Ground penetrating radar high-quality imaging method based on image recovery diffusion model

Abstract

The invention discloses a ground penetrating radar high-quality imaging method based on an image recovery diffusion model, and the method comprises the steps: constructing a simulation image through an open source simulation tool gprMax based on a finite difference time domain method (FDTD), and constructing a contrast high-frequency clutter-free and low-frequency clutter data set through preprocessing and fusing actual measurement clutter simulation; a double-layer routing attention mechanism is introduced on a basic model to enhance association between features, conditional information is fully utilized, discrete wavelet transform is used to effectively separate low-frequency and high-frequency information, information loss is reduced, and the high-frequency information is enhanced through an adaptive wavelet transform high-frequency enhancement module. According to the method, good improvement is achieved on a synthetic data set, acceptable effects are achieved on multiple evaluation indexes including root mean square error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), perceptual image similarity (LPIPS) and depth image structure and texture similarity (DISTS), and a certain underground target structure can also be recovered in actual measurement data.

Description

Technical Field

[0001] The present invention belongs to the technical field of ground penetrating radar imaging, and specifically relates to a high-quality ground penetrating radar imaging method based on an image restoration diffusion model. Background Technique

[0002] Ground penetrating radar technology, as a non-destructive underground detection tool, is widely used in multiple fields, such as the inspection of roads, tunnels, and bridges. Especially in complex underground environments, the detection of underground structures is crucial. Ground penetrating radar detects underground targets by transmitting electromagnetic waves and analyzing the reflected signals. However, in practical applications, due to complex underground media and environmental interference, ground penetrating radar images are often affected by strong clutter and low signal-to-noise ratio, which makes it very difficult to identify targets and improve the resolution of images.

[0003] Traditional ground penetrating radar data processing methods mainly rely on techniques such as signal denoising, filtering, and echo cancellation. These methods have certain effects when dealing with simple interference, but when facing a strong clutter background or complex underground environment, the clutter removal effect is often poor, and the high-frequency details in the image cannot be restored, resulting in blurred images and loss of important structural information. Specifically, the performance of the two tasks of clutter suppression and image resolution improvement is restricted in complex environments. Especially when dealing with underground targets covered by strong clutter, it is very difficult for traditional methods to achieve accurate image reconstruction and high-resolution restoration.

[0004] In recent years, deep learning, especially generative adversarial networks and other convolutional neural network models, have gradually become the main means to solve the clutter removal and super-resolution problems of ground penetrating radar images. These methods can automatically extract important information in the image by learning features in large-scale datasets. However, although generative adversarial networks have shown strong capabilities in some tasks, their performance in dealing with complex ground penetrating radar images is still limited. Especially in the aspect of high-frequency detail restoration, although some methods can remove background clutter, the quality of detail restoration is often low. Especially in the case of blurred target edges and distorted image textures, the generated images are prone to artifacts and blurring, affecting the accuracy of target recognition.

[0005] As an emerging generative model, diffusion models have achieved remarkable results in the fields of image denoising and super-resolution in recent years. Different from traditional generative models, diffusion models generate high-quality images through a step-by-step reverse denoising process, which can effectively avoid the problem of training instability while retaining image structure and details. Although existing diffusion models have shown great potential in the tasks of clutter removal and image resolution improvement, they still face problems such as high computational cost, many time steps, and insufficient utilization of conditional information, which limit their popularization in practical applications.

[0006] To overcome the deficiencies of the prior art, the present invention proposes a ground penetrating radar high-resolution imaging method based on an image restoration diffusion model.

[0007] Technical comparison with the patent CN118587087A "A ground penetrating radar high-resolution imaging method based on a diffusion model"

[0008] Patent CN118587087A uses a conditional diffusion model. In view of the problem that in a complex underground environment, strong background clutter and low signal-to-noise ratio often lead to severely limited target recognition and resolution, and it is difficult to restore high-frequency details, this patent adopts a diffusion model based on image restoration and innovatively improves it by integrating a double-layer routing attention mechanism, discrete wavelet transform, and an adaptive wavelet transform high-frequency enhancement module.

