A highway pavement technical condition inspection disease identification method

Abstract

The application discloses a highway pavement technical condition inspection disease identification method and relates to the technical field of highway detection; the method realizes automatic synchronous data collection through distance triggering; an FCN-YOLOv8 fusion disease identification algorithm is used, P2-P5 multi-scale feature maps are extracted through a shared CSPDarkNet backbone network, and after feature sharing to the C3 stage, the features are divided into FCN segmentation branches and YOLOv8 detection branches for parallel processing; the EMA-C2f feature enhancement module, the improved PIoU loss function and the special loss function group for the diseases of facilities along the line are combined to improve the identification precision, and the FCN frame-by-frame execution and YOLOv8 real-time running time-sharing processing mechanism is adopted to adapt to the vehicle-mounted edge computing environment; finally, pixel-level masks and detection results are fused to generate a standardized detection report; the application realizes high-precision, real-time and full-element inspection of the diseases of the highway pavement and the facilities along the line, has high detection precision and fast operation speed, can effectively reduce the workload of data processing in the office, and improves the highway maintenance inspection efficiency.

Description

Technical Field

[0001] This invention relates to the field of highway pavement technical condition detection and highway roadside facility inspection technology. More specifically, it is a deep learning-based optical image analysis method for identifying defects in highway pavement technical condition detection, and is suitable for vehicle-mounted edge computing environments. Background Technology

[0002] For a long time, my country's highway industry has held the concept of "reconstruction over maintenance," which has led to the overall late start of my country's highway maintenance industry compared to Europe and the United States. Furthermore, the maintenance tasks for many early-built highways are increasing year by year, making maintenance a challenging undertaking. Currently, highway maintenance development needs to enter a stage where it is moderately ahead of economic and social development. Therefore, the green maintenance concept of "intensive, professional, energy-saving, and environmentally friendly" will be the main direction for the upgrading of my country's highway maintenance development in the future.

[0003] To ensure the rationality and compliance of highway maintenance, various localities have begun to strengthen the inspection of highway pavement technical conditions and the daily patrol of roadside facilities. The core purpose is to understand the current real-time information of the road area, so as to optimize maintenance plans and improve the effectiveness of daily patrols. In response to this need, pavement technical condition inspection equipment has emerged.

