Semantic segmentation positioning method for apparent diseases of urban subway tunnel

Abstract

The invention provides an urban subway tunnel apparent disease semantic segmentation positioning method. The method comprises the following steps: collecting tunnel interval images; cutting the tunnel interval image, and constructing an apparent disease data set; a deep learning model is built, the deep learning model comprises an input layer, a coding layer, a decoding layer and an output layer, the input layer receives image data in the apparent disease data set, and the coding layer comprises an apparent disease feature encoder for extracting features of apparent diseases and a lining feature encoder for extracting features of lining positions; the decoding layer comprises a disease semantic segmentation decoder for generating a semantic segmentation result of an apparent disease and a lining position feature decoder for generating a semantic segmentation result of a lining position, and the output layer outputs the semantic segmentation result of the apparent disease and the semantic segmentation result of the lining position; and training and testing the deep learning model by using the apparent disease data set. The method can effectively improve the accuracy of semantic segmentation and positioning of the apparent diseases of the subway tunnel, and reduces the situation of missing detection or false detection.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel structure disease detection, and particularly to a semantic segmentation and localization method for the apparent diseases of urban subway tunnels. Background Art

[0002] Due to natural factors such as water penetration and geological conditions, aging, corrosion and wear of the internal structure materials of the tunnel, improper operation and maintenance of the tunnel, underground construction activities, temperature and humidity changes, traffic vibration and chemical pollution, etc., the apparent diseases of the tunnel may occur. Apparent diseases may lead to the damage, corrosion or weakening of the tunnel structure, thus endangering the overall structural safety of the tunnel. If not repaired in time, these diseases may cause the tunnel to collapse or structural damage, threatening the safety of tunnel users and the surrounding environment; it may also damage the infrastructure such as equipment, tracks, cables, etc. in the tunnel, resulting in equipment failures, train operation interruptions or delays, and affecting the normal operation of the subway system. Therefore, it is crucial to identify and solve the problem of apparent diseases at an early stage to maintain the reliability and safety of the tunnel system.

[0003] At present, the investigation of tunnel apparent diseases mainly adopts the method of manual inspection, taking photos and on-site records, focusing on checking cracks, corrosion, leakage, spots, damage, etc., and organizing the recorded data into an inspection report, including photos and written descriptions of the location, degree and type of the diseases. The method of manual inspection has a large workload, requires a lot of time and labor costs, low efficiency, subjectivity, and is greatly interfered by human factors. Therefore, many advanced technologies such as automated inspection, image processing and deep learning have begun to be explored and used.

[0004] The semantic segmentation and localization method is one of the technologies used to extract the information of the apparent diseases of urban subway tunnels. By segmenting the tunnel images, detailed information about the diseases can be obtained, thus providing important information support for tunnel maintenance and repair. Currently, deep learning models are usually used to identify the diseases in the images and distinguish them from the normal tunnel parts.

[0005] Chinese invention CN109767426A proposes a shield tunnel leakage detection method based on image feature recognition. Before leakage detection, detailed preprocessing of the image is carried out, that is, the tunnel feature recognition rules are first determined through the statistical analysis of the gray-scale image of the tunnel appearance, and then the tunnel features are identified one by one. After all the features are removed, the leakage detection is carried out.

[0006] Chinese invention CN106841216A discloses an automatic tunnel disease recognition device based on panoramic image CNN. It mainly uses a panoramic vision sensor to obtain panoramic images of the tunnel inner wall, extracts suspected disease areas through panoramic image unfolding, image preprocessing, binarization processing, etc., and finally uses a convolutional neural network to automatically detect, classify, and recognize diseases.

[0007] Chinese invention CN109615653A discloses a method for detecting and recognizing the leakage water area based on deep learning and a field of view projection model. It mainly collects video data and point cloud data of the tunnel, and uses deep learning methods to detect the leakage water area pictures from the video data.