[0009] Patent CN118587087A uses the reverse process of the diffusion model to recover a high-resolution image from noise, consuming a large amount of computing resources and having a long time step. This patent, on the other hand, directly recovers from a low-quality image through conditional guidance, greatly reducing the consumption of computing resources. Only 4 time steps are required to estimate the prior representation of image restoration in the diffusion model part, improving the computing efficiency.

[0010] Patent CN118587087A is for the application of the diffusion model in the ground penetrating radar high-resolution imaging task, while this patent is for the tasks of ground penetrating radar clutter removal, high-resolution imaging, reducing information loss, and optimizing high-frequency details.

[0011] Technical comparison with the patent CN119107236A "Ground penetrating radar Bscan image super-resolution processing method and device"

[0012] Patent CN119107236A uses a conditional denoising diffusion model with a residual self-attention Unet network as the backbone network, with a large overall computational amount. Considering the problem of large computational consumption of the traditional diffusion model, this patent adopts a diffusion model based on image restoration (DiffIR) as the basic architecture, and only 4 time steps are required to estimate the prior representation of image restoration in the diffusion model part, greatly reducing the computational amount.

[0013] When making the simulation data set in patent CN119107236A, measured noise was not introduced. Considering the complex phenomenon of clutter and noise in the measured environment, this patent introduced 50 pieces of measured clutter and noise backgrounds when making the low-frequency data to simulate the real situation.

[0014] The high-frequency data in patent CN119107236A is obtained by the preprocessing module through bilinear scaling and bicubic interpolation, while the high-frequency data in this patent is collected and generated by ground penetrating radars with different frequencies on both the simulation data set and the measured data set.

[0015] Patent CN119107236A did not conduct index evaluation, simulation, and effect display of the measured dataset, while this patent presents the generation results of the simulation dataset and the measured dataset and gives the evaluation data of five important indicators.

[0016] Based on the improved image restoration diffusion model, the present invention has completed the tasks of ground penetrating radar clutter removal and high-resolution imaging. Summary of the Invention

[0017] To solve the above problems and improve the imaging quality of ground penetrating radar, the present invention proposes a high-quality imaging method for ground penetrating radar based on an image restoration diffusion model. By using gprMax to construct a simulation image, a control high-frequency clutter-free and low-frequency clutter-containing dataset is constructed through preprocessing and fusing measured clutter simulation. A double-layer routing attention mechanism is introduced on the basic model to strengthen the association between features. The discrete wavelet transform is used to effectively separate low-frequency and high-frequency information, reduce information loss, and the high-frequency information is enhanced by an adaptive wavelet transform high-frequency enhancement module.

[0018] To achieve the above object, the technical solution adopted by the present invention is:

[0019] A high-quality imaging method for ground penetrating radar based on an image restoration diffusion model, comprising the following steps:

[0020] S1. In the gprMax simulation environment, three typical underground scenes of roads, tunnels, and mines are constructed. In each scene, the targets are distributed in different quantities and positions. Subsequently, a 900 MHz low-frequency ground penetrating radar and a 1500 MHz high-frequency ground penetrating radar are used to collect radar echo signals for these scenes respectively, generating a 3000-group control radar echo signal dataset. This dataset presents the signal characteristics collected by low-frequency and high-frequency radars in the same scene only through the difference in frequency.

[0021] S2. For the dataset in S1, first, it is cropped and rescaled, and then the simulation background clutter is removed from the processed data to obtain a target image without background clutter, constructing a control dataset containing low-frequency clutter and high-frequency clutter-free, providing a basis for subsequent clutter removal and resolution improvement tasks.

