Road disease detection method based on unmanned aerial vehicle, electronic equipment and program product

Abstract

The invention discloses a road disease detection method based on an unmanned aerial vehicle, electronic equipment and a program product. The method is realized based on a trained road disease detection model, a C3k2-MDDSC module is introduced into a backbone network of the model, the feature multiplexing capability is enhanced through gradient shunting and multi-scale fusion, the gradient disappearance problem is relieved, and the robustness of the model is improved by means of jump connection and packet convolution. An ACFP module is introduced into the tail end of the backbone network, dynamic fusion of local and global features is realized by using multi-scale cavity convolution and a channel-space attention mechanism, and the complex scene modeling capability is improved. And the neck network is integrated with an SGF module, so that the spatial perception of the model to a tiny target can be improved. Besides, the ES-FPN proposed based on the SGF module not only can enhance the utilization of shallow spatial information, but also can optimize the complementarity of cross-level features. During training, regression loss, namely fast high-quality intersection-to-union ratio loss, is proposed, and angle punishment is introduced to improve the alignment precision of the rotating frame and the convergence speed of the model.

Description

Technical Field

[0001] The present application belongs to the field of image processing technology, and in particular relates to a road damage detection method based on drones, an electronic device, and a computer program product. Background Art

[0002] Road defects such as cracks and potholes, if not discovered in time, can affect driving safety and accelerate road aging. Traditional manual inspections are inefficient and have limited coverage, while on-board inspection systems are costly and inflexible.

[0003] In recent years, drones, with their flexibility, efficiency, and low cost, have been increasingly used for road damage detection, particularly in remote or dangerous sections of road. Combined with deep learning algorithms, they can automatically identify defects, improving detection efficiency and accuracy while reducing maintenance costs.

[0004] However, in practical applications, small-scale defects occupy a small proportion of the imagery captured by drones and have weak features, making them easily missed and affecting recognition. Therefore, improving the detection capabilities of small defects in drone images has become a key issue in the development of this technology. Summary of the Invention The present application provides a road disease detection method, electronic equipment and computer program product based on drones, which can effectively alleviate the recognition difficulties caused by the low proportion of disease target pixels and unclear texture features, thereby improving the detection accuracy and stability of small-scale diseases.

[0005] In a first aspect, the present application provides a method for detecting road damage based on a drone, comprising: Based on the backbone network of the pre-trained road disease detection model, feature extraction is performed on the image to be detected collected from the perspective of the drone to obtain image features; the image to be detected includes roads; The neck network based on the road damage detection model fuses image features to obtain target fusion features; The detection network based on the road damage detection model detects the target fusion features and obtains the detection results of road damage in the image to be detected; The backbone network includes feature extraction modules of m scales, where m is a positive integer and m≥2. Each feature extraction module is equipped with a C3k2-MDDSC submodule. The C3k2-MDDSC submodule includes the first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer. For the first input feature of the C3k2-MDDSC submodule: Performing a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature; Splitting the first compressed feature into a first segmentation feature and a second segmentation feature through a segmentation layer; Performing a multi-level progressive extraction operation on the first segmentation feature through a multi-level progressive stacking extraction layer to obtain hierarchical features corresponding to each level; The first segmentation feature, the second segmentation feature, and the features of each level are spliced ​​together through the first splicing layer to obtain a first splicing feature; Perform a channel expansion operation on the first concatenated feature through the second CBS layer to obtain a first output feature corresponding to the first input feature; For each feature extraction module that extracts shallow features, when the corresponding C3k value in C3k2-MDDSC is False, the multi-level progressive stacking extraction layer is improved based on MDDSC; for each feature extraction module that extracts deep features, when the corresponding C3k value in C3k2-MDDSC is True, the multi-level progressive stacking extraction layer is improved based on C3k-MDDSC.

[0006] In a second aspect, the present application provides a road disease detection device, comprising: An extraction module is used to extract features from the image to be detected collected from the perspective of the drone based on the backbone network of the pre-trained road disease detection model to obtain image features; the image to be detected includes roads; A fusion module is used to fuse image features based on the neck network of the road damage detection model to obtain target fusion features; A detection module is used to detect target fusion features based on the detection network of the road damage detection model to obtain the detection results of road damage in the image to be detected; The backbone network includes feature extraction modules of m scales, where m is a positive integer and m≥2. Each feature extraction module is equipped with a C3k2-MDDSC submodule. The C3k2-MDDSC submodule includes the first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer. For the first input feature of the C3k2-MDDSC submodule: Performing a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature; Splitting the first compressed feature into a first segmentation feature and a second segmentation feature through a segmentation layer; Performing a multi-level progressive extraction operation on the first segmentation feature through a multi-level progressive stacking extraction layer to obtain hierarchical features corresponding to each level; The first segmentation feature, the second segmentation feature, and the features of each level are spliced ​​together through the first splicing layer to obtain a first splicing feature; Perform a channel expansion operation on the first concatenated feature through the second CBS layer to obtain a first output feature corresponding to the first input feature; For each feature extraction module that extracts shallow features, when the corresponding C3k value in C3k2-MDDSC is False, the multi-level progressive stacking extraction layer is improved based on MDDSC; for each feature extraction module that extracts deep features, when the corresponding C3k value in C3k2-MDDSC is True, the multi-level progressive stacking extraction layer is improved based on C3k-MDDSC.

[0007] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of the first aspect when executing the computer program.

[0008] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method of the first aspect are implemented.

[0009] In a fifth aspect, the present application provides a computer program product, which includes a computer program. When the computer program is executed by one or more processors, it implements the steps of the method of the first aspect.

[0010] Compared with the prior art, the present application has the following beneficial effects: by introducing C3k2-MDDSC submodules into the feature extraction modules of each scale of the backbone network, adaptive optimization of features at different levels can be achieved: the shallow feature extraction module adopts the C3k2-MDDSC submodule corresponding to C3k=False, and the multi-level progressive stacking extraction layer in this submodule is obtained based on MDDSC. This structure not only enhances the multi-scale progressive extraction capability of local details and edge textures, but also enables small-scale defects to be fully characterized in the early stages despite their small proportion and weak features in the image; the deep feature extraction module adopts the C3k2-MDDSC submodule corresponding to C3k=True, and the multi-level progressive stacking extraction layer in this submodule is obtained based on the C3k-MDDSC structure. This structure retains shallow spatial detail information while maintaining global semantic modeling capabilities, thereby avoiding the loss of small target information by deep features. The multi-level progressive stacking and splicing operations based on gradient diversion enhance the complementarity of features at each level, thereby increasing the weight of small-scale defects in multi-scale feature fusion and significantly reducing the probability of missed detection. From the perspective of a drone, this design can effectively alleviate the recognition difficulties caused by the low proportion of disease target pixels and unclear texture features, thereby improving the detection accuracy and stability of small-scale diseases.

[0011] It can be understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0013] Figure 1 Schematic diagram of the network structure of the road damage detection model provided in the embodiment of the present application; Figure 2 This is an example graph in the UAV-SMRDD dataset provided by the embodiment of the present application; Figure 3 Schematic diagram of the process of the drone-based road damage detection method provided in the embodiment of the present application; Figure 4 2 is a schematic diagram of the network structure of the C3k2-MDDSC submodule when the value of C3k is False provided in an embodiment of the present application; Figure 5 2 is a schematic diagram of the network structure of the C3k2-MDDSC submodule when the C3k value is True provided in an embodiment of the present application; Figure 6 Schematic diagram of the network structure of C3k-MDDSC provided in an embodiment of the present application; Figure 7 Schematic diagram of the network structure of the MDDSC provided in the embodiment of the present application; Figure 8 Schematic diagram of the network structure of the ACFP module provided in an embodiment of the present application; Figure 9 Schematic diagram of the network structure of the SGF submodule provided in the embodiment of the present application; Figure 10 is a schematic diagram of the ES-FPN structure provided in an embodiment of the present application; Figure 11 Schematic diagram of the structure of the road disease detection device provided in an embodiment of the present application; Figure 12 It is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0015] Drones, due to their flexibility, high efficiency, and low cost, are widely used for road damage detection, particularly in remote or dangerous sections. Combined with deep learning algorithms, they can automatically identify defects, significantly improving detection efficiency and accuracy while reducing maintenance costs. However, in practice, small defects in drone-viewed images have a low pixel count and weak feature representation, making them prone to missed detection and compromising overall recognition effectiveness. Therefore, improving the ability to detect small defects in drone imagery has become a critical issue in this field, requiring urgent breakthroughs.

[0016] This study found that traditional detection methods struggle to fully exploit subtle features of targets for two main reasons: First, continuous downsampling in deep networks gradually weakens the detailed information of small targets, making them difficult to effectively represent in deep feature maps, thus affecting positioning accuracy; second, while shallow features retain rich spatial details, they lack semantic expression and are susceptible to noise interference in complex backgrounds, further increasing the difficulty of extracting small target features. Therefore, how to balance detail preservation and semantic expression during feature extraction and achieve an efficient fusion and balance between the two has become a key challenge in improving small target detection performance.

