Geological radar profile horizontal interference suppression method based on diffusion model

Abstract

The invention discloses a ground penetrating radar (GPR) section horizontal interference suppression method based on a diffusion model. The method comprises the following steps: (1) constructing a standardized profile interference data set containing an actual measurement sample and a simulation sample; wherein an actual measurement sample comes from an urban road detection profile, and proper samples are screened out through mean value reduction preprocessing operation; a simulation sample is subjected to forward modeling based on a time domain finite difference method (FDTD), and various road structures and underground anomaly features are simulated; (2) a comprehensive ResBlock is introduced into the diffusion model, a lightweight Agent Agent module is introduced between ResBlocks of all layers, a space attention mechanism is deployed at the middle connection position of the Unet architecture, and a strategy based on Cosine theta schema is introduced in the forward diffusion stage; (3) training by using an improved diffusion model based on the constructed data set to obtain a trained model; and (4) sending test set data into the trained model, and completing the horizontal interference suppression test reasoning of the GPR profile. According to the method, construction of the horizontal interference data set and improvement of the diffusion model are combined together, the problems that a current traditional low-rank decomposition method is insufficient in intelligent degree, difficult to consider multi-source data and depends on manual parameter adjustment are solved, and technical support is provided for GPR high-resolution imaging and anomaly detection.

Description

Technical Field

[0001] The present invention belongs to the field of graphics and image processing, and in particular relates to a method for suppressing horizontal interference of geological radar profiles based on a diffusion model. Background Art

[0002] Horizontal interference is a common and unavoidable type of noise, typically manifesting as high-amplitude bands covering the entire GPR profile with fluctuating characteristics. This interference arises from a variety of sources, including electronic design issues with the radar antenna, impedance mismatch between the antenna and the ground, and multiple reflections of electromagnetic waves at stratum interfaces. This interference not only affects image clarity but, in severe cases, can obscure the true reflection characteristics of subsurface targets, reducing the image quality provided by the GPR system. Therefore, effective methods for suppressing horizontal interference are crucial for improving GPR image quality and enhancing target detection accuracy.

[0003] Traditional GPR data processing methods, such as principal component analysis (PCA), robust principal component analysis (RPCA), non-negative matrix factorization (NMF), and GoDecomposition (GoDec), have been proposed for separating and removing horizontal low-rank interference. However, these methods often suffer from the following problems: First, they are not compatible with various types of GPR data and have limited effectiveness in suppressing complex interference; second, these methods often rely on manual design and parameter adjustment, resulting in low intelligence and difficulty in adapting to the complex variations found in real-world scenarios.

[0004] In recent years, deep learning techniques have made significant progress in image restoration and noise suppression. Generative adversarial networks (GANs), in particular, have been widely used in GPR image restoration. Through adversarial training between a generator and a discriminator, GANs are able to restore more realistic and detailed image information. However, GAN training often faces instability issues, particularly prone to mode collapse, resulting in suboptimal generated image quality or artifacts.

[0005] Diffusion models have recently demonstrated remarkable performance in various image generation tasks. Their fundamental principle is to model the diffusion process and learn its inverse. However, current diffusion model training networks are mostly based on variants of the UNet architecture, which lacks deep structural innovations in feature extraction and context modeling. Furthermore, existing deep learning models for horizontal interference suppression in GPR profiles often suffer from insufficient noise immunity and excessive energy loss in target reflection wave groups. Summary of the Invention

[0006] Purpose of the invention: To address the issues of traditional low-rank decomposition methods, which lack intelligence, struggle to account for multi-source data, and rely on manual parameter adjustment, this invention suppresses horizontal interference in GPR data profiles based on a constructed horizontal interference dataset and an improved diffusion model.