[0004] There are various methods for road surface detection and inspection, including physical analysis and optical analysis, with optical imaging analysis being one of the most common methods currently in use. Optical imaging analysis offers high accuracy and a short lifespan, but existing optical imaging detection devices have the following drawbacks: 1. The degree of automation is relatively low, requiring staff to manually control the opening and closing of the camera, and special areas need to be manually marked in the software; 2. Data obtained from on-site inspections needs to be manually processed and analyzed at the workplace. Manual data processing and analysis is inefficient and cannot quickly complete the annotation and analysis of inspection data in batches. 3. Existing AI detection algorithms mostly use a single model, such as FCN or YOLO. FCN focuses on pixel-level segmentation, but has a large computational load and insufficient real-time performance. YOLO is good at real-time target detection, but has limited accuracy in identifying thin cracks and small target defects. When used alone, it is difficult to achieve both precise localization and rapid identification of road damage. 4. Existing algorithms are mainly designed for road surface defects, and their ability to identify defects in roadside facilities (such as guardrails, signs, poles, boxes, drainage facilities, etc.) is poor, which cannot meet the needs of full-element road inspection. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying road surface defects. By combining the FCN-YOLOv8 dual-branch fusion architecture with a dedicated vehicle-mounted inspection platform, it can achieve high-precision, real-time identification of road surface and roadside facility defects, solving the problems of low automation, low processing efficiency, susceptibility to background interference, and the inability of a single model to balance accuracy and real-time performance in traditional detection methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying road surface defects uses a mobile vehicle as an inspection platform, equipped with a fixed frame, an optical acquisition module, an optical ranging module, an optical scanning analysis module, a BeiDou positioning acquisition module, a distance-triggered acquisition device, an integrated controller, an acquisition control computer, and an uninterruptible power supply. The optical acquisition module, optical ranging module, optical scanning analysis module, BeiDou positioning acquisition module, and distance-triggered acquisition device are all mounted on the fixed frame and electrically connected to the integrated controller. The integrated controller is communicatively connected to the acquisition control computer. The method includes the following steps: Step 1: The distance-triggered acquisition device detects the vehicle's movement distance in real time. When the vehicle moves a preset distance, the integrated controller is automatically triggered to synchronously drive the optical acquisition module, optical ranging module, optical scanning analysis module, and Beidou positioning acquisition module to collect optical image data, distance data, and Beidou positioning data of the road surface and roadside facilities. Step 2: The acquisition and control computer receives and stores the above-mentioned acquired data, performs preprocessing operations such as noise reduction and cropping on the image data, and summarizes the basic detection data of the left and right road widths; Step 3: Input the preprocessed image into the FCN-YOLOv8 fusion disease identification algorithm deployed on the acquisition control computer. The P2-P5 multi-scale feature maps are extracted by sharing the C-step PDarkNet backbone network. After the features are shared to the C3 stage, they are divided into the FCN segmentation branch and the YOLOv8 detection branch. Step 4: The FCN segmentation branch and the YOLOv8 detection branch process the multi-scale feature maps in parallel, and output the pixel-level mask of the disease and the bounding box, category and confidence information of the disease respectively. Step 5: Perform mask-detection fusion between the pixel-level mask output by the FCN segmentation branch and the bounding box and category information output by the YOLOv8 detection branch to obtain a disease quantification result that includes the location, type and area of ​​the disease. Step 6: Summarize the quantitative results of defects from multiple road sections and generate the final highway pavement technical condition inspection report.

[0007] Furthermore, in step 3, the shared feature extraction satisfies the formula: ; Where I is the input image, The i-th layer is the feature map; the P2 layer is dedicated to subpixel crack detection, and the backbone network can be replaced with MobileOne for lightweight deployment.

[0008] Furthermore, in step 4, the FCN segmentation branch uses a transposed convolution + skip connection decoding structure to recover spatial details, and enhances the crack feature channel weights and suppresses asphalt texture noise through the EMA-C2f feature enhancement module. The segmentation output satisfies the formula: ; Where M is the pixel-level mask. To share the multi-scale feature maps output by the backbone network.

[0009] Furthermore, in step 4, the YOLOv8 detection branch outputs detection results based on the mid-layer features of P3-P5, adjusts the anchor frame scale to 32×32 to adapt to the crack aspect ratio, and optimizes the Task-AlignedAssignments loss. The overall loss satisfies the formula: L total =λ bbox ·L bbox +λ cls ·L cls +λ dfl ·L dfl; Default weight λ bbox =6.5、λ cls =0.6、λ dfl =1.4.

[0010] Furthermore, the YOLOv8 branch employs an improved PIOU loss to optimize the positioning accuracy of slender cracks. The PIOU loss satisfies the formula: ;

[0011] Where C gt Let d be the coordinates of the center point of the bounding box. 2 diagonal This is the length of the diagonal.

[0012] Furthermore, a bidirectional feature guidance mechanism is adopted between the two branches to realize feature interaction and enhancement, and the loss weights of the segmentation task and the detection task are dynamically adjusted through an adaptive multi-task balancer.

[0013] Furthermore, the algorithm introduces a set of dedicated loss functions for roadside facility defects. These dedicated loss functions are different from the road surface defect loss functions and are specifically designed for the morphological characteristics, size distribution, and occlusion of roadside facilities such as guardrails, signs, poles, boxes, and drainage facilities. This is used to improve the classification and location accuracy of roadside facility defects.

[0014] Furthermore, the weight coefficients of the overall loss function can be dynamically and adaptively adjusted based on the training accuracy according to the different road types in different regions.

