Detection method of internal cavity and non-compact defects in tunnel lining based on GPR data
By using the improved DETR deep learning framework and Hybrid Vision Transformers, combined with data augmentation and transfer learning, intelligent detection of internal cavities and non-compact defects in tunnel linings was achieved. This solved the problems of reliance on manual interpretation and insufficient model generalization ability in traditional methods, and improved the recognition accuracy and robustness.
Patent Information
- Application Number
- CN202410710750.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-04
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2044-06-04
Smart Images

Figure CN118628452B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tunnel safety, specifically a method for detecting internal cavities and non-compact defects in tunnel lining based on GPR data. Background Technology
[0002] Tunnels play a vital role in transportation, water conservancy projects, and infrastructure. However, due to factors such as increased service life, changes in geological conditions, and natural weathering, tunnel linings often develop various defects, including cavities and loose linings. These defects can lead to tunnel instability, thereby jeopardizing tunnel operational safety. Therefore, during tunnel construction, accurately classifying and locating cavities and loose lining defects is crucial for maintaining tunnel safety. Traditional GPR data interpretation relies heavily on manual work, which suffers from significant drawbacks such as high subjectivity, long interpretation cycles, and high technical requirements. Therefore, automatic identification and interpretation methods for radar spectra are extremely necessary in scientific research and practical applications, and are gradually becoming a hot topic in the industry.
[0003] In recent years, deep learning networks have been widely used for identifying and classifying ground-penetrating radar (GPR) data images, with the Convolutional Neural Network (CNN) architecture being the most popular. While CNN-based research has effectively identified the location, category, and probability of defects in GPR data images, it suffers from drawbacks such as requiring large datasets, weak generalization ability, and a lack of global information.
[0004] Recently, the emergence of a new deep learning model, the Transformer, has brought new solutions to GPR data processing and defect identification. Utilizing an encoder-decoder architecture, the Transformer has distinguished itself in natural language processing due to its self-attention mechanism, parallel computing power, flexibility, and scalability, and is gradually gaining widespread application in computer vision. The Vision Transformer (ViT) is a typical Transformer model used for visual processing. However, the global receptive field of the ViT model also leads to high computational costs, limiting its practical application. Therefore, Hybrid Vision Transformers (HVT), also known as CNN-Transformers, were introduced to combine the power of CNN and ViT. These hybrid models utilize the convolutional layers of CNN to capture local features and then feed them into ViT to obtain global context using a self-attention mechanism. HVT has demonstrated better performance in many image recognition tasks.
[0005] End-to-End Object Detection with Transformers (DETR) is an HVT architecture proposed in 2020 for performing object detection in natural images. The entire network implements end-to-end object detection, abandoning the previously heavily knowledge-dependent Non-Maximum Suppression (NMS) and anchor parts, greatly simplifying the object detection pipeline and making DETR easily transferable to other tasks. Compared to other representative detection models (such as Faster R-CNN), DETR exhibits better performance.
[0006] The differences compared to existing technologies are as follows;
[0007] Technical Comparison with Patent CN115222999B "Method and Device for Classifying Images of Intrinsic Defects in Concealed Structures of Tunnels"
[0008] Patent CN115222999B utilizes simulated image datasets and actual image datasets for transfer learning. This patent, however, uses the publicly available COCO2017 dataset and measured datasets of two tunnel components (steel bars and steel arches) as samples, resulting in a lower cost for transfer learning.
[0009] Patent CN115222999B does not cover data augmentation, but this patent combines traditional data augmentation methods to perform augmentation operations on the GPR dataset, such as random mirror flipping, Gaussian blurring, brightness transformation, scale transformation, and adding noise, to further improve the model's generalization ability and increase the diversity of samples.
[0010] Patent CN115222999B uses YOLOv3 as the tunnel defect detection model. This patent improves upon the DETR-based deep learning framework by embedding a novel and efficient multi-scale attention module—MCA—into the original architecture. This allows the model to dynamically focus on the most useful features based on task requirements and content context, thereby discarding the non-maximum suppression and anchor parts that heavily rely on prior knowledge. This significantly simplifies the target detection process and enables end-to-end automatic detection of internal defects in tunnel linings.
[0011] Patent CN115222999B does not specify which particular tunnel defect it applies to. However, this patent is specifically designed for the detection of tunnel cavities and deficiencies.
