Tunnel lining disease detection method and device and processing equipment

Abstract

The invention provides a tunnel lining disease detection method and device and processing equipment, designs a specific tunnel lining disease detection network architecture, on the basis of introducing DetCo self-supervised training, on one hand, effective fusion between a self-supervised backbone network and other structures in the network can be ensured, and on the other hand, the detection efficiency is improved; and on the other hand, it is guaranteed that parameters in the self-supervision pre-training stage can effectively act on the network, so that the sensitivity and accuracy of the network on tunnel lining disease detection are improved, high-quality data support is provided for tunnel structure safety monitoring work, and tunnel operation safety and service life are guaranteed.

Description

Technical Field

[0001] This application relates to the field of tunnel structure safety monitoring technology, specifically to a method, device, and treatment equipment for detecting tunnel lining defects. Background Technology

[0002] Tunnels, as an important component of modern transportation infrastructure, play a crucial role in transportation, urban construction, and underground engineering. However, due to long-term exposure to loads, changes in geological conditions, and environmental factors, tunnel lining structures are prone to defects such as cracks, voids, and delamination. If these defects are not detected and repaired in a timely manner, they may lead to serious safety hazards such as water seepage and collapse, threatening the safety and service life of the tunnel.

[0003] In subsequent maintenance work, different types of defects require targeted repair measures. By accurately identifying the types of defects and implementing corresponding repair methods, not only can maintenance efficiency be improved, but the service life of road and tunnel structures can also be extended.

[0004] However, accurately identifying the type of damage faces many difficulties. Due to the complex structure of tunnels and roads and the variety of damage forms, accurate judgment of damage often relies on experienced professionals, which is inefficient and cannot ensure timeliness, thus placing higher demands on the effectiveness of subsequent repair measures.

[0005] Traditional ground-penetrating radar (GPR) target detection and classification methods primarily rely on signal processing techniques and classical machine learning algorithms. While these methods achieve high classification and detection accuracy on small-scale, simple background simulation datasets, their performance is limited in real-world complex backgrounds or when targets have irregular shapes. They are particularly limited in accurately identifying feature signals with different morphologies, locations, and types of defects in GPR images. Furthermore, these methods heavily depend on human experience in feature engineering and parameter settings, requiring complex feature extraction methods, resulting in low efficiency and making them unsuitable for tunnel defect detection projects that demand high levels of automation and real-time detection.

[0006] In recent years, deep learning has made significant progress in the application of ground-penetrating radar (GPR) image target detection. Its stronger feature learning and generalization capabilities better address problems that traditional methods struggle with due to their limited generalization abilities. Supervised deep learning methods utilize large amounts of labeled data to train models, enabling them to automatically learn key features from numerous complex GPR images, achieving efficient and accurate target detection. Different deep learning models have been applied to target object detection scenarios in GPR images. The widespread application of supervised learning in GPR image recognition tasks demonstrates that the robustness of deep learning image algorithms can be transferred to target recognition tasks in GPR images.

[0007] While supervised deep learning methods have shown great potential in ground-penetrating radar detection, their reliance on large amounts of high-quality labeled data has also brought challenges to their application. In recent years, self-supervised learning, which does not rely on large amounts of labeled data, has become a hot research direction. Self-supervised learning first showed relatively robust applications in natural language processing, but unlike words, which are concise and have strong semantic information, visual signals in computer vision exist in a continuous, high-dimensional space and require more memory resources during training. Unlike natural language processing, it cannot use a single memory bank to store all negative samples. Therefore, the development of self-supervised learning in computer vision has evolved from image classification to object detection tasks.

[0008] With the rise of contrastive learning, self-supervised methods have made groundbreaking progress in the application of image processing. Initially, contrastive learning mainly focused on learning global representations, maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, enabling the model to learn more robust representations. While these methods performed well in classification tasks, their feature extraction methods primarily focused on global information and did not explicitly consider local region features in object detection tasks.

[0009] Due to the unique characteristics and industry nature of ground-penetrating radar (GPR) images, current research in this field lags significantly behind that in natural image processing. However, the technological potential has begun to emerge. Although existing work has preliminarily verified the feasibility of self-supervised learning in GPR image processing, significant challenges remain for the specific task of detecting tunnel lining defects.

[0010] 1) Locality and weak significance of disease characteristics: Cracks, voids and other defects inside the lining often appear as weak local reflection signals in B-Scan images, and traditional global contrastive learning is easily affected by background clutter.

[0011] 2) Spatiotemporal signal coupling characteristics: Ground penetrating radar images have dual physical attributes of temporal fluctuations and spatial distribution, requiring simultaneous modeling of wave propagation laws and spatial topological correlations of the disease. Summary of the Invention

[0012] This application provides a method, apparatus, and processing equipment for detecting tunnel lining defects. It designs a specific network architecture for detecting tunnel lining defects. Based on the introduction of DetCo self-supervised training, it can ensure the effective integration between the self-supervised backbone network and other structures in the network. On the other hand, it can ensure that the parameters in the self-supervised pre-training stage can effectively affect the network, thereby improving the network's sensitivity and accuracy in detecting tunnel lining defects. This provides high-quality data support for tunnel structural safety monitoring and ensures the safety and service life of tunnel operation.

[0013] Firstly, this application provides a method for detecting tunnel lining defects, the method comprising:

[0014] Acquire ground-penetrating radar images of the target obtained from the tunnel lining;

[0015] Ground-penetrating radar (GPR) images are input into a pre-configured tunnel lining defect detection network. This network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features at different scales from the input GPR images, including first-scale, second-scale, third-scale, and fourth-scale features. The feature pyramid network fuses features at different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has undergone self-supervised pre-training within the DetCo self-supervised framework.

[0016] Extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

[0017] Secondly, this application provides a tunnel lining defect detection device, the device comprising:

[0018] The acquisition unit is used to acquire ground-penetrating radar images of the target collected by the tunnel lining.

