Road crack detection method based on HSW-YOLOv8

Abstract

The invention provides a road crack detection method based on HSW-YOLOv8, and aims to improve the accuracy and speed of crack detection. According to the method, firstly, ground buildings, roads and the like are shot through high-altitude flight of an unmanned aerial vehicle, a high-resolution image data set containing cracks is obtained, and the data set is marked and divided; then, on the basis of the YOLOv8n network, the HGNetV2 is used for replacing a backbone network of the YOLOv8n network, and efficient and lightweight feature extraction is achieved through the combination of HGStem, HGBlock and DWConv. An LSKA attention mechanism is introduced into the SPPF module to enlarge the receptive field of the network and realize long dependency, and meanwhile, a Wise-IoU loss function is introduced to optimize the regression precision of the crack boundary. Through the trained and optimized HSW-YOLOv8 network model, the category, position and bounding box information of the crack in the image can be rapidly and accurately output, and the recognition precision of the algorithm is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of crack image recognition and analysis, and specifically relates to a road crack detection method based on HSW-YOLOv8. Background Art

[0002] In fields such as civil engineering, building structures, and infrastructure maintenance, crack detection is crucial for ensuring structural safety and stability. Traditional crack detection methods rely primarily on visual inspection or simple optical devices. These methods are not only inefficient and subjective, but also struggle to accurately locate and classify cracks. These shortcomings are particularly evident in large or inaccessible areas, such as bridges, high-rise building facades, and tunnels. With the rapid development of deep learning in recent years, the use of drones equipped with high-resolution cameras for crack detection has become an efficient and viable solution. However, existing deep learning-based crack detection methods still have some shortcomings. Many existing deep learning models have high computational complexity when processing high-resolution images, resulting in slow inference speed and difficulty meeting the requirements of real-time detection. Furthermore, these models typically require significant computing resources and storage space, limiting their application on resource-constrained devices.

[0003] Crack features are often subtle and complex, especially in low-contrast or noisy images. Traditional convolutional neural networks (CNNs) can struggle to effectively extract crack feature information. Furthermore, the diverse shapes and orientations of cracks complicate detection. Existing deep learning-based crack detection methods mostly use traditional loss functions, which are inadequate when dealing with fuzzy crack boundaries. This results in low-accuracy regression of the model's crack boundaries, thus impacting detection accuracy. Data labeling is time-consuming and labor-intensive, while existing models have limited generalization capabilities across different scenarios, making them difficult to adapt to diverse crack types and environmental conditions.

[0004] To address the above issues, the present invention proposes a road crack detection method based on HSW-YOLOv8. This method aims to improve the efficiency, accuracy, and generalization ability of crack detection while reducing the computational complexity and resource consumption of the model by optimizing the network architecture, introducing an efficient feature extraction module, and an improved loss function. By acquiring crack image data using a high-resolution camera mounted on an unmanned aerial vehicle and combining it with efficient annotation tools, the present invention can quickly generate a high-quality training dataset. In terms of model design, the present invention uses HGNetV2 as the backbone network and introduces the LSKA attention mechanism in the SPPF module to further enhance feature extraction capabilities. In addition, by optimizing the model training process using the Wise-IoU loss function, the present invention can more accurately locate crack boundaries and improve detection accuracy and robustness. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a road crack detection method based on HSW-YOLOv8 with high detection accuracy and fast detection speed.

[0006] To achieve the above objectives, the present invention proposes a road crack detection method based on HSW-YOLOv8, comprising the following steps:

[0007] 1) Using a drone to fly high above the ground to capture images of the ground, obtaining a training dataset and dividing the dataset. The training dataset includes multiple drone-photographed image samples labeled with the type and location of the target to be detected. Image enhancement is then performed on the images to enhance the image feature information.

[0008] 2) Based on the YOLOv8n network, HGNetV2 is used to replace the YOLOv8n backbone network. The LSKA attention mechanism is introduced in the SPPF module to expand the network receptive field and achieve long-term dependency correlation. The module that introduces the LSKA attention mechanism is called the SPPF_LSKA module. Finally, the Wise-IoU loss function is introduced to obtain the HSW-YOLOv8 network model.