[0012] Technical comparison with patent CN117538863A "Method, device, equipment and medium for improving the accuracy of ground-penetrating radar detection of tunnel lining defects"
[0013] Patent CN117538863A does not involve transfer learning. However, this patent uses the publicly available COCO2017 dataset and measured datasets of two types of tunnel components (steel bars and steel arches) as samples, which improves the model's convergence speed and enhances its robustness.
[0014] Patent CN115222999B uses a modified convolutional neural network based on VGG-RESNET chirp as a tunnel defect detection model. This patent improves upon the DETR-based deep learning framework by embedding a novel and efficient multi-scale attention module—MCA—into the original architecture. This allows the model to dynamically focus on the most useful features based on task requirements and content context, thereby abandoning the non-maximum suppression and anchor parts that heavily rely on prior knowledge. This significantly simplifies the target detection process and enables end-to-end automatic detection of internal defects in tunnel linings.
[0015] Patent CN117538863A does not mention which specific tunnel defect it applies to. However, this patent is specifically designed for the detection of tunnel cavities and deficiencies. Summary of the Invention
[0016] To address the above problems, this invention proposes a method for detecting internal cavities and non-compact defects in tunnel linings based on GPR data. By improving the deep learning framework based on the DETR model and combining two types of tunnel component datasets—steel reinforcement and steel arch—for transfer learning on different backbone networks, the model performance is improved, enabling intelligent detection of cavities and non-compact defects in tunnel ground-penetrating radar images.
[0017] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0018] A method for detecting internal cavities and non-compactness defects in tunnel lining based on GPR data includes the following steps:
[0019] S1, Ground-penetrating radar image acquisition inside the tunnel:
[0020] Images representing a wide variety of elements inside the tunnel lining, including steel bars, steel arches, non-compact defects, and cavities, were collected as pre-training data.
[0021] S2, Image Preprocessing and Construction of Tunnel Components—Defect Dataset:
[0022] For the images acquired by S1, the image quality is checked, and images with poor quality or damage are filtered out or repaired. The image size and resolution are adjusted to ensure the consistency and validity of the dataset. Then, data augmentation is performed. Traditional data augmentation methods are used to perform random image mirroring, Gaussian blur, brightness transformation, scale transformation, and noise enhancement operations on the training set data to further improve the model's generalization ability and increase the diversity of samples. A tunnel component-defect dataset is established for model training. The images in the dataset are divided into training set, validation set, and test set in a 7:2:1 ratio.
[0023] S3, improve the DETR-based deep learning network framework and use the tunnel component dataset for transfer learning training to obtain a pre-trained model;
[0024] S4, Defect Detection:
[0025] Images from the tunnel defect dataset are input into a segmentation model to obtain corresponding detection results. Then, these segmentation results are comprehensively analyzed to obtain the final result, thus achieving intelligent detection of internal defects in the tunnel lining.
[0026] Furthermore, the S2 data enhancement involves the following specific steps:
[0027] The data augmentation described in S2 uses the following transformation methods on the tunnel component-defect dataset to further improve the model's generalization ability and increase sample diversity:
[0028] Mirror flip: Mirrors the image horizontally;
[0029] Gaussian blur: Makes the image more blurred by using a Gaussian blur filter of a specified size;
[0030] Brightness Transformation: Increases or decreases the brightness of the image;
[0031] Add noise: Add an appropriate amount of salt and pepper noise or Gaussian noise to the image;
[0032] Scale transformation: Preserves the feature regions of the image and performs cropping and enlargement;
[0033] For each transformation method, different parameters and ranges are used to generate different transformed images: for brightness transformation, two transformed images are generated using two ratios of 0.8 and 1.2; for adding noise, two transformed images are generated using two noise ratios of 0.001 and 0.002; for scale transformation, three transformed images are generated using three methods: image center, image edge, and random region as the transformation center; finally, multiple transformed images are obtained.
