Shield tunnel bolt health degree evaluation method based on images

Through the image-based deep learning model, the corrosion detection and health assessment of shield tunnel bolts is solved, and the problems of low manual detection efficiency and high safety risks in the existing technology are realized, automated detection and evaluation are achieved, and detection efficiency and accuracy are improved. extends the service life of the tunnel.

CN118982529AActive Publication Date: 2024-11-19CHINA RAILWAY SIYUAN SURVEY & DESIGN GRP CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411084931.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-19
Estimated Expiration
2044-08-08

AI Technical Summary

Technical Problem

The existing shield tunnel bolt corrosion detection mainly relies on manual methods, and there are problems such as low efficiency, high safety risks, high cost and strong subjectivity, making it difficult to meet the needs of large-scale, long-term and large-scale regular inspections.

Method used

The image-based shield tunnel bolt health evaluation method is used to pre-process the bolt scanned image, feature extraction, rust detection and health evaluation are achieved through deep learning models to achieve automated detection and evaluation. The method includes HSV color space conversion, exponential operation preprocessing, feature pyramid networks and dual branch network structures, which can automatically detect rust areas and evaluate bolt health.

Benefits of technology

It significantly improves detection efficiency and accuracy, reduces manual subjectivity and safety risks, realizes automated detection and health assessment of bolt rust in shield tunnels, provides reliable technical support for the maintenance and management of shield tunnels, and extends the tunnel. Service life.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982529A_ABST
    Figure CN118982529A_ABST
Patent Text Reader

Abstract

The invention provides an image-based shield tunnel bolt health degree evaluation method, and relates to the technical field of shield tunnel engineering, and the method comprises the steps: obtaining a bolt scanning image in a tunnel; constructing a deep learning model which comprises a backbone network and a segmentation network; inputting the preprocessed bolt scanning image into a deep learning model for corrosion detection, extracting a feature map by using a backbone network, setting regions of interest for the feature map, screening the regions of interest, and inputting the screened regions of interest into a segmentation network to obtain a corrosion detection result of the bolt scanning image. Comprise rust position information, category information and segmentation information; and calculating a corrosion area according to a corrosion detection result, and evaluating the bolt based on the corrosion area according to a health degree evaluation rule to obtain a health degree evaluation result of the bolt scanning image. The corrosion efficiency of the duct piece bolt is well improved, and the influence of manual subjectivity on accuracy and the time and labor cost are reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of shield tunnel engineering, and in particular to an image-based shield tunnel bolt health evaluation method. Background Art

[0002] With the acceleration of China's urbanization process and the rapid development of the national economy, urban rail transit systems, especially subway networks, have shown explosive growth. As of December 2023, 306 urban rail transit lines have been opened and operated nationwide, with a total mileage of more than 10,000 kilometers, marking that my country's subway construction has entered the large-scale operation stage.

[0003] As a core component of subway infrastructure, shield tunnels have the characteristics of huge investment scale and high difficulty in maintenance and reconstruction, which poses severe challenges to their long-term service performance and service life. However, shield tunnels are exposed to complex underground environments for a long time and face multiple factors such as groundwater seepage and ion erosion, which leads to the aggravation of the deterioration process of the segments. These factors work together to cause a series of diseases in the tunnel structure during operation, including but not limited to: lining cracking and peeling, segment misalignment, water leakage, etc.

[0004] Domestic case studies highlight potential risks to the long-term performance of shield tunnels:

[0005] Hong Kong MTR, China: The section tunnel built in the 1970s had serious problems of internal steel bar corrosion and concrete protective layer peeling in the 1990s.

[0006] Dapu Road Cross-River Tunnel in Shanghai, China: Built in the early 1970s, it was found to have problems with peeling of the protective layer and corrosion of the joint bolts and steel connections less than 20 years after it was put into operation.

[0007] Kaohsiung Port Cross-Harbor Undersea Tunnel in Taiwan, China: After 20 years of operation, it has suffered from various problems such as lining cracks, corrosion of steel bars and steel plates, water leakage and deterioration of the beam protective layer.

[0008] These cases highlight that the durability of shield tunnel structures has become a key factor affecting their long-term safe operation.

[0009] At present, the corrosion detection of shield tunnel segment bolts still mainly relies on manual methods, which has many limitations:

[0010] Inefficiency: Manual inspection is slow and cannot meet the inspection needs of large-scale tunnel networks. Safety risks: Tunnel vault inspection requires special equipment, which poses potential safety hazards. High cost: Manual inspection is costly and is not conducive to long-term, large-scale regular inspection. High subjectivity: Health assessment relies too much on workers’ experience and lacks objective and unified evaluation standards. High professional requirements: The professional quality of inspectors is high and the training cost is high.

[0011] In view of the important position of shield tunnels in urban rail transit and the long-term service performance challenges they face, it has become an urgent task to develop an efficient, objective and safe method for detecting and evaluating the corrosion of segment bolts. Summary of the invention

[0012] In view of this, the present invention proposes an image-based shield tunnel bolt health evaluation method, which uses a computer to quickly obtain the corrosion location and corrosion area of ​​the segment bolts and propose a method for the bolt health grade, which effectively solves the corrosion efficiency of the segment bolts, reduces the influence of human subjectivity on accuracy, and reduces time and labor costs.

[0013] The technical solution of the present invention is implemented as follows: The present invention provides an image-based shield tunnel bolt health evaluation method, comprising:

[0014] S1 obtains a bolt scanning image in the tunnel and pre-processes the bolt scanning image;

[0015] S2 builds a deep learning model, including a backbone network and a segmentation network;

[0016] S3 inputs the preprocessed bolt scanning image into the deep learning model for corrosion detection, uses the backbone network to extract feature maps, sets regions of interest for the feature maps, and screens the regions of interest. The screened regions of interest are input into the segmentation network to obtain the corrosion detection results of the bolt scanning image, including the location information, category information, and segmentation information of the corrosion;

[0017] S4 calculates the corrosion area according to the corrosion detection result, evaluates the bolt based on the corrosion area according to the health evaluation rule, and obtains the health evaluation result of the bolt scanning image.

[0018] On the basis of the above technical solution, preferably, the backbone network includes a feature extraction module and a region generation module, and the segmentation network includes a first network branch and a second network branch, wherein:

[0019] The feature extraction module includes a 7*7 convolutional layer, a 3*3 maximum pooling layer, 4 residual blocks, and a feature pyramid. Each residual block contains multiple residual units. Each residual unit includes 2 1*1 convolutional layers and 1 3*3 convolutional layer. The feature pyramid adopts a top-down path.

