Unmanned aerial vehicle-based road disease detection method, electronic device, and program product

Abstract

The application discloses a kind of road disease detection method based on unmanned plane, electronic equipment and program product. The method is realized based on the road disease detection model of training completion, the C3k2-MDDSC module introduced in the backbone network of the model, the feature reuse capability is enhanced by gradient shunt and multi-scale fusion, the gradient vanishing problem is alleviated, and the model robustness is improved with the aid of jump connection and grouping convolution. The ACFP module is introduced at the end of the backbone network, then the dynamic fusion of local and global features is realized using multi-scale hollow convolution and channel-space attention mechanism, and the complex scene modeling capability is improved. The SGF module is integrated in the neck network, which can improve the spatial perception of the model for small targets. In addition, the ES-FPN proposed based on the SGF module not only can strengthen the utilization of shallow spatial information, but also can optimize the complementarity of cross-level features. During training, a fast and high-quality intersection over union loss is proposed as a regression loss, which improves the rotation box alignment accuracy and model convergence speed by introducing angle penalty.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a road disease detection method based on a UAV, an electronic device, and a computer program product. BACKGROUND

[0002] If road diseases such as cracks and pits are not found in time, driving safety may be affected and the road surface may be aged. Traditional manual inspection is low in efficiency and limited in coverage, and a vehicle-mounted detection system has problems such as high cost and poor flexibility.

[0003] In recent years, a UAV has gradually been applied to road disease detection due to its advantages such as flexibility, high efficiency, and low cost, and is particularly suitable for remote or dangerous road sections. In combination with a deep learning algorithm, automatic identification of diseases can be achieved, and the detection efficiency and accuracy can be improved, and the maintenance cost can be reduced.

[0004] However, in actual application, small-scale diseases have a small proportion in images under the UAV perspective and weak features, and are easily missed, which affects the identification effect. Therefore, how to improve the detection capability of small diseases in UAV images has become a key problem in the development of the technology. SUMMARY

[0005] The application provides a road disease detection method based on a UAV, an electronic device, and a computer program product, which can effectively alleviate the difficulty in identification caused by a low proportion of disease target pixels and unobvious texture features, and further improve the detection accuracy and stability of small-scale diseases.

[0006] In a first aspect, the application provides a road disease detection method based on a UAV, comprising:

[0007] extracting features of a to-be-detected image collected under the UAV perspective based on a backbone network of a pre-trained road disease detection model, to obtain image features; the to-be-detected image comprises a road;

[0008] fusing the image features based on a neck network of the road disease detection model, to obtain target fusion features;

[0009] detecting the target fusion features based on a detection network of the road disease detection model, to obtain a detection result of road diseases in the to-be-detected image;

[0010] The backbone network comprises m scale feature extraction modules, m is a positive integer, and m≥2; each feature extraction module is provided with a C3k2-MDDSC submodule; the C3k2-MDDSC submodule comprises a first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer, and for a first input feature input into the C3k2-MDDSC submodule:

[0011] perform a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature;

[0012] segment the first compressed feature into a first segmented feature and a second segmented feature through the segmentation layer;

[0013] perform a multi-level progressive extraction operation on the first segmented feature through the multi-level progressive stacking extraction layer to obtain a level feature corresponding to each level;

[0014] splice the first segmented feature, the second segmented feature and the level features through the first splicing layer to obtain a first spliced feature;

[0015] perform a channel expansion operation on the first spliced feature through the second CBS layer to obtain a first output feature corresponding to the first input feature;

[0016] Each feature extraction module for extracting a shallow layer feature has a C3k value of False in the corresponding C3k2-MDDSC, and the multi-level progressive stacking extraction layer is improved based on the MDDSC. When the C3k value of each feature extraction module for extracting a deep layer feature is True in the corresponding C3k2-MDDSC, the multi-level progressive stacking extraction layer is improved based on the C3k-MDDSC.

[0017] In a second aspect, the present application provides a road disease detection device, comprising:

[0018] The extraction module is configured to extract features from a to-be-detected image collected from the perspective of the unmanned aerial vehicle based on a backbone network of a pre-trained road disease detection model to obtain image features, wherein the to-be-detected image includes a road.

[0019] The fusion module is configured to fuse the image features based on a neck network of the road disease detection model to obtain target fusion features.

[0020] The detection module is configured to detect the target fusion features based on a detection network of the road disease detection model to obtain a detection result of a road disease in the to-be-detected image.

[0021] The backbone network includes m scale feature extraction modules, m is a positive integer, and m is greater than or equal to 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, and for a first input feature input into the C3k2-MDDSC submodule:

[0022] perform a channel compression operation on the first input feature through the first CBS layer to obtain a first compressed feature;

[0023] The first compressed feature is split into a first split feature and a second split feature by a split layer;

[0024] A multi-level progressive extraction operation is performed on the first split feature by a multi-level progressive stacking extraction layer to obtain a level feature corresponding to each level;

[0025] The first split feature, the second split feature and the level features are spliced by a first splicing layer to obtain a first splicing feature;

[0026] A channel expansion operation is performed on the first splicing feature by a second CBS layer to obtain a first output feature corresponding to the first input feature;

[0027] Each feature extraction module for extracting a shallow layer feature has a C3k value of False in the corresponding C3k2-MDDSC, and the multi-level progressive stacking extraction layer is improved based on the MDDSC; when each feature extraction module for extracting a deep layer feature has a C3k value of True in the corresponding C3k2-MDDSC, the multi-level progressive stacking extraction layer is improved based on the C3k-MDDSC.

[0028] In a third aspect, the present application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method of the first aspect when executing the computer program.

[0029] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the method of the first aspect when executed by a processor.

[0030] In a fifth aspect, the present application provides a computer program product, which comprises a computer program, and the computer program implements the steps of the method of the first aspect when executed by one or more processors.

[0031] Compared with the prior art, the beneficial effects of the present application are: by introducing the C3k2-MDDSC sub-module in the feature extraction module of each scale of the backbone network, the adaptive optimization of different levels of features can be realized: the C3k2-MDDSC sub-module corresponding to the shallow feature extraction module with C3k=False, the multi-level progressive stacking extraction layer in the sub-module is obtained based on MDDSC, the structure not only strengthens the multi-scale progressive extraction ability of local details and edge textures, so that small-scale diseases can be fully represented in the early stage although they occupy a small proportion and have weak features in the image; the C3k2-MDDSC sub-module corresponding to the deep feature extraction module with C3k=True, the multi-level progressive stacking extraction layer in the sub-module is obtained based on the C3k-MDDSC structure, the structure retains the spatial detail information of the shallow layer while maintaining the global semantic modeling capability, thereby avoiding the loss of small target information by deep features. The multi-level progressive stacking and splicing operation based on gradient shunting enhances the complementarity of features at each level, so that the weight of small-scale diseases is improved in multi-scale feature fusion, significantly reducing the probability of missed detection. Under the perspective of the unmanned aerial vehicle, the design can effectively alleviate the difficulty of recognition caused by low pixel proportion of disease targets and unobvious texture features, thereby improving the detection accuracy and stability of small-scale diseases.

[0032] It can be understood that the beneficial effects of the above-mentioned second aspect to the fifth aspect can be referred to the related description in the first aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0034] Figure 1 is a network structure schematic diagram of a road disease detection model provided by the embodiments of the present application;

[0035] Figure 2 is an example diagram in the UAV-SMRDD data set provided by the embodiments of the present application;

[0036] Figure 3 is a flowchart of the UAV-based road disease detection method provided by the embodiments of the present application;

[0037] Figure 4 is a network structure schematic diagram of the C3k2-MDDSC sub-module when C3k is False;

[0038] Figure 5is a network structure schematic diagram of a C3k2-MDDSC sub-module when a C3k value is True, provided by an embodiment of the present application;

[0039] Figure 6 is a network structure schematic diagram of a C3k-MDDSC, provided by an embodiment of the present application;

[0040] Figure 7 is a network structure schematic diagram of a MDDSC, provided by an embodiment of the present application;

[0041] Figure 8 is a network structure schematic diagram of an ACFP module, provided by an embodiment of the present application;

[0042] Figure 9 is a network structure schematic diagram of an SGF sub-module, provided by an embodiment of the present application;

[0043] Figure 10 is a schematic diagram of an ES-FPN structure, provided by an embodiment of the present application;

[0044] Figure 11 is a structure schematic diagram of a road disease detection device, provided by an embodiment of the present application;

[0045] Figure 12 is a structure schematic diagram of an electronic device, provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] In the following description, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the present embodiments. However, persons skilled in the art will understand that the present application can be practiced without these specific details. In other instances, well-known structures, devices, circuits, and methods have not been described in detail in order to avoid obscuring the present application.

