Railway tunnel lining disease detection and evaluation method based on anchor-frame-free instance segmentation

Abstract

The invention relates to a railway tunnel lining disease detection and evaluation method based on anchor-frame-free instance segmentation, and belongs to the technical field of railway tunnel disease detection and evaluation. The method comprises the steps that firstly, marked railway tunnel lining disease data sets based on ground penetrating radar images are manufactured, a training set, a verification set and a test set are divided according to the proportion, and the images of all the data sets are not repeated; then, building a model based on deep learning, and carrying out iteration on model parameters through the training set to obtain a training result; thirdly, evaluating a result obtained by training through a verification set, and evaluating the reasoning speed of the training model through a test set; and finally, calculating the minimum length of the disease based on an image processing technology and an optimization algorithm, and evaluating the degradation level of the tunnel. According to the method, the ground penetrating radar image is used as a data source, the category and contour of the tunnel lining disease are output through the deep learning model, the tunnel degradation grade is evaluated, and compared with existing manual inspection, the working efficiency and accuracy are greatly improved.

Description

Technical Field

[0001] The invention relates to a railway tunnel lining defect detection and evaluation method based on anchor-free frame instance segmentation, and belongs to the technical field of railway tunnel defect detection and evaluation. Background Art

[0002] Railway tunnels, as important underground infrastructure, play a vital role in ensuring the smooth and safe operation of trains. Due to environmental factors such as long-term cyclic loads, insufficient maintenance, and groundwater ablation, various structural defects will inevitably occur in tunnel linings during construction and operation. Therefore, regular inspection of tunnel lining defects is essential. The type and size of lining defects are important indicators for evaluating the safety level of tunnels. At present, tunnel defects are mainly inspected manually, but they are highly subjective and inefficient. It is urgent to develop an intelligent and automated method for tunnel inspection. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a railway tunnel lining defect detection and evaluation method based on anchor-free instance segmentation. In the current situation where tunnel defects are mainly inspected manually, there are problems such as low work efficiency and inaccurate detection and evaluation.

[0004] The technical solution of the present invention is: a railway tunnel lining disease detection and evaluation method based on anchor-free instance segmentation, which integrates an efficient detection module and an evaluation system to solve the technical problems of low efficiency and high labor intensity of existing manual inspections. The specific steps are as follows:

[0005] S1: The railway tunnel lining images were collected by ground penetrating radar (GPR), and the categories were divided according to the different characteristics of various diseases. The real distribution of the diseases (Ground Truth, GT) was manually annotated to generate label files, which were randomly divided into training set, validation set and test set according to 7:2:1. Among them, the labeling tool was completed with the help of the target detection tool LabelImg. The main label categories include: voids, water-containing voids and looseness. The training set and validation set are divided one by one by images and label files. The test set only contains images, and the images are not repeated between the data sets.

[0006] S2: Input the training set and the validation set into a deep learning model, where the deep learning model includes an input end, a detection layer, a training layer, and an output end.

[0007] The input end is used to receive an image and a corresponding label file.

[0008] The detection layer extracts and fuses features of the input image. Feature extraction extracts high-level features layer by layer through a series of convolutional layers, pooling layers, and residual connections, while gradually reducing the spatial resolution. Specifically, feature extraction locates the coordinates of the detection target through the center point, predicts the category of the target at each positive sample in the prediction box, and uses the offset distance relative to the upper, lower, left, and right boundaries of the bounding box as the regression target, and then retains the center point with higher confidence through non-maximum suppression, filters out redundant points, and outputs the feature map.

[0009] The feature fusion consists of an embedding layer, three fusion layers, four convolution combinations and a spatial pyramid pooling layer, which fuse feature maps from different scales. Specifically, the embedding layer is a regular 4×4 convolution module with a step size of 4, and the fusion layer is a regular 2×2 convolution module with a step size of 2. Each convolution combination consists of a partial convolution, a point-by-point convolution, a normalized network, a ReLU activation function and a point-by-point convolution, which are connected in series in sequence. The spatial pyramid pooling layer consists of a regular 5×5 convolution module with a step size of 1, three maximum pooling layers, a Concat splicing layer and a residual connection. Specifically, partial convolution only acts on part of the input channels of the feature map to reduce redundant information, while the remaining channels use point-by-point convolution to filter limited semantic information to retain feature diversity and achieve lower latency. The normalized network and activation function are only after the point-by-point convolution to further improve the speed of network feature fusion.