[0019] The detection unit is used to input ground-penetrating radar images into a pre-configured tunnel lining defect detection network. The tunnel lining defect detection network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features of different scales from the input ground-penetrating radar images. These features include first-scale features, second-scale features, third-scale features, and fourth-scale features. The feature pyramid network fuses features of different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has undergone self-supervised pre-training under the DetCo self-supervised framework.

[0020] The extraction unit is used to extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

[0021] Thirdly, this application provides a processing device, including a processor and a memory, wherein a computer program is stored in the memory, and when the processor invokes the computer program in the memory, it executes the method provided by the first aspect of this application or any possible implementation of the first aspect of this application.

[0022] Fourthly, this application provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to perform the method provided in the first aspect of this application or any possible implementation thereof.

[0023] From the above, it can be concluded that this application has the following beneficial effects:

[0024] For the detection of tunnel lining defects based on ground-penetrating radar images, this application designs a specific network architecture for tunnel lining defect detection. By introducing DetCo self-supervised training, it can ensure the effective integration between the self-supervised backbone network and other structures in the network, and ensure that the parameters in the self-supervised pre-training stage can be effectively applied to the network. This improves the network's sensitivity and accuracy in detecting tunnel lining defects, thereby providing high-quality data support for tunnel structural safety monitoring and ensuring the safety and service life of tunnel operation. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic flowchart of a tunnel lining defect detection method according to this application;

[0027] Figure 2 This is a schematic diagram of a tunnel lining defect detection network according to this application;

[0028] Figure 3 This is a schematic diagram illustrating the working logic of the self-supervised backbone network of this application;

[0029] Figure 4 This is a schematic diagram of an example encoder of this application;

[0030] Figure 5 This is a schematic diagram of a tunnel lining defect detection device according to this application.

[0031] Figure 6 This is a schematic diagram of one type of processing equipment used in this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. The naming or numbering of steps appearing in this application does not imply that the steps in the method flow must be performed in the chronological / logical order indicated by the naming or numbering. The execution order of named or numbered process steps can be changed according to the desired technical purpose, as long as the same or similar technical effect is achieved.

[0034] The module division described in this application is a logical division. In practical applications, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection between modules shown or discussed may be through some interfaces, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules. Some or all of the modules may be selected to achieve the purpose of the solution in this application according to actual needs.

[0035] Before introducing the tunnel lining defect detection method provided in this application, the background content of this application will be introduced first.

[0036] The tunnel lining defect detection method, device, and computer-readable storage medium provided in this application can be applied to processing equipment. A specific tunnel lining defect detection network architecture is designed. Based on the introduction of DetCo self-supervised training, it can ensure the effective integration between the self-supervised backbone network and other structures in the network. On the other hand, it can ensure that the parameters in the self-supervised pre-training stage can be effectively applied to the network, thereby improving the network's sensitivity and accuracy in detecting tunnel lining defects. This provides high-quality data support for tunnel structural safety monitoring and ensures the safety and service life of tunnel operation.

[0037] The tunnel lining defect detection method mentioned in this application can be implemented by a tunnel lining defect detection device, or by different types of processing devices such as servers, physical hosts, or user equipment (UE) that integrate the tunnel lining defect detection device. The tunnel lining defect detection device can be implemented in hardware or software. The UE can be a terminal device such as a smartphone, tablet, laptop, desktop computer, or personal digital assistant (PDA). The processing devices can be configured in a device cluster.

[0038] It is understandable that in specific applications, the solution of this application is usually based on the data processing carried out on the basis of the acquired or ready-made ground penetrating radar images. Therefore, the processing equipment that implements the tunnel lining defect detection method of this application or is equipped with the application service corresponding to the ground penetrating radar image method of this application usually only needs to meet the required data processing capabilities. The specific equipment type and equipment deployment form are quite flexible and can be flexibly configured according to the actual situation.

[0039] If the acquisition of ground-penetrating radar images is also involved, the processing equipment needs to be configured with the necessary hardware and software for ground-penetrating radar image acquisition.

[0040] Similarly, if there is a need to display the results, the processing device needs to be equipped with a corresponding display screen (including a touch screen), or the results can be displayed through an external display device or other devices with a display screen.

[0041] The tunnel lining defect detection method provided in this application will now be introduced.

[0042] First, refer to Figure 1 , Figure 1 This paper illustrates a flowchart of a tunnel lining defect detection method according to the present application. The tunnel lining defect detection method provided by the present application may specifically include the following steps S101 to S10:

[0043] Step S101: Acquire the ground-penetrating radar image of the target obtained from the tunnel lining;

[0044] Understandably, in the actual use of the tunnel lining defect detection network configured for the tunnel lining defect detection needs of this application, it is necessary to acquire the ground-penetrating radar image (also called ground-penetrating radar B-Scan image) collected from the tunnel lining of the target tunnel that needs to be detected. For ease of explanation, the image that needs to be processed is referred to as the target ground-penetrating radar image, so as to distinguish it from the sample ground-penetrating radar images designed during the network training process.

[0045] This typically involves the extraction and processing of existing images, which can be done through manual input, local retrieval, or remote retrieval. Alternatively, it may involve the acquisition and processing of real-time images.

[0046] Since ground-penetrating radar (GPR) images and how to acquire them are within the scope of existing technology, they will not be discussed in detail here.

[0047] Step S102: The ground-penetrating radar image is input into a pre-configured tunnel lining defect detection network. The tunnel lining defect detection network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features of different scales from the ground-penetrating radar image input to the network. The features of different scales include first-scale features, second-scale features, third-scale features, and fourth-scale features. The feature pyramid network fuses the features of different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has been pre-trained in the DetCo self-supervised framework.

[0048] It is easy to see that the tunnel lining defect detection network or tunnel lining defect detection model specially designed in this application is configured on the basis of DetCo self-supervised training. The DetCo self-supervised training specifically corresponds to the self-supervised backbone network in the network. This self-supervised backbone network has been pre-trained in the DetCo self-supervised framework, and then the overall network is trained after the pre-training.

