Road defect target detection method and system based on SLCD-YOLO network

Abstract

This invention discloses a road defect target detection method and system based on the SLCD-YOLO network. The method comprises the following specific steps: S1. Selecting and preprocessing a road damage dataset; S2. Configuring a training environment; S3. Improving the YOLOv8 architecture to obtain an SLCD-YOLO network model; S4. Integrating the SLCD-YOLO network model obtained in step S3 into the training environment configured in step S2 and adding the specified parameter file to the network model; and finally, using a partitioned training set and validation set of images for model training and validation. This invention is suitable for real-time, efficient road defect monitoring and identification.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology and relates to technologies such as target detection, defect detection and road damage classification. In particular, it relates to a road defect target detection method and system based on the SLCD-YOLO network. Background Art

[0002] Problems such as cracks and potholes in road use have a serious impact on overall road quality and safety. Transportation agencies need to maintain high-quality roads in good condition to ensure road traffic safety and improve the efficiency of trunk roads. Timely detection and repair of cracks and potholes is a key step in ensuring driving safety. Traditional methods for detecting road defects include manual inspection and the use of multi-purpose road inspection vehicles. Manual inspection is time-consuming and labor-intensive, and can compromise the accuracy of inspection results due to fatigue, potentially leading to missed or false positives. In contrast, multi-purpose road inspection vehicles utilize high-tech equipment such as GPS, cameras, LiDAR, and ground-based radar to more accurately and efficiently detect road cracks. However, the high cost of these high-tech devices limits their widespread application. With the rapid development of computer hardware and distributed computing, deep learning has surpassed traditional algorithms in multiple fields. To address this issue, deep learning has made significant progress in road defect detection in recent years. Deep learning eliminates the need for manual feature extraction and simply feeds a large number of training images into the model, thereby improving detection efficiency and accuracy. However, the large number of parameters in deep learning models leads to performance bottlenecks in real-time applications. This results in relatively slow detection speeds in demanding real-time applications such as traffic monitoring systems and autonomous vehicles, potentially leading to latency issues and limiting the model's practicality. Therefore, addressing the large number of parameters and slow detection speeds in deep learning models is a key area of ​​focus and improvement, aiming to make them more suitable for real-time, efficient road defect monitoring and identification. Summary of the Invention

[0003] In view of the above problems in the prior art, the present invention proposes a road defect target detection method and system based on the SLCD-YOLO network. In order to reduce the computational complexity of the model, the present invention introduces a new lightweight cross-scale feature fusion module CCFM in the neck network (Neck). The CCFM module compresses the feature map extracted by the backbone network (Backbone) in the channel dimension, reducing the computational complexity of Neck; at the same time, in order to further improve the inference speed, the SimSPPF module is introduced in Backbone. The ReLU activation function used by SimSPPF has a faster inference speed than SiLU. In order to improve the detection accuracy of targets such as road cracks and small potholes, the present invention introduces the LSKA feature selection module in Backbone. This module can expand the receptive field of the model and more effectively extract the multi-scale features of the target; in order to further obtain multi-scale contextual information, the present invention constructs the C2f_DWRSeg module in Neck. This module introduces an efficient multi-scale feature extraction method DWRSeg, which uses multi-rate depth-direction dilated convolution to improve the efficiency of obtaining multi-scale information in road detection. In order to further preserve the required original detail information of the target at multiple angles, the present invention introduces an adaptive lightweight upsampling operator CARAFE at the neck. This improvement improves the detail preservation and reconstruction quality during the upsampling process, enabling the network to generate clearer and more accurate output.

[0004] The technical solutions adopted by the present invention to achieve its technical purpose are as follows:

[0005] The road defect target detection method based on the SLCD-YOLO network has the following specific steps:

[0006] S1. Select a road damage dataset and preprocess it;

[0007] S2. Configure the training environment;

[0008] S3. Improve the YOLOv8 structure to obtain the SLCD-YOLO network model;

[0009] S4. Integrate the SLCD-YOLO network model obtained in step S3 into the training environment configured in step S2, and add the set parameter file to the network model; use the divided training set and validation set images to train and verify the model.

