A deep learning ground penetrating radar non-uniform clutter suppression method fusing subspace projection

Abstract

The application discloses a kind of fusion subspace projection's deep learning ground penetrating radar non-uniform clutter suppression method, under the modeling of non-uniform wall body concrete and aggregate, the CBNet network architecture for clutter suppression is proposed, the advantages of projection mechanism and data-driven are used, the clutter suppression is carried out by image adaptive projection, the subspace base vector close to real target is provided, the information loss is reduced while realizing the improvement of high signal-to-clutter ratio, effectively suppresses superficial layer non-uniform clutter.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a ground penetrating radar non-uniform clutter suppression method fusing subspace projection and deep learning. BACKGROUND

[0002] Ground penetrating radar (GPR) technology is a geophysical detection method for detecting underground targets, and is one of the most promising and promising methods in current non-destructive detection technology. Compared with conventional non-destructive detection methods such as resistivity method and low-frequency electromagnetic induction method, ground penetrating radar technology has the advantages of fast detection speed, continuous detection process, high resolution, convenient and flexible operation, and low detection cost, and has been an international research hotspot. As an important non-destructive detection method, it has been widely used in civil engineering, archaeology, mineral exploration, military target detection and other fields.

[0003] In practical applications, there are many hazards affecting the safety of houses: thermal expansion and cold contraction caused by temperature changes around the house, uneven settlement of the foundation, and many other reasons that cause a large number of cracks in the wall; Corrosion of steel bars in the wall caused by various reasons can also cause damage to reinforced concrete; In addition, the detection and safety evaluation of wall heat preservation rivets or other wall materials are also one of the focuses of house detection. As an effective means for detecting cracks and steel bars in the wall, ground penetrating radar can be affected by wall reflection, direct coupling of transmitting and receiving antennas, and scattering of wall bonding layer and non-uniform cement mortar, which can cause clutter interference and affect the identification performance of ground penetrating radar for cracks and steel bars. Therefore, in order to solve the complex actual detection environment, the research on the shallow layer ground penetrating radar clutter suppression algorithm has very important practical significance.

[0004] Many researches have been made on clutter suppression algorithm of ground penetrating radar in the shallow layer at home and abroad. Subspace decomposition method mainly includes principal component analysis (PCA), independent component analysis (ICA) and singular value decomposition (SVD), and the basic idea is to decompose the clutter and target signals in the echo data into clutter components and target components, and then select the target components to form the target subspace. Such algorithms have good adaptability to different environments, but when there are multiple target signals in the echo data or the environmental clutter is complex, the clutter cannot be completely removed, and the selection of target components needs to be improved. The method based on low-rank and sparse matrix decomposition (LRSD) theory regards the echo data as the sum of low-rank matrix and sparse matrix. The clutter and target response are contained in the low-rank matrix and sparse matrix respectively, which can be separated by solving the LRSD problem. Representative methods include robust principal component analysis (RPCA), robust nonnegative matrix factorization (RNMF) and robust orthonormal subspace learning (ROSL), which have good effect and wide adaptability. However, these methods have certain limitations, such as the selection of regularization parameter has a great influence on the algorithm effect, and it is easy to distort the target signal, and the effect of removing non-uniform clutter is general.

[0005] In recent years, the method of deep learning has shown strong feature learning ability in practical application, including target recognition, medical image semantic segmentation, etc. In the aspect of ground penetrating radar clutter suppression, there are methods based on Convolutional Autoencoder, Gan and U-Net network architecture for clutter suppression. Driven by large data, the network has strong supervised learning ability and has achieved high signal-to-clutter ratio. However, the current data set model is still simple, and the non-uniform clutter in the shallow layer cannot be well characterized and suppressed, and the network's ability to represent target signals needs to be improved. SUMMARY

[0006] Therefore, the application provides a deep learning ground penetrating radar non-uniform clutter suppression method fusing subspace projection.