[0020] The region generation module includes a shared convolutional layer, a classification layer, and a bounding box regression layer;

[0021] The first network branch includes two fully connected layers;

[0022] The second network branch includes a 3*3 convolutional layer and a transposed convolutional layer.

[0023] Based on the above technical solution, preferably, step S1 includes:

[0024] S11 converts the bolt scanning image from the original RGB space to the HSV color space to obtain an HSV image;

[0025] S12 separates the H, S and V channels of the HSV image;

[0026] S13 performs exponential operation on the S and V channels respectively, wherein the exponential factor ranges from 0.25 to 4;

[0027] S14 keeps the H channel unchanged, and recombines the S and V channels processed in step S13 with the original H channel;

[0028] S15 converts the enhanced HSV image obtained in step S14 back to the RGB space to obtain a preprocessed bolt scanning image.

[0029] Based on the above technical solution, preferably, step S3 includes:

[0030] S31 inputs the preprocessed bolt scan image into the feature extraction module, uses the 7*7 convolution layer to extract the first feature map C1, and then sequentially inputs 4 residual blocks to extract 4 feature maps, namely the second feature map C2, the third feature map C3, the fourth feature map C4, and the fifth feature map C5, wherein the scale of each feature map is different, and the scale of the feature map increases in sequence according to the data flow direction;

[0031] S32 uses a top-down path to construct a pyramid and obtains a five-layer pyramid feature map;

[0032] S33 uses a fixed-size sliding window on the five-layer pyramid feature map, generates multiple anchor boxes of preset sizes and aspect ratios at the center of the sliding window, applies a 3*3 shared convolutional layer to each sliding window position, uses a classification layer for binary classification, predicts the probability of each anchor box being the foreground or background, uses a bounding box regression layer for bounding box regression, predicts the adjustment parameters of each anchor box, selects the N anchor boxes with the highest scores according to the classification scores, applies NMS to the N anchor boxes, and uses the NMS-posted anchor boxes as the filtered regions of interest;

[0033] S34 applies RoI Align to each filtered region of interest to map regions of interest of different sizes to feature maps of fixed size;

[0034] S35 inputs the feature map of the region of interest after RoI Align into the first network branch, uses the first fully connected layer to output a multi-category classification, that is, the category information of the rust, and uses the second fully connected layer to output a bounding box result, that is, the location information of the rust;

[0035] S36 inputs the feature map of the region of interest after RoI Align into the second network branch, uses a 3*3 convolutional layer to enhance the feature map, uses a transposed convolutional layer to upsample the feature map, and generates a mask prediction for each category, that is, the segmentation information of the rust.

[0036] Based on the above technical solution, preferably, step S32 includes:

[0037] After performing 1*1 convolution on C5, the second-layer pyramid feature map P5 is constructed first: P5 = Conv1*1(C5);

[0038] P5 is subjected to 3*3 convolution and maximum pooling to construct the highest-level pyramid feature map P6: P6 = Maxpool2*2(Conv3*3(P5));

[0039] Upsample P5 to the size of C4, add the upsampled P5 and C4 after 1*1 convolution, and construct the third-layer pyramid feature map P4: P4=Conv3*3(Upsample(P5)+Conv1*1(C4));

[0040] Upsample P4 to the size of C3, add the upsampled P4 and C3 after 1*1 convolution, and construct the fourth-layer pyramid feature map P3: P3 = Conv3*3(Upsample(P4)+Conv1*1(C3));

[0041] Upsample P3 to the size of C2, add the upsampled P3 and C2 after 1*1 convolution, and construct the fifth-layer pyramid feature map P2: P2 = Conv3*3(Upsample(P3)+Conv1*1(C2));

[0042] The final five-layer pyramid feature map includes P2, P3, P4, P5, and P6, and each layer of the feature map has the same number of channels.

[0043] Based on the above technical solution, preferably, step S34 includes:

[0044] Get the coordinates of the filtered region of interest on the original image;

[0045] Set the target feature map size to N*N;

[0046] Map the coordinates of each region of interest from the original image scale to the scale of the target feature map;

[0047] Each mapped region of interest is evenly divided into grids of the size of the target feature map, and each grid is used as a subwindow;

[0048] Select sampling points in each sub-window, and use the interpolation algorithm to calculate the eigenvalue of each sampling point;

[0049] Perform an average operation on the characteristic values ​​of the sampling points in each sub-window to generate a single characteristic value;

[0050] The feature values ​​of all sub-windows are combined to form the final feature map of the region of interest after RoI Align, which is the same size as the target feature map.

[0051] Based on the above technical solution, preferably, step S35 includes:

[0052] Flatten the feature map of the region of interest after RoI Align into a one-dimensional vector;

[0053] The one-dimensional vector is input into the first fully connected layer, and the original category score is output after the activation function;

[0054] Apply the softmax function to the output of the first fully connected layer to convert the original category scores into probability distributions and obtain the probability of each category;

[0055] The flattened one-dimensional vector is fed into the second fully connected layer to predict four bounding box adjustment values ​​(Δx, Δy, Δw, Δh) for each category.

[0056] Fine-tune the position and size of the region of interest based on the bounding box adjustment value corresponding to the highest value in the probability of each class;

[0057] The final output is the category and location information of the rust.

[0058] Based on the above technical solution, preferably, step S36 includes:

[0059] The feature map of the region of interest after RoI Align is input into the second network branch, and two 3*3 convolutional layers are used to extract and enhance features to obtain the enhanced feature map;

[0060] Use the transposed convolution layer to upsample the enhanced feature map;

[0061] The upsampled feature map is fed into the final convolutional layer to generate mask predictions for each category.

[0062] Apply the sigmoid activation function to the output of the final convolutional layer to compress the value of each pixel to between 0 and 1, indicating the probability that the pixel belongs to the rusted area;

[0063] Align the activated mask with the original ROI and adjust it to the size of the original ROI;

[0064] Output the final mask, i.e. the segmentation information of the rust.

[0065] On the basis of the above technical solution, preferably, step S4 includes:

[0066] S41 obtains the segmentation result of the rust, sets a threshold, and binarizes the mask so that the value in the mask is 0 or 1;

[0067] S42 establishes a mapping relationship from pixels to actual sizes;

[0068] S43 calculates the number of values ​​1 in the mask, converts and calculates the corrosion area using a mapping relationship;

[0069] S44 obtains the total area of ​​the bolts and calculates the corrosion percentage according to the corrosion area and the total area of ​​the bolts;

[0070] S45 divides the health level into health level 1, health level 2, health level 3, health level 4 and health level 5. Among them, the health level 1 bolts have the lowest degree of rust, and the health level 5 bolts have the highest degree of rust.

