Method for detecting pavement cracks by joint cutter based on YOLOv5-CSS model

Abstract

A method for detecting pavement cracks by a joint cutter based on a YOLOv5-CSS model comprises the following steps: acquiring a pavement image, and performing normalization processing; a pavement image is identified by using a trained YOLOv5-CSS model to obtain an identification result, the YOLOv5-CSS model comprises a backbone network introduced with a C2fRepGhost module, a neck structure introduced with an SE attention mechanism module and a detection head introduced with a SimAM attention mechanism module, and the identification method comprises the steps of inputting the pavement image into the C2fRepGhost module for feature extraction to obtain an intermediate feature map, and performing feature extraction on the intermediate feature map; the intermediate feature maps are processed by at least one RepGhostBottleneck module and then spliced to form an output feature map; an SE attention mechanism module is adopted to dynamically adjust the importance and adaptability of the features of the output feature map, and a fused feature map is obtained; a SimAM attention mechanism module is adopted to analyze the fused feature map to extract an important feature region; according to the method, the model calculation complexity is reduced, and the detection speed and accuracy are improved.

Description

Technical Field

[0001] The invention relates to the technical field of road crack recognition by a seam cutter, and in particular to a method for detecting road cracks by a seam cutter based on a YOLOv5-CSS model. Background Art

[0002] While the development of roads brings convenience, it also brings challenges to road maintenance and management. Various factors in daily life can cause damage to roads, such as cracks and potholes, which pose hidden dangers to traffic safety. Among them, road cracks are common damages that have an important impact on road safety, service life and driving comfort. Therefore, it is very important to identify road cracks in a timely and accurate manner. In the existing technology, road crack cutters are important equipment for identifying road cracks and road maintenance.

[0003] Traditional crack detection algorithms include ultrasonic detection and fiber optic sensing, but these methods are often costly and have low recognition accuracy. Therefore, target detection algorithms based on deep learning and neural networks are widely used, such as GooleNet, VGGNet, or SegNet. These algorithms can automatically detect targets and enable the seam cutter to have crack learning capabilities. However, the above algorithms have high requirements for data sets, often requiring huge data sets to meet training requirements, and have low detection accuracy.

[0004] The YOLOv5 model is a single-stage target detection algorithm with the advantages of high accuracy and fast training speed. However, the industrial scene of seam cutter detection contains a large number of complex details, requiring the detection system to respond quickly and accurately identify the key components and cutting lines of the equipment. For irregular targets such as road cracks, in order to ensure their recognition accuracy, more computing resources are required, thereby increasing the amount of calculation. When the seam cutter identifies road cracks, the captured road cracks are highly similar to the background features, so the detection process is easily disturbed, which leads to false detection. Summary of the invention

[0005] In order to solve the problems of low accuracy and large calculation amount in identifying pavement cracks by the YOLOv5 model in the prior art, the present invention provides a method for detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model, which reduces the calculation complexity of the model and improves the detection speed and accuracy.

[0006] In order to achieve the above object, the specific scheme adopted by the present invention is: a method for detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model, comprising the following steps:

[0007] Acquire road surface images and perform normalization processing;

[0008] The trained YOLOv5-CSS model is used to identify the road image and obtain the recognition result. The YOLOv5-CSS model includes a backbone network with a C2fRepGhost module, a neck structure with a SE attention mechanism module, and a detection head with a SimAM attention mechanism module. The recognition method includes:

[0009] The road surface image is input to the C2fRepGhost module for feature extraction to obtain an intermediate feature map. Some of the intermediate feature maps are processed by all RepGhostBottleneck modules and then concatenated with the remaining intermediate feature maps that are not processed by the RepGhostBottleneck module to form an output feature map.

[0010] The SE attention mechanism module is used to dynamically adjust the importance and adaptability of the output feature map to obtain a fused feature map;

[0011] The SimAM attention mechanism module is used to analyze the fused feature map and extract important feature areas;

[0012] The detection head is used to identify the feature area and output the recognition result.

[0013] As an optimization solution for the above-mentioned method of detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model: the C2fRepGhost module includes at least one RepGhostBottleneck module.