[0010] The training layer includes a segmentation head and a loss function calculation control gradient for iterative training.

[0011] The segmentation head processes the feature map into a high-resolution output by combining convolutional layers, F-C2f modules and upsampling layers to gradually restore the spatial resolution, and finally generates a segmentation mask of the target object. Specifically, it includes extracting multi-scale features using multiple convolutional layers, and restoring the low-resolution feature map to the original resolution of the input image through layer-by-layer upsampling and fusion, and outputting a high-quality mask containing the segmentation information of the target object.

[0012] The loss function includes the positive and negative sample allocation strategy and loss calculation. The core formula of the positive and negative sample allocation strategy is defined as:

[0013] t=s α ×u β

[0014] Where s and u are the classification score and IoU (intersection over union) value, respectively, and α and β are weight hyperparameters (set to 0.5 and 6.0 by default, respectively). The core formula is used to calculate the degree of alignment for each GT box and the predicted box, and the final predicted box is obtained through the t value.

[0015] The GT box comes from manual annotation, and the four vertex coordinates, center point coordinates and category of the real distribution box are determined for each type of instance object during annotation. The predicted box is a rectangular box inferred from the location of the target object by the model extracting image information layer by layer through a convolutional neural network. The model directly predicts the center point and bounding box of the target without pre-defining the anchor box, and takes the offset of each pixel relative to the target box (in four directions: left, right, top and bottom) as the prediction target.

[0016] The loss calculation includes two parts: classification loss and regression loss, which are defined as:

[0017] L=γ1L cla +γ2L reg

[0018] Among them, γ1 and γ2 are the classification loss L cla and the regression loss L reg The weight coefficient of . Specifically, the classification loss L cla The binary cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the actual label, which is defined as:

[0019]

[0020] Among them, yi represents the label of sample i, the positive sample is 1, and the negative sample is 0; p i represents the probability that sample i is a positive sample, and N represents the number of label categories. Classification loss L cla Used to judge "whether it is of this type" and output the maximum confidence value under each category.

[0021] The regression loss is calculated using CIoU+DFL. CIoU (Complete Intersection over Union) is an improved version of IoU (Intersection over Union) loss, which considers the complete intersection between target boxes and is defined as:

[0022]

[0023] Among them, b and b gt Represents the predicted box B and the true distribution box B respectively GT The center point of , ρ(,) represents the Euclidean distance, c represents the diagonal distance between the closed area of ​​the predicted box and the true distribution box, v is used to measure the consistency of the relative proportions of the two boxes, and δ is the weight coefficient. They are defined as:

[0024]

[0025] Among them, (w gt ,h gt ) and (w, h) represent the predicted box B and the true distribution box B respectively GTThe width and height of the image. DFL (Distribution Focal Loss) is also calculated using cross entropy to obtain the probability of the left and right positions closest to the label y, which is defined as:

[0026] DFL(S i , S i+1 )=-[(y i+1 -y)log(S i )+(yy i )log(S i+1 )]

[0027] Among them, S i and S i+1 are the predicted value and the near-predicted value of the network, y, y i and i+1 They are the true value of the label, the label integral value and the adjacent label integral value respectively. The DFL converts the center point coordinates (x, y) and size (w, h) of the label value into four distances (l, t, r, b) from the predicted point to the left, top, right and bottom of the predicted box, and converts (l, t, r, b) into the integral value required by the DFL.

[0028] The calculation of the loss function controls the gradient and iterates the training, and is characterized in that the loss function calculates the error between the predicted value and the true value, and transmits the error back to each layer of the network through the back propagation algorithm to adjust the weights and biases in the network. Specifically, the gradient of the loss function is used to guide the weight update, and the gradient is calculated layer by layer through the chain rule and back propagated to each layer. The weight is adjusted according to the gradient, and the training process is iterated to gradually reduce the loss function value and improve the prediction accuracy of the model.