[0017] Based on this concept, this application proposes a road damage detection model, which includes a backbone network, a neck network, and a detection network. Specifically, the backbone network is used to extract features from images of roads to be detected, collected by drones, to obtain image features; the neck network is used to fuse these image features to obtain target fusion features; and the detection network is used to detect the target fusion features to obtain road damage detection results in the processed images.

[0018] To balance detail preservation and semantic expression, achieving an efficient fusion and balance between the two, the C3k2-MDDSC module was introduced into the backbone network of the road defect detection model. This module improves upon the C3k2 architecture by incorporating multi-scale dilated depthwise separable convolution (MDDSC). This module leverages C3k2's flexibility and efficiency in extracting both deep and shallow features, enabling the model to maintain its lightweight architecture while enhancing the expressive power of multi-scale features.

[0019] Specifically, the backbone network contains feature extraction modules at m scales (m ≥ 2), each equipped with a C3k2-MDDSC submodule. This submodule consists of a first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first concatenation layer, and a second CBS layer. The segmentation layer (split) divides the first compressed feature into two segmentation features, denoted as the first segmentation feature and the second segmentation feature. The first segmentation feature, in addition to being progressively extracted at multiple levels to capture details and semantic information under different receptive fields, can also be directly passed to the subsequent concatenation layer; the second segmentation feature retains the original information. The two segmentation features and features from each level are then fused in the first concatenation layer, and the number of channels is restored through the second CBS layer to achieve a balanced representation of details and semantics. This structure can retain more spatial details in the shallow feature extraction module and enhance global semantic perception in the deep feature extraction module, thereby significantly improving the detection accuracy and robustness of small-scale diseases from the drone's perspective.

[0020] The C3k2 module's multi-level progressive stacking extraction layers are dynamically adjusted based on the feature pyramid level (P). At shallow layers (e.g., P3 / 8), C3k is set to False, and the multi-level progressive stacking extraction layers are based on an improved MDDSC. At deeper layers (e.g., P5 / 32), C3k is set to True, and the multi-level progressive stacking extraction layers are based on an improved C3k-MDDSC.

[0021] For example, assume a multi-level progressive stacking extraction layer consists of n levels, consisting of n MDDSC / C3k-MDDSCs connected in series. Except for the first level, which directly takes the first segmentation feature as input, the remaining levels all take the output of the previous level as input. Each level not only passes its output to the next level but also feeds it into the first concatenation layer for fusion. This design effectively alleviates the vanishing gradient problem of deep networks through multi-path gradient propagation and improves feature reuse through a multi-branch structure, enabling the network to learn complex patterns more efficiently.

[0022] In some embodiments, when C3k takes the value of True, in order to enhance the expressiveness of deep features in multi-level progressive stacking extraction layers, the structure of C3k-MDDSC includes a third CBS layer, a series of dual MDDSCs, a second splicing layer, and a fourth CBS layer to achieve full extraction and fusion of deep features.

[0023] By designing a continuous stacking of two MDDSCs within the multi-level progressive stacking extraction layer of C3k-MDDSC, not only is the staged multi-scale extraction and refinement of deep features achieved, but the gradient propagation path is also enriched with the cooperation of the multi-branch structure, alleviating the gradient vanishing problem of deep networks. The first MDDSC focuses on expanding the receptive field to capture global semantics and long-range dependencies, while the second MDDSC builds on this to enhance the ability to depict local spatial details. The combination of the two makes the features more hierarchical and discriminative in multi-scale space. This design not only improves the model's ability to recognize small-scale defects in complex backgrounds, but also enhances the robustness and generalization performance of feature expression, thereby effectively improving the accuracy and stability of road defect detection from the perspective of drones.

[0024] In some embodiments, the MDDSC includes a first convolutional layer, a multi-scale dilated depthwise separable convolutional layer, a third splicing layer, a second convolutional layer, and a fourth splicing layer. Its core idea is to capture rich contextual diversity at a very low computational cost. Therefore, when C3k is False, there is no need to disable residual links. The multi-scale dilated depthwise separable convolutional layer uses parallel depthwise separable convolutions with multiple different dilation rates (for example, 5 scales with dilation rates of 1, 2, 3, 4, and 5, respectively) to effectively extract multi-scale contextual information and significantly enhance the model's adaptability to complex scenarios.

[0025] The small-void ratio branch focuses on local detail features (such as edges and textures), while the large-void ratio branch captures wide-area semantic information (such as object structure and scene layout). This approach achieves multi-scale fusion while maintaining feature map resolution, avoiding the information loss associated with traditional downsampling. The depthwise separable convolutional architecture effectively reduces computational complexity and enables efficient collaborative representation of local details and global semantics.

[0026] In drone road defect inspection missions, due to the complex background interference on the road surface (such as asphalt texture, markings, shadows, and lighting changes), traditional single-scale feature extraction methods often find it difficult to effectively distinguish real defects from background noise.

[0027] In some embodiments, to effectively distinguish between defects and noise in complex backgrounds, this application designs an Adaptive Contextual Feature Pyramid (ACFP) module, placed at the end of the backbone network. This module includes a third convolutional layer, a multi-scale dilated deep convolutional structure, a fifth concatenation layer, a fourth convolutional layer, a channel-spatial attention mechanism, and a residual fusion layer, forming an adaptive multi-scale feature fusion framework.

[0028] The multi-scale atrous deep convolutional architecture utilizes multi-scale atrous convolutions (with dilation rates of 2, 4, and 6) to extract local details and global contextual features, respectively, forming a hierarchical feature representation. Subsequently, a channel-spatial attention mechanism and a residual fusion layer dynamically calibrate the weights of features at each scale, suppressing noise and enhancing the response of key features. The ACFP module dynamically captures contextual information of varying granularity through a multi-scale receptive field, enhancing defect identification capabilities. Adaptive feature weighting adjusts fusion weights based on local characteristics, and cross-scale feature interaction enables the coordinated optimization of local details and global semantics. This adaptive multi-scale fusion mechanism effectively overcomes the local limitations of traditional convolution, significantly improving the model's ability to model complex scenes and robustness in detecting small-scale defects, while maintaining computational efficiency. This provides a solid technical foundation for intelligent road defect identification in complex environments.

[0029] In some embodiments, the channel-spatial attention structure includes a channel attention branch, a spatial attention branch, and an enhanced fusion layer. The channel attention branch consists of a global average pooling layer, a global maximum pooling layer, a sixth concatenation layer, a fifth convolution layer, a nonlinear transformation layer, a sixth convolution layer, and a first activation layer; the spatial attention branch includes a mean layer, a max layer, a seventh concatenation layer, a seventh convolution layer, and a second activation layer.

[0030] The channel-spatial attention structure achieves precise weighting of multi-scale features by fusing a dual attention mechanism: channel and spatial. Specifically, the channel attention mechanism adaptively strengthens the expression of disease-related feature channels, while the spatial attention mechanism focuses on the spatial distribution of disease targets, effectively suppressing interference responses in areas of background noise. This dual-attention-guided fusion mechanism significantly improves the network's ability to identify true disease features in complex backgrounds, particularly in environments with strong interference such as asphalt textures, road markings, and shadows. It enhances the detection of small-target diseases and provides an effective technical approach to solving the problem of detecting small-scale diseases in complex background interference.

[0031] It is worth noting that the channel attention branch statistically aggregates channel features through global average pooling and global maximum pooling layers, extracting global channel information from the perspectives of average response and extreme value response, respectively, and then enhancing the expressive power of disease-related channels through cross-channel fusion operations. The spatial attention branch averages and maximum pools the feature maps in the channel dimension through the Mean layer and Max layer, generating two single-channel spatial description maps to capture important information about spatial location, thereby highlighting the spatial distribution characteristics of disease targets and suppressing interference responses in background noise areas.

[0032] In drone aerial photography, due to the high altitude, road defects typically occupy only a few pixels in the image, resulting in small objects and blurred details. This makes it difficult to effectively extract and identify key visual features. Furthermore, existing detection methods often rely on high-level semantic features when fusing multi-scale features, while ignoring the rich spatial details found in shallower features. This fusion mechanism results in insufficient alignment and complementarity between features at different levels, leading to insufficient cross-scale information fusion. In particular, in complex backgrounds and low-resolution conditions, the model's spatial perception capabilities are limited, severely impacting the localization and recognition of small objects.

[0033] In some embodiments, the aforementioned issues with conventional fusion mechanisms arise from a lack of spatially guided fusion. Based on this research, this embodiment employs m feature fusion modules at corresponding scales in the neck network for each of the m feature extraction modules in the backbone network. Each fusion module is equipped with a spatially guided fusion (SGF) submodule, comprising a weighted fusion layer, a mean-max fusion layer, an eighth convolutional layer, a third activation layer, and an enhancement layer.

