Ground penetrating radar B-scan image super-resolution processing method and device

Abstract

The present invention discloses a method for processing ground penetrating radar data, and specifically relates to a method and device for super-resolution processing of ground penetrating radar B-scan images. The method comprises: constructing a ground penetrating radar data set, the ground penetrating radar data set comprises a plurality of original B-scan images; pre-processing each of the plurality of original B-scan images to obtain a target data set; constructing a conditional denoising diffusion model, the conditional denoising diffusion model comprises a Gaussian noise diffusion stage and an inference stage, the Gaussian noise diffusion stage comprises a residual self-attention U-net network; training the conditional denoising diffusion model based on the target data set to obtain a trained conditional denoising diffusion model; and performing super-resolution reconstruction processing on the image to be processed using the trained conditional denoising diffusion model. The present invention can achieve detail recovery and noise suppression, thereby improving the data quality and analysis effect of the B-scan image.

Description

Technical Field

[0001] The invention relates to a ground penetrating radar data processing method, and in particular to a ground penetrating radar B-scan image super-resolution processing method and device. Background Art

[0002] Ground Penetrating Radar (GPR) is an instrument that uses radar pulses to detect and image the conditions below the surface. It is a non-invasive underground detection method. It is often used to detect underground facilities such as concrete, asphalt, metal, pipelines, cables or masonry. This non-destructive method uses microwave (UHF / VHF frequency) band electromagnetic waves on the radio spectrum to detect and receive electromagnetic echoes caused by changes in electromagnetic parameters of underground objects. Under appropriate conditions, practitioners can use ground penetrating radar to detect underground objects, changes in material properties, and gaps and cracks.

[0003] Ground penetrating radar uses broadband electromagnetic waves for detection, generally in the range of 10MHz to 2.6GHz. The ground penetrating radar transmitting antenna transmits electromagnetic waves to the ground. When the electromagnetic waves encounter objects buried under the surface or reach the boundary where the dielectric constant changes, they may be reflected, refracted or scattered. The receiving antenna receives the echo and records the difference in the change of the echo signal. The principles involved are similar to seismology, except that the ground penetrating radar method uses electromagnetic energy instead of acoustic energy, and the energy of the electromagnetic wave will be reflected at the boundary where the electrical properties of the underground object change, rather than generating impedance at the boundary where the mechanical properties of the underground change like seismic energy.

[0004] The B-scan image information of the ground penetrating radar intuitively presents the characteristics of the underground structure in the form of a two-dimensional matrix. However, on the one hand, in actual measurements, due to the limitations of hardware equipment, such as receivers and transmitters, the A-scans are not completely consistent, resulting in inconsistency in the B-scan. On the other hand, due to the influence of non-uniform underground media and the efficiency in actual measurements, it is inconvenient to set a very dense spatial sampling interval for sampling. These factors cause B-scan images to usually have a low spatial resolution, which limits the subsequent signal analysis of B-scan signal processing in target detection, positioning, imaging, recognition, and electromagnetic inversion.

[0005] At present, methods based on convolutional neural networks (CNN) and super-resolution methods based on generative adversarial networks (GAN) have been applied to super-resolution reconstruction of ground penetrating radar B-scan images, but they still have certain limitations. For example, natural regression models are very expensive for high-resolution image generation, variational autoencoders and normalized flows usually produce suboptimal sample quality, and GANs require carefully designed regularization and optimization techniques to tame optimization instability and modal collapse. It can be seen that the existing technology still has the problems of insufficient spatial resolution of ground penetrating radar B-scan images, severe noise coverage and insufficient detail information. Summary of the invention

[0006] The technical problem to be solved by the present invention is that the above-mentioned methods commonly used in the prior art have the problems of insufficient spatial resolution of ground penetrating radar B-scan images, serious noise coverage and insufficient detail information. In order to solve the above-mentioned problems, the present invention provides a ground penetrating radar B-scan image super-resolution processing method and device.

[0007] The content of the present invention includes:

[0008] In a first aspect, an embodiment of the present invention provides a ground penetrating radar B-scan image super-resolution processing method, comprising:

[0009] Constructing a ground penetrating radar dataset, wherein the ground penetrating radar dataset includes a plurality of original B-scan images;

[0010] For each of the multiple original B-scan images, preprocess the original B-scan image to obtain high-resolution data, low-resolution data, and inference data corresponding to the original B-scan image;

[0011] Constructing a conditional denoising diffusion model, wherein the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, wherein the Gaussian noise diffusion stage includes a residual self-attention U-net network;

[0012] Training the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data, and inference data corresponding to the plurality of original B-scan images;

[0013] Performing super-resolution reconstruction on the image to be processed by using the trained conditional denoising diffusion model;

[0014] Wherein, the Gaussian noise diffusion stage is used for:

[0015] Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1;

[0016] Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images;

[0017] Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data;

[0018] Inputting T of the dual-channel noisy image data into the residual self-attention U-net network to perform noise estimation to obtain estimated noise;

[0019] The inference phase is used to:

[0020] Generate a pure noise image of the same size as the input image;

[0021] The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes:

[0022] Perform noise estimation on the image obtained by the i-1th iterative processing based on the residual self-attention U-net network to obtain the i-th estimated noise;

[0023] The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

[0024] Optionally, the residual self-attention U-net network includes:

[0025] The encoding stage includes a first convolutional layer, a first processing block and a residual module, wherein the first convolutional layer is used to map the data into a high-dimensional feature space, and the first processing block is used to reduce the spatial dimension, increase the number of feature channels, extract multi-scale features of the input data, and increase the abstraction level of feature representation;

[0026] An intermediate stage, comprising two residual modules with a self-attention mechanism, for extracting global context information and capturing long-distance dependencies through the self-attention mechanism;

[0027] The decoding stage includes a second processing block and a second convolutional layer, the second processing block is used to gradually restore the spatial dimension while reducing the number of feature channels, and the second convolutional layer is used to convert the number of feature channels in the decoding stage into the number of channels of the final output and output the result.

[0028] Optionally, the residual module provided with a self-attention mechanism includes:

[0029] Two target modules, the target modules include a group normalization layer, an activation function and a third convolutional layer, one of the target modules is used to process low-level features, and the other target module is used to process high-level features;

[0030] A feature adjustment layer, the feature adjustment layer is located between the two target modules,

[0031] The fourth convolutional layer is used to adjust the number of channels of the feature map.

