Super-resolution generation method for BScan image of ground penetrating radar

Abstract

The invention belongs to the field of ground penetrating radar detection and identification, and particularly relates to a super-resolution generation method for a B-Scan image of a ground penetrating radar. The method comprises the following steps: setting geometric and physical parameter distribution of a stratum, setting an excitation source and a detection path, and obtaining B-Scan images containing physical information of the stratum of different stratum media; a direct coupling signal in the B-Scan image is removed through an SVD method, an original high-resolution B-Scan image data pair and an original low-resolution B-Scan image data pair are established, and data input is provided for neural network training; designing a cost function related to the stratum depth, and carrying out network training; and carrying out super-resolution image generation of a given B-Scan image by using the super-resolution prediction network. According to the method, super-resolution generation of the B-Scan image of the underground target can be effectively carried out, more accurate input is provided for underground target inversion, and the requirement for underground target inversion with smaller granularity is met.

Description

Technical Field

[0001] This invention belongs to the field of radar detection technology, specifically, it relates to a super-resolution generation method for ground-penetrating radar B-Scan images. Background Technology

[0002] Ground penetrating radar (GPR) is a non-destructive testing technology that uses high-frequency electromagnetic waves to detect underground structures and targets. Its core principle is to emit electromagnetic pulses into the ground and receive reflected waves from the interfaces of different underground media. By analyzing the amplitude, waveform, and time delay of the reflected waves, the location, depth, and nature of underground targets can be inferred.

[0003] The effectiveness of GPR detection is limited by its resolution, including vertical resolution (the minimum distance between two adjacent reflective surfaces) and horizontal resolution (the minimum distance between two adjacent lateral targets). It is constrained by factors such as the transmitted pulse width, the antenna center frequency, and the attenuation characteristics of the underground medium. Increasing the resolution usually means using a higher frequency antenna, but high-frequency signals attenuate faster underground, and the detection depth will be greatly reduced. There is an inherent contradiction between GPR resolution and detection depth.

[0004] Currently, resolution enhancement methods for GPR (Gas Reflection Profile) echo B-Scan images (radar echo profiles) mainly include two approaches: longitudinal resolution enhancement and lateral resolution enhancement. Longitudinal resolution enhancement techniques include deconvolution, spectral whitening and equalization, Q-value compensation, and stepped frequency synthesis. Deconvolution assumes the subsurface reflection coefficient sequence is white noise and designs a filter to minimize the mean square error of the output signal. This effectively compresses the wavelet, but requires accurate estimation of the wavelet shape. Spectral whitening and equalization make the energy distribution of the signal more uniform across the entire frequency band, thereby widening the effective bandwidth. By designing a filter that is the reciprocal of the signal spectrum, the amplitude of the high-frequency and low-frequency components is increased, making the spectrum "flatter." However, this requires the use of bandpass filtering to avoid amplifying noise outside the frequency band. Q-value compensation compensates for high-frequency energy loss exponentially based on depth or time, performing amplitude or phase compensation to improve B-Scan image quality. Stepped frequency synthesis (B-Scan) does not emit short pulses, but instead emits a series of single-frequency continuous waves with progressively increasing frequency steps. By synthesizing a large frequency span, it achieves super-resolution of B-Scan images. Lateral resolution enhancement techniques include migration imaging, synthetic aperture radar (SAR), and super-resolution spectral estimation. Migration imaging, based on the Huygens-Fresnel principle, propagates the recorded wavefield back underground. It can handle complex structures well, significantly reduce diffraction waves, and improve lateral focusing capability. SAR technology uses multichannel data collected by GPR during its movement to synthesize a virtual large-aperture antenna. It uses coherent superposition of signals received from the same target at different locations to narrow the beamwidth. Super-resolution spectral estimation assumes the signal consists of a finite number of sine waves and uses a parametric model instead of the traditional Fourier transform to provide higher frequency / spatial resolution than FFT. Although these methods have achieved good results in certain special scenarios, they generally suffer from problems such as noise sensitivity, poor real-time performance, and low contrast of weak target signals, failing to meet the increasingly complex engineering application requirements of ground-penetrating radar. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a super-resolution generation method for ground-penetrating radar echo B-Scan images with fast generation speed and high prediction accuracy. This method is based on convolutional neural networks (CNN) to predict the super-resolution distribution of low-resolution echo images, thereby enhancing the details of the formation. The final results show that this method can effectively carry out super-resolution generation of B-Scan images of subsurface targets, providing more accurate input for subsurface target inversion and meeting the needs of subsurface target inversion with smaller granularity.

