Tunnel crack defect assessment method and system

Abstract

The present invention relates to a tunnel crack defect assessment method, system, medium and device, and relates to the field of image recognition technology. The method comprises obtaining crack image data of a tunnel to be assessed, labeling and preprocessing the data, and obtaining labeled crack image data; using a U-Net architecture as a basis, combining a feature pyramid network, introducing an attention module, constructing a multi-scale crack defect assessment model, obtaining crack image data of an assessed tunnel as a data set, training and optimizing the parameters of the multi-scale crack defect assessment model, and obtaining an optimized multi-scale crack defect assessment model; obtaining preprocessed crack image data of the tunnel to be assessed, inputting the data into the optimized multi-scale crack defect assessment model, and obtaining a crack defect assessment result.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a tunnel crack defect assessment method and system. Background Art

[0002] As a vital component of modern transportation infrastructure, tunnels are crucial for their structural safety. Over the course of long-term operation, tunnels are subject to a variety of factors, including geological conditions, environmental factors, and vehicle loads, which can lead to structural defects such as cracks. If these cracks are not detected and repaired promptly, they can lead to further damage to the tunnel structure and even cause accidents.

[0003] With the development of machine vision and image processing technology, research on intelligent detection of tunnel cracks has continued to expand. Current research on automatic detection of tunnel defects mainly uses visual inspection technology to collect and intelligently process tunnel defect images.

[0004] Currently, tunnel crack detection relies primarily on the following technologies: Visual inspection: Professionals conduct visual inspections inside the tunnel. This method is simple and easy, but inefficient, difficult to detect small cracks, and poses safety risks. Ultrasonic testing: Utilizes the reflection principle of ultrasound to detect cracks within materials. However, it requires a coupling agent and high operator skill, and its effectiveness is limited for complex cracks and non-uniform materials. Infrared thermal imaging: Detects cracks by measuring temperature differences on the tunnel surface, but is significantly affected by ambient temperature and cannot detect deep cracks. Laser scanning: Obtains three-dimensional data of the tunnel surface for crack detection, but the equipment is expensive, data processing is complex, and detection is ineffective for cracks covered by dirt. Acoustic emission: Detects the acoustic signals generated by crack propagation within the material, but can only detect active cracks and is not suitable for detecting static cracks. Electromagnetic detection: Detects cracks by changes in the electromagnetic field, but is not suitable for non-conductive materials and has limited detection depth.

[0005] The limitations of the above technologies are mainly reflected in the following aspects: Low detection efficiency: Traditional methods require a lot of manpower and time for detection, especially in long-distance tunnels. Insufficient detection accuracy: The detection ability of small cracks or deep cracks is insufficient, which can easily lead to missed detection. Complex data processing: Although some technologies can collect a large amount of data, they lack effective data processing methods and cannot quickly and accurately analyze crack information. Poor environmental adaptability: Many detection technologies are greatly affected by environmental factors such as temperature, humidity, light, etc., which affects the accuracy of detection. Therefore, it is necessary to develop a new tunnel crack defect assessment method to improve the efficiency and accuracy of detection, reduce the impact of environmental factors, and realize automated and intelligent crack detection. Summary of the Invention

[0006] In order to overcome the above-mentioned shortcomings of the prior art, the main purpose of the present invention is to provide a tunnel crack defect assessment method and system.

[0007] To achieve the above object, the present invention adopts the following technical solution, a tunnel crack defect assessment method, comprising the following steps:

[0008] Using the U-Net architecture as a foundation, a feature pyramid network was added to obtain a hybrid network structure. CBAM was inserted after the convolutional block of the hybrid network structure, SENet was added after the bottleneck layer between the encoder and decoder, and a non-local module was placed after the deepest convolutional block of the encoder to construct a multi-scale crack defect assessment model.

[0009] Acquire the evaluated annotated crack image data as a data set, train and optimize the parameters of the multi-scale crack defect assessment model, and obtain an optimized multi-scale crack defect assessment model;

[0010] After preprocessing, the acquired tunnel crack image data to be evaluated is input into the optimized multi-scale crack defect evaluation model to obtain a crack defect evaluation result.

[0011] The method of obtaining the hybrid network structure comprises the following steps:

[0012] The U-Net architecture includes a downsampling path and an upsampling path;

[0013] The FPN is implemented through a top-down path and a bottom-up path;

[0014] After each downsampling stage of the U-Net, the top-down path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the previous layer; before each upsampling stage of the U-Net, the bottom-up path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the next layer to obtain a hybrid network structure.