[0022] S3. Based on the image restoration diffusion model, first, a double-layer routing attention mechanism is incorporated into the compact prior vector extraction module to enhance the association of global semantic information. In the dynamic image restoration module, the discrete wavelet transform is introduced to replace the traditional downsampling operation to reduce information loss, and the separated high-frequency information is enhanced by the adaptive wavelet transform high-frequency enhancement module. In the upsampling stage, the inverse discrete wavelet transform is used to fuse the low-frequency and high-frequency information again. The training process of the model includes two stages. In the pre-training stage, the compact prior vector extraction module and the dynamic image restoration module are mainly trained. In the training stage, the dynamic image restoration module and the diffusion model are jointly trained. During the prediction process, the reverse processes of the compact prior vector extraction module, the dynamic image restoration module, and the diffusion model are comprehensively utilized to perform clutter removal on the ground penetrating radar image and achieve high-resolution improvement;

[0023] S4. Tests are carried out on the synthetic dataset and the measured dataset, and evaluations are conducted on five key evaluation indicators: mean square error MSE, peak signal-to-noise ratio PSNR, structural similarity index SSIM, perceptual image similarity LPIPS, and depth image structure and texture similarity DISTS, and the underground target structure is restored in the actual measurement data.

[0024] As a further improvement of the present invention, in step S2, the dataset in S1 is preprocessed, and the specific steps are as follows:

[0025] S21. Differential processing to obtain the target image:

[0026] The target object is removed by modifying the simulation.in file, and the simulation is carried out again. Then, the echo data with the target object removed is subjected to differential processing with the original simulation result to obtain the target image with the clutter background removed;

[0027] S22. Adding measured clutter and noise to the low-frequency image:

[0028] To simulate the complex underground structure in the real environment, 50 pieces of measured clutter data are selected and fused with the target image of the low-frequency clutter background removed. The measured clutter and the low-frequency clutter background removed image are weighted and fused according to the ratio of 0.6:0.4 to generate a low-frequency dataset with clutter.

[0029] As a further improvement of the present invention, in step S3, the model is based on the image restoration diffusion model, and the specific steps are as follows:

[0030] S31. Incorporating a double-layer routing attention mechanism into the compact prior vector extraction module:

[0031] In the compact prior vector extraction module, a double-layer routing attention mechanism is incorporated. The double-layer routing attention mechanism module is placed before the main convolutional network as a guide for input features. First, the input low-quality image and high-quality image undergo pixel rearrangement operations for subsequent processing. Then, these two images are concatenated in the channel dimension to form the input of the double-layer routing attention mechanism module. This module is processed using the following design: at the first level, an inter-region similarity map is constructed and pruned using the top-k routing algorithm to retain the most relevant global context information; at the second level, token-to-token attention calculation is performed within the selected regions to form a sparse and adaptive attention map. By placing this module before the main convolution, global semantic information can be fused before the feature compression stage, and it is ensured that the spatial structure and target saliency in the image are better retained and enhanced.

[0032] S32. In the dynamic image restoration module, discrete wavelet transform and adaptive wavelet transform high-frequency enhancement are introduced:

[0033] First, the input image is decomposed into frequency sub-bands through discrete wavelet transform to obtain the low-frequency sub-band and high-frequency sub-bands. In the processing of the high-frequency sub-bands, the adaptive wavelet transform high-frequency enhancement module plays an important role.

[0034] This module is processed through two branches:

[0035] Standard 3×3 convolution is used to capture local features, while dilated convolution expands the receptive field and enhances the model's sensitivity to detailed textures.

[0036] The convolution results are fused and compressed through 1×1 convolution and then returned through a residual connection.

[0037] Next, a learnable two-dimensional Gaussian filter is used to suppress random clutter while retaining detailed textures.

[0038] Next, channel attention and spatial attention are adopted to improve the perceptual quality and perform context modulation through global average pooling.

