Tunnel disease detection system based on candidate region dynamic guidance and confidence fusion

Abstract

The present application belongs to the technical field of track tunnel structure health detection, and especially is a tunnel disease detection system based on candidate region dynamic guidance and confidence fusion, which adopts a cascaded two-stage architecture. After image preprocessing optimizes the image quality, the system generates disease candidate regions and calculates the preliminary confidence through multi-scale gradient analysis, constraint determination and clustering. With the help of the embedded self-attention ShuffleNetV2 light-weight network, the system completes the fine detection, and outputs the fine confidence and segmentation mask. According to the edge device load, the system dynamically adjusts the scheduling threshold, and fuses the double confidence to obtain the final result. The cloud collects negative samples, and uses the EWC algorithm for incremental training and remote model updating. The present application can reduce the false detection and missed detection problems under complex light, relies on adaptive scheduling to reduce the reasoning time consumption, and takes into account the accuracy and real-time performance, and is suitable for automatic inspection of tunnel cracks and water leakage diseases.

Description

Technical Field

[0001] This invention relates to the field of structural health detection technology for rail transit tunnels, specifically a tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion. Background Technology

[0002] With the continuous construction and operation of my country's rail transit network, tunnel linings are constantly affected by factors such as load, temperature, humidity, and groundwater, making them prone to defects such as cracks and water leakage. If these defects are not detected and addressed in a timely manner, they can seriously threaten train safety. Currently, tunnel defect detection methods are mainly divided into three categories: manual inspection, traditional image processing, and end-to-end deep learning detection, all of which have significant shortcomings. Manual inspection: low work efficiency, high labor costs, poor lighting and confined working space inside the tunnel pose great safety hazards, and the inspection results are affected by the subjective experience of the personnel, resulting in poor consistency; Traditional image processing methods, such as Canny edge detection and morphological analysis, have low accuracy in identifying minute cracks and point-like water seepage in complex environments such as low illumination, strong shadows, and low texture in tunnels, and cannot meet the requirements for refined detection. End-to-end deep learning models, represented by single-stage models such as YOLO and SSD, have a high degree of automation, but a high false detection rate in complex lighting environments; lightweight models deployed on edge devices suffer from large inference latency and a high rate of missing detection of minor defects.

[0003] Furthermore, existing hybrid detection schemes mostly involve simple serial and parallel stacking of modules. There is a lack of collaborative guidance mechanisms between the candidate region generation module and the verification module. They also fail to perform adaptive resource scheduling based on edge device load and lack a confidence feedback closed-loop optimization system, thus failing to balance detection accuracy and real-time performance at the edge. To address these shortcomings of existing technologies, this invention proposes a novel intelligent detection system for tunnel defects.

[0004] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion, which solves the current problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion, comprising: Image acquisition module, which is used to acquire raw images of the tunnel lining surface; An image preprocessing module, which is connected to the image acquisition module, is used to perform contrast enhancement and noise reduction on the original image to obtain a preprocessed image; The structure-aware candidate region generation module is used to perform multi-scale gradient analysis, structural continuity constraints, and linear target aggregation operations on the preprocessed image to generate a set of candidate regions for tunnel defects and to calculate the preliminary confidence of each candidate region. The lightweight target detection and verification module is used to perform fine classification and segmentation of the tunnel defect area, output the defect category, fine confidence level and pixel-level segmentation mask; The dual-confidence decision and adaptive adjustment module is connected to the structure-aware candidate region generation module and the lightweight target detection verification module, respectively. It is used to monitor the load rate of edge devices in real time, dynamically adjust the threshold of the number of candidate regions, determine whether to enable the lightweight target detection verification module based on the threshold, and complete the dynamic weighted fusion of the preliminary confidence and fine confidence to obtain the final confidence. The result output and cloud transmission module is used to output qualified tunnel defect detection results and upload the detection data to the cloud server. The cloud-based incremental training module is used to collect negative sample data uploaded from the edge device, perform incremental training to update the model parameters, and then distribute the updated model to the edge device.