[0029] The iterative training includes: data preprocessing, feature extraction and fusion, prediction output, loss function calculation, gradient back propagation, parameter update, repeated iteration and model preservation and evaluation. Specifically, the data preprocessing stage performs preprocessing such as scaling and cropping on the image to generate image data and corresponding annotation data consistent with the network input size. The feature extraction and fusion stage relies on convolutional neural networks and feature pyramid networks to predict and generate feature maps and predict the confidence of the target. By calculating the loss function, gradient back propagation and parameter update, the mid-term model is saved regularly, and then a new round of iteration is repeated.

[0030] The output includes non-maximum suppression, generating a mask and an output image.

[0031] The non-maximum suppression (NMS) is a post-processing technique for detection tasks, which is used to remove duplicate detection boxes. Specifically, all detection boxes are sorted according to the predicted confidence scores, starting from the highest score, and each detection box is checked one by one, and other detection boxes whose IoU (intersection over union) exceeds a set threshold are removed.

[0032] The generated mask and output image are characterized in that each predicted box retained after non-maximum suppression has a corresponding mask map. The output includes a bounding box, a category label, a confidence level and a mask map of each target.

[0033] The confidence is obtained by combining the target probability of the prediction box and the classification probability of the prediction box. Specifically, the network first calculates the probability that the target box contains the target object (i.e., whether there is a target at that location), and then predicts the probability that the target belongs to different categories. The confidence is the product of the above two probabilities, indicating the possibility that the prediction box is both a target box and belongs to a specific category.

[0034] Furthermore, in step S2, precision, recall, F1 score, average precision (mAP) and frames per second (FPS) are used as evaluation indicators, and the deep learning model with the highest mAP and the fastest FPS is selected as the model for the final application.

[0035] S3: The evaluation system shown is characterized by, based on image processing technology, A* algorithm, pixel mapping processing depth model output mask image, calculating the actual length of the disease, and evaluating the tunnel degradation level.

[0036] The image processing technology includes binarization, opening and closing operations, A* algorithm and pixel mapping. Specifically, the mask image output by the deep learning model is binarized into black and white, and the diseased area is filled with white pixels and the background is filled with black pixels. The opening and closing operations are used to remove the redundant branches and burrs on the binary mask image, and solve the problem of discontinuity of the diseased area between the mask image and the original image due to uneven pixel distribution.

[0037] The A* algorithm is a heuristic search algorithm used for graph-based search and path planning problems. It uses a heuristic function to guide the search process, thereby effectively finding the shortest path to the target node. Specifically, the A* algorithm regards black pixel areas as obstacles and white pixels as areas where paths can be planned. It plans the shortest path between the starting and ending points of each disease on the binary mask image, and finally calculates the actual length by pixel mapping.

[0038] The tunnel degradation level assessment is carried out according to the "China State Railway Group Co., Ltd. Enterprise Standard (Railway Bridge and Tunnel Building Degradation Assessment Part 2: Tunnels) Q / CR 405.2-2019". Taking the lining void disease as the object, the actual length calculated in step S3. is used as the input of the assessment system, and the degradation level is assessed according to the interval to which the length belongs: {AA: extremely serious, A: serious, B: relatively serious, C: relatively light, D: light}.

[0039] Beneficial effects of the invention: Compared with the prior art, the present invention provides a railway tunnel lining defect detection and assessment method based on deep learning. It uses ground-penetrating radar to collect railway tunnel lining defect images as input source, accurately identifies and locates various common defects, and uses the A* algorithm to calculate the actual length of the void to evaluate the tunnel degradation level. Compared with the prior art that relies on manual identification and assessment of defects, its work efficiency is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the steps of the present invention;

[0041] Figure 2 It is a structural diagram of the deep learning model in the present invention;

[0042] Figure 3 It is the structural diagram of the A* algorithm of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0044] Example 1: Figure 1 As shown, the tunnel lining disease detection and assessment method based on deep learning in this embodiment includes the following steps:

[0045] S1: A set of image data is collected by ground penetrating radar (GPR) along the arch crown, arch span, arch haunch and wall edge line in the upward and downward directions of the railway tunnel.