[0034] Furthermore, step S3 improves the DETR-based deep learning network framework and uses the tunnel component dataset for transfer learning training. The specific steps are as follows:
[0035] S31, Improved DETR-based deep learning network framework:
[0036] First, the original DETR network framework is built. The original DETR network's image processing consists of three parts: first, it passes through the CNN backbone for extracting compact feature representations; then, it is input into the Transformer encoder and decoder; finally, the decoder output passes through a simple feedforward network responsible for the final detection and prediction. For the improved DETR network, ResNet-101 is used as the backbone network to extract high-level semantic features of the image and output feature maps at different scales. Two sequential attention modules are added to the output feature map of the backbone network: channel attention and spatial attention. The output feature map of the channel attention module serves as the input of the spatial attention module. After the features pass through the two modules in sequence, the final attention feature map is obtained. The subsequent processing is consistent with the original DETR model. The obtained attention features are added to the position encoding and then input into the Transformer architecture. The Decoder layer obtains N query vectors, which are then passed through the FFN layer to obtain N prediction boxes and their corresponding categories.
[0037] S32, using the tunnel component dataset for transfer learning training:
[0038] By utilizing transfer learning, the dataset of internal components of tunnel lining is used as an auxiliary domain for knowledge transfer. First, a coarse-grained model is obtained using the COC02017 dataset, and the model parameters are transferred to the tunnel component dataset to obtain a fine-grained model. Finally, the model parameters are transferred to the tunnel defect identification task, which improves the model convergence speed and enhances its robustness.
[0039] S33. Adjust training parameters to determine the optimal model: Training parameters include learning rate, batch size, optimizer, and number of training rounds, which affect the convergence speed and stability of the model, and finally determine the optimal training model.
[0040] Furthermore, the addition of two serial attention modules in step S31 is described in the following steps:
[0041] Channel attention module:
[0042] First, the spatial information of the feature map is aggregated through average pooling and max pooling operations to generate two different spatial context descriptors. and Let's represent the average pooling feature and the max pooling feature, respectively. Then, these two features are fed into a shared layer to generate a feature attention map. This shared layer consists of a multilayer perceptron and a hidden layer. To reduce the parameter size, the hidden activation size is set to... Where r is the reduction ratio, then the feature vectors are summed and merged, and the channel attention is calculated as follows:
[0043] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))
[0044] Where σ represents the Sigmoid function, AvgPool and MaxPool represent average pooling and max pooling operations respectively, and F is the input feature map;
[0045] Spatial attention module:
[0046] First, some channels are reshaped to the batch dimension and the channel dimension is grouped into multiple sub-features, so that the spatial semantic features are well distributed in each feature group. Then, a parallel sub-network combining 1x1 and 3x3 convolutional kernels is used to perform feature grouping and multi-scale structure processing in different spatial dimensions. Finally, the output feature maps of the two parallel sub-networks are fused.
[0047] The overall MCA attention calculation process is as follows:
[0048]
[0049]
[0050] in M represents element-wise multiplication. c and M s The parts represent channel attention computation and spatial attention computation, respectively. F, F′, and F″ represent the feature map extracted by ResNet-101, the feature map after channel attention computation, and the final output feature map, respectively. The access order of the MCA module in the improved DETR network is: ResNet-101-MCA-Encoder-Decoder-FFN.
[0051] Furthermore, the specific steps for the S4 defect detection are as follows:
[0052] S41, Defect dataset input:
[0053] For the enhanced tunnel defect image in S2, it is input into the target detection model to obtain the corresponding detection results. The detection results are N prediction boxes and their corresponding categories, where N is the fixed number of prediction boxes output each time in the model, N=100.
[0054] S42, Comprehensive Analysis:
[0055] For the multiple prediction results obtained in S41, the category with the highest probability is output by the test program in the model as the final recognition result, thus realizing the intelligent detection of cavities and non-compact defects in tunnel ground penetrating radar images.
[0056] Beneficial effects:
[0057] 1. Based on the end-to-end Transformer neural network, the global receptive field can effectively capture important information and details in the image, improving the recognition accuracy of cavities and imperfections;
[0058] 2. The use of parameter fine-tuning transfer learning solves the problem of insufficient data in the past and helps to build a more effective tunnel component-defect dataset, thereby helping the network to converge quickly;
[0059] 3. By combining traditional data augmentation methods to perform operations such as random mirror flipping, Gaussian blur, brightness transformation, and scale transformation on images, the model can better adapt to different scenarios and environments during testing, improving the model's generalization ability and increasing sample diversity.
[0060] 4. The DETR network framework was improved by embedding a novel and efficient multi-scale attention module—MCA module—which improved the training efficiency and performance of the model.