[0009] 3) Use the training set images to train and optimize the HSW-YOLOv8 network model, and obtain the final HSW-YOLOv8 network model after 300 iterations;

[0010] 4) Input the test set into the model to obtain evaluation indicators such as precision and recall;

[0011] 5) Input the image taken by the drone to be detected into the final HSW-YOLOv8 network model, and use the optimized HSW-YOLOv8 network model to obtain the category and location of the target to be detected in the drone image to be detected.

[0012] In step 2), the HGNetV2 backbone network consists of three parts: HGStem, HGBlock, and DWConv. HGStem, as the initial preprocessing layer of the network, contains multiple convolutional layers, which are used to extract initial features from the original input image and provide high-quality feature representation for subsequent modules. HGBlock is the core of HGNetV2. Its design is inspired by the concept of hierarchical feature extraction. Each HGBlock can be regarded as a small subnetwork that can process different levels of abstraction of data, allowing the network to gradually learn rich semantic information from low-level to high-level. HGBlock internally combines depthwise separable convolution (DWConv) and pointwise convolution (PWConv), which maintains a high feature extraction capability while reducing the number of parameters, and introduces a channel attention mechanism to dynamically adjust channel weights to further enhance feature expression capabilities. DWConv is widely used throughout the network. By decomposing the standard convolution into depth-wise convolution and point-by-point convolution, it significantly reduces the number of parameters and computational complexity while retaining sufficient feature expression capabilities. This allows HGNetV2 to maintain high accuracy while significantly reducing inference latency, making it particularly suitable for use in resource-constrained environments.

[0013] These design features of HGNetV2 give it a significant advantage in crack detection tasks. Crack features are often subtle and complex, especially in low-contrast or noisy images. Traditional convolutional neural networks can struggle to effectively extract these features. HGNetV2, through its hierarchical feature extraction and lightweight convolution operations, efficiently processes high-resolution images while maintaining high-precision capture of crack features. This design not only improves the model's detection accuracy but also significantly reduces computing resource requirements, making it more adaptable to the crack detection needs of real-world projects.

[0014] In step 2), the SPPF_LSKA module introduces the LSKA attention mechanism into the original SPPF module to expand its network receptive field and achieve long-term dependency correlation. This effectively improves the feature extraction capabilities of the backbone feature extraction network and enhances its ability to identify the target to be detected, thereby improving detection results. The LSKA mechanism also introduces depthwise separable convolution, which ensures that while network performance is improved, its computational complexity and memory usage are not excessive. The core of LKA is to decompose a large convolution kernel into three parts: depthwise convolution, depthwise dilated convolution, and pointwise convolution. This decomposition strategy not only inherits the advantages of large kernel convolution in terms of large receptive field, achieving detailed attention to local areas of the image, but also optimizes the model size and makes it more lightweight. Compared with traditional attention mechanisms, LKA places greater emphasis on the weight allocation of each pixel in the feature map, which is particularly effective when processing crack image datasets. Large kernel decomposition helps alleviate the quadratic increase in computational cost caused by using large kernel sizes for deep convolution alone. The output of LKA can be obtained as follows:

[0015]

[0016] A C =W 1×1 *Z C (3)

[0017]

[0018] Where C is the number of input channels, H and W represent the height and width of the feature map respectively, and d is the dilation rate. Denotes the output of the depthwise convolution, whose kernel size is (2d-1)×(2d-1), which captures the local spatial information and compensates for the grid effect of the subsequent depthwise convolution. The kernel size of the depthwise convolution is in Denotes the floor operation. The dilated depthwise convolution is responsible for capturing the global spatial information of the depthwise convolution output. is the attention map A C And the input feature map F C The Hadamard product.

[0019] By splitting the 2D weight kernel of depthwise convolution and depthwise dilated convolution into two cascaded 1D separable weight kernels, an equivalent improved LKA structure can be obtained. This modified configuration of the LKA module is called LSKA. The output of LSKA is shown below:

[0020]

[0021] A C =W 1×1 *Z C(7)

[0022]

[0023] In step 2), the Wise-IoU loss function optimizes the regression of the fuzzy crack boundary by dynamically adjusting the anchor box weights. The calculation formula is:

[0024]

[0025] L WIOUv1 =R WIOU L IOU (10)

[0026]

[0027] As a preferred solution of the road crack detection system based on HSW-YOLOv8 described in the present invention, it includes: an image acquisition module, an image enhancement module, a model design module, a loss function optimization module, a model training module and a target detection module.