[0071] S46 divides the range of the corrosion percentage into five ranges, which correspond to health level 1, health level 2, health level 3, health level 4, and health level 5 from low to high;

[0072] S47 obtains the health level of the bolt scan image according to the corrosion percentage calculated in step S44.

[0073] Based on the above technical solution, preferably, step S42 includes:

[0074] Get camera parameters, including focal length f, sensor size Sensor width and Sensor height , Image resolution width and Image height ;

[0075] Use a rangefinder to measure the shooting distance D from the camera lens to the detection bolt;

[0076] Calculate the horizontal and vertical field of view FOV:

[0077] FOV horizontal =2*arctan(Sensor width / (2*f));

[0078] FOV vertical =2*arctan(Sensor height / (2*f));

[0079] Calculate the actual field of view size:

[0080] Real width =2*D*tan(FOV horizontal / 2);

[0081] Real height =2*D*tan(FOV vertical / 2);

[0082] Calculate the mapping from pixels to millimeters:

[0083] X scale =Real width / Image width ;

[0084] Y scale =Real height / Image height .

[0085] Compared with the prior art, the present invention has the following beneficial effects:

[0086] (1) Automatic detection and health assessment of bolt corrosion are achieved through deep learning models. This method not only significantly improves detection efficiency and accuracy, but also overcomes the subjectivity and safety risks of traditional manual detection. By combining image preprocessing, deep learning models and health assessment rules, the entire process from image input to health assessment is automated, providing reliable technical support for the maintenance and management of shield tunnels and helping to extend the service life of tunnels.

[0087] (2) The HSV color space conversion and exponential operation preprocessing method adopted by the present invention effectively enhances the contrast between the rusted area of ​​the bolt and the background. By performing exponential operations on the S and V channels, the saturation and brightness of the image are flexibly adjusted, making the rusted area more visually prominent while maintaining the overall hue of the image. This preprocessing method not only improves the detection accuracy of the subsequent deep learning model, but also enhances the model's adaptability to bolt images taken under different lighting conditions, thereby improving the robustness and versatility of the entire system;

[0088] (3) The deep learning model designed by the present invention includes a feature extraction module, a region generation module and a dual-branch network structure, which realizes the precise positioning, classification and segmentation of bolt corrosion. The introduction of the feature pyramid network effectively integrates feature information of different scales and improves the model's ability to detect rusted areas of different sizes. The application of RoI Align solves the problem of feature misalignment and improves position accuracy. The dual-branch network structure simultaneously outputs category information, location information and segmentation information, realizing a comprehensive analysis of the rusted area;

[0089] (4) The health evaluation rule established by the present invention realizes the quantitative evaluation of the health status of the bolts by mapping the percentage of the corrosion area to five health levels. This evaluation method not only objectively reflects the degree of corrosion of the bolts, but also provides a clear grading standard, which is convenient for managers to quickly determine the maintenance needs of the bolts. By establishing a mapping relationship from pixels to actual sizes, the practical significance of the evaluation results is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0091] Figure 1 is a flow chart of the method of the present invention;

[0092] Figure 2 It is a technical flow chart of the present invention;

[0093] Figure 3 It is the structure diagram of ResNet101 of the present invention;

[0094] Figure 4 It is a schematic diagram of a feature pyramid of the present invention;

[0095] Figure 5 It is a diagram of the segmentation network structure of the present invention;

[0096] Figure 6 It is a schematic diagram of RoI Align of the present invention. DETAILED DESCRIPTION

[0097] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0098] like Figure 1 As shown, the present invention provides an image-based shield tunnel bolt health evaluation method, comprising:

[0099] S1 obtains a bolt scanning image in the tunnel and pre-processes the bolt scanning image;

[0100] S2 builds a deep learning model, including a backbone network and a segmentation network;

[0101] S3 inputs the preprocessed bolt scanning image into the deep learning model for corrosion detection, uses the backbone network to extract feature maps, sets regions of interest for the feature maps, and screens the regions of interest. The screened regions of interest are input into the segmentation network to obtain the corrosion detection results of the bolt scanning image, including the location information, category information, and segmentation information of the corrosion;

[0102] S4 calculates the corrosion area according to the corrosion detection result, evaluates the bolt based on the corrosion area according to the health evaluation rule, and obtains the health evaluation result of the bolt scanning image.

[0103] like Figure 2 As shown in the figure, the technical implementation process of the present invention is as follows: 1) obtaining a shield tunnel segment bolt corrosion image; 2) performing contrast transformation on the image, i.e. preprocessing, to perform image enhancement; 3) performing feature extraction on the image, and using the feature extraction module of the deep learning model to realize multi-scale feature extraction and pyramid feature construction; 4) inputting the extracted features into a segmentation network, which has two branches, branch one is a class, box branch, i.e. a target detection branch, which is used to obtain the corrosion position and category information results in the shield tunnel bolt corrosion image, and branch two is a mask branch, i.e. a segmentation branch, which is used to obtain the image. The segmentation information result of the object in the image; 5) According to the output of the segmentation network, the segmentation and positioning result of the bolt corrosion is obtained; 6) The mask of the corroded area is binarized and the mask is set to 1 or 0, where the value of the mask is 1, which means that the pixel belongs to the corroded area; 7) The pixel points of the binarized mask are calculated, and the number of white pixels whose value is 1 is counted to obtain the corroded area; 8) A mapping relationship between pixels and actual sizes is established to obtain the actual corroded area and the total area of ​​the bolts. According to the proportion of the corroded area to the total area of ​​the bolts, the corrosion of the shield tunnel segment bolts is rated for safety.

[0104] Specifically, in one embodiment of the present invention, step S1 includes:

[0105] S11 converts the bolt scanning image from the original RGB space to the HSV color space to obtain an HSV image;

[0106] S12 separates the H, S and V channels of the HSV image;

[0107] S13 performs exponential operation on the S and V channels respectively, wherein the exponential factor ranges from 0.25 to 4;

[0108] S14 keeps the H channel unchanged, and recombines the S and V channels processed in step S13 with the original H channel;

[0109] S15 converts the enhanced HSV image obtained in step S14 back to the RGB space to obtain a preprocessed bolt scanning image.