[0014] As another optimization scheme of the above-mentioned method for detecting road cracks by a seam cutter based on the YOLOv5-CSS model: the steps of forming an output feature map are:

[0015] The road surface image is input into the C2fRepGhost module, which generates an intermediate feature map through 1x1 convolution;

[0016] The intermediate feature map is divided into two parts, one of which is directly passed to the Concat module, and the other part of the intermediate feature map is passed to the Concat module after passing through at least one RepGhostBottleneck module to generate an indirect feature map;

[0017] The Concat module concatenates the intermediate feature map and the indirect feature map to obtain the output feature map.

[0018] As another optimization scheme of the above-mentioned method of detecting road cracks by a seam cutter based on the YOLOv5-CSS model: a method of fusing the output feature map using the SE attention mechanism module includes:

[0019] The output feature map is mapped and aggregated in the spatial dimension to form a C-dimensional vector through global average pooling. Each element in the C-dimensional vector represents the average value of the corresponding channel, and the C-dimensional vector serves as a global channel descriptor.

[0020] The descriptor is input into the fully connected layer and the channel weights are generated through the ReLU and Sigmoid activation functions;

[0021] The contribution of each channel is dynamically adjusted according to the channel weight, and finally a fusion feature map is formed.

[0022] As another optimization scheme of the above-mentioned method for detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model: the training process of the YOLOv5-CSS model includes the following steps:

[0023] Collect and preprocess pavement crack images to obtain a preprocessed data set;

[0024] The preprocessed dataset is divided into training set, validation set and test set according to 7:1:2;

[0025] The training set is input into the YOLOv5-CSS model for training, and the trained YOLOv5-CSS model is verified using the validation set to obtain a trained YOLOv5-CSS model;

[0026] Input the test set into the trained YOLOv5-CSS model.

[0027] As another optimization scheme for the above-mentioned method of detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model: the process of collecting pavement images is to use mobile imaging equipment to collect pavement crack images under different lighting, environments and positions.

[0028] As another optimization scheme for the above-mentioned method of detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model: the preprocessing method includes using the Labelme tool to label the collected pavement crack images and converting the labels into an XML format.

[0029] As another optimization scheme for the above-mentioned method of detecting road cracks by a seam cutter based on the YOLOv5-CSS model: the loss function of the YOLOv5-CSS model includes a confidence loss function and a bounding box regression loss function.

[0030] As another optimization scheme of the above-mentioned method of detecting road cracks by a seam cutter based on the YOLOv5-CSS model:

[0031] The confidence loss function is:

[0032]

[0033] Where N is the number of prediction boxes, y i is the true value, Indicates the confidence of the prediction;

[0034] The bounding box regression loss function is:

[0035]

[0036] Among them, l box is the bounding box regression loss, b is the predicted bounding box, is the true bounding box, IoU is the intersection over union ratio, ρ is the Euclidean distance between the center points of the predicted box and the true box, c is the diagonal distance of the smallest external box surrounding the predicted box and the true box, αv is the factor used to penalize the difference in aspect ratio, w / h is the aspect ratio of the target box, is the aspect ratio of the prediction box.

[0037] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the program is executed by a processor, the steps in the above-mentioned method of detecting pavement cracks by a seam cutter based on the YOLOv5-CSS model are implemented.

[0038] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a method for detecting pavement cracks in a seam cutter based on a YOLOv5-CSS model, wherein a C2fRepGhost module is introduced into the backbone network of the YOLOv5-CSS model, an SE attention mechanism is introduced into the neck structure, and a SimAM attention mechanism is introduced into the detection head, wherein the introduction of the C2fRepGhost module can reduce the amount of calculation and memory usage, speed up the detection and recognition speed, and enhance the gradient flow information, reduce the interference of redundant background features, and make the model adaptable to embedded devices and low-power environments; the introduction of the SE attention mechanism improves the model's ability to focus on key feature channels, reduces background noise interference, and further enhances the model's feature expression; the introduction of the SimAM attention mechanism strengthens the model's spatial feature capture, and improves the model's detection accuracy. That is, the present invention improves the detection performance and real-time performance of the YOLOv5-CSS model, and ensures its efficient application in the seam cutter. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the structure of the YOLOv5-CSS model;