[0007] In some embodiments, the image acquisition module uses a line scan industrial camera with a resolution of 1920×1080 and a frame rate of 30fps to complete image acquisition. The image preprocessing module uses the CLAHE adaptive histogram equalization algorithm to enhance contrast, sets clipLimit=2.0 and gridSize=8×8, and uses a 3×3 Gaussian kernel and σ=0.8 Gaussian filter to denoise the image.

[0008] In some embodiments, the structure-aware candidate region generation module specifically performs the following steps: 1) Multiscale gradient analysis: In scale parameters The following formula is used to calculate the image structure tensor: ; in, For scale The corresponding Gaussian kernel, To preprocess the image, local maxima of gradient magnitude at each scale are extracted as candidate features; 2) Structural continuity constraints: The continuity of a line segment is determined using a decision function, the formula of which is as follows: ; Among them, attenuation coefficient Let be the direction angle of the line segment, when A value greater than 0.6 indicates the line segment is continuous; 3) Linear target aggregation: The DBSCAN clustering algorithm is used to collect line segments, and a neighborhood radius is set. , minimum number of points The consistency of the filtering direction is greater than the threshold. Line segments converge to form a set of candidate disease areas; 4) Preliminary confidence level calculation: Calculate the preliminary confidence level using the following formula: ; in, To ensure gradient consistency in candidate regions, For the aspect ratio of the candidate region, a maximum aspect ratio threshold is preset. =50.

[0009] In some embodiments, the lightweight object detection verification module uses ShuffleNetV2 as the backbone network, combines it with the Feature Pyramid Network (FPN), and embeds self-attention modules in each layer of the FPN. The calculation formula for the self-attention module is as follows: ; in, This is a global average pooling feature. This is the weight matrix of the fully connected layer. The lightweight target detection verification module takes a candidate region image cropped and scaled to 224×224 pixels as input (using the Sigmoid function) and outputs the disease category and fine-grained confidence score. With pixel-level segmentation mask.

[0010] In some embodiments, the adaptive resource scheduling logic of the dual-confidence decision and adaptive scheduling module is as follows: Real-time acquisition of edge device GPU utilization ,when When the percentage is greater than 80%, set a threshold for the number of candidate regions. ;when At that time, set a threshold for the number of candidate regions. If the current total number of candidate regions If the lightweight target detection verification module is skipped, the preliminary confidence level is directly used as the final confidence level. If so, the lightweight target detection verification module will be activated.

[0011] In some embodiments, the dual-confidence fusion formula for the dual-confidence decision and adaptive scheduling module is as follows: ; ; in, For the initial confidence level, For fine-grained confidence, only output A test result of ≥0.5.

[0012] In some embodiments, the result output and cloud transmission module use the MQTT communication protocol to upload the image ID, candidate region coordinates, final confidence score, disease category, and segmentation mask data to the cloud server.

[0013] In some embodiments, the workflow of the cloud-based incremental training module is as follows: 1) Screening negative samples: Mark the missed detection areas with a confidence level below 0.3 and manually confirmed as disease, and the false alarm areas with a confidence level above 0.7 and manually confirmed as no disease as negative samples, and at the same time, push the samples with a confidence level in the range of 0.3-0.6 to the manual annotation platform. 2) Incremental training: Model training is performed every 1000 accumulated image samples and every two weeks. The EWC elastic weight consolidation algorithm is used for incremental training to avoid catastrophic forgetting of the model. 3) Model push: The parameters of the trained lightweight object detection verification module are pushed to the edge device via OTA to complete the model update.

[0014] This invention also discloses an intelligent detection method for tunnel defects based on dynamic guidance of candidate regions and dual confidence fusion, applied to the above-mentioned system, comprising the following steps: Step 1: Acquire the original image of the tunnel lining using a linear industrial camera, and then perform CLAHE enhancement and Gaussian filtering for noise reduction to obtain the preprocessed image; Step 2: The structure-aware candidate region generation module is used to complete multi-scale gradient analysis, structural continuity constraints, and DBSCAN clustering to obtain a set of disease candidate regions and their corresponding preliminary confidence scores. Step 3: Monitor the GPU utilization of edge devices in real time, dynamically set the threshold for the number of candidate regions, and determine whether to enable the lightweight object detection verification module based on the relationship between the total number of candidate regions and the threshold. Step 4: If the verification module is skipped, the preliminary confidence level is directly output as the final result; if the verification module is enabled, the fine confidence level is obtained and then the dynamic weighted fusion of the two confidence levels is completed, and the final disease detection result with a confidence level ≥ 0.5 is output. Step 5: Upload the detection data to the cloud. The cloud collects negative samples and completes incremental training. After updating the model, it is distributed to the edge devices.