[0007] The technical scheme of the application is as follows:

[0008] The application provides a deep learning ground penetrating radar non-uniform clutter suppression method fusing subspace projection. First, a ground penetrating radar received signal model is established as a concrete background model containing aggregate, and a concrete non-uniform medium data set based on a random aggregate (pebble, gravel) throwing model is constructed. Then, the original data set is preprocessed based on median cancellation of clutter, and the direct wave and surface reflected wave are preliminarily removed to form a final data set after preprocessing. A CBNet network architecture for clutter suppression is constructed, specifically a U-Net neural network architecture combined with a subspace projection attention (SPA) gate. The SPA module is added in the process of skip connection from a low feature map to a high feature map of the U-Net network. The original data is supervised learning based on the data driving and projection mechanism. The input is the final data after preprocessing, and the label is the target data under the ideal condition without clutter. Finally, the trained network is used to output the clutter suppression result containing only the target in the radar image.

[0009] Advantages:

[0010] 1. The application is applied to the field of ground penetrating radar signal processing. The application compares various typical methods and deep learning methods (robust non-negative matrix factorization, robust principal component analysis, methods based on U-Net and CR-Net networks, etc.) through experiments. When the echo signal contains target signals and non-uniform clutter exists at the same time, the application can provide high-definition target signals and obtain a high signal-to-clutter ratio, and belongs to an effective clutter suppression method.

[0011] 2. The application simulates non-uniform clutter in an actual scene by using a concrete non-uniform simulation based on a random aggregate throwing model, which is close to the real scene and enhances the effectiveness of the data set.

[0012] 3. The application combines the subspace projection attention gate with the network, uses the advantages of the projection mechanism and data driving, performs clutter suppression through image adaptive projection, provides subspace base vectors close to the real target, reduces information loss, and realizes the improvement of the high signal-to-clutter ratio. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a processing flowchart of an embodiment of the present application.

[0014] Figure 2 is a schematic diagram of a set of simulation data in the method of the present application. Wherein, (a) is the original data diagram,

[0015] (b) is a marker diagram.

[0016] Figure 3 is a schematic diagram of the value cancellation preprocessing result in the method of the present application.

[0017] Figure 4 is a schematic diagram of the network structure proposed in the method of the present application. Wherein, (a) is a general block diagram, and (b) is a schematic diagram of the SPA module.

[0018] Figure 5 is a clutter suppression result diagram of the measured data 1 by different methods of the present application; wherein, (a) is the original image, (b) is RNMF, (c) is RPCA, (d) is U-Net, (e) is CR-Net, and (f) is CBNet.

[0019] Figure 6 is a clutter suppression result diagram of the measured data 2 by different methods of the present application. Wherein, (a) is the original image, (b) is RNMF, (c) is RPCA, (d) is U-Net, (e) is CR-Net, and (f) is CBNet. DETAILED DESCRIPTION

[0020] The present application will be described in detail below with reference to the accompanying drawings and examples.

[0021] The present application provides a deep learning ground penetrating radar non-uniform clutter suppression method fusing subspace projection, Figure 1 is a signal processing flowchart of an embodiment of the present application. As shown in Figure 1 , the present application is realized by the following steps:

[0022] Step 1: Establish a ground penetrating radar receiving signal model, and construct a concrete non-uniform medium data set based on a random aggregate mesoscopic model;

[0023] The received echo signal can be expressed as:

[0024] X=S+C+B+E

[0025] Assume that the received signal X is an M x N matrix, M is the number of sampling points of each detection point, N is the number of detection points, that is, the number of acquisition channels, S is the target signal, this part of the signal is relatively weak, and the waveform and size of the signal depend on the size, shape and material of the target and the environment in which the target is buried. C is the direct coupling wave of the transmitting and receiving antenna, and the size depends on the design and position of the antenna. For ground penetrating radar commonly using dipole antenna, the antenna mutual coupling wave can be generally considered as a constant value, which is independent of the measurement position. B is the surface reflection echo and the reflection echo of non-uniform background medium, this part of the signal is relatively strong, and is also the part that needs to be suppressed most. The surface clutter is usually related to the antenna height from the ground, soil type and uniformity, and ground roughness. The non-uniform medium wave usually has certain randomness, and is usually related to the uniformity of the detection scene and the distribution of dielectric constant. The direct coupling wave of the transmitting and receiving antenna and the ground reflection echo usually occupy a large energy in the echo data. E is the receiver noise signal, which is also a random noise.