[0040] Figure 2 It is the network structure of the C2fRepGhost module;

[0041] Figure 3 It is the structure of the SE attention mechanism module;

[0042] Figure 4 It is the structure of SimAM attention mechanism module;

[0043] Figure 5 It is the unlabeled images in the dataset during training;

[0044] Figure 6 It is a partially labeled image of the dataset during the training process;

[0045] Figure 7 is the loss function for training and validation;

[0046] Figure 8 is the mAP result;

[0047] Fig. 9 It is the PR curve;

[0048] Fig.10 is the confusion matrix after training the YOLOv5-CSS model;

[0049] Fig.11 This is the detection result of the YOLOv5-CSS model DETAILED DESCRIPTION

[0050] The technical solution of the present invention is further elaborated in detail below in conjunction with specific embodiments. The parts of the present invention that are not described and disclosed in detail in the following embodiments should be understood as the prior art known or should be known to those skilled in the art.

[0051] Example

[0052] A method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model comprises the following steps:

[0053] Acquire road surface images and perform preprocessing;

[0054] The trained YOLOv5-CSS model is used to identify the road image and obtain the recognition result. The YOLOv5-CSS model includes a backbone network with a C2fRepGhost module, a neck structure with an SE attention mechanism module, and a detection head with a SimAM attention mechanism module. The C2fRepGhost module (the C2f module in the figure) is introduced in the backbone network. This module reduces the amount of calculation and memory usage and speeds up the reasoning speed through multi-scale feature fusion and efficient convolution operations; at the same time, it enhances the gradient flow information and eliminates the interference of redundant background features, so that the YOLOv5-CSS model can be smoothly embedded in the seam cutting machine equipment. The SE attention mechanism module is introduced in the neck structure (Backbone in the figure) to improve the YOLOv5-CSS model's ability to focus on key feature channels, reduce background noise interference, and enhance the feature expression of the YOLOv5-CSS model in multi-scale detection tasks. The SinAM attention mechanism module is introduced in front of the detection head to strengthen the spatial feature capture of the YOLOv5-CSS model, further improve the detection accuracy of the YOLOv5-CSS model for complex backgrounds and multi-scale cracks, improve the detection performance and real-time performance of the YOLOv5-CSS model, and ensure its efficient application on resource-constrained seam cutting machines.

[0055] Identification methods include:

[0056] The road surface image is input to the C2fRepGhost module for feature extraction to obtain an intermediate feature map. Part of the intermediate feature map is processed by all RepGhostBottleneck modules and then concatenated with the remaining intermediate feature maps that are not processed by the RepGhostBottleneck module to form an output feature map. In this embodiment, the C2fRepGhost module includes at least one RepGhostBottleneck module, specifically:

[0057] The road surface image is input into the C2fRepGhost module, which generates an intermediate feature map through 1x1 convolution;

[0058] The intermediate feature map is divided into two parts, one of which is directly passed to the Concat module, and the other part of the intermediate feature map is passed to the Concat module after passing through at least one RepGhostBottleneck module to generate an indirect feature map, such as Figure 2As shown in the figure, the intermediate feature map is divided into two parts after passing through the Split module. One part of the intermediate feature map is directly passed to the Concat module, and the other part of the intermediate feature first passes through at least one RepGhostBottleneck module to generate an indirect feature map and then enters the Concat module. The RepGhostBottleneck module uses a combination of deep convolution and point convolution to significantly reduce redundant calculations.

[0059] The Concat module concatenates the direct feature map and the indirect feature map, and obtains the output feature map through convolution. In the present invention, the C2fRepGhost module reduces redundant feature maps, reduces the amount of calculation, improves the detection speed, and ensures that the YOLOv5-CSS model can achieve real-time reasoning on limited computing resources. At the same time, the number of RepGhostBottleneck modules in the C2fRepGhost module can be flexibly adjusted to meet different needs, so that the seam cutter can operate efficiently; at the same time, the multi-scale feature fusion capability of the C2fRepGhost module can enhance the detection performance of the seam cutter under different crack scales and complex backgrounds, and reduce missed detection and false detection. During the operation of the seam cutter, accurate crack detection is crucial for the path planning of the seam cutter. The C2fRepGhost module effectively improves the detection accuracy of the YOLOv5-CSS model and the focus on key areas, ensures the accuracy of the cutting path, and reduces operational errors.