[0006] The technical solution of the present invention is described in detail below.

[0007] This invention provides a super-resolution generation method for ground-penetrating radar B-Scan images. It utilizes a neural network to predict the super-resolution distribution of low-resolution echo images, thereby enhancing stratigraphic details. The specific steps are as follows: Step 1: Set the distribution of formation geometry and physical parameters, place the excitation source and probe path, and obtain the B-Scan image; According to the simulation requirements, the center frequency, bandwidth, polarization and other observation parameters of the radar's working electromagnetic waves are set. Then, the scattered echoes of different strata are simulated to form B-Scan images of different strata media (containing the physical information of the strata). Step 2: Use the SVD method to remove direct-coupled signals from the B-Scan image, and retain high-frequency information of the formation by setting a threshold. Then, form a low-resolution dataset by weighted averaging to establish the original high-resolution data and low-resolution data pair, providing data input for neural network training. For the simulated B-Scan grayscale image, it is first converted into an RGB three-channel image with a [0~255] distribution. Then, SVD decomposition is performed on the data of each channel, and the straight wave signal with a large energy proportion is filtered out by calculating the cumulative probability distribution of the singular values. (1) In the formula, For B-Scan images, The row vectors and column vectors are obtained from the SVD decomposition, respectively. For singular value matrices, calculate... Cumulative probability distribution: (2) In the formula, for The non-zero element number; The number of all non-zero elements; a threshold is maintained based on a preset probability. Determined to be retained The number of non-zero elements; then obtain the result using formula (1). Image after removing direct-coupled waves (Only includes the scattered echo image distribution of underground target echoes). To ensure image contrast, [the image will be...]. Pixel mean normalized to The mean value of the corresponding channel. The data from the three channels are combined to form an image with direct coupling removed. Pairing different formation media and This forms trainable data labels; Step 3: Design a cost function related to the formation depth and conduct network training; The super-resolution network model uses three CNN layers as feature extraction layers, one transposed convolutional layer as an upsampling layer, and finally uses Sigmoid normalization to the [0~1] interval to conform to the normalized B-Scan pixel distribution. The network model is as follows: Figure 2 As shown. The weighted cost function is defined as: (3) In the formula, Weighting coefficients for different locations in the image; The number of pixels in the image; The grayscale images with direct-coupled waves removed are respectively in The pixel values ​​of the network-predicted output image and the pixel values ​​of the network-predicted output image.

[0008] This invention addresses the physical characteristics of ground-penetrating radar echoes by employing a method of reducing the weight of direct-coupled signals and strengthening the weight of deep-target echoes. Specifically: (4) In the formula, This provides information on formation thickness. The training data from step 2 is input into the network for training, resulting in a super-resolution prediction network.

[0009] Step 4: Use the network model trained in Step 3 to generate super-resolution images of the target B-Scan image under a certain stratum.

[0010] First, the B-Scan image of a target under a certain stratum is used to remove direct-coupled signals using the SVD method, downsampled to a low-resolution image, and then input into a trained super-resolution prediction network to obtain a predicted high-resolution image.

[0011] The present invention relates to direct-coupled wave filtering of B-Scan images, construction of accurate prediction network models from low-resolution images to high-resolution images, and super-resolution image inference of B-Scan images; compared with the prior art, the beneficial effects of the present invention are as follows: (1) The method of the present invention is based on high-value deep target echo information, which can better predict the super-resolution signal distribution of weak target reflection signals;

[0012] (2) The method of the present invention uses simulation data to develop the network, and retains the contrast of the signal and the echo intensity of targets in different strata through data calibration, which can better adapt to the needs of different scenarios. Attached Figure Description

[0013] Figure 1 Flowchart of the generation method of this invention.

