enriched, and dynamic convolution DC is added.

[0014] Preferably, in the step (2), the CMUNeXt module replaces the ordinary convolution kernel with a large-kernel depth convolution and two point-by-point convolutions. In the CMUNeXt module network architecture, the smoothing function in the jump-out fusion block is used to connect the encoder and the decoder. The CMUNeXt Block module is divided into three convolutions. First, the global information of each channel is extracted by depth convolution, and the next convolution layer is connected through the residual. The last two convolutions are point-by-point convolutions. Among them, the two point-by-point convolutions are designed through an inverted bottleneck to extract higher-dimensional feature information.

[0015] Preferably, in step (2), DANet is a dual attention network that combines spatial attention and channel attention. The spatial attention module focuses on the key information areas in the image by modeling the relationship between different positions in the image, thereby strengthening the expression of spatial features; while the channel attention module optimizes the dependency between channels by weighting the features of each channel, thereby improving the representation ability of global features.

[0016] Preferably, in the step (2), in order to enhance the adaptability of FPN to deformed objects, dynamic convolution DC is introduced into the hierarchy of FPN, thereby forming an FPN-DC model, wherein the feature extraction capability is enhanced by introducing dynamic convolution DC in the P3 and P5 layers of FPN, that is, DC is used instead of the standard 3×3 convolution in the P3 layer to improve the adaptability of the receptive field to small targets and enhance the alignment capability of edge details; DC is introduced in the P5 layer so that high-level features can dynamically adjust the sampling position, better model global semantic information, and improve the robustness of target detection; at the same time, the P4 layer still uses the standard 3×3 convolution to maintain the stability of information transmission.

[0017] Preferably, in step (3), the Adam optimizer is used to adjust the model parameters to minimize the loss function, and the dynamic loss scaling technique is used to avoid the gradient explosion problem.

[0018] The beneficial effects of the damage detection method of a smart track road of the present invention are:

[0019] Experimental verification shows that the model provided by the present invention has excellent detection performance, regardless of whether it is a constructed dataset or a public dataset. In actual use, images can be collected through devices such as mobile phones, high-definition cameras and drones, and input into the GUI interface developed by the present invention for automatic damage detection. It can output detection anchor frames based on the damage, and can also output detection losses based on the detection results, providing data support for subsequent staff to carry out reinforcement and repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Image annotation interface for Labelme.

[0021] Figure 2 Schematic diagram for data augmentation.

[0022] Figure 3 This is the YOLOv8n network architecture (Note: Backbone: backbone network; Head: head structure; Bottleneck: bottleneck structure; Conv: convolutional layer).

[0023] Figure 4 It is the CMUNeXt Block architecture (Note: Depthwise Conv: depthwise separable convolution; GELU: activation function; Pointwise Conv: point-by-point convolution).

[0024] Figure 5 This is the C2f-CSP DANet module (Note: Conv: convolutional layer; Bottleneck: bottleneck structure).

[0025] Figure 6It is the FPN-DC structure (Note: Conv: convolution layer; DC: dynamic convolution).

[0026] Figure 7 This is the YOLOv8n-CMDA network structure diagram (Note: Backbone: backbone network; Head: head structure).

[0027] Figure 8 This is the training result of the YOLOv8n-CMDA model.

[0028] Figure 9 It is a graphical user interface (GUI). DETAILED DESCRIPTION

[0029] The following description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0030] The following embodiments may be understood as individually expressing a part of a local structure or method of the present invention, or may be understood as a combination of the embodiments to explain the connotation of a larger structure or method of the present invention.

[0031] Example 1:

[0032] A damage detection method for a smart track road, comprising:

[0033] (1) Define magnetic nail damage, marking line damage, cracks, and potholes in smart tracks as damage types, extract the features of each damage type, and construct a dataset for these damage types for training and verification of the YOLO network;

[0034] (2) Based on multiple YOLO models (including YOLOv3, YOLOv5, and YOLOv8), an improved intelligent rail pavement damage detection network YOLOv8-CMDA was constructed;

[0035] (3) The network YOLOv8-CMDA was trained and its accuracy was verified. By training and testing on the same dataset, the experimental results showed that the network achieved an accuracy of 0.894. To verify the generalization ability of the model, a public crack dataset was also used for testing. The results showed that the detection accuracy on the Crack500 dataset was 0.7 to 0.8, and the accuracy on other public datasets was between 0.8 and 0.85, which provides solid technical support for future damage detection.

[0036] (4) Build the Gui interface.