[0032] Optionally, the preprocessing of the original B-scan image to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image includes:

[0033] Using a bicubic interpolation method to interpolate the original B-scan image to obtain the high-resolution data;

[0034] Downsampling the original B-scan image using a bilinear downsampling method to obtain the low-resolution data;

[0035] The low-resolution data is interpolated using a bicubic interpolation method to obtain the inference data, and the size of the inference data is the same as that of the high-resolution data.

[0036] Optionally, the ground penetrating radar data set includes a simulation data set and a measured data set, and the step of constructing the ground penetrating radar data set includes:

[0037] Establish the detection scenario of underground targets, debug the parameters under the condition of simulation operation, obtain the critical value of the parameters, and generate low-resolution ground penetrating radar simulation data;

[0038] A sand box was built to simulate the detection environment of real scenes. By controlling the environmental conditions and target configuration, echo signals in various real scenes were captured to generate low-resolution ground-penetrating radar measurement data.

[0039] The simulation data set is constructed based on the ground penetrating radar simulation data, and the measured data set is constructed based on the ground penetrating radar measured data.

[0040] In a second aspect, an embodiment of the present invention provides a ground penetrating radar B-scan image super-resolution processing device, comprising:

[0041] A data set construction module, used to construct a ground penetrating radar data set, wherein the ground penetrating radar data set includes a plurality of original B-scan images;

[0042] a preprocessing module, configured to preprocess each of the plurality of original B-scan images to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image;

[0043] A model building module, used to build a conditional denoising diffusion model, wherein the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, wherein the Gaussian noise diffusion stage includes a residual self-attention U-net network;

[0044] A training module, used for training the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data and inference data corresponding to the plurality of original B-scan images;

[0045] A reconstruction processing module, used for performing super-resolution reconstruction processing on the image to be processed by using the trained conditional denoising diffusion model;

[0046] Wherein, the Gaussian noise diffusion stage is used for:

[0047] Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1;

[0048] Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images;

[0049] Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data;

[0050] Inputting T of the dual-channel noisy image data into the residual self-attention U-net network to perform noise estimation to obtain estimated noise;

[0051] The inference phase is used to:

[0052] Generate a pure noise image of the same size as the input image;

[0053] The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes:

[0054] Perform noise estimation on the image obtained by the i-1th iterative processing based on the residual self-attention U-net network to obtain the i-th estimated noise;

[0055] The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

[0056] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is used to read the program in the memory to implement the steps in the ground penetrating radar B-scan image super-resolution processing method as described in the first aspect.

[0057] In a fourth aspect, an embodiment of the present invention provides a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the ground penetrating radar B-scan image super-resolution processing method as described in the first aspect.

[0058] In an embodiment of the present invention, a ground penetrating radar data set is constructed, and the ground penetrating radar data set includes multiple original B-scan images; each of the multiple original B-scan images is preprocessed to obtain a target data set; a conditional denoising diffusion model is constructed, and the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, and the Gaussian noise diffusion stage includes a residual self-attention U-net network; the conditional denoising diffusion model is trained based on the target data set to obtain a trained conditional denoising diffusion model; and the trained conditional denoising diffusion model is used to perform super-resolution reconstruction processing on the image to be processed.

[0059] The beneficial effect of the present invention is that it solves the problems of insufficient spatial resolution, severe noise coverage and insufficient detail information of ground penetrating radar B-scan images, realizes the effect of fast moving antenna in the horizontal direction, and achieves the effect of high-frequency antenna detection in the vertical direction. At the same time, the embodiment of the present invention realizes detail recovery and noise suppression by enhancing the echo signal of the target of interest, thereby improving data quality and analysis effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Attached Figure 1 A flowchart of a ground penetrating radar B-scan image super-resolution processing method provided by an embodiment of the present invention;

[0061] Attached Figure 2 A schematic diagram of a process for preprocessing an original B-scan image provided by an embodiment of the present invention;

[0062] Attached Figure 3 A schematic diagram of a Gaussian noise diffusion stage provided by an embodiment of the present invention;

[0063] Attached Figure 4 A schematic diagram of the reasoning stage provided by an embodiment of the present invention;

[0064] Attached Figure 5 A schematic diagram of the structure of a residual self-attention U-net network provided in an embodiment of the present invention;

[0065] Attached Figure 6a A schematic diagram of the internal structure of a residual module provided in an embodiment of the present invention;

[0066] Attached Figure 6b A schematic diagram of the internal structure of a self-attention residual module provided in an embodiment of the present invention;

[0067] Attached Figure 6c A schematic diagram of the internal structure of a feature adjustment layer provided in an embodiment of the present invention;

[0068] Attached Figure 7 A schematic diagram of a ground penetrating radar B-scan image super-resolution processing device provided by an embodiment of the present invention;

[0069] Attached Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In the embodiments of the present invention, the term "and / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent three situations: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. In the embodiments of the present invention, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first", "second", etc. in the specification and claims of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more.

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] See also Figure 1 , Figure 1 The present invention provides a flow chart of a method for super-resolution processing of ground penetrating radar B-scan images. The method specifically comprises the following steps:

[0073] Step 101, constructing a ground penetrating radar dataset, wherein the ground penetrating radar dataset includes a plurality of original B-scan images;

[0074] Step 102, for each of the multiple original B-scan images, preprocess the original B-scan image to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image;

[0075] Step 103, constructing a conditional denoising diffusion model, wherein the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, wherein the Gaussian noise diffusion stage includes a residual self-attention U-net network;

[0076] Step 104, training the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data, and inference data corresponding to the plurality of original B-scan images;

[0077] Step 105: Use the trained conditional denoising diffusion model to perform super-resolution reconstruction on the image to be processed.

[0078] It should be understood that the constructed conditional denoising diffusion model can be trained through the above steps 101 to 104 to obtain a trained conditional denoising diffusion model. In step 105, the trained conditional denoising diffusion model can be applied to ground penetrating radar B-scan image super-resolution processing without repeating the training model. Among them, the constructed ground penetrating radar data set is used to construct a target data set for model training. In the specific implementation, the target data set can be divided according to a preset ratio and used for model training and testing respectively. The details are not repeated here.

[0079] Specifically, the ground penetrating radar detects in each detection scene and generates a corresponding B-scan image. Each B-scan image is a two-dimensional array, which can be represented as an image. Let the number of sampling points in the longitudinal depth dimension be n and the number of sampling points in the transverse space dimension be m. When m≥n, the method provided by the present invention can generate a super-resolution B-scan image with a size of m×m; when n>m, the method provided by the present invention can generate a super-resolution B-scan image with a size of n×n.

