A method for detecting road defects in a UAV remote sensing image based on SCDLW-YOLO

Abstract

This invention belongs to the field of road defect detection technology, specifically relating to a method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO. The method includes the following steps: acquiring UAV road remote sensing images, performing data augmentation processing, constructing an expanded road defect dataset and dividing it proportionally into training, validation, and test sets; constructing an SCDLW-YOLO road defect detection model based on the YOLOv8s detection framework; inputting the training set into the SCDLW-YOLO road defect detection model for training; using the validation set to supervise and iteratively optimize the model training process; using the test set to further test and confirm the performance of the converged model; finally, obtaining the optimal SCDLW-YOLO road defect detection model and deploying it on a UAV or embedded device. This invention can effectively improve the detection accuracy and real-time performance of road defects such as cracks, potholes, and repair areas in complex scenarios, while also maintaining a lightweight model, making it suitable for road inspection and intelligent defect recognition using UAVs and edge devices.

Description

Technical Field

[0001] This invention belongs to the field of road defect detection technology, specifically relating to a method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO. Background Technology

[0002] Roads are a crucial component of transportation networks, and their surface condition directly impacts vehicle safety, traffic efficiency, and maintenance effectiveness. With ongoing urbanization and increasing traffic loads, roads are susceptible to surface defects such as cracks, potholes, and unrepaired areas due to repeated vehicle traffic, climate change, and environmental factors. Failure to detect and address these defects promptly not only reduces road performance and shortens its lifespan but also poses traffic safety hazards and increases subsequent maintenance costs. Therefore, efficient and accurate detection of road defects has significant practical and engineering application value.

[0003] Existing road defect detection methods mainly include manual inspection and detection methods based on traditional digital image processing. While manual inspection is more direct, it suffers from low efficiency, high labor intensity, high labor costs, and significant susceptibility to subjective factors, making it difficult to meet the application requirements of rapid, large-scale road inspection. Traditional digital image processing methods, such as threshold segmentation, edge extraction, and region growing, can identify road defects in some simple scenarios, but these methods typically rely on manually designed features, are sensitive to changes in lighting, shadows, and complex background textures, and have poor robustness, making them unsuitable for the automated detection of various types of road defects in complex environments. With the development of UAV remote sensing and deep learning technologies, using UAVs to acquire high-resolution road images and combining them with target detection algorithms for road defect identification has become an important research direction in the field of road inspection.

[0004] However, existing deep learning-based road defect detection methods still have certain shortcomings in UAV remote sensing image applications. On the one hand, road images acquired by UAVs often have complex backgrounds, significant lighting variations, numerous shadows, and many interfering targets such as vehicles, pedestrians, and trees, which can easily lead to false positives and false negatives during detection. On the other hand, road defects vary greatly in shape, scale, and texture. Cracks are typically elongated and have unclear boundaries, while potholes often have indistinct local textures and irregular shapes, increasing the difficulty of feature extraction and target localization. Furthermore, UAVs and edge devices place high demands on the parameter scale, computational complexity, and real-time performance of the detection model. Existing technologies often struggle to simultaneously achieve detection accuracy, lightweight models, and cross-scenario generalization capabilities, thus failing to meet the needs of practical road inspection applications. Therefore, there is an urgent need to provide a road defect detection method suitable for UAV remote sensing image scenarios that balances lightweight design with high accuracy, addressing the aforementioned problems in existing technologies.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a road defect detection method based on SCDLW-YOLO for UAV remote sensing images, so as to overcome the shortcomings of existing road defect detection technologies, such as low detection accuracy, high false negative rate, high model complexity, and difficulty in balancing lightweight and high accuracy. This invention improves the detection accuracy, positioning accuracy, and real-time detection capability of road defects in complex scenarios, while reducing the number of model parameters and computational complexity, and enhancing its applicability in UAVs and edge devices.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO includes the following steps: S1. Acquire UAV road remote sensing images as raw data, perform data augmentation processing, and construct an expanded road defect dataset; S2. Divide the expanded road defect dataset into three independent training, validation, and test sets in a 7:2:1 ratio. S3. Based on the YOLOv8s detection framework, construct the SCDLW-YOLO road defect detection model; S4. Input the divided training set into the SCDLW-YOLO road defect detection model for training, and use the validation set to supervise and iteratively optimize the model training process. Use the test set to further test and confirm the performance of the converged model, and finally obtain the SCDLW-YOLO road defect detection model with the best performance. S5. Deploy the best-performing SCDLW-YOLO road defect detection model on drones or embedded devices to achieve real-time road defect detection.