[0028] The image acquisition module is responsible for using drones equipped with high-resolution cameras to capture ground buildings and roads, acquiring a dataset of drone images containing cracks. This module ensures that the captured images are high-resolution and can clearly capture the details of the cracks, providing high-quality raw data for subsequent crack detection.

[0029] The image enhancement module is used to perform data enhancement operations on the acquired dataset images to enhance the feature information of the images and improve the robustness of the subsequent models;

[0030] The model design module is based on the YOLOv8n network, replacing its backbone network with HGNetV2 and introducing the LSKA attention mechanism in the SPPF module to form the SPPF_LSKA module. This module optimizes the network structure to enhance the model's ability to extract crack features, while reducing computational complexity and improving model efficiency and accuracy.

[0031] The loss function optimization module replaces the original CIOU loss function with the Wise-IoU loss function. This module optimizes the regression of fuzzy crack boundaries by dynamically adjusting the anchor box weights, improving the model's detection accuracy and enhancing the model's robustness.

[0032] The model training module is responsible for inputting the divided training set into the optimized network model for training and optimization. After 300 rounds of iteration, the final network model is obtained. This module uses multiple rounds of iterative training to ensure that the model can learn the complex patterns of crack characteristics and improve the model's generalization ability.

[0033] The target detection module is used to input the crack image to be detected into the trained network model and output the category, location and bounding box information of the crack in the image.

[0034] A computer device includes a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the road crack detection method based on HSW-YOLOv8 when executing the computer program.

[0035] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the road crack detection method based on HSW-YOLOv8 are implemented.

[0036] Based on the YOLOv8 target detection algorithm, this paper introduces the LSKA attention mechanism and the HGNetV2 backbone network, significantly improving the model's ability to extract crack features. By introducing the LSKA attention mechanism in the SPPF module, the model can effectively expand the receptive field and capture long-term dependency correlations, which is particularly important for the subtle and complex texture features that need to be identified in crack detection. At the same time, HGNetV2 is used to replace the YOLOv8n backbone network, leveraging its efficient feature extraction and lightweight design to significantly reduce the number of model parameters and computational complexity. In addition, the Wise-IoU loss function is introduced to optimize the regression accuracy of crack boundaries, further improving the model's accuracy in locating crack boundaries.

[0037] The present invention significantly enhances the performance of YOLOv8 in crack detection tasks by optimizing the network architecture and introducing an efficient feature extraction module. The LSKA attention mechanism effectively expands the network's receptive field and enhances the ability to capture crack features through depthwise separable convolution and dilated convolution. The lightweight design of the HGNetV2 backbone network enables the model to maintain high accuracy while significantly reducing the demand for computing resources and improving the model's operating efficiency. The introduction of the Wise-IoU loss function further optimizes the regression accuracy of crack boundaries, enabling the model to more accurately locate crack boundaries. Experimental results on a drone image crack detection dataset show that the present invention has achieved significant improvements in both detection accuracy and efficiency.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The method of the present invention effectively improves the accuracy and speed of road crack detection and significantly reduces the memory usage and computational complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1Flowchart of the road crack detection method based on HSW-YOLOv8.

[0041] Figure 2 This is the network model structure diagram of HSW-YOLOv8.

[0042] Figure 3 This is the HGStem structure and HGBlock structure diagram.

[0043] Figure 4 This is the structure diagram of the SPPF-LSKA module.

[0044] Figure 5 This is the principle diagram of the Wise-IoU loss function.

[0045] Figure 6 This is a partial display of the data sets used in the experiments in the specific embodiments.

[0046] Figure 7 This is a performance comparison chart of HSW-YOLOv8 and other models in a specific embodiment.

[0047] Figure 8 This is a training loss graph of the YOLOv8n model in a specific embodiment.

[0048] Figure 9 This is a training loss graph of the HSW-YOLOv8 model in a specific embodiment.

[0049] Figure 10 This is a comparison chart of the PR curves of the YOLOv8n and HSW-YOLOv8 models in a specific embodiment.