[0034] The SGF submodule dynamically adjusts the weights of features at different scales through a weighted fusion layer. Combined with the diverse information aggregation of the Mean-Max fusion layer in the spatial dimension, this effectively enhances the expressive richness and spatial consistency of features. Information is further refined through convolution and activation layers, while the enhancement layer strengthens the discriminative power of the fused features. The overall design fully leverages the spatial guidance mechanism to achieve coordinated optimization of cross-scale details and semantics, significantly improving the model's ability to perceive and identify small-scale diseases.

[0035] In some embodiments, in order to further solve the problems of small road disease targets, insufficient feature expression, and insufficient multi-scale feature fusion from the perspective of drones, the SGF submodule includes SGF-2 and SGF-3 submodules with the same network structure but different settings; The SGF-2 submodule is set in the feature fusion module of each scale; for the i-th scale, the SGF-2 submodule is used to extract the target features corresponding to the i-th scale and the alignment features corresponding to the i+1-th scale; i∈[1,m].

[0036] The feature extraction modules of the 2nd to m-1th scales are all equipped with SGF-3 submodules; for the jth scale, the SGF-3 submodule is used to fuse the target extraction features corresponding to the jth scale, the initial fusion features corresponding to the jth scale, and the final fusion features corresponding to the j-1th scale; the initial fusion features are obtained based on the feature fusion output by the SGF-2 submodule of the jth scale; the final fusion features are obtained based on the fusion of the outputs of the SGF-2 submodule of the j-1th scale or the SGF-3 submodule of the j-1th scale; j∈[2,m-1].

[0037] This architectural design introduces the SGF submodule within the shallow feature pyramid levels P2–P4 (corresponding to feature maps with higher resolution and rich spatial detail). This enhances the model's spatial perception of small objects by performing finer spatial alignment and information fusion of multi-scale features at an early stage. Compared to traditional FPN, which relies more heavily on the fusion of deep semantic features from the P3–P5 levels (lower resolution but higher semantic abstraction), the ES-FPN of this embodiment prioritizes the utilization of detailed information within shallow features and leverages spatial information to enhance cross-level feature complementarity, effectively alleviating recognition difficulties caused by small object size and image resolution limitations.

[0038] In some embodiments, given the many advantages of the YOLO series models, the road damage detection model can be improved based on the YOLO series models. In the road damage detection task, the lightweight YOLOv11s is selected as the basic model, which significantly improves the detection accuracy and small target recognition capabilities while ensuring real-time performance. Therefore, when constructing a road damage detection model, the lightweight YOLOv11s can be preferred as the basic network. On this basis, the C3k2-MDDSC module, ACFP module and SGF submodule can be combined, or the SGF submodule can be embedded in the ES-FPN structure to make improvements in the form of a single or multi-module combination, thereby further improving the model's multi-scale feature modeling capabilities and detection accuracy in complex backgrounds while maintaining real-time performance.

[0039] For example, the network structure of the road damage detection model including the C3k2-MDDSC module, the ACFP module, and the SGF submodule embedded in the ES-FPN structure improvement point can be found in Figure 1 .

[0040] based on Figure 1In this road defect detection model, the backbone network extracts multi-scale features from the input image using a sequentially connected GCBS–4×[GCBS+C3k2-MDDSC]-SPPF-ACFP algorithm. The GCBS module is used. This backbone network has four feature extraction modules at four scales, one for each dotted box. These are clearly numbered as the 1st, 2nd, 3rd, and 4th feature extraction modules along the data transmission direction.

[0041] Based on the feature extraction of each scale in the backbone network, the neck network is equipped with feature fusion modules corresponding to four scales. In the 1-4 feature fusion modules, each feature fusion module is equipped with an SGF-2 submodule; the 3rd and 4th fusion modules are also equipped with an SGF-3 submodule. For each scale feature fusion module, the features of the other two scales are first unified to the current scale, then spliced, and then output to the corresponding decoupled detection head through the c2f module for detection. When aligning scales, the large scale is aligned with the small scale using a convolution operation; the small scale is aligned with the large scale using an upsampling operation.

[0042] The corresponding detection network detects the target fusion features at each scale and outputs the detection results.

[0043] During the training process, in addition to using the 4×reg_max distributed regression strategy to make the detection box regression more detailed and stable, thereby improving positioning accuracy, the total number of categories (nc) is set to guide the output layer to generate a corresponding number of category prediction results, ensuring that the detection task can cover all target categories.

[0044] When the trained road defect detection model is put into application, redundant detection frame removal (such as non-maximum suppression, NMS) and confidence filtering operations can be further introduced to eliminate overlapping detection frames and screen out high-confidence prediction results, thereby improving the accuracy and reliability of the final detection results.

[0045] In this embodiment, YOLOv11s is used as the basic detection network, and its backbone network (Backbone) and neck network (Neck) both introduce multiple optimization modules to improve the accuracy and robustness of road disease detection. Specifically, they include: GCBS module (Grouped Convolution + BatchNorm + SiLU): Reduces computational complexity and parameter size through grouped convolution, improving feature decoupling capabilities. It combines batch normalization to stabilize training and utilizes the smooth gradient characteristics of the SiLU activation function to prevent gradient vanishing, thereby enhancing multi-scale feature processing efficiency while ensuring inference speed.

[0046] C3k2-MDDSC module: Combining gradient splitting with multi-scale dilated depthwise separable convolution strategies, it improves the gradient transfer efficiency and feature reuse capabilities of deep networks, while achieving efficient collaborative representation of local details and global semantics, significantly enhancing the model's adaptability to complex scenarios.

[0047] ACFP module (Adaptive Contextual Feature Pyramid): Placed at the end of the backbone, it captures contextual information of different granularities through multi-scale dilated convolution and combines the channel-spatial attention mechanism to adaptively adjust feature weights. This achieves cross-scale collaborative optimization of details and semantics, effectively suppressing background interference such as asphalt texture and road marking shadows.

[0048] The SGF submodule and the SGF-based ES-FPN structure perform fine spatial alignment and information fusion in the shallow feature pyramid (P2–P4) stage to enhance the spatial perception ability of small targets; improve the cross-level feature complementarity through the weighted fusion and Mean–Max fusion mechanisms, thereby alleviating the recognition difficulties caused by resolution and scale limitations in small target detection.

[0049] Through the organic combination and targeted optimization of the above modules, the improved YOLOv11 network of this embodiment significantly improves the accuracy and robustness of road damage detection under complex backgrounds, changing lighting, and multi-scale target conditions while maintaining lightweight and real-time performance.

[0050] In some embodiments, to ensure that the trained road disease detection model can meet expectations and accurately understand and detect road diseases at different scales in different scenarios, a specialized dataset can be created for training the road disease detection model. Specifically, a specialized UAV-based Specialized Multi-scale Road Disease Detection Dataset (UAV-SMRDD) was created. All samples in this dataset were collected through actual drone aerial photography, fully preserving the characteristics of road diseases from a low-altitude perspective, including real-world scene characteristics such as lighting variations, multi-scale objects, and complex backgrounds.

[0051] The UAV-SMRDD dataset contains a total of 5,000 high-resolution aerial images, covering five typical road damage categories: transverse cracks, longitudinal cracks, network cracks, potholes, and rutting.

[0052] For example, to intuitively present the morphological characteristics of road damage from the perspective of drones, refer to Figure 2 , Figure 2Representative sample images and their annotations from the UAV-SMRDD dataset are shown. Through systematic data collection and annotation, this dataset provides important benchmark data support for drone road inspections. Its construction method meets the requirements of dataset specifications in the field of computer vision.

[0053] For example, in order to provide a high-quality dataset, in addition to ensuring the number of images, images covering a variety of scenes including commercial streets, residential streets, and urban-rural fringe streets can also be collected.

[0054] For example, the image may contain common obstructions, such as trees, vehicles, buildings, etc.

[0055] For example, each scene may include multiple time periods and weather conditions such as day, night, sunny, rainy and foggy days to increase the complexity and comprehensiveness of the scene.

[0056] Such datasets can provide data support for road repairs and urban management, thereby enhancing the capabilities of the urban management platform and helping to improve urban management efficiency and intelligent decision-making.

[0057] To ensure the effectiveness of model training and the reliability of evaluation results, thereby safeguarding the model's robustness and generalization capabilities, the 5,000 images in the dataset were divided into three parts in a ratio of 7:1:2: First, a training set containing 3,500 images was used to train and optimize the road deterioration detection model, enabling the model to learn the various characteristics and patterns of road deterioration. Second, a validation set containing 500 images was designed to independently verify the algorithm's detection capabilities and accuracy, ensuring the model's performance on unseen data, thereby ensuring the reliability and practicality of the research results. In addition, a test set containing 1,000 images was used to finally evaluate the performance of the trained and validated algorithm.