[0026] Concrete is a kind of heterogeneous composite material, mainly composed of coarse aggregate, fine aggregate, cement hydration paste, micro-cracks and pores, etc. The internal dielectric constant of the composite material also changes randomly. Therefore, it is necessary to preliminarily model the concrete medium.

[0027] The model is based on the aggregate gradation curve, which describes the probability of the occurrence of any diameter aggregate in the interface. The probability of the occurrence of any diameter aggregate in the cross section of concrete is:

[0028]

[0029] In the formula: P C (D) is the probability of any point on the section being located on the aggregate particle with a particle size less than D, that is, the percentage of the cross-sectional area of the cross-sectional circle with a diameter less than D to the cross-sectional area, P K is the percentage of the volume of the aggregate to the total volume of the concrete; D max is the maximum particle size of the aggregate. The model is based on the gradation curve.

[0030] The above formula is used to simulate the required aggregate particle size, and the function value is subjected to Monte Carlo sampling to obtain a particle size sequence that meets the gradation requirements. Then, the aggregate is placed in the simulated scene area according to the aggregate stacking rules. The main steps include determining the aggregate control parameters (center coordinates and radius) that meet the random distribution, judging whether the newly generated aggregate overlaps with the generated aggregate. Among them, the aggregate is a random polygon, and the area of the aggregate is recorded as a percentage of the total area during the placement process. When the content meets the preset value, the placement is stopped.

[0031] After the completion of the delivery, since the final random medium model array is to be imported into gprMax (electromagnetic simulation software based on the finite difference time domain (FDTD) method), background mesh partitioning is required to adjust the grid size to simulate different accuracies. A grid with uniform distribution but a scale much smaller than the aggregate is established in the simulation area, and the grid cells are assigned properties (dielectric constant) according to the distribution of aggregates, mortar and interface space. Subsequent calculations will only be based on the translated grid after assignment. Finally, the obtained concrete simulation data is imported into gprMax to generate a B-scan image of non-uniform concrete medium with random characteristics. Scripted simulation control can be achieved by writing a Python file to call the gprMax software to generate batch training data under different scenarios. The following table shows the specific simulation parameters:

[0032] Scenario size 110 cm x 30 cm Aggregate size 0.6-2 cm Rebar radius 0.3-1.2 cm Aggregate content 30％-60％ Burial depth 3-8 cm Concrete ε 7.0 Antenna frequency 2.6 G GSSI Aggregate ε 8.5

[0033] As Figure 2 (a) is the generated non-uniform medium concrete aggregate B-scan, a total of 4000 original data sets are generated.

[0034] Step 2: Preprocess the original data set to preliminarily remove direct waves and surface reflected waves to form the final data set after preprocessing;

[0035] First, the input data is preprocessed using the clutter suppression method based on median cancellation. The basic principle of median cancellation is to first take the median of all channels at each sampling time to form a "median channel", and then sequentially cancel all channels to highlight the hyperbolic target characteristics and residual non-uniform clutter, so that the algorithm has high applicability to complex scenes. The results of the clutter suppression method based on median cancellation are shown in Figure 3 .