[0050] Figure 11 This is a comparison chart of the prediction results of the YOLOv8n and HSW-YOLOv8 models in a specific embodiment.

[0051] Figure 12 This is a comparison chart of the prediction results of the YOLOv8n and HSW-YOLOv8 models in a specific embodiment. DETAILED DESCRIPTION

[0052] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] Example

[0054] See also Figure 1 The road crack detection method based on HSW-YOLOv8 provided in this embodiment includes the following steps:

[0055] 1) Build a dataset

[0056] The experiment uses the open source UAV-PDD2023 dataset for road damage detection, which contains 2,440 high-resolution images covering a variety of weather conditions and road types. The dataset is annotated into six typical road damage types: Longitudinal crack, Transverse crack, Alligator crack, Oblique crack, Repair, and Pothole. Some images were taken within 1 hour after the rain, which increased the complexity and diversity of the dataset. The training set, validation set, and test set were divided into a ratio of 7:1:2;

[0057] 2) Based on the YOLOv8n network, HGNetV2 is used to replace the YOLOv8n backbone network. The LSKA attention mechanism is introduced in the SPPF module to expand the network receptive field and achieve long-term dependency correlation. The module that introduces the LSKA attention mechanism is called the SPPF_LSKA module. Finally, the Wise-IoU loss function is introduced to obtain the HSW-YOLOv8 network model.

[0058] 3) Use the dataset constructed in step 1) to train and optimize the HSW-YOLOv8 network model, and obtain the final HSW-YOLOv8 network model after 300 iterations;

[0059] 4) Input the test set into the model to obtain evaluation indicators such as precision and recall;

[0060] 5) Input the crack image to be detected into the trained network model, and output the category, location and bounding box information of the crack in the image.

[0061] The experimental environment of this example is as follows: operating system: Ubuntu 20.04; CPU: 15vCPU Intel(R) Xeon(R) Platinum 8474C; GPU: NVIDIA GeForce RTX 4090D; video memory: 24G; memory: 80GB; CUDA version: 11.8; algorithm integration framework using Pytorch 2.0; programming language: Python 3.8;

[0062] The present embodiment will be further described below with reference to the accompanying drawings:

[0063] like Figure 1As shown in the figure, this paper proposes a complete process for road crack detection based on HSW-YOLOv8. This process begins with the acquisition of drone images, continues through dataset construction and partitioning, model optimization and improvement, training optimization, and finally crack detection, clearly presenting the entire process from data collection to practical application.

[0064] like Figure 2 As shown in the figure, the overall architecture of the HSW-YOLOv8 network model consists of three main modules: the backbone, neck, and head. In the backbone, HGNetV2 replaces the traditional YOLOv8n backbone network. Through the combination of HGStem, HGBlock, and DWConv, efficient and lightweight feature extraction is achieved. In the neck, the LSKA attention mechanism is introduced in the SPPF module, forming the SPPF_LSKA module. This significantly expands the network's receptive field and enhances its ability to capture crack features. The head is responsible for the final crack detection output.

[0065] like Figure 3 As shown in the figure, HGStem, as the initial preprocessing layer of the network, quickly extracts initial image features through efficient convolution operations, laying the foundation for subsequent feature extraction. HGBlock, the core of the network, uses a hierarchical feature extraction approach combined with depthwise separable convolution and channel attention mechanisms to enhance feature expression while reducing the number of parameters. This is particularly important for the subtle and complex texture features that need to be identified in crack detection.

[0066] like Figure 4 As shown in Figure 2, the present invention introduces the LSKA attention mechanism to construct the SPPF-LSKA module structure. By utilizing depthwise separable convolutions and dilated convolutions, this module significantly expands the network's receptive field and implements long-range dependency correlations. This design enables the model to better capture the global characteristics of cracks while maintaining attention to local details, thereby improving crack detection accuracy.

[0067] like Figure 5 As shown in Figure 2, this paper introduces the Wise-IoU loss function. Compared to the traditional CIOU loss function, Wise-IoU introduces an additional penalty mechanism and an improved calculation method to more accurately reflect the geometric differences between the predicted and true bounding boxes, thereby improving positioning accuracy. Especially when processing bounding boxes with little overlap, Wise-IoU provides a larger gradient, which promotes faster model convergence during training and enhances robustness in various complex scenarios.