[0047] Drones are widely used in road disease detection due to their advantages of flexibility, high efficiency, low cost, etc., and are particularly suitable for remote or dangerous sections. Combined with deep learning algorithms, automatic disease identification can be achieved, significantly improving detection efficiency and accuracy and reducing maintenance costs. However, in actual applications, small-scale diseases have low pixel ratio and weak feature expression in drone visual images, which can easily lead to missed detection and affect overall recognition effect. Therefore, improving the detection capability of small-scale diseases in drone images has become a key problem that needs to be broken through in this field.

[0048] The present application research finds that the traditional detection method is difficult to fully excavate the subtle characteristics of the target, mainly for two reasons: firstly, the continuous downsampling of the deep network gradually weakens the details of the small target, making it difficult to be effectively represented in the deep feature map, thereby affecting the positioning accuracy; secondly, although the shallow features retain rich spatial details, the semantic expression ability is insufficient, and in a complex background, it is easy to be disturbed by noise, further increasing the difficulty of small target feature extraction. Therefore, how to balance detail preservation and semantic expression in the feature extraction process, and realize efficient fusion and balance of the two, has become a key challenge to improve the performance of small target detection.

[0049] Based on this idea, the present application proposes a road disease detection model, which includes a backbone network, a neck network and a detection network. Specifically, the backbone network is used for feature extraction of the image containing the road to be detected collected by the unmanned aerial vehicle, to obtain image features; the neck network is used for feature fusion of the image features to obtain target fusion features; the detection network is used for detection of the target fusion features to obtain the detection result of the road disease in the image to be processed.

[0050] In order to balance detail preservation and semantic expression, and realize efficient fusion and balance of the two, a C3k2-MDDSC module is introduced in the backbone network of the road disease detection model. The module is improved based on the C3k2 structure combined with multi-scale dilated depthwise separable convolution (MDDSC), which fully utilizes the flexibility and efficiency of C3k2 in deep and shallow feature extraction, so that the model can strengthen the expression ability of multi-scale features while maintaining lightweight.

[0051] Specifically, the backbone network contains m scale feature extraction modules (m≥2), each of which is configured with a C3k2-MDDSC submodule. The submodule includes a first CBS layer, a split layer, a multi-level progressive stacking extraction layer, a first splicing layer and a second CBS layer: the split layer (split) divides the first compressed feature into two split features, denoted as the first split feature and the second split feature. Among them, the first split feature can be directly transmitted to the subsequent splicing layer after being progressively extracted by multiple levels to capture the details and semantic information under different receptive fields; the second split feature directly retains the original information; then the two split features and the features of each level are fused in the first splicing layer, and the channel number is restored through the second CBS layer to realize 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 under the perspective of the unmanned aerial vehicle.

[0052] The multi-level progressive stacked extraction layer of the C3k2 module can be dynamically adjusted according to a feature pyramid level (P). In a shallow layer (such as P3 / 8), the C3k takes a value of False, and the multi-level progressive stacked extraction layer is improved based on the MDDSC; in a deep layer (such as P5 / 32), the C3k takes a value of True, and the multi-level progressive stacked extraction layer is improved based on the C3k-MDDSC.

[0053] By way of example, it is assumed that the multi-level progressive stacked extraction layer contains n levels, that is, is composed of n MDDSCs / C3k-MDDSCs connected in series. Except that the first level directly takes the first segmented feature as input, the remaining levels all take the output of the previous level as input; each level, while passing the output to the next level, also sends the output to the first concatenation layer to participate in fusion. This design effectively alleviates the gradient vanishing problem of the deep network through multi-path gradient propagation, and improves the feature reuse rate with the multi-branch structure, so that the network can more efficiently learn complex patterns.

[0054] In some embodiments, when the C3k takes a value of True, to enhance the expression ability of deep features in the multi-level progressive stacked extraction layer, the structure of the C3k-MDDSC includes a third CBS layer, a double MDDSC connected in series, a second concatenation layer, and a fourth CBS layer, to realize sufficient extraction and fusion of deep features.

[0055] By designing two MDDSCs connected in series in the multi-level progressive stacked extraction layer of the C3k-MDDSC, not only is the multi-scale extraction and refinement of deep features realized in stages, but also the gradient propagation path is enriched with the cooperation of the multi-branch structure, alleviating the gradient vanishing problem of the deep network. The first MDDSC focuses on expanding the receptive field to capture global semantics and long-range dependency relationships, and the second MDDSC enhances the ability to depict local spatial details on this basis, and the combination of the two makes the features have stronger hierarchy and discriminability in the multi-scale space. This design not only improves the recognition ability of the model for small-scale diseases in complex backgrounds, but also enhances the robustness and generalization performance of feature expression, thereby effectively improving the precision and stability of road disease detection under the perspective of an unmanned aerial vehicle.

[0056] In some embodiments, the MDDSC includes a first convolutional layer, a multi-scale hollow depth separable convolutional layer, a third concatenation layer, a second convolutional layer, and a fourth concatenation layer. The core idea is to capture rich contextual diversity at very low computational cost, so when the C3k takes a value of False, the residual link does not need to be disabled. The multi-scale hollow depth separable convolutional layer uses multiple depth separable convolutions with different hole rates (for example, 5 scales, with hole rates of 1, 2, 3, 4, and 5) in parallel to effectively extract multi-scale context information, significantly enhancing the adaptability of the model to complex scenes.

[0057] The small cavity rate branch focuses on local detail features (such as edges and textures), and the large cavity rate branch is responsible for capturing global semantic information (such as object structure and scene layout), thereby realizing multi-scale fusion while maintaining feature map resolution, and avoiding information loss caused by traditional down-sampling. The depth separable convolution structure effectively reduces the computational complexity, and realizes efficient collaborative representation of local details and global semantics.

[0058] In the unmanned aerial vehicle road disease inspection task, due to the complex background interference (such as asphalt texture, marking line, shadow and illumination change, etc.) on the road surface, the traditional single-scale feature extraction method often cannot effectively distinguish the real disease from the background noise.

[0059] In some embodiments, in order to effectively distinguish the disease from the noise in the complex background, an adaptive contextual feature pyramid (ACFP) module is designed at the end of the backbone network. The module includes a third convolutional layer, a multi-scale dilated depth convolution structure, a fifth concatenation layer, a fourth convolutional layer, a channel-spatial attention mechanism and a residual fusion layer, and an adaptive multi-scale feature fusion framework is constructed.

[0060] The multi-scale dilated depth convolution structure uses multi-scale dilated convolution (with expansion rates of 2, 4 and 6) to extract local details and global context features respectively, forming a hierarchical feature representation; then, the channel-spatial attention mechanism and the residual fusion layer dynamically calibrate the feature weights of each scale, suppress noise and enhance key feature responses. The ACFP module dynamically captures context information of different granularities through multi-scale receptive fields, improves the disease identification capability; with adaptive feature weighting, the fusion weight is adjusted according to the local characteristics; and through cross-scale feature interaction, the local details and global semantics are collaboratively optimized. This adaptive multi-scale fusion mechanism effectively overcomes the local limitation of traditional convolution, significantly improves the modeling ability of the model for complex scenes and the detection robustness of small-scale diseases, while keeping the calculation lightweight, thereby providing a solid technical guarantee for intelligent identification of road diseases in complex environments.

[0061] 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 is composed of a global average pooling layer, a global maximum pooling layer, a sixth concatenation 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 concatenation layer, a seventh convolutional layer and a second activation layer.

[0062] The channel-space attention structure realizes precise weighting of multi-scale features by fusing channel and space double attention mechanisms. Specifically, the channel attention mechanism adaptively enhances the expression of disease-related feature channels, and the space attention mechanism focuses on the spatial distribution of disease targets, effectively suppressing the interference response of background noise regions. This double attention guided fusion mechanism significantly improves the network's ability to recognize real disease features in complex backgrounds, especially in strong interference environments such as asphalt texture, marking, and shadow, enhancing the detection of small target diseases. This provides an effective technical approach to solving the problem of small-scale disease detection in complex background interference.

[0063] Notably, the channel attention branch aggregates each channel feature through a global average pooling layer and a global maximum pooling layer, respectively extracting global channel information from average response and extreme value response, and then enhancing the expression ability of disease-related channels through cross-channel fusion operation. The space attention branch performs average and maximum pooling on the channel dimension of the feature map through the Mean layer and the Max layer to generate two single-channel spatial description maps to capture important information in the spatial position, thereby highlighting the spatial distribution characteristics of the disease target and suppressing the interference response of the background noise region.