[0008] The above patents all use deep learning methods to study tunnel disease recognition. Usually, general deep learning models are used or fine-tuned on this basis, such as the YOLO series models for object detection and Unet for image segmentation. The structural designs of these general models do not consider the prior laws of the apparent diseases of subway tunnels. For example, due to factors such as structural deformation, the structural leakage water disease is likely to occur at the splicing position of the lining structure, and the structural defect is mainly caused by the mutual extrusion of adjacent segments, resulting in the crushing of the edge concrete. Therefore, the general model method is prone to missed detection or false detection. Summary of the Invention

[0009] The purpose of the present invention is to provide a semantic segmentation and localization method for the apparent diseases of urban subway tunnels to solve the above problems, which can effectively improve the accuracy of the semantic segmentation and localization of the apparent diseases of urban subway tunnels and greatly reduce the situation of missed detection or false detection.

[0010] The present invention proposes a semantic segmentation and localization method for the apparent diseases of urban subway tunnels, including the following steps:

[0011] Step S1, collect tunnel section images;

[0012] Step S2, crop the tunnel section images to construct an apparent disease data set, and the apparent disease data set includes various apparent disease labels;

[0013] Step S3: Build a deep learning model, including an input layer, an encoding layer, a decoding layer, and an output layer. The input layer receives the image data in the apparent disease dataset. The encoding layer is connected to the input layer and includes an apparent disease feature encoder and a lining feature encoder. The apparent disease feature encoder extracts the features of apparent diseases, and the lining feature encoder extracts the features of the lining position. The decoding layer includes a disease semantic segmentation decoder and a lining position feature decoder. The disease semantic segmentation decoder generates the semantic segmentation result of apparent diseases, and the lining position feature decoder generates the semantic segmentation result of the lining position. The output layer outputs the semantic segmentation results of apparent diseases and the lining position respectively;

[0014] Step S4: Use the apparent disease dataset to train and test the deep learning model.

[0015] In one embodiment, the tunnel section images are acquired by a device equipped with a three-dimensional laser scanner and an area array camera, and the tunnel section images are single-channel grayscale images.

[0016] In one embodiment, in step S2, cropping the tunnel section images specifically includes using a sliding window with a size of 512×512 pixels to crop the tunnel section images.

[0017] In one embodiment, the apparent disease labels include wet marks, water leakage, dripping, line leakage, sediment leakage, and defects, and the proportion of the image data of various apparent disease labels is balanced.

[0018] In one embodiment, the apparent disease dataset includes a training set, a validation set, and a test set. The training set is used for training the deep learning model, the validation set is used for validating the deep learning model, and the test set is used for testing the deep learning model. Among them, the training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%.

[0019] In one embodiment, the basic architecture of the deep learning model is a convolutional neural network.

[0020] In one embodiment, there are two input layers, including a first input layer and a second input layer. The first input layer is connected to the apparent disease feature encoder, and the second input layer is connected to the lining feature encoder.

[0021] In one embodiment,

[0022] The apparent disease feature encoder uses the residual convolution module of the residual network as the basic architecture and includes three residual convolution modules;

[0023] The lining feature encoder uses the depthwise separable convolution module of MobileNet as the basic architecture and includes three depthwise separable convolution modules.

[0024] In one embodiment,

[0025] The disease semantic segmentation decoder receives the outputs of the apparent disease feature encoder and the lining feature encoder respectively. The disease semantic segmentation decoder performs pyramid feature decoding and includes three transposed convolution modules and one upsampling module;

[0026] The lining position feature decoder receives the output of the lining feature encoder and includes one upsampling module.

[0027] In one embodiment, the output layer includes a first output layer and a second output layer;

[0028] The first output layer is a softmax layer, which is connected to the disease semantic segmentation decoder to output the semantic segmentation result of the apparent disease and judge whether different regions in the image are affected by the apparent disease;

[0029] The second output layer is a sigmoid layer, which is connected to the lining position feature decoder to output the semantic segmentation result of the lining position and judge whether different positions in the image belong to the lining position.

[0030] In one embodiment, in step S4, when training the deep learning model using the apparent disease dataset, the parameters of the deep learning model are adjusted by using the learning rate adaptive stochastic gradient algorithm, and the loss function L total is used to characterize the performance of the deep learning model;

[0031] The calculation formula of the loss function is:

[0032] where represents the weighted cross-entropy loss of the semantic segmentation result of the apparent disease, represents the weighted cross-entropy loss of the semantic segmentation result of the lining position, α represents the weighting coefficient, and the value range of α is 0.3 - 0.5.