[0015] Furthermore, the FCN segmentation branch adopts an inter-frame execution mode, while the YOLOv8 detection branch adopts a real-time execution mode, forming a time-sharing processing mechanism to reduce the memory consumption of the dual-branch computation superposition and adapt to the vehicle edge computing hardware resources.

[0016] Furthermore, in step 6, the detection data from multiple road sections are summarized, evaluated, and comprehensively analyzed to automatically generate a standardized highway pavement technical condition inspection report that includes defect statistics, road condition rating, and maintenance recommendations.

[0017] Compared with the prior art, the present invention has the following significant advantages: 1. This invention achieves fully automated data acquisition through a distance-triggered acquisition device, eliminating the need for manual control of the shooting start and stop. At the same time, the acquisition control computer can perform AI data analysis in real time on-site, greatly reducing the workload of internal data processing after the inspection is completed and improving inspection efficiency.

[0018] 2. The FCN-YOLOv8 dual-branch fusion architecture proposed in this invention reduces computational redundancy by sharing the CSPDarkNet backbone network. Combining the pixel-level segmentation capability of FCN with the efficient detection capability of YOLOv8, it achieves high-precision disease identification. On the RDD2022 dataset, the mIoU reaches 87.2%, a 106% improvement over pure YOLO.

[0019] 3. In view of the characteristics of slender cracks and small-target defects in road surfaces, this invention designs an EMA-C2f feature enhancement module and an improved PIOU loss function, which effectively enhances the weight of crack feature channels, suppresses asphalt texture noise, and improves the positioning accuracy of slender cracks.

[0020] 4. This invention introduces a set of loss functions specifically designed for roadside facility defects, which are optimized to address the defect characteristics of roadside facilities. This solves the problem of poor defect identification performance of existing algorithms and enables full-element inspection of highways.

[0021] 5. This invention designs a time-sharing processing mechanism that combines FCN segmentation branch execution across frames with real-time execution of YOLOv8 detection branch, reducing the theoretical peak memory consumption of dual-branch parallel computing by about 40%-50%, enabling the algorithm to run stably on mainstream vehicle edge computing devices, with inference speed maintained above 20 FPS, meeting the requirements of real-time inspection of highway surfaces. Attached Figure Description

[0022] Figure 1This is a flowchart of the overall FCN-YOLOv8 fusion algorithm described in this invention; Figure 2 This is a comparison chart of the efficiency of loss optimization described in this invention; Figure 3 This is a schematic diagram of the overall structure of the inspection platform described in this invention; Figure 4 This is a diagram showing the overall structure of the YOLOv8 model described in this invention.

[0023] Reference numerals: 1-Mobile vehicle, 2-Uninterruptible power supply, 3-Acquisition and control computer, 4-Distance trigger acquisition device, 5-Optical ranging module, 6-Optical scanning and analysis module, 7-BeiDou positioning acquisition module, 8-Fixed frame, 9-Optical acquisition module, 10-Wireless gateway, 11-Outer shell. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 4 The highway pavement technical condition inspection device of the present invention includes a mobile vehicle 1, an uninterruptible power supply 2, a data acquisition and control computer 3, a distance triggering data acquisition device 4, an optical ranging module 5, an optical scanning analysis module 6, a Beidou positioning data acquisition module 7, a fixed frame 8, an optical data acquisition module 9, a wireless gateway 10, and a housing 11. The fixed frame 8 is set on the top of the mobile vehicle 1. The optical ranging module 5, the optical scanning analysis module 6, the Beidou positioning data acquisition module 7, and the optical data acquisition module 9 are all installed on the fixed frame 8. The housing 11 covers the fixed frame 8 and each data acquisition module, providing waterproof, dustproof, and collision-proof protection. The uninterruptible power supply 2, the data acquisition and control computer 3, and the wireless gateway 10 are set inside the carriage of the mobile vehicle 1.