[0036]

[0037] where s(n, t) represents the nth A-scan data, MED represents the median operation, The nth A-scan data after clutter suppression by median cancellation. The essence of median cancellation to suppress clutter is to subtract the median of the row in which each element of the radar B-scan echo data is located from each element, so as to remove the horizontal phase axis and eliminate the strong horizontal interference of the B-scan echo. In the case of detecting a flat and uniform surface, the delay and amplitude variation of the direct wave and surface reflected wave components in each A-scan are small, and the target signal and non-uniform clutter have different amplitudes in each A-scan data. Therefore, the median of the row of the B-scan data matrix is approximately equal to the direct wave value at each measurement point in the row, so that the median cancellation method can effectively suppress the direct clutter component. It can be seen from the principle of median cancellation that this method cannot effectively suppress the non-uniform scattering clutter of the medium and the noise component, because these clutter components are randomly distributed in each A-scan data and have large spatial variation. The median cancellation operation is performed to eliminate the direct wave and the flat surface reflected wave, and the residual non-uniform scattering wave still needs to be processed by subsequent methods.

[0038] Step 3: Construct the CBNet network architecture, and perform supervised learning on the original data based on the data driving and projection mechanism. The input is the final data after preprocessing, and the label is the target data under ideal conditions without clutter.

[0039] The CBNet network structure is a U-Net neural network combined with a subspace projection attention (SPA). The SPA is added in the process of the jump connection of the low feature map to the high feature map of the U-Net, and the advantages of the projection mechanism and the data driving are utilized to perform clutter suppression through image adaptive projection, to provide a subspace basis vector close to the real target, and to eliminate irrelevant and noisy responses generated by the jump connection.

[0040] U-Net is an algorithm for semantic segmentation using a fully convolutional network proposed in 2015. It uses a symmetric U-shaped structure containing a compression path and an expansion path. The name of the network is also derived from its U-shaped shape. The left side of the network is a series of down-sampling operations composed of convolution and max-pooling, called the compression path. The right part of the network is called the expansion path, which adopts an encoder-decoder structure. Due to its good feature representation characteristics, it is widely used in medical images and other fields.

[0041] The subspace projection attention door SPA is shown in Figure 4As shown in (b), the concept of image projection is mainly utilized. Projection mainly involves two steps: first, basis generation: generating subspace basis vectors from the image feature map, and then reconstructing the image within the subspace covered by these basis vectors; second, projection: converting the feature map into a signal subspace. Since natural images usually belong to a low-rank signal subspace, by correctly learning and generating basis vectors, the reconstructed image can retain most of the original information and suppress clutter and noise unrelated to the generated basis set. However, the subspace basis in image analysis is artificially constructed, and image information and noise often reside in a single basis. Removing this basis set results in a loss of image details. Therefore, this invention utilizes deep learning to automatically learn subspace basis vectors, which can better distinguish the distribution of targets and clutter.

[0042] We assume X1, X2 ∈ R H×W×C These are two feature maps from different layers of a single image in the network. The low-level feature map is X1, and the high-level feature map is X2. H, W, and C represent the height, width, and number of channels of the feature maps, respectively. First, based on X1 and X2, r basis vectors [a1; a2; ...; a...] are estimated. r ], each a i ∈R N These are the basis vectors of the signal subspace, where N = HW. Then X1 is transformed into a. i The subspace represented is the projection of the original feature map.

[0043] Specifically applied to network architecture, the subspace basis A is derived from X1 and X2 through the function f. θ Mapping generation:

[0044] A = f θ (X1, X2)

[0045] Where A∈R HW×r The function mapping f with θ as parameter θ :(R H×W×C R H×W×C → R HW×r Generated by residual convolution blocks, where r is the number of subspace bases, and A = [a1; a2; ...; a...]. r ] is a matrix composed of basis vectors. We implement the function f using a small convolutional network. θ First, concatenate X1 and X2 along the channel dimension to obtain X∈R. H×W×2C The data is then fed into a shallow residual convolutional block with r output channels, and its output dimension can then be converted to HW×r. The weights and biases of the basis vector generation block are updated end-to-end during training.