[0008] Preferably, S1 performs data augmentation processing on the road remote sensing image, specifically including: brightness transformation, random rotation, horizontal flipping, and vertical flipping.

[0009] Preferably, the expanded road defect dataset constructed in S1 includes: longitudinal cracks, transverse cracks, network cracks, diagonal cracks, repair areas, and potholes.

[0010] Preferably, the SCDLW-YOLO road defect detection model constructed in S3 includes: a backbone network, a neck network, a detection head, and a loss function optimization module; the backbone network integrates the StarNet structure and CAA attention mechanism into the C2f module to construct the C2f_Star_CAA module; the neck network uses the DySample dynamic upsampling operator to replace the traditional UpSample upsampling operator; the detection head introduces the LSCD lightweight shared convolutional detection head, which performs joint processing of feature maps at different scales through shared convolutional layers; the loss function optimization module uses the WIoUv2 loss function to replace the bounding box regression loss function in the basic detection framework.

[0011] Preferably, the C2f_Star_CAA module uses depthwise separable convolution, pointwise convolution, and attention weighting mechanisms to jointly extract and enhance local texture information and long-distance contextual information in road defect areas.

[0012] Preferably, the DySample dynamic upsampling operator performs an adaptive upsampling operation on the input feature map by dynamically generating a sampling offset. The specific calculation steps are as follows: Given input feature map And a sampling set S of size 2g×sH×sW; where s represents the upsampling scaling factor and g represents the number of coordinates corresponding to the sampling grid points; The input feature map is resampled using the positions in the sampling set S to obtain the output feature map χ. Its expression is: ; Among them, grid_Sample() is the grid sampling function; For the input feature map An offset O is generated through a linear layer and superimposed on the original sampling grid G ​​to obtain a dynamic sampling set S, the expression of which is: ; To avoid overlapping local sampling positions, a static range factor is introduced to constrain the offset, and its expression is as follows: ; A point-by-point dynamic range factor is generated by linearly projecting input features, and the offset is dynamically adjusted using the Sigmoid function. Its expression is further given as: .

[0013] Preferably, the WIoUv2 loss function introduces a dynamic non-monotonic focusing mechanism during training, adaptively allocating gradient weights for different samples based on sample quality. The specific calculation steps are as follows:

[0014]

[0015]

[0016] ; in, This represents the dynamically adjusted IoU loss value. The exponential moving average representing IoU loss. For hyperparameters, For dynamic weighting factors; and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. and This represents the width and height of the smallest bounding rectangle.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The road defect detection method based on SCDLW-YOLO for UAV remote sensing images of the present invention enhances the model’s ability to extract detailed features of road surface and global context information in complex backgrounds by constructing the C2f_Star_CAA module in the backbone network, which is conducive to improving the recognition accuracy of road defects such as cracks, potholes and repair areas.

[0018] (2) The road defect detection method based on SCDLW-YOLO for UAV remote sensing images of the present invention improves the multi-scale feature fusion effect by using the DySample dynamic upsampling operator in the neck network, enhances the model's ability to detect small target road defects, and helps to reduce the false negative rate.

[0019] (3) The road defect detection method based on SCDLW-YOLO for UAV remote sensing images of the present invention, by introducing the LSCD lightweight shared convolutional detection head, realizes the joint processing and parameter sharing of feature maps of different scales. It can reduce the number of model parameters and computational complexity while ensuring detection accuracy, and is more suitable for deployment in UAVs and edge devices.

[0020] (4) The road defect detection method based on SCDLW-YOLO for UAV remote sensing images of the present invention improves the model’s localization accuracy for road defect targets with blurred boundaries and irregular shapes by using the WIoUv2 loss function to optimize the bounding box regression process, and enhances the detection stability in complex scenarios.

[0021] (5) The road defect detection method based on SCDLW-YOLO of the present invention improves the road defect detection accuracy, positioning accuracy and real-time detection capability while taking into account the lightweight model, and can meet the actual application needs of road inspection and intelligent defect identification of UAV remote sensing images. Attached Figure Description

[0022] Figure 1 This is a flowchart of a road defect detection method based on SCDLW-YOLO for UAV remote sensing images according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall network structure of the SCDLW-YOLO road defect detection model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the C2f_Star_CAA module structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the DySample dynamic upsampling module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the LSCD lightweight shared convolutional detection head structure according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of this invention patent will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0024] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention.