[0064] In the unmanned aerial vehicle aerial scene, due to the high shooting height, the road disease usually only occupies a few pixels in the image, showing the characteristics of small target size and blurred details, making it difficult to effectively extract and identify key visual features. At the same time, existing detection methods often rely on high-level semantic features when performing multi-scale feature fusion, while ignoring the rich spatial detail information in shallow features. This fusion mechanism makes the alignment and complementarity between different levels of features insufficient, causing insufficient cross-scale information fusion. Especially in complex background and low resolution conditions, the spatial perception ability of the model is limited, which seriously affects the positioning and recognition performance of the small target.

[0065] In some embodiments, the above problems in the conventional fusion mechanism are due to the lack of spatial guided fusion. Based on this research basis, m scale feature fusion modules corresponding to the backbone network and the neck network are set. Each fusion module is equipped with a spatial guided fusion (SGF) submodule, including a weighting fusion layer, a Mean-Max fusion layer, an eighth convolution layer, a third activation layer, and an enhancement layer.

[0066] The SGF sub-module dynamically adjusts the weights of different scale features through the empowerment fusion layer, and combines the Mean-Max fusion layer to aggregate the diversified information in the spatial dimension, thereby effectively improving the expression richness and spatial consistency of the features. Further, the information is refined through the convolution and activation layer, and the enhancement layer strengthens the discrimination ability of the fused features. The overall design fully utilizes the spatial guidance mechanism to realize the collaborative optimization of cross-scale details and semantics, and significantly improves the perception and recognition ability of the model for small-scale diseases.

[0067] In some embodiments, in order to further solve the problems of small road disease targets, insufficient feature expression, and insufficient multi-scale feature fusion under the perspective of the unmanned aerial vehicle, the SGF sub-module includes SGF-2 sub-modules and SGF-3 sub-modules which have the same network structure but are arranged at different positions;

[0068] The SGF-2 sub-module is arranged in each scale feature fusion module; for the ith scale, the SGF-2 sub-module is used for target extraction features corresponding to the ith scale and aligned features corresponding to the ith+1 scale; i ∈ [1, m].

[0069] The SGF-3 sub-module is arranged in the feature extraction module of the 2nd to the m-1th scale; for the jth scale, the SGF-3 sub-module is used for fusing target extraction features corresponding to the jth scale, initial fusion features corresponding to the jth scale, and 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 sub-module of the jth scale; the final fusion features are obtained based on the fusion output by the SGF-2 sub-module of the j-1th scale or the SGF-3 sub-module of the j-1th scale; j ∈ [2, m-1].

[0070] The structure design introduces the SGF sub-module based on the shallow feature pyramid level P2-P4 (corresponding to the feature map with higher resolution and rich spatial details), enhances the spatial perception ability of the model for small targets by performing more fine spatial alignment and information fusion on multi-scale features in the early stage. Compared with the traditional FPN which relies more on the fusion of deep semantic feature levels P3-P5 (lower resolution but higher semantic abstraction), the ES-FPN of the embodiment pays more attention to the utilization of detail information in the shallow features, and utilizes the information guidance in the spatial dimension to improve the complementarity of cross-level features, thereby effectively alleviating the recognition difficulty caused by the small target size and the limitation of image resolution.

[0071] In some embodiments, in view of the advantages of the YOLO series model, the road disease detection model can be improved based on the model of the YOLO series. In the road disease detection task, a lightweight YOLOv11s is selected as the basic model to significantly improve the detection accuracy and small target recognition ability while ensuring real-time. Therefore, when constructing the road disease detection model, the lightweight YOLOv11s can be preferably selected as the basic network. On this basis, the C3k2-MDDSC module, the ACFP module, and the SGF sub-module can be combined, or the SGF sub-module can be embedded into the ES-FPN structure, and the model can be improved in a single or multi-module combination manner, so as to further improve the multi-scale feature modeling capability of the model and the detection accuracy in a complex background while maintaining real-time.

[0072] Exemplarily, the network structure of the road disease detection model including the C3k2-MDDSC module, the ACFP module, and the SGF sub-module embedded into the ES-FPN structure improvement point can be referred to Figure 1 .

[0073] Based on the road disease detection model Figure 1 , for the input image to be detected, the backbone network performs multi-scale feature extraction on it through the sequentially connected GCBS-4x[GCBS+C3k2-MDDSC]-SPPF-ACFP. Among them, the GCBS module. There are four scale feature extraction modules in the backbone network, one dashed box corresponds to one, and they are numbered as the 1st, 2nd, 3rd, and 4th feature extraction modules along the data transmission direction.

[0074] Based on the feature extraction of each scale in the backbone network, the neck network is provided with four scale feature fusion modules. In the 1-4 feature fusion modules, each feature fusion module is provided with an SGF-2 sub-module; and the 3rd and 4th fusion modules are further provided with an SGF-3 sub-module. 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 the scales, align the large scale to the small scale using convolution operation; and align the small scale to the large scale using up-sampling operation.

[0075] For the target fusion features of each scale, the corresponding detection network detects them, i.e. outputs the detection results.

[0076] Among them, in the training process, in addition to the 4xreg_max distributed regression strategy, which makes the detection frame regression more detailed and stable, thereby improving the positioning accuracy, the total number of classes (nc) is also set to guide the output layer to generate a corresponding number of class prediction results, so as to ensure that the detection task can cover all target classes.

[0077] When the trained road disease detection model is put into the application stage, a redundant detection box removal (such as non-maximum suppression, NMS) and confidence filtering operation can be further introduced to eliminate overlapping detection boxes and select high-confidence prediction results, thereby improving the accuracy and reliability of the final detection results.

[0078] In this embodiment, YOLOv11s is used as the basic detection network, and various optimization modules are introduced into the backbone network and neck network to improve the accuracy and robustness of road disease detection. Specifically, they include:

[0079] GCBS module (Grouped Convolution + BatchNorm + SiLU): Through grouped convolution, the computational complexity and parameter size are reduced, and the feature decoupling capability is improved; combined with batch normalization to stabilize training, and using the smooth gradient characteristics of SiLU activation function to avoid gradient disappearance, thereby enhancing the multi-scale feature processing efficiency while ensuring the inference speed.

[0080] C3k2-MDDSC module: combined with gradient shunting and multi-scale hollow depth separable convolution strategy, the gradient transmission efficiency and feature reuse capability of deep network are improved, and efficient collaborative representation of local details and global semantics is realized, significantly enhancing the adaptability of the model to complex scenes.

[0081] ACFP module (Adaptive Contextual Feature Pyramid): placed at the end of the backbone, it captures context information at different granularities through multi-scale hollow convolution, and combines channel-spatial attention mechanism for adaptive adjustment of feature weights, realizing cross-scale detail and semantic collaborative optimization, effectively suppressing background interference such as asphalt texture and marking shadow.

[0082] SGF sub-module and ES-FPN structure based on SGF: fine spatial alignment and information fusion are performed in the shallow feature pyramid (P2-P4) stage to enhance the spatial perception ability of small targets; through weighting fusion and Mean-Max fusion mechanism, the cross-level feature complementarity is improved, thereby alleviating the recognition difficulty caused by limited resolution and scale in small target detection.

[0083] Through the organic combination and targeted optimization of the above modules, the YOLOv11 improved network of this embodiment significantly improves the accuracy and robustness of road disease detection in complex background, light change and multi-scale target conditions while maintaining lightweight and real-time.

[0084] In some embodiments, in order to ensure that the trained road disease detection model can meet the expectations and accurately understand and detect different scales of road diseases in different scenarios, a specialized dataset can be created for training the road disease detection model. Specifically, a UAV-based specialized multi-scale road disease detection dataset (UAV-SMRDD) is created, all samples of which are collected by actual aerial photography of a UAV, complete road disease features under low-altitude perspective are retained, including real scene characteristics such as light changes, multi-scale targets and complex backgrounds.

[0085] The UAV-SMRDD dataset contains a total of 5000 high-resolution aerial images, covering five typical road diseases: transverse cracks, longitudinal cracks, network cracks, potholes and ruts.

[0086] Exemplarily, to intuitively present the road disease form features under the UAV perspective, refer to Figure 2 , Figure 2 Representative sample images and their annotation information in the UAV-SMRDD dataset are shown. Through systematic data collection and annotation work, this dataset provides important benchmark data support for UAV road inspection, and its construction method meets the dataset specification requirements in the field of computer vision.

[0087] Exemplarily, in order to provide high-quality datasets, in addition to ensuring the number of images, images of multiple scenes including commercial streets, residential streets and urban-rural junction streets can be collected.

[0088] Exemplarily, the images can contain common coverings such as trees, vehicles, buildings, etc.

[0089] Exemplarily, each scene can include multiple time periods and weather conditions such as day, night, sunny, rainy and foggy, to improve the complexity and comprehensiveness of the scene.