[0046] After the basis vector generation, the projection part follows. Given the matrix A mentioned above, whose columns are the basis vectors of the r-dimensional signal subspace γ, the image feature X1 can be projected onto γ by the orthogonal linear projection. Assume P: R N → γ is the orthogonal projection matrix of the signal subspace, P s can be calculated from A:

[0047] P s = A(A T A) -1 A T

[0048] where the normalization term (A T A) -1 is to ensure the orthogonality of the basis vectors during the basis vector generation process. Finally, the image feature map X1 can be reconstructed in the signal subspace as Y:

[0049] Y = P s X1

[0050] Y is the output after the subspace projection. The operation in the projection is a pure linear matrix operation, which is differentiable and can be easily implemented in the neural network framework. Combining the basis vector generation and the subspace projection, the SPA module structure used is shown in Figure 4 (b).

[0051] The CBNet architecture used is shown in Figure 4 (a). The overall structure is based on the classic U-Net architecture. CBNet has 4 encoders and 4 corresponding decoders, where the feature maps are down-sampled by a convolution with a step of 2 to 1 / 2 scale at the end of each encoder, and up-sampled to 2 times scale by deconvolution before each decoder. The skip connection transmits the large-scale low-level feature map from each encoder to its corresponding decoder. ReLU is used as the activation function of each convolution layer.

[0052] The SPA module is placed in each skip connection. Since the feature map from the low level contains more detailed original image information, the low-level feature map is taken as X1, the high-level feature is taken as X2, and they are fed into the SPA module. The low-level feature map from the skip connection is projected into the signal subspace guided by the up-sampled high-level feature. Then, the projected feature is fused with the original high-level feature, and output to the next decoder stage. Compared with the traditional U-Net-like architecture which directly fuses the low-level and high-level feature maps at each decoder stage, the main difference of CBNet is that the low-level feature is projected by the SPA module before fusion. After that, the output of the last decoder is converted to 1 channel by a 1x1 convolution layer to estimate the clutter-free data and output the result after clutter suppression.

[0053] Step 4: The trained network is used to finally output the clutter suppression result containing only the target in the radar image;

[0054] The network input is the data after median cancellation preprocessing, the label is the target data under ideal no-clutter (obtained by cancellation with the empty scene without target), and the output is the clutter suppression result containing only the target. The training data set is to adjust the size of the B-scan data and its corresponding label to 256x80 to input the network, and through the subspace projection based on the neural network, the CBNet learns to generate a set of base vectors for the signal subspace. By projecting the input feature map into this space, the signal can be enhanced after reconstruction to facilitate separation from clutter and noise. This method solves the problem of base vector selection and target component selection in traditional subspace methods, uses the subspace projection physical model rule, and at the same time has the excellent performance of deep learning methods, greatly improves the image quality, effectively suppresses non-uniform clutter and noise, and has important significance for subsequent target recognition and other processing.

[0055] Embodiment

[0056] In order to verify the deep learning ground penetrating radar non-uniform clutter suppression method based on subspace projection fusion proposed in the present application, experiments are performed on GPR simulation data and real data of measured environment to verify the effectiveness of various algorithms, including robust non-negative matrix factorization (RNMF), robust principal component analysis (RPCA), method based on U-Net architecture, method based on CR-Net architecture, and method based on CBNet proposed in the present application. The performance of each algorithm is quantitatively compared and analyzed by improving the factor performance index.

[0057] Since visual inspection is not sufficient to verify the performance of the algorithm in most cases, multiple performance evaluation indexes are needed to verify the effect of clutter suppression. The present application adopts the index of improvement factor. The improvement factor compares the signal-to-clutter ratio (signal-to-clutter ratio is the energy ratio of target signal and pseudo-target signal to background signal) of the suppressed image with the signal-to-clutter ratio of the image before suppression to determine the effectiveness of clutter suppression and the degree of improvement of signal-to-clutter ratio. The larger the value, the greater the improvement.