[0061] 5. It can realize the automatic identification of internal defects in tunnel lining, improve the efficiency and accuracy of tunnel safety early warning and prevention, and help reduce casualties and major economic losses;
[0062] In summary, the intelligent detection of internal cavities and non-compactment defects in tunnel lining based on GPR data has high practical value and social benefits. Attached Figure Description
[0063] Figure 1 It is the overall intelligent detection process for internal defects in tunnel lining;
[0064] Figure 2 It is based on the improvement and training process of the DETR network;
[0065] Figure 3 This describes the implementation process of spatial attention in the MCA module;
[0066] Figure 4 These are the detection results of internal defects in the tunnel lining. Detailed Implementation
[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0068] The overall process for intelligent detection of internal defects in tunnel lining according to the present invention is as follows: Figure 1 As shown.
[0069] This invention is achieved through the following technical solution:
[0070] S1, Ground Penetrating Radar Image Acquisition Inside the Tunnel: Acquire representative images of various elements inside the tunnel lining, including steel bars, steel arches, non-compact defects, and cavities, as pre-training data.
[0071] S2, Image Preprocessing and Construction of Tunnel Components—Defect Dataset: For the images acquired in S1, the image quality is checked and images with poor quality or damage are filtered out or repaired. The image size and resolution are adjusted to ensure the consistency and validity of the dataset. Then, data augmentation is performed. Traditional data augmentation methods are used to perform enhancement operations on the training set data, such as random image mirroring, Gaussian blur, brightness transformation, scale transformation, and adding noise, to further improve the generalization ability of the model and increase the diversity of samples. A tunnel component—defect dataset is established for model training. The images in the dataset are divided into training set, validation set, and test set in a ratio of 7:2:1.
[0072] The data augmentation described in S21 uses the following transformation methods on the tunnel component-defect dataset to further improve the model's generalization ability and increase sample diversity:
[0073] Mirror flip: Mirrors the image horizontally;
[0074] Gaussian blur: Makes the image more blurred by using a Gaussian blur filter of a specified size;
[0075] Brightness Transformation: Increases or decreases the brightness of the image;
[0076] Add noise: Add an appropriate amount of salt and pepper noise or Gaussian noise to the image;
[0077] Scale transformation: Preserves the feature regions of the image and performs cropping and enlargement;
[0078] For each transformation method, different parameters and ranges are used to generate different transformed images: for brightness transformation, two transformed images are generated using two ratios of 0.8 and 1.2; for adding noise, two transformed images are generated using two noise ratios of 0.001 and 0.002; for scale transformation, three transformed images are generated using three methods: image center, image edge, and random region as the transformation center; finally, multiple transformed images are obtained.
[0079] S3, improves the DETR-based deep learning network framework and uses the tunnel component dataset for transfer learning training to obtain a pre-trained model. The improved DETR network is shown below. Figure 2 .
[0080] S31, Improved DETR-based Deep Learning Network Framework: First, the original DETR network framework is built. The original DETR network's image processing mainly consists of three parts: first, the image passes through a CNN backbone for extracting compact feature representations; then, it is input into the Transformer encoder and decoder; finally, the decoder output passes through a simple feedforward network (FFN) responsible for the final detection and prediction. For the improved DETR network, ResNet-101 is used as the backbone network to extract high-level semantic features of the image and output feature maps at different scales. Two sequential attention modules are added to the output feature maps of the backbone network: channel attention and spatial attention. The output feature map of the channel attention module serves as the input of the spatial attention module. After the features pass through the two modules sequentially, the final attention feature map is obtained. Subsequent processing is consistent with the original DETR model: the obtained attention features are added to the positional encoding and then input into the Transformer architecture. The Decoder layer obtains N query vectors, which, after passing through the FFN layer, yield N prediction boxes and their corresponding categories.
[0081] Channel attention module: First, it aggregates the spatial information of the feature map through average pooling and max pooling operations to generate two different spatial context descriptors. and These represent the average pooling feature and the max pooling feature, respectively. These two features are then fed into a shared layer to generate our feature attention map. This shared layer consists of a multilayer perceptron (MLP) and a hidden layer. To reduce the parameter size, the hidden activation size is set to... Where r is the reduction ratio, the feature vectors are then summed and merged. In short, channel attention is calculated as follows:
[0082] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))
[0083] Where σ represents the Sigmoid function, AvgPool and MaxPool represent average pooling and max pooling operations respectively, and F is the input feature map.