[0037] Example 2:

[0038] The step (1) includes: the data set used in the present invention is obtained by field photography, and a total of 2,000 damage pictures are collected, including 500 pictures of road cracks and cracks, 1,000 pictures of road spike damage (protrusions, exposed spikes, and depressions), and 500 pictures of mixed road and spike damage; the acquisition equipment is a drone DJI Air2s and a HUAWEI P40 mobile phone; the drone is equipped with a Sony 1 / 2-inch image sensor with an aperture of F2.8, which can take high-quality photos and videos and supports 48-megapixel photos. During the shooting process, fixed-track flight shooting is adopted; the HUAWEI P40 mobile phone adopts the Ultra Vision Leica triple-camera system, including a 50-megapixel super-sensing camera (1 / 1.28-inch sensor); the pictures taken are all square, and in order to improve the generalization ability of the model and avoid overfitting in the training process, the data set is expanded using image augmentation technology; and the data set is divided into a training set: validation set = 7:3 ratio. The expanded data set is processed using labelme, and the data set is calibrated by a closed graph connected end to end, such as Figure 1 As shown;

[0039] Data augmentation is to generate new training samples by applying multiple transformation strategies to the original data, thereby effectively expanding the size of the data set, improving the generalization ability of the model, and enhancing its robustness to data changes and noise in complex environments; the present invention uses the following four data augmentation methods, such as Figure 2 shown.

[0040] (1) Rotation transformation: Rotate the original image at a set angle (e.g., 30°, 150°, 270°) to obtain sample images at different angles, thereby improving the model's adaptability to direction changes;

[0041] (2) Flip transformation: randomly flip the image horizontally or vertically to increase data diversity and reduce the model's dependence on specific direction information;

[0042] (3) Color transformation: Adjust image parameters such as brightness, contrast, hue, and saturation to simulate the effects of different lighting conditions and imaging devices on the data, and enhance the robustness of the model in various visual environments;

[0043] (4) Random cropping: Randomly crop the image to simulate the changes of the target object at different sizes and positions, thereby improving the robustness of the model to scale changes and displacement changes.

[0044] Example 3:

[0045] In step (2), the improved network architecture is as follows: Figure 7As shown, the YOLOv8n backbone network incorporates the CMUNeXt module to enhance feature extraction capabilities and improve computational efficiency. The CMUNeXt module, through efficient convolution operations and multi-scale feature extraction, makes the network highly adaptable to various visual tasks. To adapt intelligent rail damage detection to complex scenarios, the CSP-C2f module was improved and DANet was added, enabling YOLOv8n to more intelligently select key feature areas and enhance its ability to identify important targets. This enables the model to maintain high detection accuracy in complex scenarios, especially when the target is small or partially occluded. Due to the variability in the position, shape, and size of the detected target, the FPN structure was enriched and a dynamic convolutional structure was added, enabling YOLOv8n-CMDA to better handle targets of varying sizes and shapes, particularly in detection tasks against complex backgrounds.

[0046] Currently, although the YOLO series performs well in target detection tasks, there is still room for improvement in accuracy and generalization capabilities in complex scenarios. To this end, YOLOv8n introduces a self-attention mechanism, an adaptive anchor box mechanism, and an optimized loss function strategy, and improves the backbone network, adding a CMUNeXt module to enhance multi-scale target detection capabilities. At the same time, YOLOv8n uses dynamic label allocation and efficient data augmentation technology to improve detection accuracy while maintaining real-time advantages, making it more valuable in autonomous driving, intelligent monitoring and other fields. Figure 3 shown.

[0047] The YOLOv8n architecture consists of three main components: Backbone, Neck, and Head. The YOLOv8 network's primary detection process begins with the backbone network extracting grid features at three different scales (80*80, 40*40, and 20*20). After feature extraction, the network enters the Neck structure, where the extracted features are fused and subsequently decoded. Decoding involves two types of decoding: bonding box decoding and class loss decoding, enabling detection of objects with varying damage types. The YOLOv8 network has five model versions with different scales: n, s, l, m, and x. With each version, the network depth and detection accuracy of these modules continue to improve. To ensure lightweight detection of road studs and road damage, this study selected the lightweight YOLOv8n version. Using a model with a depth of n effectively reduces design parameters and significantly improves object detection speed.

