Road disease detection algorithm based on unmanned aerial vehicle inspection

Abstract

The application provides a road disease detection algorithm based on unmanned aerial vehicle inspection, comprising the following steps: S1, decoupling and normalization convolution operation: inputting a road image collected by an unmanned aerial vehicle into an improved YOLOv11 model for decoupling and normalization operation; S2, backbone network feature extraction: the normalized image first enters a backbone feature extraction network; S3, Neck level expansion and multi-scale feature fusion: the multi-layer features output by the backbone feature extraction network are sent to a feature fusion network for multi-scale feature fusion; S4, frequency domain enhancement fusion: the multi-scale features after the Neck fusion are respectively input into a HEADF module; S5, multi-scale detection output: the features after the HEADF processing are respectively input into a detection head, and finally, the model outputs complete road disease detection results. The C3k2_WTConv module designed by the application not only enables the multi-scale feature extraction capability of wavelet transformation to be flexibly embedded into the original C3k2 module, but also ensures the structural compatibility of the module through a channel number judgment mechanism, so that the detection precision of the model on multi-scale targets can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of road defect detection technology, and more specifically, to a road defect detection algorithm based on unmanned aerial vehicle (UAV) inspection. Background Technology

[0002] Road inspection is a core function to ensure the safety, smooth flow, and extended service life of road infrastructure. Regular inspections, recording, assessments, and handling of the main road structure and its ancillary facilities help identify and resolve potential problems in a timely manner, preventing traffic accidents or road malfunctions. Currently, the most common road problem is age-related cracking, which, if not repaired promptly, increases the probability of traffic accidents. China's total highway mileage has been increasing year by year. As of the end of 2024, the total national highway mileage reached 5.4904 million kilometers, including 190,700 kilometers of expressways (see "Ministry of Transport. (2025). 2024 Statistical Bulletin on the Development of the Transportation Industry [EB / OL]. https: / / xxgk.mot.gov.cn / 2020 / jigou / zhghs / 202506 / t20250610_4170228.html."). Currently, road maintenance departments mainly rely on manpower for road defect detection (see "Cao Yan. Problems and Countermeasures in Municipal Road Engineering Maintenance [J]. China Building Decoration and Renovation, 2022(4):120-121."). However, due to the inability of manpower to conduct frequent inspections and the uncertainty of the location of cracks, timely repair of road defects is not possible. With the development of the times, in the field of transportation, drone inspection (see "Quan Deqiang. Exploration of the Application of '5G+Drone' Technology in Highway Inspection [J]. China Transportation Informationization, 2023, No.289(11)") is gradually becoming a key means to solve the problem of road monitoring. Drones, with their excellent flexibility, wide coverage and more economical cost, effectively improve inspection efficiency and reduce resource waste.

