Tunnel lining defect ground penetrating radar image enhancement method, medium and equipment

Abstract

The invention relates to the technical field of image processing, in particular to a tunnel lining defect ground penetrating radar image enhancement method, medium and equipment, and the method comprises the following steps: constructing a data set; constructing a generative adversarial network model based on improved deep convolution; the obtained tunnel lining structure defect ground penetrating radar image data set is used for training in the constructed improved deep convolution-based generative adversarial network model, and hyper-parameters of a generator and a discriminator are continuously adjusted through a test result, so that the adversarial network model is globally optimal; and generating a tunnel lining structure defect ground penetrating radar image by using the improved deep convolution generative adversarial network model. According to the invention, the technical problem that the difference between the image generated by the existing image processing method and the real ground penetrating radar image is large for the ground penetrating radar image of the internal defect of the tunnel lining is solved.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method, medium, and device for enhancing ground-penetrating radar images of tunnel lining defects. Background Technology

[0002] As a crucial load-bearing structure of tunnels, tunnel lining plays a vital role in maintaining overall structural stability and ensuring safe tunnel operation. Due to factors such as construction techniques, operating environment, and geological conditions, various defects inevitably appear within the tunnel lining structure, such as voids, cavities, honeycomb texture, loose areas, and cracks. Once these internal defects develop and accumulate, they will lead to a decrease in structural load-bearing capacity and weakened waterproofing performance. In severe cases, they may even induce major engineering accidents such as structural failure, leakage, or collapse, posing a serious threat to the long-term safe operation of the tunnel.

[0003] Ground-penetrating radar (GPR) technology, with its advantages of being non-destructive, rapid, and efficient, has become a core technical means for detecting internal defects in tunnel linings. However, due to the complex layers and significant differences in dielectric constants of tunnel lining materials, coupled with environmental noise interference and limitations imposed by electromagnetic wave propagation characteristics, GPR images often exhibit complex textures, severe noise, and blurred defect boundaries. This not only increases the difficulty of manually interpreting radar image data but also makes it difficult to guarantee the accuracy and stability of defect identification.

[0004] Compared to traditional interpretation methods based on human experience, deep learning technology can automatically extract high-dimensional features from a large number of samples, significantly improving the efficiency and objectivity of detection. However, the quantity and quality of the ground-penetrating radar (GPR) image dataset of internal defects in tunnel linings also have a significant impact on the recognition performance of deep learning algorithms. If the number of images in the dataset is insufficient, it is difficult to extract enough detailed features for deep learning. In actual tunnel engineering, acquiring a sufficient number of high-quality GPR images of internal defects in tunnel linings with rich features is not easy. It not only requires a large number of tunnels as engineering support, but also the tunnel lines generally cannot be in a state of maintenance for too long, with limited maintenance windows.

[0005] ① Traditional data augmentation methods adjust and transform existing images, but cannot generate entirely new image samples.

[0006] ② Although numerical simulation methods can accurately simulate ground-penetrating radar images of internal defects in tunnel lining structures, the generated images are too idealized, completely devoid of noise interference, and differ significantly from real ground-penetrating radar images.

[0007] ③ Conventional generative adversarial networks still have certain limitations in the application of ground-penetrating radar image data enhancement for internal defects in tunnel linings: the generated images have low resolution, the details are relatively blurry, and it is difficult to clearly show the subtle features and structural textures of the defects, failing to achieve the high resolution level of real ground-penetrating radar images; the generated images differ from real ground-penetrating radar images in terms of noise distribution and the naturalness of defect features, and may have problems such as distorted defect morphology and unnatural background texture, resulting in generated images lacking the realism and credibility of real data; during the training process, it is difficult for the generator and discriminator to achieve a good balance, with large fluctuations, affecting the convergence effect and stability of the generation performance of the model. Summary of the Invention

[0008] The main objective of this invention is to provide a method, medium, and device for enhancing ground-penetrating radar images of tunnel lining defects, aiming to solve the technical problem that existing image processing methods generate images with significant discrepancies compared to actual ground-penetrating radar images of internal defects in tunnel linings.