[0090] Such a dataset can provide data support for road repair and urban management, thereby enhancing the capabilities of the urban management platform, and helping to improve the efficiency and intelligent decision-making level of urban management.

[0091] To ensure the effectiveness of model training and the reliability of evaluation results, and further to guarantee the robustness and generalization ability of the model, the 5000 images in the dataset are divided into three parts according to the ratio of 7:1:2. First, the training set containing 3500 images is used to train and optimize the road disease detection model, so that the model can learn various features and patterns of road diseases. Second, the validation set containing 500 images is designed to independently verify the detection ability and accuracy of the algorithm, to ensure the performance of the model on unseen data, and to guarantee the reliability and practicality of the research results. In addition, the test set containing 1000 images is used to evaluate the performance of the trained and validated algorithm.

[0092] Through the above strict data allocation strategy, especially in the test process close to the actual application scene, not only the training efficiency and generalization ability of the model are improved, but also the accuracy and stability of the model in dealing with unknown data are effectively verified. The setting of the test set is of key significance for testing the performance of the model in real environment, which provides an objective and fair evaluation basis for the performance of the model, and further reflects the feasibility and technical leading of the research results in practical application.

[0093] Preferably, during the process of dividing the dataset, special attention is paid to maintaining the balance of sample distribution among the three datasets, ensuring that the images in each dataset have similar statistical characteristics in terms of road disease type, scale distribution, background environment, etc. This can avoid evaluation bias caused by uneven sample distribution, making the validation results more representative and reliable.

[0094] In some embodiments, in order to comprehensively and accurately measure the performance of each version of the road disease detection model, after training at least one version of the road disease detection model based on the training set to convergence, the converged model can be verified by the validation set and 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.

[0095] Specifically, the performance of the road disease detection model on the test set can be evaluated according to preset conditions.

[0096] For example, these preset conditions can include the intersection over union, detection accuracy and recall rate of the road disease detection model, and other performance indicators.

[0097] That is, after verifying each version of the road disease detection model by the validation set, each version of the road disease detection model can be comprehensively evaluated based on the above several indicators, so as to determine the best performance road disease detection model from each version as the trained road disease detection model.

[0098] In some embodiments, the model running environment includes an Intel Xeon Platinum 8255C processor, 314 GB of memory, an NVIDIA Tesla V100 32 GB graphics card, and a CentOS 8.5.2 (64-bit) operating system. 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. The experimental setup has a batch size of 64, 300 epochs of training, an SDG optimizer, an initial learning rate of 0.01, and a cosine decay strategy for optimization.

[0099] In some embodiments, the road disease detection model training is optimized in combination with a classification loss and a regression loss. The classification loss uses a binary cross-entropy loss (BCE Loss) to determine the anchor box class; the regression loss includes a complete intersection over union loss (CIoU Loss) and a distribution focal loss (DFL loss), which are used to measure the error between the predicted bounding box and the true bounding box, respectively. The CIoU Loss is improved based on commonly used loss functions such as intersection over union (IoU), generalized intersection over union (GIoU), and distance intersection over union (DIoU). Compared with previous loss functions, it adds a penalty term for the aspect ratio, can better distinguish the error under different conditions when the center points of the predicted bounding box and the true bounding box coincide, and has scale invariance. Its formula is as follows:

[0100]

[0101] wherein, L CIoU CIoU Loss, b and represent the center points of the true box and the predicted box, respectively; p is the Euclidean distance between the predicted bounding box and the true bounding box; c is the distance between the opposite diagonal lines of the closed area of the predicted bounding box and the true bounding box; v is the consistency of the relative proportion of the predicted bounding box and the true bounding box, IoU is the intersection over union of the predicted bounding box and the true bounding box; α is a weight coefficient; w and These are the widths of the predicted border and the actual border, respectively; h and These represent the heights of the predicted bounding box and the actual bounding box, respectively.

[0102] Based on CIoU Loss, the introduction of DFL loss can further improve the regression accuracy of bounding boxes. DFL loss discretizes the coordinates to address the uncertainty in bounding box coordinate prediction, making the regression process more refined and robust, and effectively reducing prediction errors. The formula for DFL loss is as follows:

[0103]

[0104] in, S i The cross-entropy loss is the difference between the true and predicted bounding boxes on the left. S i+1 The cross-entropy loss is calculated between the true bounding box and the predicted bounding box on the right.

[0105] However, although CIoU loss takes into account three important factors of localization loss: overlap area, center point distance, and aspect ratio, the CIoU loss formula has α... v However, there are still problems with the design of this item, which has slowed down the convergence speed.

[0106] To improve convergence speed, a new IoU loss, Fast Quality IoU Loss (FQ-IoU Loss), is proposed based on the penalty term of the original CIoU loss. FQ-IoU Loss consists of four components: overlap loss, center distance loss, width-height loss, and angle loss.

[0107] The core innovation of FQ-IoU Loss lies in optimizing bounding box rotation alignment by introducing an angle penalty term, thereby significantly improving the model's convergence speed. Simultaneously, this method abandons the traditional joint aspect ratio penalty mechanism, instead decoupling the aspect ratio into independent width and height penalty terms. This decoupling design allows the model to finely adjust prediction errors in both the width and height directions separately, avoiding the dimensional coupling bias that might arise from a single aspect ratio penalty. Through this dual-dimensional independent optimization strategy, FQ-IoU Loss can more accurately capture localization errors in different directions, thus significantly improving bounding box localization accuracy while maintaining angular consistency.

[0108] Introducing the form of exponential functions The IoU is enhanced by a nonlinear mapping, which improves the sensitivity and adaptability of the loss function. It assigns high gradients to small deviations (e.g., small target shifts), accelerating convergence, while smoothing large deviations to avoid gradient explosion and improve optimization stability. Meanwhile, the exponential transformation compresses the distance to the range (0, 1], ensuring numerical stability and strengthening the optimization of geometric features through exponential difference ratios (e.g. ). In addition, combining trigonometric functions (e.g., the Λ term) explicitly handles angular symmetry, making the detection more robust to rotation and scale changes. Finally, this design dynamically balances multi-objective weights, improving precision while maintaining fast convergence, making it particularly suitable for complex scene detection tasks.

[0109] Specifically, the formula of FQ-IoU Loss is as follows:

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116] wherein, is the fast and high-quality intersection over union loss; IoU is the intersection over union between the predicted box and the real box; L IoU is the intersection over union loss; L dis is the center distance loss; L shp is the width-height loss; L ang is the angle loss; Δ is the center distance loss multiplied by 2; b and represent the center point of the real box and the center point of the predicted box, respectively; and represent the square of the Euclidean distance between the center points of the real box and the predicted box, respectively; and w represent the width and height of the minimum enclosing rectangle, respectively; Ω is the width-height loss multiplied by 2; and represent the width of the predicted box and the real box, respectively; h represents the square of the Euclidean distance between the width of the real box and the predicted box; and square of the Euclidean distance between the real box and the predicted box height; is 2 times the angle loss; and respectively represent the width and height of the rectangular box constructed by the center points of the real box and the predicted box.

[0117] Therefore, the formula of the total loss Loss is as follows:

[0118]

[0119] wherein λ1 and λ2 are balance coefficients.

[0120] Based on the total loss function, the road disease detection model described in any of the preceding embodiments is trained, which can improve the convergence speed of the model and thus improve the overall training efficiency. By continuously iterating and optimizing the road disease detection model, the disease recognition effect of the model in a complex scene is more optimal and tends to be stable. Finally, a version with higher robustness can be selected from the converged multiple model versions as the final trained model.

[0121] Based on the network structure of the road disease detection model of the preceding embodiments, the present application proposes a road disease detection method based on a UAV.

[0122] The road disease detection method based on a UAV provided by 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, notebook computers, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. The embodiments of the present application do not make any limitation on the specific type of electronic device.

[0123] In order to illustrate the technical solutions proposed by the present application, the electronic device will be taken as the execution subject to describe each embodiment.

[0124] Figure 3 A schematic flowchart of the road disease detection method based on a UAV provided by the present application is shown, which includes:

[0125] In step 310, the electronic device performs feature extraction on the to-be-detected image collected under the perspective of the UAV based on the backbone network of the pre-trained road disease detection model, to obtain image features.

[0126] In step 320, the electronic device fuses the image features based on the neck network of the road disease detection model to obtain target fusion features.

[0127] In step 330, the electronic device detects the target fusion features based on the detection network of the road disease detection model to obtain a detection result of the road disease in the to-be-detected image.