[0003] In the field of crack detection, the non-patent literature “Ding K, Ding Z, Zhang Z, et al. SCD-YOLO: anovel object detection method for efficient road crack detection[J].Multimedia Systems, 2024, 30(6): 351.” discloses that by drawing on shift operators, a sparse mechanism and shifted standard convolution are used to simulate the effect of large convolution kernels, which significantly expands the receptive field of the model while controlling the amount of computation; at the same time, a built-in scale-aware attention is adopted to dynamically adjust the feature representation to adapt to small objects of different scales. The non-patent literature “Pei J, Wu X, Liu X. YOLO-RDD: A road defect detection algorithm based on YOLO[C] / / 2024 27th International Conference on Computer Supported Cooperative Work in Design(CSCWD). IEEE, 2024: 1695-1703.” discloses a method based on the YOLOv8 framework that achieves multi-scale detection by adding P2 and P6 feature layers, and combines a coordinate attention mechanism to build cross-layer connections, thereby strengthening the interaction of semantic information between deep and shallow layers to improve the crack feature capture capability. The non-patent literature “Yanfei P, Yue J. Road crack detection algorithm based on improved YOLOv8[C] / / 2023 5th International Conference on Artificial Intelligence and Computer Applications (ICAICA). IEEE, 2023: 28-32.” discloses a method that enhances the crack feature extraction capability at different scales and improves detection accuracy by introducing a GAM attention mechanism into the backbone network and neck network and using a DIoU loss function to optimize the bounding box loss. However, the above three methods increase the complexity of the model, making them difficult to apply to the field of drone detection.The non-patent literature “Ma D, Fang H, Wang N, et al. Automatic detection and counting system for pavement cracks based on PCGAN and YOLO-MF[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(11):22166-22178.” discloses a method to generate crack images using generative adversarial networks to compensate for insufficient training data, and to construct a YOLO-MF model using an improved YOLOv3 and median flow algorithm to achieve crack detection, tracking, and accurate counting in video sequences. The non-patent technology literature “Wang W, Li L, Han Y. Crack detection in shadowed images ongray level deviations in a moving window and distance deviations between connected components[J]. Construction and Building Materials, 2021, 271:121885.” discloses a method to automatically determine a suitable window based on image resolution, segment the image by setting a threshold through the local window grayscale standard deviation to initially extract crack segments, then use the distance standard deviation of connected regions to filter cracks, and finally combine the crack segment gap length and adjacent segment direction information to complete crack connection. The non-patent literature “Hu H, Li Z, He Z, et al. Roadsurface crack detection method based on improved YOLOv5 and vehicle-mountedimages[J]. Measurement, 2024, 229: 114443.” discloses a method that integrates Slim-Neck, C2f, Decoupled Head and SPPCSPC structure, combined with Silu activation function and CIoU loss function, to solve the problems of blurring of small cracks and incomplete information extraction, while balancing detection accuracy and real-time performance.The non-patent literature “Meng A, Zhang X, Yu X, et al. Investigation on lightweight identification method for pavement cracks[J]. Construction and Building Materials, 2024, 447: 138017.” proposes a knowledge distillation model enhanced by multiple teaching assistants and incorporates a random learning strategy to achieve lightweight crack detection. Although the above method improves detection accuracy, the model may produce false detections in noisy scenes such as strong light and shadow. The non-patent literature “Chai H, Zeng S, Zhang C. Enhanced Real-time Pavement Damage Analysis: Integrating CNN Edge Detection with YOLO for Accurate Classification[C] / / 2024 2nd International Conferenceon Computer, Vision and Intelligent Technology (ICCVIT). IEEE, 2024: 1-8.” discloses a method that preprocesses the image by reducing its size through minimum filtering and grayscale conversion, while using edge detection techniques such as Laplacian filters to extract longitudinal, transverse, and mesh crack features, and combining the feature learning capabilities of CNN with the real-time detection advantages of YOLO to achieve crack classification. However, this method has difficulty estimating the original size of the pit associated with the crack, and is limited by the mobility of the processing equipment, making it difficult to achieve large-scale deployment. Summary of the Invention

[0004] In view of this, this application provides a road defect detection algorithm based on UAV inspection to overcome the above-mentioned defects in the prior art.

[0005] To achieve the above objectives, the technical solution adopted in this application is as follows: Road defect detection algorithms based on drone inspection include: S1, Decoupled Normalized Convolution Operation: The road image acquired by the UAV is input into the improved YOLOv11 model, and the Conv_nbn module (decoupled normalized convolution module) is used to perform a decoupled normalized convolution operation on the input road image. The improved YOLOv11 model refers to the YOLOv11 model with the addition of the Conv_nbn module, the C3k2_WTConv module, and the HEADF module. The YOLOv11 model includes an input end, a backbone network, a neck feature fusion network, and a detection head connected in sequence. The Conv_nbn module is set at the input end to replace the original input convolution module and perform decoupled normalized convolution processing on the input road image. The C3k2_WTConv module is set in the neck feature fusion network to introduce wavelet convolution feature enhancement in the fusion stage after upsampling and feature concatenation. The HEADF module is set between the output end of the neck feature fusion network and the detection head to perform frequency domain enhancement on the fused features of adjacent scales before feeding them into the detection head for prediction. S2, Backbone Network Feature Extraction: The image after decoupling and normalizing convolution enters the backbone network for feature extraction. That is, the image after decoupling and normalizing convolution passes through the Conv and C3k2 basic modules for layer-by-layer downsampling and semantic feature extraction, resulting in multiple feature maps of different resolutions. S3, Neck layer expansion and multi-scale feature fusion: The Neck layer is expanded from four layers to six layers; the deep high semantic features of the multi-layer features extracted by the backbone network are fused with the shallow high resolution features; at the same time, wavelet convolution is added to the C3k2_WTConv module in the neck to enhance the fusion effect. After multiple upsampling and feature concatenation (Upsample+Concat), a multi-scale fused feature map is constructed. S4, Frequency Domain Enhanced Fusion: Two multi-scale fused feature maps from adjacent layers are input into the HEADF module. The HEADF module uses wavelet transform to decompose the input features into low-frequency structural information and high-frequency detail information; whereby the low-frequency structural information is used to characterize the overall structure of the road area, and the high-frequency detail information is used to characterize the detailed response of crack edges and pothole contours.