[0009] To achieve the above objectives, this invention proposes a method for enhancing ground-penetrating radar images of tunnel lining defects, comprising the following steps: S1. Construct a dataset of ground-penetrating radar images of defects in tunnel lining structures; S2. Construct an improved deep convolutional generative adversarial network model. Specifically, based on the original deep convolutional generative adversarial network model, increase the depth of the network structure, introduce spectral normalization operation in the discriminator, and introduce a multi-scale efficient attention module to obtain the improved deep convolutional generative adversarial network model. S3. Improved training and testing optimization of deep convolutional generative adversarial network model: Specifically, the ground-penetrating radar image dataset of tunnel lining structure defects obtained in S1 is used to train the improved deep convolutional generative adversarial network model built in S2. The hyperparameters of the generator and discriminator are continuously adjusted through test results so that the adversarial network model reaches the global optimum. S4. Use an improved deep convolutional generative adversarial network model to generate ground-penetrating radar images of tunnel lining structural defects. Specifically, save the training weights of the adversarial network model when it reaches the global optimum in S3, and use the adversarial network model to generate ground-penetrating radar images of internal structural defects in the tunnel lining.

[0010] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention is that S1 specifically includes: collecting original ground-penetrating radar image data of tunnel lining structures, performing denoising processing and traditional data enhancement on the original image data to obtain ground-penetrating radar images of lining structure defects, and forming a ground-penetrating radar image dataset of tunnel lining structure defects for model training.

[0011] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention is that, when increasing the depth of the network structure, the convolutional layer of the original depth convolution generative adversarial network model includes convolution, batch normalization, and leaky rectified linear units; the deconvolutional layer of the original depth convolution generative adversarial network model includes deconvolution, batch normalization, and rectified linear units. The improved deep convolutional generative adversarial network model's convolutional layers include convolution, multi-scale efficient attention modules, spectral normalization, and leaky rectified linear units; The deconvolutional layers of the improved deep convolutional generative adversarial network model include: deconvolution, multi-scale efficient attention modules, batch normalization, and rectified linear units.

[0012] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention lies in controlling the gradient variation range of the network by constraining the weight matrix of each convolutional layer in the discriminator when introducing a spectral normalization operation. The specific steps are as follows: weight matrix The maximum singular value is: ; weight matrix Normalization is performed. ; in: Weight matrix The maximum singular value; This indicates that for all non-zero input vectors Find the maximum value; The feature vector input to this linear layer; This is the weight matrix after spectral normalization.

[0013] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention lies in the introduction of a multi-scale efficient attention module to divide the input feature vector into multiple sets of sub-feature vectors along the channel dimension, as expressed below: ; Among them, The size of the input feature vector. The size of the sub-feature vector. g The number of sub-eigenvector groups.

[0014] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention lies in that, after dividing the input feature vector into multiple sets of sub-feature vectors along the channel dimension, features at different scales are processed through two sets of parallel branches. The global branch employs one-dimensional global average pooling and 1×1 convolution. Activation functions capture long-range dependencies, and the process is as follows: ; Local branches use 3×3 convolution to extract local spatial information, and utilize... The function is standardized based on its short-range characteristics, as follows: ; in: for Activation function for function, For 1×1 convolution, It is a 3×3 convolution. and For one-dimensional global average pooling, For global branches, For local branches, For connection operations, F This is the sub-feature vector.

[0015] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention lies in that the output feature vectors of the global branch and the local branch are respectively subjected to two-dimensional global average pooling. Encode spatial information and use matrix multiplication. Achieving cross-scale feature interaction by combining global semantic guidance with local detail enhancement The process is as follows: ; in: For two-dimensional global average pooling, For matrix multiplication, This is the feature vector after the interaction.

[0016] A further improvement of the ground-penetrating radar image enhancement method for tunnel lining defects in this invention lies in utilizing... The feature vectors after function standardization interaction are reweighted and then output as the corresponding feature vectors. The process is as follows: ; in: Weighted by repetition.

[0017] The present invention also provides a readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above for enhancing the ground-penetrating radar image of tunnel lining defects.

[0018] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, runs the ground-penetrating radar image enhancement method for tunnel lining defects as described above.