[0033] In one embodiment, in step S4, when testing the deep learning model using the apparent disease dataset, it includes setting accuracy metrics to evaluate the deep learning model, and the accuracy metrics include the IOU metric.

[0034] In one embodiment, the method for semantic segmentation and localization of apparent diseases in urban subway tunnels further includes step S5: selecting the deep learning model with the best IOU metric as the final application model.

[0035] Compared with the prior art, the beneficial effects of the method for semantic segmentation and localization of apparent diseases in urban subway tunnels of the present invention are as follows:

[0036] 1) The method for semantic segmentation of apparent diseases in tunnel structures guided by the lining splicing position of the present invention introduces prior knowledge of the influence of the splicing position on diseases in model design, guiding the model to focus on the splicing position of the lining structure as the key area of concern, helping the model to more accurately extract disease texture features, thereby effectively improving the accuracy of semantic segmentation and localization of apparent diseases in urban subway tunnels and greatly reducing the situation of missed detection or misdetection.

[0037] 2) The present invention can automatically detect apparent diseases, improve efficiency, reduce workload, reduce labor and time costs, and will not be affected by human subjective factors, ensuring data consistency and accuracy.

[0038] 3) Timely and accurate disease detection and maintenance contribute to improving the safety and reliability of the urban subway system, reducing potential risks, and ensuring the safety of passengers.

[0039] 4) Through the method in the present invention, historical data of apparent diseases in tunnels can be easily recorded and tracked, which helps to analyze the evolution trend of diseases and formulate maintenance plans in advance. Description of the Drawings

[0040] Figure 1 It is a schematic structural diagram of a deep learning model according to an embodiment of the present invention;

[0041] Figure 2 It is a convolutional structure diagram of an apparent disease feature encoder in a deep learning model according to an embodiment of the present invention;

[0042] Figure 3 It is a convolutional operation structure diagram of a lining feature encoder in a deep learning model according to an embodiment of the present invention;

[0043] Figure 4 It is a structural diagram of a disease semantic segmentation decoder in a deep learning model according to an embodiment of the present invention.

[0044] Reference Signs

[0045] 11. First input layer, 12. Second input layer, 21. Apparent disease feature encoder, 22. Lining feature encoder, 31. Disease semantic segmentation decoder, 32. Lining position feature decoder, 41. First output layer, 42. Second output layer. Detailed Embodiments

[0046] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted here that the specific embodiments described herein are used to help understand the present invention, but do not limit the present invention.

[0047] The present invention provides a method for semantic segmentation and localization of apparent diseases in urban subway tunnels, which includes the following steps:

[0048] Step S1: Collect images of tunnel sections;

[0049] Step S2: Crop the images of tunnel sections to construct an apparent disease dataset, which includes various apparent disease labels for describing different types of apparent disease problems in tunnels;

[0050] Step S3: Build a deep learning model, as Figure 1 shown, including an input layer, an encoding layer, a decoding layer and an output layer. The input layer receives the image data in the apparent disease dataset. The encoding layer is connected to the input layer and includes an apparent disease feature encoder 21 and a lining feature encoder 22. The apparent disease feature encoder 21 extracts the features of apparent diseases, and the lining feature encoder 22 extracts the features of the lining position. The decoding layer includes a disease semantic segmentation decoder 31 and a lining position feature decoder 32. The disease semantic segmentation decoder 31 generates the semantic segmentation result of apparent diseases, and the lining position feature decoder 32 generates the semantic segmentation result of the lining position. The output layer outputs the semantic segmentation results of apparent diseases and the lining position respectively;

[0051] Step S4: Use the apparent disease dataset to train and test the deep learning model.

[0052] In short, after the two-way input of image data, one encoder learns the features of apparent diseases, and the other encoder learns the features of the lining position; for the two-way decoder output, one decoder fuses the features of the previous two encoders and outputs the semantic segmentation result Output1 of apparent diseases. The input of the other decoder is the feature output by the lining feature encoder, and the output is the semantic segmentation result Output2 of the lining position in the image. This is because the vast majority of apparent diseases in urban subway tunnels occur at the positions of lining joints. Using the convolutional neural network architecture with two encoders and two decoders can enable it to better capture the disease features at the lining joint positions.