[0049] Specifically, for the self-supervised backbone network, its self-supervised pre-training is performed using unlabeled ground-penetrating radar images, while the overall tunnel lining defect detection network is trained using labeled ground-penetrating radar images.

[0050] During the main network operation, the self-supervised backbone network extracts features of different scales from ground-penetrating radar images through its network structure design. These features include four main types: first-scale features, second-scale features, third-scale features, and fourth-scale features. It is worth noting that scale can also be referred to as hierarchy. The subsequent Feature Pyramid Network (FPN) is configured to correspond to the output of the self-supervised backbone network, thereby fusing the features of different scales extracted by the self-supervised backbone network. The Region Proposal Network (RPN) then generates candidate regions based on the fused features obtained from the Feature Pyramid Network. Finally, a customized detection head determines the type and location of tunnel lining defects.

[0051] Furthermore, you can refer to Figure 2 The schematic diagram shown here is a structural representation of the tunnel lining defect detection network of this application, for a more illustrative purpose.

[0052] It is understandable that the tunnel lining defect detection network built on the Faster R-CNN framework in this application is designed to meet the physical characteristics of ground penetrating radar images and defect detection requirements. Through multi-scale feature fusion and task-driven module optimization, it achieves high-precision positioning and classification.

[0053] Step S103: Extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

[0054] After the tunnel lining defect detection network completes the corresponding tunnel lining defect detection processing on the current ground penetrating radar image, it is easy to understand that the tunnel lining defect detection results output by the tunnel lining defect detection network can be extracted to complete a tunnel lining defect detection task. This corresponds to the situation in practical applications where the solution processing is usually initiated in the form of a task.

[0055] The tunnel lining defect detection task can be initiated manually or by the system according to the corresponding task initiation strategy, which can be configured according to the actual situation.

[0056] At the same time, ground-penetrating radar images of targets can be directly carried in mission information or indirectly indicate the location stored. Of course, in some cases, there is also the possibility of real-time image acquisition.

[0057] Once the tunnel lining defects detection results are obtained, further data processing can be performed based on the corresponding data application needs in tunnel management.

[0058] For example, it could involve local storage, off-site storage, result display, output of a notification indicating completion of the test, result forwarding, alerts, or further data analysis and processing.

[0059] Obviously, the specific data application content that can be involved can be flexibly adjusted according to the pre-configured and real-time configured single data application strategies.

[0060] As can be seen from the above embodiments, for the target of tunnel lining defect detection based on ground penetrating radar images, this application designs a specific tunnel lining defect detection network architecture. On the basis of introducing DetCo self-supervised training, on the one hand, it can ensure the effective integration between the self-supervised backbone network and other structures in the network, and on the other hand, it can ensure that the parameters in the self-supervised pre-training stage can be effectively applied to the network, thereby improving the network's sensitivity and accuracy for tunnel lining defect detection, and thus providing high-quality data support for tunnel structure safety monitoring, ensuring the safety and service life of tunnel operation.

[0061] Next, we will further explain the detailed optimization settings of the tunnel lining defect detection network in this application.

[0062] (1) Self-supervised backbone network

[0063] Due to the shortcomings of existing technologies in supervised learning for tunnel lining defect detection, which relies heavily on high-quality labels with extensive signal processing experience, and the fact that existing self-supervised methods still suffer from insufficient detection accuracy, this application introduces the DetCo self-supervised framework to address these issues. This involves self-supervised pre-training of the self-supervised backbone network, thereby avoiding the limitations of manually designed decomposition models. Through data-driven contrastive learning, defect discrimination features are automatically mined. Since the self-supervised pre-training uses unlabeled images, it fully utilizes these images, reducing reliance on labeled data and improving the accuracy of detection results even when the number of labels is insufficient.

[0064] correspond Figure 2 The self-supervised pre-trained DetCo backbone network is used as the feature extraction backbone network. Stage 1 (i.e., the first feature extraction layer) and Stage 2 (i.e., the second feature extraction layer) capture shallow local features, such as the pulse reflection waveform at the crack edge and the weak signal disturbance in the void region. Stage 3 (i.e., the third feature extraction layer) and Stage 4 (i.e., the fourth feature extraction layer) encode deep global semantics, representing the difference between the damage caused by dielectric constant jump and the electromagnetic properties of the background medium. By preserving the complete feature hierarchy of GPR-DetCo, the network can make full use of the cross-scale spatiotemporal patterns learned by self-supervised learning, which significantly improves the robustness of the network.

[0065] Furthermore, in one exemplary embodiment, reference is made to... Figure 3 The diagram shown illustrates the working logic of a self-supervised backbone network of this application. The self-supervised backbone network can adopt a dual-branch encoder architecture (each branch includes an online encoder and a momentum encoder), specifically including a first online encoder, a first momentum encoder, a second online encoder, and a second momentum encoder.

[0066] Correspondingly, a self-supervised backbone network can include the following tasks:

[0067] 1. Perform global image enhancement on the ground-penetrating radar image input from the network to obtain the first view image (corresponding to I in the figure). q ) and second view image (corresponding to I in the figure) k The third-view image (corresponding to P in the figure) is obtained by performing local image patch enhancement on the ground-penetrating radar image input from the network. q ) and the fourth view image (corresponding to P in the figure) k );

[0068] Global image enhancement can be specifically random image enhancement, while local image patch enhancement involves obtaining local image patches from the original image and then performing random image enhancement.

[0069] 2. The first online encoder extracts a first multi-scale feature map from the first view image, the first momentum encoder extracts a second multi-scale feature map from the second view image, the second online encoder extracts a third multi-scale feature map from the third view image, and the second momentum encoder extracts a fourth multi-scale feature map from the fourth view image, wherein each of the four multi-scale feature maps includes a corresponding first feature map, second feature map, third feature map, and fourth feature map.

[0070] It is important to note that the four multi-scale feature maps mentioned here are not the first-scale, second-scale, third-scale, and fourth-scale features (i.e., features of different scales) extracted from the ground-penetrating radar images input to the network by the self-supervised backbone network mentioned earlier.