[0110] In this embodiment, the cvtColor function can be used to convert the RGB image into the HSV format, such as: HSV image = cv2.cvtColor(rgb_image,cv2.COLOR_RGB2HSV); Use the split function to decompose the HSV image into three channels: H, S, and V: H, S, V = cv2.split(HSV image ); perform exponential operation on S and V channels, with the exponential factor γ ranging from 0.25 to 4: S enhanced =np.power(S / 255.0,γ)*255,V enhanced = np.power(V / 255.0,γ)*255, convert the result back to uint8 type; use the Merge function to recombine the processed S and V channels with the original H channel: HSV enhanced =cv2.merge([H,S enhanced ,V enhanced ]); Use the cvtColor function to convert the enhanced HSV image back to RGB format: RGB enhanced =cv2.cvtColor(HSV enhanced ,cv2.COLOR_HSV2RGB).

[0111] This embodiment can effectively enhance the contrast of the image by performing exponential operations on the S (saturation) and V (brightness) channels. This processing method is particularly useful for highlighting the rusted area on the surface of the bolt, making it more clearly distinguishable from the background. Keeping the H (hue) channel unchanged ensures that the basic color information of the original image is not changed. This is very important for maintaining the characteristic color of the rusted area, which helps with subsequent identification and classification. By adjusting the value of the exponential factor γ, it can be flexibly adapted to images taken under different lighting conditions. This adaptability enables the preprocessing method to process bolt images acquired in various environments. By adjusting the V channel, the impact of uneven lighting on image quality can be reduced to a certain extent.

[0112] Specifically, in one embodiment of the present invention, the backbone network includes a feature extraction module and a region generation module, and the segmentation network includes a first network branch and a second network branch, wherein:

[0113] The feature extraction module includes a 7*7 convolutional layer, a 3*3 maximum pooling layer, 4 residual blocks, and a feature pyramid. Each residual block contains multiple residual units. Each residual unit includes 2 1*1 convolutional layers and 1 3*3 convolutional layer. The feature pyramid adopts a top-down path.

[0114] The region generation module includes a shared convolutional layer, a classification layer, and a bounding box regression layer;

[0115] The first network branch includes two fully connected layers;

[0116] The second network branch includes a 3*3 convolutional layer and a transposed convolutional layer.

[0117] Specifically, in the deep learning model of this embodiment, the feature extraction module includes a feature extraction unit and a pyramid unit. The feature extraction unit adopts the ResNet101 architecture, removes the last average pooling layer and the fully connected layer, and only retains the feature extraction architecture. The ResNet101 structure is as follows: Figure 3 As shown, the pyramid unit adopts the FPN structure, such as Figure 4 As shown in the figure, the segmentation network consists of two branches, one is the target detection branch, namely the class and box branch, and the other is the segmentation branch. The structure of the segmentation network is shown in the figure Figure 5 shown.

[0118] Specifically, after the deep learning model is built, model training is required. The training process is as follows:

[0119] 1. Prepare a dataset of shield tunnel bolt images with corrosion location annotations, category labels, and segmentation masks, and enhance the dataset using the preprocessing method of step S1 to increase illumination changes.

[0120] 2. Set the training strategy:

[0121] Initial stage: freeze the feature extraction module and only train the region generation module and segmentation network.

[0122] Fine-tuning stage: Unfreeze the high layers of the feature extraction module, i.e., the third and fourth residual blocks, and perform end-to-end training.

[0123] Fully fine-tuning: Unfreeze all layers and fine-tune with a small learning rate.

[0124] 3. Set training parameters:

[0125] Learning rate setting: Use the warmup strategy to gradually increase the learning rate to the initial value, and use the cosine annealing strategy to reduce the learning rate.

[0126] Regularization: Use weight decay to prevent overfitting and selectively apply Dropout in each fully connected layer.

[0127] 4. Perform joint training, the comprehensive loss function is: Ltotal = L RPN +L det +L seg , where L RPN is the loss function of the backbone network, L det is the loss function of the first network branch, L segis the loss function of the second network branch.

[0128] Backbone network training:

[0129] Use the pre-trained ResNet101 as the initial weight, fine-tune ResNet101 to adapt to the bolt corrosion detection task, build a feature pyramid network, namely FPN, on top of ResNet101, and use the output feature map of FPN to train the region generation network, namely RPN. The loss function of the backbone network is: L RPN =L cls +λ1*L reg , where L cls is the binary cross entropy loss, i.e. the prediction of foreground or background, L reg is the smooth L1 loss for bounding box regression.

[0130] First network branch training:

[0131] Using the region of interest generated by RPN and combining it with RoI Align to extract features, classification and border fine-tuning are performed through two fully connected layers. The loss function of the first network branch is: L det =L clsdet +λ2*L regdet , where L clsdet is the multi-class cross entropy loss, used to predict multiple categories, L regdet is the smoothed L1 loss for rusted bounding box regression.

[0132] Second network branch training:

[0133] Using the region of interest generated by RPN, combined with RoI Align to extract features, upsampling is performed through 3*3 convolution and transposed convolution. The loss function of the second network branch is: L seg =α1*L boundary +α2*L dice .

[0134] L boundary =ω·BCE(P boundary ,T boundary )+(1-ω)·BCE(P non-boundary ,T non-boundary )

[0135]

[0136] Among them, L boundary is the boundary perception loss, ω is the boundary weight, which is dynamically adjusted according to the proportion of boundary pixels, P boundary is the predicted boundary area, T boundary is the true boundary area, P non-boundary is the predicted non-boundary area, Tnon-boundary is the real non-boundary area, L dice is the Dice loss, X is the predicted segmentation mask, Y is the actual segmentation mask, ε is the smoothing term, and BCE is the binary cross entropy loss.

[0137] The entire model is optimized using the back-propagation algorithm and stochastic gradient descent until the comprehensive loss function reaches the minimum value, that is, the training is completed and a trained deep learning model is obtained.

[0138] Specifically, in one embodiment of the present invention, step S3 includes:

[0139] S31 inputs the preprocessed bolt scan image into the feature extraction module, uses the 7*7 convolution layer to extract the first feature map C1, and then sequentially inputs 4 residual blocks to extract 4 feature maps, namely the second feature map C2, the third feature map C3, the fourth feature map C4, and the fifth feature map C5, wherein the scale of each feature map is different, and the scale of the feature map increases in sequence according to the data flow direction;

[0140] S32 uses a top-down path to construct a pyramid and obtains a five-layer pyramid feature map;

[0141] S33 uses a fixed-size sliding window on the five-layer pyramid feature map, generates multiple anchor boxes of preset sizes and aspect ratios at the center of the sliding window, applies a 3*3 shared convolutional layer to each sliding window position, uses a classification layer for binary classification, predicts the probability of each anchor box being the foreground or background, uses a bounding box regression layer for bounding box regression, predicts the adjustment parameters of each anchor box, selects the N anchor boxes with the highest scores according to the classification scores, applies NMS to the N anchor boxes, and uses the NMS-posted anchor boxes as the filtered regions of interest;

[0142] S34 applies RoI Align to each filtered region of interest to map regions of interest of different sizes to feature maps of fixed size;

[0143] S35 inputs the feature map of the region of interest after RoI Align into the first network branch, uses the first fully connected layer to output a multi-category classification, that is, the category information of the rust, and uses the second fully connected layer to output a bounding box result, that is, the location information of the rust;

[0144] S36 inputs the feature map of the region of interest after RoI Align into the second network branch, uses a 3*3 convolutional layer to enhance the feature map, uses a transposed convolutional layer to upsample the feature map, and generates a mask prediction for each category, that is, the segmentation information of the rust.

