An unmanned aerial vehicle perspective-oriented road subtle disease intelligent detection method and device

Abstract

The application discloses a kind of unmanned aerial vehicle visual angle-oriented road subtle disease intelligent detection method and device, pre-treatment and input adaptation are carried out to road overhead image, and standardized input image is obtained;Non-destructive down-sampling feature extraction is carried out to standardized input image, and multi-scale feature map is obtained;Global aggregation and distribution are executed to multi-scale feature map, and coordinate attention mechanism is introduced in feature fusion process, and enhanced multi-scale feature map is obtained;Angle perception bounding box regression is carried out to enhanced multi-scale feature map, and candidate road subtle disease detection result is obtained and output detection result is output.The application can solve the problem of high road subtle disease miss detection rate and false detection rate, improve the sensitivity to slender crack disease, and realize accurate positioning of road subtle disease.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection of road traffic infrastructure and computer vision, and in particular relates to an intelligent detection method for minor road defects from the perspective of unmanned aerial vehicles. Background Technology

[0002] With the continuous expansion of road infrastructure and its increasing service life, the early detection and timely maintenance of pavement defects are crucial for ensuring traffic safety, extending road lifespan, and reducing maintenance costs. Traditional road inspection methods mainly rely on manual visual inspection or vehicle-mounted inspection equipment, which suffer from low efficiency, limited coverage, significant traffic disruption, and insufficient adaptability to special scenarios such as mountainous areas, bridges, and narrow road sections. In recent years, drones have gradually become an important tool for road defect inspection due to their advantages such as mobility, non-contact data acquisition, wide coverage, and relatively low deployment costs.

[0003] However, there are significant differences between UAV top-down images and traditional ground-based planar images in terms of target scale distribution, background composition, and imaging geometry. First, road defects, especially fine cracks, in UAV images typically occupy only a small number of pixels, and may even appear as weak textures in some areas. Second, road surfaces are often accompanied by complex background interference such as tree shadows, manhole covers, tire tracks, road repair marks, and debris, resulting in small grayscale differences between the defective area and the background. Third, crack-type defects usually have a distinct elongated shape and a large aspect ratio, with prominent geometric principal directions, making the bounding box localization stability of traditional general-purpose target detection algorithms poor on this type of target.

[0004] While existing general object detection models based on convolutional neural networks perform well in conventional vision tasks, they still have significant shortcomings in scenarios involving subtle road defects encountered by drones. Firstly, conventional backbone networks typically use stride convolution or pooling for downsampling, which compresses spatial resolution but directly loses fine-grained texture information of small-scale cracks, leading to a higher false negative rate. Secondly, conventional multi-scale feature fusion methods often prioritize information transfer between local adjacent layers, lacking collaborative modeling of global semantic and orientation-sensitive features, making it difficult to effectively suppress false targets in complex backgrounds. Thirdly, traditional IoU-type bounding box loss functions mainly focus on overlapping area, center distance, and aspect ratio, resulting in insufficient regression constraints for highly directional targets such as thin cracks, easily leading to localization shifts. Summary of the Invention

[0005] To address the above problems, this invention provides an intelligent detection method, device, electronic equipment, and storage medium for minor road defects from the perspective of unmanned aerial vehicles (UAVs). This method can reduce the rate of missed detection and false detection of road defects, improve the sensitivity to defects such as slender cracks, and achieve accurate location of minor road defects.