[0128] The backbone network includes m scale feature extraction modules, 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, and for the first input feature input into the C3k2-MDDSC submodule:

[0129] In step A11, the electronic device performs a channel compression operation on the first input feature through the first CBS layer to obtain first compressed features. The electronic device performs channel compression through the first CBS layer (convolution+BN+activation) to reduce the number of feature channels, reduce the amount of calculation while retaining the main information, obtain first compressed features, and further provide efficient feature input for subsequent processing.

[0130] In step A12, the electronic device divides the first compressed features into first and second segmented features through the segmentation layer.

[0131] For the first compressed features, the electronic device divides the compressed features into two parts (first and second segmented features) through the segmentation layer, creates conditions for multi-path feature processing, and realizes differential extraction of information.

[0132] In step A13, the electronic device performs a multi-level progressive extraction operation on the first segmented features through the multi-level progressive stacking extraction layer to obtain level features corresponding to each level.

[0133] The electronic device performs a multi-level progressive feature extraction operation on the first segmented features through the multi-level progressive stacking extraction layer. It is assumed that the extraction layer includes n levels, i.e., it is composed of n MDDSC or C3k-MDDSC submodules connected in series.

[0134] The first level directly takes the first segmented features as input to complete preliminary feature enhancement; each level from the n-1th to the nth level takes the output of the previous level as input to gradually deepen the feature representation; while the output is transmitted layer by layer, the output of each level is also directly sent to the first splicing layer to participate in fusion, realizing cross-level feature aggregation.

[0135] The multi-level progressive stacking extraction layer effectively alleviates the gradient vanishing problem of deep network by gradient propagation of multi-path, and significantly improves the feature reuse rate through multi-branch structure, so that the network can more efficiently capture and learn complex patterns and multi-scale information under the premise of ensuring computational efficiency.

[0136] Step A14, the electronic device splices the first split feature, the second split feature and the hierarchical features through the first splicing layer to obtain the first splicing feature.

[0137] Step A15, 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.

[0138] Subsequently, the electronic device fuses the first split feature, the second split feature and the hierarchical output features through the first splicing layer to obtain the first splicing feature; finally, a channel expansion operation is performed on the first splicing feature through the second CBS layer to obtain the first output feature corresponding to the first input feature.

[0139] In this embodiment, not only is the gradual deep enhancement of the input feature and the cross-level information fusion achieved, but also the multi-scale, multi-path and attention guiding mechanism is introduced in the early stage of feature extraction, effectively improving the feature expression and discriminability of the network; while maintaining the computational efficiency, the modeling ability of the model for complex patterns is significantly enhanced, thereby providing higher quality feature support for subsequent detection tasks.

[0140] In some embodiments, Figure 4 and Figure 3 respectively show the specific network structure of the C3k2-MDDSC submodule when C3k takes the value of False and True.

[0141] In some embodiments, the C3k-MDDSC includes a third CBS layer, a series of double MDDSCs, a second splicing layer and a fourth CBS layer, and for the second input feature input into the C3k-MDDSC:

[0142] Step B11, the electronic device performs a feature compression operation on the second input feature through the third CBS layer to obtain the second compressed feature.

[0143] Step B12, the electronic device performs multi-scale feature extraction on the second compressed feature through the series of double MDDSCs to obtain the multi-scale feature.

[0144] Step B13, the electronic device performs a splicing operation on the second compressed feature and the multi-scale feature through the second splicing layer to obtain the second splicing feature.

[0145] Step B14, 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.

[0146] 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; then, performs multi-scale feature extraction on the second compressed feature via a double MDDSC structure in series, which utilizes convolution combinations with different receptive fields and hole rates to simultaneously 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 and 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.

[0147] It can be understood that when C3k takes the value True, the second output feature corresponding to the i-th level is the level feature of the i-th level.

[0148] Exemplarily, referring to Figure 6 , Figure 6 A network structure schematic diagram of C3k-MDDSC is shown.

[0149] In this embodiment, through the progressive design of compression, multi-scale extraction and fusion, the diversity and semantic level richness of the features are significantly improved while reducing the computational amount; especially, the introduction of the double MDDSC structure enables the network to retain fine texture and global context at the same stage, providing high-quality feature input with both detail sensitivity and semantic expressiveness for the subsequent detection network.

[0150] In some embodiments, the MDDSC includes a first convolutional layer, a multi-scale hollow depth separable convolutional layer, a third splicing layer, a second convolutional layer, and a fourth splicing layer; for a third input feature input into the MDDSC:

[0151] 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.

[0152] Step C12, the electronic device performs at least two scales of hollow depth separable convolution operations on the third compressed feature through the multi-scale hollow depth separable convolutional layer to obtain hollow depth separable convolution features corresponding to each scale.

[0153] Step C13, the third splicing layer performs a splicing operation on the hollow depth separable convolution features of each scale to obtain a third spliced feature.

[0154] Step C14: Perform channel expansion operation on the third concatenated feature through the second convolutional layer; to obtain the expanded feature.

[0155] Step C15: Perform a concatenation operation on the third input feature and the extended feature through the fourth concatenation layer to obtain the third output feature corresponding to the third input feature; the third output feature is the hierarchical feature of the corresponding MDDSC level.

[0156] The electronic device first performs channel compression on the third input feature through a first convolutional layer to reduce the number of feature channels, lower the computational burden, and remove redundant information, thus obtaining the third compressed feature. Then, it performs at least two scales of dilatational convolution on the third compressed feature through a multi-scale dilatational convolutional layer. Different scales correspond to different dilatation rates, thereby capturing local texture details and global contextual information under different receptive fields at the same stage, resulting in dilatational convolutional features at each scale. Next, the electronic device uses a third concatenation layer to concatenate the features at each scale, fusing them into a third concatenated feature. Then, a second convolutional layer performs channel expansion on the third concatenated feature to restore the number of feature channels and enhance feature expressiveness, resulting in an expanded feature. Finally, the electronic device uses a fourth concatenation layer to concatenate the original third input feature and the expanded feature, forming the third output feature. This third output feature corresponds to the hierarchical feature of its MDDSC level and will enter the subsequent multi-level feature fusion process.

[0157] In some embodiments, Figure 4 A schematic diagram of the MDDSC network structure is shown. Based on... Figure 4 The network structure is designed for the third input feature. Third output feature , B It refers to the batch size. C 1 represents the number of input channels. C 2 represents the number of output channels. H It's about height. W It's the width. The electronic device first uses a 1×1 ordinary convolution (the first convolutional layer) to... X Halving the number of channels yields the first compression feature. ,in , e For channel compression ratio ( The specific expression is as follows:

[0158]

[0159] Then, for each void ratio , and sequentially performs a dilated depthwise separable convolution (DDSConv). The DDSConv features corresponding to each scale are extracted :

[0160]

[0161] The DDSConv includes two parts:

[0162] (a) a dilated depthwise convolution (DDWConv)

[0163] The depthwise convolution independently performs a dilated convolution on each channel of the feature map:

[0164]

[0165] (b) a pointwise convolution (PWConv)

[0166] The number of channels is then expanded to C 2:

[0167]

[0168] Next, the feature maps corresponding to all dilated rates are spliced along the channel dimension to obtain :

[0169]

[0170] Finally, the multi-scale features are integrated through a 1x1 convolution, and the final output is generated through a residual connection :

[0171]

[0172] In the C3k2-MSDDC module, the gradient shunting mechanism and the multi-scale feature fusion strategy are cooperatively optimized. By constructing a multi-branch parallel structure and introducing deep separable convolution operations with different receptive fields, multi-level representation and interactive fusion of local texture features and global context information are achieved, thereby significantly enhancing the precision performance of the model in road small-scale disease detection. While maintaining the spatial resolution of the feature map, the lightweight characteristics of the deep separable convolution are utilized to fully extract multi-scale features with low computational complexity, effectively alleviating the problem of easy loss of micro-disease features in complex road scenes.

[0173] In some embodiments, an ACFP module is arranged at the end of the backbone network, and the ACFP module comprises a third convolutional layer, a multi-scale hollow deep convolutional structure, a fifth splicing layer, a fourth convolutional layer, a channel-space attention structure, and a residual fusion layer; for the fourth input feature input into the ACFP module:

[0174] 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.

[0175] 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 operations, thereby obtaining a fourth compressed feature.

[0176] D12, the electronic device performs a hollow deep convolution operation and a nonlinear transformation operation on the fourth compressed feature through the multi-scale hollow deep convolutional structure under different scales to obtain a deep convolution feature corresponding to each scale.

[0177] The electronic device performs a hollow deep convolution operation on the fourth compressed feature under different hollow rates through the multi-scale hollow deep convolutional structure, and combines a nonlinear transformation to enhance the expression nonlinear capability of the feature, thereby obtaining a deep convolution feature corresponding to each scale, and realizing feature capture of different receptive fields.