[0006] S5, Multi-scale detection output: The features processed by HEADF are input into the detection head. The detection head performs classification prediction and bounding box regression on the enhanced features at different scales to generate road defect categories, confidence scores, and target bounding box positions. Finally, the model outputs complete road defect detection results.

[0007] Furthermore, the decoupling normalization formula in step S1 is: (1); in, Point to The unit vector obtained by L2 normalization retains only the directional information of γ, eliminating the difference in numerical magnitude. It is a learnable scalar parameter that uniformly controls the amplitude of γ across all channels; Point to The unit vector obtained by L2 normalization retains only the directional information of β, eliminating the difference in numerical magnitude. It is a learnable scalar parameter that uniformly controls the magnitude of β across all channels; The change in the gradient formula of the loss function demonstrates that decoupling normalization can achieve more uniform weights. and bias β, to For example, its gradient calculation formula is as follows: (2); in, It is a positive scaling factor across different channels; , is the key adjustment term in the gradient formula; It is a loss function For parameters The gradient; When the value is greater than 0, the loss function will affect... Punishment is imposed, but the severity of punishment varies depending on the category; the more frequent the category, the greater the severity of the punishment. When <0, the loss function will affect Incentives are provided, but the level of incentive varies across different categories; the more frequently a category is used, the greater the incentive. This method generates a more balanced weight and bias.

[0008] Furthermore, the operation method of the C3k2_WTConv module is as follows: When C3k=False, the input features are first reduced in dimensionality by initial convolution, and then split into two parallel branches by the Split operation. One branch is directly retained, and the other branch is processed by several WTBottleneck modules for multi-component feature extraction. Finally, the two features are fused by Concat, and the output is integrated by the final convolution. When C3k=True, the convolution result is also split as a preprocess, but the subsequent branch is replaced by WT_C3k which incorporates wavelet transform. Finally, the feature fusion and output are also completed by Concat and the final convolution.

[0009] Furthermore, the WT_C3k module operates as follows: the input features are first divided into two parallel processing branches. The first branch performs channel compression via 1×1 convolution and then inputs a feature extraction unit composed of multiple WTBottlenecks. In each WTBottleneck, the features are first processed by convolution to extract local spatial information, and then enhanced in the frequency domain using WTConv wavelet convolution to obtain low-frequency structural information and high-frequency edge detail information. Residual connections are used to maintain the stability of feature propagation. The second branch directly extracts and retains the shallow semantic information of the input features via 1×1 convolution. Subsequently, the deep enhanced features output from the first branch and the shallow preserved features output from the second branch are concatenated along the channel dimension, and channel fusion is completed through a final 1×1 convolution, ultimately outputting the fused feature representation.

[0010] Furthermore, the HEADF module's computation method is as follows: Two adjacent-scale features, X1 and X2, are simultaneously input into the HEADF module. First, the two input features are size-aligned to meet the subsequent fusion requirements in both channel and spatial dimensions. Then, wavelet transforms are performed on X1 and X2 respectively, decomposing each feature into a low-frequency component LL and high-frequency components LH, HL, and HH. LL primarily preserves the overall structure and semantic information of the road area, while LH, HL, and HH primarily preserve details such as crack edges, pothole outlines, and local texture variations. Next, in the frequency domain feature reconstruction stage, the HEADF module selects the high-frequency component HH and adds it to the low-frequency component LL to obtain the LL+HH frequency domain feature, which simultaneously represents the overall road structure information and strong edge detail information. The HL and LH directional high-frequency components are also preserved to describe the crack edge response in different directions. Subsequently, HL, LH, and LL+HH are used as three main frequency domain feature branches, and global max pooling is applied to obtain the corresponding global response features, denoted as X. HL X LH and X LL+HH Among them, X HL and X LH X is used to highlight the crack edge response in different directions. LL+HH It is used to simultaneously preserve the overall road structure information and strong edge detail information.

[0011] Then, in the feature interaction phase, the HEADF module uses X LL+HH As a comprehensive guiding feature of structure and edge, respectively related to X HL and X LH Matrix multiplication yields X hl and X lhThis establishes a correlation between directional high-frequency details and structural enhancement features. Through this operation, the HEADF module can enhance responses related to road distress edges, crack orientation, and pothole contours, while reducing irrelevant background interference such as shadows, water stains, and complex pavement textures.