[0007] To achieve the above purpose, the present invention adopts the following technologies:

[0008] A method for suppressing horizontal interference of geological radar profile based on diffusion model includes the following steps:

[0009] (1) Construct a standardized GPR horizontal interference dataset: collect measured profile samples and perform mean subtraction preprocessing to select appropriate data; at the same time, perform forward simulation based on FDTD to generate simulation samples containing different road structures and underground anomaly shapes to form a unified standard dataset;

[0010] (2) Building the diffusion model network structure: The residual module is fully introduced into the original Unet framework, a lightweight Agent Attention mechanism is inserted between the residual modules of each layer, and a spatial attention connection structure is set in the middle of the network; a noise scheduling strategy based on Cosineθ is adopted in the forward diffusion process;

[0011] (3) Performing model training: Using the constructed data set to train the improved diffusion model, a model for horizontal interference suppression is obtained;

[0012] (4) Complete interference suppression reasoning: Input the GPR profile to be processed into the trained model to obtain the output result after interference suppression.

[0013] Furthermore, the specific steps of constructing the dataset in step 1 include:

[0014] Step 1: Obtain measured samples: Use ground penetrating radar with a center frequency of 400MHz or 100MHz to detect urban roads, with a sampling interval of 0.023m; perform mean subtraction processing on the collected original GPR profiles, and select samples with better processing effects;

[0015] Step 2: Generate simulation samples: Based on the FDTD method, forward modeling is performed to construct a heterogeneous underground model containing different particles and cavity structures. A variety of simulation profile images are generated by setting several random control parameters. The image containing the defect is the noise sample, and the profile obtained by subtracting the GPR profile without the defect from the GPR profile with the defect is considered the ground truth sample.

[0016] Step 3: Summarize and organize the measured samples and simulated samples to construct a standardized GPR profile horizontal interference data set; then divide the samples into training set, test set and validation set to provide data support for the training and validation of the diffusion model.

[0017] Furthermore, the residual module in step 2 consists of two layers of 3×3 convolution and ReLU activation function, and the feature learning effect is enhanced through jump connections.

[0018] Furthermore, in step 2, the Agent Attention module introduces the agent vector A to construct a four-element attention mechanism (Q, A, K, V).

[0019] Furthermore, in step 2, the spatial attention module performs maximum pooling and average pooling operations on the input feature map in the channel dimension and then splices it, and generates a spatial attention map through 7×7 convolution and Sigmoid function.

[0020] Furthermore, the loss function of the network in step 2 draws on the maximum likelihood learning strategy to improve training stability. To address the difficulty in predicting the instantaneous noise distribution in the diffusion model, the optimization objective is set to minimize the Euclidean distance between the reconstructed states:

[0021]

[0022] where γ t represents positive weight, According to the current state x t and noise estimation network The optimal previous state estimate obtained by the closed-form calculation is calculated by the analytical expression to ensure that the network can more accurately fit the intermediate state in the back-diffusion process. T Start sampling and iteratively solve the diffusion SDE by combining numerical solutions such as Euler–Maruyama to generate a high-resolution map.

[0023] Furthermore, the specific configuration of the network parameters in step 2 is described as follows: the input map is divided into 128×128 patches, the batch size is set to 16, and the Adam optimizer is used for training with hyperparameters β1=0.9, β2=0.99, and the initial learning rate is set to 1×10 -4 , and decay the learning rate to half of its original value after every 200,000 iterations; the entire network was trained for a total of 500,000 iterations. The training process was run on an RTX 4090 GPU with 24GB of video memory and took about 4 days to complete, ensuring that the model achieved a good balance between stability and convergence performance.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] 1. Construct a dataset for GPR horizontal interference, combining both measured and simulated data. In the simulation phase, a numerical model of the highway subgrade was first constructed based on the actual structure. The dataset was then expanded through forward modeling. This combined approach of measured and simulated dataset construction fills a gap in current GPR horizontal interference dataset construction methods.