[0006] According to one aspect of the present invention, a method for intelligent detection of minor road defects from the perspective of unmanned aerial vehicles (UAVs) is provided, comprising: The road top-view image is preprocessed and adapted for input to obtain a standardized input image; the road top-view image is acquired by an image acquisition device carried by an unmanned aerial vehicle (UAV) flight platform. The standardized input image is fed into the feature extraction network. In at least one shallow downsampling stage of the backbone network, a lossless downsampling structure combining spatial depth transformation and convolutional fusion is used to process the feature map to obtain a multi-scale feature map. Global aggregation and distribution are performed on the multi-scale feature maps. The features at each scale are scale-aligned and feature-fused to obtain global aggregated features. The global aggregated features are then injected back into the feature maps at each scale. A coordinate attention mechanism is introduced during the feature fusion process. Position-sensitive information is extracted from the feature map after feature fusion along the horizontal and vertical directions, and corresponding attention weights are generated. The input features are then weighted with direction sensitivity to enhance the multi-scale feature map. The enhanced multi-scale feature map is input into the detection head to perform disease category prediction and bounding box regression, and candidate results are output; the bounding box regression adopts a loss function that includes an angle cost term; The candidate results are post-processed to output the category, location, and confidence level information of minor road defects.

[0007] When preprocessing and adapting the road top-view image, the preprocessing includes at least one or more operations such as image denoising, brightness or color normalization, size normalization, overlapping sliding window cropping, and label synchronous mapping. The road top-view image is acquired by an image acquisition device mounted on an UAV flight platform. Image acquisition can be performed using either single-frame image acquisition or frame-by-frame video stream capture. Image data can be input into an airborne edge computing terminal in real time or stored for offline processing.

[0008] In the process of feeding the standardized input image into the feature extraction network, at least one shallow downsampling stage of the backbone network is replaced by a lossless downsampling structure that combines spatial depth transformation and convolution fusion to replace the traditional stride convolution or pooling operation. use Represents the input feature map, where, For feature map height, The width of the feature map. Let be the number of channels; let the step size of the spatial depth transformation be . The output feature map after lossless rearrangement is:

[0009] Its mapping relationship can be expressed as:

[0010] in, and This represents the spatial location index of the rearranged feature map. Indicates the original channel index. and This represents the slice offset; the above process essentially rearranges the spatial pixels of the local neighborhood in the original feature map to the channel dimension according to rules, thus reducing spatial resolution without directly discarding any pixel sampling points; subsequently, convolutional fusion is performed on the rearranged feature map: This is to achieve lossless transfer and feature integration of spatial information to the channel dimension.

[0011] When performing global aggregation and distribution on the multi-scale feature maps obtained by lossless downsampling, let the features at different scales be as follows: , and First, adjust the channels and align the spatial scale to obtain:

[0012] in, This represents the features after scale alignment. The scale alignment mapping function can be composed of upsampling, downsampling, convolution adjustment, or a combination thereof; the scale-aligned features are then fused to obtain globally aggregated features.

[0013] in, Indicates global clustering characteristics. The aggregation function can be implemented using methods such as concatenated convolution, weighted summation, linear transformation, and MLP mapping; then, the global aggregation features are injected back into the feature maps at each scale through the corresponding mapping function.

[0014] in, This represents the enhanced multi-scale feature map. This is a distribution mapping function used to redistribute global semantic information to various scale branches.

[0015] In the feature fusion process, the input features are represented as follows: One-dimensional aggregation is performed along two directions respectively:

[0016] in, This represents the horizontal description obtained by aggregation along the width direction. This represents the vertical description obtained by aggregating along the height direction; the two sets of directional descriptions are then mapped using a mapping function to generate corresponding attention weights:

[0017] in, This represents the attention weights in the width direction. The attention weights represent the height direction. and For mapping functions, The activation function is used; finally, the input features are subjected to orientation-sensitive weighting to obtain the enhanced multi-scale feature map:

[0018] in, This is the enhanced multi-scale feature map.

[0019] This orientation-sensitive enhancement method can highlight crack skeleton features and suppress background spurious responses.

[0020] When the enhanced multi-scale feature map is input into the detection head, the bounding box regression uses a loss function that includes an angle cost term:

[0021] in, Represents the loss function. This represents the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. This represents the center distance loss term. Represents the shape difference loss term. This represents the angle cost item; further, the angle cost item can be expressed as:

[0022] in, The normalized angle parameter is related to the offset direction of the prediction box center; this loss function can improve the directional consistency and geometric fit of slender crack targets.