[0058] This rigorous data allocation strategy, particularly during testing that closely mirrors real-world application scenarios, not only improved the model's training efficiency and generalization capabilities, but also effectively verified the model's accuracy and stability when dealing with unknown data. The setup of the test set is crucial for testing model performance in real-world environments. It provides an objective and unbiased evaluation of model performance, thereby demonstrating the feasibility and technological leadership of the research results in practical applications.

[0059] Preferably, when dividing the datasets, pay special attention to maintaining a balanced sample distribution across the three datasets, ensuring that the images in each dataset have similar statistical characteristics in terms of road hazard type, scale distribution, background environment, etc. This can avoid evaluation bias caused by uneven sample distribution and make the verification results more representative and reliable.

[0060] In some embodiments, in order to comprehensively and accurately measure the performance of each version of the road damage detection model, after at least one version of the road damage detection model is converged based on the training set training, it can be verified by the validation set, and the converged model can be evaluated by the test set to avoid model overfitting, verify the generalization of the model, and ensure that the model can run stably after deployment.

[0061] Specifically, the performance of the road damage detection model on the test set can be evaluated according to preset conditions.

[0062] For example, these preset conditions may include performance indicators such as the intersection-over-union ratio, detection accuracy, and recall rate of the road damage detection model for road damage.

[0063] That is to say, after verifying each version of the road damage detection model through the validation set, the road damage detection model of each version can be comprehensively evaluated based on the above-mentioned indicators, so as to determine the road damage detection model with the best performance from each version as the trained road damage detection model.

[0064] In some embodiments, the model runtime environment includes an Intel Xeon Platinum 8255C processor, 314 GB of memory, an NVIDIA Tesla V100 32 GB graphics card, and the operating system is CentOS 8.5.2 (64-bit). The deep neural network is built based on the PyTorch framework, with an input image size of [640, 640], and a multi-scale training strategy is used. The experiment sets the batch size to 64, trains for 300 epochs, and uses the SDG optimizer with an initial learning rate of 0.01 and a cosine decay strategy for optimization.

[0065] In some embodiments, the road disease detection model training is optimized by combining classification loss and regression loss. The classification loss uses binary cross-entropy loss (BCE Loss) to determine the anchor box category; the regression loss includes complete intersection over Union loss (CIoULoss) and dynamic category loss (DFL loss), which are used to measure the error between the predicted bounding box and the true bounding box respectively. CIoU Loss is improved on the basis of commonly used loss functions such as intersection over Union (IoU), generalized intersection over Union (GIoU), and distance intersection over Union (DIoU). Compared with the previous loss function, it adds a penalty term for the aspect ratio. When the center points of the predicted border and the true border coincide, it can better distinguish the errors in different situations and has scale invariance. Its formula is as follows:

[0066] in, L CIoU is CIoU Loss, b and Represent the center point of the real box and the center point of the predicted box respectively; ρ is the Euclidean distance between the predicted bounding box and the true bounding box; c is the diagonal distance between the predicted bounding box and the true bounding box; v To ensure the consistency of the relative proportions between the predicted bounding box and the true bounding box, IoU is the intersection-over-union ratio of the predicted bounding box and the true bounding box; α is the weight coefficient; w and are the width of the predicted border and the true border respectively; h and are the heights of the predicted bounding box and the true bounding box, respectively.

[0067] Based on CIoU Loss, the introduction of DFL loss can further improve the bounding box regression accuracy. DFL loss discretizes the uncertainty in the bounding box coordinate prediction, making the regression process more refined and robust, effectively reducing the prediction error. The formula of DFL loss is as follows:

[0068] in, S i is the cross entropy loss between the real border and the predicted border on the left, S i+1 is the cross entropy loss between the true box and the predicted box on the right.

[0069] However, although the CIoU loss takes into account the three important factors of positioning loss: overlapping area, center point distance and aspect ratio, the α in the CIoU loss formula is v ,There are still problems with the design of this item, which slows down the convergence speed.

[0070] To speed up convergence, we propose a new IoU loss, Fast Quality IoU Loss (FQ-IoU Loss), based on the penalty term of the original CIoU loss. FQ-IoU Loss consists of four components: overlap loss, center distance loss, width and height loss, and angle loss.

[0071] The core innovation of FQ-IoU Loss lies in optimizing the rotational alignment of the bounding box by introducing an angle penalty term, thereby significantly improving the convergence speed of the model. At the same time, this method abandons the traditional aspect ratio joint penalty mechanism and instead decouples the aspect ratio into independent width and height penalty terms. This decoupling design enables the model to make fine adjustments to the prediction errors in the width and height directions respectively, avoiding the dimensional coupling bias that may be caused by a single aspect ratio penalty. Through this dual-dimensional independent optimization strategy, FQ-IoU Loss can more accurately capture positioning errors in different directions, thereby significantly improving the positioning accuracy of the bounding box while maintaining angle consistency.

[0072] Introducing the exponential function form The IoU of θ enhances the sensitivity and adaptability of the loss function through nonlinear mapping. It assigns high gradients to small deviations (such as small target offsets) to accelerate convergence, while smoothing large deviations to avoid gradient explosion and improve optimization stability. At the same time, the exponential transformation compresses the distance to the range of (0,1] to ensure the stability of numerical calculations and improve the stability of the optimization through the exponential difference ratio (such as ) strengthens the optimization of geometric features. Furthermore, by incorporating trigonometric functions (such as the Λ term) to explicitly handle angular symmetry, detection is more robust to rotation and scale changes. Ultimately, this design dynamically balances the weights of multiple targets, improving accuracy while maintaining rapid convergence, making it particularly suitable for detection tasks in complex scenes.

[0073] Specifically, the formula of FQ-IoU Loss is as follows:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] in, is a fast and high-quality intersection-over-union loss; IoU is the intersection-over-union ratio between the predicted box and the true box; L IoU is the intersection and comparison loss; L dis is the center distance loss; L shp is the width and height loss; L ang is the angle loss; Δ is 2 times the center distance loss; b and Represent the center point of the real box and the center point of the predicted box respectively; Represents the square of the Euclidean distance between the center point of the real box and the predicted box; and Respectively represent the width and height of the minimum bounding rectangle; Ω is 2 times the width and height loss; w and are the width of the predicted border and the true border respectively; Represents the square of the Euclidean distance between the true box and the predicted box width; h and are the heights of the predicted bounding box and the true bounding box respectively; Represents the square of the Euclidean distance between the true box and the predicted box height; The angle loss is 2 times; and Represent the width and height of the rectangular box constructed by the center points of the true box and the predicted box, respectively.

[0080] Therefore, the formula for the total loss Loss is as follows:

[0081] Among them, λ1 and λ2 are balance coefficients.

[0082] Training the road defect detection model described in any of the aforementioned embodiments based on this total loss function can accelerate the model's convergence speed, thereby improving overall training efficiency. By continuously iteratively optimizing the road defect detection model, its defect recognition performance in complex scenarios becomes more robust and stable. Ultimately, the most robust version can be selected from multiple converged model versions as the final trained model.

[0083] Based on the network structure of the road damage detection model in the previous embodiment, this application proposes a road damage detection method based on drones.

[0084] The drone-based road defect detection method provided in the embodiments of the present application can be applied to electronic devices such as mobile phones, tablet computers, vehicle-mounted devices, augmented reality (AR) / virtual reality (VR) devices, laptop computers, ultra-mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs). The embodiments of the present application do not impose any restrictions on the specific types of electronic devices.

[0085] In order to illustrate the technical solution proposed in this application, each embodiment will be described below using an electronic device as the execution entity.

[0086] Figure 3 The schematic flow chart of the UAV-based road damage detection method provided by the present application is shown. The UAV-based road damage detection method includes: Step 310: The electronic device extracts features of the image to be detected collected from the perspective of the drone based on the backbone network of the pre-trained road defect detection model to obtain image features.

[0087] Step 320: The electronic device fuses the image features based on the neck network of the road damage detection model to obtain target fusion features.

[0088] Step 330: The electronic device detects the target fusion features based on the detection network of the road damage detection model to obtain a detection result of the road damage in the image to be detected.

[0089] The backbone network includes feature extraction modules of m scales, where m is a positive integer and m≥2. Each feature extraction module is equipped with a C3k2-MDDSC submodule. The C3k2-MDDSC submodule includes the first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer. For the first input feature of the C3k2-MDDSC submodule: Step A11: The electronic device performs a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature. The electronic device performs channel compression through the first CBS layer (convolution + batch normalization + activation), reducing the number of feature channels, reducing the amount of computation while retaining key information, and obtaining the first compressed feature, thereby providing efficient feature input for subsequent processing.

[0090] Step A12: The electronic device segments the first compressed feature into a first segmentation feature and a second segmentation feature through a segmentation layer.

[0091] For the first compressed feature, the electronic device segmentation layer divides the compressed feature into two parts (first segmentation feature and second segmentation feature) according to the channel, creating conditions for multi-path feature processing and realizing differentiated information extraction.