[0012] Finally, the X obtained through interaction hl and X lh After dimensionality transformation, the features are concatenated, and corresponding enhancement weights are generated through an activation function. These enhancement weights are multiplied by the original input features X1 and X2 respectively to achieve adaptive enhancement of the effective disease areas. Finally, the two enhanced features are added and fused, and local features are integrated through a 3×3 convolution to output a frequency domain enhanced fused feature map.

[0013] Compared with the prior art, the beneficial effects of this application are: 1. The C3k2_WTConv module designed in this application not only allows the multi-scale feature extraction capability of wavelet transform to be flexibly embedded into the original C3k2 module, but also ensures the structural compatibility of the module through the channel number judgment mechanism, which can effectively improve the detection accuracy of the model for multi-scale targets.

[0014] 2. The decoupling normalization method introduced in this application greatly alleviates the problem of poor hole detection caused by the uneven number of road defects in the dataset.

[0015] 3. This application introduces the HEADF module, which enhances the model's detection performance for targets of different scales and small targets. Test results.

[0016] 4. This application expands the number of layers in the Neck section, increasing the feature fusion path from the original number of layers to more layers, allowing shallow high-resolution features to participate again in subsequent fusion processes. By increasing the number of Neck layers, the model can more fully fuse shallow detail features with deep semantic features, thereby improving the problem that small-scale road damage targets are easily lost during multiple downsampling processes from the perspective of UAVs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the road defect detection algorithm based on UAV inspection in this application.

[0019] Figure 2 This is a structural diagram of the YOLOv11 model.

[0020] Figure 3 This is a structural diagram of the road defect detection model (improved YOLOv11 model) based on UAV inspection in this application.

[0021] Figure 4 This is a structural diagram of the Conv_Nbn module in this application.

[0022] Figure 5 This is a structural diagram of the WTBottleneck module in this application.

[0023] Figure 6 This is a diagram of the internal structure of the WT_C3k module in this application.

[0024] Figure 7 This is a structural diagram of the C3k2_WTConv module in this application.

[0025] Figure 8 This is a structural diagram of the HEADF module in this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0027] like Figure 1 and Figure 3 As shown, the road defect detection algorithm based on UAV inspection includes: S1, Decoupled Normalized Convolution Operation: The road image acquired by the UAV is input into the improved YOLOv11 model, and the Conv_nbn module is used to perform a decoupled normalized convolution operation on the input road image. The improved YOLOv11 model refers to the YOLOv11 model with the addition of the Conv_nbn module, the C3k2_WTConv module, and the HEADF module. The YOLOv11 model includes an input end, a backbone network, a neck feature fusion network, and a detection head connected in sequence. The Conv_nbn module is set at the input end to replace the original input convolution module and perform decoupled normalized convolution processing on the input road image. The C3k2_WTConv module is set in the neck feature fusion network to introduce wavelet convolution feature enhancement in the fusion stage after upsampling and feature concatenation. The HEADF module is set between the output end of the neck feature fusion network and the detection head to perform frequency domain enhancement on the fused features of adjacent scales before feeding them into the detection head for prediction. Figure 3 This is a structural diagram of the improved YOLOv11 model in this application. Figure 3 As shown, this method is intended to be implemented in YOLOv11 (internal structure as shown in the figure). Figure 2 Based on the above, we improve upon the previous method by applying wavelet transform as wavelet convolution in C3k2_WTConv, and also by applying wavelet transform in HEADF for background suppression.

[0028] This application employs a decoupled normalization method to balance the feature strengths of frequent and rare classes. Furthermore, to improve the detection performance of the model for small targets, wavelet convolution is used to separate high-frequency and low-frequency information in the image, while simultaneously expanding the receptive field to improve the detection efficiency for holes.

[0029] In most datasets, a long-tail distribution problem exists, meaning that a few categories (such as cracks) have a large number of samples, while most categories (such as holes) have only a small number of samples. To balance the poor detection performance of holes due to the uneven distribution of road defects in the dataset, a decoupling normalization method is proposed, with the following formula: (1); in, Point to The unit vector obtained by L2 normalization retains only the directional information of γ, eliminating the difference in numerical magnitude. It is a learnable scalar parameter that uniformly controls the amplitude of γ across all channels; Point to The unit vector obtained by L2 normalization retains only the directional information of β, eliminating the difference in numerical magnitude. It is a learnable scalar parameter that uniformly controls the amplitude of β across all channels.