[0015] The method of obtaining the annotated crack image data comprises the following steps:

[0016] Define the tunnel crack criteria, select annotation tools to annotate crack image data of the tunnel to be evaluated, identify the crack locations, draw bounding boxes for each crack, and perform pixel-level annotation using semantic segmentation to obtain preliminarily annotated crack image data;

[0017] The initially annotated crack image data is subjected to image enhancement, size adjustment, normalization and cleaning, and then data segmentation is performed to obtain the annotated crack image data.

[0018] The method of obtaining the optimized multi-scale crack defect assessment model comprises the following steps:

[0019] The evaluated annotated tunnel crack image data is obtained as a dataset and divided into a training set, a validation set, and a test set.

[0020] Select binary cross entropy as the loss function, SGD as the optimizer, and set hyperparameters, including training batch size, learning rate, number of iterations, regularization parameter, dropout ratio, activation function, and architecture parameters of the multi-scale crack defect assessment model;

[0021] The validation set is used to perform model evaluation and hyperparameter adjustment on the multi-scale crack defect assessment model to obtain the optimized multi-scale crack defect assessment model.

[0022] The method of obtaining the crack defect assessment result comprises the following steps:

[0023] Extracting features from the pre-processed tunnel crack image data to be evaluated using an encoder to obtain extracted features, and enhancing the extracted features using CBA, SENet, and Non-local to obtain multi-level enhanced features;

[0024] The feature maps of the lower levels in the multi-level enhanced features are upsampled to higher levels to obtain the sampled feature maps, and then the upsampled feature maps are spliced ​​with the original feature maps of the higher levels using side connections to obtain spliced ​​features;

[0025] Apply convolution operation on the spliced ​​feature map to fuse features of different scales. Repeat the above fusion process for each pyramid level until the top level to obtain multi-scale fused feature map data.

[0026] The multi-scale fusion feature mapping data is subjected to contour detection, classification and quantification to obtain the crack defect assessment results.

[0027] The contour detection, classification and quantification processing of the multi-scale fusion feature map data includes the following steps:

[0028] The multi-scale fusion feature map data is used to identify possible crack areas by setting the activation function Sigmoid to obtain multiple crack areas;

[0029] Applying conditional random fields and region growing to refine the multiple crack regions to obtain multiple refined crack regions;

[0030] Perform contour extraction on multiple refined crack regions to obtain the boundaries of multiple cracks;

[0031] Verifying and obtaining specific boundaries of multiple cracks through morphological operations and geometric constraints to obtain multiple detected crack regions; the geometric constraints include length and width ratios of the cracks;

[0032] Extract feature vectors for each detected crack region based on multi-scale fusion feature mapping, including texture, shape, and size;

[0033] Normalize the feature vector to obtain a standardized feature vector;

[0034] Input the normalized feature vector into the pre-trained classification model to obtain the crack classification result;

[0035] The classification results are processed by limiting the overlap using non-maximum suppression to generate crack defect assessment results.

[0036] A tunnel crack defect assessment system, comprising:

[0037] A data acquisition module is used to acquire crack image data of the tunnel to be evaluated, perform labeling and preprocessing, and obtain labeled crack image data;

[0038] An evaluation model construction module is used to use the U-Net architecture as a basis, add a feature pyramid network, and obtain a hybrid network structure. The CBAM is inserted after the convolution block of the hybrid network structure, SENet is added after the bottleneck layer between the encoder and decoder, and a non-local module is placed after the deepest convolution block of the encoder to construct a multi-scale crack defect assessment model. The module obtains evaluated annotated crack image data as a dataset, trains and optimizes the parameters of the multi-scale crack defect assessment model, and obtains an optimized multi-scale crack defect assessment model.

[0039] The evaluation and analysis module is used to pre-process the tunnel crack image data to be evaluated and input it into the optimized multi-scale crack defect evaluation model to obtain the crack defect evaluation result.