[0071] The first, second, third, and fourth feature maps in each multi-scale feature map are different in scale, so the four feature maps of different scales make up the multi-scale feature map.

[0072] As an example, the first feature map can be denoted as MFM1, with a resolution of 160×160; the second feature map can be denoted as MFM2, with a resolution of 80×80; the third feature map can be denoted as MFM3, with a resolution of 40×40; and the fourth feature map can be denoted as MFM4, with a resolution of 20×20.

[0073] These multi-scale feature maps can capture hierarchical information from local waveform details to global contextual semantics, providing a foundation for subsequent comparative learning.

[0074] 3. For the first, second, third, and fourth feature maps of the four multi-scale feature maps, adaptive average pooling and nonlinear projection are performed respectively to obtain the first-scale feature, the second-scale feature, the third-scale feature, and the fourth-scale feature in sequence, forming features of different scales.

[0075] Taking the first multi-scale feature map extracted from the first view image by the first online encoder as an example, adaptive average pooling and nonlinear projection are performed on the first feature map, second feature map, third feature map and fourth feature map included in the first multi-scale feature map to obtain the first scale feature. By analogy, the first scale feature, second scale feature, third scale feature and fourth scale feature corresponding to the four encoders are obtained. These are the features of different scales extracted from the ground penetrating radar image input to the network by the self-supervised backbone network mentioned above.

[0076] Furthermore, focusing on the specific encoder structure of the four encoders, for ease of explanation, the specific encoder currently being addressed among the four encoders will be referred to as the current encoder, and the explanation will be carried out from the perspective of the current encoder.

[0077] In one exemplary embodiment, the current encoder includes an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The current encoder is a first online encoder, a first momentum encoder, a second online encoder, or a second momentum encoder.

[0078] Understandably, under this encoder structure, this application can also be configured to: freeze the parameters of the first feature extraction layer, the second feature extraction layer, and the third feature extraction layer after training, while unfreezing the parameters of the fourth feature extraction layer.

[0079] By freezing the parameters of the first, second, and third feature extraction layers, the fourth feature extraction layer participates in the subsequent training process. In terms of details, this can further ensure the effective integration of the self-supervised backbone network with other modules in the disease detection network, and also ensure that the parameters in the self-supervised pre-training stage can effectively affect the overall network, thereby improving the accuracy of detection.

[0080] On the other hand, under this encoder structure, the processing of the first feature map, second feature map, third feature map, and fourth feature map can specifically include:

[0081] The input layer preprocesses the view image input to the encoder to obtain a preprocessed feature map. The view image input to the encoder is either the first view image, the second view image, the third view image, or the fourth view image (corresponding to the case where the current encoder is one of four encoders).

[0082] The first feature extraction layer extracts the first feature map from the preprocessed feature map;

[0083] The second feature extraction layer extracts the second feature map from the first feature map;

[0084] The third feature extraction layer extracts the third feature map from the second feature map;

[0085] The fourth feature extraction layer extracts the fourth feature map from the third feature map.

[0086] To make it easier to understand, you can also combine Figure 4 The schematic diagram of an example encoder of this application is shown for a more intuitive understanding.

[0087] As can be seen, for the input layer, the size of the convolution kernel can be 7×7, and the number of output channels can be 64.

[0088] The first feature extraction layer, the second feature extraction layer, the third feature extraction layer, and the fourth feature extraction layer all include multiple residual blocks.

[0089] Furthermore, it can be noted that, as shown in part (a) of the figure, this application specifically uses a ResNet50 network to build the backbone network. For the current encoder, its input layer adjusts the original three-channel convolutional layer of ResNet50 to a single-channel input to adapt to the grayscale characteristics of ground-penetrating radar images. The improved input layer convolutional kernel size is 7×7, and the number of output channels is 64. While preserving spatial downsampling capability, it avoids redundant information introduced by multiple channels and is equipped with BN layers, ReLU functions, and MaxPool layers. The backbone network consists of four stages (Stage 1-Stage 4) of ResNet50, each stage consisting of multiple residual blocks, according to... The number of features is 3, 4, 6, and 3, and the outputs are a first feature map (MFM1) with a resolution of 160×160, a second feature map (MFM2) with a resolution of 80×80, a third feature map (MFM3) with a resolution of 40×40, and a fourth feature map (MFM4) with a resolution of 20×20 in sequence. The residual block is shown in part (b) of the figure. In the Bottleneck structure, it includes a cascaded 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer. The input of the first 1×1 convolutional layer is also connected to the output of the second 1×1 convolutional layer (i.e., the two are added together) to form a residual connection, and then output through the ReLU activation function.

[0090] Furthermore, regarding parameter updates, as an exemplary embodiment, it also corresponds to... Figure 3 As shown, the weights of the first momentum encoder are synchronized with the parameters of the first online encoder, and the weights of the second momentum encoder are synchronized with the parameters of the second online encoder.

[0091] The weights of the online encoder (either the first or second online encoder) are updated in real time via backpropagation, while the weights of the momentum encoder (either the first or second momentum encoder) are slowly synchronized with the parameters of the corresponding online encoder using an exponential moving average (EMA) strategy. The update strategy is as follows:

[0092]

[0093] in, These are the parameters of the momentum encoder at time t. These are the parameters of the momentum encoder at time t-1. Here, represents the parameters of the online encoder at time t, and m is the momentum coefficient.

[0094] As an example, m can specifically be 0.999.

[0095] The multi-level contrastive learning head is DetCo's core innovative module. Addressing the sparsity and cross-scale characteristics of defects in ground-penetrating radar images, the network performs adaptive average pooling and nonlinear projection on the feature maps output from each of the four stages. Features from each level are mapped to a low-dimensional embedding space through two fully connected layers. A Spatial Attention Module (SAM) is introduced after the projection head in the third stage (Stage 3). This module dynamically focuses on key regions of the reflected signal through channel compression and spatial weight calculation.