[0060] The SE attention mechanism module is used to dynamically adjust the importance and adaptability of the features of the output feature map to obtain a fused feature map. The SE attention mechanism module uses a channel-level attention mechanism to give the network the ability to dynamically adjust the importance and adaptability of features. Specifically:

[0061] The SE attention mechanism module contains two key operations. The first is the Squeeze operation, which aggregates the output feature map (H×W×C) in the spatial dimension to form a C-dimensional vector through global average pooling. Each element in the C-dimensional vector represents the average value of the corresponding channel as a global channel descriptor.

[0062] Then comes the Excitation operation, which inputs the descriptor into the fully connected layer and generates channel weights through ReLU and Sigmoid activation functions;

[0063] The contribution of each channel is dynamically adjusted according to the channel weight, key features are highlighted, irrelevant information is suppressed, and finally a fused feature map is formed. The SE attention mechanism module makes the grid more focused on effective features in the task by learning channel-level weights, thereby optimizing feature representation and improving recognition accuracy and efficiency.

[0064] The SimAM attention mechanism module is used to analyze the fused feature map, highlight the important feature areas, the detection head identifies the feature areas, and outputs the recognition results. During the working process of the seam cutter, the SimAM attention mechanism module can adaptively adjust the network's attention to different feature areas, so that the YOLOv5-CSS model focuses on the key position of the crack, reduces background interference, and ensures that tiny cracks can be accurately located under complex backgrounds, thereby improving the accuracy and stability of detection. The SimAM attention mechanism module analyzes the fused feature map, evaluates the importance of each channel and spatial position, generates feature weight distribution to reflect the activation degree of each feature area, and fuses these weights with the original feature map. The SimAM attention mechanism module dynamically adjusts the network's attention to different areas, enhances the response of important areas, and suppresses irrelevant areas, thereby optimizing the feature representation, effectively improving the accuracy and robustness of feature extraction, and enabling the YOLOv5-CSS model to better handle small target recognition under complex backgrounds. In this embodiment, the network parameters of the YOLOv5-CSS model are shown in Table 1.

[0065] Table 1 Network parameters of the YOLOv5-CSS model in this embodiment

[0066] The training process of the YOLOv5-CSS model includes the following steps:

[0067] The pavement crack images are collected and preprocessed to obtain the preprocessed data set. The process of collecting pavement images is to use mobile devices to collect pavement crack images under different lighting, environments and locations. The Labelme tool is used to label the collected pavement crack images and convert the labels into XML format to adapt to the YOLOv5-CSS model of the seam cutter. AD shows the data set from different lighting, interference and crack morphology, and E and F are the pavement crack data simulated by ink. The labeled data set is shown in Figure 1. Figure 6 shown.

[0068] The preprocessed dataset is divided into training set, validation set and test set according to 7:1:2;

[0069] The training set is input into the YOLOv5-CSS model for training, and the trained YOLOv5-CSS model is verified using the validation set to obtain a trained YOLOv5-CSS model.

[0070] Input the test set into the trained YOLOv5-CSS model and output the recognition result.

[0071] In the training of the YOLOv5-CSS model of the seam cutter, the loss function needs to be optimized according to the characteristics of crack detection to improve the detection accuracy and robustness of the YOLOv5-CSS model. The loss function is generally composed of classification loss, confidence loss and bounding box regression loss. Different modules are combined according to weights to adapt to the complexity of crack targets. Among them, the confidence loss function is used to calculate the error between the predicted box and the true box, and the bounding box regression loss function is used to calculate the confidence of the YOLOv5-CSS model. In this embodiment, the loss function includes the confidence loss function and the bounding box regression loss function, specifically:

[0072] Loss is:

[0073] Loss = λ1l obj +λ2l box

[0074] Among them, λ1 is 0.5 and λ2 is 1.