[0014] Figure 2 The neural network model used.

[0015] Figure 3 This is a two-dimensional simulation medium distribution map of a certain stratum.

[0016] Figure 4 The target B-Scan image and the processed high and low resolution images as well as the high resolution image generated by the network.

[0017] Figure 5 The results are directly predicted from B-Scan images of new scenes on lower resolution images.

[0018] Figure 6 A-Scan curves for intermediate locations in different B-Scan images of a new scene. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments and accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this invention provides a super-resolution generation method for ground-penetrating radar echo B-Scan images, combining noise suppression, direct-coupled wave cancellation, and fast neural network prediction processes. The specific steps are as follows:

[0021] Step 1: Set the distribution of formation geometry and physical parameters, place the excitation source and probe path, and obtain the B-Scan image.

[0022] The center frequency, bandwidth, polarization, and other observation parameters of the radar's electromagnetic waves are set according to the simulation requirements. Then, the scattered echoes from different strata are simulated, with the signal receiving position being the same as the transmitting position. The reflected signal is a Gaussian-modulated broadband signal. The simulated received echo signal at each position forms an A-Scan echo. Multiple A-Scan images formed along the observation direction are synthesized into a B-Scan image of the current stratum distribution. In a specific embodiment of the present invention, the observation parameters include: center frequency: 1.5 GHz, bandwidth: 0.2 GHz, polarization: electric field vector direction, using TM polarization, with the electric field direction perpendicular to the scanning plane.

[0023] Step 2: Use the SVD method to remove direct-coupled signals from the B-Scan image and retain high-frequency formation information by setting a threshold.

[0024] For the singular values ​​after SVD decomposition, low-frequency energy is truncated according to 75% cumulative probability density, and the remaining 25% of energy is used to recover high-frequency echoes from deeper layers. To ensure the contrast of the decoupled signal and the intensity correlation of the formation medium echo, the signal is normalized according to the mean of the B-Scan image before removing direct coupling, resulting in a B-Scan image with strong target correlation and high contrast. Then, the signal is scaled according to a scaling factor. Weighted average downsampling is performed. Specifically: (5)

[0025] Step 3: Train the network parameters using high- and low-resolution B-Scan datasets simulated with different stratigraphic parameters. The neural network model structure is as follows: Figure 2 As shown, the solver uses the Adam method, with an initial training step size of 0.01, and employs an exponentially decreasing step size training method (the step size decreases by 0.002 every 100 iterations). Training terminates after 3000 iterations or when the cost function is less than 0.001.

[0026] Step 4: Change the target geometry and medium parameters in the formation, and simulate the new scene to obtain a B-Scan image under the same excitation. Then, SVD is used to remove the direct-coupled signal, followed by downsampling to obtain a low-resolution B-Scan image of the new scene. The input is fed into the network trained in step 3 to generate a new high-resolution image, and the calculation is performed. and The cost function is used to evaluate the network's generalization ability. If the error meets the requirements, the trained model is used in the real-world scenario to directly predict the super-resolution results of the measured B-Scan images; otherwise, data from the new scenario is added to the training set, and the network is retrained to enhance its prediction accuracy for the new scenario.

[0027] The following are specific examples.

[0028] Example 1

[0029] Assuming the radar's operating center frequency is 1.5 GHz, the detector's lateral movement interval is 2.5 mm, and the simulation area is a 1m × 1m region. Simulated targets include PVC pipes, steel bars, and cavities of varying radii. The scene settings are as follows: Figure 3 As shown.

[0030] Ignoring formation loss, with an echo signal interval of 0.1 ns, a total of 400 data points were collected, moved 10 cm laterally, forming a 400×400 sampling dataset. With fixed surface formation, layered formations, soil layer thickness, and medium parameters, the size and location of PVC pipes, reinforcing bars, and voids were randomly moved to simulate and obtain multiple B-Scan images, which were used to establish the network training dataset. The training scene consisted of 3000 B-Scan images.