[0025] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to a connection within two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] See attached document Figure 1 This invention provides a road defect detection method based on UAV remote sensing images using SCDLW-YOLO, suitable for intelligent detection of road surface defects in complex backgrounds. The method includes the following steps: Step 1: Select UAV road remote sensing images from the publicly available road defect dataset as the raw data to construct a road defect dataset; perform data augmentation processing on the road remote sensing images, including brightness transformation, random rotation, horizontal flipping, and vertical flipping, to obtain the expanded road defect dataset; the road defect dataset includes various defect types such as longitudinal cracks, transverse cracks, mesh cracks, diagonal cracks, repair areas, and potholes; Step 2: Divide the expanded road defect dataset into three independent sets—training, validation, and test—in a 7:2:1 ratio for subsequent model training, validation, and detection. Step 3: Based on the YOLOv8s detection framework, construct the SCDLW-YOLO road defect detection model; Step 4: Input the road remote sensing images from the training set into the SCDLW-YOLO road defect detection model for training, and use the validation set to supervise and iteratively optimize the model training process. Use the test set to further test and confirm the performance of the converged model, and finally obtain the SCDLW-YOLO road defect detection model with the best performance. Step 5: Deploy the best-performing SCDLW-YOLO road defect detection model on a drone or embedded device, and then input the remote sensing image of the road to be detected into the deployed model in real time. The model will quickly output the category, location and confidence information of the road defect target, realize real-time road defect detection, and complete the entire process of road defect detection of drone remote sensing images.

[0027] In this embodiment, the SCDLW-YOLO road defect detection model can be used to detect various types of road defects, including longitudinal cracks, transverse cracks, network cracks, diagonal cracks, repair areas, and potholes. It is also applicable to UAV remote sensing image scenarios with complex backgrounds, significant lighting variations, and occlusion interference. (See attached image.) Figure 2 As shown, the SCDLW-YOLO road defect detection model includes a backbone network, a neck network, a detection head, and a loss function optimization module. The backbone network is used to extract deep features from the input image, the neck network is used to achieve multi-scale feature fusion, the detection head is used to output the category and location information of the road defect target, and the loss function optimization module is used to optimize the bounding box regression performance during the model training process.

[0028] In the backbone network, the StarNet structure and CAA attention mechanism are integrated into the C2f module to construct the C2f_Star_CAA module, thereby enhancing the model's ability to extract road defect features in complex scenarios and model global contextual information. (See attached image) Figure 3 As shown, the input features are first mapped through convolutional operations, and then enter the Star-CAABlock module through a branching structure. In the Star-CAA Block module, depthwise separable convolution and pointwise convolution are used to achieve feature extraction and channel expansion, and feature interaction is enhanced by element-wise multiplication. Furthermore, the CAA attention mechanism is combined to model local features and long-distance contextual information, generate attention weights and enhance features, thereby improving the model's ability to represent road defects such as thin cracks, irregular potholes and complex background interference.

[0029] In the neck network, the DySample dynamic upsampling operator is used instead of the traditional UpSample upsampling operator to improve multi-scale feature fusion capabilities and the ability to detect small-target road defects. (See attached image) Figure 4 As shown, the DySample dynamic upsampling operator dynamically generates sampling offsets and performs adaptive upsampling on the input feature map. This allows the model to dynamically adjust the sampling position based on the semantic information of the input feature map, thereby more effectively preserving the detailed features of road defect areas and improving the detection capability of small-scale cracks and potholes. The specific calculation process is as follows: Given input feature map And a sampling set S of size 2g×sH×sW; where s represents the upsampling scaling factor and g represents the number of coordinates corresponding to the sampling grid points; The input feature map is resampled using the positions in the sampling set S to obtain the output feature map χ. Its expression is: ; Among them, grid_Sample() is the grid sampling function; For the input feature map An offset O is generated through a linear layer and superimposed on the original sampling grid G ​​to obtain a dynamic sampling set S, the expression of which is: ; To avoid overlapping local sampling positions, a static range factor is introduced to constrain the offset, and its expression is as follows: ; A point-by-point dynamic range factor is generated by linearly projecting input features, and the offset is dynamically adjusted using the Sigmoid function. Its expression is further given as: ; In this way, the DySample dynamic upsampling operator can adaptively generate sampling positions based on the content of the input feature map, thereby improving the feature representation ability of small target road defect areas in complex backgrounds.

[0030] In the detection head, a lightweight shared convolutional detection head (LSCD) is introduced. This head uses shared convolutional layers to jointly process feature maps at different scales, thereby reducing the number of model parameters, memory usage, and computational complexity. For example... Figure 5 As shown, the LSCD lightweight shared convolutional detection head uses shared convolutional layers to jointly process feature maps at different scales and improves the stability of classification and localization tasks through a normalization mechanism. Unlike traditional detection heads that process feature maps at different scales independently, the LSCD lightweight shared convolutional detection head can achieve cross-scale feature interaction and parameter sharing, thereby reducing the number of model parameters, reducing memory usage, and improving the model's deployment capability on resource-constrained devices.