[0046] S2: After the GPR images are manually annotated by the LabelImg tool, they are divided into training set, validation set and test set according to the ratio of 7:2:1; the data set is input into the deep learning model to obtain the mask image and confidence after segmentation. The deep learning model includes the image input end, detection layer, training layer and output end;

[0047] The input end is used to input pictures and corresponding label files.

[0048] The detection layer is used for image preprocessing, feature extraction and feature fusion. Specifically, for feature extraction and feature fusion of the input image, feature extraction extracts high-level features layer by layer through a series of convolutional layers, pooling layers and residual connections, while gradually reducing the spatial resolution. Specifically, feature extraction locates the coordinates of the detection target through the center point, predicts the category of the target at each positive sample in the prediction box, and uses the offset distance relative to the upper, lower, left and right boundaries of the bounding box as the regression target, and then retains the center point with higher confidence through non-maximum suppression, filters out redundant points, and outputs the feature map.

[0049] The training layer is used to calculate the loss function, control the gradient and iterate the training. Specifically, the training layer includes a segmentation head and a loss function calculation control gradient for iterative training.

[0050] The segmentation head processes the feature map into a high-resolution output by combining convolutional layers, F-C2f modules and upsampling layers to gradually restore the spatial resolution, and finally generates a segmentation mask of the target object. Specifically, it includes extracting multi-scale features using multiple convolutional layers, and restoring the low-resolution feature map to the original resolution of the input image through layer-by-layer upsampling and fusion, and outputting a high-quality mask containing the segmentation information of the target object.

[0051] The loss function includes a positive and negative sample allocation strategy and a mixed loss function calculation. The positive and negative sample allocation strategy is defined as:

[0052] t=s α ×u β

[0053] It is used to calculate the degree of alignment between each GT box and the predicted box, and obtain the final predicted box through the t value.

[0054] The GT box comes from manual annotation, and the four vertex coordinates, center point coordinates and category of the real distribution box are determined for each type of instance object during annotation. The predicted box is a rectangular box inferred from the location of the target object by the model extracting image information layer by layer through a convolutional neural network. The model directly predicts the center point and bounding box of the target without pre-defining the anchor box, and takes the offset of each pixel relative to the target box (in four directions: left, right, top and bottom) as the prediction target.

[0055] The hybrid loss calculation includes two parts: classification loss and regression loss, which are defined as:

[0056] L=γ1L cla +γ2L reg

[0057] Among them, γ1 and γ2 are the classification loss L cla and the regression loss L regThe weight coefficient of . Specifically, the classification loss L cla The binary cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the actual label, which is defined as:

[0058]

[0059] Among them, y i represents the label of sample i, positive sample is 1, negative sample is 0; p i represents the probability that sample i is a positive sample, and N represents the number of label categories. Classification loss L cla Used to judge "whether it is of this type" and output the maximum confidence value under each category.

[0060] The regression loss is calculated using CIoU+DFL. CIoU (Complete Intersection over Union) is an improved version of IoU (Intersection over Union) loss, which considers the complete intersection between target boxes and is defined as:

[0061]

[0062] Among them, b and bg t Represents the predicted box B and the true distribution box B respectively GT The center point of , ρ(,) represents the Euclidean distance, c represents the diagonal distance between the closed area of ​​the predicted box and the true distribution box, v is used to measure the consistency of the relative proportions of the two boxes, and δ is the weight coefficient. They are defined as:

[0063]

[0064] Among them, (w gt ,h gt ) and (w, h) represent the predicted box B and the true distribution box B respectively GT The width and height of the image. DFL (Distribution Focal Loss) is also calculated using cross entropy to obtain the probability of the left and right positions closest to the label y, which is defined as:

[0065] DFL(S i , S i+1 )=-[(y i+1 -y)log(S i )+(yy i )log(S i+1 )]

[0066] Among them, S i and S i+1 are the predicted value and the near-predicted value of the network, y, y i and i+1They are the true value of the label, the label integral value and the adjacent label integral value respectively. The DFL converts the center point coordinates (x, y) and size (w, h) of the label value into four distances (l, t, r, b) from the predicted point to the left, top, right and bottom of the predicted box, and converts (l, t, r, b) into the integral value required by the DFL.

