Intelligent tunnel disease inspection method based on multi-model fusion and consistency verification

Abstract

The invention belongs to the technical field of tunnel engineering safety monitoring, and relates to a tunnel disease intelligent inspection method based on multi-model fusion and consistency verification. A first deep learning model is used for outputting a fixed boundary disease candidate frame, a second deep learning model is used for outputting a non-fixed boundary disease pixel-level mask, and in combination with consistency judgment of spatial position association, the limitation that a single model cannot give consideration to recognition of different forms of diseases is broken through; comprehensive and accurate identification of discrete targets such as chipping and foreign matters and irregular region diseases such as water leakage and water accumulation is realized, and the breadth and precision of disease identification are remarkably improved; and thirdly, through space-time consistency verification, space filtering is performed based on a preset monitoring area, and then the stability of a candidate result is analyzed and verified by using a time sequence of a sliding window, so that false alarms caused by instantaneous environmental interference such as vehicle lamps and light and shadow changes in the tunnel are effectively filtered out, and the reliability of an inspection result is greatly improved.

Description

Technical Field

[0001] This application belongs to the field of tunnel engineering safety monitoring technology, and more specifically, it relates to an intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification. Background Technology

[0002] As a key node in the transportation network, the structural safety of tunnels directly affects the safety of traffic operations, making tunnel structural health monitoring crucial. Currently, tunnel structural health monitoring mainly relies on regular manual inspections and traditional video surveillance systems. Regular manual inspections are not only inefficient and costly, but also susceptible to subjective influences and cannot achieve 24 / 7 real-time monitoring. Furthermore, they suffer from significant delays in responding to sudden hazards (such as rockfalls). Traditional video surveillance systems typically employ pixel-based motion detection technology for anomaly detection. However, this technology lacks semantic understanding of the video content, often misinterpreting normal phenomena such as vehicles driving normally within the tunnel, changes in light and shadow caused by headlights, and water vapor condensation as anomalies. This leads to a flood of alarms with little practical value, placing a heavy burden on monitoring personnel.

[0003] In recent years, the single deep learning target detection algorithm inspection scheme has improved the level of automation to some extent, but its model has poor versatility. It can only identify defects with clear outlines and fixed shapes (such as falling blocks and foreign objects), and it is difficult to accurately identify defects with irregular shapes and no fixed boundaries, such as water seepage and road water accumulation. In addition, it is prone to false alarms due to interference from complex lighting conditions in tunnels. Furthermore, it cannot quantitatively assess the severity of defects, which is not conducive to operators making scientific maintenance decisions and risk level assessments. Summary of the Invention

[0004] This invention provides an intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification, which aims to solve the technical problems mentioned in the background.

[0005] The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification includes the following steps: Step 1: Acquire the monitoring video stream inside the tunnel and preprocess it to obtain standardized video frames; Step 2: Start two parallel deep learning models to analyze the standardized video frames. The first deep learning model outputs candidate boxes of lesions with fixed boundary features, and the second deep learning model outputs pixel-level masks of lesions without fixed boundary features. Then, based on the spatial position correlation between the two, a consistency judgment is made to obtain the preliminary fusion candidate results. Step 3: Spatial filtering is performed on the preliminary fusion candidate results to retain the candidate results located within the preset monitoring area. Then, the stability of the spatially filtered candidate results is verified by time series analysis to obtain stable suspected disease events. Step 4: Based on the preset camera calibration parameters, establish the mapping relationship between pixel coordinates and physical world coordinates, and convert the pixel dimension information corresponding to stable suspected disease events into actual physical scale information; Step 5: Compare the actual physical scale information with the preset disease threshold standard to determine whether an alarm is triggered. If an alarm is triggered, generate structured alarm information, encapsulate the structured alarm information into a message conforming to the message queue telemetry transmission protocol, and publish it to the monitoring platform.