[0023] When post-processing the candidate results output by the detection head, the system outputs information on the category, location, and confidence level of minor road defects. Preferably, the output categories include longitudinal cracks, transverse cracks, diagonal cracks, alligator cracks, potholes, and repair areas. The detection results can be displayed in real time on a ground monitoring terminal or output to a defect database or a road maintenance decision-making system.

[0024] According to another aspect of the present invention, an intelligent detection device for minor road defects from the perspective of unmanned aerial vehicles (UAVs) is provided, comprising an image acquisition module, a preprocessing module, a non-destructive downsampling feature extraction module, a global-local collaborative enhancement module, and a detection output module. The device can be integrated into an airborne edge computing terminal, a ground server, or other electronic processing platform.

[0025] According to another aspect of the present invention, an electronic device is provided, including a processor, a memory, and a computer program stored in the memory, wherein when the computer program is executed by the processor, it implements the above-described intelligent detection method for minor road defects.

[0026] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described intelligent detection method for minor road defects.

[0027] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention addresses the unique imaging defects of UAV-based overhead road images by proposing three core innovations: First, it proposes a lossless downsampling mechanism based on spatial depth transformation, adapting to the imaging characteristics of UAV aerial photography where the pixel ratio of minute cracks is extremely low and fine-grained textures are easily lost, achieving complete transfer of spatial information of minute cracks to the channel dimension. Second, it proposes a global-local collaborative enhancement strategy combining global aggregation and distribution with coordinate attention, adapting to global complex background interference scenarios such as tree shadows, repair marks, and manhole covers on UAV-based overhead road images, achieving global semantic coordination and sensitive focusing on crack direction. Third, it proposes a bounding box loss function including an angle cost term, adapting to the geometric features of UAV-based cracks with large aspect ratios and random tilt, improving the directional consistency and positioning accuracy of elongated crack targets. The overall solution is specifically adapted to UAV road inspection overhead image acquisition scenarios, significantly reducing the missed detection rate and false detection rate of minor road defects in UAV aerial images, achieving accurate positioning and classification of minor defects under complex aerial road conditions. Attached Figure Description

[0028] Figure 1 This is the overall flowchart of the intelligent detection method for minor road defects from the perspective of unmanned aerial vehicles (UAVs) as described in this invention.

[0029] Figure 2 This is a schematic diagram of the lossless downsampling structure based on spatial depth transformation described in this invention.

[0030] Figure 3 This is a schematic diagram of the multi-scale feature global aggregation and distribution structure described in this invention.

[0031] Figure 4 This is a schematic diagram of the coordinate attention direction enhancement structure described in this invention.

[0032] Figure 5 This is a schematic diagram of the angle-sensing bounding box regression principle described in this invention.

[0033] Figure 6 This is a block diagram of the overall device structure described in this invention.

[0034] Figure 7 This is a schematic diagram illustrating the application of the present invention in a drone road inspection system.

[0035] Figure 8 This is a comparison chart of the training performance of the method described in this invention and existing methods in the task of detecting minor road defects. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] The purpose of this invention is to provide an intelligent method for detecting minor road defects from the perspective of unmanned aerial vehicles (UAVs), comprising the following steps: S1, preprocess and adapt the road top-view image to the input to obtain a standardized input image; the road top-view image is acquired by an image acquisition device mounted on an unmanned aerial vehicle (UAV) flight platform; S2, the standardized input image is fed into the feature extraction network. In at least one shallow downsampling stage of the backbone network, a lossless downsampling structure combining spatial depth transformation and convolution fusion is used to process the feature map to obtain a multi-scale feature map. S3, perform global aggregation and distribution on the multi-scale feature map, perform scale alignment and feature fusion on each scale feature to obtain global aggregated features, and then inject the global aggregated features back into each scale feature map; S4. In the feature fusion process, a coordinate attention mechanism is introduced to extract position-sensitive information along the horizontal and vertical directions of the feature map after feature fusion, generate corresponding attention weights, apply direction-sensitive weighting to the input features, and enhance the multi-scale feature map. S5, input the enhanced multi-scale feature map into the detection head, perform disease category prediction and bounding box regression, and output candidate results; the bounding box regression adopts a loss function that includes an angle cost term; S6. Post-process the candidate results and output the category, location, and confidence information of minor road defects.