[0026] 2. The model disclosed in this invention fully introduces ResBlocks in the U-Net encoder, intermediate layers, and decoder to replace traditional dual convolutional units, improving training stability. Agent Attention Modules are added between ResBlocks in each layer to enhance global information modeling capabilities. This attention module itself has low computational complexity and low video memory usage, making it more efficient in processing high-resolution GPR data. Furthermore, a SpatialAttention mechanism is introduced between the encoder and decoder. This module infers an attention map along the spatial dimension and achieves adaptive feature optimization by multiplying it with the input feature map. During the forward diffusion noise addition phase, a noise scheduling strategy based on a cosine function is used to ensure that the noise variance varies smoothly over time. This invention comprehensively supplements the field of horizontal interference suppression algorithms for GPR data and provides new research ideas. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of the method for suppressing horizontal interference of geological radar profile based on the diffusion model of the present invention;

[0028] Figure 2 Examples of measured samples and simulated samples required for HIDiff training of the present invention;

[0029] Figure 3 This is a diagram showing the internal structure of the Resblock module of the present invention;

[0030] Figure 4 This is the internal structure diagram of the Agent Attention module of the present invention;

[0031] Figure 5 This is a diagram showing the internal structure of the spatial attention module of the present invention;

[0032] Figure 6 It is the overall framework diagram of HIDiff of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0034] The specific operation process of the horizontal interference suppression method of geological radar profile based on diffusion model is as follows: Figure 1 As shown, the following steps are included:

[0035] (1) Construct a standardized GPR horizontal interference dataset: collect measured profile samples and perform mean subtraction preprocessing to select appropriate data; at the same time, perform forward simulation based on FDTD to generate simulation samples containing different road structures and underground anomaly shapes to form a unified standard dataset;

[0036] (2) Building the diffusion model network structure: The residual module is fully introduced into the original Unet framework, a lightweight Agent Attention mechanism is inserted between the residual modules of each layer, and a spatial attention connection structure is set in the middle of the network; a noise scheduling strategy based on Cosineθ is adopted in the forward diffusion process;

[0037] (3) Performing model training: Using the constructed data set to train the improved diffusion model, a model for horizontal interference suppression is obtained;

[0038] (4) Complete interference suppression reasoning: Input the GPR profile to be processed into the trained model to obtain the output result after interference suppression.

[0039] In step 1, the specific steps of dataset construction include:

[0040] Step 1: The dataset used consists of two parts: measured samples and simulated samples. Among them, the high-quality images of the measured samples are the results after mean subtraction processing. However, since the processing effect of this method is not always ideal, the present invention only selects image data with better effects. In order to expand the database and obtain more convincing horizontal interference suppression samples, the study uses forward modeling simulation to expand the data. According to the urban road structure, the study randomly stacks small particles of different sizes and dielectric properties in the road layer to simulate the non-uniform underground environment and randomly generates abnormal cavity shapes. In addition, the simulation process involves several random control variables to enhance the algorithm's ability to suppress horizontal interference under different conditions. In the simulated samples, the GPR profile with defects is regarded as a low-quality image, while the image obtained by subtracting the GPR profile without defects (eliminating horizontal interference) from the GPR profile with defects is regarded as a high-quality image. For the specific settings of the dielectric parameters and signal parameters of the medium in the model, please see Table 1 and Table 2;

[0041] Table 1 Material parameters of GPRMax

[0042]

[0043] Table 2 Signal channel setting parameters of GPRMax

[0044]

[0045] Step 2: Summarize and organize the measured samples and simulated samples to construct a standardized GPR profile horizontal interference data set; then divide the samples into training set, test set and validation set to provide data support for the training and validation of the diffusion model.

[0046] Figure 2 Some data sample pairs are listed, representing all categories in the sample library, among which (a) and (b) belong to 400M GPR measured sample pairs; (c) and (d) belong to 100M GPR measured sample pairs; and (e) belongs to 500M GPR simulation sample pairs.