[0006] This invention employs a multi-model parallel inference and preliminary fusion strategy. A first deep learning model outputs candidate bounding boxes for defects with fixed boundaries, while a second deep learning model outputs pixel-level masks for defects without fixed boundaries. Combined with consistency judgment based on spatial location correlation, this overcomes the limitation of a single model in recognizing defects of different morphologies. It achieves comprehensive and accurate identification of discrete targets such as falling blocks and foreign objects, as well as defects in irregular areas such as water leakage and water accumulation, significantly improving the breadth and accuracy of defect identification. Furthermore, through spatiotemporal consistency verification, spatial filtering is first performed based on a preset monitoring area, and then time-series analysis using a sliding window is used to verify the stability of the candidate results. This effectively filters out false alarms caused by instantaneous environmental interference such as vehicle headlights and changes in light and shadow within the tunnel. This significantly improves the reliability of inspection results. Simultaneously, by establishing a mapping relationship between pixel coordinates and physical world coordinates through physical scale quantification, pixel dimensional information is converted into actual physical size information, enabling a quantitative assessment of the severity of defects. Finally, through alarm judgment and protocol encapsulation and release, the quantified results are compared with threshold standards to generate structured alarm information, which is then encapsulated into messages conforming to the message queue telemetry transmission protocol and released to the monitoring platform. This allows for seamless integration with existing IoT platforms or tunnel integrated monitoring systems without requiring large-scale modifications to the upper-level platform. It significantly enhances the engineering integration and scalability of the technical solution, driving the transformation of tunnel inspection from traditional manual, qualitative judgment to automated, precise, and quantitative intelligent inspection.

[0007] Preferably, the preprocessing includes video decoding, frame rate uniform adjustment, resolution resampling, image enhancement, and non-monitoring area masking. The image enhancement is used to improve the video frame quality in low-light, overexposed, or underexposed scenes, and the non-monitoring area masking is achieved by configuring a preset monitoring area mask.

[0008] Preferably, the first deep learning model is a lightweight object detection network, and the second deep learning model is a semantic segmentation network. The lightweight object detection network is used to identify defects such as missing blocks and foreign objects; the semantic segmentation network is used to identify defects such as water leakage and road surface water accumulation.

[0009] Preferably, the lightweight object detection network is a YOLO series network, and the semantic segmentation network is any one of the following: U-Net network, U-Net network with EfficientNet as the backbone, or DeepLabv3+ network.

[0010] Preferably, the consistency determination includes: When the candidate bounding box for a defect overlaps with the pixel-level mask in the same spatial location, defect candidate bounding box optimization or noise removal is used for processing. The optimization of the disease candidate box includes: adjusting the boundary coordinates of the disease candidate box based on the bounding rectangle of the pixel-level mask, so that the adjusted disease candidate box fits the actual disease area better; Noise removal includes: counting the number of valid pixels that meet the disease characteristics after being filtered by a preset mask threshold in the pixel-level mask, and judging the results whose mask area corresponding to the number of valid pixels is lower than a preset relative threshold as noise and removing them.

[0011] Preferably, the time series analysis is implemented by maintaining a sliding window, which is used to store candidate results from the most recent consecutive frames. The temporal stability of the candidate results is verified by judging the frequency of occurrence and positional overlap of candidate results of the same category and similar position within the sliding window.

[0012] Preferably, the camera calibration parameters are obtained using the Zhang Zhengyou calibration method, including camera intrinsic parameters, distortion coefficients, and a homography matrix estimated based on the set of planar feature points within the tunnel. The homography matrix is ​​used to realize the mapping transformation from pixel coordinates to physical world coordinates.

[0013] Preferably, the structured alarm information includes disease type, disease location information, disease physical scale information, model analysis confidence level, event timestamp, and event snapshot.

[0014] Preferably, the message publishing of the message queue telemetry transmission protocol adopts hierarchical topic management and supports a message retransmission mechanism.

[0015] Preferably, the intelligent tunnel defect inspection method is deployed on an edge computing device, and the deployment and optimization of two parallel deep learning models are carried out through an inference acceleration framework.