[0053] The method for semantic segmentation and localization of apparent diseases in urban subway tunnels described above will be elaborated in detail below.

[0054] The tunnel section images of an embodiment of the present invention are acquired by a device equipped with a 3D laser scanner and an area array camera, and are used as data sources to construct a deep learning model. The tunnel section images are single-channel grayscale images.

[0055] In step S2 of an embodiment of the present invention, the tunnel section images are cropped, specifically including cropping the tunnel section images using a sliding window with a size of 512×512 pixels.

[0056] The apparent disease labels of an embodiment of the present invention include wet marks, seepage water, dripping, line leakage, sediment seepage, and defects, and the image data ratios of various apparent disease labels are balanced. In order to ensure that the data ratios of various label data in the training process dataset are generally balanced, for disease categories with too little data ratio such as defects, an offline sample enhancement method is used to generate additional data to expand the dataset.

[0057] The apparent disease dataset of an embodiment of the present invention can be divided into three parts: a training set, a validation set, and a test set. The training set is used for training the deep learning model, the validation set is used for validating the deep learning model, and the test set is used for testing the deep learning model. Among them, the training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%. Such a division helps to use different parts of the dataset in the training and evaluation testing of the deep learning model.

[0058] The basic architecture of the deep learning model of an embodiment of the present invention is a convolutional neural network.

[0059] The input layer Input of an embodiment of the present invention has two paths, and the two paths of input are both the original images. The input layer includes a first input layer 11 and a second input layer 12. The first input layer 11 is connected to the apparent disease feature encoder 21, and the second input layer 12 is connected to the lining feature encoder 22.

[0060] The apparent disease feature encoder 21 of an embodiment of the present invention uses the residual convolution module (ResNetBlock) of the residual network (Resnet) as the basic architecture, as Figure 2 shown, including three residual convolution modules Res Block1, ResBlock2, and Res Block3, which are used to learn and extract apparent disease feature representations at different levels and abstraction degrees, and output three paths of feature vectors after encoding. It should be noted that the residual convolution module of ResNet is a commonly used deep learning structure, which aims to help the deep learning model train and learn features more effectively.

[0061] The lining feature encoder 22 of an embodiment of the present invention uses the depthwise separable convolution module (DS Block) of the lightweight MobileNet as the basic architecture, as Figure 3As shown, it includes three depthwise separable convolution modules DS block1, DS block2, and DS block3, which are used to learn and extract lining position features at different levels and degrees of abstraction. It should be noted that MobileNet is a lightweight convolutional neural network structure designed to achieve efficient computing in environments with limited computing resources.

[0062] In one embodiment of the present invention, the disease semantic segmentation decoder 31 receives the outputs of the apparent disease feature encoder 21 and the lining feature encoder 22 respectively. The disease semantic segmentation decoder 31 performs pyramid feature decoding, as Figure 4 shown, it includes three transposed convolution modules AC Block1, AC Block2, and AC Block3, as well as a three-layer AC Block stacking structure and an upsampling module (UpSample). The input of AC Block1 is the feature map obtained by concatenating the output feature maps of the two modules Res Block3 and DS block3 in the encoder. The input of AC Block2 is the output of AC Block1 and the output feature maps of the two modules Res Block2 and DSblock2 in the encoder, with a total of three outputs concatenated. The input of AC Block3 is the output of AC Block2 and the output feature maps of the two modules Res Block1 and DS Block1 in the encoder, with a total of three outputs concatenated. The transposed convolution modules are used to expand the spatial dimension of the feature map, thereby restoring it to the resolution of the original input. The upsampling module is used to increase the size of the feature map output by ACBlock3, thereby achieving an increase in resolution.

[0063] In one embodiment of the present invention, the lining position feature decoder 32 receives the output of the lining feature encoder 22 and includes an upsampling module (UpSample).

[0064] In one embodiment of the present invention, the output layer includes a first output layer 41 and a second output layer 42.