[0010] Preferably, step S1, preprocessing of the data set:

[0011] The public RDD2022 dataset is used. The RDD2022 dataset is a dataset about road damage. However, since the format cannot be directly used by YOLO and a large number of images in it do not have the required annotations, a script needs to be written to clean and format the dataset before the experiment.

[0012] Preferably, step S2, configure the training environment:

[0013] Build the environment according to the environment required by the model and the basic hardware conditions, and reasonably configure parameters such as the number of training rounds and the category of the detection target.

[0014] Preferably, step S3 improves the existing YOLOv8 model structure:

[0015] The improvement of YOLOv8 structure results in the SLCD (where S stands for SimSPPF, L for LSKA, C for CARAFE, and D for DWRSeg)-YOLO network model. The specific improvements are as follows:

[0016] S3.1. Use the SimSPPF module in Backbone for faster inference. This module uses the ReLU activation function to solve the problem of gradient vanishing or gradient exploding when the data is too large or too small. The expression of the ReLU activation function is as follows:

[0017]

[0018]

[0019] S3.2. A new lightweight cross-scale feature fusion module CCFM is used to integrate features of different scales through fusion operations to enhance the model's adaptability to scale changes and its ability to detect small-scale objects.

[0020] S3.3. Use the LSKA feature selection module as the output layer after the backbone. LSKA adaptively aggregates feature information from large kernels in the spatial dimension, which enables the model to quickly, efficiently, and adaptively focus on the characteristics of key areas. The LSKA feature output expression is as follows:

[0021]

[0022] in, represents the output of global spatial feature information, d is the expansion rate, H and W are the height and width of the feature map respectively, C is the number of input channels, and F represents the input feature map.

[0023] S3.4. Use the DWRSeg dilated residual module. The DWRSeg module uses multi-rate depth-wise dilated convolution and performs morphological filtering using the desired receptive field on each concise feature map, simplifying the process of ordinary convolution to obtain complex semantic information; a two-step method is used to effectively extract multi-scale contextual information, and then the feature maps generated by the multi-scale receptive field are fused. This module is then fused with C2f to form a new C2f_DWRSeg module, so that the feature model can more effectively integrate features at different levels.

[0024] The two-step method is as follows:

[0025] The first step generates relevant residual features from the input features, called regional residuals;

[0026] The second step uses multi-rate extended depth-wise convolution to perform morphological filtering on regional features of different sizes, which is called semantic residualization. The formula for the second step is as follows:

[0027] F=[d1,d2,d3...,dk]

[0028] Where k represents the dilation rate, dk represents the convolution block group with dilation rate k, and F represents the output feature.

[0029] S3.5. A lightweight upsampling operator (CARAFE) is used for upsampling. This module consists of a kernel prediction module and a content-aware reassembly module. Given an input feature map of shape C*H*W, the kernel prediction module first predicts the upsampling kernel, and then the content-aware reassembly module performs the upsampling. Where C is the number of input channels, and H and W are the width and height of the feature map, respectively.

[0030] The kernel prediction module and content-aware reconstruction module are formulated as follows:

[0031]

[0032]

[0033] Among them, ψ is the kernel prediction module, φ is the content-aware reconstruction module, represents the k*k region centered on l, k encoder represents the size of the recombinant nucleus at the target location, k up represents the recombinant nuclear size.

[0034] CARAFE performs the following steps: First, the recombinant kernel of each target position is predicted, and then the predicted kernel is used to recombinant features; given a feature map of size C*H*W and an upsampling ratio μ, CARAFE generates a new feature map of size C*μH*μW; secondly, the kernel prediction module of CARAFE generates a position-specific kernel based on the content of the input feature, and for each specific kernel The feature reorganization module reorganizes the features in the local area through the function φ. The kernel prediction module maps each position in the output feature map back to the input feature map and extracts the position centered on it. The content-aware reconstruction module uses the kernel to reconstruct features.

[0035] Preferably, step S4 trains the network model:

[0036] After obtaining the improved SLCD-YOLO network model, we integrated it into the previously configured training environment and added the pre-set parameter file to the network model. We then used the divided training and validation sets for model training and validation. During training, we tested images selected from the test set on a computer to assess the effectiveness of each training phase.

[0037] Preferably, after step S4, step S5 is performed to analyze the results:

[0038] After training, the model generates a corresponding weight file. Then, you import these trained weights, along with the image to be tested and its corresponding labels. After running the program, you can obtain the tested data and images and compare the recognition results and detection accuracy to see if they meet the expected requirements.