[0040] Compared with existing technologies, the present invention achieves the following benefits: By improving the U-Net architecture and introducing a feature pyramid network to achieve multi-scale feature fusion, the model can simultaneously capture global semantic features and local detail features of cracks, making it particularly suitable for detecting cracks of various scales and complex shapes. Furthermore, by inserting CBAM, SENet, and Non-local modules, the model focuses on prominent crack regions, suppresses background noise interference on feature extraction, emphasizes the importance of feature channels, improves crack detection sensitivity, captures long-range dependencies, and enhances the model's ability to detect crack discontinuities and discontinuities. This improves crack detection accuracy, particularly under complex backgrounds and low-contrast conditions. This helps the model focus on key crack regions, enhances feature characterization capabilities, and enables precise crack classification and quantification. Detected cracks can be classified by length, width, depth, and quantified by defect grade, providing detailed information for tunnel maintenance. This significantly increases the speed of crack detection, particularly in large-scale tunnel inspection projects, reducing the time and labor costs required for manual inspections while also reducing the risk of false detections and missed detections due to human error. Accurately assessing tunnel crack defects allows for timely maintenance, ensuring the safety and reliability of the tunnel structure. This systematic approach makes the model easy to integrate and deploy, and subsequent maintenance relatively straightforward.

[0041] In summary, the tunnel crack defect assessment method and technical system implemented above not only improves the automation and efficiency of detection, but also enhances the accuracy and safety of detection, providing strong technical support for tunnel maintenance and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0043] Figure 1 It is a schematic diagram of the process structure of the present invention;

[0044] Figure 2 It is a schematic diagram of the process of constructing a multi-scale crack defect assessment model according to the present invention;

[0045] Figure 3 It is a schematic diagram of the process of obtaining crack image data according to the present invention;

[0046] Figure 4 Schematic diagram of the process of optimizing the multi-scale crack defect assessment model of the present invention;

[0047] Figure 5 It is a schematic diagram of the main process of crack defect evaluation results of the present invention;

[0048] Figure 6 It is a schematic diagram of the process framework of the crack defect evaluation result of the present invention;

[0049] Figure 7 Schematic diagram of the framework structure of the tunnel crack defect assessment system of the present invention;

[0050] Figure 8 Schematic diagram of the network structure of the multi-scale crack defect assessment model in the present invention.

[0051] 1.CBAM module, 2.SENet module, 3.Non-local module. DETAILED DESCRIPTION

[0052] The limitations of existing tunnel crack detection technologies are mainly reflected in the following aspects: Low detection efficiency: Traditional methods require a lot of manpower and time for detection, especially in long-distance tunnels. Insufficient detection accuracy: The detection ability of small cracks or deep cracks is insufficient, which can easily lead to missed detection. Complex data processing: Although some technologies can collect a large amount of data, they lack effective data processing methods and cannot quickly and accurately analyze crack information. Poor environmental adaptability: Many detection technologies are greatly affected by environmental factors such as temperature, humidity, light, etc., which affects the accuracy of detection. Therefore, it is necessary to develop a new tunnel crack defect assessment method to improve the efficiency and accuracy of detection, reduce the impact of environmental factors, and realize automated and intelligent crack detection.

[0053] In order to overcome the above shortcomings of the prior art, the main purpose of the present invention is to provide a tunnel crack defect assessment method. Figures 1-8 , including the following steps:

[0054] Acquire crack image data of a tunnel to be evaluated, perform labeling and preprocessing, and obtain labeled crack image data;

[0055] A hybrid network structure is constructed by adding a feature pyramid network to the U-Net architecture. The CBAM is inserted after the convolutional blocks of the hybrid network structure, and SENet is added after the bottleneck layer between the encoder and decoder. Furthermore, a non-local module is placed after the deepest convolutional block of the encoder to construct a multi-scale crack defect assessment model. Crack image data of an evaluated tunnel is obtained as a dataset, and the parameters of the multi-scale crack defect assessment model are trained and optimized to obtain an optimized multi-scale crack defect assessment model.

[0056] The pre-processed crack image data of the tunnel to be evaluated is obtained and input into the optimized multi-scale crack defect evaluation model to obtain a crack defect evaluation result.

[0057] Among them, the construction of multi-scale crack defect assessment model, see Figure 2 , including the following steps:

[0058] Design the U-Net architecture, such as Figure 8 As shown in the figure, the encoder implements convolutional and pooling layers, while the decoder implements upsampling and convolutional layers. The core of the U-Net architecture design lies in deep feature extraction in the encoder and precise localization recovery in the decoder. Convolutional layers extract image features, while pooling layers reduce feature dimensionality and expand the receptive field. In the encoder, stacking convolutional and pooling layers achieves feature abstraction from low-level to high-level layers, providing rich context for subsequent defect recognition. Pooling layers reduce computational complexity while helping the model capture a wider range of crack defect distributions. In the decoder, upsampling layers restore the spatial resolution of the image, while convolutional layers refine the feature map, enhancing the model's ability to recognize crack details. Upsampling layers gradually restore the size of the feature map through interpolation or transposed convolution, supporting precise crack location. Convolutional layers in the decoder further refine features, enhance edge information of crack defects, and improve the model's localization accuracy.