[0084] Spatial Attention Module: First, some channels are reshaped to the batch dimension and grouped into multiple sub-features, ensuring a good distribution of spatial semantic features within each feature group. Then, parallel sub-networks combining 1x1 and 3x3 convolutional kernels perform feature grouping and multi-scale structural processing across different spatial dimensions. Finally, the output feature maps of the two parallel sub-networks are fused. See [link to implementation details]. Figure 3 .
[0085] The overall MCA attention calculation process is as follows:
[0086]
[0087]
[0088] in M represents element-wise multiplication. c and M s The terms "F" and "F'" represent channel attention computation and spatial attention computation, respectively. F, F', and F'" represent the feature map extracted by ResNet-101, the feature map after channel attention computation, and the final output feature map, respectively. The MCA module is accessed in the improved DETR network in the following order: ResNet-101-MCA-Encoder-Decoder-FFN.
[0089] S32, Transfer Learning Training Using a Tunnel Component Dataset: Addressing the issues of scarce internal defect data in tunnel linings under small sample conditions, slow convergence speed, and susceptibility to overfitting, transfer learning is employed, using the internal component dataset of the tunnel lining as an auxiliary domain for knowledge transfer. First, a coarse-grained model is obtained using the COCO2017 dataset, and its parameters are transferred to the tunnel component dataset to obtain a fine-grained model. Finally, its model parameters are transferred to the tunnel defect identification task, improving both the model's convergence speed and robustness.
[0090] S33. Adjust training parameters to determine the optimal model: Training parameters include learning rate, batch size, optimizer, and number of training rounds, which affect the convergence speed and stability of the model, and finally determine the optimal training model.
[0091] S4, Defect Detection: Input the images in the tunnel defect dataset into the segmentation model to obtain the corresponding detection results. Then, perform comprehensive analysis on these segmentation results to obtain the final result, thereby realizing intelligent detection of internal defects in the tunnel lining.
[0092] S41, Defect Dataset Input: The enhanced tunnel defect image in S2 is input into the target detection model to obtain the corresponding detection results. The detection results are N prediction boxes and their corresponding categories, where N is the fixed number of prediction boxes output in each output (N=100) set in the model.
[0093] S42, Comprehensive Analysis: For the multiple prediction results obtained in S41, the class with the highest probability is output by the test program in the model as the final recognition result, ultimately achieving intelligent detection of cavities and non-compact defects in tunnel ground-penetrating radar images. The detection results are shown in [link to relevant documentation]. Figure 4 .
[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.
Claims
1. A method for detecting internal cavities and non-compact defects in tunnel lining based on GPR data, comprising the following steps, characterized in that: S1, Ground-penetrating radar image acquisition inside the tunnel: Images representing a wide variety of elements inside the tunnel lining, including steel bars, steel arches, non-compact defects, and cavities, were collected as pre-training data. S2, Image Preprocessing and Construction of Tunnel Components—Defect Dataset: For images acquired by S1, check the image quality and filter or repair images with poor quality or damage, adjust the image size and resolution, and ensure the consistency and validity of the dataset. Next, data augmentation was performed. Traditional data augmentation methods were used to perform random image mirroring, Gaussian blur, brightness transformation, scale transformation, and noise enhancement on the training set data to further improve the generalization ability of the model and increase the diversity of the samples. A tunnel component-defect dataset was established for model training. The images in the dataset were divided into training set, validation set, and test set in a ratio of 7:2:
1. S3, improve the DETR-based deep learning network framework and use the tunnel component dataset for transfer learning training to obtain a pre-trained model; Step S3, which improves the DETR-based deep learning network framework and uses the tunnel component dataset for transfer learning training, includes the following specific steps: S31, Improved DETR-based deep learning network framework: First, the original DETR network framework is built. The original DETR network's image processing consists of three parts: first, it passes through the CNN backbone for extracting compact feature representations; then, it is input into the Transformer encoder and decoder; finally, the decoder output passes through a simple feedforward network responsible for the final detection and prediction. For the improved DETR network, ResNet-101 is used as the backbone network to extract high-level semantic features of the image and output feature maps at different scales. Two sequential attention modules are added to the output feature map of the backbone network: channel attention and spatial attention. The output feature map of the channel attention module serves as the input of the spatial attention module. After the features pass through the two modules in sequence, the final attention feature map is obtained. The subsequent processing is consistent with the original DETR model. The