[0015] Compared with existing technologies, this invention provides a tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion, which has the following beneficial effects: 1. This tunnel defect detection system based on dynamic candidate region guidance and confidence fusion adopts a structure-aware candidate region generation mechanism. It designs multi-scale gradients and structural continuity constraints for linear defects in tunnels, and combines DBSCAN clustering to accurately locate micro-cracks and water leakage areas. It is equipped with a lightweight attention network for fine-grained verification, and then makes decisions through dynamic fusion of dual confidence. Compared with existing models such as YOLOv8 and STCYOLO, the false detection rate is reduced by 35%-50%, the recall rate of micro-defects smaller than 32×32 pixels is improved by 18.6%, the mAP@0.5 index is significantly improved in low light (≤10 lux) and strong shadow environments, and the robustness to complex lighting is stronger. 2. This tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion introduces an adaptive resource scheduling strategy. It dynamically decides whether to skip the verification module based on the load of edge devices and the number of candidate regions. In scenarios with low candidate regions, the inference time can be as low as 30ms, and the computational overhead is reduced by about 47% in high-load scenarios. The overall inference latency is reduced by 25%, which fully meets the real-time detection requirements of edge devices, and the accuracy loss is less than 1%. 3. This tunnel defect detection system based on candidate region dynamic guidance and confidence fusion adopts a cascaded architecture of "candidate region dynamic guidance + lightweight network verification". The two modules work together, which is different from the traditional simple stacking scheme. The front end selects effective areas and the back end performs fine verification. The division of labor is clear and the overall detection logic is more in line with the characteristics of tunnel defects. At the same time, a cloud-based negative sample collection and incremental training module is built, which, combined with the EWC algorithm, avoids model forgetting and updates the edge model regularly. The system can be continuously iterated and optimized according to the actual working conditions on site. With long-term use, the detection performance is steadily improved, reducing the later maintenance costs. 4. This tunnel defect detection system, based on dynamic guidance of candidate regions and confidence fusion, is flexible in overall deployment and can be mounted on mainstream edge computing platforms such as NVIDIA Jetson AGX Xavier. It is suitable for automated inspection operations of various rail transit tunnels and municipal tunnels, with minimal manual intervention and strong practicality. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall module structure of the tunnel defect detection system of the present invention. Detailed Implementation

[0017] 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.

[0018] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0019] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0021] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0022] Please see Figure 1 The intelligent tunnel defect detection system proposed in this invention based on dynamic guidance of candidate regions and dual confidence fusion includes: an image acquisition module, an image preprocessing module, a structure-aware candidate region generation module, a lightweight target detection verification module, a dual confidence decision and adaptive scheduling module, a result output and cloud transmission module, and a cloud incremental training module.

[0023] Image acquisition module: used to acquire raw images of the tunnel lining surface; Linear industrial cameras were used to acquire images on site. The camera resolution was set to 1920×1080 and the working frame rate was 30fps. The camera was mounted on the tunnel inspection equipment and moved at a constant speed along the tunnel driving route to continuously acquire full-area images of the tunnel lining surface. All acquired raw images were transmitted to the next module in real time to complete the construction of the tunnel raw image data source.