[0080] First, a ground penetrating radar data set for super-resolution reconstruction is constructed. Optionally, in some embodiments, the ground penetrating radar data set includes a simulation data set and a measured data set, and step 101 includes:

[0081] Establish the detection scenario of underground targets, debug the parameters under the condition of simulation operation, obtain the critical value of the parameters, and generate low-resolution ground penetrating radar simulation data;

[0082] A sand box was built to simulate the detection environment of real scenes. By controlling the environmental conditions and target configuration, echo signals in various real scenes were captured to generate low-resolution ground-penetrating radar measurement data.

[0083] The simulation data set is constructed based on the ground penetrating radar simulation data, and the measured data set is constructed based on the ground penetrating radar measured data.

[0084] In this embodiment, the ground penetrating radar data set includes a simulation data set and a measured data set. By generating low-resolution ground penetrating radar data through the above method, the generalization ability of the data-driven model in reconstructing the resolution of ground penetrating radar data can be significantly improved.

[0085] The measured data set uses the soil types built into the ground penetrating radar data acquisition system, including dry sand, dry soil, wet soil, and wet sand. Underground targets include steel bars and water pipes. Set the acquisition scene and target burial conditions, adjust the system parameters, and complete the acquisition of the measured data set.

[0086] As a specific embodiment, the constructed data set includes several soil types such as concrete, inhomogeneous soil, and several underground targets such as buried steel bars, buried pipes and cracks in simulated and real detection scenarios. Through different combinations of soil types and underground targets, multiple different detection scenarios can be designed. By generating multiple detection scenarios, a ground penetrating radar data set containing a certain number of images is produced.

[0087] The simulation data set was generated using the gprMax simulation software. The relative dielectric constant of concrete is 10, the conductivity is 0, the relative magnetic permeability is 1, and the magnetic loss is 0. The heterogeneous soil consists of 50 different materials with a sand content of 0.5, a clay content of 0.5, and a bulk density of 2g / cm 3 , the density of sand is 2.66g / cm 3 , the water volume fraction is 0.1%-25%.

[0088] For the simulation data construction of buried steel bars, a cylindrical perfect conductor (PEC) is used to simulate steel bars during simulation. In this embodiment, a Rick wavelet with unit amplitude and center frequency of 500MHz is used, the transmitting source adopts z polarization, and the antenna movement step is 0.01. Under the premise of meeting the low resolution requirement, the critical space discretization is set to (0.01, 0.01, 0.01). At the same time, when the target echo is fully received, the minimum longitudinal sampling point is set to 512. The simulation scene sizes corresponding to a single steel bar, two steel bars and four steel bars are (1.2, 1.2, 0.01), (1.8, 1.2, 0.01) and (3, 1.2, 0.01) respectively. The target burial depth is random between 0.05-0.25. The number of antenna movements corresponding to different target numbers, that is, the number of A-scan channels, are 64, 128 and 256 respectively. Finally, a total of 1290 sets of low-resolution buried steel bar data with sizes of 64×512, 128×512, and 256×512 were constructed under the two soil settings.

[0089] For the simulation data construction of buried water pipes, the water pipe is made of PVC, with a relative dielectric constant of 4, a conductivity of 0, a relative magnetic permeability of 1, and a magnetic loss of 0. The relative dielectric constant of water is 80, the conductivity is 0, the relative magnetic permeability is 1, and the magnetic loss is 0. The antenna movement step is 0.005. On the premise of meeting the low resolution requirement, the critical space discretization is set to (0.005, 0.005, 0.005). When the target echo is fully received, the minimum longitudinal sampling point is set to 1024. The radius of the water pipe is 0.03, and the wall thickness is 0.005. The simulation scene sizes corresponding to a single water pipe, two water pipes, and four water pipes are (1.2, 1.2, 0.005), (1.6, 1.2, 0.005), and (3, 1.2, 0.005), respectively. The remaining parameters are the same as those for constructing buried steel bars. Finally, a total of 1290 sets of low-resolution buried water pipe data with sizes of 128×1024, 256×1024, and 512×1024 were constructed under the two soil settings.

[0090] For the simulation data construction of surface visible cracks, multiple orthogonal parallelepipeds with random lengths and widths between 0.01 and 0.02 are used, with air inside, to spread downward from the surface to simulate cracks, and the crack direction is randomly left or right. Since it is a surface visible crack, the distance between the transmitting source and the receiving antenna is set to 0.05 from the target to avoid the overlap of the echo and direct wave of the nearest target. The simulation scene sizes corresponding to a single crack, two cracks, and three cracks are (1.2, 1.2, 0.01), (1.8, 1.2, 0.01), and (3, 1.2, 0.01), respectively. The remaining parameters are the same as those for constructing buried steel bars. Finally, in the context of concrete soil, a total of 645 sets of low-resolution surface visible crack data with sizes of 64×512, 128×512, and 256×512 were constructed.

[0091] In this embodiment, for the measured data set, a sand box with dimensions of 1.5 m long, 0.5 m wide, and 0.5 m deep was built. By controlling environmental conditions and target configurations, echo signals in a variety of real scenes were captured to generate low-resolution GPR data.

[0092] Regarding the acquisition equipment, a high-performance GSSI SIR 4000 ground penetrating radar data acquisition system equipped with a 2GHz antenna is used in this embodiment. In distance mode, the range measuring wheel is calibrated to 22182 per meter. Four signal gain points are set, namely 20, 30, 35 and 35. The delay is set to 12 nanoseconds to avoid interference from the initial part of the transmitted signal. The surface emission position is set to 10 cm to avoid interference from the transparent acrylic plate to the radar wave. The IIR low-pass filter is set to 3500MHz to eliminate high-frequency noise and retain low-frequency information. The IIR high-pass is set to 400MHz to eliminate low-frequency noise and retain high-frequency signals. IIR horizontal smoothing and IIR background removal remain off or set to 0.

[0093] The measured data set uses the soil types built into the SIR 4000, including dry sand, dry soil, wet soil, and wet sand. Underground targets include steel bars and water pipes. Take the measured data collected in a size of 64×256 as an example. The antenna movement distance is 60 cm. In this case, the sampling / scan value needs to be set to 256 and the scan / meter to 106 to ensure that 64 A-scans can be collected within the collection distance of 60 cm, and 256 points can be sampled in the depth direction. Under the four soil types, the collection depths are set to 40 cm, 30 cm, 25 cm, and 20 cm, respectively, and the corresponding time windows are 5.93 ns, 6.67 ns, 8.28 ns, and 7.41 ns, respectively.