[0039] The present invention also discloses a road defect target detection system based on the SLCD-YOLO network, which includes the following modules based on the above method:

[0040] Data acquisition module: selects road damage dataset and performs preprocessing;

[0041] Training environment configuration module: configure the training environment;

[0042] Network model construction module: Improve the YOLOv8 structure to obtain the SLCD-YOLO network model;

[0043] Training and validation module: Integrate the obtained SLCD-YOLO network model into the configured training environment and add the set parameter file to the network model; use the divided training set and validation set images to train and validate the model.

[0044] The beneficial effects of the present invention are:

[0045] This paper discloses a road defect target detection method and system based on the SLCD-YOLO network. To improve the model's inference speed, the present invention introduces SimSPPF into Backbone. The inference speed of a single ConvBNReLU module in SimSPPF is 18% faster than that of ConvBNSiLU in SPPF. Furthermore, to further reduce the model's computational complexity, the present invention introduces a lightweight cross-scale feature fusion module (CCFM) into Neck. CCFM downsamples the feature maps extracted by Backbone in the channel dimension, reducing the number of channels in the output feature maps and effectively reducing the model's computational complexity. To address the accuracy issues in small target defect detection, the present invention introduces the LSKA feature selection module into Backbone. The LSKA feature selection module dynamically selects appropriate convolution kernels by considering the local information of the input feature map to adapt to the contextual information of various target types. To better obtain multi-scale information, the present invention introduces the DWRSeg module into Neck and fuses it with C2f to form a new C2f_DWRSeg module. C2f_DWRSeg uses multi-rate depth-wise dilated convolution to effectively extract multi-scale contextual information, while utilizing the residual network to more effectively fuse features at different levels. To further improve defect detection at multiple angles, the present invention introduces the lightweight upsampling operator CARAFE into Neck. CARAFE divides sampling into two steps: prediction and reorganization. During the sampling process, more original low-level feature information, such as texture, color, and structure, is retained. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a diagram of the overall model structure of a preferred embodiment of the present invention.

[0047] Figure 2 This is a comparison chart of the effects of the LSKA feature selection module.

[0048] Figure 3 This is the CCFM structure diagram.

[0049] Figure 4 This is the C2f_DWRSeg module structure diagram.

[0050] Figure 5 This is a schematic diagram of the CARAFE lightweight upsampling operator.

[0051] Figure 6 This is a comparison chart of detection effects.

[0052] Figure 7 This is a block diagram of a road defect target detection system based on an SLCD-YOLO network in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention is further described below with reference to the accompanying drawings and specific examples. However, it should be noted that the present invention is not limited to the following examples. Those skilled in the art will recognize that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of protection of the present invention should not be limited to the contents disclosed in the examples.

[0054] This embodiment provides a road defect target detection method based on the SLCD-YOLO network, and the specific steps are as follows:

[0055] S1. Data acquisition:

[0056] First, we obtain a dataset of road defects. This example uses the RDD2022 dataset. RDD2022 contains 47,420 images from six different countries, with varying sizes and angles. The resulting dataset is diverse and includes four categories of defects, totaling 55,007 labels. After data cleaning and formatting, 23,767 images remain, split into a training set and a validation set with a 7:3 ratio. The RDD2022 dataset is about road damage, but because its format cannot be directly used by YOLO and a large number of images lack the required annotations, a script is required to clean and format the dataset.

[0057] S2.yolov8 model dataset configuration

[0058] After processing the dataset, create a Road_yaml file for the dataset, write the paths of the training set and validation set and the categories of road defects in Road_yaml, and then modify the parameters such as the number of training times and batch-size under default.yaml as required. The environment of the present invention is: cuda12.2, deep learning framework pytorch1.12.1, Intel corei9-13900ks CPU, 128G memory, GPU is NVIDIA GeForce RTX 4090, and video memory is 24G.