[0026] The optical acquisition module 9 and the optical scanning analysis module 6 are electrically connected to the integrated controller, which controls the synchronous operation of each module. The distance-triggered acquisition device 4 is also electrically connected to the integrated controller. When the vehicle moves a preset distance, the distance-triggered acquisition device 4 sends a trigger signal, and the integrated controller controls the optical ranging module 5, the optical scanning analysis module 6, the optical acquisition module 9, and the Beidou positioning acquisition module 7 to operate synchronously. The acquisition control computer 3 communicates with each module via electrical and network connections, receiving and processing the acquired data. The wireless gateway 10 is used to realize data transmission between the acquisition control computer 3 and the remote server, enabling the real-time uploading of inspection reports to the highway maintenance management platform.

[0027] This invention provides a method for identifying road surface defects, the specific steps of which are as follows: Step 1: Automatic Data Collection The mobile vehicle 1, equipped with the aforementioned inspection device, is driven to the section of highway to be inspected, and the inspection system is activated. The distance-triggered acquisition device 4 monitors the vehicle's movement distance in real time. When the vehicle moves a preset distance (e.g., 1 meter), the integrated controller is automatically triggered, synchronously driving the optical acquisition module 9, optical ranging module 5, optical scanning analysis module 6, and BeiDou positioning acquisition module 7 to collect optical image data, distance data, and BeiDou positioning data of the highway surface and roadside facilities. The standard resolution of the images acquired by the optical acquisition module 9 is 2048×1024, which can clearly capture minute cracks in the road surface and detailed features of roadside facilities. Step 2: Data Preprocessing and Summarization The acquisition and control computer 3 receives and stores the acquired data, performs preprocessing operations such as denoising, cropping, and grayscale correction on the image data to remove noise and invalid areas in the image, and summarizes the basic detection data of the left and right road widths to prepare for subsequent AI analysis.

[0028] Step 3: Feature Extraction of Shared Backbone Network The preprocessed image is input into the FCN-YOLOv8 fusion disease identification algorithm deployed on the acquisition and control computer 3. The algorithm first extracts multi-scale feature maps through a shared backbone network. The general mathematical model for feature extraction is: ; Where I is the input image, This is a set of multi-scale feature maps, where H is the feature map height, W is the feature map width, and C is the number of feature map channels.

[0029] This invention employs CSPDarkNet as a shared backbone network to extract multi-scale feature maps from layers P2 to P5, specifically satisfying the formula: ; in The i-th layer is the feature map; the P2 layer is dedicated to sub-pixel crack detection and can capture the most subtle crack features of the road surface; after the features are shared to the C3 stage, they are divided into the FCN segmentation branch and the YOLOv8 detection branch. This design can effectively reduce computational redundancy and improve the algorithm's running efficiency.

[0030] As an optional lightweight deployment solution, the shared CSPDarkNet backbone network can be replaced with MobileOne. MobileOne is a high-efficiency backbone network designed specifically for mobile devices, characterized by its small parameter count and fast inference speed. After the replacement, the number of algorithm parameters is reduced by approximately 60%, the inference speed is improved by approximately 30%, and the mIoU decreases by only approximately 1.5%, which can meet the application scenarios with higher real-time requirements.

[0031] S4: Two-branch parallel processing The FCN segmentation branch and the YOLOv8 detection branch process multi-scale feature maps in parallel. A bidirectional feature guidance mechanism is used between the two branches to achieve feature interaction and enhancement; that is, features from the FCN segmentation branch can guide the YOLOv8 detection branch to focus on disease areas, and features from the YOLOv8 detection branch can guide the FCN segmentation branch to optimize segmentation boundaries. Simultaneously, an adaptive multi-task balancer dynamically adjusts the loss weights of the segmentation and detection tasks, enabling the two branches to optimize collaboratively and preventing either task from dominating the training process.