[0059] Taking the U-Net architecture as the basis, after adding the feature pyramid network, multiple attention modules are introduced. Specifically, CBAM is placed after the convolution block of the U-Net architecture, SENet is placed after the bottleneck layer between the encoder and decoder, and a non-local module is placed after the deepest convolution block of the encoder to build a multi-scale crack defect assessment model.

[0060] The hybrid network architecture incorporates attention modules, including the Convolutional Block Attention Module (CBAM). CBAM is an effective attention mechanism that enhances the model's focus on key features by focusing on both channel and spatial dimensions. The channel attention mechanism assesses the importance of different channels and assigns weights to each, allowing the model to focus on features that contribute to crack detection. This helps the model accurately identify crack defects in complex backgrounds, improving the robustness of defect detection. The spatial attention mechanism focuses on the importance of different locations in the image, assigning weights to each pixel to precisely locate cracks. Guided by spatial attention, the model is able to focus more on crack regions, reduce interference from non-crack areas, and improve crack detection accuracy. The introduction of the spatial attention mechanism enables the model to better adapt to changes in crack shape and position when processing crack defects, improving its generalization. Furthermore, SENet (Squeeze-and-Excitation Networks) adaptively recalibrates the importance of feature channels, enhancing the model's ability to represent crack characteristics. The SE module learns the correlation between channels through global average pooling and fully connected layers, generates an attention weight for each channel, and realizes dynamic adjustment of feature channels. The introduction of this module enables the model to focus more on feature channels that are important for crack identification, improving the accuracy and robustness of crack defect assessment. There is also the fusion of the Non-local module. The Non-local module uses the self-attention mechanism to model the long-distance dependency between any two locations in the image, enhancing the model's utilization of global contextual information. By calculating the correlation between each location and all other locations, the Non-local module can capture the global information of crack defects and improve the model's ability to recognize cracks. The introduction of this module enables the model to take into account more distant contextual information when processing crack defects, improving the accuracy and robustness of crack detection.

[0061] Among them, refer to 8, the addition position of CBAM, SENet and Non-local modules in the hybrid network structure

[0062] Where to add the CBAM (Convolutional Block Attention Module): The CBAM module is inserted after each convolution block in the hybrid network structure. This means that after each convolution operation, channel attention and spatial attention are calculated to enhance feature representation.

[0063] SENet (Squeeze-and-Excitation Network) is added after the bottleneck layer between the encoder and decoder.

[0064] The bottleneck layer is the part of the network that connects the encoder and decoder. Typically, the feature maps have a smaller spatial size but a larger number of channels. Adding SENet here can effectively model the dependencies between channels and improve feature representation capabilities.

[0065] Where to add the non-local module: The non-local module is placed after the deepest convolutional block in the encoder. This deepest convolutional block typically contains the most abstract and discriminative features. Adding the non-local module here can capture long-range dependencies and enhance the model's ability to utilize global information.

[0066] The specific structure is as follows:

[0067] Input->Preprocessing->Annotation->Hybrid Network Structure->Output

[0068] See 8, where the hybrid network structure is:

[0069]

[0070]

[0071] Each convolutional block is followed by a CBAM module to enhance feature representation.

[0072] SENet is added after the bottleneck layer to model the dependencies between channels.

[0073] The non-local module is placed after the deepest convolutional block of the encoder to capture long-distance dependencies.

[0074] This structural design enables the multi-scale crack defect assessment model to fully utilize local and global information and improve the accuracy of crack defect assessment.

[0075] To obtain the annotated crack image data, refer to Figure 3 , including the following steps:

[0076] Define the tunnel crack criteria, select annotation tools to annotate crack image data of the tunnel to be evaluated, identify the crack locations, draw bounding boxes for each crack, and perform pixel-level annotation using semantic segmentation to obtain preliminarily annotated crack image data;

[0077] The initially annotated crack image data is subjected to image enhancement, size adjustment, normalization and cleaning, and then data segmentation is performed to obtain the annotated crack image data.

[0078] Obtain the optimized multi-scale crack defect assessment model, see Figure 4 , including the following steps:

[0079] The crack image data of the evaluated tunnel is obtained as a dataset and divided into a training set, a validation set, and a test set.