[0059] S3. Improve the existing yolov8 model structure and combine Figure 1 Description: The main improvements are as follows:

[0060] (1) This paper introduces SimSPPF and LSKA into Backbone, which reduces the amount of computation while enabling the model to adaptively focus on road defect areas. The ReLU activation function used by SimSPPF is more concise than the SiLU activation function used by SPPF in activating regional features. The activation function expression is as follows:

[0061]

[0062]

[0063] (2) The present invention introduces a lightweight cross-scale feature fusion module CCFM, which performs a downsampling operation on the feature maps of different scales extracted from Backbone to reduce the number of output channels of the feature maps, thereby reducing the amount of calculation, and then uses upsampling and splicing operations to fuse the feature maps of different scales, thereby effectively integrating feature information. Figure 3 As shown in the figure, CCFM integrates features of different scales through fusion operations to enhance the model's adaptability to scale changes and its ability to detect small target objects. At the same time, CCFM can also effectively integrate detailed features and contextual information, thereby improving the overall performance of the model.

[0064] (3) The LSKA feature selection module is used to select the most important areas in the image, i.e., the areas with road defects, e.g. Figure 2 As shown, Figure 2 (a) is without adding LSKA module, Figure 2 (b) After adding, the color depth of the region clearly shows where the model is focusing in the image. LSKA is a feature selection module that uses large-kernel convolution to extract features. Large-kernel convolution (LKA) is a convolution operation with a large kernel. This large kernel can capture more contextual information at once, helping the model recognize large-scale features and patterns in the image. Large-kernel convolution combines the advantages of convolution and self-attention, not only considering local contextual information and a large receptive field, but also incorporating dynamic processes. However, large convolution kernels require multiplication and addition operations on more elements, which increases the computational resource requirements. To address this issue, LSKA introduces a series of large separable kernel attention modules, which decompose the 2D convolution kernels of the depthwise convolution layer into cascaded horizontal and vertical 1-D kernels. Compared to the standard large-kernel attention mechanism, this design can directly use the depthwise convolution layers with large kernels without any additional modules, significantly reducing computational complexity and memory usage. By decomposing the 2D weight kernels of depthwise convolution and depthwise dilated convolution into two cascaded 1D separable weight kernels, an equivalent and improved configuration of LKA can be obtained. This innovative design enables LSKA to maintain the advantages of large kernel convolution while reducing the computational burden. The feature output expressions of LKA and LSKA are as follows:

[0065]

[0066]

[0067] in, represents the output of global spatial feature information, d is the dilation rate, H and W are the height and width of the feature map, respectively, C is the number of input channels, and F represents the input feature map. As shown in the above formula, LSKA uses (2d-1)×1 and 1×(2d-1) convolution kernels instead of the (2d-1)×(2d-1) convolution kernel, which can effectively reduce the quadratic growth of parameters in the calculation without degrading performance. In this embodiment, the value of d is 11.

[0068] (4) The present invention introduces a DWR module that efficiently obtains multi-scale information, such as Figure 4 As shown in the figure, this module can help the model better understand and process targets and scenes at different scales, and put this module in the residual network of Bottleneck to fuse into a new module C2f_DWRSeg, which can better fuse the extracted feature information and thus obtain better detection performance.

[0069] The DWR module is designed in a residual manner. Within the residual, a two-step method is used to effectively extract multi-scale contextual information, and then the feature maps generated by the multi-scale receptive fields are fused. The first step is to generate relevant residual features from the input features, which is called regional residual. In this step, a series of concise feature maps in the form of regions of different sizes are generated. The second step is to perform morphological filtering on regional features of different sizes using multi-rate extended depth-wise convolution, which is called semantic residualization. Only a single desired receptive field is applied to each channel feature to prevent possible redundant receptive fields. In practical applications, the required concise regional feature map can be judiciously learned in the first step according to the size of the receptive field in the second step to quickly match the receptive field.

[0070] To achieve this step, the regional feature maps are first divided into several groups, and then different groups are subjected to dilated depth convolution at different rates. Figure 4 As shown, the second step uses three different colors of convolution to represent convolutions with expansion rates of 1, 3, and 5, respectively, to simply extract the regional features after the output of the first convolution. The formula for the second step is as follows:

[0071] F=[d1,d2,d3...,dk]

[0072] Where k represents the expansion rate, dk represents the convolution block group with expansion rate k, and F represents the output feature. The present invention adopts three groups of convolutions with different expansion rates: d1, d3, and d5.