[0145] In this embodiment, step S31 includes:

[0146] The preprocessed bolt scan image is input into the feature extraction module, and the 7*7 convolution layer is used to capture the low-level features of the image, such as edges, textures, etc. The 7*7 convolution layer includes 64 convolution kernels with a step size of 2. The first feature map C1 is output here, and its size is 1 / 2 of the original image.

[0147] A 3*3 maximum pooling layer is connected after the 7*7 convolutional layer with a step size of 2 to reduce the size of the first feature map C1 and increase the receptive field.

[0148] The first feature map C1 after pooling is input into the residual structure. The residual structure includes 4 residual blocks. Each residual block contains multiple residual units. The first residual block includes 3 residual units. Its input is the feature map from the pooling layer, and its output is the second feature map C2, whose size is 1 / 4 of the original image; the second residual block includes 4 residual units, its input is C2, and its output is the third feature map C3, whose size is 1 / 8 of the original image; the third residual block includes 23 residual units. Its input is C3, and its output is the fourth feature map C4, which is 1 / 16 of the original image; the fourth residual block includes 3 residual units, its input is C4, and its output is the fifth feature map C5, which is 1 / 32 of the original image; each residual unit includes 1*1 convolution (dimensionality reduction), 3*3 convolution (feature extraction), 1*1 convolution (dimensionality increase), residual connection (adding the input directly to the output), and these 4 residual blocks generate feature maps C2, C3, C4, and C5 in turn. The feature map scale C1>C2>C3>C4>C5, smaller feature maps (such as C5) have larger receptive fields and are suitable for detecting large-sized rusted areas; larger feature maps (such as C2) retain more details and are suitable for detecting small-sized rusted areas.

[0149] Specifically, in one embodiment of the present invention, step S32 includes:

[0150] After performing 1*1 convolution on C5, the second-layer pyramid feature map P5 is constructed first: P5=Conv1*1(C5); wherein, the 1*1 convolution is to adjust the number of channels, and at this time, high-level semantic information is retained to match the subsequent layers.

[0151] P5 is subjected to 3*3 convolution and maximum pooling to construct the highest-level pyramid feature map P6: P6=Maxpool2*2(Conv3*3(P5)); This embodiment introduces an additional P6 on top of the deepest feature output by the ResNet101 structure as the highest-level feature of the feature pyramid in order to further increase the receptive field and increase the feature information.

[0152] P5 is upsampled to the size of C4, and the upsampled P5 and C4 are added together after 1*1 convolution to construct the third-layer pyramid feature map P4: P4=Conv3*3(Upsample(P5)+Conv1*1(C4)); this layer of features integrates high-level semantic information and mid-level features to enhance feature expression.

[0153] P4 is upsampled to the size of C3, and the upsampled P4 and C3 are added after 1*1 convolution to construct the fourth-layer pyramid feature map P3: P3=Conv3*3(Upsample(P4)+Conv1*1(C3)); this layer further integrates the mid- and low-level features to enhance the expression of details.

[0154] P3 is upsampled to the size of C2, and the upsampled P3 is added to the result of 1*1 convolution of C2 to construct the fifth-layer pyramid feature map P2: P2=Conv3*3(Upsample(P3)+Conv1*1(C2)); this layer integrates low-level features to enhance fine-grained expression.

[0155] The final five-layer pyramid feature map includes P2, P3, P4, P5, and P6, and each layer of the feature map has the same number of channels.

[0156] This embodiment constructs a pyramid network after the feature extraction unit. The generated feature pyramid contains multiple resolution levels, which is conducive to detecting targets of different sizes. By upsampling high-level semantic information and transferring it to the low-level layers, the semantic expression ability of the low-level features is enhanced, and the detail information of the low-level features is retained and utilized, which helps to improve the performance of small target detection and fine segmentation.

[0157] Specifically, in one embodiment of the present invention, step S33 includes:

[0158] Apply a sliding window of the same size, such as a 3*3 window, to each layer of feature maps of P2, P3, P4, P5, and P6. This sliding window moves on each layer of feature maps to cover all positions.

[0159] For each sliding window position, multiple anchor boxes are generated, such as 3 scales, 3 aspect ratios, and a total of 9 preset anchor boxes. The size of the anchor box varies with the level of the feature map, with larger anchor boxes at higher levels and smaller anchor boxes at lower levels.

[0160] A 3*3 shared convolution layer is applied to each sliding window position. This shared convolution layer shares parameters across all feature maps and all positions and outputs a feature vector of fixed dimension.

[0161] For each anchor box, the probability of predicting whether it is a foreground (including the target) or background is predicted. Specifically, this can be achieved using a 1*1 convolution, and the number of output channels is twice the number of anchor boxes.

[0162] Predict the adjustment parameters of each anchor box, also using 1x1 convolution, and the number of output channels is 4 times the number of anchor boxes.

[0163] Sort the foreground scores of all anchor boxes and select the first N anchor boxes with the highest scores. In this embodiment, N is 1000.

[0164] Apply NMS (non-maximum suppression) to N anchor boxes, calculate the IoU between the anchor boxes, retain the anchor box with the highest score, suppress other anchor boxes with high overlap, and repeat this process until all anchor boxes are processed.

[0165] The anchor box after NMS is used as the filtered region of interest.

[0166] Specifically, Figure 6 As shown, in one embodiment of the present invention, step S34 includes:

[0167] Get the coordinates of the filtered region of interest on the original image;

[0168] Set the target feature map size to N*N;

[0169] Map the coordinates of each region of interest from the original image scale to the scale of the target feature map;

[0170] Each mapped region of interest is evenly divided into grids of the size of the target feature map, and each grid is used as a subwindow;

[0171] Select sampling points in each sub-window, and use the interpolation algorithm to calculate the eigenvalue of each sampling point;

[0172] Perform an average operation on the characteristic values ​​of the sampling points in each sub-window to generate a single characteristic value;

[0173] The feature values ​​of all sub-windows are combined to form the final feature map of the region of interest after RoI Align, which is the same size as the target feature map.