[0092] Step A13: The electronic device performs a multi-level progressive extraction operation on the first segmentation feature through a multi-level progressive stacking extraction layer to obtain hierarchical features corresponding to each level.

[0093] The electronic device performs a multi-level progressive feature extraction operation on the first segmentation feature by progressively stacking multiple levels of extraction layers. Assume that the extraction layer includes n levels, that is, is composed of n serially connected MDDSC or C3k-MDDSC submodules.

[0094] The first level directly uses the first segmentation feature as input to complete preliminary feature enhancement; each level from the n-1th to the nth level uses the output of the previous level as input, deepening the feature representation step by step; while transmitting the output layer by layer, the output of each level will also be directly sent to the first splicing layer for fusion, realizing cross-level feature aggregation.

[0095] Multi-level progressive stacking extraction layers utilize multi-path gradient propagation to effectively alleviate the gradient vanishing problem of deep networks, and significantly improve the feature reuse rate through a multi-branch structure, so that the network can more efficiently capture and learn complex patterns and multi-scale information while ensuring computational efficiency.

[0096] In step A14, the electronic device stitches the first segmentation feature, the second segmentation feature, and the features of each level together through a first stitching layer to obtain a first stitching feature.

[0097] Step A15: The electronic device performs a channel expansion operation on the first splicing feature through the second CBS layer to obtain a first output feature corresponding to the first input feature.

[0098] Subsequently, the electronic device fuses the first segmentation feature, the second segmentation feature and the output features of each level through the first splicing layer to obtain the first splicing feature; finally, the electronic device performs a channel expansion operation on the first splicing feature through the second CBS layer to obtain the first output feature corresponding to the first input feature.

[0099] In this embodiment, not only is the step-by-step depth enhancement and cross-level information fusion of input features achieved, but also multi-scale, multi-path and attention guidance mechanisms are introduced in the early stage of feature extraction, effectively improving the feature expressiveness and discriminability of the network; while maintaining computational efficiency, the model's ability to model complex patterns is significantly enhanced, thereby providing higher-quality feature support for subsequent detection tasks.

[0100] In some embodiments, Figure 4 and Figure 3 , respectively showing the specific network structure of the C3k2-MDDSC sub-module when the value of C3k is False and True.

[0101] In some embodiments, the C3k-MDDSC includes a third CBS layer, a dual MDDSC connected in series, a second splicing layer, and a fourth CBS layer. For a second input feature of the C3k-MDDSC: Step B11: The electronic device performs a feature compression operation on the second input feature through the third CBS layer to obtain a second compressed feature.

[0102] Step B12: The electronic device performs multi-scale feature extraction on the second compressed features through the dual MDDSCs connected in series to obtain multi-scale features.

[0103] Step B13: The electronic device performs a splicing operation on the second compressed feature and the multi-scale feature through a second splicing layer to obtain a second splicing feature.

[0104] Step B14: The electronic device performs a channel expansion operation on the second splicing feature through the fourth CBS layer to obtain a second output feature corresponding to the second input feature.

[0105] The electronic device first performs a feature compression operation on the second input feature through the third CBS layer to reduce the channel dimension and remove redundant information, thereby obtaining a second compressed feature. Subsequently, the second compressed feature is extracted using a series of dual MDDSC structures. The dual MDDSC structure uses a convolution combination with different receptive fields and void rates to capture local detail information and long-range dependency features in a single structure, thereby forming a multi-scale feature representation. Next, the electronic device fuses the second compressed feature with the multi-scale feature along the channel dimension through the second splicing layer to obtain a second spliced ​​feature. Finally, the electronic device performs a channel expansion operation on the second spliced ​​feature through the fourth CBS layer to obtain a second output feature corresponding to the second input feature.

[0106] It can be understood that when the value of C3k is Ture, the second output feature corresponding to the i-th level is the hierarchical feature of the i-th level.

[0107] For example, see Figure 6 , Figure 6 A schematic diagram of the network structure of C3k-MDDSC is shown.

[0108] In this embodiment, through the progressive design of compression, multi-scale extraction and fusion, the diversity of features and the richness of semantic levels are significantly improved while reducing the amount of computation; in particular, the introduction of the dual MDDSC structure enables the network to retain both fine texture and global context at the same stage, providing subsequent detection networks with high-quality feature input that is both detail-sensitive and semantically expressive.

[0109] In some embodiments, the MDDSC includes a first convolutional layer, a multi-scale dilated depth-separable convolutional layer, a third convolutional layer, a second convolutional layer, and a fourth convolutional layer; for the third input feature of the MDDSC: Step C11: The electronic device performs a channel compression operation on the third input feature through the first convolutional layer to obtain a third compressed feature.

[0110] Step C12: The electronic device performs at least two scales of atrous depth-separable convolution operations on the third compressed feature through a multi-scale atrous depth-separable convolution layer to obtain atrous depth-separable convolution features corresponding to each scale.

[0111] Step C13: Perform a splicing operation on the hole depth separable convolution features of each scale through a third splicing layer to obtain a third splicing feature.

[0112] Step C14: Perform a channel expansion operation on the third concatenated feature through the second convolutional layer to obtain an expanded feature.

[0113] Step C15: Perform a splicing operation on the third input feature and the extended feature through the fourth splicing layer to obtain a third output feature corresponding to the third input feature; the third output feature is a hierarchical feature of the level where the corresponding MDDSC is located.

[0114] The electronic device first performs a channel compression operation on the third input feature through the first convolution layer to reduce the number of feature channels, reduce the computational burden and remove redundant information, thereby obtaining a third compressed feature; then, a multi-scale hole depth separable convolution layer is used to perform hole depth separable convolution operations of at least two scales on the third compressed feature, and different scales correspond to different hole rates, thereby capturing local texture details and global context information under different receptive fields at the same stage, and obtaining hole depth separable convolution features of each scale; then, the electronic device uses the third splicing layer to perform channel splicing on the features of each scale and fuse them into a third splicing feature; thereafter, the second convolution layer is used to perform a channel expansion operation on the third splicing feature to restore the number of feature channels and enhance the feature expression capability to obtain an extended feature; finally, the electronic device splices the original third input feature with the expanded feature through the fourth splicing layer to form a third output feature, which corresponds to the hierarchical feature of the MDDSC level in which it is located, and will enter the subsequent multi-level feature fusion process.

[0115] In some embodiments, Figure 4 The network structure diagram of MDDSC is shown. Figure 4 The network structure, for the third input feature , the third output feature , B is the batch size, C 1 is the number of input channels, C 2 is the number of output channels, H is the height, W is the width. The electronic device first uses a 1×1 ordinary convolution (the first convolution layer) to X The number of channels of 1 is halved to obtain the first compression feature ,in , e is the channel compression ratio ( ), the specific expression is as follows:

[0116] Then, for each void rate , and perform dilated depthwise separable convolution (DDSConv) in sequence. The DDSConv features corresponding to each scale have been extracted :

[0117] DDSConv consists of two parts: (a) Dilated Depthwise Convolution (DDWConv) The depth convolution performs hole convolution on each channel of the feature map independently:

[0118] (b) Pointwise Convolution (PWConv) The number of channels is then expanded to C 2:

[0119] Next, all feature maps corresponding to the void ratio are concatenated along the channel dimension to obtain :

[0120] Finally, the multi-scale features are integrated through 1×1 convolution and the final output is generated through residual connection :

[0121] The C3k2-MSDDC module utilizes a collaborative optimization design approach combining a gradient diversion mechanism with a multi-scale feature fusion strategy. By constructing a multi-branch parallel structure and introducing depthwise separable convolution operations with different receptive fields, it achieves multi-level representation and interactive fusion of local texture features and global contextual information, significantly enhancing the model's accuracy in detecting small-scale road defects. While maintaining the spatial resolution of feature maps, this module leverages the lightweight nature of depthwise separable convolution to fully extract multi-scale features with low computational complexity, effectively alleviating the problem of easily lost features of small defects in complex road scenes.

[0122] In some embodiments, an ACFP module is provided at the end of the backbone network. The ACFP module includes a third convolutional layer, a multi-scale hole depth convolution structure, a fifth splicing layer, a fourth convolutional layer, a channel-spatial attention structure, and a residual fusion layer. For the fourth input feature of the ACFP module: D11. The electronic device performs a channel compression operation on the fourth input feature through the third convolutional layer to obtain a fourth compressed feature.

[0123] The electronic device performs a channel compression operation on the fourth input feature through the third convolutional layer to reduce redundant channel information and reduce the amount of subsequent calculations, thereby obtaining a fourth compressed feature.

[0124] D12. The electronic device performs dilated depth convolution operations and nonlinear transformation operations at different scales on the fourth compressed feature through a multi-scale dilated depth convolution structure to obtain depth convolution features corresponding to each scale.

[0125] The electronic device uses a multi-scale dilated deep convolution structure to perform dilated deep convolution operations on the fourth compressed feature at different dilation rates, and combines it with nonlinear transformation to enhance the nonlinear expression capability of the feature, thereby obtaining the deep convolution features corresponding to each scale and realizing feature capture of different receptive fields.