[0038] For step S1, the corresponding Figure 1 In the image preprocessing stage of the overall flowchart shown, the road top view image is acquired by the image acquisition device carried by the UAV flight platform. The image acquisition can be carried out by single frame image acquisition or by frame-by-frame capture of video stream. The image data can be input into the airborne edge computing terminal in real time or stored and processed offline.

[0039] For the specific implementation of the lossless downsampling structure based on spatial depth transformation in step S2, the corresponding Figure 1 The lossless downsampling step in the overall flowchart shown, and Figure 2 The diagram shows a lossless downsampling structure based on spatial depth transformation.

[0040] like Figure 2 As shown, the input feature map The size is ,in, For feature map height, The width of the feature map. Let be the number of channels; let be the step size for spatial depth transformation. =2, perform spatial slicing rearrangement to obtain Slices 1-4, and rearrange the four slices in a predetermined order to the channel dimension to obtain the output feature map. The size becomes The mapping relationship can be expressed as:

[0041] in, and This represents the spatial location index of the rearranged feature map. Indicates the original channel index. and This represents the slice offset; subsequently, the rearranged feature map... Perform convolutional fusion: This process achieves lossless transfer and feature integration of spatial information to the channel dimension; finally, the output features are generated. Step S3 rearranges local neighboring pixels to the channel dimension after spatial slicing, and then achieves lossless information transfer through convolutional fusion.

[0042] For step size The value of can be adjusted according to the actual application scenario. The convolutional fusion layer can use 1×1 convolution, 3×3 convolution or concatenated convolution.

[0043] The non-destructive downsampling mechanism significantly improves the detection capability of small-target defects such as fine cracks, and reduces the problem of fine-grained texture loss caused by traditional strided convolution.

[0044] For the specific implementation of global-local collaborative feature enhancement of multi-scale feature maps in steps S3 and S4, the corresponding... Figure 1 The overall flowchart shown includes the global aggregation and distribution stage and the directional attention enhancement stage, as well as... Figure 3 The diagram shows the global aggregation and distribution structure of multi-scale features. Figure 4 The diagram shows a coordinate attention direction enhancement structure.

[0045] The scale-aligned features are then fused to obtain globally aggregated features:

[0046] in, This can be achieved using methods such as convolution after concatenation, weighted summation, linear transformation, and MLP mapping; then, the globally aggregated features are injected back into the feature maps at each scale through the corresponding mapping function.

[0047] in, This is a distribution mapping function used to redistribute global semantic information to various scale branches.

[0048] like Figure 3 As shown, the backbone network outputs feature maps at three scales. , and First, adjust the channels and align the spatial scale:

[0049] in, This represents the features after scale alignment. The scale-aligned mapping function is then used; then global aggregation is performed:

[0050] in, Indicates global clustering characteristics. The aggregation function is used; then the global aggregation features are injected back into the feature maps at each scale through the corresponding mapping function:

[0051] in, This represents the enhanced multi-scale feature map. The distribution mapping function is used; finally, the enhanced multi-scale feature map is output.

[0052] The scale alignment method in global aggregation and distribution can be any one or more of upsampling, downsampling, interpolation scaling, and convolution transformation; the aggregation method can be splicing, weighted summation, MLP mapping, or convolution fusion; the distribution method can be residual summation, gated injection, or layer-by-layer mapping back injection.

[0053] like Figure 4 As shown, the input feature map The size is ,in, For feature map height, The width of the feature map. The number of channels; one-dimensional aggregation is performed along the horizontal (width) and vertical (height) directions respectively:

[0054] in, This represents the horizontal description obtained by aggregation along the width direction. This represents the vertical description obtained by aggregating along the height direction; the two sets of directional descriptions are then mapped using a mapping function to generate corresponding attention weights:

[0055] in, This represents the attention weights in the width direction. The attention weights represent the height direction. and For mapping functions, The activation function is used to finally apply orientation-sensitive weighting to the input features.