[0174] Specifically, in this embodiment, an adaptive sampling strategy is adopted when selecting sampling points, as follows:

[0175] The number of sampling points is calculated based on the area and aspect ratio of the region of interest. The calculation function is: Number of sampling points = base samples *(1+α*log(area / base area )+β*|aspect ratio -1|), where base samples is the number of basic sampling points, area is the area of ​​the area of ​​interest, base areais a base area, aspect ratio is the aspect ratio of the region of interest, α and β are adjustable hyperparameters. The number of sampling points is rounded up to the nearest square.

[0176] In this embodiment, the interpolation algorithm is as follows:

[0177] Define a learnable interpolation kernel K(x,y) for each sampling point. K(x,y) is a small neural network. The input is the offset of the sampling point relative to its nearest four integer coordinates. The network outputs four weights w1, w2, w3, w4, corresponding to the nearest four integer coordinates. Eigenvalue calculation formula: f(x,y) = w1*f 11 +w2*f 21 +w3*f 12 +w4*f 22 , where f 11 、f 21 、f 12 、f 22 are the eigenvalues ​​of the four nearest integer coordinate points.

[0178] In this embodiment, an adaptive sampling strategy is used to dynamically adjust according to the size and shape of the region of interest, making the extraction of different features more flexible and accurate. The interpolation process is transformed into a learnable process, allowing the network to automatically adjust the best feature extraction method according to task requirements. By adjusting the α and β parameters, the sensitivity of the sampling strategy to the size and shape of the region of interest can be controlled in a fine-grained manner.

[0179] Specifically, in one embodiment of the present invention, step S35 includes:

[0180] Flatten the feature map of the region of interest after RoI Align into a one-dimensional vector; for example, flatten the 7*7*256 feature map into a one-dimensional vector of length NC=12544 to convert the spatial information into a sequence form.

[0181] The one-dimensional vector is input into the first fully connected layer, and the original category score is output after the activation function; the weight matrix size of the first fully connected layer is: (NC)*(num classes +1), and the activation function uses ReLu to increase nonlinearity.

[0182] Apply the softmax function to the output of the first fully connected layer to convert the original category scores into probability distributions and obtain the probability of each category; ensure that the sum of all category probabilities is 1. Formula: Its output is the probability of each class, including the background class.

[0183] The flattened one-dimensional vector is input into the second fully connected layer to predict 4 bounding box adjustment values ​​(Δx, Δy, Δw, Δh) for each category; the weight matrix size of the second fully connected layer is: (NC)*(4*num classes ). This fully connected layer does not apply an activation function, and the resulting adjustment value can be positive or negative.

[0184] Fine-tune the position and size of the region of interest based on the bounding box adjustment value corresponding to the highest value in the probability of each category; select the non-background category with the highest probability, extract the bounding box adjustment value corresponding to this category, and use these values ​​to adjust the coordinates and size of the original region of interest. The adjustment formula is: new =x+w*Δx,y new =y+h*Δy,w new =w*e Δw ,h new =h*e Δh , where (x, y, w, h) are the original center coordinates, width, and height of the region of interest.

[0185] The final output is the category information and location information of the rust. The category information is the non-background category with the highest probability and the softmax probability of the category, and the location information is the fine-tuned bounding box coordinates (x new ,y new ,w new ,h new ).

[0186] Specifically, in one embodiment of the present invention, step S36 includes:

[0187] The feature map of the region of interest after RoI Align is input into the second network branch, and two 3*3 convolutional layers are used to extract and enhance features to obtain the enhanced feature map.

[0188] Specifically, the two convolutional layers are 3*3 convolutions with 256 filters and ReLu activation. Two convolutional layers are used to further extract and enhance features and increase the receptive field of the network.

[0189] The enhanced feature maps are upsampled using transposed convolutional layers.

[0190] Specifically, the transposed convolution layer is a 2*2 transposed convolution with a step size of 2 and 256 filters. This layer enlarges the size of the feature map and restores higher-resolution spatial information, which is conducive to generating more refined masks.

[0191] The upsampled feature maps are fed into the final convolutional layer to generate mask predictions for each class.

[0192] Specifically, the final convolutional layer uses 1*1 convolution, the number of output channels is the number of categories (excluding background), and each channel corresponds to the mask prediction of one category.

[0193] The sigmoid activation function is applied to the output of the final convolutional layer to compress the value of each pixel to between 0 and 1, indicating the probability that the pixel belongs to the rusted area. Sigmoid function: f(x) = 1 / (1+e -x ), compressing the pixel value to between 0 and 1, where 0 means the pixel does not belong to the rust area and 1 means it completely belongs to the rust area.

[0194] Align the activated mask with the original ROI and resize it to the original ROI size.

[0195] The mask can be adjusted to the size of the original region of interest using bilinear interpolation to ensure that the mask accurately matches the detected rust areas.

[0196] Output the final mask, i.e. the segmentation information of the rust.

[0197] In this embodiment, after sigmoid activation, morphological operations can also be performed on the mask to effectively refine the segmentation result, remove noise, fill small holes, and make the mask smoother and more accurate. The morphological process is as follows:

[0198] The mask is binarized using Otsus, and structural elements of different sizes are used for opening and closing operations. Structural elements are gradually applied from small to large to handle noise points and holes of different sizes.

[0199] Noise removal: Use opening operation, first erosion, then expansion.

[0200] Small hole filling: Use closing operation, first dilation, then erosion.

[0201] Specifically, the corrosion operation is: D(x,y)=min{D(x+i,y+j)+d(i,j)}, and the expansion operation is: D(x,y)=max{D(xi,yj)-d(i,j)}, where D is the distance map and d is the distance from the center of the structural element to (i,j).

[0202] Median filtering is used to smooth the boundaries, and connected region analysis is performed to remove regions that are too small.

[0203] The processed mask is fused with the original sigmoid output to obtain the morphologically processed mask.

[0204] Specifically, in one embodiment of the present invention, step S4 includes:

[0205] S41 obtains the segmentation result of the rust, sets a threshold, and binarizes the mask so that the value in the mask is 0 or 1;

[0206] S42 establishes a mapping relationship from pixels to actual sizes;

[0207] S43 calculates the number of values ​​1 in the mask, converts and calculates the corrosion area using a mapping relationship;

[0208] S44 obtains the total area of ​​the bolts and calculates the corrosion percentage according to the corrosion area and the total area of ​​the bolts;

[0209] S45 divides the health level into health level 1, health level 2, health level 3, health level 4 and health level 5. Among them, the health level 1 bolts have the lowest degree of rust, and the health level 5 bolts have the highest degree of rust.