[0019] The technical solution of the present invention has the following beneficial effects: The present invention provides a method for enhancing ground-penetrating radar (GPR) images of tunnel lining defects. This method utilizes a deep convolutional generative adversarial network (GAN) based on unsupervised learning. The generator is responsible for producing realistic GPR images from random noise, while the discriminator distinguishes between the generator's output and real images. During training, the generator continuously optimizes to attempt to "deceive" the discriminator, while the discriminator continuously improves to accurately distinguish between real and fake images. Both systems progress together in this adversarial game, enabling the generator to produce realistic images. This addresses the technical problem that existing image processing methods for GPR images of internal tunnel lining defects often produce images that differ significantly from real GPR images.

[0020] This invention deepens the structure of the original network model, doubles the resolution of the generated images, and introduces a multi-scale efficient attention module that combines global semantic guidance with local detail enhancement. This improves the model's ability to extract detailed features from ground-penetrating radar images of lining defects, and solves the problem that existing networks generate images with low resolution, blurry details, and difficulty in clearly displaying the subtle features and structural textures of defects.

[0021] This invention uses spectral normalization in the model to effectively control the gradient variation range of the network, prevent gradient explosion or vanishing, thereby improving the stability of model training. It solves the problem that the generator and discriminator are difficult to achieve a good balance during the training process of existing networks, resulting in large fluctuations that affect the convergence effect and generation performance stability of the model. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the ground-penetrating radar image enhancement method for tunnel lining defects according to the present invention; Figure 2 This is a diagram of a traditional deconvolution structure. Figure 3 This is an improved deconvolution structure diagram of the ground-penetrating radar image enhancement method for tunnel lining defects according to the present invention; Figure 4 This is a diagram of a traditional convolutional structure. Figure 5 This is an improved convolutional structure diagram of the ground-penetrating radar image enhancement method for tunnel lining defects according to the present invention; Figure 6 This is a flowchart of the multi-scale efficient attention module of the ground-penetrating radar image enhancement method for tunnel lining defects according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0026] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0027] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0029] like Figure 1As shown, this invention proposes a method for enhancing ground-penetrating radar images of tunnel lining defects, comprising the following steps: S1. Construct a ground-penetrating radar image dataset of tunnel lining structure defects. Specifically, this includes: collecting raw ground-penetrating radar image data of tunnel lining structures, performing denoising processing and traditional data augmentation (random cropping, mirroring, scaling, size adjustment, etc.) on the raw image data to obtain ground-penetrating radar images of lining structure defects, and forming a ground-penetrating radar image dataset of tunnel lining structure defects for model training.

[0030] S2. Construct an improved deep convolutional generative adversarial network model. Specifically, based on the original deep convolutional generative adversarial network model, increase the depth of the network structure, introduce spectral normalization operation in the discriminator, and introduce multi-scale efficient attention mode to obtain the improved deep convolutional generative adversarial network model. S201. When increasing the depth of the network structure, it means increasing the original deep convolutional generative adversarial network model structure (five convolutional layers and five deconvolutional layers) to six convolutional layers (Convolution, Conv) and six deconvolutional layers (Deconvolution, Deconv); S20101, such as Figures 2-5 As shown, the convolutional layers of the original deep convolutional generative adversarial network model include convolution (Conv), batch normalization (BN), and leaky rectified linear units (Leaky ReLU); the deconvolutional layers of the original deep convolutional generative adversarial network model include deconvolution (Deconv), batch normalization (BN), and rectified linear units (ReLU). S20102. The improved deep convolutional generative adversarial network model consists of convolutional layers (Conv), efficient multi-scale attention (EMA), spectral normalization (SN), and leaky rectified linear units (ReLU). The improved deep convolutional generative adversarial network model also consists of deconvolutional layers (Deconv_EMA), efficient multi-scale attention (EMA), batch normalization (BN), and rectified linear units (ReLU).

[0031] Table 1. Network Structure of the Adversarial Network Model of this Invention

[0032] This invention deepens the generator and discriminator network structures in the original deep convolutional generative adversarial network model, thereby expanding the resolution of the generated images from 64×64 to 128×128.