[0065] The first output layer 41 is a softmax layer, which is connected to the disease semantic segmentation decoder 31 and outputs the semantic segmentation result of the apparent disease, determining whether different regions in the image are affected by the apparent disease. It should be noted that softmax (soft maximum) is an activation function, usually used in multi-class classification problems, which converts the original output of the model into a probability distribution for each class, so as to determine which class the input data is most likely to belong to.

[0066] The second output layer 41 is a sigmoid layer, which is connected to the lining position feature decoder 32 and outputs the semantic segmentation result of the lining position to determine whether different positions in the image belong to the lining position. It should be noted that sigmoid (hyperbolic tangent) is an activation function, usually used for binary classification problems, mapping the input data to values between 0 and 1, representing the probability of each category.

[0067] In step S4 of an embodiment of the present invention, when training the deep learning model using the preprocessed apparent disease dataset, the loss function L is used. total Characterize the performance of the deep learning model, and adopt the stochastic gradient algorithm with adaptive learning rate to adjust the parameters of the deep learning model for optimization. It should be noted that the stochastic gradient algorithm is a method for adjusting the parameters of a neural network. Adaptive learning rate means that the learning rate will be automatically adjusted according to the situation during the training process to help the model converge to the optimal solution faster. Cross-entropy loss is usually used for classification problems to measure the difference between the output of the model and the actual target. Weighted cross-entropy loss is to weight the cross-entropy loss, usually used to handle the situation of class imbalance. Since the label data of various apparent diseases are not completely balanced, the weighted cross-entropy loss is used to calculate the output result of each path. The deep learning model finally uses the weighted sum of the weighted cross-entropy losses of the two paths of output as the objective optimization.

[0068] Among them, the loss function L total The calculation formula of is:

[0069] Among them, Represents the weighted cross-entropy loss of the semantic segmentation result of the apparent disease, Represents the weighted cross-entropy loss of the semantic segmentation result of the lining position, α represents the weighting coefficient, and the value range of α is 0.3 - 0.5.

[0070] In step S4 of an embodiment of the present invention, when testing the deep learning model using the apparent disease dataset, it includes setting accuracy metrics to evaluate the deep learning model. The accuracy metrics include the IOU metric. The IOU metric is an evaluation metric commonly used in image segmentation tasks to measure the overlap degree between the segmentation result generated by the model and the true segmentation. A high IOU indicates that the prediction of the model matches the true segmentation better, that is, the model performs best on the test data and is considered a model more suitable for actual image segmentation applications.

[0071] An embodiment of the method for semantic segmentation and localization of apparent diseases in urban subway tunnels of the present invention further includes step S5: Select the deep learning model with the best IOU metric as the final application model, that is, select the model with the highest IOU score on the test data as the model used in actual applications.

[0072] It should be noted that in this application, unless otherwise clearly specified and defined, similar terms such as "connection" should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can also be the communication inside two components. Those skilled in the art can understand the specific meanings of the above terms in this application according to specific situations. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0073] The present invention has the following beneficial effects:

[0074] 1) The method for semantic segmentation of apparent diseases of tunnel structures based on the guiding of the lining splicing position in the present invention introduces prior knowledge of the influence of the splicing position on diseases from the model design, guiding the model to regard the splicing position of the lining structure as the key attention area, helping the model to more accurately extract the texture features of diseases, thereby effectively improving the accuracy of semantic segmentation and localization of apparent diseases in urban subway tunnels and greatly reducing the situation of missed detection or misdetection.

[0075] 2) The present invention can automatically detect apparent diseases, improving the efficiency, reducing the workload, reducing the human and time costs, and not being affected by subjective human factors, and can ensure the consistency and accuracy of data.

[0076] 3) Timely and accurate disease detection and maintenance contribute to improving the safety and reliability of the urban subway system, reducing potential risks, and ensuring the safety of passengers.

[0077] 4) Through the method in the present invention, the historical data of apparent diseases in tunnels can be easily recorded and tracked, which helps to analyze the evolution trend of diseases and formulate maintenance plans in advance.