[0030] By decoupling the normalization layers γ and β into scalar × unit vectors, the direction and amplitude do not interfere with each other. In this case, frequent classes cannot interfere with the features of rare classes, thus improving the detection efficiency of rare classes. This decoupling of γ and β decomposes class and amplitude information. Furthermore, the change in the gradient formula of the loss function demonstrates that decoupling normalization achieves more uniform weights. and bias β, to For example, its gradient calculation formula is as follows: (2); in, It is a positive scaling factor across different channels; , is the key adjustment term in the gradient formula; It is a loss function For parameters The gradient.

[0031] When the value is greater than 0, the loss function will affect... Punishment is imposed, but the severity of punishment varies depending on the category; the more frequent the category, the greater the severity of the punishment. When <0, the loss function will affect Incentives are provided, but the level of incentive varies across different categories; the more frequently a category is used, the greater the incentive. This method generates a more balanced weight and bias.

[0032] Based on the above principles, a decoupled normalization module is constructed. This module performs normalization operations according to the decoupled normalization formula. This module is then applied to the convolution module (whose workflow is 2D convolution -> normalization -> activation function), replacing the normalization method therein. The Conv_nbn module structure is as follows: Figure 4 As shown.

[0033] The decoupled normalized convolution module separates the direction information and amplitude information of the normalized parameters to reduce the impact of uneven distribution of road defect categories on the feature learning process.

[0034] S2, Backbone Network Feature Extraction: The image after decoupling and normalizing convolution enters the backbone network for feature extraction. That is, the image after decoupling and normalizing convolution passes through the Conv and C3k2 basic modules for layer-by-layer downsampling and semantic feature extraction, resulting in multiple feature maps of different resolutions. Subsequently, the features are sequentially downsampled and semantic feature extracted layer by layer through basic modules such as Conv and C3k2, resulting in multiple feature maps of different resolutions. Among them, shallow features retain more detailed information such as the edges of road cracks and the contours of potholes, while deep features contain richer semantic information.

[0035] S3, Neck layer expansion and multi-scale feature fusion: The Neck layer is expanded from four layers to six layers; the deep high semantic features of the multi-layer features extracted by the backbone network are fused with the shallow high resolution features; at the same time, wavelet convolution is added to the C3k2_WTConv module in the neck to enhance the fusion effect. After multiple upsampling and feature concatenation (Upsample+Concat), a multi-scale fused feature map is constructed. Unlike the shallower Neck structure in the original YOLOv11, this application extends the hierarchy of the Neck section, such as... Figure 3As shown, the four-layer (two upsample layers + two Conv layers) model was expanded to six layers (three upsample layers + three Conv layers), increasing the feature fusion path from the original number of layers to more layers, allowing shallow high-resolution features to participate again in subsequent fusion processes. By increasing the number of Neck layers, the model can more fully fuse shallow detail features with deep semantic features, thereby improving the problem that small-scale road damage targets are easily lost during multiple downsampling processes from the perspective of UAVs.

[0036] Wavelet transform is a common technique used in detection models. Wavelet transform can extract... (Low-frequency component) (Horizontal high-frequency components) (Vertical high-frequency components) (Diagonal high-frequency components) Four types of information. By extracting features from these four types of information and repeatedly stacking them, more detailed features can be obtained.

[0037] like Figure 5 As shown, to efficiently integrate the advantages of wavelet transform while retaining the compatibility of the original modules, this paper designs the WTBottleneck module (i.e., the improved Bottleneck). This module uses whether the number of channels in the convolutional hidden layer is the same as the number of channels in the output layer as the core judgment condition. When the number of channels in the convolutional hidden layer is the same as the number of channels in the output layer, wavelet transform is used to replace the convolution operation in the original Bottleneck, and multi-component feature extraction is used to enhance the expressive power; when the number of channels in the convolutional hidden layer is different from the number of channels in the output layer, the original convolutional structure is retained to ensure the adaptability of the channel dimension.

[0038] Based on the WTBottleneck module, this application constructs the C3k2_WTConv module, whose structure is as follows: Figure 7 As shown, there are two branching structures (operational flow): When C3k=False, the input features are first reduced in dimensionality through initial convolution, and then split into two parallel branches through a Split operation. One branch is directly retained, while the other branch undergoes multi-component feature extraction through several WTBottleneck modules. Multiple stacks of WTBottleneck modules can continuously refine the feature details. Finally, the two feature paths are fused through Concat, and the output is integrated through a final convolution. When C3k=True, the convolution result is also split as a preliminary process, but subsequent branches are replaced with a WT_C3k structure that incorporates wavelet transform (e.g., ...). Figure 6As shown in the diagram, feature fusion and output are finally achieved through Concat and final convolution. This design not only allows the multi-scale feature extraction capability of wavelet transform to be flexibly embedded into the original C3k2 module, but also ensures the structural compatibility of the module through the channel number judgment mechanism, which can effectively improve the detection accuracy of the model for multi-scale targets.