[0033] The increased resolution directly enhances the visual effect of ground-penetrating radar (GPR) images. Compared to 64×64, 128×128 images can present more detailed information (such as texture, edges, and local structures), making them closer to the expressive power of real GPR images. The addition of network structure expands the model's feature extraction and mapping capabilities. The generator can learn more complex feature combinations, and the discriminator can more accurately identify real and fake features in the image, improving the overall generation quality of the model. It also provides sufficient network depth to support the subsequent introduction of more complex modules, avoiding the problem of insufficient feature expression capabilities caused by module insertion.

[0034] S202. When introducing spectral normalization into the discriminator, the gradient variation range of the network is controlled by constraining the weight matrix of each convolutional layer in the discriminator, preventing gradient explosion or vanishing, thereby improving the stability of model training. The specific steps are as follows: weight matrix The maximum singular value is: ; weight matrix Normalization is performed. ; in: Weight matrix The maximum singular value; This indicates that for all non-zero input vectors Find the maximum value; The feature vector input to this linear layer; This is the weight matrix after spectral normalization.

[0035] This invention stabilizes the training process by introducing a spectral normalization operation into the discriminator, effectively alleviating the problems of gradient vanishing and gradient exploding.

[0036] Significantly improves training stability by constraining the discriminator's Pushtz constant, effectively suppressing gradient vanishing and gradient explosion phenomena, making the loss curve smoother during training and avoiding training collapse (mode collapse) problems; reduces sensitivity to hyperparameters. Traditional deep convolutional generative adversarial networks require fine-tuning of hyperparameters such as learning rate and batch size to maintain training stability. After introducing spectral normalization, the adjustable range of hyperparameters is wider, reducing the difficulty of model parameter tuning; improves the learning efficiency of the generator. Stable gradient feedback allows the generator to more clearly capture the discriminator's evaluation signal, accelerating model convergence, while reducing the problem of pattern uniformity in generated images and improving the diversity of generated results.

[0037] S203, such as Figure 6 As shown, the feature vectors output by the multi-scale efficient attention module include: S20301. Divide the input feature vector into multiple sets of sub-feature vectors along the channel dimension to reduce computation while preserving complete channel information. The expression is as follows: ; Among them, The size of the input feature vector. The size of the sub-feature vector. g The number of sub-eigenvector groups.

[0038] S20302. After dividing the input feature vector into multiple sets of sub-feature vectors along the channel dimension, two sets of parallel branches process features of different scales respectively. The global branch uses one-dimensional global average pooling and 1×1 convolution. Activation functions capture long-range dependencies, and the process is as follows: ; Local branches use 3×3 convolution to extract local spatial information, and utilize... The function is standardized based on its short-range characteristics, as follows: ; in: for Activation function for function, For 1×1 convolution, It is a 3×3 convolution. and For one-dimensional global average pooling, For global branches, For local branches, For connection operations, F This is the sub-feature vector.

[0039] S20303, the output feature vectors of the global branch and the local branch are respectively processed by two-dimensional global average pooling. Encode spatial information and use matrix multiplication. Achieving cross-scale feature interaction by combining global semantic guidance with local detail enhancement The process is as follows: ; in: For two-dimensional global average pooling, For matrix multiplication, This is the feature vector after the interaction.

[0040] S20304, Utilization The feature vectors after function standardization interaction are reweighted and then output as the corresponding feature vectors. The process is as follows: ; in: Weighted by repetition.

[0041] This invention enhances the global feature extraction capabilities of the generator and discriminator by introducing a multi-scale efficient attention module, thereby improving the quality of image detail feature representation.

[0042] Enhance global feature extraction capabilities by expanding the receptive field through multi-scale branches to solve the "local field of view limitation" problem of traditional convolution; improve the quality of detailed feature representation by increasing the weight of key details (such as texture and edges) in the attention module, making the details of the generated image clearer and more realistic, and reducing blurriness and false features; improve the model's feature utilization efficiency by avoiding feature information loss through multi-scale fusion and reducing redundant feature interference through attention weight allocation, thereby improving model performance without significantly increasing computational cost.