[0067] The calculation of the loss function controls the gradient and iterates the training, and is characterized in that the loss function calculates the error between the predicted value and the true value, and transmits the error back to each layer of the network through the back propagation algorithm to adjust the weights and biases in the network. Specifically, the gradient of the loss function is used to guide the weight update, and the gradient is calculated layer by layer through the chain rule and back propagated to each layer. The weight is adjusted according to the gradient, and the training process is iterated to gradually reduce the loss function value and improve the prediction accuracy of the model.

[0068] The iterative training includes: data preprocessing, feature extraction and fusion, prediction output, loss function calculation, gradient back propagation, parameter update, repeated iteration and model preservation and evaluation. Specifically, the data preprocessing stage performs preprocessing such as scaling and cropping on the image to generate image data and corresponding annotation data consistent with the network input size. The feature extraction and fusion stage relies on convolutional neural networks and feature pyramid networks to predict and generate feature maps and predict the confidence of the target. By calculating the loss function, gradient back propagation and parameter update, the mid-term model is saved regularly, and then a new round of iteration is repeated.

[0069] The output includes non-maximum suppression, generating a mask and an output image.

[0070] S3: The evaluation system shown is characterized in that, based on image processing technology, A* algorithm and pixel mapping process the output mask image of the deep model, the actual length of the disease is calculated, and the tunnel degradation level is evaluated.

[0071] The image processing technology includes binarization, opening and closing operations, A* algorithm and pixel mapping. Specifically, the mask image output by the deep learning model is binarized into black and white, and the diseased area is filled with white pixels and the background is filled with black pixels. The opening and closing operations are used to remove the redundant branches and burrs on the binary mask image, and solve the problem of discontinuity of the diseased area between the mask image and the original image due to uneven pixel distribution.

[0072] The A* algorithm is a heuristic search algorithm used to effectively find the shortest path to the target node. Specifically, as shown in the figure, the A* algorithm regards the black pixel area as an obstacle and the white pixel as the plannable path area. It plans the shortest path between the starting and ending points of each disease on the binary mask image, and finally calculates the actual length by pixel mapping.

[0073] The tunnel degradation level assessment is carried out according to the "China State Railway Group Co., Ltd. Enterprise Standard (Railway Bridge and Tunnel Building Degradation Assessment Part 2: Tunnels) Q / CR 405.2-2019". Taking the lining void disease as the object, the actual length calculated in step S3. is used as the input of the assessment system, and the degradation level is assessed according to the interval to which the length belongs: {AA: extremely serious, A: serious, B: relatively serious, C: relatively light, D: light}.

[0074] The intelligent detection method for tunnel lining defects based on deep learning in this embodiment is characterized in that tunnel lining GPR images are used as data sources, tunnel lining defects are intelligently identified, located and classified through a deep learning model, and confidence levels and defect contours are output in the form of masks. Compared with the existing method of relying on manual identification and evaluation of defects, the work efficiency is greatly improved.

[0075] This embodiment is only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation, characterized by: S1: collect GPR image data; S2: Manually label various defects in GPR image data and create the data set required for training; S3: Input the dataset into the deep learning model to obtain the segmented mask image; S4: Using the obtained mask image, calculate the minimum length of each defect and complete the tunnel degradation level assessment.

2. According to claim 1, a railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation is characterized in that: Specifically, S1 includes: the GPR is a ground penetrating radar, and the GPR image data is collected by using the GPR along the arch crown, arch shoulder, arch waist and wall edge of the railway tunnel.

3. The railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation according to claim 1 is characterized in that: Specifically, S2 includes: inputting the GPR image data into the target detection tool platform, manually annotating the real distribution of various defects, and generating a label file. The data set required for training is to randomly divide the manually annotated GPR image data and the label file in proportion to formulate a data set, and ensure that the GPR image data is not repeated between the data sets.

4. The railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation according to claim 1 is characterized in that: The S3 is specifically: S3.1: Input the annotated GPR image data and label file at the input end of the deep learning model, then input the training set into the detection layer of the deep learning model, complete the feature map extraction, and then output the fused feature map; S3.2: Use the fused feature map to calculate the loss function and control the gradient, iterate the training, and complete the model training; S3.3: Use the trained model to output the segmented mask image.