[0075] Use BCE to calculate the confidence loss:

[0076]

[0077] Where N is the number of prediction boxes, y i is the true value, Indicates the confidence of the prediction

[0078] In this embodiment, the bounding box regression loss function takes into account distance, overlap area, and aspect ratio:

[0079]

[0080] Among them, l box is the bounding box regression loss, b is the predicted bounding box, is the true bounding box, IoU is the intersection over union ratio, ρ is the Euclidean distance between the center points of the predicted box and the true box, c is the diagonal distance of the smallest external box surrounding the predicted box and the true box, αv is the factor used to penalize the difference in aspect ratio, w / h is the aspect ratio of the target box, is the aspect ratio of the prediction box.

[0081] In this embodiment, PyCharm software is used, and the algorithm is run using the Pytorch network framework for training, stochastic gradient descent (SGD), the momentum coefficient Momentum is set to 0.937, the weight decay Weight decay is set to 0.0005, the initial learning rate is set to 0.01, the cosine annealing adjustment strategy is used, and the CIoU threshold is set to 0.45. The input image resolution is 224*224, the batch size is set to 16, and the number of training iterations is 100, as shown in Table 2.

[0082] Table 2 Training YOLOv5-CSS model parameter settings

[0083] After the YOLOv5-CSS model is trained, the performance of the YOLOv5-CSS model is evaluated by evaluation indicators, including Recall, Precision, and mAP. The Recall evaluation indicator is:

[0084]

[0085]

[0086] Among them, TP is an example that the classifier considers to be a positive sample and is actually a positive sample, FP is an example that the classifier considers to be a positive sample but is actually not a positive sample, and FN is an example that the classifier considers to be a negative sample but is actually not a negative sample. The parameter definitions are shown in Table 3.

[0087] Table 3 Evaluation index parameter definition

[0088] mAP is one of the commonly used evaluation indicators in self-label detection. It is the average of Average Precision (APMAP). AP reflects the accuracy of detection of a single category, and the average of AP of all categories is mAP. The calculation formulas of mAP and AP are as follows:

[0089]

[0090]

[0091] Among them, N represents the total number of detection samples of a certain category, P represents the accuracy (Precision), C represents the number of all categories, and mAP measures the detection effect of all categories.

[0092] In this embodiment, during the training of the YOLOv5-CSS model, the losses and mAP of the training set and the validation set change with iterations as follows: Figure 7 and Figure 8As shown in the figure. From the curve changes, it can be seen that the loss decreases rapidly within the first 50 iterations, while the rate of decrease between 50 and 100 iterations gradually slows down and stabilizes, indicating that the YOLOv5-CSS model is close to convergence. According to the trend of the loss of the training and validation sets, no obvious signs of overfitting are found, indicating that the training effect of the YOLOv5-CSS model is relatively ideal. mAP (average precision) continues to improve with the progress of training, and eventually approaches 0.3, indicating that the YOLOv5-CSS model has good detection ability and stability when handling given tasks. The overall performance of the loss and mAP curves during the training process shows that the YOLOv5-CSS model performs consistently on the training and validation data, and the training process is smooth and converges well.

[0093] The PR curve and confusion matrix of the experimental test are as follows Fig. 9 and Fig.10 As shown in the figure. The PR curve shows that the mean average precision (mAP) of the YOLOv5-CSS model on all categories is about 0.3, indicating that the overall performance of the YOLOv5-CSS model in the detection task is relatively stable, and a good balance between precision and recall is achieved between different categories. The confusion matrix further reveals the classification performance of the YOLOv5-CSS model. The values ​​on the main diagonal represent the proportion of correctly classified samples, among which the detection accuracy of the crack category reaches 0.65, showing that the model has high reliability in complex scenes where cracks and background features are highly similar. At the same time, the misclassification rate of the background category is low, indicating that the YOLOv5-CSS model has a strong ability to distinguish between target cracks and backgrounds, effectively reducing the occurrence of false alarms.