[0096] Furthermore, it is understandable that the contrastive loss function used in self-supervised training employs an improved hierarchical weighted (Information Noise Contrastive Estimation, InfoNCE) loss. By constructing positive and negative sample pairs to optimize feature representation, for two randomly augmented views of the same input image, the online encoder and momentum encoder extract multi-scale features respectively and calculate the contrastive loss of each level. The negative sample queue dynamically stores historical momentum encoder features to ensure the diversity of negative samples.

[0097] Specifically, in one exemplary embodiment, for a self-supervised backbone network, the loss function used for self-supervised pre-training can be:

[0098]

[0099] in, For the global contrast loss at the i-th scale, For the local contrast loss at the i-th scale, Let ω be the global-local contrast loss at the i-th scale. i The preset weights are for the i-th scale.

[0100] The loss function consists of the global contrastive loss for each level (i.e., the i-th scale). Local contrast loss Global-local contrast loss The weighted sum is used to construct the feature set, with shallow features having smaller weights and deep features having larger weights.

[0101] Specifically, there are:

[0102]

[0103] The first view image is denoted as I. q The second view image is denoted as I.k For the i-th layer, From I q Query features generated by an online encoder Corresponding to positive sample features, and from I k , correspond Negative sample features, from the feature queue of the i-th layer The sampled data is obtained, where τ is a temperature coefficient that controls the smoothness of the distribution.

[0104]

[0105] in, Enhancing spatial fine-grained information and improving the feature representation capability of object detection tasks involves dividing the image into local blocks for enhancement. Similarly, for the i-th layer, local features are extracted. As a local query feature, Positive sample features corresponding to local blocks For the corresponding Negative sample features, from the local feature queue of the i-th layer Obtained by sampling from the middle;

[0106]

[0107] Among them, cross-modal comparison of global features and local block features is obtained. Force global features through cross-modal alignment With local features Alignment is used to enhance the discriminative power of local blocks for tunnel lining defects while maintaining consistency in the global representation.

[0108] Specifically, this embodiment uses the DetCo self-supervised network to implement the DetCo backbone network based on the improved ResNet50 architecture, and makes a customized design for the characteristics of ground penetrating radar images. In this way, the discriminative features of tunnel lining defects are mined through a multi-level contrastive learning strategy. At the same time, the multi-level contrast mechanism of DetCo is more suitable for processing the spatiotemporal coupling mode across scales in ground penetrating radar images, which is conducive to improving detection accuracy.

[0109] (2) Feature Pyramid Network

[0110] The Feature Pyramid Network merges the DetCo Stage1-Stage4 feature maps (i.e., first-scale features, second-scale features, third-scale features, and fourth-scale features (features at different scales)) through a top-down path and lateral connections to generate a feature pyramid with progressively decreasing resolution. High-resolution features focus on detailed textures, while low-resolution features enhance the global context. The combination of the two can effectively detect disease targets with significant size differences.

[0111] (3) Regional Proposal Network

[0112] The region proposal network receives the fused features output by the feature pyramid network and generates candidate regions through a sliding window, which can be configured as follows:

[0113] Each anchor point is associated with nine anchor frames, the size of which is customized according to the typical physical dimensions of tunnel defects: three basic sizes are set: 8×8, 16×16, and 32×32, corresponding to the typical signal range of small target cracks, medium target cavities and rebar removal, and large target voids; the aspect ratio is fixed at 1:1, 1:2, and 2:1 to cover the diversity of defect morphology such as horizontally extending cracks and circular cavities. The classification score and bounding box offset of the anchor frame are output through two parallel convolutional layers to screen out high-confidence candidate regions for subsequent detection head refinement.

[0114] (4) Detection head

[0115] In a specific design, as an exemplary embodiment, the detection head may specifically include an alignment unit (ROI Align) and a dual-branch prediction unit.

[0116] Improvements to physical perception ROI alignment based on ground-penetrating radar image characteristics: An electromagnetic wave attenuation compensation factor is introduced on the basis of standard ROI Align.

[0117] The alignment unit dynamically adjusts the feature weights based on the estimated burial depth of the candidate region, and uses these feature weights to extract features of the region of interest from the candidate region.

[0118] Among them, feature weights d represents the relative position of the ground-penetrating radar image input to the network corresponding to the current feature map, and α represents the medium attenuation coefficient.

[0119] The weighted feature map can suppress the intensity attenuation of deep signals and improve the sensitivity to deep, hidden diseases.

[0120] The dual-branch prediction unit predicts the probability of the type of tunnel lining defects based on the characteristics of the region of interest, and generates a prediction bounding box of the location of the defects, thus obtaining the type and location of the tunnel lining defects.

[0121] Furthermore, the dual-branch prediction unit can specifically include a classification branch and a regression branch. The classification branch predicts the probability of the type of defects in the tunnel lining, and the regression branch generates the predicted bounding box of the defect location, such as... Figure 2 As shown, the two can also share a fully connected layer.

[0122] The classification branch can use a two-layer fully connected network to output the disease category probability, with the activation function being the Softmax function, as shown in the following formula:

[0123]

[0124] The loss function used in the classification branch is cross-entropy loss, denoted as . The regression branch refines the prediction of the bounding box's center coordinates, width, and height through linear transformation. The loss function used in the regression branch employs the smooth L1 norm, denoted as... The total loss used by the detection head during the training of the tunnel lining defect detection network, which combines the loss functions of regression and classification branches, is as follows:

[0125]

[0126] Where C is the total number of categories, z i To predict the probability of the i-th type of disease, p c To predict the probability of the obtained category c defects, which correspond to cracks, voids, and delamination, as well as the rebar area, y c This is a real label;

[0127]

[0128] Where β is the preset value and x is the error between the predicted value and the actual value.

[0129] At the same time, the network configuration process may also involve a comprehensive evaluation of network performance.

[0130] To comprehensively and objectively evaluate the performance of the lining defect detection network in detecting lining defects, this application specifically adopts four indicators (which are existing indicators): precision, recall, F1 score, and mean precision (mAP). These are used to quantitatively analyze the network from the dimensions of classification confidence, missed detection risk, overall effectiveness, and location accuracy. The corresponding evaluation indicator formulas are as follows:

[0131]

[0132]

[0133] Where TP represents the network correctly identifies real disease areas, FP represents the network misidentifies normal areas as diseases, FN represents diseases that the network misses, and C represents the number of disease categories.