[0039] In the upsampling stage, a symmetric inverse discrete wavelet transform-based upsampling mechanism is adopted. Each level receives four frequency sub-bands from the encoder, reconstructs the spatial resolution through inverse discrete wavelet transform, and the reconstructed features pass through 1×1 convolution and are fused with the encoder features through skip connections. This process realizes the effective fusion of low-frequency and high-frequency information, enables the reconstruction of image details, and maximally retains the structure and texture of the target.

[0040] As a further improvement of the present invention, good improvement has been achieved on the synthetic dataset in step S4, and certain underground target structures can also be restored in actual measurement data. The specific steps are as follows:

[0041] S41, Synthetic dataset test:

[0042] First, tests were carried out on the synthetic dataset, and corresponding tests were carried out on image quality and various evaluation metrics, including root mean square error, peak signal-to-noise ratio, structural similarity index, perceptual image similarity, and depth image structure and texture similarity;

[0043] S42, Measured dataset test:

[0044] On the actual measurement dataset, in the complex clutter background of the real scene, background clutter and noise are suppressed and target information is restored.

[0045] Beneficial effects: The benefits brought by the present invention and the achieved indicators.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. It provides a new solution for solving the dual tasks of ground penetrating radar clutter removal and high-resolution improvement;

[0048] 2. The image restoration diffusion model is used as the basic model, and the diffusion model part is used to estimate the prior representation of image restoration. This process only requires 4 time steps, improving the calculation efficiency;

[0049] 3. The double-layer routing attention mechanism is incorporated into the model compact prior vector extraction module to enhance the association of global semantic information and make full use of conditional information;

[0050] 4. In the dynamic image restoration module, the discrete wavelet transform is introduced to replace the traditional downsampling operation to reduce information loss, and the separated high-frequency information is enhanced by the adaptive wavelet transform high-frequency enhancement module to optimize the high-frequency details;

[0051] 5. It provides a solid foundation for downstream tasks such as target recognition, improving the accuracy and reliability of subsequent tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is the implementation process of a ground penetrating radar high-quality imaging method based on an image restoration diffusion model of the present invention patent;

[0053] Figure 2 are the steps of dataset processing;

[0054] Figure 3It is the structure diagram of the discrete wavelet transform, inverse discrete wavelet transform, and adaptive wavelet transform high-frequency enhancement module innovatively improved in the dynamic image restoration module;

[0055] Figure 4 It is the display of the effect of the present invention on the synthetic dataset;

[0056] Figure 5 It is the display of the effect of the present invention on the measured dataset. Specific implementation manners

[0057] The following further describes the present invention in detail in conjunction with the accompanying drawings and specific implementation manners:

[0058] The implementation process of this aspect is as Figure 1 shown.

[0059] The present invention is realized through the following technical solutions:

[0060] S1. In the gprMax simulation environment, three typical underground scenarios of roads, tunnels, and mines were constructed. In each scenario, the targets were distributed with different quantities and positions. Subsequently, a 900 MHz low-frequency ground-penetrating radar and a 1500 MHz high-frequency ground-penetrating radar were used to collect radar echo signals for these scenarios respectively, generating a control radar echo signal dataset of 3000 groups. This dataset only presents the signal characteristics collected by the low-frequency and high-frequency radars in the same scenario through the difference in frequency.

[0061] S2. For the dataset in S1, first, it was cropped and rescaled, and then the processed data was subjected to simulation background clutter removal to obtain a target image without background clutter. In order to more accurately simulate the complex underground structures in the real environment, 50 measured clutter background images were selected and fused with the low-frequency ground-penetrating radar images at a ratio of 0.6:0.4. This process is as Figure 2 shown, constructing a control dataset containing low-frequency band clutter and high-frequency clutter-free, providing a basis for subsequent clutter removal and resolution improvement tasks.

[0062] S21. Removing the clutter background from the simulation image:

[0063] The target objects were removed by modifying the simulation.in file, and then the simulation was re-performed. Then, the echo data with the target objects removed was subjected to differential processing with the original simulation results, thereby obtaining a target image with the clutter background removed.