[0056] The enhanced multi-scale feature map is obtained. And output. The entire step S4 models long-distance dependencies along the horizontal and vertical directions respectively, preserving location information and enhancing crack orientation-sensitive features.

[0057] In directional attention, directional feature extraction can employ average pooling, max pooling, or a combination of both; the mapping function can consist of convolutional layers, fully connected layers, normalization layers, and activation functions.

[0058] By combining global aggregation and distribution with coordinate attention, we can improve the anti-interference ability under complex background conditions and reduce false detections caused by spurious responses such as tree shadows, repair marks and stains.

[0059] For the specific implementation of the bounding box loss function that includes the angle cost term in step S5, the corresponding Figure 1 The detection head regression stage in the overall flowchart shown, and Figure 5 The diagram shown illustrates the principle of angle-sensing bounding box regression.

[0060] like Figure 5 As shown, for targets with slender cracks, there is not only a positional offset but also an angular deviation between the predicted and ground truth bounding boxes. Traditional IoU loss functions only consider the overlapping area and are insensitive to angular deviations, leading to insufficient bounding box fit. The bounding box regression in this invention employs a loss function that includes an angular cost term:

[0061] in, Represents the loss function. This represents the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. This represents the center distance loss term. Represents the shape difference loss term. This represents the angle cost item; further, the angle cost item can be expressed as:

[0062] in, This is a normalized angle parameter related to the offset direction of the prediction box center; this loss function can improve the directional consistency and geometric fit of the slender crack target. Throughout step S5, the directional deviation of the slender crack target is explicitly modeled, making the prediction box fit the target's main direction better.

[0063] By introducing an angle cost term into the bounding box regression, the positioning accuracy and frame fit of slender targets such as longitudinal cracks, transverse cracks, and diagonal cracks can be improved.

[0064] The bounding box regression step can use any regression loss function that includes an angle modeling mechanism, as long as it can explicitly constrain the deviation of the target principal direction, it can be used as an equivalent alternative.

[0065] For step S6, the corresponding Figure 1 The overall flowchart shown includes output steps for the category, location, and confidence level of road defects; post-processing of the candidate results output by the detection head outputs the category, location, and confidence level information of minor road defects. Preferably, the output categories include longitudinal cracks, transverse cracks, diagonal cracks, alligator cracks, potholes, and repair areas.

[0066] like Figure 6 As shown, the present invention also provides a detection device for a road minor defect intelligent detection method from the perspective of UAVs, comprising: an image acquisition module, a preprocessing module, a non-destructive downsampling feature extraction module, a global-local collaborative enhancement module, a detection output module, and a result output module.

[0067] The image acquisition module is used to acquire overhead road images captured by the image acquisition equipment mounted on the UAV flight platform.

[0068] The preprocessing module is used to preprocess and adapt the road top-view image to the input to obtain a standardized input image. The operations performed by this module include one or more of the following: image denoising, brightness or color normalization, size normalization, overlapping sliding window cropping, and synchronized label mapping. For ultra-high resolution UAV images, preferably, the preprocessing module employs a sliding window cropping strategy with overlapping boundaries to reduce the risk of missed detections caused by cracks crossing the window.

[0069] The lossless downsampling feature extraction module processes feature maps using a lossless downsampling structure that combines spatial depth transformation and convolutional fusion. This module receives the normalized image output from the preprocessing module, extracts features through a backbone network, and replaces traditional stride convolution or pooling operations with spatial depth transformation in at least one shallow downsampling stage of the backbone network, outputting multi-scale feature maps.

[0070] The global-local collaborative enhancement module is used to perform global aggregation and distribution on the multi-scale feature maps obtained by lossless downsampling, and introduces a coordinate attention mechanism for orientation-sensitive weighted enhancement, outputting the enhanced multi-scale features.