[0126] D13. The electronic device performs a splicing operation on the fourth compressed features and the depth convolution features of each scale through the fifth splicing layer to obtain a multi-scale splicing feature.

[0127] The electronic device splices the fourth compression feature with the depth convolution features of each scale in the channel dimension through the fifth splicing layer to obtain multi-scale splicing features, thereby realizing the fusion of local and global information.

[0128] D14. The electronic device performs a channel expansion operation on the multi-scale splicing features through the fourth convolutional layer.

[0129] D15. The electronic device performs spatial attention operations and channel attention operations on the multi-scale splicing features after the expanded channels through the channel-spatial attention structure, and fuses the obtained first spatial enhancement features and channel enhancement features to obtain spatial-channel enhancement features.

[0130] The electronic device performs a channel expansion operation on the multi-scale spliced ​​features through the fourth convolutional layer, restoring the number of channels to match the original features. Using the channel-spatial attention structure, the electronic device performs spatial attention and channel attention operations on the expanded multi-scale spliced ​​features to strengthen the response of key locations and key channels. The electronic device then fuses the obtained first spatial enhancement features with the channel enhancement features to generate spatial-channel enhancement features.

[0131] D16. The electronic device enhances the multi-scale splicing features after the expanded channel and the space-channel enhancement features through a residual fusion layer to obtain a fourth output feature corresponding to the fourth input feature.

[0132] The electronic device uses the residual fusion layer to fuse the multi-scale splicing features after channel expansion with the space-channel enhancement features element by element, achieves feature enhancement and efficient gradient transfer, and obtains the fourth output feature corresponding to the fourth input feature.

[0133] In this embodiment, this embodiment has achieved optimization in multiple links such as multi-scale feature extraction, channel-space attention guidance and residual fusion. On the one hand, the multi-scale hole deep convolution structure can capture local details and global context information under different receptive fields while maintaining the spatial resolution of the feature map, significantly enhancing the model's ability to represent targets with large size changes and subtle defects; on the other hand, the channel-space attention structure, with the advantages of fusion channel weight adjustment and spatial position focusing, effectively suppresses irrelevant responses in complex background areas and enhances the significance of defect area features. Ultimately, the residual fusion layer not only ensures the efficient transmission of key information and alleviates the gradient attenuation problem in the deep feature learning process, but also improves the overall detection accuracy and robustness through complementary feature enhancement, thereby achieving stable and high-precision detection of small-scale defect targets in complex road environments.

[0134] In some embodiments, the channel-spatial attention structure includes a channel attention branch, a spatial attention branch, and an enhanced fusion layer. The channel attention branch includes a global average pooling layer, a global maximum pooling layer, a sixth splicing layer, a fifth convolutional layer, a nonlinear transformation layer, a sixth convolutional layer, and a first activation layer. The spatial attention branch includes a Mean layer, a Max layer, a seventh splicing layer, a seventh convolutional layer, and a second activation layer. The aforementioned step D15 specifically includes: Step D151: For the channel attention branch: The electronic device performs average pooling and maximum pooling operations on the multi-scale spliced ​​features after channel expansion through a global average pooling layer and a global maximum pooling layer, respectively, to obtain average pooling features and maximum pooling features. The electronic device performs a splicing operation on the average pooling features and the maximum pooling features through a sixth splicing layer to obtain a fused pooling feature. The electronic device performs a channel compression operation, a nonlinear transformation operation, a channel expansion operation, and an activation operation on the fused pooling features through a fifth convolutional layer, a nonlinear transformation layer, a sixth convolutional layer, and a first activation layer, in sequence, to obtain a channel enhancement feature.

[0135] The channel attention branch aims to adaptively adjust feature weights from the channel dimension to highlight key feature channels and suppress redundant channel information. The global average pooling layer and the global maximum pooling layer statistically aggregate the multi-scale spliced ​​features after channel expansion, extracting different types of channel information representations from a global perspective: the average pooling feature reflects the overall response trend, and the maximum pooling feature reflects the significant response channel. The sixth splicing layer splices the two types of features to form a fused pooling feature to ensure information complementarity. The fifth convolutional layer performs channel compression to reduce dimensionality, reduce computational complexity, and aggregate information; the nonlinear transformation layer improves feature expression capabilities; the sixth convolutional layer performs channel expansion to restore the original number of channels; and the first activation layer introduces nonlinear mapping to obtain the final channel-enhanced features.

[0136] Step D152, for the spatial attention branch: the electronic device performs Mean operation and Max operation on the multi-scale splicing features after the expansion channel through the Mean layer and the Max layer respectively to obtain the average fusion feature and the maximum fusion feature; splices the average fusion feature and the maximum fusion feature through the seventh splicing layer to obtain the target fusion feature; performs convolution operation and activation operation on the target fusion feature in sequence through the seventh convolution layer and the second activation layer to obtain the first spatial enhancement feature.

[0137] The spatial attention branch aims to locate key information areas in the spatial dimension, allowing the network to focus on the spatial location of the disease target. The Mean layer and Max layer perform mean and maximum projections on the feature map in the channel dimension, respectively, to obtain the average fusion feature and the maximum fusion feature. The former preserves the overall spatial distribution information, while the latter highlights significant spatial locations. The seventh splicing layer concatenates the two types of features to form a more discriminative target fusion feature. The seventh convolutional layer and the second activation layer perform convolution extraction and nonlinear transformation on the target fusion feature to obtain the final first spatial enhancement feature.

[0138] In this embodiment, the collaborative design of two branches achieves dual attention guidance in both the channel and spatial dimensions: the channel attention branch ensures that the network adaptively emphasizes key information related to disease detection in the feature channel allocation, while the spatial attention branch enables the model to focus on the disease target area in the spatial domain of the feature map, significantly suppressing background interference. The combination of these two branches forms a complementary mechanism, which not only improves the discriminability and robustness of feature representation, but also enhances the detection accuracy and stability of small-scale disease targets in complex road scenes while maintaining the model's lightweight.

[0139] In some embodiments, Figure 8 The network structure diagram of the ACFP module is shown. Figure 5 The network structure, for the fourth input feature , first pass a 1×1 ordinary convolution (the third convolution layer) on X Perform channel compression to obtain the second compression feature ,in, , is the scaling factor, which takes a value of 0.5.

[0140]

[0141] Then, different void ratios are performed in turn. The hole depth convolution is performed, and the convolution results of each scale are activated by ReLU. Then the nonlinear transformation is performed by SiLU activation function to obtain the depth convolution features corresponding to each scale. Its expression is:

[0142] The scales of and After splicing, we get multi-scale splicing features :

[0143] Then, through the 1×1 ordinary convolution Perform channel adjustment to obtain multi-scale splicing features after expanding the channel :

[0144] Finally, Z Input the channel attention branch and the spatial attention branch respectively to perform channel-spatial attention enhancement, thereby more fully promoting the fusion of multi-scale features to obtain the fourth output feature :

[0145]

[0146]

[0147] In this embodiment, the ACFP module adopts a multi-branch dilated convolution structure and realizes efficient aggregation of multi-scale contextual information through differentiated receptive field design. Based on this, a channel-space dual attention mechanism is introduced to strengthen the expression of key information: the channel attention mechanism adaptively improves the weight of feature channels that are highly correlated with the disease target based on the correlation analysis between feature channels; the spatial attention mechanism focuses on the spatial position distribution of the disease target, effectively suppressing the interference response of the background noise area. This design of multi-scale feature fusion and dual attention collaborative guidance enables the network to accurately distinguish between real disease features and complex background information in strong interference environments such as asphalt texture and road marking shadows on the road surface, thereby significantly improving the detection accuracy and robustness of small-scale disease targets, and providing an efficient technical solution to solve the problem of road disease detection under complex background interference.

[0148] In some embodiments, corresponding to the m feature extraction modules, the neck network is provided with m scale feature fusion modules; each scale feature fusion module is provided with an SGF submodule. For example, the network structure diagram of the SGF submodule can be found in Figure 9 The SGF submodule includes a weighted fusion layer, a Mean-Max fusion layer, an eighth convolutional layer, a third activation layer, and an enhancement layer; for at least two fifth input features input to the SGF submodule: Step E11: The electronic device performs a weighting operation on each fifth feature through a weighted fusion layer, and concatenates the weighted fifth features to obtain a fused weighted feature.

[0149] The electronic device performs a weight assignment operation on each fifth input feature through a weighted fusion layer to highlight key information channels, and concatenates the weighted fifth input features to obtain a fused weighted feature.

[0150] Step E12: The electronic device performs Mean operation and Max operation on the fused weighted features through the Mean-Max fusion layer, and concatenates the fused weighted features after the Mean operation and the fused weighted features after the Max operation to obtain the Mean-Max fused features.