[0068] like Figure 6 As shown in Figure 2, some images of the dataset used in the experiment are shown.

[0069] like Figure 7 The figure below compares the accuracy of the HSW-YOLOv8 model of the present invention with other common object detection models. The experimental results show that HSW-YOLOv8 significantly reduces the computational complexity and memory usage of the model while maintaining high detection accuracy.

[0070] like Figure 8 and Figure 9 As shown in the figure, the training losses of the YOLOv8n and HSW-YOLOv8 models are shown. It can be seen that both models have reached convergence after 300 rounds of training.

[0071] like Figure 10 As shown in the figure, the PR curves of the YOLOv8n and HSW-YOLOv8 models are shown. The area enclosed by the PR curve and the horizontal and vertical axes is the average precision (AP) of the model. It can be seen that the area enclosed by the PR curve of the HSW-YOLOv8 model and the horizontal and vertical axes is significantly larger than the area enclosed by the PR curve of the YOLOv8n model and the horizontal and vertical axes. This shows that the recognition accuracy of the HSW-YOLOv8 model is significantly higher than that of the YOLOv8n model.

[0072] like Figure 11 and Figure 12 The figure below shows a comparison of the prediction results of the YOLOv8n and HSW-YOLOv8 models. The left figure shows the prediction results of the YOLOv8n model, and the right figure shows the prediction results of the HSW-YOLOv8 model. Figure 10 It can be seen that the prediction accuracy of the HSW-YOLOv8 model is significantly better than that of the YOLOv8n model; Figure 11 It can be seen that the YOLOv8n model has some cracks that are missed, while the HSW-YOLOv8 model can more comprehensively detect all cracks in the image.

[0073] Through the detailed description of the above drawings, the innovation and advantages of the road crack detection method based on HSW-YOLOv8 of the present invention are clearly demonstrated.

[0074] The above content is a further detailed description of the present invention in conjunction with specific embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A road crack detection method based on HSW-YOLOv8, characterized in that: The following steps are involved: 1) Use a drone equipped with a high-resolution camera to photograph ground buildings and roads to obtain a dataset of drone images containing cracks. Each image obtained is annotated using labelimg. The dataset is then divided into a training set, a test set, and a validation set. Each image in the dataset is annotated with the location and category of the crack. 2) Based on the YOLOv8n network, HGNetV2 was used to replace the YOLOv8n backbone network. The Large Separable Kernel Attention (LSKA) mechanism was introduced in the SPPF module to expand the network receptive field and achieve long-term dependency correlation. The module that introduced the LSKA attention mechanism was named SPPF_LSKA module. Finally, the Wise-IoU loss function was used to replace the original CIOU loss function. 3) Input the divided training set into the network model for training and optimization, and obtain the final network model after 300 iterations; 4) Input the test set into the model to obtain evaluation indicators such as precision and recall; 5) Input the crack image to be detected into the trained network model, and output the category, location and bounding box information of the crack in the image.

2. A road crack detection method based on HSW-YOLOv8 as claimed in claim 1, characterized in that: In step 2), the HGNetV2 backbone network consists of three parts: HGStem, HGBlock, and DWConv. HGStem serves as the initial preprocessing layer of the network. It efficiently extracts initial features by stacking multiple small convolution kernels, while reducing the amount of parameters and computation, and providing high-quality feature representation for subsequent modules. HGBlock is the core of HGNetV2. Its design is inspired by the concept of hierarchical feature extraction. Each HGBlock can be regarded as a small subnetwork that can process different levels of abstraction of data, allowing the network to gradually learn rich semantic information from low to high levels. HGBlock internally combines depthwise separable convolution (DWConv) and pointwise convolution (PWConv), which reduces the amount of parameters while maintaining a high feature extraction capability. It also introduces a channel attention mechanism to dynamically adjust channel weights to further enhance feature expression capabilities. DWConv is widely used throughout the network. By decomposing standard convolution into depthwise and pointwise convolutions, it significantly reduces the number of parameters and computational complexity while retaining sufficient feature representation. This allows HGNetV2 to maintain high accuracy while significantly reducing inference latency, making it particularly suitable for use in resource-constrained environments. The overall architectural design of HGNetV2 fully considers the balance between computational efficiency and accuracy. Through hierarchical feature extraction, lightweight convolution operations, and channel attention mechanisms, it achieves efficient application in a variety of tasks.