[0064] S22. Adding measured clutter to the low-frequency image:

[0065] To simulate complex underground structures in a real environment, 50 pieces of measured clutter data were selected and fused with target images with the low-frequency clutter background removed. To better reflect the interference characteristics of the actual environment, the measured clutter and the low-frequency clutter-removed background image were weighted and fused at a ratio of 0.6:0.4 to generate a data set with clutter in the low-frequency band.

[0066] S3. Based on the image restoration diffusion model (DiffIR), first, a double-layer routing attention mechanism was incorporated into the compact prior vector extraction module to enhance the association of global semantic information. In the dynamic image restoration module, the discrete wavelet transform was introduced to replace the traditional downsampling operation to reduce information loss, and the separated high-frequency information was enhanced by the adaptive wavelet transform high-frequency enhancement module. In the upsampling stage, the inverse discrete wavelet transform was used to fuse the low-frequency and high-frequency information again. The training process of the model includes two stages. In the pre-training stage, the compact prior vector extraction module and the dynamic image restoration module were mainly trained. In the joint training stage, the dynamic image restoration module and the diffusion model were jointly trained. In the prediction process, the reverse processes of the compact prior vector extraction module, the dynamic image restoration module, and the diffusion model were comprehensively utilized to remove clutter from the ground-penetrating radar image and achieve high-resolution improvement.

[0067] S31. Incorporating a double-layer routing attention mechanism into the compact prior vector extraction module:

[0068] In the compact prior vector extraction module, a double-layer routing attention mechanism was incorporated. This mechanism aims to enhance the extraction ability of global semantic information and improve the efficiency and accuracy of feature extraction through dynamic routing. The double-layer routing attention mechanism module is placed before the main convolutional network as a guide for input features. First, the input low-quality image and high-quality image undergo pixel rearrangement operations for subsequent processing. Then, these two images are concatenated in the channel dimension to form the input of the double-layer routing attention mechanism module. The module is processed using the following design: at the first level, a similarity map between regions is constructed and cropped using the top-k routing algorithm to retain the most relevant global context information; at the second level, token-to-token attention calculation is performed within the selected regions to form a sparse and adaptive attention map. By placing this module before the main convolution, global semantic information can be fused before the feature compression stage, and the spatial structure and target saliency in the image can be better retained and enhanced.

[0069] S32. Introducing discrete wavelet transform and adaptive wavelet transform high-frequency enhancement in the dynamic image restoration module:

[0070] In the dynamic image restoration module, a discrete wavelet transform and an adaptive wavelet transform high-frequency enhancement module are introduced. The purpose of this module is to improve the detail restoration and target recognition capabilities of images through frequency-domain enhancement and high-frequency information reconstruction. First, the input image is decomposed into frequency sub-bands by the discrete wavelet transform to obtain low-frequency and high-frequency sub-bands, which carry different frequency information of the image respectively. In particular, the high-frequency sub-band contains important edges and abnormal features, which can provide key information for interpreting underground targets in ground penetrating radar images. In the processing of the high-frequency sub-band, the adaptive wavelet transform high-frequency enhancement module plays an important role. This module is processed through two branches: a standard 3×3 convolution is used to capture local features, while the dilated convolution expands the receptive field and enhances the model's sensitivity to detailed textures. The convolution results are fused and compressed through a 1×1 convolution, and then returned through a residual connection. Next, a two-dimensional Gaussian filter is used to suppress random clutter while retaining detailed textures. To further improve the perceptual quality of the image, a dual attention mechanism is adopted, including channel attention (using global average pooling and max pooling combined with fully connected layers) and spatial attention (using 7×7 convolution), and global context modulation is performed through global average pooling. These operations effectively enhance the model's sensitivity to weak targets in complex scenes and help remove background clutter. In the upsampling stage, a symmetric upsampling mechanism based on the inverse discrete wavelet transform is adopted. Four frequency sub-bands received from the encoder are used to reconstruct the spatial resolution through the inverse discrete wavelet transform. The reconstructed features pass through a 1×1 convolution and are fused with the encoder features through a skip connection. This process realizes the effective fusion of low-frequency and high-frequency information, enables the reconstruction of image details, and maximally retains the structure and texture of the target. The structure of this part is as Figure 3 shown.