[0032] FCN branching processing The FCN segmentation branch utilizes shallow high-resolution features and deep semantic features, recovering spatial details through a transposed convolutional layer and skip connections to generate an initial pixel-level mask. Then, an EMA-C2f feature enhancement module strengthens the weights of crack feature channels and suppresses asphalt texture noise, ultimately generating a precise pixel-level mask. The segmentation output satisfies the formula: ; Where M is the pixel-level mask. To share the multi-scale feature maps output by the backbone network.

[0033] The FCN segmentation branch is trained using a pixel-level cross-entropy loss function to measure the difference between the predicted segmentation map and the ground truth label. For each pixel, the loss function is defined as: ; Where N is the total number of pixels, C is the number of categories, and y i,c It is a true label encoded in one-hot encoding, p i,c It is the probability that pixel i belongs to category c as predicted by the model.

[0034] YOLOv8 Branch Detection Processing The YOLOv8 detection branch enhances multi-scale features through Neck modules such as PANet / FPN, and uses the detection head to predict target bounding boxes, categories, and confidence scores. Based on the mid-layer features of P3-P5, it outputs bounding boxes and categories, adjusts the anchor box scale to 32×32 to fit the crack aspect ratio, and optimizes the Task-AlignedAssignments loss to improve the accuracy of target assignment.

[0035] The YOLOv8 detection branch uses a combination of Distributed Focal Loss (DFL) and CIoU loss as the basic bounding box regression loss. The overall loss satisfies the formula: L total =λ bbox ·L bbox +λ cls ·L cls +λ dfl ·L dfl; Default weight λ bbox =6.5、λ cls =0.6、λ dfl =1.4.

[0036] YOLOv8 uses a combination of binary cross-entropy loss and focus loss: ; Among them, the focus loss adjustment factor is: ; Distributed Focus Loss (DFL) formula: ; Meanwhile, this invention employs an improved PIOU loss to optimize the positioning accuracy of slender cracks, and the PIOU loss satisfies the formula: ; Where C gt Let d be the coordinates of the center point of the bounding box. 2 diagonal The length of the diagonal; Finally, the bounding box and category output are completed, which facilitates mask-detection fusion; the PioU loss can effectively solve the problem of inaccurate localization of slender targets by traditional IoU loss, and significantly improve the detection accuracy of slender cracks in the road surface.

[0037] In addition, the algorithm introduces a set of loss functions specifically for roadside facility defects. This set of loss functions is different from the road surface defect loss function. It is specially designed for the morphological characteristics, size distribution and occlusion of roadside facilities such as guardrails, signs, poles, boxes, and drainage facilities. It integrates classification loss, bounding box regression loss, occlusion adaptive loss and small target enhancement loss, which effectively improves the classification and location accuracy of roadside facility defects.

[0038] The weight coefficients of the overall loss function can be dynamically and adaptively adjusted based on the training accuracy, according to different road types in different regions. For example, for asphalt pavement, the weight of the loss related to crack detection can be appropriately increased; for cement pavement, the weight of the loss related to pothole and broken slab detection can be appropriately increased; and for mountain roads, the weight of the loss related to roadside facility defects detection can be appropriately increased.