[0094] During the collection, two targets formed four underground situations, including single steel bar, double steel bar, single water pipe and double water pipe. In any case, the target was 15 cm away from the acrylic plate. The target positions of single target and double target were located at 30 cm, 20 cm and 40 cm of the antenna movement distance, respectively. Four soil types and four underground situations were combined in sequence, and 10 sets of data were collected for each group, and a total of 160 sets of 64×256 measured data were collected. Using the same method, 160 sets of measured data with sizes of 64×512, 128×256 and 128×512 were collected, respectively, and finally 640 sets of measured data constituted the measured data set.

[0095] It should be understood that in this embodiment, background elimination is performed on the data in the constructed data set to remove the influence of the direct wave on the target echo. Assume that the number of sampling points in the longitudinal depth dimension of the constructed data is n, and the number of sampling points in the transverse spatial dimension is m. When m≥n, the method of the present invention can generate a super-resolution B-scan image with a size of m×m; when n>m, the present invention can generate a super-resolution B-scan image with a size of n×n.

[0096] Design a data preprocessing method to preprocess the constructed data set. Optionally, in some embodiments, step 102 includes: interpolating the original B-scan image using a bicubic interpolation method to obtain the high-resolution data; downsampling the original B-scan image using a bilinear downsampling method to obtain the low-resolution data; interpolating the low-resolution data using a bicubic interpolation method to obtain the inference data, wherein the size of the inference data is the same as the size of the high-resolution data.

[0097] Assume that the number of sampling points in the vertical depth dimension of the constructed original B-scan image is n, and the number of sampling points in the horizontal spatial dimension is m. Take a B-scan with n>m as an example, record it as d, and use the bicubic interpolation method to interpolate d into n×n high-resolution data d h , using the bilinear downsampling method to downsample d to m×m low-resolution data d l , and then use the bicubic interpolation method to convert d l Interpolate to n×n size inference data d f Assume that a total of N original B-scan images are collected, and each GPRB-scan image is preprocessed to generate 3×N data. The 3×N GPR B-scans are divided according to the proportion and can be used for model training and testing respectively.

[0098] Specifically, Figure 2As shown in the figure, taking a certain original B-scan image (i.e., original data) in the constructed simulation data with a size of 64×512 as an example, the number of sampling points in the horizontal space is 64, and the number of sampling points in the vertical depth dimension is 512, which is recorded as d. The bicubic interpolation method is used to interpolate d into a high-resolution data d of 512×512. h In this embodiment, the bicubic interpolation method is used because it calculates the difference by considering 16 adjacent pixels and by using a cubic polynomial, it can fit the data more accurately, reduce the interpolation error, and provide a reconstruction effect closer to the original data while retaining more data details.

[0099] The bilinear downsampling method is used to downsample d to a low-resolution data d of 64×64. l Among them, the reason for using the bilinear interpolation method is that in low-resolution data, there is no need to pay attention to too much detail information, which provides a good balance between accuracy and efficiency.

[0100] Then use the bicubic interpolation method to convert d l Interpolate to 512×512 size inference data d f (Right now Figure 2 d in inf ). At this time, d f Although the size is similar to d h The two methods are consistent, but there are significant differences in data quality, detail information, and smoothing effect.

[0101] In the constructed dataset, a total of 3865 B-scan images were collected, and the above process was repeated for each GPR B-scan, generating 11595 data sets. These 11595 GPR B-scans were divided into training set, validation set and test set in a ratio of 7:1:2.

[0102] The constructed conditional denoising diffusion model is different from the ordinary diffusion probability model. The generated data is not random with probability, but generated according to conditional input. The conditional denoising diffusion model mainly includes Gaussian noise diffusion stage and inference stage.

[0103] Optionally, the Gaussian noise diffusion stage is used to:

[0104] Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1;

[0105] Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images;

[0106] Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data;

[0107] The T dual-channel noisy image data are input into the residual self-attention U-net network for noise estimation to obtain estimated noise.

[0108] See also Figure 3 , with the aforementioned d h (ie d h As an example, in the Gaussian noise diffusion stage, T images with d are generated. h Gaussian noise of the same size {p 1 ,p 2 ,…,p t ,…p T}, the noise intensity increases successively. h With each noise data in the noise sequence [p t ,t=1,2,…,T] are added separately to obtain T single-channel noisy images, that is, a sequence of images with gradually added noise Right now in is approximated to a pure noise image. In this process, each single-channel noisy image of size n×n in the sequence is With single-channel image inference data d of size n×n f Channels are merged, and the merged data is a two-channel noisy image data of size n×n. It is input into the residual self-attention U-net network to estimate the noise added at each step.

[0109] During the training process, this embodiment designs a loss function to measure the prediction value of the network. and the actual noise p t The network is trained to estimate the added noise by taking the difference between the two images, so that each step can restore a noise-suppressed B-scan image.

[0110] In this embodiment, the random diffusion of Gaussian noise is no longer directly parameterized based on the sampled time series. Instead, after a time series is sampled, the Gaussian noise is uniformly distributed between the time series and the previous time series for selection, making the random diffusion process of Gaussian noise continuous and captureable for calculation. In the training phase of the conditional denoising diffusion model, the time series is no longer directly input into the residual self-attention U-net network, but the noise level is directly input into the network to train the network to improve its noise suppression capability.

[0111] Optionally, the reasoning phase is used to:

[0112] Generate a pure noise image of the same size as the input image;

[0113] The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes:

[0114] Performing noise estimation on the image obtained by the i-1th iterative processing based on the residual self-attention U-net network to obtain an i-th estimated noise;

[0115] The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

[0116] In the inference phase, we first generate a pure noise image of size n×n Inverse diffusion starts from a pure noise image and estimates the current image through the residual self-attention U-net network trained in the Gaussian diffusion stage The noise in the image is subtracted and the noise is suppressed. After T iterations, the noise is finally suppressed. The image is then super-resolution reconstructed. It should be understood that the image obtained after the T-th iteration is the image after noise suppression.