[0080] Select binary cross entropy as the loss function, SGD as the optimizer, and set hyperparameters, including training batch size, learning rate, number of iterations, regularization parameter, dropout ratio, activation function, and architecture parameters of the multi-scale crack defect assessment model;

[0081] The validation set is used to perform model evaluation and hyperparameter adjustment on the multi-scale crack defect assessment model to obtain the optimized multi-scale crack defect assessment model.

[0082] To obtain the crack defect assessment results, the following steps are included:

[0083] Extracting features from the pre-processed tunnel crack image data to be evaluated using an encoder to obtain extracted features, and enhancing the extracted features using CBA, SENet, and Non-local to obtain multi-level enhanced features;

[0084] The feature maps of the lower levels in the multi-level enhanced features are upsampled to higher levels to obtain the sampled feature maps, and then the upsampled feature maps are spliced ​​with the original feature maps of the higher levels using side connections to obtain spliced ​​features;

[0085] Apply convolution operation on the spliced ​​feature map to fuse features of different scales. Repeat the above fusion process for each pyramid level until the top level to obtain multi-scale fused feature map data.

[0086] The multi-scale fusion feature mapping data is subjected to contour detection, classification and quantification to obtain the crack defect assessment results.

[0087] and contour detection, classification, and quantification of multi-scale fused feature maps, see Figure 6 , including the following steps:

[0088] The multi-scale fusion feature map data is used to identify possible crack areas by setting the activation function Sigmoid to obtain multiple crack areas;

[0089] Applying conditional random fields and region growing to refine the multiple crack regions to obtain multiple refined crack regions;

[0090] Perform contour extraction on multiple refined crack regions to obtain the boundaries of multiple cracks;

[0091] Verifying and obtaining specific boundaries of multiple cracks through morphological operations and geometric constraints to obtain multiple detected crack regions; the geometric constraints include length and width ratios of the cracks;

[0092] Extract feature vectors for each detected crack region based on multi-scale fusion feature mapping, including texture, shape, and size;

[0093] Normalize the feature vector to obtain a standardized feature vector;

[0094] Use the training data set to train the classification model constructed by the neural network.

[0095] The normalized feature vector of the detected crack is input into the trained classification model to obtain the classification result of the crack.

[0096] The classification results are subjected to non-maximum suppression to merge the overlapping crack classification results and generate crack defect assessment results.

[0097] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0098] Example 1:

[0099] The following is a specific implementation case to illustrate the steps of the above tunnel crack defect assessment method:

[0100] The first step is data acquisition and preprocessing

[0101] Step 1: Use a high-definition camera mounted on a drone to photograph a city subway tunnel and obtain a series of tunnel surface images.

[0102] Step 2: Define the crack marking standard, such as only cracks with a width greater than 0.5mm will be marked.

[0103] Step 3: Use annotation tools such as LabelImg to annotate the cracks in the image, including crack location, bounding box drawing, and pixel-level semantic segmentation.

[0104] Step 4: Enhance the annotated image by rotating, flipping, scaling, resizing it to a uniform size of 256x256 pixels, normalizing it, scaling the pixel values ​​to the range of 0-1, and cleaning it to remove blurry or unclear images.

[0105] Step 5: Split the image data into 70% training set, 15% validation set and 15% test set.

[0106] 2. Constructing a multi-scale crack defect assessment model

[0107] Step 1: Design a U-Net-based model, including the convolutional and pooling layers of the encoder, and the upsampling and convolutional layers of the decoder.

[0108] Step 2: Add side connections in the encoder and integrate the feature pyramid network to fuse features at different scales.

[0109] Step 3: Introduce the CBAM attention module after the convolution block of the U-Net architecture, introduce the SENet module after the bottleneck layer, and introduce the Non-local module after the convolution block of the deepest layer of the encoder.

[0110] 3. Model training and optimization

[0111] Step 1: Use the evaluated tunnel crack image dataset and perform model training according to the previous data segmentation.

[0112] Step 2: Select binary cross entropy as the loss function, SGD as the optimizer, and set hyperparameters such as batch size (32), learning rate (0.001), number of iterations (10000), etc.

[0113] Step 3: Use the validation set to evaluate the model and adjust hyperparameters, such as the learning rate decay strategy, based on the evaluation results until the model performance reaches the best.

[0114] 4. Crack defect assessment

[0115] Step 1: Preprocess the new tunnel image to be evaluated, including resizing and normalization.

[0116] Step 2: Build an image pyramid and apply convolutional neural networks to extract features at each level.

[0117] Step 3: Concatenate feature maps of different scales through upsampling and side connections.