[0134] The above four indicators can comprehensively and quantitatively assess the reliability and robustness of the network, reflecting the network's ability to identify the location and extent of damage.

[0135] To further highlight the advantages of the proposed solution, a comparative test was conducted between this solution and existing technologies. Specifically, the following tests were performed:

[0136] The basic backbone network was pre-trained using 9404 unlabeled datasets under self-supervised conditions to obtain a self-supervised backbone network. Then, the basic detection network constructed from this self-supervised backbone network was further trained and fine-tuned using approximately 10% (1000) of the unlabeled dataset with labeled data to obtain the disease detection network described in this embodiment. Simultaneously, the same 1000 labeled datasets were used to train the existing DenseCL model, Faster-RCNN model, and YOLOv8 model. The metrics of each trained model are shown in Table 1 below.

[0137] Table 1 - Examples of Network Performance Comparison

[0138] The method used Accuracy Recall rate F1 score mAP This application proposal 0.8411 0.9806 0.9055 0.8224 DenseCL 0.7947 0.8544 0.8234 0.8074 Faster-RCNN 0.7822 0.8155 0.7985 0.7808 YOLOv8 0.5690 0.8449 0.6800 0.8032

[0139] To balance safety and cost, the network needs to have both extremely high recall and high precision, ensuring accurate detection of tunnel lining defects while tolerating a certain number of false positives that mistake ground-penetrating radar image background for defects. As shown in the table above, the Faster-RCNN model has high precision, but its supervised learning nature, relying on a large amount of labeled data, limits its capabilities. Overfitting during training can lead to poor robustness, resulting in a recall rate as low as 81.55%, indicating poor robustness. The YOLOv8 model has a relatively high recall rate of 84.49%, but its precision is low at only 56.9%, indicating that the model generates many false positives, increasing the manpower and material costs of tunnel inspection and making it less economical. DenseCL's various metrics are relatively average. The network in this application has an accuracy of 84.11%, a recall of 98.06%, an F1 score of 0.9055, and an mAP of 0.8224. Its performance in all indicators is better than that of the prior art. It can be considered that the solution in this application has high accuracy and good robustness. This application introduces self-supervised pre-training, thereby making full use of unlabeled data for training and reducing the dependence of the disease detection network on a large amount of labeled data.

[0140] This application also provides code for self-supervised pre-training of the basic backbone network and fine-tuning of the basic detection network. The code for self-supervised pre-training of the basic backbone network is shown in Table 2 below:

[0141] Table 2 - Examples of pre-trained code for self-supervised backbone networks

[0142]

[0143]

[0144] After pre-training, the self-supervised backbone network is extracted as the feature extraction network. The parameters of Stages 1-3 (i.e., the first, second, and third feature extraction layers) are frozen to preserve the self-supervised learning representations. The parameters of Stage 4 (i.e., the fourth feature extraction layer) are unfrozen to adapt to the detection task. During training, only 625 labeled images (approximately 1% of the dataset) are used. After temporal alignment and normalization preprocessing, these images are input into the detection network for end-to-end optimization. The code for fine-tuning the overall network is shown in Table 3 below.

[0145] Table 3 - Code Examples for Overall Network Fine-tuning

[0146]

[0147]

[0148] The above is an introduction to the tunnel lining defect detection method provided in this application. In order to facilitate better implementation of the tunnel lining defect detection method provided in this application, this application also provides a tunnel lining defect detection device from the perspective of functional modules.

[0149] See Figure 5 , Figure 5 This is a schematic diagram of a tunnel lining defect detection device according to this application. In this application, the tunnel lining defect detection device 500 may specifically include the following structure:

[0150] Acquisition unit 501 is used to acquire ground-penetrating radar images of the target collected by the tunnel lining;

[0151] The detection unit 502 is used to input ground-penetrating radar images into a pre-configured tunnel lining defect detection network. The tunnel lining defect detection network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features of different scales from the ground-penetrating radar images input to the network. The features of different scales include first-scale features, second-scale features, third-scale features, and fourth-scale features. The feature pyramid network fuses features of different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has been pre-trained in the DetCo self-supervised framework.

[0152] Extraction unit 503 is used to extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

[0153] In yet another exemplary embodiment, the self-supervised backbone network adopts a dual-branch encoder architecture, specifically including a first online encoder, a first momentum encoder, a second online encoder, and a second momentum encoder.

[0154] The self-monitoring backbone network includes the following tasks:

[0155] Global image enhancement is performed on the ground-penetrating radar image input from the network to obtain the first view image and the second view image. Local image patch enhancement is performed on the ground-penetrating radar image input from the network to obtain the third view image and the fourth view image.

[0156] The first online encoder extracts a first multi-scale feature map from the first view image, the first momentum encoder extracts a second multi-scale feature map from the second view image, the second online encoder extracts a third multi-scale feature map from the third view image, and the second momentum encoder extracts a fourth multi-scale feature map from the fourth view image. Each of the four multi-scale feature maps includes a corresponding first feature map, second feature map, third feature map, and fourth feature map.

[0157] Adaptive average pooling and nonlinear projection are performed on the first, second, third, and fourth feature maps of the four multi-scale feature maps, respectively, to obtain the first-scale feature, second-scale feature, third-scale feature, and fourth-scale feature in sequence, forming features of different scales.

[0158] In another exemplary embodiment, for the current encoder, the current encoder includes an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The current encoder is a first online encoder, a first momentum encoder, a second online encoder, or a second momentum encoder. After training, while freezing the parameters of the first, second, and third feature extraction layers, the parameters of the fourth feature extraction layer are unfrozen. The processing of the first, second, third, and fourth feature maps includes:

[0159] The input layer preprocesses the view image input to the encoder to obtain a preprocessed feature map, wherein the view image input to the encoder is a first view image, a second view image, a third view image, or a fourth view image;

[0160] The first feature extraction layer extracts the first feature map from the preprocessed feature map;

[0161] The second feature extraction layer extracts the second feature map from the first feature map;

[0162] The third feature extraction layer extracts the third feature map from the second feature map;

[0163] The fourth feature extraction layer extracts the fourth feature map from the third feature map.