[0031] In the loss function optimization module, the WIoUv2 loss function is used to replace the original bounding box regression loss function in YOLOv8s to optimize the bounding box regression process. The WIoUv2 loss function introduces a dynamic non-monotonic focusing mechanism during training, which adaptively allocates gradient weights to different samples based on sample quality. This reduces the interference of low-quality samples on parameter updates and improves the model's localization accuracy for targets with blurred boundaries, large scale variations, and irregular road defects. The specific calculation process of the WIoUv2 loss function includes:

[0032]

[0033]

[0034] ; in, This represents the dynamically adjusted IoU loss value. The exponential moving average representing IoU loss. For hyperparameters, For dynamic weighting factors; and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. and This represents the width and height of the smallest bounding rectangle.

[0035] The road defect detection method based on SCDLW-YOLO for UAV remote sensing images of the present invention can achieve high-precision detection of road defects in UAV remote sensing images in complex scenarios, while taking into account the requirements of lightweight model and practical deployment. On the basis of ensuring lightweight model, it can improve the accuracy of road defect detection, positioning accuracy and real-time detection performance, and is suitable for application in UAVs, edge devices and intelligent road inspection systems.

[0036] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO, characterized in that, Includes the following steps: S1. Acquire UAV road remote sensing images as raw data, perform data augmentation processing, and construct an expanded road defect dataset; S2. Divide the expanded road defect dataset into three independent sets—training, validation, and test—in a 7:2:1 ratio. S3. Based on the YOLOv8s detection framework, construct the SCDLW-YOLO road defect detection model; S4. Input the divided training set into the SCDLW-YOLO road defect detection model for training, and use the validation set to supervise and iteratively optimize the model training process. Use the test set to further test and confirm the performance of the converged model, and finally obtain the SCDLW-YOLO road defect detection model with the best performance. S5. Deploy the best-performing SCDLW-YOLO road defect detection model on drones or embedded devices to achieve real-time road defect detection.

2. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 1, characterized in that, S1 performs data augmentation processing on road remote sensing images, specifically including brightness transformation, random rotation, horizontal flipping, and vertical flipping.

3. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 1, characterized in that, The expanded road defect dataset constructed in S1 includes: longitudinal cracks, transverse cracks, network cracks, diagonal cracks, repair areas, and potholes.

4. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 1, characterized in that, The SCDLW-YOLO road defect detection model constructed in S3 includes: a backbone network, a neck network, a detection head, and a loss function optimization module. The backbone network integrates the StarNet structure and CAA attention mechanism into the C2f module to construct the C2f_Star_CAA module. The neck network uses the DySample dynamic upsampling operator to replace the traditional UpSample upsampling operator. The detection head introduces the LSCD lightweight shared convolutional detection head, which performs joint processing of feature maps at different scales through shared convolutional layers. The loss function optimization module uses the WIoUv2 loss function to replace the bounding box regression loss function in the basic detection framework.

5. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 4, characterized in that, The C2f_Star_CAA module uses depthwise separable convolution, pointwise convolution, and attention weighting mechanisms to jointly extract and enhance local texture information and long-range contextual information in road defect areas.

6. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 4, characterized in that, The DySample dynamic upsampling operator performs adaptive upsampling on the input feature map by dynamically generating a sampling offset. The specific calculation steps are as follows: Given an input feature map And a sampling set S of size 2g×sH×sW; where s represents the upsampling scaling factor and g represents the number of coordinates corresponding to the sampling grid points; The input feature map is resampled using the positions in the sampling set S to obtain the output feature map χ. Its expression is: ； Among them, grid_Sample() is the grid sampling function; For the input feature map An offset O is generated through a linear layer and superimposed on the original sampling grid G ​​to obtain a dynamic sampling set S, the expression of which is: ； To avoid overlapping local sampling positions, a static range factor is introduced to constrain the offset, and its expression is as follows: ； A point-by-point dynamic range factor is generated by linearly projecting input features, and the offset is dynamically adjusted using the Sigmoid function. Its expression is further given as: 。 7. The method for detecting road defects in UAV remote sensing images based on SCDLW-YOLO according to claim 4, characterized in that, The WIoUv2 loss function introduces a dynamic non-monotonic focusing mechanism during training, adaptively allocating gradient weights for different samples based on sample quality. The specific calculation steps are as follows: ； ； ； ； in, This represents the dynamically adjusted IoU loss value. The exponential moving average representing IoU loss. For hyperparameters, For dynamic weighting factors; and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. and This represents the width and height of the smallest bounding rectangle.

Images