[0117] As a specific example, see Figure 4 ,like Figure 4 In the Gaussian diffusion stage shown in the figure, a B-scan image d in the target data set h As the input of the overall model, let x 0 =d h First, a time series t~{0,1,2,…,T} is uniformly sampled. The variable β is introduced, which is defined as the diffusion rate, which is in a decaying state. As the noise is added, the noise will increase. At any time, x t must be the same as x at the previous moment t-1 , and the added noise, so the noise sampled at this moment is expressed by β t =1-α t Represented by using the weight term and To measure x t Which one is more affected by the state at the previous moment or the added noise:

[0118]

[0119] Among them, ∈ is the added Gaussian noise, which obeys Gaussian distribution

[0120] With the β t Getting bigger, α t Getting smaller and smaller, Getting smaller and smaller, It means that as time goes by, the influence of noise on the result will become greater and greater. At this time, a problem arises: any moment can be obtained from the previous moment. But currently only x 0 The time is known, of course, we can first use x 0 Time to calculate x 1 time, and then by x 1 Calculate x at any time 2 time, and so on, until x is calculated T , which is similar to a recursive calculation, but the recursion performs serial calculations, which results in redundant calculations and inability to transfer gradients, leading to inability to perform training. Therefore, how to get from x 0 How to calculate any time at any moment, or how to directly calculate the final time x T It is worth noting that in the reverse diffusion process, only the final moment x generated by the diffusion process is concerned T The distribution results of , no need to pay attention to other moments.

[0121] Known x 0 , x t is a distribution at any time. Formula (1) shows that x t With x t-1 The relationship between t-1 With x t-2 The relationship between them is as follows:

[0122]

[0123] Substituting formula (2) into formula (1) yields:

[0124]

[0125] From the above formula, we know that all the terms related to α are known, and the only unknown term is x t-2 , expanding formula (3) yields:

[0126]

[0127] Because ∈ t With ∈ t-1 are all sampled noises that obey Gaussian distribution, so for ∈ t With ∈ t-1 For , the distributions of the two are exactly the same. Expanding the above formula (4), according to the principle of Gaussian probability addition under the same distribution, ∈ t With ∈ t-1 The ride and The term changes the variance of its Gaussian distribution and becomes and Then the two noises that follow the Gaussian distribution follow This principle, so formula (4) can be simplified and combined to obtain:

[0128]

[0129] Observe formula (1) and formula (5), respectively, x t With x t-1 , x t-2 The relationship between x and x can be found to be a rule. t With x 0 When the relationship is t Multiply until α 1 , and then take the square root, we can get x 0 The weight at the moment, similarly, the noise weight at this time can also be calculated. Therefore, when the original distribution x is known 0 In the case of , the distribution at any time can be obtained by the following formula:

[0130]

[0131] In the diffusion process, we focus on how the input continuously adds noise through the Gaussian diffusion process, how to change from time t = 0 to time t = T, and we can also get the noise distribution at any time. In the reverse process, we focus more on how to remove noise step by step from time t = T to generate noise-free data at time t = 0.

[0132] In the reverse process, it is impossible to directly infer from x t Calculate x at any time 0 moment, can only be obtained through iteration from x t Infer x at any time t-1 time, from x t-1 Infer x at any time t-2 moment, and iterate in this way until x is calculated 0 Therefore, the solution task changes from solving p(x 0 |x t ) is converted into solving p(x t-1 |x t ) problem, using the Bayesian formula, and then solving p(x t-1 |x t ) is converted to solve p(x t |x t-1 ) problem, and solve p(x t |x t-1 ) has been solved in the forward process. Therefore, the first step in the reverse process is to solve p(x t-1 |x t), and converted into the solution of the following formula through the Bayesian formula:

[0133]

[0134] According to the calculation of the forward process, q(x t |x t-1 ) is a known term, given by Solve. And q(x t-1 ) and q(x t ) is solved according to formula (6), when x is known 0 Under the condition of , the calculation is performed at any time, so formula 7) is further solved into the following form:

[0135]

[0136] Therefore, q(x t |x t-1 ,x 0 ), q(x t-1 |x 0 ) and q(x t |x 0 ) can be solved by formula (6):

[0137]

[0138] Among them, all terms are known and computable. And ∈ is Gaussian noise that follows a normal distribution, and we only need to determine the mean μ and variance σ 2 From the above formula, we can see that ∈ is multiplied by one term first, which is equivalent to changing the variance of the normal distribution, and then one term is added, which is equivalent to changing the mean of the normal distribution, but the whole still obeys the normal distribution.

[0139] A normal distribution has the form Based on its characteristics, we can understand that multiplication is equivalent to an addition operation, and division is equivalent to a subtraction operation. Combining formulas (8) and (9), this solution can be transformed into the solution of the following formula:

[0140]

[0141] In the process of solving the above equation, we only need to t-1 Therefore, when expanding the above formula and combining similar terms, we pay more attention to x t-1 Related items:

[0142]

[0143] Among them, C is a constant term.

[0144] Expanding the normal distribution yields:

[0145]

[0146] Combining formula (11) and (12), we can find the corresponding relationship, x 2 and Correspondingly, x and x t-1 Correspondingly, therefore:

[0147]

[0148] Among them, each term in the first formula is a known condition, so the obtained σ 2 is a constant value. According to the known conditions and the obtained σ 2 , formulate and simplify the second formula, and find that μ and x t , x 0 Related, means x t-1 It is also related to the distribution of these two moments:

[0149]

[0150] like Figure 4 In the reverse process shown, x 0 is unknown, but can be converted to x by the forward process formula (6) 0 Use x t Express:

[0151]

[0152] Finally, by substituting formula (15) into formula (14), we can calculate the mean μ, which is only related to x t Related:

[0153]

[0154] The calculation ends here. From the beginning, we know the last moment x t In the case of t-1 The mean of the distribution at the moment, the variance is a constant, and (1-α t ) means, then x t-1 for:

[0155]

[0156] By x t The time x is calculated t-1 distribution of moments, but introduces a new problem ∈ t , that is, in the forward process x t-1 Time to x tThe noise added at the moment. t In the case of , an approximate solution can be obtained. Therefore, the neural network f θ (x 0 ,x t ,∈ t ) to predict at x t The noise added at time ∈ t The predicted value of .

[0157] In the forward process, the distribution at the next moment is obtained by adding noise to the distribution at the previous moment, where the noise is a known condition, obtained through sampling, and used as the true value for training. In other words, the forward diffusion process is a process of continuously adding noise and a process of continuously constructing noise labels. Returning to formula (16) itself, each item at this moment is a known condition, and the noise ∈ t Therefore, during the training process, the forward process provides the label of the noise at each moment, and the reverse process fits the noise label.