[0118] Step 4: Apply convolution operation on the concatenated feature map to fuse the features, and repeat this process until the top layer.

[0119] Step 5: Perform contour detection, classification (crack length, width) and quantified defect level processing on the multi-scale fusion feature map.

[0120] 5. Result Output

[0121] Step 1: Extract features of the detected crack area, including texture, shape, size, etc.

[0122] Step 2: Normalize the feature vector and input it into the trained classification model to obtain the crack classification result.

[0123] Step 3: Use non-maximum suppression to merge overlapping crack classification results to generate the final crack defect assessment results.

[0124] Data acquisition module: automatically acquires tunnel images taken by drones and performs preprocessing.

[0125] Evaluation model building module: Build and train a multi-scale crack defect evaluation model.

[0126] Evaluation and analysis module: input pre-processed image data and output crack defect evaluation results.

[0127] Through the above implementation cases, this method can automatically and accurately evaluate tunnel crack defects and provide decision support for tunnel maintenance and management.

[0128] Example 2

[0129] A tunnel crack defect assessment system, see Figure 7 ,include:

[0130] An evaluation model construction module uses a U-Net architecture as a foundation and adds a feature pyramid network to obtain a hybrid network structure. CBAM is inserted after the convolutional block of the hybrid network structure, SENet is added after the bottleneck layer between the encoder and decoder, and a non-local module is placed after the deepest convolutional block of the encoder to construct a multi-scale crack defect assessment model. Evaluated annotated crack image data is obtained as a dataset, and parameters of the multi-scale crack defect assessment model are trained and optimized to obtain an optimized multi-scale crack defect assessment model.

[0131] The evaluation and analysis module is used to obtain the marked tunnel crack image data to be evaluated, input it into the optimized multi-scale crack defect evaluation model, and obtain the crack defect evaluation result.

[0132] That is, the system improves the accuracy and efficiency of crack detection by integrating advanced image processing and deep learning technologies. The following are the specific implementation steps of the system:

[0133] Data acquisition module:

[0134] Crack image acquisition: First, crack image data is collected from the tunnel. This image data needs to cover different stages and types of cracks so that the system can learn and identify various crack patterns.

[0135] Image annotation and preprocessing: The collected images are annotated to identify the specific location and morphology of cracks. In addition, necessary preprocessing operations such as cropping, denoising, and normalization are performed to optimize the subsequent image analysis process.

[0136] Evaluation Model Building Module:

[0137] Hybrid network architecture: A U-Net-based architecture is combined with a feature pyramid network to enhance the model's ability to identify cracks of varying scales and complexity, resulting in a hybrid network structure. Furthermore, a CBAM convolutional block attention module, a SENet compression-excitation network, and a non-local module are added to enhance the model's ability to extract features and focus on key areas.

[0138] Model training: Using labeled tunnel crack image data as a training dataset, the model parameters are trained and optimized. This process involves multiple iterations, where the network parameters are continuously adjusted to minimize prediction error, resulting in more accurate crack detection performance.

[0139] Evaluation and analysis module:

[0140] Crack Identification and Analysis: The preprocessed tunnel crack image is fed into a trained multi-scale crack defect assessment model. The model detects and identifies the presence of cracks by analyzing features in the image.

[0141] Evaluation output: The model outputs the type, size, shape, and possible growth trend of the crack. These outputs can be directly used for decision support, such as tunnel repair and reinforcement planning.

[0142] This advanced crack identification system enables rapid and efficient crack detection, significantly improving the efficiency and accuracy of tunnel safety monitoring. This system is particularly well-suited for long-term, continuous structural health monitoring, accurately diagnosing cracks in their early stages of development and effectively preventing larger-scale structural damage.

[0143] This system, which integrates multiple deep learning and image processing technologies, provides strong technical support for the maintenance of infrastructure such as tunnels, and demonstrates the application prospects of modern engineering technology in traditional industries.

[0144] Example 3

[0145] The following is a specific example of a tunnel crack defect assessment method, demonstrating the actual operation steps and corresponding experimental data, to verify the detection and classification effectiveness of the proposed multi-scale tunnel crack defect assessment model in complex environments.

[0146] Raw data: 5,000 crack images collected from different tunnels. The images have a resolution of 1024 × 1024 pixels and cover a variety of crack types, including surface cracks, through-hole cracks, and spiderweb cracks.

[0147] Annotation Standard: Use bounding boxes to annotate the crack area and accurately record the shape and boundaries of the crack.