[0039] Specifically, the wavelet transform uses the Haar wavelet basis function, and the wavelet decomposition has two layers. The internal operation flow of the WT_C3k module is as follows: The input features are first divided into two parallel processing branches. The first branch performs channel compression through 1×1 convolution and then inputs a feature extraction unit composed of multiple WTBottlenecks. In each WTBottleneck, the features are first processed by convolution to extract local spatial information, and then enhanced in the frequency domain through WTConv wavelet convolution to obtain low-frequency structural information and high-frequency edge detail information. The stability of feature propagation is maintained through residual connections. The second branch directly extracts and retains the shallow semantic information of the input features through 1×1 convolution. Subsequently, the deep enhanced features output by the first branch and the shallow retained features output by the second branch are concatenated in the channel dimension, and channel fusion is completed through a terminal 1×1 convolution, finally outputting the fused feature representation.

[0040] S4, Frequency Domain Enhanced Fusion: Two multi-scale fused feature maps from adjacent layers are input into the HEADF module. The HEADF module uses wavelet transform to decompose the input features into low-frequency structural information and high-frequency detail information; whereby the low-frequency structural information is used to characterize the overall structure of the road area, and the high-frequency detail information is used to characterize the detailed response of crack edges and pothole contours.

[0041] By interactively modeling high- and low-frequency features, HEADF can enhance edge and texture features associated with road defects while reducing interference from background information such as road surface texture, shadows, stains, and repair marks.

[0042] Furthermore, such as Figure 8As shown, the HEADF module operates as follows: Two adjacent-scale features, X1 and X2, are simultaneously input into the HEADF module. First, the two input features are size-aligned to meet the subsequent fusion requirements in both channel and spatial dimensions. Then, wavelet transforms are performed on X1 and X2 respectively, decomposing each feature into a low-frequency component LL and high-frequency components LH, HL, and HH. LL primarily preserves the overall structure and semantic information of the road area, while LH, HL, and HH primarily preserve details such as crack edges, pothole outlines, and local texture variations. Next, in the frequency domain feature reconstruction stage, the HEADF module selects the high-frequency component HH and adds it to the low-frequency component LL to obtain the LL+HH frequency domain feature, which simultaneously represents the overall road structure information and strong edge detail information. The HL and LH directional high-frequency components are also preserved to describe the crack edge response in different directions. Subsequently, HL, LH, and LL+HH are used as three main frequency domain feature branches, and global max pooling is applied to obtain the corresponding global response features, denoted as X. HL X LH and X LL+HH Among them, X HL and X LH X is used to highlight the crack edge response in different directions. LL+HH It is used to simultaneously preserve the overall road structure information and strong edge detail information.

[0043] Then, in the feature interaction phase, the HEADF module uses X LL+HH As a comprehensive guiding feature of structure and edge, respectively related to X HL and X LH Matrix multiplication yields X hl and X lh This establishes a correlation between directional high-frequency details and structural enhancement features. Through this operation, the HEADF module can enhance responses related to road distress edges, crack orientation, and pothole contours, while reducing irrelevant background interference such as shadows, water stains, and complex pavement textures.

[0044] Finally, the X obtained through interaction hl and X lh After dimensionality transformation, the features are concatenated, and corresponding enhancement weights are generated through an activation function. These enhancement weights are multiplied by the original input features X1 and X2 respectively to achieve adaptive enhancement of the effective disease areas. Finally, the two enhanced features are added and fused, and local features are integrated through a 3×3 convolution to output a frequency domain enhanced fused feature map.

[0045] Through the above processing, the HEADF module can enhance high-frequency defect features such as crack edges, pothole outlines, and local texture changes while preserving the overall road structure information, thereby improving the detection capability of road defect targets in complex backgrounds.

[0046] Specifically, the dimensionality transformation rule is Bmm (Batch Matrix Multiplication), and the two features X hl and X lh The dimension after the dimension transformation is C×1.