[0164] In yet another exemplary embodiment, the weights of the first momentum encoder are synchronized with the parameters of the first online encoder, and the weights of the second momentum encoder are synchronized with the parameters of the second online encoder.

[0165] The weights of the online encoder are updated in real time through backpropagation, and the update strategy is as follows:

[0166]

[0167] in, These are the parameters of the momentum encoder at time t. These are the parameters of the momentum encoder at time t-1. Here, represents the parameters of the online encoder at time t, and m is the momentum coefficient.

[0168] In yet another exemplary embodiment, for a self-supervised backbone network, the loss function used for self-supervised pre-training is:

[0169]

[0170] in, For the global contrast loss at the i-th scale, For the local contrast loss at the i-th scale, Let ω be the global-local contrast loss at the i-th scale.i The preset weights for the i-th scale;

[0171]

[0172] The first view image is denoted as I. q The second view image is denoted as I. k For the i-th layer, From I q Query features generated by an online encoder Corresponding to positive sample features, and from I k , correspond Negative sample features, from the feature queue of the i-th layer The temperature coefficient τ is obtained from the sampling.

[0173]

[0174] For the i-th layer, local features are extracted. As a local query feature, Positive sample features corresponding to local blocks For the corresponding Negative sample features, from the local feature queue of the i-th layer Obtained by sampling from the middle;

[0175]

[0176] Among them, global features are forced through cross-modal alignment. With local features Alignment.

[0177] In yet another exemplary embodiment, the detection head includes an alignment unit and a dual-branch prediction unit;

[0178] The alignment unit dynamically adjusts the feature weights based on the estimated burial depth of the candidate region, and uses these feature weights to extract features of the region of interest from the candidate region.

[0179] Among them, feature weights d represents the relative position of the ground-penetrating radar image input to the network corresponding to the current feature map, and α represents the medium attenuation coefficient;

[0180] The dual-branch prediction unit predicts the probability of the type of tunnel lining defects based on the characteristics of the region of interest, and generates a prediction bounding box of the location of the defects, thus obtaining the type and location of the tunnel lining defects.

[0181] In yet another exemplary embodiment, the dual-branch prediction unit includes a classification branch and a regression branch, wherein the classification branch predicts the probability of the type of defects in the tunnel lining, and the regression branch generates a predicted bounding box of the location of the defects.

[0182] The loss function used in the classification branch is denoted as The loss function used in the regression branch is denoted as The total loss used by the detection head during the training of the tunnel lining defect detection network is:

[0183]

[0184] Where C is the total number of categories, z i To predict the probability of the i-th type of disease, p c To predict the probability of the obtained category c defects, which correspond to cracks, voids, and delamination, as well as the rebar area, y c This is a real label;

[0185]

[0186] Where β is the preset value and x is the error between the predicted value and the actual value.

[0187] This application also provides a processing device from a hardware architecture perspective, see [link / reference]. Figure 6 , Figure 6 This diagram illustrates a structural schematic of the processing device of this application. Specifically, the processing device may include a processor 601, a memory 602, and an input / output device 603. The processor 601 executes the computer program stored in the memory 602 to implement, for example... Figure 1 The steps of the tunnel lining defect detection method in the corresponding embodiment; or, when the processor 601 executes the computer program stored in the memory 602, it implements as follows: Figure 5 Corresponding to the functions of each unit in the embodiment, the memory 602 is used to store the functions executed by the processor 601 as described above. Figure 1 The computer program required for the tunnel lining defect detection method in the corresponding embodiment.

[0188] For example, a computer program may be divided into one or more modules / units, one or more of which are stored in memory 602 and executed by processor 601 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a computer device.

[0189] The processing device may include, but is not limited to, processor 601, memory 602, and input / output device 603. Those skilled in the art will understand that the illustrations are merely examples of the processing device and do not constitute a limitation on the processing device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the processing device may also include network access devices, buses, etc., and processor 601, memory 602, input / output device 603, etc., are connected via a bus.

[0190] Processor 601 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the processing device, connecting various parts of the device through various interfaces and lines.

[0191] The memory 602 can be used to store computer programs and / or modules. The processor 601 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 602 and by calling data stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the processing device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0192] When processor 601 executes a computer program stored in memory 602, it can specifically perform the following functions:

[0193] Acquire ground-penetrating radar images of the target obtained from the tunnel lining;

[0194] Ground-penetrating radar (GPR) images are input into a pre-configured tunnel lining defect detection network. This network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features at different scales from the input GPR images, including first-scale, second-scale, third-scale, and fourth-scale features. The feature pyramid network fuses features at different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has undergone self-supervised pre-training within the DetCo self-supervised framework.

[0195] Extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

[0196] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device, tunnel lining defect detection and treatment equipment, and its corresponding units can be found by referring to, for example... Figure 1 The description of the tunnel lining defect detection method in the corresponding embodiment will not be repeated here.

[0197] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0198] Therefore, this application provides a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute the present application. Figure 1 The steps of the tunnel lining defect detection method in the corresponding embodiment can be referred to as follows for specific operations. Figure 1 The description of the tunnel lining defect detection method in the corresponding embodiment will not be repeated here.

[0199] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0200] Because of the instructions stored in the computer-readable storage medium, the present application can be executed as described above. Figure 1 The steps of the tunnel lining defect detection method in the corresponding embodiment can therefore achieve the results of this application. Figure 1 The beneficial effects that the tunnel lining defect detection method in the corresponding embodiment can achieve are detailed in the preceding description and will not be repeated here.