[0151] The electronic device performs mean (Mean) operation and maximum (Max) operation on the fused weighted features through the Mean-Max fusion layer to obtain feature expressions under different statistical characteristics, and splices the two to obtain the Mean-Max fusion feature.

[0152] Step E13: The electronic device sequentially performs convolution operations and activation operations on the Mean-Max fusion features through the eighth convolution layer and the third activation layer to obtain a second spatial enhancement feature.

[0153] The electronic device performs convolution operations and nonlinear activation operations on the Mean-Max fusion features through the eighth convolution layer and the third activation layer in sequence to obtain the second spatial enhancement features.

[0154] Step E14: The electronic device fuses the second spatial enhancement feature with the fused weighted feature through the enhancement layer to obtain a fifth output feature corresponding to the fifth input feature.

[0155] The electronic device performs feature fusion on the second spatial enhancement feature and the fusion weighted feature through the enhancement layer to form a comprehensive feature representation that has both local response enhancement and global context preservation, and obtains the fifth output feature corresponding to the fifth input feature.

[0156] In this embodiment, the SGF submodule weights the multi-scale input features through the weighted fusion layer, dynamically adjusts the importance of different features, and realizes effective screening and highlighting of information. Subsequently, the Mean-Max fusion layer is used to extract the mean and maximum response of the features respectively, capture the spatial information under different statistical characteristics, and enhance the richness and discriminability of feature expression. The fusion features are further refined through the convolution and activation layers to improve the spatial semantic expression ability. Finally, the enhancement layer combines the spatial enhancement features with the original fusion weighted features to achieve the coordinated optimization of local details and global context. This design significantly improves the model's spatial perception and detail capture capabilities for small targets, effectively alleviates the problems of information loss and noise interference in traditional multi-scale fusion, thereby improving the accuracy and robustness of small-scale road disease detection and ensuring the stable performance of the model in complex environments.

[0157] The SGF submodules include the SGF-2 submodule and the SGF-3 submodule, which have the same network structure but are set at different locations; The SGF-2 submodule is set in the feature fusion module of each scale; for the i-th scale, the SGF-2 submodule is used to extract the target features corresponding to the i-th scale and the alignment features corresponding to the i+1-th scale; i∈[1,m]; The feature extraction modules of the 2nd to m-1th scales are all equipped with SGF-3 submodules; for the jth scale, the SGF-3 submodule is used to fuse the target extraction features corresponding to the jth scale, the initial fusion features corresponding to the jth scale, and the final fusion features corresponding to the j-1th scale; the initial fusion features are obtained based on the feature fusion output by the SGF-2 submodule of the jth scale; the final fusion features are obtained based on the fusion of the outputs of the SGF-2 submodule of the j-1th scale or the SGF-3 submodule of the j-1th scale; j∈[2,m-1].

[0158] For example, Figure 10 A schematic diagram of the network structure of ES-FPN designed based on the SGF-2 submodule and the SGF-3 submodule is shown.

[0159] In some embodiments, for the UAV road defect detection task, based on the YOLOv11s architecture, a number of key improvements such as the C3k2-MDDSC submodule, ACFP module, and SGF submodule (or ES-FPN structure based on SGF) are systematically introduced to achieve a significant improvement in model performance.

[0160] First, the C3k2-MDDSC module enhances the model's multi-level representation ability of local texture details and global context by combining gradient splitting and multi-scale depth-separable convolution, effectively alleviating the gradient disappearance problem of deep networks. Figure 1GCBS in

[15] reduces the risk of overfitting, ensuring lightweight while improving feature reuse rate and generalization performance.

[0161] Secondly, the ACFP module adopts a multi-branch dilated convolution structure at the end of the backbone network, uses differentiated receptive fields to realize multi-scale contextual information aggregation, and integrates the channel-space dual attention mechanism to dynamically adjust the weights of different feature channels and spatial positions, significantly enhancing the ability to recognize small-scale diseases in complex backgrounds, especially in strong interference environments such as asphalt textures and road marking shadows.

[0162] Finally, the SGF submodule is responsible for the weighted fusion and spatial alignment of multi-scale features in the neck network. The weighted fusion layer highlights key information channels, the Mean-Max fusion layer captures spatial features with different statistical characteristics, and the convolutional activation and enhancement layers are combined to achieve collaborative optimization of details and context. This greatly improves the model's spatial perception and detail capture capabilities for tiny disease targets, alleviating the problems of information loss and noise interference in traditional multi-scale fusion.

[0163] In summary, the synergistic effect of various modules not only comprehensively improves the model's detection accuracy and robustness for small-scale road defects, but also enhances its ability to suppress diverse background noise in complex scenarios, optimizes the model's training efficiency and inference speed, ensures the stability and practicality of intelligent identification of road defects from the perspective of drones, and provides strong technical support for practical applications.

[0164] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0165] Corresponding to the road damage detection method based on drone in the above embodiment, Figure 11 A structural block diagram of a road defect detection device 11 provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0166] Reference Figure 11 , the road disease detection device 11 comprises: The extraction module 111 is used to extract features from the image to be detected collected from the perspective of the drone based on the backbone network of the pre-trained road disease detection model to obtain image features; the image to be detected includes roads; A fusion module 112 is used to fuse image features based on the neck network of the road damage detection model to obtain target fusion features; The detection module 113 is used to detect the target fusion features based on the detection network of the road damage detection model to obtain the detection result of the road damage in the image to be detected; The backbone network includes feature extraction modules of m scales, where m is a positive integer and m≥2. Each feature extraction module is equipped with a C3k2-MDDSC submodule. The C3k2-MDDSC submodule includes the first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer. For the first input feature of the C3k2-MDDSC submodule: Performing a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature; Splitting the first compressed feature into a first segmentation feature and a second segmentation feature through a segmentation layer; Performing a multi-level progressive extraction operation on the first segmentation feature through a multi-level progressive stacking extraction layer to obtain hierarchical features corresponding to each level; The first segmentation feature, the second segmentation feature, and the features of each level are spliced ​​together through the first splicing layer to obtain a first splicing feature; Perform a channel expansion operation on the first concatenated feature through the second CBS layer to obtain a first output feature corresponding to the first input feature; For each feature extraction module that extracts shallow features, when the corresponding C3k value in C3k2-MDDSC is False, the multi-level progressive stacking extraction layer is improved based on MDDSC; for each feature extraction module that extracts deep features, when the corresponding C3k value in C3k2-MDDSC is True, the multi-level progressive stacking extraction layer is improved based on C3k-MDDSC.

[0167] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0168] Figure 12 This is a schematic diagram of the physical structure of an electronic device provided in one embodiment of the present application. Figure 8 As shown, the electronic device 12 of this embodiment includes: at least one processor 120 ( Figure 12 Only one processor is shown), a memory 121, and a computer program 122 stored in the memory 121 and executable on at least one processor 120. When the processor 120 executes the computer program 122, the steps of any of the above-mentioned embodiments of the method for detecting road defects based on a drone are implemented, such as Figure 3 Steps 310-330 are shown.

[0169] The processor 120 may be a central processing unit (CPU), or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0170] In some embodiments, the memory 121 may be an internal storage unit of the electronic device 12, such as a hard disk or memory of the electronic device 12. In other embodiments, the memory 121 may also be an external storage device of the electronic device 12, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 12.

[0171] Furthermore, the memory 121 may include both an internal storage unit of the electronic device 12 and an external storage device. The memory 121 is used to store operating devices, applications, boot loaders, data, and other programs, such as program code of computer programs. The memory 121 may also be used to temporarily store data that has been output or is about to be output.

[0172] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0173] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the steps in the above-mentioned various method embodiments.

[0174] An embodiment of the present application provides a computer program product. When the computer program product is run on an electronic device, the electronic device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0175] If this integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to the camera / electronic device, a recording medium, computer memory, read-only memory (ROM), random access memory (RAM), an electrical carrier signal, a telecommunications signal, and a software distribution medium. Examples include a USB flash drive, a removable hard drive, a magnetic disk, or an optical disk.

[0176] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0177] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0178] In the embodiments provided in this application, it should be understood that the disclosed devices / network equipment and methods can be implemented in other ways. For example, the device / network equipment embodiments described above are merely illustrative. For example, the division of the above modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0179] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0180] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A road disease detection method based on drone, characterized in that: include: Based on the backbone network of the pre-trained road damage detection model, feature extraction is performed on the image to be detected collected from the perspective of the drone to obtain image features; The image to be detected includes a road; fusing the image features based on the neck network of the road damage detection model to obtain target fusion features; Detecting the target fusion features based on the detection network of the road damage detection model to obtain a detection result of road damage in the image to be detected; The backbone network includes feature extraction modules of m scales, where m is a positive integer and m≥2; each feature extraction module is provided with a C3k2-MDDSC submodule; the C3k2-MDDSC submodule includes a first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer. For the first input feature of the C3k2-MDDSC submodule: Performing a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature; Segmenting the first compressed feature into a first segmentation feature and a second segmentation feature by the segmentation layer; Performing a multi-level progressive extraction operation on the first segmentation feature through the multi-level progressive stacking extraction layer to obtain hierarchical features corresponding to each level; Splicing the first segmentation feature, the second segmentation feature, and each of the hierarchical features through the first splicing layer to obtain a first splicing feature; Performing a channel expansion operation on the first concatenated feature through the second CBS layer to obtain a first output feature corresponding to the first input feature; For each feature extraction module that extracts shallow features, when the corresponding C3k value in C3k2-MDDSC is False, the multi-level progressive stacking extraction layer is obtained based on the improvement of MDDSC; for each feature extraction module that extracts deep features, when the corresponding C3k value in C3k2-MDDSC is True, the multi-level progressive stacking extraction layer is obtained based on the improvement of C3k-MDDSC.