[0016] The beneficial effects of this invention include: This invention employs a multi-model parallel inference and preliminary fusion strategy. A first deep learning model outputs candidate bounding boxes for defects with fixed boundaries, while a second deep learning model outputs pixel-level masks for defects without fixed boundaries. Combined with consistency judgment based on spatial location correlation, this overcomes the limitation of a single model in recognizing defects of different morphologies. It achieves comprehensive and accurate identification of discrete targets such as falling blocks and foreign objects, as well as defects in irregular areas such as water leakage and water accumulation, significantly improving the breadth and accuracy of defect identification. Furthermore, through spatiotemporal consistency verification, spatial filtering is first performed based on a preset monitoring area, and then time-series analysis using a sliding window is used to verify the stability of the candidate results. This effectively filters out false alarms caused by instantaneous environmental interference such as vehicle headlights and changes in light and shadow within the tunnel. This significantly improves the reliability of inspection results. Simultaneously, by establishing a mapping relationship between pixel coordinates and physical world coordinates through physical scale quantification, pixel dimensional information is converted into actual physical size information, enabling a quantitative assessment of the severity of defects. Finally, through alarm judgment and protocol encapsulation and release, the quantified results are compared with threshold standards to generate structured alarm information, which is then encapsulated into messages conforming to the message queue telemetry transmission protocol and released to the monitoring platform. This allows for seamless integration with existing IoT platforms or tunnel integrated monitoring systems without requiring large-scale modifications to the upper-level platform. It significantly enhances the engineering integration and scalability of the technical solution, driving the transformation of tunnel inspection from traditional manual, qualitative judgment to automated, precise, and quantitative intelligent inspection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a system architecture diagram provided for an embodiment of the present invention.

[0019] Figure 2 The overall flowchart provided for embodiments of the present invention. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] See Figure 1 and Figure 2 As shown, the intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification specifically includes the following steps: Step 1: Connect to the front-end camera deployed in the tunnel via Real-Time Streaming Protocol (RTSP) to acquire real-time video stream. Decode the video stream using H.264 or H.265 encoding standards to obtain the original video frame sequence. Utilize the high compression ratio and low latency characteristics of H.264 or H.265 encoding to meet the real-time requirements of tunnel inspection.

[0022] The decoded original video frames are uniformly adjusted to a preset frame rate (preferably 25 FPS) to avoid increasing the hardware computing power burden due to excessively high frame rates or causing missed instantaneous defects due to excessively low frame rates. At the same time, the resolution is resampled and uniformly adjusted to 1920×1080 or 1280×720 (the specific resolution is adapted according to the computing power of the edge computing device) to ensure that the image size of the input model is consistent and to improve the stability of model inference.

[0023] For common tunnel scenarios such as low illumination (e.g., insufficient lighting in the middle of the tunnel), overexposure (e.g., backlighting at the tunnel entrance), and underexposure, targeted enhancement processing is performed on video frames: in low illumination scenarios, brightness enhancement algorithms and Gaussian denoising algorithms are used to remove image noise while improving image clarity; in overexposure / underexposure scenarios, exposure normalization algorithms are used to adjust the image grayscale range to be consistent with the grayscale range of tunnel video frames under normal lighting conditions, ensuring that the defects can be clearly presented.

[0024] Based on the actual tunnel structure drawings, ROI mask files are configured to shield non-monitored areas such as cable troughs, signs, and tunnel edge guardrails through mask calculations. After shielding, the amount of invalid calculations in the model can be reduced, and the normal structures in these areas can be avoided from being misjudged as defects, thus reducing the risk of false alarms.

[0025] Step 2: Due to the limitations of single deep learning models in terms of generality—object detection models excel at identifying discrete targets with clear outlines and fixed shapes, while semantic segmentation models are adept at delineating irregularly shaped regional targets without fixed boundaries—tunnel defects precisely encompass both types of targets (e.g., loose rocks and foreign objects are discrete targets, while water seepage and road surface water are irregular regional targets). Therefore, this embodiment employs a multi-model parallel inference strategy combining object detection and semantic segmentation, along with consistency judgment to achieve preliminary fusion, ensuring the breadth and accuracy of defect identification. The specific steps are as follows: The preprocessed video frames are synchronously input into two parallel deep learning models: The first deep learning model uses a lightweight object detection network such as YOLOv8 or YOLOv11, with an input size of 640×640, a confidence threshold of 0.2~0.4 (preferably 0.25), and a non-maximum suppression (NMS) threshold of 0.4~0.6 (preferably 0.45). The lightweight model reduces computational consumption while ensuring recognition accuracy, meeting the requirements of real-time inference. The first deep learning model is used to identify defects with relatively fixed boundaries, such as missing blocks and foreign objects, and outputs candidate results including defect category, confidence level, and candidate box coordinates (x1, y1, x2, y2).