[0048] The following describes in detail the improvement of the backbone network according to the present invention:

[0049] (1) Improvement of backbone network:

[0050] In the task of intelligent track damage detection, global information helps the model better locate and identify specific objects. In order to effectively extract global context information and reduce calculation parameters, the present invention uses the CMUNeXt module. The CMUNeXt module replaces the ordinary convolution kernel with a large-kernel depth convolution and two point-by-point convolutions, fully considering the randomness of the intelligent track damage location, and uses the smoothing function in the jump-out fusion block in the CMUNeXt network architecture to connect the encoder and decoder. The CMUNeXt Block module is mainly divided into three convolutions. First, the depth convolution is used to extract the global information of each channel, and the next convolution layer is connected through the residual. The last two are point-by-point convolutions. The two point-by-point convolutions are designed with an inverted bottleneck to extract higher-dimensional feature information.

[0051] (2) Dual Attention Network (DANet):

[0052] DANet (Dual Attention Network) is a network architecture that combines the Position Attention Module (PAM) and the Channel Attention Module (CAM). The PAM models the relationship between different positions in an image, focusing on key information areas and enhancing the representation of spatial features. The CAM, on the other hand, weights the features of each channel to optimize inter-channel dependencies and improve the representation of global features. By combining these two attention mechanisms, DANet significantly improves the accuracy and robustness of feature extraction in multi-scale object detection and complex scenes.

[0053] The present invention adds DANet to the YOLOv8n backbone network, where CSP-C2f is the core module for feature extraction in YOLOv8n. By embedding PAM and CAM into this module, the model can simultaneously enhance the expression capabilities of spatial features and channel features. In each CSP-C2f module, DANet applies spatial and channel attention modules to the feature map after preliminary convolution processing, thereby focusing on important feature areas and strengthening key channel information. First, convolution dimensionality reduction (3×3 convolution + BN + SiLU) is performed, and DANet is applied to the feature map after convolution to enhance the focus on the target area. After the output feature is completed, it is fused with another branch of CSP (Shortcut connection). For detailed operations, see Figure 5 .

[0054] (3) Enriching feature pyramid structure:

[0055] Enrich the feature pyramid structure and introduce dynamic convolution (DC). The Feature Pyramid Network (FPN) is a network architecture that enhances image representation through multi-scale feature fusion. FPN adopts a top-down structure and constructs a feature pyramid containing information at different scales by upsampling and fusing high-level semantic feature maps to low-level spatial feature maps layer by layer. Although FPN performs well in multi-scale feature learning, its fixed convolution operation still has certain limitations when processing deformed objects. To enhance FPN's adaptability to deformed objects, dynamic convolution is introduced into the FPN layer to form the FPN-DC model. This combination aims to simultaneously leverage the advantages of FPN in multi-scale feature extraction and the flexibility of DC in dynamic sampling, thereby significantly improving the model's performance in complex visual tasks.

[0056] Deformable convolution is an improved version of the convolution operation. It introduces learnable offsets so that the convolution kernel can dynamically adjust the sampling position to adapt to the deformation of the target. Its mathematical expression is as follows:

[0057] ;

[0058] Where, Represents the standard sampling position of the convolution kernel, is the learnable offset.

[0059] The improved FPN-DC proposed in this paper primarily introduces deformable convolution (DC) in the P3 and P5 layers to enhance feature extraction. Specifically, DC replaces the standard 3×3 convolution in the P3 layer to improve receptive field adaptability for small objects and enhance the alignment of edge details. The introduction of DC in the P5 layer enables dynamic adjustment of sampling positions for high-level features, better modeling of global semantic information and improving the robustness of object detection. Meanwhile, the P4 layer retains the standard 3×3 convolution to maintain stable information transfer. By rationally arranging the DC structure, this paper improves detection accuracy while maintaining superior computational efficiency, making it particularly suitable for multi-scale object detection in complex scenarios.

[0060] Example 4:

[0061] During the training process of the YOLOv8n-CMDA model, selecting appropriate training parameters plays a crucial role in model performance. By properly setting key parameters such as the learning rate, batch size, number of iterations, and loss function, combined with strategies such as data augmentation and optimizers, the model's accuracy and robustness in object detection tasks can be significantly improved. This paper uses the Adam optimizer to adjust model parameters to minimize the loss function and utilizes dynamic loss scaling to avoid the exploding gradient problem. The momentum parameter of the Adam optimizer is typically controlled by two hyperparameters: β1 and β2. The decay rate of the first-order moment estimate controls the mean estimate of the gradient. It determines the length of the optimizer's memory of past gradient values. To avoid excessive reliance on historical gradient information, the value of β1 is reduced. The decay rate of the second-order moment estimate controls the variance estimate of the gradient. It determines the optimizer's responsiveness to gradient changes. To provide more accurate second-order moment estimates, the value of β1 is increased. Therefore, β1 is set to 0.937 and β2 to 0.999. The initial learning rate was set to 0.01, and the learning rate decay was 0.0001, decreasing by a factor of 0.1 every 10 epochs. Model checkpointing was enabled to save the best model during training, and training was stopped when the validation set accuracy did not improve for 10 consecutive epochs.