[0047] S5, Multi-scale detection output: The features processed by HEADF are input into the detection head. The detection head performs classification prediction and bounding box regression on the enhanced features at different scales to generate road defect categories, confidence scores, and target bounding box positions. Finally, the model outputs complete road defect detection results.

[0048] Higher resolution features are mainly used to detect fine cracks and small-scale pits, while lower resolution features are used to identify diseased areas that are larger in scale or have more obvious semantic features.

[0049] In summary, this application first inputs UAV road images into an improved YOLOv11 network; then, it extracts multi-layer features by decoupling normalized convolutions and the backbone network; next, it enhances the fusion of shallow detail features and deep semantic features through a deepened Neck structure; subsequently, it utilizes the HEADF module for frequency domain decomposition, edge enhancement, and background suppression; finally, it outputs the category and location information of road damage targets through a detection head. This process can improve the model's ability to detect road damage targets with small scale, weak texture, and complex backgrounds.

[0050] Example: (1) Application scenario: This application improves the YOLOv11 model and is used for road defect detection. Existing detection methods have problems such as high false negative rate of small holes, many false background detections, and low detection efficiency, which cannot meet the needs of industrial batch detection. (2) Application process: The improved model in this application is deployed on a drone to detect road cracks. The drone flies at a speed of 5 m / min. The input image is collected by the drone and preprocessed to 640×640 pixels before being input into the model. The batch processing size is set to 40. During the detection process, the model automatically completes Conv_nbn preprocessing, feature extraction, feature fusion and detection output. (3) Application Results: With a P100 GPU and the UAV-PDD2023 dataset, YOLOv11 achieved an accuracy of 0.84, a recall of 0.7, an mAP50 of 0.744, and an mAP50-95 of 0.453. The model in this application achieved an accuracy of 0.889, a recall of 0.788, an mAP50 of 0.846, and an mAP50-95 of 0.579. The model significantly improved the mAP50 and mAP50-95 for various road defects; results for different categories are shown in Table 1.

[0051] Table 1. Comparison of Results ; (4) Application prospects: The method of this application can be widely used in defect detection scenarios of industrial products such as metal plates, glass, and PCB boards, and is especially suitable for batch detection of small-sized defects (such as holes, cracks, and pinholes).

[0052] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A road defect detection algorithm based on unmanned aerial vehicle (UAV) inspection, characterized in that, include: S1, Decoupled Normalized Convolution Operation: The road image acquired by the UAV is input into the improved YOLOv11 model, and the Conv_nbn module is used to perform a decoupled normalized convolution operation on the input road image. The improved YOLOv11 model refers to the YOLOv11 model with the addition of the Conv_nbn module, the C3k2_WTConv module, and the HEADF module. The YOLOv11 model includes an input end, a backbone network, a neck feature fusion network, and a detection head connected in sequence. The Conv_nbn module is set at the input end to replace the original input convolution module and perform decoupled normalized convolution processing on the input road image. The C3k2_WTConv module is set in the neck feature fusion network to introduce wavelet convolution feature enhancement in the fusion stage after upsampling and feature concatenation. The HEADF module is set between the output end of the neck feature fusion network and the detection head to perform frequency domain enhancement on the fused features of adjacent scales before feeding them into the detection head for prediction. S2, Backbone Network Feature Extraction: The image after decoupling and normalizing convolution enters the backbone network for feature extraction. That is, the image after decoupling and normalizing convolution passes through the Conv and C3k2 basic modules for layer-by-layer downsampling and semantic feature extraction, resulting in multiple feature maps of different resolutions. S3, Neck layer expansion and multi-scale feature fusion: The Neck layer is expanded from four layers to six layers; the deep high semantic features of the multi-layer features extracted by the backbone network are fused with the shallow high resolution features; at the same time, wavelet convolution is added to the C3k2_WTConv module in the neck to enhance the fusion effect. After multiple upsampling and feature concatenation (Upsample+Concat), a multi-scale fused feature map is constructed. S4, Frequency Domain Enhanced Fusion: Two multi-scale fused feature maps from adjacent layers are input into the HEADF module. The HEADF module uses wavelet transform to decompose the input features into low-frequency structural information and high-frequency detail information; whereby the low-frequency structural information is used to characterize the overall structure of the road area, and the high-frequency detail information is used to characterize the detailed response of crack edges and pothole contours. S5, Multi-scale detection output: The features processed by HEADF are input into the detection head. The detection head performs classification prediction and bounding box regression on the enhanced features at different scales to generate road defect categories, confidence scores, and target bounding box positions. Finally, the model outputs complete road defect detection results.