3. The method according to claim 1, wherein The SPPF_LSKA module introduces the LSKA attention mechanism into the original SPPF module to expand its network receptive field and achieve long-term dependency correlation. It can effectively improve the feature extraction capability of the backbone feature extraction network and enhance the recognition capability of the target to be detected, thereby improving the detection effect. The LSKA mechanism also introduces depthwise separable convolution, which ensures that while the network performance is improved, the computational complexity and memory usage will not be too large. The core of LKA is to decompose a large convolution kernel into three parts: depthwise convolution, depthwise dilated convolution, and pointwise convolution. This decomposition strategy not only inherits the advantages of large kernel convolution in terms of large receptive field, achieving detailed attention to local areas of the image, but also optimizes the model size, making it more lightweight. Compared to traditional attention mechanisms, LKA places greater emphasis on weighting each pixel in the feature map, and its effectiveness is particularly pronounced when processing crack image datasets. Large kernel decomposition helps alleviate the quadratic increase in computational cost caused by using large kernel sizes for depthwise convolution alone. The output of LKA can be obtained as follows: AND C =In 1×1 *WITH C (3) Where C is the number of input channels, H and W represent the height and width of the feature map respectively, and d is the dilation rate. Denotes the output of the depthwise convolution, whose kernel size is (2d-1)×(2d-1), which captures the local spatial information and compensates for the grid effect of the subsequent depthwise convolution. The kernel size of the depthwise convolution is in Denotes the floor operation. The dilated depthwise convolution is responsible for capturing the global spatial information of the depthwise convolution output. is the attention map A C And the input feature map F C The Hadamard product. By splitting the 2D weight kernel of depthwise convolution and depthwise dilated convolution into two cascaded 1D separable weight kernels, an equivalent improved LKA structure can be obtained. This modified configuration of the LKA module is called LSKA. The output of LSKA is shown below: AND C =In 1×1 *WITH C (7) 4. The method according to claim 1, wherein The Wise-IoU loss function optimizes the regression of the fuzzy crack boundary by dynamically adjusting the anchor box weights. The calculation formula is: L WIOUv1 =R WIOU L IOU (10) Among them, Wg and Hg represent the width and height of the minimum bounding box of the predicted box and the real box respectively, W i and H i It describes the width and height of the intersection area of ​​the two frames, S u Represents the overlapping area between the predicted box and the true box.

5. A road crack detection system based on HSW-YOLOv8 according to any one of claims 1 to 4, characterized in that: It includes image acquisition module, image enhancement module, model design module, loss function optimization module, model training module and target detection module. The image acquisition module is responsible for using drones equipped with high-resolution cameras to capture ground buildings, roads, or bridges, generating a dataset of drone images containing cracks. This module ensures that the captured images are high-resolution and clearly capture the details of the cracks, providing high-quality raw data for subsequent crack detection. The image enhancement module is used to perform data enhancement operations on the acquired dataset images to enhance the feature information of the images and improve the robustness of the subsequent models; The model design module is based on the YOLOv8n network, replacing its backbone network with HGNetV2 and introducing the LSKA attention mechanism in the SPPF module to form the SPPF_LSKA module. This module optimizes the network structure to enhance the model's ability to extract crack features, while reducing computational complexity and improving model efficiency and accuracy. The loss function optimization module replaces the original CIOU loss function with the Wise-IoU loss function. This module optimizes the regression of fuzzy crack boundaries by dynamically adjusting the anchor box weights, improving the model's detection accuracy and enhancing the model's robustness. The model training module is responsible for inputting the divided training set into the optimized network model for training and optimization. After 300 rounds of iteration, the final network model is obtained. This module uses multiple rounds of iterative training to ensure that the model can learn the complex patterns of crack characteristics and improve the model's generalization ability. The target detection module is used to input the crack image to be detected into the trained network model and output the category, location and bounding box information of the crack in the image.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the road crack detection method based on HSW-YOLOv8 are implemented as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the road crack detection method based on HSW-YOLOv8 are implemented according to any one of claims 1 to 5.

Images