[0071] The detection output module is used to input the enhanced multi-scale features into the detection head to perform disease category prediction and bounding box regression, and to post-process the candidate results to output the category, location and confidence information of minor road diseases; the bounding box regression adopts a loss function that includes an angle cost term.

[0072] The final output module will output the detection results to the terminal, the disease database, or the road maintenance decision-making system.

[0073] like Figure 1 and Figure 7 As shown, the present invention also provides an application of a road minor defect intelligent detection method from the perspective of UAVs in a UAV road inspection system.

[0074] Specifically, such as Figure 7As shown, this invention is deployed in an unmanned aerial vehicle (UAV) road inspection system. 701 represents the road surface and affected areas. After the UAV platform (702) equipped with an image acquisition device acquires road images, the acquired image data is input into the onboard edge computing terminal (703) in real time, or it can be stored and processed offline. After receiving the images, the onboard edge computing terminal executes the above steps S1-S6. The detection results are transmitted in real time to the ground monitoring station (705) via the wireless transmission module (704). Ground operators can view the location, type, and confidence level information of the defects in real time. The detection results can also be output to a defect database or a road maintenance decision-making system. The detection results can also be stored in the onboard storage device, and the defect detection results and alarm information can be further analyzed after the UAV returns to the ground monitoring station (705). The UAV platform, camera model, computing chip, and communication method can all be replaced, and do not constitute a substantial limitation of this invention.

[0075] This invention has good engineering deployment value while taking into account both accuracy and real-time performance, and is suitable for intelligent road inspection applications at the drone, edge, and ground levels.

[0076] Figure 8 This is a comparison chart of the training performance of the method of this invention and existing methods in the task of detecting minor road defects. Figure 8 (a) shows the curve of accuracy as a function of training rounds. Figure 8 (b) shows the curve of recall rate as a function of training rounds. Figure 8 (c) represents the curve of mAP@0.5 changing with the number of training rounds. Figure 8 (d) represents the curve showing the change of mAP@0.5:0.95 with training rounds. The method of this invention outperforms the comparative method in terms of precision, recall, mAP@0.5, and mAP@0.5:0.95 in the road minor defect detection task. Furthermore, the method converges faster and exhibits less fluctuation in the later stages, indicating that it not only improves detection accuracy but also possesses better training stability and overall performance. Specifically, in the early stages of training, the various indicators of the method of this invention increase more rapidly, indicating higher learning efficiency for defect features. In the later stages of training, the indicators remain at higher levels, demonstrating that the method of this invention has better accuracy and generalization ability in the minor defect identification task.

[0077] This invention addresses the problem of high false negative rates caused by the loss of high-frequency texture information in traditional downsampling processes for small targets such as minute cracks in road images viewed from above by UAVs. It also addresses the issue of strong interference from tree shadows, repair marks, manhole covers, stains, and other background textures on road surfaces in UAV scenarios, leading to insufficient focus on the affected areas and high false positive rates in existing detection models. Furthermore, this invention addresses the problem that cracks, with their large aspect ratio, slender shape, and significant directionality, are not sufficiently sensitive to angles in existing bounding box regression mechanisms, resulting in frame offsets and inaccurate positioning. Existing methods that rely on deepening the network and stacking complex 2D attention modules to improve detection accuracy significantly increase the number of parameters and floating-point computational overhead, making it difficult to balance real-time performance with edge device deployment and unsuitable for real-time inspection scenarios on low-computing-power UAV-borne terminals. This invention utilizes a lightweight, lossless downsampling structure, a one-dimensional coordinate attention mechanism, and a single-round global aggregation and distribution multi-scale fusion strategy. It adds a low-cost angle loss constraint only to the detection head, thereby improving the accuracy of detecting minor defects while controlling the overall inference computing power consumption. This effectively resolves the contradiction between accuracy improvement and computational overhead, meeting the engineering implementation requirements of real-time inspection at UAV and edge devices.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent detection of minor road defects from the perspective of unmanned aerial vehicles (UAVs), characterized in that, include: The road top view image is preprocessed and adapted to the input to obtain a standardized input image; The overhead view of the road was acquired using image acquisition equipment carried on a drone flight platform; The standardized input image is fed into the feature extraction network. In at least one shallow downsampling stage of the backbone network, a lossless downsampling structure combining spatial depth transformation and convolutional fusion is used to process the feature map to obtain a multi-scale feature map. Global aggregation and distribution are performed on the multi-scale feature maps. The features at each scale are scale-aligned and feature-fused to obtain global aggregated features. The global aggregated features are then injected back into the feature maps at each scale. A coordinate attention mechanism is introduced during the feature fusion process. Position-sensitive information is extracted from the feature map after feature fusion along the horizontal and vertical directions, and corresponding attention weights are generated. The input features are then weighted with direction sensitivity to enhance the multi-scale feature map. The enhanced multi-scale feature map is input into the detection head to perform disease category prediction and bounding box regression, and candidate results are output; the bounding box regression adopts a loss function that includes an angle cost term; The candidate results are post-processed to output the category, location, and confidence level information of minor road defects.