obtained attention features are added to the position encoding and then input into the Transformer architecture. The Decoder layer obtains N query vectors, which are then passed through the FFN layer to obtain N prediction boxes and their corresponding categories. The steps for adding two sequential attention modules in step S31 are as follows: Channel attention module: First, the spatial information of the feature map is aggregated through average pooling and max pooling operations to generate two different spatial context descriptors. and Let represent the average pooling feature and the max pooling feature, respectively. Then, these two features are fed into a shared layer to generate a feature attention map. This shared layer consists of a multilayer perceptron and a hidden layer. To reduce the parameter size, the hidden activation size is set to... Where r is the reduction ratio, then the feature vectors are summed and merged. The channel attention is calculated as follows: ; in This represents the Sigmoid function. and These represent average pooling and max pooling operations, respectively. Input feature map; Spatial attention module: First, some channels are reshaped to the batch dimension and the channel dimension is grouped into multiple sub-features, so that the spatial semantic features are well distributed in each feature group. Then, a parallel sub-network combining 1x1 and 3x3 convolutional kernels is used to perform feature grouping and multi-scale structure processing in different spatial dimensions. Finally, the output feature maps of the two parallel sub-networks are fused. The overall MCA attention calculation process is as follows: ; ; in This indicates element-wise multiplication. and The parts represent channel attention computation and spatial attention computation, respectively. , and These represent the feature map extracted by ResNet-101, the feature map calculated after channel attention, and the final output feature map, respectively. The MCA module is accessed in the improved DETR network in the following order: ResNet-101-MCA-Encoder-Decoder-FFN. S32, using the tunnel component dataset for transfer learning training: By utilizing transfer learning, the dataset of internal components of tunnel lining is used as an auxiliary domain for knowledge transfer. First, a coarse-grained model is obtained using the COCO2017 dataset, and the model parameters are transferred to the tunnel component dataset to obtain a fine-grained model. Finally, the model parameters are transferred to the tunnel defect identification task, which improves the model convergence speed and enhances its robustness. S33, Adjust training parameters to determine the optimal model: Training parameters include learning rate, batch size, optimizer, and number of training epochs, which affect the model's convergence speed and stability, ultimately determining the optimal training model. S4, Defect Detection: Images from the tunnel defect dataset are input into a segmentation model to obtain corresponding detection results. Then, these segmentation results are comprehensively analyzed to obtain the final result, thus achieving intelligent detection of internal defects in the tunnel lining.
2. The method for detecting internal cavities and non-compactness defects in tunnel lining based on GPR data according to claim 1, characterized in that: The S2 data enhancement process involves the following steps: The data augmentation described in S2 uses the following transformation methods on the tunnel component-defect dataset to further improve the model's generalization ability and increase sample diversity: Mirror flip: Mirrors the image horizontally; Gaussian blur: Makes the image more blurred by using a Gaussian blur filter of a specified size; Brightness Transformation: Increases or decreases the brightness of the image; Add noise: Add an appropriate amount of salt and pepper noise or Gaussian noise to the image; Scale transformation: Preserves the feature regions of the image and performs cropping and enlargement; For each transformation method, different parameters and ranges are used to generate different transformed images: for brightness transformation, two transformed images are generated using two ratios of 0.8 and 1.2; for adding noise, two transformed images are generated using two noise ratios of 0.001 and 0.
002. For scaling transformation, three transformed images are generated using three methods: the image center, the image edge, and a random region as the transformation center; finally, multiple transformed images are obtained.
3. The method for detecting internal cavities and non-compactness defects in tunnel lining based on GPR data according to claim 1, characterized in that: The specific steps for S4 defect detection are as follows: S41, Defect dataset input: For the enhanced tunnel defect image in S2, it is input into the target detection model to obtain the corresponding detection results. The detection results are N prediction boxes and their corresponding categories, where N is the fixed number of prediction boxes output each time in the model, N=100. S42, Comprehensive Analysis: For the multiple prediction results obtained in S41, the category with the highest probability is output by the test program in the model as the final recognition result, thus realizing the intelligent detection of cavities and non-compact defects in tunnel ground penetrating radar images.
Citation Information
Patent Citations
Method and apparatus for classifying images of internal defects in concealed tunnel structures
CN115222999B
Method, device, equipment and medium for improving geological radar detection precision of tunnel lining defects
CN117538863A