2. The road damage detection method according to claim 1, wherein: The C3k-MDDSC includes a third CBS layer, a dual MDDSC connected in series, a second splicing layer, and a fourth CBS layer. For the second input feature of the C3k-MDDSC: Performing a feature compression operation on the second input feature through the third CBS layer to obtain a second compressed feature; Performing multi-scale feature extraction on the second compressed features by the series-connected dual MDDSC to obtain multi-scale features; Performing a splicing operation on the second compressed feature and the multi-scale feature through the second splicing layer to obtain a second splicing feature; A channel expansion operation is performed on the second splicing feature through the fourth CBS layer to obtain a second output feature corresponding to the second input feature; the second output feature is a hierarchical feature of the corresponding layer of the C3k-MDDSC.

3. The road damage detection method according to claim 1, wherein: The MDDSC includes a first convolutional layer, a multi-scale hole depth-separable convolutional layer, a third splicing layer, a second convolutional layer, and a fourth splicing layer; for the third input feature of the MDDSC: Performing a channel compression operation on the third input feature through the first convolutional layer to obtain a third compressed feature; Performing at least two scales of atrous depth-separable convolution operations on the third compression feature through the multi-scale atrous depth-separable convolution layer to obtain atrous depth-separable convolution features corresponding to each scale; atrous depth-separable convolution operations of different scales correspond to different atrous rates; Performing a splicing operation on the hole depth separable convolution features of each scale through the third splicing layer to obtain a third splicing feature; Performing a channel expansion operation on the third concatenated feature through the second convolutional layer to obtain an expanded feature; A splicing operation is performed on the third input feature and the extended feature through the fourth splicing layer to obtain a third output feature corresponding to the third input feature; the third output feature is a hierarchical feature of the level where the corresponding MDDSC is located.

4. The road damage detection method according to claim 1, wherein: The backbone network is provided with an ACFP module at the end, which includes a third convolutional layer, a multi-scale hole depth convolution structure, a fifth splicing layer, a fourth convolutional layer, a channel-spatial attention structure, and a residual fusion layer; for the fourth input feature of the ACFP module: Performing a channel compression operation on the fourth input feature through the third convolutional layer to obtain a fourth compressed feature; Performing a dilated depth convolution operation and a nonlinear transformation operation at different scales on the fourth compressed feature through the multi-scale dilated depth convolution structure to obtain a depth convolution feature corresponding to each scale; Performing a splicing operation on the fourth compressed features and the depth convolution features of each scale through the fifth splicing layer to obtain a multi-scale splicing feature; Performing a channel expansion operation on the multi-scale splicing feature through the fourth convolutional layer; Performing spatial attention operations and channel attention operations on the multi-scale splicing features after the channel expansion through the channel-spatial attention structure, and fusing the obtained first spatial enhancement features and channel enhancement features to obtain spatial-channel enhancement features; The multi-scale splicing feature after the channel expansion and the space-channel enhancement feature are enhanced through the residual fusion layer to obtain a fourth output feature corresponding to the fourth input feature.

5. The road damage detection method according to claim 4, characterized in that: The channel-spatial attention structure includes a channel attention branch, a spatial attention branch and an enhanced fusion layer. The channel attention branch includes a global average pooling layer, a global maximum pooling layer, a sixth splicing layer, a fifth convolutional layer, a nonlinear transformation layer, a sixth convolutional layer, and a first activation layer. The spatial attention branch includes a Mean layer, a Max layer, a seventh splicing layer, a seventh convolutional layer, and a second activation layer. The performing of a spatial attention operation and a channel attention operation on the multi-scale splicing features after the channel expansion through the channel-spatial attention structure includes: For the channel attention branch: Performing an average pooling operation and a maximum pooling operation on the multi-scale splicing features after the expansion channel through the global average pooling layer and the global maximum pooling layer, respectively, to obtain an average pooling feature and a maximum pooling feature; Performing a splicing operation on the average pooling feature and the maximum pooling feature through the sixth splicing layer to obtain a fused pooling feature; Performing a channel compression operation, a nonlinear transformation operation, a channel expansion operation, and an activation operation on the fused pooling feature in sequence through the fifth convolution layer, the nonlinear transformation layer, the sixth convolution layer, and the first activation layer to obtain a channel enhancement feature; For the spatial attention branch: Performing a Mean operation and a Max operation on the multi-scale splicing features after the expansion channel through the Mean layer and the Max layer respectively to obtain an average fusion feature and a maximum fusion feature; The average fusion feature and the maximum fusion feature are concatenated through the seventh concatenation layer to obtain a target fusion feature; The convolution operation and the activation operation are sequentially performed on the target fusion feature through the seventh convolution layer and the second activation layer to obtain the first spatial enhancement feature.

6. The road damage detection method according to any one of claims 1 to 5, characterized in that: Corresponding to the m feature extraction modules, the neck network is correspondingly provided with feature fusion modules of m scales; the feature fusion module of each scale is provided with an SGF submodule, and the SGF submodule includes a weighted fusion layer, a Mean-Max fusion layer, an eighth convolutional layer, a third activation layer, and an enhancement layer; for at least two fifth input features input into the SGF submodule: Performing a weighting operation on each of the fifth features through the weighted fusion layer, and concatenating the weighted fifth features to obtain a fused weighted feature; Performing Mean and Max operations on the fused weighted features respectively through the Mean-Max fusion layer, and concatenating the fused weighted features after the Mean operation with the fused weighted features after the Max operation to obtain a Mean-Max fused feature; Performing convolution operations and activation operations on the Mean-Max fusion features in sequence through the eighth convolution layer and the third activation layer to obtain a second spatial enhancement feature; The second spatial enhancement feature is fused with the fusion weighted feature through the enhancement layer to obtain a fifth output feature corresponding to the fifth input feature.

7. The road damage detection method according to claim 6, characterized in that: The SGF submodules include an SGF-2 submodule and an SGF-3 submodule with the same network structure but different settings; The SGF-2 submodule is provided in the feature fusion module of each scale; for the i-th scale, the SGF-2 submodule is used to extract the target features corresponding to the i-th scale and the alignment features corresponding to the i+1-th scale; the i∈[1,m]; The feature extraction modules of the 2nd to m-1th scales are all provided with an SGF-3 submodule; for the jth scale, the SGF-3 submodule is used to fuse the target extraction features corresponding to the jth scale, the initial fusion features corresponding to the jth scale, and the final fusion features corresponding to the j-1th scale; The initial fusion feature is obtained based on the feature fusion output by the SGF-2 submodule of the j-th scale; The final fusion feature is obtained based on the fusion of the output of the SGF-2 submodule at the j-1th scale or the SGF-3 submodule at the j-1th scale; Said j∈[2,m-1].

8. The road damage detection method according to any one of claims 1 to 5, characterized in that: The road damage detection model is trained based on regression loss and classification loss; the regression loss includes fast and high-quality intersection-over-union loss, and the formula is as follows: Among them, the is a fast and high-quality intersection-over-union loss; the IoU is the intersection-over-union ratio between the predicted box and the real box; the L IoU is the intersection and union loss; L dis is the center distance loss; L shp is the width and height loss; L ang is the angle loss; the Δ is 2 times the center distance loss; the b and stated Represent the center point of the real box and the center point of the predicted box respectively; Represents the square of the Euclidean distance between the center point of the real box and the predicted box respectively; and stated Respectively represent the width and height of the minimum circumscribed rectangular frame; the Ω is 2 times the width and height loss; the w and are the widths of the predicted border and the true border respectively; Represents the square of the Euclidean distance between the true box and the predicted box width; h and are the heights of the predicted border and the true border respectively; Represents the square of the Euclidean distance between the true box and the predicted box height; The angle loss is 2 times; and stated Represent the width and height of the rectangular box constructed by the center points of the true box and the predicted box, respectively.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the drone-based road damage detection method according to any one of claims 1 to 8 is implemented.

10. A computer program product, wherein the computer program product stores a computer program, characterized in that: When the computer program is executed by a processor, the drone-based road damage detection method according to any one of claims 1 to 8 is implemented.

Images