[0031] First, the ground direct-coupled signal is removed using the SVD method to obtain the target echo image under the layered strata. Then, a low-resolution image is formed by a 4x weighted downsampling sliding window. Finally, the obtained high- and low-resolution images are used to train the network model, where the training set and validation set are dynamically selected in a 9:1 ratio, and the cutoff error is set to 0.001.

[0032] Figure 4 The prediction results for B-Scan images in a certain validation set are presented. The generated high-resolution image has an error of 0.0013, which shows good prediction accuracy.

[0033] Example 2

[0034] By changing the placement of targets such as PVC pipes, steel bars, and voids (different from the 3000 scenarios in Example 1), the B-Scan direction corresponding to the new scenario is re-simulated. Following step 2, the corresponding downsampled low-resolution image is obtained, and then input into the network model to directly predict the new high-resolution image, such as... Figure 5 As shown. The A-Scan contrast curves at different midpoints of the B-Scan images are as follows. Figure 6 As shown, the high-resolution images generated by the network match well with the high-resolution images after removing direct coupling, especially showing good accuracy in the deep subsurface areas of interest.

Claims

1. A super-resolution generation method for ground-penetrating radar B-Scan images, characterized in that, A neural network is used to achieve super-resolution generation of low-resolution B-Scan images, thereby enhancing stratigraphic details. The specific steps are as follows: Step 1: Set the distribution of formation geometry and physical parameters, set the excitation source and probe path, and obtain B-Scan images containing formation physical information for different formation media. Step 2: Use the SVD method to remove the direct-coupled signal in the B-Scan image, and retain the high-frequency information of the formation by setting a threshold. Then, form the corresponding low-resolution image according to the weighted average method, establish the data pair of the original high-resolution image and the low-resolution image, and provide data input for the neural network training. Step 3: Design a cost function related to formation depth and conduct network training. The super-resolution prediction network model uses three CNN layers as feature extraction layers, one transposed convolutional layer as an upsampling layer, and finally uses Sigmoid to normalize to the [0~1] interval to conform to the normalized B-Scan pixel distribution; a weighted cost function is defined, and the network is trained using the paired data obtained in step 2; Step 4: Use the super-resolution prediction network trained in Step 3 to generate super-resolution images of arbitrary underground target B-Scan images. First, the direct-coupled signal is removed from the low-resolution B-Scan image of a target under a certain stratum using the SVD method. Then, it is input into the trained super-resolution prediction network to obtain the predicted high-resolution image.

2. The super-resolution generation method according to claim 1, characterized in that, In step 1, radar observation parameters are set according to simulation requirements to simulate the scattered echoes of different strata and form corresponding B-Scan images; the observation parameters include center frequency, bandwidth and polarization mode.

3. The super-resolution generation method according to claim 1, characterized in that, In step 2, the original B-Scan image is first converted into an RGB image. Then, SVD decomposition is performed on the data of each channel. By calculating the cumulative probability distribution of singular values, straight-wave signals with a large energy proportion are filtered out, resulting in the image after removing the straight-coupled waves. It only includes the image distribution of scattered echoes from underground targets. To ensure image contrast, The pixel mean of each channel is normalized to The mean value of each channel is used; the data from the three channels are synthesized to form a grayscale image with direct-coupled waves removed. ; Pairing different formation media and This forms trainable data labels.

4. The super-resolution generation method according to claim 3, characterized in that, In step 3, the cost function is: (3) In the formula, Weighting coefficients are applied to different locations in the image. The number of pixels in the image. The grayscale images with direct-coupled waves removed are respectively... The pixel values ​​of the network-predicted output image and the pixel values ​​of the network-predicted output image.

5. The super-resolution generation method according to claim 4, characterized in that, In view of the physical characteristics of ground-penetrating radar echoes, a method is adopted to reduce the weight of direct-coupled signals and strengthen the weight of deep target echoes; Specifically: (4) In the formula, This is information about the thickness of the formation.

Images