[0024] Image preprocessing module: Performs contrast enhancement and noise reduction on the original image and outputs a preprocessed image; The CLAHE adaptive histogram equalization algorithm was used to enhance contrast, adapting to the complex lighting environment of tunnels with low intensity and strong shadows. The algorithm was configured with clipLimit=2.0 and gridSize=8×8, effectively improving the texture recognition of dark areas in the image and highlighting the characteristics of defects such as cracks and water seepage. Then perform image denoising: A Gaussian kernel with a size of 3×3 was selected and the Gaussian coefficient σ=0.8 was set to perform Gaussian filtering on the image after contrast enhancement to filter out salt and pepper noise and Gaussian noise generated during image acquisition and transmission, and finally obtain a preprocessed image with clear image quality and complete features. Structure-aware candidate region generation module: Performs multi-scale gradient analysis, structural continuity constraints, and linear target aggregation on the preprocessed image to generate a set of disease candidate regions and calculate the preliminary confidence level; First, we proceed with multi-scale gradient analysis: Set scale parameters The structure tensor of the preprocessed image is calculated at three different scales, and the formula is as follows: ; in, For scale The corresponding Gaussian kernel, To preprocess the image, local maxima of gradient magnitude at each scale are extracted sequentially, and all extreme points are summarized as candidate feature points for disease, taking into account the linear disease features of different thicknesses and scales. Then, the structural continuity constraint determination is performed: Line segments are fitted to the extracted candidate feature points, and the validity of the line segments is screened using a structural continuity determination formula. The determination function formula is as follows: ; Among them, attenuation coefficient The direction angle of a single line segment. For the coordinates of the endpoints of adjacent line segments, the system sets a judgment threshold of 0.6. When the calculation structure... If the value is greater than 0.6, the line segments are considered to have structural continuity and are retained; otherwise, they are considered to be discrete noise line segments and are directly discarded. Next, linear target aggregation is performed: The DBSCAN clustering algorithm is used to collect line segments, and a neighborhood radius is set. , minimum number of points The consistency of the filtering direction is greater than the threshold. Line segments converge to form a set of candidate disease areas; Finally, calculate the initial confidence level: Calculate the preliminary confidence level using the following formula: ; in, To ensure gradient consistency in candidate regions, For the aspect ratio of the candidate region, a maximum aspect ratio threshold is preset. It iterates through all candidate disease regions, calculates the preliminary confidence level for each region, records the coordinate information of the candidate regions, and outputs the candidate region set and the corresponding preliminary confidence level data in a unified manner.

[0025] Lightweight target detection verification module: performs fine classification and segmentation of disease candidate regions, and outputs disease category, fine confidence score and pixel-level segmentation mask; First, perform image preprocessing on the input candidate regions: Read the coordinates of the candidate regions output by the structure-aware candidate region generation module, crop each candidate region in the preprocessed image, and uniformly scale it to 224×224 pixels to form a standard input image; Then, build the lightweight network main structure: This module uses a lightweight convolutional network, ShuffleNetV2, as its backbone network for basic feature extraction. The backbone network is connected to a Feature Pyramid Network (FPN) at the back end to achieve multi-scale feature fusion. A self-attention module is embedded in each feature branch of the FPN layer to strengthen the weights of effective disease features and suppress background interference. The calculation formula for the self-attention module is as follows: ; in, This is a global average pooling feature. This is the weight matrix of the fully connected layer. For the Sigmoid function; Network training parameter configuration: The model training batch size is set to 16, the initial learning rate is 0.001, the optimizer is AdamW, the total number of training rounds is set to 100, and the pre-training is completed based on the tunnel disease dataset; After a standard-sized image is input into the network, it undergoes feature extraction, feature fusion, and attention weighting. Finally, the output layer outputs three types of data: first, the disease category, which is divided into three categories: cracks, water leakage, and no disease; second, fine confidence, with a value range of {0,1}; and third, a 224×224 pixel-level segmentation mask, with mask pixels having a value of 0 or 1, used to mark the disease pixel region. Dual-confidence decision-making and adaptive scheduling module: Real-time monitoring of edge device load, dynamic adjustment of thresholds, execution of scheduling logic and completion of dual-confidence fusion; First, collect edge device load data in real time: Online reading of edge device GPU utilization (Unit: %), when When the percentage is greater than 80%, set a threshold for the number of candidate regions. ;when At that time, set a threshold for the number of candidate regions. If the current total number of candidate regions If the lightweight target detection verification module is skipped, the preliminary confidence score will be used directly as the final confidence score. Then the lightweight target detection verification module will be enabled.