[0158] For the time series t mentioned initially during the diffusion phase, it is no longer determined by α t The noise level is defined by t With α t-1 The noise γ~U(γ t-1 ,γ t ), then the calculation of the Gaussian diffusion stage (6) and the reverse stage (17) becomes:

[0159]

[0160] And a segmented distribution method is used for the noise level in, Ensure that the weight of each segment is uniformly distributed and the total weight is 1. The whole process is guaranteed to be carried out between the uniformly distributed time t and t-1, in the corresponding interval (γ t-1 ,γ t ) is uniformly sampled within the image, thereby obtaining a continuous and smooth noise level γ, reducing the discontinuity or mutation in the generated image that may be caused in a discrete time series.

[0161] In this embodiment, the conditional denoising diffusion model is a conditional generative model that gradually covers the single-channel data with noise. The corresponding single-channel inference data d f Channel merging is performed as a condition and input into the residual self-attention U-net network to estimate the noise. The conditional input limits the probability of the model randomly generating output results.

[0162] Optionally, in some embodiments, the residual self-attention U-net network includes:

[0163] The encoding stage includes a first convolutional layer, a first processing block and a residual module, wherein the first convolutional layer is used to map the data into a high-dimensional feature space, and the first processing block is used to reduce the spatial dimension, increase the number of feature channels, extract multi-scale features of the input data, and increase the abstraction level of feature representation;

[0164] An intermediate stage, comprising two residual modules with a self-attention mechanism, for extracting global context information and capturing long-distance dependencies through the self-attention mechanism;

[0165] The decoding stage includes a second processing block and a second convolutional layer, the second processing block is used to gradually restore the spatial dimension while reducing the number of feature channels, and the second convolutional layer is used to convert the number of feature channels in the decoding stage into the number of channels of the final output and output the result.

[0166] See also Figure 5-6c The conditional denoising diffusion model constructed in this embodiment designs a residual self-attention U-net network and scales the jump connection to The number of residual modules is increased, and channel multipliers at different resolutions are increased. It should be understood that in this embodiment, for all experiments, the uniformly sampled time step t~{0,...,T} has T=2000. Generally, the larger the value of T, the better the model will be.

[0167] Encoding stage: The input first passes through the first convolutional layer to map the data to a high-dimensional feature space, providing rich feature representation for subsequent network layers. Then, the spatial dimension is gradually reduced through the first processing block (the first processing block is used to perform a series of convolution and downsampling operations), the number of feature channels is increased, the multi-scale features of the input data are extracted, and the abstraction level of the feature representation is gradually increased. Finally, it is completed through a residual module. Among them, the size of the first convolutional layer is 3×3, the step size is 1, and the padding is 1. The convolution kernel size used in the downsampling stage is 3×3, the step size is 2, and the padding is 1.

[0168] Intermediate stage: After passing through two residual modules with self-attention mechanism, global context information is extracted, long-distance dependencies are captured through the self-attention mechanism, feature expression is enhanced, high-dimensional features are further processed, and the continuity and consistency of features are maintained.

[0169] The decoding stage corresponds to the encoding stage. The second processing block (the second processing block is used to perform a series of upsampling and convolution operations) gradually restores the spatial dimension and reduces the number of feature channels. The upsampling uses nearest neighbor interpolation, the convolution kernel size is 3×3, the step size is 1, and the padding is 1. Finally, the second convolution layer converts the number of feature channels in the decoding stage to the number of channels in the final output and outputs the result.

[0170] Optionally, in some embodiments, the residual module provided with a self-attention mechanism includes:

[0171] Two target modules, the target modules include a group normalization layer, an activation function and a third convolutional layer, one of the target modules is used to process low-level features, and the other target module is used to process high-level features;

[0172] A feature adjustment layer, the feature adjustment layer is located between the two target modules,

[0173] The fourth convolutional layer is used to adjust the number of channels of the feature map.

[0174] The residual module with self-attention mechanism can extract local and global feature information in the downsampling stage and retain more context information. In the upsampling stage, the upsampled features can be further processed, and the jump connection features in the encoding stage can be integrated to improve the reconstruction quality. The residual module with self-attention mechanism consists of two target modules with the same structure, a feature adjustment layer, and a 1×1 convolution layer (i.e., the fourth convolution layer). The target module consists of a group normalization layer, a Swish activation function, and a third convolution layer with a convolution kernel size of 3×3, a step size of 1, and a padding of 1. Among them, the use of group normalization can normalize the input features so that the feature maps have the same distribution, reduce the internal covariate shift, and is suitable for small batch training of the proposed model, which can play a role in maintaining good stability. The Swish activation function is a smooth activation function, which still has an activation effect on negative inputs and can retain more information.

[0175] One target module is used to process low-level features, and the other target module is used to process higher-level features to enhance feature representation capabilities. The cascade of the two target modules improves the model's expressiveness and complexity, and can better capture complex feature relationships. The final 1×1 convolutional layer is used to adjust the number of channels in the feature map.

[0176] like Figure 6a-6cAs shown in the figure, the feature adjustment layer is between the two target modules. The feature adjustment layer first maps the noise level to a dimension that matches the feature map through a linear layer, providing a way to introduce noise information into the feature map. Secondly, shape matching adjusts the shape of the linear layer result to match the shape of the feature map. Then when using an affine transformation, the calculation result is split into 'gamma' and 'beta', and finally the feature map is scaled using gamma and translated using beta, providing the ability to dynamically adjust the feature map. The feature adjustment layer provides a way to map the noise level to the feature map, allowing the model to adapt to different noise levels, improving the model's ability to dynamically adjust the feature map, and improving the flexibility and adaptability of the model.

[0177] like Figure 6b As shown in the figure, the residual module with self-attention mechanism adds a self-attention module on the basis of the residual module without self-attention mechanism. The processed feature map is first group normalized, and then the query, key and value feature maps are generated. The softmax layer is applied to calculate the self-attention weight, and a 1×1 convolution layer is used for linear transformation. Finally, the feature map is output and added to the input feature map. By setting the residual module with self-attention mechanism, the feature selection ability can be improved, the local perception ability of convolution operation can be improved, long-distance dependencies can be captured, and the feature expression ability can be enhanced. For the network as a whole, the representation ability of the network is improved, the effect of multi-scale feature fusion is enhanced, the generalization ability of the model is improved, and the flexibility of the model is increased.