[0058] The simulation data is generated by the software gprMax based on the finite difference time domain (FDTD) method. The target is an ideal conductor PEC with a radius of 0.3-1.2 cm, a buried depth of 3-8 cm, and a non-uniform aggregate content of 30-60%. The obtained GPR image is shown in Figure 2 (a), and the training set has 4000 pairs of original data and no-clutter data (obtained by cancellation with the empty scene, as shown in Figure 2(b) shown), the training set test set division ratio is 8:2. CBNet is implemented on an NVIDIA 3090 GPU using PyTorch. The network weights are initialized using a standard Gaussian function, and a total of 100 epochs are trained. The batch size is 4, and the end-to-end training is performed based on the MSE loss function, with an initial learning rate of 1e-4 and a network optimizer of the Adam optimizer.

[0059] The measured data 1 uses a GSSI SIR4000 radar system, and the antenna center frequency is 2600MHz. The detection of steel bars in concrete is completed, the underground medium is concrete, the target is three threaded steel bars with a diameter of 16mm, and the obtained GPR image is as shown in (a) of FIG. 1. Figure 5 As can be seen from the figure, due to the large intensity of the antenna direct coupling wave and the ground surface reflection wave, and the non-uniformity of the concrete environment, the target echo signal is covered. After processing by various clutter suppression algorithms, the results are as shown in (b)-(f) of FIG. 1. Figure 5 As can be seen from the figure, the hyperbolic curve of the target has been completely revealed, that is, the antenna direct coupling wave and the flat ground surface reflection wave can be filtered out, but there are still residual non-uniform clutter near the target. Compared with the low-rank sparse decomposition-based robust non-negative matrix factorization method and the robust principal component analysis method, the UNet and CR-Net methods based on network architecture can further remove the non-uniform clutter, but there are still some residual noise and clutter, and the target echo is not clear enough. The method based on the CBNet architecture proposed in the present application has the best processing effect, improves the target clarity, and basically retains the target signal while suppressing most of the non-uniform clutter, with the highest improvement factor index.

[0060] Table 1: Index comparison of measured data 1 after processing by different algorithms

[0061] RNMF RPCA U-Net CR-Net CBNet IF / dB 14.77 13.30 20.68 17.05 22.84

[0062] The measured data 2 uses a GSSI SIR4000 radar system, and the antenna center frequency is 2600MHz. The detection of steel bars in concrete is completed, the underground medium is concrete, the target is a steel pipe with a diameter of 25mm and a steel bar with a diameter of 20mm, and the obtained GPR image is as shown in (a) of FIG. 2. Figure 6 As can be seen from the figure, the two approximate horizontal lines are the upper and lower surfaces of the concrete medium, and there are many clutters. After processing by various clutter suppression algorithms, the results are as shown in (b)-(f) of FIG. 2. Figure 6The performance indicators are shown in Table 2. Although the robust non-negative matrix decomposition method and the robust principal component analysis method can remove the direct coupling wave and the surface reflection wave, there are still more residual reflection waves, and the non-uniform medium reflection wave cannot be effectively processed. The method based on the U-Net and CR-Net architecture can remove the surface reflection wave to a greater extent, and has a certain inhibitory effect on the non-uniform medium reflection wave, thereby improving the clarity of the target. The method based on the CBNet architecture proposed in the application not only further removes the surface reflection wave, but also ideally suppresses the non-uniform medium reflection wave, highlights the target information, and has the highest improvement factor, thereby verifying the effectiveness of the algorithm.