[0071] S4. Tests are carried out on the synthetic dataset and the measured dataset. The experimental results show that the method of the present invention not only achieves good improvement on the synthetic dataset, reaches acceptable effects on multiple key evaluation indicators, but also can restore a certain underground target structure in the actual measurement data.

[0072] S41. Synthetic dataset test:

[0073] First, the method of the present invention is tested on the synthetic dataset. The test results show that the method of the present invention performs well in terms of image quality and various evaluation indicators. Especially in the restoration of target images under the background of low-frequency band clutter, it can suppress background clutter and restore details, and the generated effect is as Figure 4 shown, and acceptable effects are achieved in the root mean square error, peak signal-to-noise ratio, structural similarity index, perceptual image similarity, and depth image structure and texture similarity. The index evaluation results are shown in Table 1.

[0074] Table 1: Test result metrics on the synthetic dataset;

[0075]

[0076] S42, Test on the measured dataset:

[0077] On the actual measured dataset, the model of the present invention also performs well and can recover a certain underground target structure. Although it is only trained on the synthetic dataset, in the complex clutter background of the real scene, the method of the present invention can effectively suppress the background noise and recover the target information. The generated image effect is as Figure 5 shown.

[0078] The above are only the preferred embodiments of the present invention, and do not constitute any other form of limitation to the present invention. Any modification or equivalent change made according to the technical essence of the present invention still belongs to the scope protected by the present invention.

Claims

1. A high-quality imaging method for ground penetrating radar based on an image restoration diffusion model, characterized in that: It includes the following steps: S1. In the gprMax simulation environment, three typical underground scenarios of roads, tunnels, and mines were constructed. In each scenario, targets were distributed with different quantities and positions. Subsequently, a 900 MHz low-frequency ground-penetrating radar and a 1500 MHz high-frequency ground-penetrating radar were used to collect radar echo signals for these scenarios respectively, generating a control radar echo signal dataset of 3000 groups. This dataset presents the signal characteristics collected by low-frequency and high-frequency radars in the same scenario only through the difference in frequency; S2. For the dataset in S1, first, it was cropped and rescaled, and then the processed data was subjected to simulation background clutter removal to obtain a target image without background clutter, constructing a control dataset containing low-frequency band clutter and high-frequency clutter-free, providing a basis for subsequent clutter removal and resolution improvement tasks; S3. Using the image restoration diffusion model as the basis, first, a double-layer routing attention mechanism was incorporated into the compact prior vector extraction module to enhance the association of global semantic information. In the dynamic image restoration module, the traditional downsampling operation was replaced by introducing discrete wavelet transform to reduce information loss, and the separated high-frequency information was enhanced through the adaptive wavelet transform high-frequency enhancement module. In the upsampling stage, the inverse discrete wavelet transform was used to fuse the low-frequency and high-frequency information again. The training process of the model includes two stages. Among them, the pre-training stage focuses on training the compact prior vector extraction module and the dynamic image restoration module, and the training stage jointly trains the dynamic image restoration module and the diffusion model. In the prediction process, the reverse processes of the compact prior vector extraction module, the dynamic image restoration module, and the diffusion model are comprehensively utilized to perform clutter removal on the ground-penetrating radar image and achieve high-resolution improvement; S4. Testing on the synthetic dataset and the measured dataset, evaluating on five key evaluation indicators of root mean square error MSE, peak signal-to-noise ratio PSNR, structural similarity index SSIM, perceptual image similarity LPIPS, and depth image structure and texture similarity DISTS, and restoring the underground target structure in the actual measurement data.