[0178] In an embodiment of the present invention, a series of targeted modules are designed according to the characteristics of GPR data to improve the effects of denoising and super-resolution reconstruction. On the one hand, the feature extraction capability of the network can be improved by introducing a residual module. Compared with the traditional U-net, the introduction of a residual module can more effectively capture the detail information in the image and reduce information loss. On the other hand, a self-attention mechanism is embedded in the encoder and decoder of U-net. The self-attention mechanism can capture long-distance dependencies in the image and enhance the understanding of global information, which is particularly suitable for processing complex B-scan images. On the third hand, the network architecture is optimized according to the characteristics of GPR data, including special processing modules for underground target features, which can better identify and reconstruct the detail information of underground targets. On the fourth hand, multi-scale feature fusion technology is introduced in the jump connection of U-net, combining feature information of different scales to improve the robustness and reconstruction accuracy of the network.

[0179] In the above specific embodiment, the super-resolution task of converting 64×512 buried single steel bar data from 64×64 to 512×512 can be completed by the above method. In the horizontal direction, the number of antenna movements is increased from 64 times to 512 times, achieving fast antenna movement. In the vertical direction, although the number of sampling points remains unchanged, the super-resolution result with richer details achieves the detection effect at a higher center frequency.

[0180] The present invention can effectively improve the resolution of the two-dimensional B-scan image of the ground penetrating radar. When n>m, interpolation and super-resolution in the airspace are realized in the lateral spatial survey line direction. When m≥n, interpolation and super-resolution in the time domain are realized in the longitudinal time dimension sampling. By super-resolution processing of the B-scan image, the echo signal of the target of interest is enhanced, and the super-resolution construction of the detail features of the weak echo signal and the noise suppression of the B-scan image are realized, thereby improving the data quality of the B-scan and the subsequent processing and analysis effect.

[0181] In addition to improving the perceived quality of images, super-resolution reconstruction of GPR B-scan images can also help improve the application of B-scan images in other tasks. First, high-resolution GPR B-scan images can provide more accurate and detailed underground structural information, which helps geologists and engineers better understand the structure of underground rock layers, soils, and waters, and is of great significance for geological exploration, underground resource development, and engineering design. By observing and analyzing subtle changes in high-resolution B-scan images, it is possible to more accurately determine underground geological features, hydrological processes, and potential underground hazards. Second, super-resolution reconstructed B-scan images can also provide more accurate non-invasive underground detection and monitoring. For example, in urban construction and soil pollution monitoring, high-resolution B-scan images can help locate underground pipelines, underground facilities, and groundwater levels, as well as detect and track the spread and migration of pollutants. This is of great significance for urban planning and environmental management, and can reduce unnecessary damage and environmental risks.

[0182] like Figure 7 As shown, the embodiment of the present invention further provides a ground penetrating radar B-scan image super-resolution processing device 700, comprising:

[0183] A data set construction module 701 is used to construct a ground penetrating radar data set, wherein the ground penetrating radar data set includes a plurality of original B-scan images;

[0184] A preprocessing module 702 is used to preprocess each of the multiple original B-scan images to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image;

[0185] A model construction module 703 is used to construct a conditional denoising diffusion model, wherein the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, wherein the Gaussian noise diffusion stage includes a residual self-attention U-net network;

[0186] A training module 704 is used to train the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data and inference data corresponding to the multiple original B-scan images;

[0187] Reconstruction processing module 705, used for performing super-resolution reconstruction processing on the image to be processed by using the trained conditional denoising diffusion model;

[0188] Wherein, the Gaussian noise diffusion stage is used for:

[0189] Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1;

[0190] Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images;

[0191] Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data;

[0192] The T dual-channel noisy image data are input into the residual self-attention U-net network for noise estimation to obtain estimated noise.

[0193] The inference phase is used to:

[0194] Generate a pure noise image of the same size as the input image;

[0195] The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes:

[0196] Perform noise estimation on the image obtained by the i-1th iterative processing based on the residual self-attention U-net network to obtain the i-th estimated noise;

[0197] The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

[0198] Optionally, the residual self-attention U-net network includes:

[0199] The encoding stage includes a first convolutional layer, a first processing block and a residual module, wherein the first convolutional layer is used to map the data into a high-dimensional feature space, and the first processing block is used to reduce the spatial dimension, increase the number of feature channels, extract multi-scale features of the input data, and increase the abstraction level of feature representation;

[0200] An intermediate stage, comprising two residual modules with a self-attention mechanism, for extracting global context information and capturing long-distance dependencies through the self-attention mechanism;

[0201] The decoding stage includes a second processing block and a second convolutional layer, the second processing block is used to gradually restore the spatial dimension while reducing the number of feature channels, and the second convolutional layer is used to convert the number of feature channels in the decoding stage into the number of channels of the final output and output the result.

[0202] Optionally, the residual module provided with a self-attention mechanism includes:

[0203] Two target modules, the target modules include a group normalization layer, an activation function and a third convolutional layer, one of the target modules is used to process low-level features, and the other target module is used to process high-level features;

[0204] A feature adjustment layer, the feature adjustment layer is located between the two target modules;

[0205] The fourth convolutional layer is used to adjust the number of channels of the feature map.

[0206] Optionally, the preprocessing module 702 includes:

[0207] A first processing unit, configured to perform interpolation processing on the original B-scan image using a bicubic interpolation method to obtain the high-resolution data;

[0208] A second processing unit is used to downsample the original B-scan image using a bilinear downsampling method to obtain the low-resolution data;

[0209] The third processing unit is used to perform interpolation processing on the low-resolution data using a bicubic interpolation method to obtain the inference data, and the size of the inference data is the same as the size of the high-resolution data.

[0210] Optionally, the ground penetrating radar data set includes a simulation data set and a measured data set, and the data set construction module 701 includes:

[0211] The first generation unit is used to establish a detection scene of underground targets, debug parameters under the condition of simulation operation, obtain critical values ​​of parameters, and generate low-resolution ground penetrating radar simulation data;

[0212] The second generation unit is used to build a sand box to simulate the detection environment of real scenes. By controlling environmental conditions and target configuration, echo signals in various real scenes are captured to generate low-resolution ground penetrating radar measured data;

[0213] A construction unit is used to construct the simulation data set based on the ground penetrating radar simulation data, and to construct the measured data set based on the ground penetrating radar measured data.

[0214] The ground penetrating radar B-scan image super-resolution processing device 700 provided in the embodiment of the present invention can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be repeated in this embodiment.