[0178] D13, the electronic device performs a splicing operation on the fourth compressed feature and the deep convolution features of each scale through the fifth splicing layer to obtain a multi-scale splicing feature.

[0179] The electronic device splices the fourth compressed feature and the deep convolution features of each scale in the channel dimension through the fifth splicing layer to obtain a multi-scale splicing feature, thereby realizing fusion of local and global information.

[0180] D14, the electronic device performs a channel expansion operation on the multi-scale splicing feature through the fourth convolutional layer.

[0181] D15, the electronic device respectively performs a spatial attention operation and a channel attention operation on the multi-scale splicing feature after channel expansion through the channel-space attention structure, and fuses the obtained first spatial enhanced feature and channel enhanced feature to obtain a spatial-channel enhanced feature.

[0182] The electronic device performs a channel expansion operation on the multi-scale splicing feature through the fourth convolutional layer to restore to a channel number matching the original feature. The electronic device respectively performs a spatial attention operation and a channel attention operation on the expanded multi-scale splicing feature through the channel-space attention structure to strengthen the response of key positions and key channels; and then fuses the obtained first spatial enhanced feature and channel enhanced feature to generate a spatial-channel enhanced feature.

[0183] D16, the electronic device enhances the multi-scale spliced features after channel expansion and the spatial-channel enhanced features through a residual fusion layer to obtain fourth output features corresponding to the fourth input features.

[0184] The electronic device element-wise fuses the multi-scale spliced features after channel expansion and the spatial-channel enhanced features through a residual fusion layer to realize feature enhancement and efficient gradient transmission, and obtains fourth output features corresponding to the fourth input features.

[0185] In the embodiment, the embodiment optimizes multiple links such as multi-scale feature extraction, channel-spatial attention guidance, and residual fusion. On the one hand, the multi-scale hollow depth 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 representation ability for targets with large size changes and subtle diseases. On the other hand, the channel-spatial attention structure effectively suppresses irrelevant responses in complex background regions and enhances the saliency of disease region features by combining the advantages of channel weight adjustment and spatial position focusing. Finally, the residual fusion layer not only ensures efficient transmission of key information and alleviates the gradient attenuation problem in deep feature learning, but also enhances the overall detection accuracy and robustness through complementary feature enhancement, thereby realizing stable and high-precision detection of small-scale disease targets in complex road environments.

[0186] In some embodiments, the channel-spatial attention structure includes a channel attention branch, a spatial attention branch, and an enhancement fusion layer. The channel attention branch includes a global average pooling layer, a global maximum pooling layer, a sixth splicing 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 splicing layer, a seventh convolution layer, and a second activation layer. The foregoing step D15 specifically includes:

[0187] 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 the global average pooling layer and the global maximum pooling layer, respectively, to obtain average pooling features and maximum pooling features; performs splicing operations on the average pooling features and the maximum pooling features through the sixth splicing layer to obtain fused pooling features. The fused pooling features are sequentially subjected to channel compression, nonlinear transformation, channel expansion, and activation through the fifth convolution layer, the nonlinear transformation layer, the sixth convolution layer, and the first activation layer to obtain channel enhanced features.

[0188] The channel attention branch aims to adaptively adjust the 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 respectively perform statistical aggregation on the multi-scale spliced features after channel expansion, extracting different types of channel information representations from a global range: 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 fusion pooling feature, ensuring information complementarity. The fifth convolutional layer performs channel compression to reduce dimensionality, reduce computational complexity, and aggregate information; the nonlinear transformation layer improves the feature expression capability; and the sixth convolutional layer performs channel expansion to restore the original number of channels; the first activation layer introduces a nonlinear mapping to obtain the final channel enhanced feature.

[0189] In step D152, for the spatial attention branch: the electronic device performs Mean operation and Max operation on the multi-scale spliced features after channel expansion through the Mean layer and the Max layer respectively to obtain average fusion features and maximum fusion features; splices the average fusion features and the maximum fusion features through the seventh splicing layer to obtain target fusion features; and performs convolution operation and activation operation on the target fusion features through the seventh convolutional layer and the second activation layer in sequence to obtain the first spatial enhanced feature.

[0190] The spatial attention branch aims to locate the key information region from the spatial dimension, so that the network focuses on the spatial position of the disease target. Among them, the Mean layer and the Max layer respectively project the feature map in the channel dimension to obtain the average fusion feature and the maximum fusion feature: the former retains the overall spatial distribution information, and the latter highlights the significant spatial position. The seventh splicing layer splices the two types of features to form a target fusion feature with better discriminability. 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 enhanced feature.

[0191] In this embodiment, the cooperative design of the two branches realizes double attention guidance in the channel dimension and the spatial dimension: the channel attention branch ensures that the network adaptively emphasizes key information related to disease detection in feature channel allocation, and the spatial attention branch makes the model focus on the disease target area in the spatial domain of the feature map, significantly suppressing background interference. The combination of the 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 keeping the model lightweight.

[0192] In some embodiments, Figure 8 The network structure schematic diagram of the ACFP module is shown. Based on the network structure of Figure 5 , for the fourth input feature , first, a 1x1 ordinary convolution (third convolutional layer) is performed onX channel compression to obtain a second compressed feature wherein, , is a scaling factor, and takes a value of 0.5.

[0193]

[0194] Then, a cavity depth convolution of different cavity rates is sequentially performed, and the convolution results of each scale are activated by a ReLU activation function to obtain Then, a cavity depth convolution of different cavity rates is sequentially performed, and the convolution results of each scale are activated by a ReLU activation function to obtain

[0195]

[0196] After the multi-scale spliced features of each scale are spliced with and , a multi-scale spliced feature is obtained:

[0197]

[0198] Then, the multi-scale spliced features are subjected to channel adjustment by a 1x1 ordinary convolution to obtain multi-scale spliced features after channel expansion:

[0199]

[0200] Finally, the multi-scale spliced features are respectively input into a channel attention branch and a spatial attention branch for channel-spatial attention enhancement, so as to more fully promote the fusion of multi-scale features to obtain a fourth output feature: Z

[0201]

[0202]

[0203]

[0204] ​​​​In this embodiment, the ACFP module adopts a multi-branch hollow convolution structure to realize efficient aggregation of multi-scale context information through differential receptive field design. Based on this, a channel-spatial dual attention mechanism is introduced to strengthen key information expression: among them, the channel attention mechanism is based on the correlation analysis between feature channels to adaptively improve the feature channel weight highly related to the disease target; the spatial attention mechanism focuses on the spatial position distribution of the disease target, effectively suppressing the interference response of the background noise area. The design of multi-scale feature fusion and dual attention guided by the network can accurately distinguish real disease features from complex background information in the strong interference environment of road surface asphalt texture, marking shadow, etc., thereby significantly improving the detection accuracy and robustness of small-scale disease targets, and providing an efficient technical solution to the problem of road disease detection under complex background interference.

[0205] In some embodiments, corresponding to the m feature extraction modules, the neck network is correspondingly 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 referred to as Figure 9 . The SGF submodule includes a weighted fusion layer, a Mean-Max fusion layer, an eighth convolution layer, a third activation layer, and an enhancement layer; for at least two fifth input features input into the SGF submodule:

[0206] Step E11, the electronic device performs a weighting operation on each fifth feature through the weighted fusion layer, and splices the weighted fifth features to obtain a fused and weighted feature.

[0207] The electronic device performs a weight distribution operation on each fifth input feature through the weighted fusion layer to highlight the key information channel, and splices the weighted fifth input features to obtain a fused and weighted feature.

[0208] Step E12, the electronic device performs Mean operation and Max operation on the fused and weighted feature through the Mean-Max fusion layer, and splices the fused and weighted feature after the Mean operation with the fused and weighted feature after the Max operation to obtain a Mean-Max fusion feature.

[0209] The electronic device performs a Mean operation and a Max operation on the fused and weighted feature through the Mean-Max fusion layer to obtain feature expression under different statistical characteristics, and splices the two to obtain a Mean-Max fusion feature.

[0210] Step E13, the electronic device performs convolution operation and activation operation on the Mean-Max fusion feature through the eighth convolution layer and the third activation layer in sequence to obtain a second spatial enhancement feature.

[0211] The electronic device sequentially performs convolution operation and non-linear activation operation on the Mean-Max fusion feature through the eighth convolutional layer and the third activation layer to obtain the second spatial enhancement feature.

[0212] In step E14, the electronic device fuses the second spatial enhancement feature and the fusion weighting feature through the enhancement layer to obtain the fifth output feature corresponding to the fifth input feature.

[0213] The electronic device fuses the second spatial enhancement feature and the fusion weighting feature through the enhancement layer to form a comprehensive feature representation with local response enhancement and global context preservation, thereby obtaining the fifth output feature corresponding to the fifth input feature.