[0062] The present invention uses the above-mentioned mAP50 and mAP50-95 curves to represent the training results. The training results of the improved YOLOv8n-CMDA model are shown as follows:

[0063] Figure 8 (a) shows the model's detection performance using the mAP50 and mAP50-95 metrics. The data in the figure shows that the mAP50 (Box) metric is close to 0.9, demonstrating the model's excellent performance under conventional evaluation criteria. Furthermore, under the more stringent mAP50-95 (IoU 50-95%) evaluation, the model still achieves a high score exceeding 0.8, further validating its detection capability and robustness under this high-standard evaluation system. Figure 8 (b) Represents the recognition accuracy of anchor boxes detected by Cls-loss, box-loss, and dfl-loss. Figure 8 (cd) The training results are presented as a chaos matrix. The chaos matrix shows the performance of the classification model and uses a standardized matrix to evaluate the model's prediction accuracy for different categories. Each row of the matrix represents the true category, while each column represents the predicted category. Figure 8(c) The confusion matrix shows that the model has high classification accuracy for some categories (such as "crack" and "pit"), but there is some confusion between "concavity" and "background," and between "exposed nail" and "protrusion." To address these issues, the model's classification ability for these categories can be improved by optimizing data augmentation strategies, refining damage feature extraction methods, or adjusting the loss function weights.

[0064] Figure 8 (e) shows the effect of model Labels on detecting target locations. During the training and evaluation process of YOLOv8n-CMDA, Labels usually contain the object category ID, the coordinates of the bounding box (usually normalized coordinates relative to the image size), and other possible attributes (such as confidence, etc.). Figure 8 (e) shows a relatively even distribution of center points, but more annotation boxes are concentrated in the center of the image. Center points are less likely to be close to the edges, indicating that objects are primarily concentrated in the middle of the image. The distribution of the width and height of the annotation boxes is shown: most boxes have small widths and heights (close to 0), indicating that objects are generally small. Some boxes have widths and heights close to 1, possibly because large objects occupy the entire image. Figure 8 (f) represents the Labels-correlogram, which is used to show the correlation between labels predicted by the target detection algorithm during the training process. It is usually presented as a matrix, where the rows and columns represent different categories (or labels), and the values ​​in the matrix represent the strength of the correlation between the corresponding labels. Figure 8 As shown in (f), small objects (width and height close to 0) account for the majority of the image. There are a small number of large objects (width and height close to 1), indicating that the dataset contains large objects, but in a smaller proportion.

[0065] In summary, YOLOv8n-CMDA demonstrates exceptional performance in object detection tasks. Leveraging its efficient model architecture and optimization strategies, the model achieves remarkable results across multiple standard evaluation metrics. The improved model demonstrates exceptional accuracy and stability, both in the conventional mAP50 score and the more rigorous mAP50-95 test. Its superior detection performance not only ensures precise object recognition but also demonstrates remarkable robustness, enabling it to handle a variety of challenging scenarios in complex environments. This further validates its leading position in the field of object detection and its broad potential for application.

[0066] Example 5:

[0067] Ablation experiments were conducted on YOLOv8n-CMDA, and the results of various improvements were compared. Table 1 shows this. In this experiment, various improvements were made to the YOLOv8 model, including the CMUNeXt module, the dual attention mechanism (DANet), and the dynamic convolution (FPN-DC). The performance of these different configurations was compared. The results show that the CMUNeXt module alone improved the model's mAP@0.5 from 76.5% to 78.8%, while the dual attention mechanism increased mAP@0.5 to 78.2%. When the CMUNeXt module and the dual attention mechanism were combined, the model's mAP@0.5 significantly improved to 83.0%, demonstrating their complementary effect on feature extraction. Further adding the dynamic convolution (FPN-DC) further improved mAP@0.5 to 89.4%, demonstrating that the fusion of all modules can significantly enhance object detection capabilities. In addition, although the number of model parameters increased from 24M to 32M, the FPS remained at 98 frames per second, proving that while improving detection accuracy, the real-time performance of the model can still meet the needs of efficient target detection.