[0047] In step 2, the residual module consists of two convolutional layers and two activation function layers, such as Figure 3 As shown, specifically:

[0048] Step 1: First, the input feature map is fed into the first two-dimensional convolutional layer with a kernel size of 3×3 and a stride of (1,1) to extract local feature information.

[0049] Step 2: Then input the convolution result into the ReLU activation function layer to increase the nonlinear expression ability of the network;

[0050] Step 3: Then pass the activated feature map into the second two-dimensional convolutional layer, which also uses a 3×3 convolution kernel and a stride of (1,1);

[0051] Step 4: Processed again by ReLU activation function to further enhance the features;

[0052] Step 5: Finally, the output features processed by two convolutions and activation functions are added element by element to the original input feature map (i.e., skip connection) to form a residual connection structure to alleviate the gradient disappearance problem in deep networks and improve the efficiency and accuracy of feature learning.

[0053] In step 2, the Agent Attention Module is introduced, which innovatively introduces a set of additional agent vectors A into the attention triple (Q, K, V), defining a four-tuple agent attention paradigm (Q, A, K, V). Figure 4As shown in Figure 1, in proxy attention, instead of directly calculating the pairwise similarities between Q and K, a small number of proxy vectors A are used to collect information from K and V, which is then presented to Q, achieving global information modeling at a very low computational cost. From an overall structural perspective, proxy attention consists of two conventional Softmax attention operations and is equivalent to a generalized linear attention. This achieves a natural fusion of high-performance Softmax attention and efficient linear attention, thus combining the advantages of both: low computational complexity and strong model expressiveness.

[0054] In step 2, the spatial attention module first performs maximum pooling and average pooling operations on the channel dimension, and concatenates the results of the two to construct an efficient feature descriptor, such as Figure 5 As shown in Figure 2. Formally, the input features first undergo channel-wise max pooling and average pooling operations to generate the corresponding outputs. Subsequently, the concatenated feature descriptors are fed into a convolutional layer and coupled with a sigmoid activation function to generate a spatial attention map.

[0055] In short, the output features of the spatial attention module can be expressed as:

[0056]

[0057] Among them, f 7×7 (·) indicates the convolution operation with a kernel size of 7, and σ(·) indicates the Sigmoid function.

[0058] In summary, the overall network model is as follows Figure 6 As shown. The present invention adopts a conditional UNet structure, whose input consists of a noisy image x(t) and a low-resolution image μ. By calculating x(t)-μ and splicing it as the enhancement condition I t To the network input, it can enhance the network's perception of noise patterns; in addition, the network architecture proposed in this study uses ResBlock to replace the continuous convolution unit of traditional UNet on a global scale, which will help enhance deep information transmission and alleviate the gradient dissipation problem; Moreover, an Agent Attention Module is added between each layer of ResBlock in the network, and a Spatial Attention Module is added between the encoder and decoder to further enhance the feature modeling capability; Subsequently, the time-dependent network With external conditions I t and time t as input, aiming to predict the output pure noise

[0059]

[0060] On this basis, the present invention proposes a HINet architecture for noise prediction. Until it converges.

[0061] In step 2, the network loss function adopts the maximum likelihood learning strategy to improve training stability. To address the difficulty of predicting the instantaneous noise distribution in the diffusion model, the optimization objective is set to minimize the Euclidean distance between the reconstructed states:

[0062]

[0063] where γ t represents positive weight, According to the current state x t and noise estimation network The optimal previous state estimate obtained by the closed-form calculation is calculated by the analytical expression to ensure that the network can more accurately fit the intermediate state in the back-diffusion process. T Sampling is started, and the diffusion SDE is iteratively solved by combining numerical solutions such as the Euler–Maruyama method to generate a high-resolution map.