[0039] Time-sharing mechanism To address the limited hardware resources of automotive edge computing, this invention designs a time-sharing processing mechanism that combines FCN segmentation branch execution across frames with real-time execution of the YOLOv8 detection branch. This effectively solves the technical problems of excessive memory consumption and inference speed failing to meet real-time detection requirements caused by parallel computation of dual branches. The specific implementation is as follows: The acquisition and control computer 3 sequentially numbers the image frames continuously acquired by the optical acquisition module 9 to form a frame sequence. ; The YOLOv8 detection branch processes all numbered image frames in real time, that is, it performs real-time processing on all image frames. Each frame performs target detection, outputting the defect bounding box, category, and confidence information. This ensures real-time detection capabilities for targets requiring rapid response, such as sudden road anomalies and roadside facility defects. FCN segmentation branches only process image frames with odd-numbered frames. The process involves performing a pixel-level segmentation operation once every frame to generate a pixel-level mask for the disease. For intermediate frames not processed by the FCN branch (such as...) etc.), the system reuses the previous frame ( The segmentation results of (etc.) are fused with the output of the YOLOv8 detection branch of the current frame. Since the vehicle-mounted inspection vehicle travels at a stable speed (usually 30-60 km / h), the content changes little between adjacent frames, and this reuse method does not significantly reduce the accuracy of disease identification. The system can dynamically adjust the frame interval of the FCN branch according to the real-time load of the vehicle edge computing hardware. When the hardware load is lower than the preset threshold, the frame interval can be adjusted to 1 (i.e., full frame processing). When the hardware load is higher than the preset threshold, the frame interval can be adjusted to 2 or 3 to ensure the stable operation of the system.

[0040] Through the above-mentioned time-sharing processing mechanism, this invention reduces the theoretical peak memory consumption of dual-branch parallel computing by about 40%-50%, enabling the algorithm to run stably on mainstream vehicle edge computing devices (such as NVIDIA Jetson Xavier NX, Rockchip RK3588, etc.), with an inference speed of over 20 FPS, meeting the requirements of real-time inspection of highway surfaces.

[0041] Step 5: Masking-Detection Fusion The pixel-level mask output from the FCN segmentation branch is fused with the bounding box and category information output from the YOLOv8 detection branch using a mask-detection fusion method. The specific steps are as follows: Map the segmentation map and the detection box to the same coordinate system (2048×1024); Instance-based segmentation logic: Use detection boxes to crop the segmentation map and generate an accurate mask for each detected target (similar to MaskR-CNN). Based on panoramic segmentation logic: merge semantic segmentation map with instance detection results to distinguish between "countable objects" (from YOLO) and "background / uncountable regions" (from FCN); The segmentation results are used to optimize the detection box boundaries and add accurate labels to the identified targets.

[0042] The final result is a quantitative assessment of the disease, including its location, type, area, and severity.

[0043] Step Six: Generating the Test Report The data acquisition and control computer 3 aggregates the quantitative results of road surface defects from multiple road sections, summarizes, evaluates, and comprehensively analyzes the detection data from multiple road sections, and automatically generates a standardized highway pavement technical condition inspection report that includes defect statistics, road condition rating, and maintenance recommendations. The report can be directly exported as PDF or Excel format, or uploaded to the remote highway maintenance management platform via the wireless gateway 10, providing data support for highway maintenance decision-making.

[0044] The present invention was tested on the RDD2022 international standard pavement distress dataset. The test results show that the mIoU of the proposed FCN-YOLOv8 fusion distress identification algorithm reaches 87.2%, which is 106% higher than that of pure YOLOv8. At the same time, the inference speed is maintained above 20 FPS, achieving a good balance between detection accuracy and real-time performance.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying road surface defects during technical condition inspection, characterized in that, A mobile vehicle equipped with a fixed frame, an optical acquisition module, an optical ranging module, an optical scanning analysis module, a Beidou positioning acquisition module, a distance-triggered acquisition device, an integrated controller, an acquisition control computer, and an uninterruptible power supply is used as the inspection platform. The optical acquisition module, optical ranging module, optical scanning analysis module, Beidou positioning acquisition module, and distance-triggered acquisition device are all mounted on the fixed frame and electrically connected to the integrated controller. The integrated controller is communicatively connected to the acquisition control computer. The defect identification method includes the following steps: S1. Automatic data acquisition: The distance-triggered acquisition device detects the vehicle's movement distance in real time. When the vehicle moves a preset distance, the integrated controller is automatically triggered to synchronously drive the optical acquisition module, optical ranging module, optical scanning analysis module, and Beidou positioning acquisition module to work, and to collect optical image data, distance data, and Beidou positioning data of the road surface and roadside facilities. S2. Data Preprocessing and Summarization: The acquisition control computer receives and stores the above-mentioned acquired data, performs preprocessing operations such as noise reduction and cropping on the image data, and summarizes the basic detection data of the left and right road widths. S3. Feature extraction via shared backbone network: The preprocessed image is input into the FCN-YOLOv8 fusion disease identification algorithm deployed on the acquisition control computer. The P2-P5 multi-scale feature maps are extracted through the shared CSPDarkNet backbone network. After the features are shared to the C3 stage, they are divided into the FCN segmentation branch and the YOLOv8 detection branch. S4. Parallel processing with two branches: The FCN segmentation branch and the YOLOv8 detection branch process the multi-scale feature maps in parallel, and output the pixel-level mask of the disease and the bounding box, category and confidence information of the disease respectively. S5, Mask-Detection Fusion: The pixel-level mask output by the FCN segmentation branch is fused with the bounding box and category information output by the YOLOv8 detection branch to obtain a quantitative result of the disease including the location, type and area of ​​the disease. S6. Report Generation: Summarize the quantitative results of defects from multiple road sections and generate the final road surface technical condition inspection report.

2. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, The shared feature extraction described in step S3 satisfies the following formula: ； Where I is the input image, The i-th layer feature map; the P2 layer is dedicated to subpixel crack detection, and the backbone network is replaced with MobileOne for lightweight deployment.

3. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, In step S4, the FCN segmentation branch uses a transposed convolutional fusion skip connection decoding structure to recover spatial details, and enhances the crack feature channel weights and suppresses asphalt texture noise through the EMA-C2f feature enhancement module. The segmentation output satisfies the formula: ； Where M is the pixel-level mask. To share the multi-scale feature maps output by the backbone network.

4. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, In step S4, the YOLOv8 detection branch outputs detection results based on the features of the P3-P5 layers, adjusts the anchor frame size to 32×32 to match the crack aspect ratio, and optimizes the Task-AlignedAssignments loss. The overall loss satisfies the formula: L total =λ bbox ·L bbox +λ cls ·L cls +λ dfl ·L dfl ; Default weight λ bbox =6.5、λ cls =0.6、λ dfl =1.

4.

5. The method for identifying road surface defects during technical condition inspection according to claim 4, characterized in that, The YOLOv8 branch uses an improved PIOU loss to optimize the positioning accuracy of slender cracks. The PIOU loss satisfies the following formula: ； Where C gt The coordinates of the center point of the bounding box. This is the length of the diagonal.

6. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, A bidirectional feature guidance mechanism is used between the two branches to achieve feature interaction and enhancement, and an adaptive multi-task balancer is used to dynamically adjust the loss weights of the segmentation task and the detection task.

7. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, The algorithm introduces a set of dedicated loss functions for roadside facility defects. These dedicated loss functions are different from the road surface defect loss functions and are specifically designed for the morphological characteristics, size distribution, and occlusion of roadside facilities such as guardrails, signs, poles, boxes, and drainage facilities. This is used to improve the classification and location accuracy of roadside facility defects.

8. The method for identifying road surface defects during technical condition inspection according to claim 4, characterized in that, The weight coefficients of the overall loss function can be dynamically and adaptively adjusted based on the training accuracy, according to the different road types in different regions.

9. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, The FCN segmentation branch uses an inter-frame execution mode, while the YOLOv8 detection branch uses a real-time execution mode, forming a time-sharing processing mechanism to reduce the memory consumption of the dual-branch computation superposition and adapt to the vehicle edge computing hardware resources.

10. The method for identifying road surface defects during technical condition inspection according to claim 1, characterized in that, In step S6, the detection data of multiple road sections are summarized, evaluated and analyzed to automatically generate a standardized highway pavement technical condition inspection report that includes defect statistics, road condition rating and maintenance recommendations.

Images