[0094] In this embodiment, the cracks are annotated with multiple frames. This method can flexibly deal with cracks of various shapes. Whether they are short and obvious or slender and tortuous, they can be accurately annotated by adjusting the size and position of the annotation frame, which can more accurately capture the shape, length, width and other characteristics of the cracks. Especially for slender and tortuous cracks, multi-frame annotation can avoid the errors that may be caused by a single annotation frame, thereby improving the accuracy and reliability of annotation. The specific detection results are as follows: Fig.11 As shown in the figure. The improved YOLOv5-CSS model achieves accurate detection of cracks in various forms. Even under complex road backgrounds or noise interference, the YOLOv5-CSS model can still reliably capture key features such as the shape, length, and width of the cracks. This shows that the improved YOLOv5-CSS model has excellent robustness and adaptability in dealing with complex backgrounds and cracks of different forms. By generating a separate detection frame for each part of the crack, the YOLOv5-CSS model effectively avoids the problem of feature loss or misjudgment that may be caused by a single large frame, ensuring the accuracy and integrity of the detection.

[0095] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model, characterized in that: The following steps are involved: Acquire road surface images and perform normalization processing; The trained YOLOv5-CSS model is used to identify the road image and obtain the recognition result. The YOLOv5-CSS model includes a backbone network with a C2fRepGhost module, a neck structure with a SE attention mechanism module, and a detection head with a SimAM attention mechanism module. The recognition method includes: The road surface image is input to the C2fRepGhost module for feature extraction to obtain an intermediate feature map. Some of the intermediate feature maps are processed by all RepGhostBottleneck modules and then concatenated with the remaining intermediate feature maps that are not processed by the RepGhostBottleneck module to form an output feature map. The SE attention mechanism module is used to dynamically adjust the importance and adaptability of the output feature map to obtain a fused feature map; The SimAM attention mechanism module is used to analyze the fused feature map and extract important feature areas; The detection head is used to identify the feature area and output the recognition result.

2. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 1, characterized in that: The C2fRepGhost module includes at least one RepGhostBottleneck module.

3. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 2, characterized in that: The steps to form the output feature map are: The road surface image is input into the C2fRepGhost module, which generates an intermediate feature map through 1x1 convolution; The intermediate feature map is divided into two parts, one of which is directly passed to the Concat module, and the other part of the intermediate feature map is passed to the Concat module after passing through at least one RepGhostBottleneck module to generate an indirect feature map; The Concat module concatenates the intermediate feature map and the indirect feature map to obtain the output feature map.

4. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 1, characterized in that: The method of fusing the output feature map using the SE attention mechanism module includes: The output feature map is mapped and aggregated in the spatial dimension to form a C-dimensional vector through global average pooling. Each element in the C-dimensional vector represents the average value of the corresponding channel, and the C-dimensional vector serves as a global channel descriptor. The descriptor is input into the fully connected layer and the channel weights are generated through the ReLU and Sigmoid activation functions; The contribution of each channel is dynamically adjusted according to the channel weight, and finally a fusion feature map is formed.

5. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 1, characterized in that: The training process of the YOLOv5-CSS model includes the following steps: Collect and preprocess pavement crack images to obtain a preprocessed data set; The preprocessed dataset is divided into training set, validation set and test set according to 7:1:2; The training set is input into the YOLOv5-CSS model for training, and the trained YOLOv5-CSS model is verified using the validation set to obtain a trained YOLOv5-CSS model; Input the test set into the trained YOLOv5-CSS model.

6. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 5, characterized in that: The process of collecting road surface images is to use mobile imaging equipment to collect road surface crack images under different lighting, environments and locations.

7. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 5, characterized in that: The preprocessing method includes using the Labelme tool to label the collected pavement crack images and converting the labels into XML format.

8. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 5, characterized in that: The loss functions of the YOLOv5-CSS model include the confidence loss function and the bounding box regression loss function.

9. The method for detecting pavement cracks using a seam cutter based on a YOLOv5-CSS model as claimed in claim 8, characterized in that: The confidence loss function is: Where N is the number of prediction boxes, y i is the true value, Indicates the confidence of the prediction; The bounding box regression loss function is: Among them, l box is the bounding box regression loss, b is the predicted bounding box, is the true bounding box, IoU is the intersection over union ratio, ρ is the Euclidean distance between the center points of the predicted box and the true box, c is the diagonal distance of the smallest external box surrounding the predicted box and the true box, αv is the factor used to penalize the difference in aspect ratio, w / h is the aspect ratio of the target box, is the aspect ratio of the prediction box.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for detecting pavement cracks by a seam cutter based on a YOLOv5-CSS model as described in any one of claims 1 to 9 are implemented.

Images