2. The intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle (UAV) as described in claim 1, characterized in that: The preprocessing includes at least one or more of the following operations: image denoising, brightness or color normalization, size normalization, overlapping sliding window cropping, and label synchronization mapping.

3. The intelligent detection method for minor road defects from the perspective of unmanned aerial vehicles (UAVs) as described in claim 1, characterized in that: When the standardized input image is fed into the feature extraction network, let the input feature map be... The height is Width is The number of channels is The step size for spatial depth transformation is The feature map is then output after lossless rearrangement. The height is Width is The number of channels is Convolutional fusion is performed on the rearranged feature maps to achieve lossless transfer of spatial information to the channel dimension and feature integration.

4. The intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle (UAV) as described in claim 1, characterized in that: When performing global aggregation and distribution on the multi-scale feature maps obtained by lossless downsampling, let the features at different scales be as follows: , and First, adjust the channels and align the spatial scale to obtain: in, This represents the features after scale alignment. This represents the scale alignment mapping function; the scale-aligned features are then fused to obtain the globally aggregated features. in, Indicates global clustering characteristics. The global clustered features are then injected back into the feature maps at each scale through the corresponding mapping function: in, This represents the enhanced multi-scale feature map. This is a distribution mapping function used to redistribute global semantic information to various scale branches.

5. The intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle (UAV) as described in claim 1, characterized in that: In the feature fusion process, let the input features be... One-dimensional aggregation is performed along two directions respectively: in, This represents the horizontal description obtained by aggregation along the width direction. This represents the vertical description obtained by aggregating along the height direction; the two sets of directional descriptions are then mapped using a mapping function to generate corresponding attention weights: in, This represents the attention weights in the width direction. Attention weights in the height direction, and For mapping functions, The activation function is used; finally, the input features are subjected to orientation-sensitive weighting to obtain the enhanced multi-scale feature map: in, This is the enhanced multi-scale feature map.

6. The intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle (UAV) as described in claim 1, characterized in that: When the enhanced multi-scale feature map is input into the detector head, the loss function for bounding box regression is expressed as: in, Represents the loss function. This represents the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. This represents the center distance loss term. Represents the shape difference loss term. This represents the angle cost item.

7. The intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle (UAV) as described in claim 6, characterized in that: When the enhanced multi-scale feature map is input into the detector head, the angle cost term of the loss function for bounding box regression is expressed as: in This is a normalized angle parameter related to the direction of the center offset of the prediction box.

8. A smart detection device for minor road defects from the perspective of unmanned aerial vehicles (UAVs), characterized in that, It includes an image acquisition module, a preprocessing module, a lossless downsampling feature extraction module, a global-local collaborative enhancement module, and a detection output module.

9. An electronic device, characterized in that: The system includes a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the intelligent road defect detection method from the perspective of an unmanned aerial vehicle (UAV) as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the intelligent detection method for minor road defects from the perspective of an unmanned aerial vehicle as described in any one of claims 1-7.

Images