[0201] The tunnel lining defect detection method, apparatus, processing equipment, and computer-readable storage medium provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting tunnel lining defects, characterized in that, The method includes: Acquire ground-penetrating radar images of the target obtained from the tunnel lining; The ground-penetrating radar image is input into a pre-configured tunnel lining defect detection network. This network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features at different scales from the input ground-penetrating radar image. These features include first-scale, second-scale, third-scale, and fourth-scale features. The feature pyramid network fuses these features to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has undergone self-supervised pre-training within the DetCo self-supervised framework. Extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

2. The method according to claim 1, characterized in that, The self-supervised backbone network adopts a dual-branch encoder architecture, specifically including a first online encoder, a first momentum encoder, a second online encoder, and a second momentum encoder; The self-supervised backbone network includes the following tasks: Global image enhancement is performed on the ground-penetrating radar image input to the network to obtain a first view image and a second view image; local image patch enhancement is performed on the ground-penetrating radar image input to the network to obtain a third view image and a fourth view image. The first online encoder extracts a first multi-scale feature map from the first view image, the first momentum encoder extracts a second multi-scale feature map from the second view image, the second online encoder extracts a third multi-scale feature map from the third view image, and the second momentum encoder extracts a fourth multi-scale feature map from the fourth view image. Each of the four multi-scale feature maps includes a corresponding first feature map, second feature map, third feature map, and fourth feature map. Adaptive average pooling and nonlinear projection are performed on the first feature map, the second feature map, the third feature map, and the fourth feature map in the four multi-scale feature maps, respectively, to obtain the first scale feature, the second scale feature, the third scale feature, and the fourth scale feature in sequence, thus forming the features of different scales.

3. The method according to claim 2, characterized in that, For the current encoder, which includes an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer, and is either the first online encoder, the first momentum encoder, the second online encoder, or the second momentum encoder, after training, while freezing the parameters of the first, second, and third feature extraction layers, unfreezing the parameters of the fourth feature extraction layer, the processing of the first, second, third, and fourth feature maps includes: The input layer preprocesses the view image input by the encoder to obtain a preprocessed feature map, wherein the view image input by the encoder is the first view image, the second view image, the third view image, or the fourth view image; The first feature extraction layer extracts the first feature map from the preprocessed feature map; The second feature extraction layer extracts the second feature map from the first feature map; The third feature extraction layer extracts the third feature map from the second feature map; The fourth feature extraction layer extracts the fourth feature map from the third feature map.

4. The method according to claim 3, characterized in that, The weights of the first momentum encoder are synchronized with the parameters of the first online encoder, and the weights of the second momentum encoder are synchronized with the parameters of the second online encoder. The weights of the online encoder are updated in real time through backpropagation, and the update strategy is as follows: in, These are the parameters of the momentum encoder at time t. These are the parameters of the momentum encoder at time t-1. Here, represents the parameters of the online encoder at time t, and m is the momentum coefficient.

5. The method according to claim 2, characterized in that, For the self-supervised backbone network, the loss function used in the self-supervised pre-training is: in, For the global contrast loss at the i-th scale, The local contrast loss at the i-th scale is... Let ω be the global-local contrast loss at the i-th scale. i The preset weights for the i-th scale; Wherein, the first view image is denoted as I. q The second view image is denoted as I. k For the i-th layer, From I q Query features generated by an online encoder Corresponding to positive sample features, and from I k , correspond Negative sample features, from the feature queue of the i-th layer. The temperature coefficient τ is obtained from the sampling. For the i-th layer, local features are extracted. As a local query feature, Positive sample features corresponding to local blocks For the corresponding Negative sample features, from the local feature queue of the i-th layer. Obtained by sampling from the middle; Among them, global features are forced through cross-modal alignment. With local features Alignment.

6. The method according to claim 1, characterized in that, The detection head includes an alignment unit and a dual-branch prediction unit; The alignment unit dynamically adjusts the feature weights based on the estimated burial depth of the candidate regions, and uses these feature weights to extract features of the region of interest from the candidate regions. Among them, feature weights d represents the relative position of the current feature map to the ground-penetrating radar image input to the network, and α represents the medium attenuation coefficient; The dual-branch prediction unit predicts the probability of the type of tunnel lining defects based on the characteristics of the region of interest, and generates a prediction bounding box of the location of the defects, thereby obtaining the type of tunnel lining defects and the location of the tunnel lining defects.

7. The method according to claim 6, characterized in that, The dual-branch prediction unit includes a classification branch and a regression branch. The classification branch predicts the probability of the type of defects in the tunnel lining, and the regression branch generates a predicted bounding box of the location of the defects. The loss function used by the classification branch is denoted as... The loss function used in the regression branch is denoted as... The total loss used by the detection head during the training of the tunnel lining defect detection network is: Where C is the total number of categories, z i To predict the probability of the i-th type of disease, p c To predict the probability of the obtained category c defects, which correspond to the three defects of cracks, voids, and detachments, as well as the rebar area, y c This is a real label; Where β is the preset value and x is the error between the predicted value and the actual value.

8. A device for detecting tunnel lining defects, characterized in that, The device includes: The acquisition unit is used to acquire ground-penetrating radar images of the target collected by the tunnel lining. A detection unit is used to input the ground-penetrating radar image into a pre-configured tunnel lining defect detection network. The tunnel lining defect detection network is constructed based on a self-supervised backbone network, a feature pyramid network, a region proposal network, and a detection head. The self-supervised backbone network extracts features at different scales from the input ground-penetrating radar image. These features include first-scale features, second-scale features, third-scale features, and fourth-scale features. The feature pyramid network fuses these features at different scales to obtain fused features. The region proposal network generates candidate regions corresponding to the fused features. The detection head determines the tunnel lining defect category and location based on the candidate regions. The tunnel lining defect detection network is trained after the self-supervised backbone network has undergone self-supervised pre-training within the DetCo self-supervised framework. The extraction unit is used to extract the tunnel lining defect detection results output by the tunnel lining defect detection network.

9. A processing device, characterized in that, The method includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the method as described in any one of claims 1 to 7 when it invokes the computer program in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the method of any one of claims 1 to 7.

Images