[0043] S3. Improved training and testing optimization of the deep convolutional generative adversarial network (GAN) model: The ground-penetrating radar image dataset of tunnel lining structural defects obtained in S1 is used to train the improved deep convolutional GAN ​​model constructed in S2. The hyperparameters of the generator and discriminator are continuously adjusted based on the test results to make the adversarial network model reach the global optimum, and the training process is stable and easy to converge. Specifically, the hyperparameter adjustment of the GAN needs to be based on the test results, and the dynamic balance between the generator and discriminator is achieved through iterative optimization. Before adjustment, the performance needs to be quantified from two dimensions: generation quality and training stability, based on the test set. The distribution similarity of the generated images is evaluated using FID (Freche Initial Distance), PSNR (Peak Signal-to-Noise Ratio), and SSIM (Structural Similarity Index). At the same time, the trend of the loss curve and sample diversity are paid attention to determine the stability. The adjustment follows the logic of "test feedback → problem localization → parameter optimization → verification iteration", and is promoted step by step according to the priority of training hyperparameters, balancing hyperparameters, and structural hyperparameters. Training hyperparameters should be adjusted first. If gradient vanishing occurs during testing, the learning rate of the discriminator can be reduced, the learning rate of the generator increased, and a decay strategy can be adopted. If the loss fluctuates, the batch size can be increased or gradient accumulation can be used. If convergence is too slow, an adaptive momentum estimation optimizer can be used. Balancing hyperparameters are used to regulate the game rhythm. If the discriminator is too strong, its update frequency can be reduced or output noise can be added. If the generator is too strong and causes mode collapse, gradient penalties can be introduced. Structural hyperparameters should be adapted to task requirements. Underfitting can be addressed by adding hidden units, while overfitting can be addressed by regularization. A balance between generation quality and stability can be achieved through multiple iterations to meet the requirements.

[0044] S4. Use an improved deep convolutional generative adversarial network model to generate ground-penetrating radar images of tunnel lining structural defects. Specifically, save the training weights of the adversarial network model when it reaches the global optimum in S3, and use the adversarial network model to generate high-quality, high-definition ground-penetrating radar images of internal structural defects in the tunnel lining.

[0045] The present invention also provides a readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above for enhancing the ground-penetrating radar image of tunnel lining defects.

[0046] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, runs the ground-penetrating radar image enhancement method for tunnel lining defects as described above.

[0047] To verify the effectiveness of the method of the present invention, based on the above embodiments, a comparative experiment is conducted between the method of the present invention and several commonly used generative adversarial network algorithms: Data Preparation: A large amount of ground-penetrating radar (GPR) image data of the tunnel lining structure has been collected during intelligent tunnel operation and maintenance. After preliminary processing using specialized software, 800 valid GPR images containing two types of defects—voids and non-compactness—were selected. Data augmentation was then performed using two methods: geometric transformation and pixel-level transformation. Geometric transformation included operations such as horizontal mirroring, scaling, stretching, and cropping; pixel-level transformation included color adjustment, brightness adjustment, salt-and-pepper noise addition, Gaussian blurring, and contrast adjustment. This expanded the dataset to 1300 images, including 500 images of void defects, 500 images of non-compact defects, and 300 images of mixed defects.

[0048] Model experiments were conducted based on the above dataset, and the experimental results were compared with those of four other generative adversarial models. The results of the comparison experiments are shown in Table 2.

[0049] Table 2 Comparison of experimental results

[0050] The Fréchet Inception Distance (FID) measures the difference between the feature space vectors of generated and real images. FID considers the overall feature distribution of both generated and real images, rather than evaluating each image individually. This effectively assesses the quality of generated images and indirectly reflects the training progress of the model. A smaller FID indicates higher quality generated images.

[0051] Peak signal-to-noise ratio (PSNR) is used to measure the similarity between a generated image and a real image. The higher the PSNR, the more similar the generated image is to the real image.

[0052] The Structure Similarity Index Measure (SSIM) is used to measure the similarity between a generated image and a real image in three dimensions: brightness, contrast, and structure. The SSIM value ranges from -1 to 1, with values ​​closer to 1 indicating a higher similarity between the generated and real images.

[0053] As shown in Table 2, compared with the other four models, the model of this invention has the best performance in the three evaluation indicators of FID, PSNR and SSIM on the test set, which are 207.50, 28.45db and 0.547 respectively. This indicates that the ground penetrating radar image of the lining structure defect generated by the model of this invention has largely learned the defect features in the real image, and is more similar to the real image in terms of brightness, contrast and structure.

[0054] The above description is only a preferred embodiment of the present invention and does not limit the scope of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of the present invention.