[0026] The second deep learning model uses semantic segmentation networks such as U-Net, U-Net with EfficientNet as the backbone, or DeepLabv3+ to output pixel-level masks of irregular areas such as water seepage and road surface water accumulation. The mask value is 0 or 1, where 1 represents the affected area and 0 represents the background area. The mask binarization threshold is set to 0.45~0.55 (preferably 0.5). In this embodiment, the semantic segmentation model accurately outlines the contours of irregular areas, making up for the deficiency of the target detection model in recognizing areas without fixed boundaries.

[0027] Spatial correlation analysis is performed on the outputs of the first and second deep learning models. When candidate boxes and pixel-level masks have overlapping responses in the same spatial location (i.e., the same area is detected as a defect by both models simultaneously), one of the following two methods is selected for processing: Candidate box optimization: First, calculate the minimum bounding rectangle of the diseased area using pixel-level masking, and then obtain the coordinates of the top-left corner of this bounding rectangle. ) and the coordinates of the lower right corner ( Then, the candidate box coordinates (x1, y1, x2, y2) output by the target detection model are extracted. The boundary coordinates of the candidate boxes are directly corrected to the boundary coordinates of the bounding rectangle of the mask, or the two sets of coordinates are weighted and adjusted through a coordinate fusion algorithm (the weights are allocated according to the model confidence; the higher the confidence of the semantic segmentation model, the greater the weight of the bounding rectangle coordinates of the mask). Finally, the optimized candidate boxes completely wrap the actual disease area without any extra background area, achieving a precise fit with the actual disease area. The mask output by the semantic segmentation model accurately reflects the disease outline. Optimizing the candidate boxes in this way can solve the problem of inaccurate positioning of irregular disease boundaries by the target detection model and improve the target positioning accuracy.

[0028] Noise Removal: The number of valid pixels with a mask value of 1 (i.e., pixels matching the defect characteristics) in the pixel-level mask is counted, and the corresponding mask area is calculated. Results with a mask area lower than a preset relative threshold are identified as suspected noise and removed. The preset relative threshold is determined based on the defect classification standards in tunnel operation and maintenance specifications, specifying the minimum identifiable physical size of the defect to be monitored. For example, the minimum identifiable area for a slab in a high-speed tunnel is 0.01㎡, and the minimum identifiable area for water leakage is 0.005㎡. This is then combined with the local scale factor (per unit pixel) obtained from camera calibration. The physical area of ​​the smallest identifiable defect is converted into the corresponding pixel area, which is the preset relative threshold. For different types of defects, corresponding preset relative thresholds can be set. For example, the preset relative threshold for chipping is 0.01㎡ ÷ physical area per pixel, and the preset relative threshold for water leakage is 0.005㎡ ÷ physical area per pixel. This ensures that only small noise areas are removed without missing valid defects. Since small noise areas are mostly false detections caused by image noise or light and shadow interference, removing them can initially reduce the false alarm rate.

[0029] Step 3: Since the candidate results after the initial fusion of multiple models may still contain false detections caused by instantaneous environmental interference such as headlight sweeping, backlighting at openings, and water vapor condensation, these results lack spatial persistence and temporal stability. Therefore, in this embodiment, a dual verification of spatial filtering and temporal filtering is used to retain stable suspected defects and further suppress false alarms. The specific steps are as follows: Spatial filtering: The preliminary fusion candidate results are matched with the configured ROI mask file, and only candidate results that fall completely or partially within the ROI are retained, while candidate results located in non-monitoring areas are removed. This forms a double guarantee with the non-monitoring area masking, ensuring that only the results in the areas that need to be monitored are processed.

[0030] Temporal filtering: Time series analysis is performed by maintaining a sliding window of length M (M is 3~9, preferably 5). The sliding window stores the candidate results and their corresponding timestamps of the most recent M consecutive frames. The stability of the spatially filtered candidate results is verified according to the following rules: For candidate boxes of target detection: determine the frequency of occurrence of candidate boxes of the same category and similar location within the sliding window. When they appear in at least N frames (N is 2~5, preferably 3) and the IoU (Intersection over Union) of candidate boxes in adjacent frames is ≥0.5~0.7 (preferably 0.6), they are confirmed as stable suspected events.