[0210] S46 divides the range of the corrosion percentage into five ranges, which correspond to health level 1, health level 2, health level 3, health level 4, and health level 5 from low to high;

[0211] S47 obtains the health level of the bolt scan image according to the corrosion percentage calculated in step S44.

[0212] Specifically, in one embodiment of the present invention, step S42 includes:

[0213] Get camera parameters, including focal length f, sensor size Sensor width and Sensor height , Image resolution width and Image height ;

[0214] Use a rangefinder to measure the shooting distance D from the camera lens to the detection bolt;

[0215] Calculate the horizontal and vertical field of view FOV:

[0216] FOV horizontal =2*arctan(Sensor width / (2*f));

[0217] FOV vertical =2*arctan(Sensor height / (2*f));

[0218] Calculate the actual field of view size:

[0219] Real width =2*D*tan(FOV horizontal / 2);

[0220] Real height =2*D*tan(FOV vertical / 2);

[0221] Calculate the mapping from pixels to millimeters:

[0222] X scale =Real width / Image width ;

[0223] Y scale =Real height / Image height .

[0224] In another embodiment of the present invention, considering the lens distortion and tilt angle issues, another method is used to establish a mapping relationship from pixels to actual sizes:

[0225] Get camera parameters, including focal length f, sensor size Sensor width and Sensor height , Image resolution width and Image height , distortion coefficients k1, k2 (radial distortion) and p1, p2 (tangential distortion);

[0226] Use a rangefinder to measure the shooting distance D from the camera lens to the detection bolt; obtain the camera tilt angle θ;

[0227] Calculate the horizontal and vertical field of view FOV:

[0228] FOV horizontal =2*arctan(Sensor width / (2*f));

[0229] FOV vertical =2*arctan(Sensor height / (2*f));

[0230] Calculate the actual field of view size:

[0231] Real width =2*D*tan(FOV horizontal / 2) / cosθ;

[0232] Real height =2*D*tan(FOV vertical / 2) / cosθ;

[0233] Calculate the initial pixel to millimeter mapping:

[0234] X scale-initial=Real width / Image width ;

[0235] Y scale-initial =Real height / Image height .

[0236] For each pixel (x,y) in the image, perform the following correction:

[0237] Normalize pixel coordinates:

[0238] x normalized =(x-Image width / 2) / Image width ;

[0239] y normalized =(y-Image height / 2) / Image height ;

[0240] Calculate radial distortion:

[0241] r2=x normalized 2 +y normalized 2 ;

[0242] radial distortion =1+k1*r2+k2*r2 2 ;

[0243] Calculate the tangential distortion:

[0244] tangential x =2*p1*x normalized *y normalized +p2*(r2+2*x normalized 2 )

[0245] tangential y =p1*(r2+2*y normalized 2 )+2*p2*x normalized *y normalized ;

[0246] Apply distortion correction:

[0247] x corrected =x normalized *radial distortion +tangential x ;

[0248] ycorrected =y normalized *radial distortion +tangential y ;

[0249] Convert the rectified coordinates back to pixel coordinates:

[0250] x final =x corrected *Image width +Image width / 2;

[0251] y final =y corrected *Image height +Image height / 2;

[0252] Calculate the final pixel-to-mm mapping:

[0253] X scale-final =X scale-initial *(Image width / max(x final -min(x final )));

[0254] Y scale-final =Y scale-initial *(Image height / max(y final -min(y final ))).

[0255] After determining the mapping relationship, the rust area is calculated according to the number of pixels in the mask with a value of 1. The total area of ​​the bolt is obtained, and the rust percentage can be calculated, which is the percentage of the rust area to the total area of ​​the bolt.

[0256] The rust percentage range of 0-100% is divided into 5 intervals, 0%: health level 1; 1-10%: health level 2; 10-25%: health level 3; 25-40%: health level 4; above 40%: health level 5.

[0257] Based on the calculated rust percentage, the corresponding health level can be determined.

[0258] The above description is only a preferred embodiment of the present invention and is 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 in the protection scope of the present invention.

Claims

1. An image-based shield tunnel bolt health evaluation method, characterized in that: include: S1 obtains a bolt scanning image in the tunnel and pre-processes the bolt scanning image; S2 builds a deep learning model, including a backbone network and a segmentation network; S3 inputs the preprocessed bolt scanning image into the deep learning model for corrosion detection, uses the backbone network to extract feature maps, sets regions of interest for the feature maps, and screens the regions of interest. The screened regions of interest are input into the segmentation network to obtain the corrosion detection results of the bolt scanning image, including the location information, category information, and segmentation information of the corrosion; S4 calculates the corrosion area according to the corrosion detection result, evaluates the bolt based on the corrosion area according to the health evaluation rule, and obtains the health evaluation result of the bolt scanning image.

2. The image-based shield tunnel bolt health evaluation method according to claim 1, characterized in that: The backbone network includes a feature extraction module and a region generation module, and the segmentation network includes a first network branch and a second network branch, wherein: The feature extraction module includes a 7*7 convolutional layer, a 3*3 maximum pooling layer, 4 residual blocks, and a feature pyramid. Each residual block contains multiple residual units. Each residual unit includes 2 1*1 convolutional layers and 1 3*3 convolutional layer. The feature pyramid adopts a top-down path. The region generation module includes a shared convolutional layer, a classification layer, and a bounding box regression layer; The first network branch includes two fully connected layers; The second network branch includes a 3*3 convolutional layer and a transposed convolutional layer.

3. The image-based shield tunnel bolt health evaluation method according to claim 1, characterized in that: Step S1 includes: S11 converts the bolt scanning image from the original RGB space to the HSV color space to obtain an HSV image; S12 separates the H, S and V channels of the HSV image; S13 performs exponential operation on the S and V channels respectively, wherein the exponential factor ranges from 0.25 to 4; S14 keeps the H channel unchanged, and recombines the S and V channels processed in step S13 with the original H channel; S15 converts the enhanced HSV image obtained in step S14 back to the RGB space to obtain a preprocessed bolt scanning image.