[0214] In this embodiment, the SGF sub-module weights the multi-scale input features through the weighting fusion layer to dynamically adjust the importance of different features, thereby achieving effective screening and highlighting of information. Subsequently, the Mean-Max fusion layer is used to extract the mean and maximum value responses of the features, respectively, to capture spatial information under different statistical characteristics and enhance the richness and discriminability of feature expression. The convolution and activation layers are further used to refine the fusion features and improve the spatial semantic expression capability. Finally, the enhancement layer combines the spatial enhancement feature and the original fusion weighting feature to achieve collaborative optimization of local details and global context. This design significantly improves the spatial perception capability and detail capture capability of the model for small targets, effectively alleviates the information loss and noise interference problems in traditional multi-scale fusion, thereby improving the precision and robustness of small-scale road disease detection and ensuring the stable performance of the model in complex environments.

[0215] The SGF sub-module includes SGF-2 sub-modules and SGF-3 sub-modules that have the same network structure but are arranged at different positions;

[0216] The SGF-2 sub-module is arranged in each scale feature fusion module; for the i-th scale, the SGF-2 sub-module is used for the target extraction feature corresponding to the i-th scale and the aligned feature corresponding to the i+1-th scale; i∈[1,m];

[0217] The SGF-3 sub-module is arranged in the feature extraction module of the 2nd to (m-1)th scale; for the j-th scale, the SGF-3 sub-module is used for the target extraction feature corresponding to the j-th scale, the initial fusion feature corresponding to the j-th scale, and the final fusion feature corresponding to the j-1-th scale; the initial fusion feature is obtained based on the feature fusion output by the SGF-2 sub-module of the j-th scale; the final fusion feature is obtained based on the fusion output by the SGF-2 sub-module of the j-1-th scale or the SGF-3 sub-module of the j-1-th scale; j∈[2,m-1].

[0218] Exemplarily, Figure 10A network structure diagram of the ES-FPN designed based on the SGF-2 submodule and the SGF-3 submodule is shown.

[0219] In some embodiments, for the unmanned aerial vehicle road disease detection task, based on the YOLOv11s architecture, a plurality of key improvements such as the C3k2-MDDSC submodule, the ACFP module, and the SGF submodule (or the SGF-based ES-FPN structure) are systematically introduced, and the model performance is significantly improved.

[0220] Firstly, the C3k2-MDDSC module enhances the multi-level representation ability of the model to local texture details and global context through the combination of gradient shunting and multi-scale depth separable convolution, effectively alleviates the gradient vanishing problem of deep network, and reduces the risk of overfitting through the jump connection and the group convolution (GCBS) in the C3k2-MDDSC module, ensures the lightweight while improving the feature reuse rate and generalization performance. Figure 1

[0221] Secondly, the ACFP module adopts a multi-branch hollow convolution structure at the end of the backbone network, realizes multi-scale context information aggregation through differential receptive field, and fuses channel-spatial dual attention mechanism to dynamically adjust the weight of different feature channels and spatial positions, significantly enhances the recognition ability of small-scale diseases in complex background, especially in strong interference environment such as asphalt texture and marking shadow.

[0222] Finally, the SGF submodule is responsible for the weighted fusion and spatial alignment of multi-scale features in the neck network, highlights key information channels through the weighted fusion layer, captures spatial features with different statistical characteristics through the Mean-Max fusion layer, and realizes the collaborative optimization of details and context through the convolution activation and enhancement layer, greatly improves the spatial perception and detail capture ability of the model to small disease targets, and alleviates the problem of information loss and noise interference in traditional multi-scale fusion.

[0223] In summary, the synergistic effect of each module not only improves the detection accuracy and robustness of the model to small-scale road diseases, but also enhances the suppression ability of various background noises in complex scenes, optimizes the training efficiency and inference speed of the model, and ensures the stability and practicality of intelligent recognition of road diseases from the perspective of unmanned aerial vehicles, providing strong technical support for practical applications.

[0224] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0225] The unmanned aerial vehicle-based road disease detection method corresponding to the above embodiments, Figure 11 ​A structural block diagram of a road disease detection device 11 provided by an embodiment of the present application is shown, and only parts related to the embodiment of the present application are shown for ease of description.

[0226] With reference to Figure 11 The road disease detection device 11 comprises:

[0227] An extraction module 111 is configured to perform feature extraction on a to-be-detected image collected by a UAV from a perspective of the UAV based on a backbone network of a pre-trained road disease detection model, to obtain image features; the to-be-detected image comprises a road.

[0228] A fusion module 112 is configured to perform fusion on the image features based on a neck network of the road disease detection model, to obtain target fusion features.

[0229] A detection module 113 is configured to perform detection on the target fusion features based on a detection network of the road disease detection model, to obtain a detection result of a road disease in the to-be-detected image.

[0230] The backbone network comprises m scale feature extraction modules, m is a positive integer, and m≥2; each feature extraction module is provided with a C3k2-MDDSC submodule; the C3k2-MDDSC submodule comprises a first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer, and a second CBS layer; for a first input feature input into the C3k2-MDDSC submodule:

[0231] The first CBS layer is configured to perform a channel compression operation on the first input feature, to obtain a first compressed feature;

[0232] The segmentation layer is configured to segment the first compressed feature into a first segmented feature and a second segmented feature;

[0233] The multi-level progressive stacking extraction layer is configured to perform a multi-level progressive extraction operation on the first segmented feature, to obtain a level feature corresponding to each level;

[0234] The first splicing layer is configured to splice the first segmented feature, the second segmented feature, and the level features, to obtain a first spliced feature;

[0235] The second CBS layer is configured to perform a channel expansion operation on the first spliced feature, to obtain a first output feature corresponding to the first input feature;

[0236] Each feature extraction module for extracting a shallow layer feature has a C3k in the corresponding C3k2-MDDSC set as False, and the multi-level progressive stacking extraction layer is obtained based on an improvement of a MDDSC; each feature extraction module for extracting a deep layer feature has a C3k in the corresponding C3k2-MDDSC set as True, and the multi-level progressive stacking extraction layer is obtained based on an improvement of a C3k-MDDSC.

[0237] 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 embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0238] Figure 12 This is a schematic diagram of the physical layer structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 12 of this embodiment includes: at least one processor 120 ( Figure 12 The diagram shows only one processor, memory 121, and a computer program 122 stored in memory 121 that can run on at least one processor 120. When processor 120 executes computer program 122, it implements the steps in any of the above embodiments of the road defect detection method based on UAVs, for example... Figure 3 Steps 310-330 are shown.

[0239] The processor 120 can be a Central Processing Unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0240] In some embodiments, memory 121 may be an internal storage unit of electronic device 12, such as a hard disk or memory of electronic device 12. In other embodiments, memory 121 may also be an external storage device of electronic device 12, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 12.

[0241] Furthermore, the memory 121 may include both internal storage units and external storage devices of the electronic device 12. The memory 121 is used to store operating devices, application programs, boot loaders, data, and other programs, such as program code for computer programs. The memory 121 can also be used to temporarily store data that has been output or will be output.

[0242] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the above-mentioned device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or software functional unit. In addition, the specific name of each functional unit and module is only for the convenience of mutual distinction, and does not limit the protection scope of the present application. The specific working process of the unit and module in the above system can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0243] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to realize the steps in the above-mentioned various method embodiments.

[0244] The embodiment of the present application provides a computer program product. When the computer program product is run on an electronic device, the electronic device is caused to execute the steps in the above-mentioned various method embodiments.

[0245] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the present application realizes all or part of the processes in the above-mentioned embodiment methods, which can be completed by a computer program instructing related hardware. The above-mentioned computer program can be stored in a computer readable storage medium. The computer program is executed by a processor to realize the steps in the above-mentioned various method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium at least includes any entity or device capable of carrying the computer program code to the photographing device / electronic device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium. For example, U disk, mobile hard disk, magnetic disk or optical disk, etc.

[0246] In the above embodiments, the description of each embodiment focuses on different aspects, and the parts not described in detail or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0247] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0248] In the embodiments provided in the present application, it should be understood that the disclosed apparatus / network device and method can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely schematic, for example, the division of the above modules or units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be through some interface, indirect coupling or communication connection between the units or components, which can be electrical, mechanical or other forms.