[0031] For semantic segmentation masks: the IoU of the bounding rectangle of the mask or the pixel-by-pixel overlap ratio of the mask is used as a consistency metric. When the metric meets the same threshold requirements as the candidate box, i.e., the sliding window length M is 3~9 (preferably 5), appears in at least N frames (2~5, preferably 3), and the metric is ≥0.5~0.7 (preferably 0.6), it is confirmed as a stable suspected event. This allows for a more precise consistency judgment method to be flexibly selected for irregular diseases of different shapes.

[0032] In this embodiment, the sliding window uses a ring-shaped caching mechanism to store candidate results. When new frame data is input, it automatically overwrites the cached data of the earliest frame, achieving efficient use of storage space. At the same time, it ensures that the sliding window always stores the latest M-frame data, which meets the time-series requirements of real-time inspection.

[0033] Step 4: Use Zhang Zhengyou's calibration method to calibrate the front-end camera in the tunnel and obtain the camera intrinsic parameters (fx, fy, cx, cy), distortion coefficients, and homography matrix H. The homography matrix H is estimated by the planar feature point set of the ground or sidewall in the tunnel (such as paving stones of known size or embedded parts), and is used to establish the mapping relationship between pixel coordinates and physical world coordinates (such as meters).

[0034] For stable suspected events in target detection, extract the pixel coordinates of the four vertices of the optimized candidate box, map them to physical world coordinates through the homography matrix H, and calculate the actual length and width of the candidate box (such as the length and width of the dropped block).

[0035] For semantic segmentation-type stable suspected events: count the number of effective pixels in the mask, and combine the local scale factor (physical area corresponding to unit pixel) obtained from camera calibration to calculate the actual area of ​​the defect (such as the area of ​​water leakage or water accumulation); for areas with perspective distortion (such as the far end of the tunnel), perform partition integration according to the perspective correction algorithm to improve quantization accuracy and avoid quantization errors caused by perspective distortion.

[0036] In this embodiment, a mapping relationship between pixel coordinates and physical world coordinates is established through camera calibration, transforming pixel-level disease information into actual physical scale, thereby achieving quantitative assessment of diseases.

[0037] Step 5: Establish a disease threshold library and set corresponding physical scale thresholds for different disease types (slab falling, water seepage, road surface water accumulation, foreign objects, etc.). For example, an alarm is triggered when the maximum outer edge of the slab falling is ≥20cm; an alarm is triggered when the water seepage area is ≥1m²; and a high-level alarm is triggered when the road surface water accumulation area is ≥0.5m² and extends to the center line of the lane. The thresholds can be dynamically updated through configuration files or remote commands to adapt to the operation and maintenance needs of tunnels of different levels.

[0038] Compare the disease physical scale information obtained in step 4 with the corresponding thresholds in the threshold library. When the quantization result reaches or exceeds the threshold, it is determined as a disease event that requires an alarm, and structured alarm information is generated. The structured alarm information includes the following core fields: Disease type, disease location (image coordinates and mapped physical coordinates), disease physical scale quantization value, model comprehensive confidence level (obtained by fusing the confidence levels of two models, such as weighted average), event timestamp, event snapshot (screenshot of the corresponding video frame).

[0039] Furthermore, in this embodiment, to achieve seamless docking of the present invention with the existing tunnel integrated monitoring system or Internet of Things (IoT) platform, a standardized communication protocol needs to be used to encapsulate the alarm information to ensure the compatibility and reliability of data transmission. The specific steps are as follows: Encapsulate the generated structured alarm information into a JSON format message body. A typical example of JSON fields is as follows: {"msg":"almInfo","gateway":"Tunnel A - Gateway 1","code":"leakage","type":"Water leakage","val":1.5,"ts":"2025-10-27T10:00:00Z","picData":"<Base64 encoded snapshot data>","bbox":[x1,y1,x2,y2]}; Publish the JSON message body to the MQTT broker server through the MQTT protocol. Hierarchical topic management is adopted, and the topic format is preferably "warning / {gateway_id} / event" (such as "warning / Tunnel A - Gateway 1 / event"), which is convenient for the monitoring platform to accurately subscribe to alarm information according to the tunnel area and gateway number. In engineering deployment, the quality of service level QoS 1 is selected, and a message retransmission strategy (such as 3 retransmission times and a 2 - second retransmission interval) is set to ensure that the message can be re - transmitted when lost due to network fluctuations, improving the message delivery rate.