[0068] In summary, the combined use of the CMUNeXt module and the self-attention mechanism optimization strategy not only effectively improves the detection accuracy of the model, but also achieves higher performance at a reasonable inference speed, verifying the effectiveness of this improvement scheme in improving YOLOv8 detection results and inference efficiency.

[0069] Table 1 Comparison of ablation experiment results:

[0070]

[0071] Example 6:

[0072] To verify the robustness of the proposed model, the network model's detection performance was validated using public datasets. To more clearly demonstrate the experimental results, the number of iterations was set to 50. It can be seen that after 50 iterations, the detection results have almost stabilized, with no significant increase, fluctuating within 0.005. The proposed YOLOv8n-CMDA network achieved an mAP50 detection accuracy of 0.7-0.8 on the Crack500 public dataset and 0.8-0.85 on other public datasets, providing technical support for subsequent damage detection. Table 2 shows the mAP50 detection results on three public datasets.

[0073] Table 2. Performance of YOLOv8n-CMDA on other datasets:

[0074] Public datasets mAP50 / % mAP50-95 / % Crack500 71.2 58.4 CFD 82.6 60.8 CrackLS315 81.3 61.4

[0075] Example 7:

[0076] In order to simplify the operation process of damage target detection, the present invention designs a graphical user interface (GUI), such as Figure 9 As shown in the figure, this interface fully leverages the performance of PC devices, allowing users to launch and run program code and easily implement various functions with simple mouse operations. This design not only simplifies the process of calling and implementing programs, but also significantly improves the convenience and smoothness of human-computer interaction. There are many ways to implement a GUI (graphical user interface), the most popular of which are libraries such as Tkinter, PyQt, PySide, wxPython, and Kivy. When developing the GUI interface, Tkinter is used for interface design. Finally, the pyinstaller library is used to package the py file into an exe file, ensuring that users can seamlessly install and run this GUI interface on Windows systems.

Claims

1. A damage detection method for smart rail roads, characterized by: include: (1) Define magnetic nail damage, marking line damage, cracks, and potholes in smart tracks as damage types, extract the features of each damage type, and construct a dataset for these damage types for training and verification of the YOLO network; (2) Constructing an improved intelligent rail pavement damage detection network YOLOv8-CMDA based on multiple YOLO models; (3) Training and accuracy verification for the network YOLOv8-CMDA; (4) Build a graphical user interface (GUI); In step (2), YOLOv8n is used as the backbone network, and the CMUNeXt module is added to the backbone network; Improve the CSP-C2f module and add DANet; enrich the FPN structure and add dynamic convolution DC; In step (2), the CMUNeXt module replaces the ordinary convolution kernel with a large-kernel depth convolution and two point-by-point convolutions. In the CMUNeXt module network architecture, the smoothing function in the jump-out fusion block is used to connect the encoder and decoder. The CMUNeXt module is divided into three convolutions. First, the global information of each channel is extracted by depth convolution, and the next convolution layer is connected through the residual. The last two convolutions are point-by-point convolutions. Among them, the two point-by-point convolutions are designed with an inverted bottleneck to extract higher-dimensional feature information. In the step (2), dynamic convolution DC is introduced in the P3 and P5 layers of FPN to enhance the feature extraction capability, that is, DC is used instead of the standard 3×3 convolution in the P3 layer; DC is introduced in the P5 layer; and at the same time, the P4 layer still uses the standard 3×3 convolution.

2. A method for detecting damage to a smart track road as claimed in claim 1, characterized in that: In the step (1), the data set is expanded using image data augmentation technology; and the data set is divided into a training set: validation set ratio of 7:3, and data augmentation is performed using four methods: rotation transformation, flip transformation, color transformation, and random cropping.

3. A method for detecting damage to a smart track road as claimed in claim 2, characterized in that: In step (2), the multiple YOLO models refer to: YOLOv3, YOLOv5 and YOLOv8.

4. A method for detecting damage to a smart track road as claimed in claim 3, characterized in that: In the step (2), DANet is a dual attention network that combines spatial attention and channel attention. The spatial attention module focuses on the key information areas in the image by modeling the relationship between different positions in the image, thereby strengthening the expression of spatial features; while the channel attention module optimizes the dependency between channels by weighting the features of each channel, thereby improving the representation ability of global features.

5. A method for detecting damage to a smart track road as claimed in claim 4, characterized in that: In step (3), the Adam optimizer is used to adjust the model parameters to minimize the loss function, and the dynamic loss scaling technique is used to avoid the gradient explosion problem.

Images