5. The railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation according to claim 4 is characterized in that: The S3.2 is specifically: The loss function includes the positive and negative sample allocation strategy and loss function calculation. The positive and negative sample allocation strategy selects positive samples according to the weighted scores of classification and regression, and the number of positive samples is supplemented as training progresses. The core formula is described as follows: t=s α ×u β Where s and u are the classification score and intersection-over-union ratio respectively, α and β are weight hyperparameters, and the core formula is used to calculate the alignment degree of each true distribution box and the predicted box, and the final predicted box is obtained through the t value; The true distribution frame comes from the manually annotated fused feature map; the predicted frame is obtained through the model from the fused feature map; The loss function is used to measure the degree of overlap between the true distribution box and the predicted box. The value of the loss function is minimized through iterative training. The loss function is positively correlated with the training gradient. The loss function uses the weighted sum of classification loss and regression loss to control the training gradient, which is defined as: L=γ1L cla +γ2L reg Among them, γ1 and γ2 are the classification loss L cla and the regression loss L reg The weight coefficient of Classification loss L cla The binary cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the actual label, which is defined as: Among them, y i represents the label of sample i, positive sample is 1, negative sample is 0; p i represents the probability that sample i is a positive sample, and N represents the number of label categories; The regression loss is obtained by using the complete intersection-union ratio + distribution focus loss. The regression loss is used to measure the position error between the predicted box and the true distribution box in the target detection task, and is defined as: 1-IoU represents the overlap between the predicted box and the true distribution box. represents the Euclidean distance between the center points of the predicted box and the true distribution box, δv is the shape loss term, which is used to measure the difference in aspect ratio between the predicted box and the true distribution box. gt Represents the predicted box B and the true distribution box B respectively GT , ρ(·) represents the Euclidean distance, c represents the diagonal distance of the closed area of ​​the two rectangular boxes, v is used to measure the consistency of the relative proportions of the predicted box and the true distribution box, and δ is the weight coefficient, which are defined as: Among them, (w gt ,h gt ) and (w, h) represent the predicted box B and the true distribution box B respectively GT Width and height; In order to match the complete intersection-union ratio to measure the position error between the predicted box and the true distribution box, the distribution focus loss is calculated in the form of cross entropy to obtain the probability of the two closest left and right positions of label y, which is defined as: DFL(S i ,S i+1 )=-[(y i+1 -y)log(S i )+(y-y i )log(S i+1 )] Among them, S i and S i+1 are the predicted value and the near predicted value of the network, respectively, representing the predicted score of the model for a certain position of the target box, y, y i and i+1 They are the true value of the label, the integral value of the label and the integral value of the adjacent label respectively. The center point coordinates (x, y) and size (w, h) of the label value are converted into four distances (l, t, r, b) from the predicted point to the left, top, right and bottom of the predicted box through the distribution focus loss, and (l, t, r, b) is converted into the integral value required by the distribution focus loss; The iterative training includes data preprocessing, feature extraction and fusion, prediction output, loss function calculation, gradient back propagation, parameter update, repeated iteration and model preservation and evaluation, and finally completing the model training.

6. The railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation according to claim 4 is characterized in that: The S3.3 is specifically: Through the trained model, the feature map is first used to generate a binary mask map of each target through convolution and upsampling operations, and then the confidence is obtained by combining the target probability of the prediction box and the classification probability of the prediction box. The disease target mask is output through confidence filtering and mask post-processing steps, and finally the mask and prediction box with low confidence are filtered through the non-maximum suppression algorithm to output the final segmentation mask image.

7. The railway tunnel lining defect detection and assessment method based on anchor-free instance segmentation according to claim 1 is characterized in that: The S4 is specifically as follows: firstly, the obtained segmentation mask image is binarized to remove burrs and redundant branch processing operations, and then the shortest path between the specified starting and ending points is planned at the pixel level for the processed mask image through an optimization algorithm, and then the GPR image length scale is converted into the actual length by pixel mapping, and finally the tunnel degradation level is assessed according to the length interval.

Images