[0249] The units described above as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0250] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for detecting road disease based on a UAV, characterized in that, The application relates to a road disease detection method based on a pre-trained road disease detection model. The method comprises the following steps: An image feature is extracted from a to-be-detected image collected by a UAV through a backbone network of the road disease detection model; The to-be-detected image comprises a road; A target fusion feature is obtained by fusing the image feature through a neck network of the road disease detection model; A detection result of a road disease in the to-be-detected image is obtained by detecting the target fusion feature through a detection network of the road disease detection model; The backbone network comprises m feature extraction modules, wherein m is a positive integer and m>=2; each feature extraction module is provided with a C3k2-MDDSC submodule; the C3k2-MDDSC submodule comprises a first CBS layer, a segmentation layer, a multi-level progressive stacking extraction layer, a first splicing layer and a second CBS layer; for a first input feature input into the C3k2-MDDSC submodule: The first input feature is subjected to a channel compression operation through the first CBS layer to obtain a first compressed feature; The first compressed feature is segmented into a first segmentation feature and a second segmentation feature through the segmentation layer; The first segmentation feature is subjected to a multi-level progressive extraction operation through the multi-level progressive stacking extraction layer to obtain a level feature corresponding to each level; The first segmentation feature, the second segmentation feature and the level features are spliced through the first splicing layer to obtain a first splicing feature; The first splicing feature is subjected to a channel expansion operation through the second CBS layer to obtain a first output feature corresponding to the first input feature; For each feature extraction module for extracting a shallow layer feature, the C3k in the corresponding C3k2-MDDSC is set as False, and the multi-level progressive stacking extraction layer is obtained based on MDDSC improvement; for each feature extraction module for extracting a deep layer feature, the C3k in the corresponding C3k2-MDDSC is set as True, and the multi-level progressive stacking extraction layer is obtained based on C3k-MDDSC improvement; The C3k-MDDSC comprises a third CBS layer, a series-connected double MDDSC, a second splicing layer and a fourth CBS layer; for a second input feature input into the C3k-MDDSC: The second input feature is subjected to a feature compression operation through the third CBS layer to obtain a second compressed feature; The second compressed feature is subjected to multi-scale feature extraction through the series-connected double MDDSC to obtain a multi-scale feature; The second compressed feature and the multi-scale feature are subjected to a splicing operation through the second splicing layer to obtain a second splicing feature; The second splicing feature is subjected to a channel expansion operation through the fourth CBS layer to obtain a second output feature corresponding to the second input feature; the second output feature is a level feature of a level corresponding to the C3k-MDDSC; The MDDSC comprises a first convolution layer, a multi-scale hollow depth separable convolution layer, a third splicing layer, a second convolution layer and a fourth splicing layer; for a third input feature input into 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 the hole depth separable convolution operation on the third compressed feature through the multi-scale hole depth separable convolutional layer to obtain the hole depth separable convolution feature corresponding to each scale; the hole depth separable convolution operation of different scales corresponds to different hole rates; performing a splicing operation on the hole depth separable convolution feature of each scale through the third splicing layer to obtain a third splicing feature; performing a channel expansion operation on the third splicing feature through the second convolutional layer to obtain an expansion feature; performing a splicing operation on the third input feature and the expansion 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 corresponding to the layer level of the MDDSC.

2. The road disease detection method according to claim 1, characterized by, The end of the backbone network is provided with an ACFP module, and 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-space attention structure, and a residual fusion layer; for a fourth input feature input into 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 hole depth convolution operation and a nonlinear transformation operation under different scales on the fourth compressed feature through the multi-scale hole depth convolution structure to obtain a depth convolution feature corresponding to each scale; performing a splicing operation on the fourth compressed feature and the depth convolution feature 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 a spatial attention operation and a channel attention operation on the multi-scale splicing feature after the channel expansion through the channel-space attention structure respectively, and fusing the obtained first spatial enhancement feature and channel enhancement feature to obtain a space-channel enhancement feature; enhancing the multi-scale splicing feature after the channel expansion and the space-channel enhancement feature through the residual fusion layer to obtain a fourth output feature corresponding to the fourth input feature.

3. The road disease detection method according to claim 2, characterized by, The channel-space attention structure includes a channel attention branch, a spatial attention branch, and an enhancement 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 channel-space attention structure includes a channel attention branch, a spatial attention branch, and an enhancement 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; For the channel attention branch: performing an average pooling operation and a maximum pooling operation on the multi-scale splicing feature after the channel expansion through the global average pooling layer and the global maximum pooling layer respectively to obtain an average pooling feature and a maximum pooling feature; The sixth splicing layer is used for performing a splicing operation on the average pooling feature and the maximum pooling feature, so as to obtain a fusion pooling feature; The fifth convolution layer, the nonlinear transformation layer, the sixth convolution layer and the first activation layer are used for sequentially performing a channel compression operation, a nonlinear transformation operation, a channel expansion operation and an activation operation on the fusion pooling feature, so as to obtain a channel enhancement feature; For the spatial attention branch: The Mean layer and the Max layer are used for respectively performing a Mean operation and a Max operation on the multi-scale splicing feature after the channel expansion, so as to obtain an average fusion feature and a maximum fusion feature; The seventh splicing layer is used for splicing the average fusion feature and the maximum fusion feature, so as to obtain a target fusion feature; The seventh convolution layer and the second activation layer are used for sequentially performing a convolution operation and an activation operation on the target fusion feature, so as to obtain the first spatial enhancement feature.

4. The road disease detection method according to any one of claims 1 to 3, characterized by, Corresponding to m feature extraction modules, the neck network is correspondingly provided with m scale feature fusion modules; each scale feature fusion module is provided with an SGF submodule, and the SGF submodule includes a weighting fusion layer, a Mean-Max fusion layer, an eighth convolution layer, a third activation layer and an enhancement layer; for at least two fifth input features input into the SGF submodule: The weighting fusion layer is used for performing a weighting operation on each fifth input feature, and the weighted fifth input features are spliced, so as to obtain a fusion weighted feature; The Mean-Max fusion layer is used for respectively performing a Mean operation and a Max operation on the fusion weighted feature, and the fusion weighted feature after the Mean operation and the fusion weighted feature after the Max operation are spliced, so as to obtain a Mean-Max fusion feature; The eighth convolution layer and the third activation layer are used for sequentially performing a convolution operation and an activation operation on the Mean-Max fusion feature, so as to obtain a second spatial enhancement feature; The enhancement layer is used for fusing the second spatial enhancement feature and the fusion weighted feature, so as to obtain a fifth output feature corresponding to the fifth input feature.

5. The road disease detection method according to claim 4, characterized by, The SGF submodule includes an SGF-2 submodule and an SGF-3 submodule which have the same network structure but different positions; The SGF-2 submodule is arranged in each scale feature fusion module; for the i-th scale, the SGF-2 submodule is used for a target extraction feature corresponding to the i-th scale and an aligned feature corresponding to the i+1-th scale; i is an integer in the range of 1 to m; The SGF-3 submodule is arranged in the feature extraction module of the 2nd to the m-1th scale; for the j-th scale, the SGF-3 submodule is used for fusing a target extraction feature corresponding to the j-th scale, an initial fusion feature corresponding to the j-th scale and a final fusion feature corresponding to the j-1th scale; The initial fusion feature is obtained based on a feature fusion output by the SGF-2 submodule of the j-th scale; The final fusion feature is obtained based on fusion of outputs of the SGF-2 submodule of the j-1th scale or the SGF-3 submodule of the j-1th scale; The j is in [2, m-1].

6. The road disease detection method according to any one of claims 1 to 3, characterized by, The road disease detection model is trained based on a regression loss and a classification loss; the regression loss includes a fast quality intersection-over-union loss, and a formula is as follows: Among them, the For fast and high-quality intersection-union (IoU) loss; the IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; the L IoU For intersection-union ratio loss; the L dis For center distance loss; the L shp For width and height loss; the stated L ang The angle loss is Δ, which is twice the center distance loss. b and stated These represent the center points of the ground truth bounding box and the center points of the predicted bounding box, respectively; These represent the squares of the Euclidean distances between the center points of the ground truth bounding box and the center point of the predicted bounding box, respectively; and stated These represent the width and height of the minimum bounding rectangle, respectively; Ω represents twice the width and height loss; the... w and These are the widths of the predicted border and the true border, respectively; Represents the square of the Euclidean distance between the width of the ground truth bounding box and the width of the predicted bounding box; h and The heights of the predicted bounding box and the ground truth bounding box are respectively; The square of the Euclidean distance between the heights of the ground truth bounding box and the predicted bounding box; The angle loss is twice that of the stated angle loss; and stated These represent the width and height of the rectangle constructed from the center points of the ground truth box and the predicted box, respectively.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the unmanned aerial vehicle-based road disease detection method according to any one of claims 1 to 6.

8. A computer program product storing a computer program comprising, The computer program is executed by the processor to implement the unmanned aerial vehicle-based road disease detection method according to any one of claims 1 to 6.

Images