[0215] It should be noted that the division of units in the embodiments of the present invention is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present invention may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0216] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0217] like Figure 8As shown, an embodiment of the present invention provides an electronic device 800, including: a memory 802, a processor 801, and a program stored in the memory 802 and executable on the processor 801; the processor 801 is used to read the program in the memory 802 to implement the steps in the ground penetrating radar B-scan image super-resolution processing method as described above.

[0218] An embodiment of the present invention also provides a readable storage medium, on which a program is stored. When the program is executed by a processor, the various processes of the above-mentioned ground penetrating radar B-scan image super-resolution processing method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here. Among them, the readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (such as compact disk (CD), digital video disc (DVD), Blu-ray Disc (BD), high-definition versatile disc (HVD), etc.), and semiconductor memory (such as read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read only memory (EEPROM), non-volatile memory (NAND FLASH), solid-state drive (Solid State Disk or Solid State Drive, SSD)), etc.

[0219] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0220] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) to execute the methods described in each embodiment of the present invention.

[0221] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for super-resolution processing of ground penetrating radar B-scan images, characterized in that: include: Constructing a ground penetrating radar dataset, wherein the ground penetrating radar dataset includes a plurality of original B-scan images; For each of the multiple original B-scan images, preprocess the original B-scan image to obtain high-resolution data, low-resolution data, and inference data corresponding to the original B-scan image; Constructing a conditional denoising diffusion model, the conditional denoising diffusion model comprising a Gaussian noise diffusion stage and an inference stage, the Gaussian noise diffusion stage comprising a residual self-attention U-net network; Training the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data, and inference data corresponding to the plurality of original B-scan images; Performing super-resolution reconstruction on the image to be processed by using the trained conditional denoising diffusion model; Wherein, the Gaussian noise diffusion stage is used for: Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1; Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images; Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data; Inputting T of the dual-channel noisy image data into the residual self-attention U-net network to perform noise estimation to obtain estimated noise; The inference phase is used to: Generate a pure noise image of the same size as the input image; The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes: Based on the residual self-attention U-net network, the image obtained by the i-1th iterative processing is subjected to noise estimation to obtain the i-th estimated noise; The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

2. The method according to claim 1, characterized in that: The residual self-attention U-net network includes: The encoding stage includes a first convolutional layer, a first processing block and a residual module, wherein the first convolutional layer is used to map the data into a high-dimensional feature space, and the first processing block is used to reduce the spatial dimension, increase the number of feature channels, extract multi-scale features of the input data, and increase the abstraction level of feature representation; An intermediate stage, comprising two residual modules with a self-attention mechanism, for extracting global context information and capturing long-distance dependencies through the self-attention mechanism; The decoding stage includes a second processing block and a second convolutional layer, the second processing block is used to gradually restore the spatial dimension while reducing the number of feature channels, and the second convolutional layer is used to convert the number of feature channels in the decoding stage into the number of channels of the final output and output the result.

3. The method according to claim 2, characterized in that: The residual module with a self-attention mechanism includes: Two target modules, the target modules include a group normalization layer, an activation function and a third convolutional layer, one of the target modules is used to process low-level features, and the other target module is used to process high-level features; A feature adjustment layer, the feature adjustment layer is located between the two target modules; The fourth convolutional layer is used to adjust the number of channels of the feature map.

4. The method according to claim 1, characterized in that: The preprocessing of the original B-scan image to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image includes: Using a bicubic interpolation method to interpolate the original B-scan image to obtain the high-resolution data; Downsampling the original B-scan image using a bilinear downsampling method to obtain the low-resolution data; The low-resolution data is interpolated using a bicubic interpolation method to obtain the inference data, and the size of the inference data is the same as that of the high-resolution data.

5. The method according to claim 1, characterized in that: The ground penetrating radar data set includes a simulation data set and a measured data set, and the construction of the ground penetrating radar data set includes: Establish the detection scenario of underground targets, debug the parameters under the condition of simulation operation, obtain the critical value of the parameters, and generate low-resolution ground penetrating radar simulation data; A sand box was built to simulate the detection environment of real scenes. By controlling the environmental conditions and target configuration, echo signals in various real scenes were captured to generate low-resolution ground-penetrating radar measurement data. The simulation data set is constructed based on the ground penetrating radar simulation data, and the measured data set is constructed based on the ground penetrating radar measured data.

6. A ground penetrating radar B-scan image super-resolution processing device, characterized in that: include: A data set construction module, used to construct a ground penetrating radar data set, wherein the ground penetrating radar data set includes a plurality of original B-scan images; a preprocessing module, configured to preprocess each of the plurality of original B-scan images to obtain high-resolution data, low-resolution data and inference data corresponding to the original B-scan image; A model building module, used to build a conditional denoising diffusion model, wherein the conditional denoising diffusion model includes a Gaussian noise diffusion stage and an inference stage, wherein the Gaussian noise diffusion stage includes a residual self-attention U-net network; A training module, used for training the conditional denoising diffusion model based on a target data set to obtain a trained conditional denoising diffusion model, wherein the target data set includes high-resolution data, low-resolution data and inference data corresponding to the plurality of original B-scan images; A reconstruction processing module, used for performing super-resolution reconstruction processing on the image to be processed by using the trained conditional denoising diffusion model; Wherein, the Gaussian noise diffusion stage is used for: Generate T Gaussian noises with the same size as the input image and increasing noise intensity, where T is a positive integer greater than 1; Adding the input image and T Gaussian noises respectively to obtain T single-channel noisy images; Merge the T single-channel noisy images with the inference data corresponding to the input image to obtain T dual-channel noisy image data; Inputting T of the dual-channel noisy image data into the residual self-attention U-net network to perform noise estimation to obtain estimated noise; The inference phase is used to: Generate a pure noise image of the same size as the input image; The pure noise image is iterated T times to obtain a noise-suppressed image, wherein the i-th iterative processing includes: Based on the residual self-attention U-net network, the image obtained by the i-1th iterative processing is subjected to noise estimation to obtain the i-th estimated noise; The i-th estimated noise is subtracted from the image obtained by the i-1-th iterative processing to obtain an image obtained by the i-th iterative processing, where i and T are both positive integers greater than 1, and i is less than or equal to T.

7. An electronic device comprising: A memory, a processor, and a program stored in the memory and executable on the processor; wherein the processor is used to read the program in the memory to implement the steps in the ground penetrating radar B-scan image super-resolution processing method as described in any one of claims 1 to 5.

8. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps in the ground penetrating radar B-scan image super-resolution processing method according to any one of claims 1 to 5 are implemented.

Images