Claims

1. A method for enhancing ground-penetrating radar images of tunnel lining defects, characterized in that, Includes the following steps: S1. Construct a dataset of ground-penetrating radar images of defects in tunnel lining structures; S2. Construct an improved deep convolutional generative adversarial network model. Specifically, based on the original deep convolutional generative adversarial network model, increase the depth of the network structure, introduce spectral normalization operation in the discriminator, and introduce a multi-scale efficient attention module to obtain the improved deep convolutional generative adversarial network model. S3. Improved training and testing optimization of deep convolutional generative adversarial network model: Specifically, the ground-penetrating radar image dataset of tunnel lining structure defects obtained in S1 is used to train the improved deep convolutional generative adversarial network model built in S2. The hyperparameters of the generator and discriminator are continuously adjusted through test results so that the adversarial network model reaches the global optimum. S4. Use an improved deep convolutional generative adversarial network model to generate ground-penetrating radar images of tunnel lining structural defects. Specifically, save the training weights of the adversarial network model when it reaches the global optimum in S3, and use the adversarial network model to generate ground-penetrating radar images of internal structural defects in the tunnel lining.

2. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 1, characterized in that, S1 specifically includes: collecting raw ground-penetrating radar image data of tunnel lining structures, performing denoising processing and traditional data augmentation on the raw image data to obtain ground-penetrating radar images of lining structure defects, and forming a ground-penetrating radar image dataset of tunnel lining structure defects for model training.

3. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 1, characterized in that, When increasing the depth of the network structure, the convolutional layers of the original deep convolutional generative adversarial network model include convolution, batch normalization, and leaky rectified linear units; the deconvolutional layers of the original deep convolutional generative adversarial network model include deconvolution, batch normalization, and rectified linear units. The improved deep convolutional generative adversarial network model's convolutional layers include convolution, multi-scale efficient attention modules, spectral normalization, and leaky rectified linear units; The deconvolutional layers of the improved deep convolutional generative adversarial network model include: deconvolution, multi-scale efficient attention modules, batch normalization, and rectified linear units.

4. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 1, characterized in that, When introducing spectral normalization into the discriminator, the gradient variation range of the network is controlled by constraining the weight matrix of each convolutional layer in the discriminator. The specific steps are as follows: weight matrix The maximum singular value is: ； weight matrix Normalization is performed. ； in: Weight matrix The maximum singular value; This indicates that for all non-zero input vectors Find the maximum value; The feature vector input to this linear layer; This is the weight matrix after spectral normalization.

5. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 4, characterized in that, A multi-scale efficient attention module is introduced to divide the input feature vector into multiple sets of sub-feature vectors along the channel dimension, as shown in the following expression: ； Among them, The size of the input feature vector. The size of the sub-feature vector. g The number of sub-eigenvector groups.

6. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 5, characterized in that, After dividing the input feature vector into multiple sub-feature vectors along the channel dimension, two parallel branches process features at different scales respectively. The global branch uses one-dimensional global average pooling and 1×1 convolution. Activation functions capture long-range dependencies, and the process is as follows: ； Local branches use 3×3 convolution to extract local spatial information, and utilize... The function is standardized based on its short-range characteristics, as follows: ； in: for Activation function for function, For 1×1 convolution, It is a 3×3 convolution. and For one-dimensional global average pooling, For global branches, For local branches, For connection operations, F This is the sub-feature vector.

7. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 6, characterized in that, The output feature vectors of the global branch and the local branch are respectively processed by two-dimensional global average pooling. Encode spatial information and use matrix multiplication. Achieving cross-scale feature interaction by combining global semantic guidance with local detail enhancement The process is as follows: ； in: For two-dimensional global average pooling, For matrix multiplication, This is the feature vector after the interaction.

8. The method for enhancing ground-penetrating radar images of tunnel lining defects as described in claim 7, characterized in that, use The feature vectors after function standardization interaction are reweighted and then output as the corresponding feature vectors. The process is as follows: ； in: Weighted by repetition.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-8, which is a ground-penetrating radar image enhancement method for tunnel lining defects.

10. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, performs the ground-penetrating radar image enhancement method for tunnel lining defects according to any one of claims 1-8.

Images