[0073] Through the two-step region residualization-semantic residualization method, the role of multi-rate depth-wise dilated convolution changes from laboriously acquiring as much complex semantic information as possible to simply performing morphological filtering using a desired receptive field on each concisely expressed feature map, simplifying the processing and thus ordering the learning process. As a result, multi-scale contextual information can be more effectively preserved.

[0074] (5) To address the problem of inconsistent object scales at different angles, the present invention adopts an upsampling method called CARAFE to enhance the convolutional neural network feature map. CARAFE is mainly divided into two modules: the kernel prediction module and the content-aware reconstruction module. The formulas for the two modules are as follows:

[0075]

[0076]

[0077] Among them, ψ is the kernel prediction module, φ is the content-aware reconstruction module, represents the k*k region centered on l, k encoder represents the size of the recombinant nucleus at the target location, k up represents the recombinant nuclear size.

[0078] CARAFE has two steps: first, predict the recombinant kernel for each target position, and then recombinant the features using the predicted kernel. Given a feature map of size C*H*W and an upsampling ratio μ, CARAFE will generate a new feature map of size C*μH*μW. Secondly, the kernel prediction module of CARAFE generates a position-specific kernel based on the content of the input features. For each specific kernel The feature reorganization module reorganizes the features in the local area through the function φ. In the kernel prediction module, for each position in the output feature map, it maps it back to the input feature map and takes out the position centered on it. The area of ​​​​the point is dot-producted with the predicted upsampling kernel to obtain the output kernel map, and then the content-aware reconstruction module uses these kernels to reconstruct features.

[0079] like Figure 5 As shown in Figure 2, CARAFE predicts a recombinant kernel by leveraging the underlying content information at each location and recombines features within a predefined nearby region. Due to the introduction of content information, CARAFE can use adaptive and optimized recombinant kernels at different locations, thus performing better than mainstream upsampling operators such as interpolation or deconvolution.

[0080] S4. Train with the modified model: Integrate the obtained SLCD-YOLO network model into the configured training environment and add the set parameter file to the network model; use the divided training set and validation set images to train and verify the model.

[0081] Specifically, in this embodiment, the data set is divided into a training set and a validation set according to a ratio of 7:3, the training coefficient is set to 100 rounds, 16 pictures are input for each training, and the training process is observed in real time through wandb during training. After the training is completed, the trained weights are saved. The following is an illustration of the effects achieved by the present invention with reference to the accompanying figures and data. In order to further test the effect of the model, a comparative test was conducted. The experimental results are shown in Table 1. It can be seen that the model proposed by the present invention not only reduces the number of parameters and the amount of calculation, but also reduces the AP 50 Increased by 2 percentage points, AP 50-95 An increase of 1.2 percentage points.

[0082] Table 1

[0083]

[0084] In order to further demonstrate the effect achieved by the invention, Figure 6 Explanation: Compare the detection graph of the improved model with the existing model. The left side is the detection graph of the improved model, and the right side is the detection graph of the original model. Figure 6 As shown in (a), it can be seen that under small target detection, the cracks that cannot be identified by the existing technology can be successfully identified by the improved model of the present invention; Figure 6 As shown in (b), under complex backgrounds, the existing model may miss detection, but the improved model of the present invention can accurately identify it; Figure 6 As shown in (c), in the improved model inference results, small targets in the details of the car shooting perspective are also successfully identified; comparison Figure 6 (a) and (c), it can be seen that the detection performance of the proposed model at multiple angles is better than that of the existing model.

[0085] like Figure 7 As shown, this embodiment discloses a road defect target detection system based on the SLCD-YOLO network, based on the above method embodiment, including the following modules:

[0086] Data acquisition module: selects road damage dataset and performs preprocessing;

[0087] Training environment configuration module: configure the training environment;

[0088] Network model construction module: Improve the YOLOv8 structure to obtain the SLCD-YOLO network model;

[0089] Training and validation module: Integrate the obtained SLCD-YOLO network model into the configured training environment and add the set parameter file to the network model; use the divided training set and validation set images to train and validate the model.

[0090] For other contents of this embodiment, please refer to the above method embodiment.

[0091] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that a person skilled in the art can easily conceive within the technical scope disclosed in the present application should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. The road defect target detection method based on the SLCD-YOLO network is characterized by: The specific steps are as follows: S1. Select a road damage dataset and preprocess it; S2. Configure the training environment; S3. Design the SLCD-YOLO network model. The details of this step are as follows: S3.