[0026] Dual-confidence fusion consists of two steps: dynamic weight calculation and final confidence calculation. ; ; in, For the initial confidence level, To achieve a refined confidence level, a unified screening process is performed after the fusion is completed. Valid disease detection results with a value ≥0.5 are output to downstream modules; Results output and cloud transmission module: Outputs detection results locally and uploads all data to the cloud server; Local output stage: Displays the area coordinates of effective diseases, disease type, final confidence level, and pixel-level segmentation mask, completing the visualization output of on-site detection results; Data upload stage: Using the MQTT IoT communication protocol, complete detection data such as single frame image ID, candidate region coordinates, final confidence score, disease category, and segmentation mask are packaged and uploaded to the cloud server in real time for storage, providing a data source for subsequent sample collection and model optimization; Cloud-based incremental training module: collects negative samples, performs incremental training, and remotely pushes updated model parameters; First, sample classification and labeling are performed: two types of typical negative samples are selected from the detection data stored in the cloud: the first type is missed samples with a confidence level below 0.3, which are confirmed as disease by manual review; the second type is false positive samples with a confidence level above 0.7, which are confirmed as disease-free by manual review. At the same time, low-confidence samples with a confidence level in the range of 0.3-0.6 are pushed to the manual labeling platform for supplementary labeling. Then, incremental training is performed: Triggering rules are set: incremental training of the model is initiated every 1000 accumulated image samples in the cloud, or every fixed two-week period. The training process employs the EWC elastic weight consolidation algorithm to effectively suppress the catastrophic forgetting problem that occurs during incremental model training. The training object is a lightweight object detection verification module, and the network weight parameters are iteratively optimized.

[0027] Finally, the model is remotely pushed: After incremental training is completed, the updated model parameters are sent to the front-end edge detection device via OTA remote upgrade, automatically completing the model replacement and forming a closed-loop iterative system of "edge detection - cloud optimization - edge update" to continuously improve the system's detection performance.

[0028] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0029] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion, characterized in that, include: Image acquisition module, which is used to acquire raw images of the tunnel lining surface; An image preprocessing module, which is connected to the image acquisition module, is used to perform contrast enhancement and noise reduction on the original image to obtain a preprocessed image; The structure-aware candidate region generation module is used to perform multi-scale gradient analysis, structural continuity constraints, and linear target aggregation operations on the preprocessed image to generate a set of candidate regions for tunnel defects and to calculate the preliminary confidence of each candidate region. The lightweight target detection and verification module is used to perform fine classification and segmentation of the tunnel defect area, output the defect category, fine confidence level and pixel-level segmentation mask; The dual-confidence decision and adaptive adjustment module is connected to the structure-aware candidate region generation module and the lightweight target detection verification module, respectively. It is used to monitor the load rate of edge devices in real time, dynamically adjust the threshold of the number of candidate regions, determine whether to enable the lightweight target detection verification module based on the threshold, and complete the dynamic weighted fusion of the preliminary confidence and fine confidence to obtain the final confidence. The result output and cloud transmission module is used to output qualified tunnel defect detection results and upload the detection data to the cloud server. The cloud-based incremental training module is used to collect negative sample data uploaded from the edge device, perform incremental training to update the model parameters, and then distribute the updated model to the edge device.

2. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The image acquisition module uses a line scan industrial camera with a resolution of 1920×1080 and a frame rate of 30fps to complete image acquisition. The image preprocessing module uses the CLAHE adaptive histogram equalization algorithm to enhance contrast, sets clipLimit=2.0 and gridSize=8×8, and uses a 3×3 Gaussian kernel and σ=0.8 Gaussian filter to denoise the image.

3. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The specific execution steps of the structure-aware candidate region generation module are as follows: 1) Multiscale gradient analysis: In scale parameters The following formula is used to calculate the image structure tensor: ； in, For scale The corresponding Gaussian kernel, To preprocess the image, local maxima of gradient magnitude at each scale are extracted as candidate features; 2) Structural continuity constraints: The continuity of a line segment is determined using a decision function, the formula of which is as follows: ； Among them, attenuation coefficient Let be the direction angle of the line segment, when A value greater than 0.6 indicates the line segment is continuous; 3) Linear target aggregation: The DBSCAN clustering algorithm is used to collect line segments, and a neighborhood radius is set. , minimum number of points The consistency of the filtering direction is greater than the threshold. Line segments converge to form a set of candidate disease areas; 4) Preliminary confidence level calculation: Calculate the preliminary confidence level using the following formula: ； in, To ensure gradient consistency in candidate regions, For the aspect ratio of the candidate region, a maximum aspect ratio threshold is preset. .

4. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The lightweight target detection verification module uses ShuffleNetV2 as the backbone network, combines it with the Feature Pyramid Network (FPN), and embeds self-attention modules in each layer of the FPN. The calculation formula for the self-attention module is as follows: ； in, This is a global average pooling feature. This is the weight matrix of the fully connected layer. The lightweight target detection verification module takes a candidate region image cropped and scaled to 224×224 pixels as input (using the Sigmoid function) and outputs the disease category and fine-grained confidence score. With pixel-level segmentation mask.

5. The tunnel defect detection system based on candidate region dynamic guidance and confidence fusion according to claim 1, characterized in that, The adaptive resource scheduling logic of the dual-confidence decision and adaptive scheduling module is as follows: Real-time acquisition of edge device GPU utilization ,when When the percentage is greater than 80%, set a threshold for the number of candidate regions. ;when At that time, set a threshold for the number of candidate regions. If the current total number of candidate regions If the lightweight target detection verification module is skipped, the preliminary confidence level is directly used as the final confidence level. If so, the lightweight target detection verification module will be activated.

6. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The dual-confidence fusion formula for the dual-confidence decision-making and adaptive scheduling module is as follows: ； ； in, For the initial confidence level, For fine-grained confidence, only output The test results.

7. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The output of the results is transmitted to the cloud using the MQTT communication protocol, which uploads the image ID, candidate region coordinates, final confidence score, disease category, and segmentation mask data to the cloud server.

8. The tunnel defect detection system based on dynamic guidance of candidate regions and confidence fusion according to claim 1, characterized in that, The workflow of the cloud-based incremental training module is as follows: 1) Screening negative samples: Mark the missed detection areas with a confidence level below 0.3 and manually confirmed as disease, and the false alarm areas with a confidence level above 0.7 and manually confirmed as no disease as negative samples, and at the same time, push the samples with a confidence level in the range of 0.3-0.6 to the manual annotation platform. 2) Incremental training: Model training is performed every 1000 accumulated image samples and every two weeks. The EWC elastic weight consolidation algorithm is used for incremental training to avoid catastrophic forgetting of the model. 3) Model push: The parameters of the trained lightweight object detection verification module are pushed to the edge device via OTA to complete the model update.

9. A tunnel defect detection method based on dynamic guidance of candidate regions and confidence fusion, characterized in that, The detection system applied to any one of claims 1-8 comprises the following steps: Step 1: Acquire the original image of the tunnel lining using a linear industrial camera, and then perform CLAHE enhancement and Gaussian filtering for noise reduction to obtain the preprocessed image; Step 2: The structure-aware candidate region generation module is used to complete multi-scale gradient analysis, structural continuity constraints, and DBSCAN clustering to obtain a set of disease candidate regions and their corresponding preliminary confidence scores. Step 3: Monitor the GPU utilization of edge devices in real time, dynamically set the threshold for the number of candidate regions, and determine whether to enable the lightweight object detection verification module based on the relationship between the total number of candidate regions and the threshold. Step 4: If the review module is skipped, the preliminary confidence level is directly output as the final result; If the verification module is enabled, after obtaining the fine confidence level, the dynamic weighted fusion of the two confidence levels is completed, and the final disease detection result with a confidence level ≥ 0.5 is output. Step 5: Upload the detection data to the cloud. The cloud collects negative samples and completes incremental training. After updating the model, it is distributed to the edge devices.

Images