2. The road defect detection algorithm based on UAV inspection as described in claim 1, characterized in that, The decoupling normalization formula in step S1 is: （1） in, Point to The unit vector obtained by L2 normalization retains only the directional information of γ, eliminating the difference in numerical magnitude. It is a learnable scalar parameter that uniformly controls the amplitude of γ across all channels; Point to The unit vector obtained by L2 normalization retains only the directional information of β, eliminating the difference in numerical magnitude. These are learnable scalar parameters that uniformly control the magnitude of β across all channels; the change in the gradient formula of the loss function demonstrates that decoupling normalization achieves more uniform weights. and bias β, to For example, its gradient calculation formula is as follows: （2） in, It is a positive scaling factor across different channels; , is the key adjustment term in the gradient formula; It is a loss function For parameters The gradient; When the value is greater than 0, the loss function will affect... Punishment is imposed, but the severity of punishment varies depending on the category; the more frequent the category, the greater the severity of the punishment. When <0, the loss function will affect Encouragement is provided, but the degree of encouragement varies across different categories. The more frequent the category, the greater the degree of encouragement. This method generates a more balanced weight and bias.

3. The road defect detection algorithm based on UAV inspection as described in claim 1, characterized in that, The operation method of the C3k2_WTConv module is as follows: when C3k=False, the input features are first reduced in dimensionality by initial convolution, and then divided into two parallel branches by the Split operation. One branch is directly retained, and the other branch is processed by several WTBottleneck modules for multi-component feature extraction. Finally, the two features are fused by Concat and the output is integrated by the final convolution. When C3k=True, the convolution result is split as a preliminary process, but the subsequent branch is replaced with WT_C3k which incorporates wavelet transform. Finally, feature fusion and output are completed through Concat and the final convolution.

4. The road defect detection algorithm based on UAV inspection as described in claim 3, characterized in that, The WT_C3k module operates as follows: the input features are first divided into two parallel processing branches. The first branch performs channel compression through 1×1 convolution and then inputs a feature extraction unit composed of multiple WTBottlenecks. In each WTBottleneck, the features are first processed by convolution to extract local spatial information, and then enhanced in the frequency domain by WTConv wavelet convolution to obtain low-frequency structural information and high-frequency edge detail information. The stability of feature propagation is maintained through residual connections. The second branch directly extracts and retains the shallow semantic information of the input features through a 1×1 convolution. Then, the deep enhanced features output by the first branch and the shallow retained features output by the second branch are concatenated along the channel dimension, and channel fusion is completed through a terminal 1×1 convolution, finally outputting the fused feature representation.

5. The road defect detection algorithm based on UAV inspection as described in claim 1, characterized in that, The HEADF module operates as follows: Two adjacent-scale features, X1 and X2, are simultaneously input into the HEADF module. First, the two input features are size-aligned to meet the subsequent fusion requirements in both channel and spatial dimensions. Then, wavelet transforms are performed on X1 and X2 respectively, decomposing each feature into a low-frequency component LL and high-frequency components LH, HL, and HH. LL primarily preserves the overall structure and semantic information of the road area, while LH, HL, and HH primarily preserve details such as crack edges, pothole outlines, and local texture changes. Next, in the frequency domain feature reconstruction stage, the HEADF module selects the high-frequency component HH and adds it to the low-frequency component LL to obtain the LL+HH frequency domain feature, which simultaneously represents the overall road structure information and strong edge detail information. The HL and LH directional high-frequency components are also preserved to describe crack edge responses in different directions. Subsequently, HL, LH, and LL+HH are used as three main frequency domain feature branches, and global max pooling is applied to obtain the corresponding global response features, denoted as X. HL X LH and X LL+HH Among them, X HL and X LH X is used to highlight the crack edge response in different directions. LL+HH Used to simultaneously preserve overall road structure information and strong edge detail information; Then, in the feature interaction phase, the HEADF module uses X LL+HH As a comprehensive guiding feature of structure and edge, respectively related to X HL and X LH Matrix multiplication yields X hl and X lh To establish the correlation between directional high-frequency details and structural enhancement features; Finally, the X obtained through interaction hl and X lh After dimensionality transformation, the features are concatenated and corresponding enhancement weights are generated through activation functions. These enhancement weights are multiplied by the original input features X1 and X2 respectively to achieve adaptive enhancement of the effective disease area. Finally, the two enhanced features are added and fused, and local features are integrated through 3×3 convolution to output the frequency domain enhanced fused feature map.

Images