1. Use the SimSPPF module in Backbone. The expression of the ReLU activation function used in this module is as follows: S3.2, using a lightweight cross-scale feature fusion module CCFM to integrate features of different scales through fusion operations; S3.

3. Use the LSKA feature selection module as the output layer after the backbone. LSKA adaptively aggregates feature information from large kernels in the spatial dimension. The LSKA feature output expression is as follows: in, Represents the output of global spatial feature information, is the expansion rate, are the height and width of the feature map respectively, C is the number of input channels, Represents the input feature map; S3.

4. Use the DWRSeg dilated residual module. The DWRSeg dilated residual module uses multi-rate depth-wise dilated convolution to perform morphological filtering with the desired receptive field on each concise feature map. A two-step approach is used to extract multi-scale contextual information. Feature maps generated by the multi-scale receptive fields are then fused. This module is then fused with C2f to form the C2f_DWRSeg module. In this step, the two-step approach is as follows: The first step generates relevant residual features from the input features, called regional residuals; The second step uses multi-rate extended depth-wise convolution to perform morphological filtering on regional features of different sizes, which is called semantic residualization. The formula for the second step is as follows: Where k represents the expansion rate, dk represents the convolution block group with expansion rate k, and F represents the output feature; S3.5, use a lightweight upsampling operator CARAFE for sampling operation, the upsampling operator includes a kernel prediction module and a content-aware reconstruction module; given a shape of The input feature map is first predicted by the kernel prediction module to predict the upsampling kernel, and then the content-aware reconstruction module is used to complete the upsampling; S4. Integrate the SLCD-YOLO network model obtained in step S3 into the training environment configured in step S2, and add the set parameter file to the network model; use the divided training set and validation set images to train and verify the model.

2. The road defect target detection method based on the SLCD-YOLO network as claimed in claim 1, characterized in that: In step S1, use the public RDD2022 dataset and write a script to clean and format the dataset.

3. The road defect target detection method based on the SLCD-YOLO network as claimed in claim 1, characterized in that: In step S2, the environment is built according to the environment required by the model and the basic hardware conditions, and the training rounds and the categories of detection targets are configured.

4. The road defect target detection method based on the SLCD-YOLO network as claimed in claim 1, characterized in that: In step S3.5, the kernel prediction module and content-aware reconstruction module in CARAFE are formulated as follows: in, is the kernel prediction module, is the content-aware reorganization module, It represents the k*k area centered on l. represents the size of the recombinant nucleus at the target location, represents the recombinant nuclear size.

5. The road defect target detection method based on the SLCD-YOLO network as claimed in claim 4, characterized in that: The CARAFE execution steps are as follows: first, the recombinant core of each target position is predicted, and then the predicted core recombination features are used; Given a size of The feature map and an upsampling ratio , CARAFE generates a new size of Secondly, the kernel prediction module of CARAFE generates a location-specific kernel based on the content of the input features. For each specific kernel , the feature recombination module uses the function The features in the local area are reorganized. The kernel prediction module maps each position in the output feature map back to the input feature map and extracts the position centered on it. The content-aware reconstruction module uses the kernel to reconstruct features.

6. The road defect target detection method based on the SLCD-YOLO network as claimed in claim 1, characterized in that: In step S4, during the training process, images selected from the test set are tested to obtain the effect of training at each stage.

7. The road defect target detection method based on the SLCD-YOLO network according to any one of claims 1 to 6, characterized in that: After step S4, proceed to step S5 and analyze the results: after the model training is completed, generate a corresponding weight file, import the trained weights, the image to be tested, and the corresponding label; Get the detected data and pictures, and compare the recognition effect and detection accuracy.

8. A road defect target detection system based on an SLCD-YOLO network, based on the method according to any one of claims 1 to 7, characterized in that: Includes the following modules: Data acquisition module: selects road damage dataset and performs preprocessing; Training environment configuration module: configure the training environment; Network model building module: design SLCD-YOLO network model; Training and validation module: Integrate the obtained SLCD-YOLO network model into the configured training environment and add the set parameter file to the network model; use the divided training set and validation set images to train and validate the model.

Images