[0040] The present invention solves the problems of limited recognition range, high false alarm rate, lack of quantitative evaluation, and weak engineering ability in the existing tunnel inspection technology through video pre - processing, multi - model fusion, spatio - temporal consistency verification, physical quantization, and standardized publishing; comprehensively recognizes different forms of diseases through multi - model parallel inference, greatly reduces the false alarm rate through spatio - temporal consistency verification, realizes accurate grading evaluation through physical quantization, and improves the system integration ability through standardized protocol encapsulation. Finally, it promotes the transformation of tunnel inspection from traditional manual and qualitative judgment to automated, accurate, and quantitative intelligent inspection, with significant practical value and engineering significance.

[0041] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent inspection of tunnel defects based on multi-model fusion and consistency verification, characterized in that, Includes the following steps: Step 1: Acquire the monitoring video stream inside the tunnel and preprocess it to obtain standardized video frames; Step 2: Start two parallel deep learning models to analyze the standardized video frames. The first deep learning model outputs candidate boxes of lesions with fixed boundary features, and the second deep learning model outputs pixel-level masks of lesions without fixed boundary features. Then, based on the spatial positional relationship between the two, a consistency judgment is made to obtain preliminary fusion candidate results. Step 3: Spatial filtering is performed on the preliminary fusion candidate results to retain the candidate results located within the preset monitoring area. Then, the stability of the spatially filtered candidate results is verified by time series analysis to obtain stable suspected disease events. Step 4: Based on the preset camera calibration parameters, establish the mapping relationship between pixel coordinates and physical world coordinates, and convert the pixel dimension information corresponding to stable suspected disease events into actual physical scale information; Step 5: Compare the actual physical scale information with the preset disease threshold standard to determine whether an alarm is triggered. If an alarm is triggered, generate structured alarm information, encapsulate the structured alarm information into a message conforming to the message queue telemetry transmission protocol, and publish it to the monitoring platform.

2. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The preprocessing includes video decoding, frame rate uniform adjustment, resolution resampling, image enhancement, and non-monitoring area masking. Image enhancement is used to improve the quality of video frames in low-light, overexposed, or underexposed scenes, while non-monitoring area masking is achieved by configuring a preset monitoring area mask.

3. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The first deep learning model is a lightweight object detection network, and the second deep learning model is a semantic segmentation network. The lightweight object detection network is used to identify defects such as missing blocks and foreign objects; the semantic segmentation network is used to identify defects such as water leakage and road surface water accumulation.

4. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 3, characterized in that, The lightweight object detection network is a YOLO series network, and the semantic segmentation network is any one of the following: U-Net network, U-Net network with EfficientNet as the backbone, or DeepLabv3+ network.

5. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The consistency determination includes: When the candidate bounding box for a defect overlaps with the pixel-level mask in the same spatial location, defect candidate bounding box optimization or noise removal is used for processing. The optimization of the disease candidate box includes: adjusting the boundary coordinates of the disease candidate box based on the bounding rectangle of the pixel-level mask, so that the adjusted disease candidate box fits the actual disease area better; Noise removal includes: counting the number of valid pixels that meet the disease characteristics after being filtered by a preset mask threshold in the pixel-level mask, and judging the results whose mask area corresponding to the number of valid pixels is lower than a preset relative threshold as noise and removing them.

6. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The time series analysis is implemented by maintaining a sliding window, which is used to store candidate results from the most recent consecutive frames. The temporal stability of the candidate results is verified by judging the frequency of occurrence and positional overlap of candidate results of the same category and similar position within the sliding window.

7. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The camera calibration parameters are obtained through the Zhang Zhengyou calibration method, including camera intrinsic parameters, distortion coefficients, and homography matrix estimated based on the set of planar feature points in the tunnel. The homography matrix is ​​used to realize the mapping transformation from pixel coordinates to physical world coordinates.

8. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The structured alarm information includes disease type, disease location information, disease physical scale information, model analysis confidence level, event timestamp, and event snapshot.

9. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to claim 1, characterized in that, The message publishing of the message queue telemetry transmission protocol adopts hierarchical topic management and supports a message retransmission mechanism.

10. The intelligent inspection method for tunnel defects based on multi-model fusion and consistency verification according to any one of claims 1 to 9, characterized in that, The intelligent tunnel defect inspection method is deployed on edge computing devices and optimizes the deployment of two parallel deep learning models through an inference acceleration framework.

Images