[0148] Semantic segmentation is used for pixel-level annotation (the crack area is represented by label 1, and the rest of the area is represented by label 0).

[0149] Image enhancement methods: rotation (±30°), brightness adjustment (±20%), Gaussian noise (σ=0.01). Data segmentation: training set, validation set, and test set are divided into a ratio of 7:2:1.

[0150] The U-Net infrastructure is embedded in the Feature Pyramid Network (FPN), combined with CBAM, SENet, and Non-local modules. Specific hyperparameter settings include:

[0151] Loss function: Binary Cross Entropy.

[0152] Optimizer: SGD, initial learning rate 0.01, momentum 0.9.

[0153] Batch size: 16, number of training iterations: 100.

[0154] Regularization: L2 regularization coefficient 0.0005, dropout ratio 0.5.

[0155] Training equipment: NVIDIA Tesla V100 GPU, training time is about 12 hours.

[0156] Test steps and results:

[0157] (1) Crack detection effect

[0158] The 500 crack images in the test set are fed into the trained model to obtain the following detection metrics:

[0159] Detection accuracy (Precision): 96.3%.

[0160] Recall rate: 94.8%.

[0161] F1-Score: 95.5%.

[0162] The crack area of ​​the detection image is accurately marked, and the boundary after refinement is highly consistent with the actual crack. See the data below for details.

[0163] Image number Number of cracks detected Actual number of cracks Accuracy (%) Missed detection rate (%) 001 12 12 100 0 002 8 9 88.9 11.1 003 15 15 100 0

[0164] (2) Crack classification effect

[0165] The model classifies cracks into three categories using normalized feature vectors of crack texture, shape, and size:

[0166] Type I (shallow crack): depth less than 2mm;

[0167] Type II (medium crack): depth 2-5mm;

[0168] Category III (deep cracks): depth greater than 5 mm.

[0169] Crack type Classification accuracy (%) Misclassification rate (%) Category I 95.8 4.2 Category II 94.7 5.3 Category III 93.2 6.8

[0170] Specific detection and classification cases:

[0171] Tunnel crack image (see table below):

[0172]

[0173] Each crack region is represented by a bounding box, and the classification results are distinguished by color (green for Class I, blue for Class II, and red for Class III).

[0174] Assessment report: Total number of cracks: 35; Class I: 20, accounting for 57.1%; Class II: 10, accounting for 28.6%; Class III: 5, accounting for 14.3%.

[0175] Crack quantification analysis: Texture mean: 0.65; Texture variance: 0.12. Shape parameters: Average length: 5.2 cm, width: 0.8 cm.

[0176] After the standardized feature vector was input into the classification model, the accuracy of the quantitative analysis of cracks reached 97.4%.

[0177] The above examples verify that the tunnel crack defect assessment method proposed in the present invention can efficiently and accurately detect and classify cracks.

[0178] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises", or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or apparatus.

[0179] The above embodiments are merely examples of the present invention and do not limit the scope of protection of the present invention. Any designs that are identical or similar to the present invention fall within the scope of protection of the present invention.

Claims

1. A method for evaluating tunnel crack defects, characterized in that: The following steps are involved: Using the U-Net architecture as a foundation, a feature pyramid network was added to obtain a hybrid network structure. CBAM was inserted after the convolutional block of the hybrid network structure, SENet was added after the bottleneck layer between the encoder and decoder, and a non-local module was placed after the deepest convolutional block of the encoder to construct a multi-scale crack defect assessment model. Acquire the evaluated annotated crack image data as a data set, train and optimize the parameters of the multi-scale crack defect assessment model, and obtain an optimized multi-scale crack defect assessment model; After pre-processing, the obtained tunnel crack image data to be evaluated is input into the optimized multi-scale crack defect evaluation model to obtain a crack defect evaluation result; The method of obtaining the hybrid network structure comprises the following steps: The U-Net architecture includes a downsampling path and an upsampling path; The FPN is implemented through a top-down path and a bottom-up path; After each downsampling stage of the U-Net, the top-down path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the previous layer; before each upsampling stage of the U-Net, the bottom-up path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the next layer to obtain a hybrid network structure; The method of obtaining the crack defect assessment result comprises the following steps: Extracting features from the pre-processed tunnel crack image data to be evaluated using an encoder to obtain extracted features, and enhancing the extracted features using CBA, SENet, and Non-local to obtain multi-level enhanced features; The feature maps of the lower levels in the multi-level enhanced features are upsampled to higher levels to obtain the sampled feature maps, and then the upsampled feature maps are spliced ​​with the original feature maps of the higher levels using side connections to obtain spliced ​​features; Apply convolution operation on the spliced ​​feature map to fuse features of different scales. Repeat the above fusion process for each pyramid level until the top level to obtain multi-scale fused feature map data. The multi-scale fusion feature mapping data is subjected to contour detection, classification and quantification to obtain the crack defect assessment results.