[0064] In step 2, the specific configuration of the network parameters is described as follows: the input map is divided into 128×128 patches, the batch size is set to 16, and the Adam optimizer is used for training with hyperparameters β1=0.9, β2=0.99, and the initial learning rate is set to 1×10 -4 , and decay the learning rate to half of its original value after every 200,000 iterations; the entire network was trained for a total of 500,000 iterations. The training process was run on an RTX 4090 GPU with 24GB of video memory and took about 4 days to complete, ensuring that the model achieved a good balance between stability and convergence performance.

[0065] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation methods of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for suppressing horizontal interference of geological radar profile based on diffusion model, characterized by The method for suppressing horizontal interference of the ground penetrating radar profile is carried out according to the following steps: (1) Construct a standardized GPR horizontal interference dataset: collect measured profile samples and perform mean subtraction preprocessing to select appropriate data; at the same time, perform forward simulation based on FDTD to generate simulation samples containing different road structures and underground anomaly shapes to form a unified standard dataset; (2) Building the diffusion model network structure: The residual module is fully introduced into the original Unet framework, a lightweight Agent Attention mechanism is inserted between the residual modules of each layer, and a spatial attention connection structure is set in the middle of the network; a noise scheduling strategy based on Cosineθ is adopted in the forward diffusion process; (3) Performing model training: Using the constructed data set to train the improved diffusion model, a model for horizontal interference suppression is obtained; (4) Complete interference suppression reasoning: Input the GPR profile to be processed into the trained model to obtain the output result after interference suppression.

2. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that The specific steps of constructing the data set in step 1 include: Step 1: Obtain measured samples: Use ground penetrating radar with a center frequency of 400MHz or 100MHz to detect urban roads, with a sampling interval of 0.023m; perform mean subtraction processing on the collected original GPR profiles, and select samples with better processing effects; Step 2: Generate simulation samples: Based on the FDTD method, forward modeling is performed to construct a heterogeneous underground model containing different particles and cavity structures. A variety of simulation profile images are generated by setting several random control parameters. The image containing the defect is the noise sample, and the profile obtained by subtracting the GPR profile without the defect from the GPR profile with the defect is considered the GroundTruth sample. Step 3: Summarize and organize the measured samples and simulated samples to construct a standardized GPR profile horizontal interference data set; then divide the samples into training set, test set and validation set to provide data support for the training and validation of the diffusion model.

3. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that : The residual module in step 2 consists of two layers of 3×3 convolution and ReLU activation function, and the feature learning effect is enhanced through jump connection.

4. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that : In step 2, the Agent Attention module introduces the agent vector A and constructs a four-element attention mechanism (Q, A, K, V).

5. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that : In step 2, the spatial attention module performs maximum pooling and average pooling operations on the input feature map in the channel dimension and then splices it, and generates a spatial attention map through 7×7 convolution and Sigmoid function.

6. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that The loss function of the network in step 2 adopts the maximum likelihood learning strategy to improve training stability. To address the difficulty in predicting the instantaneous noise distribution in the diffusion model, the optimization objective is set to minimize the Euclidean distance between the reconstructed states: where γ t represents positive weight, According to the current state x t and noise estimation network The optimal previous state estimate obtained by the closed-form calculation is calculated by the analytical expression to ensure that the network can more accurately fit the intermediate state in the back-diffusion process. T Sampling begins, and the diffusion stochastic differential equation (SDE) is iteratively solved by combining numerical solutions such as Euler–Maruyama to generate a high-resolution map.

7. The method for suppressing horizontal interference of geological radar profile based on diffusion model according to claim 1 is characterized in that The specific configuration of the network parameters in step 2 is as follows: the input map is divided into 128×128 patches, the batch size is set to 16, and the Adam optimizer is used for training, with hyperparameters β1=0.9, β2=0.99, and the initial learning rate is set to 1×10 -4 , and decay the learning rate to half of its original value after every 200,000 iterations; the entire network was trained for a total of 500,000 iterations. The training process was run on an RTX 4090 GPU with 24GB of video memory and took about 4 days to complete, ensuring that the model achieved a good balance between stability and convergence performance.

Images