4. The image-based shield tunnel bolt health evaluation method according to claim 2, characterized in that: Step S3 includes: S31 inputs the preprocessed bolt scan image into the feature extraction module, uses the 7*7 convolution layer to extract the first feature map C1, and then sequentially inputs 4 residual blocks to extract 4 feature maps, namely the second feature map C2, the third feature map C3, the fourth feature map C4, and the fifth feature map C5, wherein the scale of each feature map is different, and the scale of the feature map increases in sequence according to the data flow direction; S32 uses a top-down path to construct a pyramid and obtains a five-layer pyramid feature map; S33 uses a fixed-size sliding window on the five-layer pyramid feature map, generates multiple anchor boxes of preset sizes and aspect ratios at the center of the sliding window, applies a 3*3 shared convolutional layer to each sliding window position, uses a classification layer for binary classification, predicts the probability of each anchor box being the foreground or background, uses a bounding box regression layer for bounding box regression, predicts the adjustment parameters of each anchor box, selects the N anchor boxes with the highest scores according to the classification scores, applies NMS to the N anchor boxes, and uses the NMS-posted anchor boxes as the filtered regions of interest; S34 applies RoI Align to each filtered region of interest to map regions of interest of different sizes to feature maps of fixed size; S35 inputs the feature map of the region of interest after RoI Align into the first network branch, uses the first fully connected layer to output a multi-category classification, that is, the category information of the rust, and uses the second fully connected layer to output a bounding box result, that is, the location information of the rust; S36 inputs the feature map of the region of interest after RoI Align into the second network branch, uses a 3*3 convolutional layer to enhance the feature map, uses a transposed convolutional layer to upsample the feature map, and generates a mask prediction for each category, that is, the segmentation information of the rust.

5. The image-based shield tunnel bolt health evaluation method according to claim 4, characterized in that: Step S32 includes: After performing 1*1 convolution on C5, the second-layer pyramid feature map P5 is constructed first: P5 = Conv1*1(C5); P5 is subjected to 3*3 convolution and maximum pooling to construct the highest-level pyramid feature map P6: P6 = Maxpool2*2(Conv3*3(P5)); Upsample P5 to the size of C4, add the upsampled P5 and C4 after 1*1 convolution, and construct the third-layer pyramid feature map P4: P4=Conv3*3(Upsample(P5)+Conv1*1(C4)); Upsample P4 to the size of C3, add the upsampled P4 and C3 after 1*1 convolution, and construct the fourth-layer pyramid feature map P3: P3 = Conv3*3(Upsample(P4)+Conv1*1(C3)); Upsample P3 to the size of C2, add the upsampled P3 and C2 after 1*1 convolution, and construct the fifth-layer pyramid feature map P2: P2 = Conv3*3(Upsample(P3)+Conv1*1(C2)); The final five-layer pyramid feature map includes P2, P3, P4, P5, and P6, and each layer of the feature map has the same number of channels.

6. The image-based shield tunnel bolt health evaluation method according to claim 4, characterized in that: Step S34 includes: Get the coordinates of the filtered region of interest on the original image; Set the target feature map size to N*N; Map the coordinates of each region of interest from the original image scale to the scale of the target feature map; Each mapped region of interest is evenly divided into grids of the size of the target feature map, and each grid is used as a subwindow; Select sampling points in each sub-window, and use the interpolation algorithm to calculate the eigenvalue of each sampling point; Perform an average operation on the characteristic values ​​of the sampling points in each sub-window to generate a single characteristic value; The feature values ​​of all sub-windows are combined to form the final feature map of the region of interest after RoI Align, which is the same size as the target feature map.

7. The image-based shield tunnel bolt health evaluation method according to claim 4, characterized in that: Step S35 includes: Flatten the feature map of the region of interest after RoIAlign into a one-dimensional vector; The one-dimensional vector is input into the first fully connected layer, and the original category score is output after the activation function; Apply the softmax function to the output of the first fully connected layer to convert the original category scores into probability distributions and obtain the probability of each category; The flattened one-dimensional vector is fed into the second fully connected layer to predict four bounding box adjustment values ​​(Δx, Δy, Δw, Δh) for each category. Fine-tune the position and size of the region of interest based on the bounding box adjustment value corresponding to the highest value in the probability of each class; The final output is the category and location information of the rust.

8. The image-based shield tunnel bolt health evaluation method according to claim 4, characterized in that: Step S36 includes: The feature map of the region of interest after RoI Align is input into the second network branch, and two 3*3 convolutional layers are used to extract and enhance features to obtain the enhanced feature map; Use the transposed convolution layer to upsample the enhanced feature map; The upsampled feature map is fed into the final convolutional layer to generate mask predictions for each category. Apply the sigmoid activation function to the output of the final convolutional layer to compress the value of each pixel to between 0 and 1, indicating the probability that the pixel belongs to the rusted area; Align the activated mask with the original ROI and adjust it to the size of the original ROI; Output the final mask, i.e. the segmentation information of the rust.

9. The image-based shield tunnel bolt health evaluation method according to claim 4, characterized in that: Step S4 includes: S41 obtains the segmentation result of the rust, sets a threshold, and binarizes the mask so that the value in the mask is 0 or 1; S42 establishes a mapping relationship from pixels to actual sizes; S43 calculates the number of values ​​1 in the mask, converts and calculates the corrosion area using a mapping relationship; S44 obtains the total area of ​​the bolts and calculates the corrosion percentage according to the corrosion area and the total area of ​​the bolts; S45 divides the health level into health level 1, health level 2, health level 3, health level 4 and health level 5. Among them, the health level 1 bolts have the lowest degree of rust, and the health level 5 bolts have the highest degree of rust. S46 divides the range of the corrosion percentage into five ranges, which correspond to health level 1, health level 2, health level 3, health level 4, and health level 5 from low to high; S47 obtains the health level of the bolt scan image according to the corrosion percentage calculated in step S44.

10. The image-based shield tunnel bolt health evaluation method according to claim 9, characterized in that: Step S42 includes: Get camera parameters, including focal length f, sensor size Sensor width and Sensor height , Image resolution width and Image height ; Use a rangefinder to measure the shooting distance D from the camera lens to the detection bolt; Calculate the horizontal and vertical field of view FOV: FOV horizontal =2*arctan(Sensor width / (2*f)); FOV vertical =2*arctan(Sensor height / (2*f)); Calculate the actual field of view size: Real width =2*D*tan(FOV horizontal / 2); Real height =2*D*tan(FOV vertical / 2); Calculate the mapping from pixels to millimeters: X scale =Real width / Image width ; Y scale =Real height / Image height 。

Citation Information

Patent Citations

  • Bolt gusset plate corrosion detection method, device and equipment and storage medium

    CN112419298A

  • Gate leaf surface corrosion image detection system and corrosion area rapid measurement and calculation method

    CN112767364A

  • Tunnel bolt positioning and bolt corrosion area identification method

    CN115661534A

  • Prediction method for corrosion rate of steel bridge bolt

    CN116563626A

  • Hardware fitting corrosion segmentation evaluation method and device

    CN118154980A