2. The tunnel crack defect assessment method according to claim 1, wherein: The method of obtaining the annotated crack image data comprises the following steps: Define the tunnel crack criteria, select annotation tools to annotate crack image data of the tunnel to be evaluated, identify the crack locations, draw bounding boxes for each crack, and perform pixel-level annotation using semantic segmentation to obtain preliminarily annotated crack image data; The initially annotated crack image data is subjected to image enhancement, size adjustment, normalization and cleaning, and then data segmentation is performed to obtain the annotated crack image data.

3. The tunnel crack defect assessment method according to claim 1, wherein: The method of obtaining the optimized multi-scale crack defect assessment model comprises the following steps: Obtain the evaluated annotated tunnel crack image data as a dataset and divide it into a training set, a validation set, and a test set; Select binary cross entropy as the loss function, SGD as the optimizer, and set hyperparameters, including training batch size, learning rate, number of iterations, regularization parameter, dropout ratio, activation function, and architecture parameters of the multi-scale crack defect assessment model; The validation set is used to perform model evaluation and hyperparameter adjustment on the multi-scale crack defect assessment model to obtain the optimized multi-scale crack defect assessment model.

4. The tunnel crack defect assessment method according to claim 1, wherein: The contour detection, classification and quantification processing of the multi-scale fusion feature map data includes the following steps: The multi-scale fusion feature map data is used to identify possible crack areas by setting the activation function Sigmoid to obtain multiple crack areas; Applying conditional random fields and region growing to refine the multiple crack regions to obtain multiple refined crack regions; Perform contour extraction on multiple refined crack regions to obtain the boundaries of multiple cracks; Verifying and obtaining specific boundaries of multiple cracks through morphological operations and geometric constraints to obtain multiple detected crack regions; the geometric constraints include length and width ratios of the cracks; Extract feature vectors for each detected crack region based on multi-scale fusion feature mapping, including texture, shape, and size; Normalize the feature vector to obtain a standardized feature vector; Input the normalized feature vector into the pre-trained classification model to obtain the crack classification result; The classification results are processed by limiting the overlap using non-maximum suppression to generate crack defect assessment results.

5. A tunnel crack defect assessment system, characterized in that: include: An evaluation model construction module is used to use the U-Net architecture as a basis, add a feature pyramid network, and obtain a hybrid network structure. The CBAM is inserted after the convolution block of the hybrid network structure, SENet is added after the bottleneck layer between the encoder and decoder, and a non-local module is placed after the deepest convolution block of the encoder to construct a multi-scale crack defect assessment model. The module obtains evaluated annotated crack image data as a dataset, trains and optimizes the parameters of the multi-scale crack defect assessment model, and obtains an optimized multi-scale crack defect assessment model. An evaluation and analysis module, configured to pre-process the acquired tunnel crack image data to be evaluated and then input the data into the optimized multi-scale crack defect evaluation model to obtain a crack defect evaluation result; The method of obtaining the hybrid network structure comprises the following steps: The U-Net architecture includes a downsampling path and an upsampling path; The FPN is implemented through a top-down path and a bottom-up path; After each downsampling stage of the U-Net, the top-down path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the previous layer; before each upsampling stage of the U-Net, the bottom-up path of the FPN is introduced to concatenate and fuse the features of the current layer with the features of the next layer to obtain a hybrid network structure; The method of obtaining the crack defect assessment result comprises the following steps: Extracting features from the pre-processed tunnel crack image data to be evaluated using an encoder to obtain extracted features, and enhancing the extracted features using CBA, SENet, and Non-local to obtain multi-level enhanced features; The feature maps of the lower levels in the multi-level enhanced features are upsampled to higher levels to obtain the sampled feature maps, and then the upsampled feature maps are spliced ​​with the original feature maps of the higher levels using side connections to obtain spliced ​​features; Apply convolution operation on the spliced ​​feature map to fuse features of different scales. Repeat the above fusion process for each pyramid level until the top level to obtain multi-scale fused feature map data. The multi-scale fusion feature mapping data is subjected to contour detection, classification and quantification to obtain the crack defect assessment results.

6. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 4 is implemented.

Images