2. The high-quality imaging method of ground penetrating radar based on an image restoration diffusion model according to claim 1, characterized in that: In step S2, the dataset in S1 is preprocessed, and the specific steps are as follows: S21. Differential processing to obtain the target image: The target object was removed by modifying the simulation.in file, and the simulation was carried out again. Then, the echo data after removing the target object was subjected to differential processing with the original simulation result to obtain a target image with the clutter background removed; S22. Adding measured clutter and noise to the low-frequency image: To simulate the complex underground structure in the real environment, 50 pieces of measured clutter data were selected and fused with the target image of the low-frequency clutter background removed. The measured clutter and the low-frequency clutter background removed image were weighted and fused according to a ratio of 0.6:0.4 to generate a low-frequency dataset with clutter; 3. A ground penetrating radar high-quality imaging method based on an image restoration diffusion model according to claim 1, characterized in that: In step S3, the model uses the image restoration diffusion model as the basis, and the specific steps are as follows: S31. Incorporating a double-layer routing attention mechanism into the compact prior vector extraction module: In the compact prior vector extraction module, a two-layer routing attention mechanism is incorporated. The two-layer routing attention mechanism module is placed before the main convolutional network as a guide for input features. First, the input low-quality image and high-quality image undergo pixel rearrangement operations for subsequent processing. Then, these two images are concatenated in the channel dimension to form the input of the two-layer routing attention mechanism module, which is processed using the following design: At the first level, an inter-region similarity map is constructed and pruned using the top-k routing algorithm to retain the most relevant global context information; at the second level, token-to-token attention calculation is performed within the selected regions to form a sparse and adaptive attention map. By placing this module before the main convolution, global semantic information can be fused before the feature compression stage, and it is ensured that the spatial structure and target saliency in the image are better retained and enhanced; S32. In the dynamic image restoration module, discrete wavelet transform and adaptive wavelet transform high-frequency enhancement are introduced: First, the input image is decomposed into frequency sub-bands through discrete wavelet transform to obtain the low-frequency sub-band and high-frequency sub-bands. In the processing of the high-frequency sub-bands, the adaptive wavelet transform high-frequency enhancement module plays an important role; This module is processed through two branches: Standard 3×3 convolutions are used to capture local features, while dilated convolutions expand the receptive field and enhance the model's sensitivity to detailed textures; The convolution results are fused and compressed through 1×1 convolutions and then returned through residual connections; Next, a learnable two-dimensional Gaussian filter is used to suppress random clutter while retaining detailed textures; Next, channel attention and co-spatial attention are adopted to improve the perceptual quality and perform context modulation through global average pooling; In the upsampling stage, a symmetric inverse discrete wavelet transform-based upsampling mechanism is adopted. Each level receives four frequency sub-bands from the encoder, and the spatial resolution is reconstructed through inverse discrete wavelet transform. The reconstructed features pass through 1×1 convolutions and are fused with the encoder features through skip connections. This process realizes the effective fusion of low-frequency and high-frequency information, enabling the reconstruction of image details while maximizing the retention of the structure and texture of the target.

4. A ground penetrating radar high-quality imaging method based on an image restoration diffusion model according to claim 1, characterized in that: Good improvements are achieved in the synthetic dataset in step S4, and certain underground target structures can also be restored in actual measurement data. The specific steps are as follows: S41. Synthetic dataset test: First, tests are conducted on the synthetic dataset, and corresponding tests are performed on the image quality and various evaluation metrics, namely root mean square error, peak signal-to-noise ratio, structural similarity index, perceptual image similarity, and depth image structure and texture similarity; S42. Measured dataset test: On the actual measurement dataset, under the complex clutter background in the real scene, background clutter and noise are suppressed and target information is restored.

Images