[0078] The above-described embodiments are only further descriptions of the present invention and do not impose other forms of limitations on the present invention. The present invention can also have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding modifications and changes according to the present invention, but these corresponding modifications and changes should all fall within the protection scope of the present invention.

Claims

1. A semantic segmentation and localization method for apparent defects in urban subway tunnels, characterized in that: The steps include: Step S1, collecting tunnel section images; Step S2, cropping the tunnel section image to construct an apparent disease data set, wherein the apparent disease data set includes a plurality of apparent disease labels; Step S3, building a deep learning model, including an input layer, an encoding layer, a decoding layer and an output layer, wherein the input layer receives the image data in the apparent disease data set, the encoding layer is connected to the input layer, and includes an apparent disease feature encoder and a lining feature encoder, the apparent disease feature encoder extracts the features of the apparent disease, the lining feature encoder extracts the features of the lining position, the decoding layer includes a disease semantic segmentation decoder and a lining position feature decoder, the disease semantic segmentation decoder generates a semantic segmentation result of the apparent disease, the lining position feature decoder generates a semantic segmentation result of the lining position, and the output layer outputs the semantic segmentation results of the apparent disease and the semantic segmentation results of the lining position respectively; Step S4: Use the apparent disease dataset to train and test the deep learning model.

2. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The tunnel section image is acquired by collecting data using a device equipped with a three-dimensional laser scanner and an area array camera, and the tunnel section image is a single-channel grayscale image.

3. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The step S2 of cropping the tunnel section image specifically includes cropping the tunnel section image using a sliding frame of 512×512 pixels; The apparent disease labels include wet marks, water leakage, dripping, line leakage, mud and sand seepage, and defects, and the image data of various apparent disease labels account for a balanced proportion; The apparent disease data set includes a training set, a validation set and a test set. The training set is used for training the deep learning model, the validation set is used for validating the deep learning model, and the test set is used for testing the deep learning model. The training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%.

4. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The basic architecture of the deep learning model is a convolutional neural network; The input layer has two paths, including a first input layer and a second input layer. The first input layer is connected to the apparent disease feature encoder, and the second input layer is connected to the lining feature encoder.

5. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The apparent disease feature encoder adopts the residual convolution module of the residual network as the basic architecture, including three residual convolution modules; The lining feature encoder adopts the depthwise separable convolutional module of MobileNet as the basic architecture, including three depthwise separable convolutional modules.

6. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The disease semantic segmentation decoder receives the outputs of the apparent disease feature encoder and the lining feature encoder respectively, and the disease semantic segmentation decoder is a pyramid feature decoding, including three deconvolution modules and one upsampling module; The lining position feature decoder receives the output of the lining feature encoder and includes an upsampling module.

7. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: The output layer includes a first output layer and a second output layer; The first output layer is a softmax layer, connected to the disease semantic segmentation decoder, outputs the semantic segmentation result of the apparent disease, and determines whether different areas in the image are affected by the apparent disease; The second output layer is a sigmoid layer, which is connected to the lining position feature decoder, outputs the semantic segmentation result of the lining position, and determines whether different positions in the image belong to the lining position.

8. The method for semantic segmentation and localization of apparent defects in urban subway tunnels according to claim 1 is characterized in that: In step S4, when the deep learning model is trained using the apparent disease dataset, the parameters of the deep learning model are adjusted using a stochastic gradient algorithm with adaptive learning rate, and a loss function L is used. total Characterizing the performance of the deep learning model; The calculation formula of the loss function is: in, represents the weighted cross entropy loss of the semantic segmentation result of the apparent disease, represents the weighted cross entropy loss of the semantic segmentation result of the lining position, α represents the weighting coefficient, and the value range of α is 0.3-0.

5.

9. The semantic segmentation and positioning method for apparent defects in urban subway tunnels according to claim 1 is characterized in that: In step S4, using the apparent disease dataset to test the deep learning model includes setting an accuracy index to evaluate the deep learning model, and the accuracy index includes an IOU index.

10. The method for semantic segmentation and localization of apparent defects in urban subway tunnels according to claim 9, characterized in that: The method also includes step S5, selecting the deep learning model with the best IOU indicator as the final application model.

Images