[0063] Table 2: Index comparison of measured data 2 after processing by different algorithms

[0064] RNMF RPCA U-Net CR-Net CBNet IF / dB 29.84 27.62 30.71 30.82 31.95

[0065] The above experimental results show that in the ground penetrating radar clutter suppression process, the deep learning ground penetrating radar non-uniform clutter suppression method fusing subspace projection proposed in the application can effectively suppress non-uniform clutter, obtain a high signal-to-clutter ratio, and maximize the preservation of target signals, thereby verifying the effectiveness and robustness of the algorithm. Compared with existing network methods, the method proposed in the application simulates non-uniform clutter in actual scenes by simulating non-uniform concrete based on a random aggregate placement model, which is close to the real scene and enhances the effectiveness of the data set. In terms of network architecture, the subspace projection attention door is combined with the U-Net network, which takes advantage of the projection mechanism and data-driven advantage, and suppresses the clutter through image adaptive projection, provides subspace basis vectors close to the real target, reduces information loss, and realizes the improvement of high signal-to-clutter ratio.

[0066] In conclusion, the above is only a preferred embodiment of the application, and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A deep learning ground penetrating radar non-uniform clutter suppression method based on fused subspace projection, characterized in that, The method comprises: Step one, establishing a ground penetrating radar receiving signal model as a concrete model containing aggregate, and constructing a concrete non-uniform medium data set based on a random aggregate throwing model; Step two, pre-processing the original data set to preliminarily remove direct waves and surface reflected waves, and forming a final data set after pre-processing; Step three, construct CBNet network architecture, based on data-driven and projection mechanism for original data supervised learning, input is the final data after pretreatment, label is the target data under ideal no clutter; CBNet network structure is U-Net neural network combined with subspace projection attention door SPA, add SPA in the process of U-Net low feature map to high feature map skip connection; Subspace projection attention door SPA mainly includes two stages of subspace base generation and feature map projection, assuming , are two feature maps of different layers of a single image in the network, respectively are height, width and channel number, then the subspace base is generated by , through the function mapping: ； wherein Parametric function mapping generated by the residual convolutional block, is the number of subspace bases, then the orthogonal projection matrix of the signal subspace is: ； After obtaining the projection matrix, the image feature map may be reconstructed in the signal subspace as : ； The output after the projection of the feature subspace; Step four, using the well-trained network, finally output the clutter suppression result containing only the target in the radar image.

2. The method of claim 1, wherein the method is a fusion subspace projection based deep learning GPR clutter suppression method. The pre-processing method in step two is wall clutter suppression based on median cancellation. The basic principle of median cancellation is that a "median channel" is first formed by taking the median of all channels at each sampling time, and then all channels are sequentially cancelled, thereby highlighting the hyperbolic target characteristics and residual non-uniform clutter. In the case of a flat and uniform detection surface, the direct wave and surface reflected wave components in each A-scan have little delay and amplitude variation, and have spatial slow variation, while the target signal and non-uniform clutter have different amplitudes in each A-scan data. Therefore, the median of the rows of the B-scan data matrix is approximately equal to the direct wave value at each measurement point in the row, so the median cancellation method can effectively suppress the direct clutter component, and the residual non-uniform scattered wave still needs to be processed by subsequent methods.

3. The method of claim 1, wherein the method is a fusion subspace projection based deep learning GPR clutter suppression method. The network structure of CBNet in step three is a U-Net neural network combined with a subspace projection attention door (SPA); the overall structure is based on the classic U-Net architecture, and CBNet has four encoders and four corresponding decoders, wherein the feature map is down-sampled by 1 / 2 scale with convolution in steps of 2 at the end of each encoder, and up-sampled to 2 times scale by deconvolution before each decoder; the skip connection transmits large-scale low-level feature maps from each encoder to its corresponding decoder; at the same time, ReLU is used as the activation function of each convolutional layer; in addition, the SPA module is added in the process of the skip connection of the low feature map to the high feature map of the U-Net, and the low-level feature map is taken as , the high-level feature is taken as , fed into the SPA module, then the projected feature is fused with the original high-level feature and output to the next decoder stage; the output of the last decoder is converted to 1 channel by a 1x1 convolutional layer to estimate the clutter-free data, and the result after clutter suppression is output; by using the advantages of the projection mechanism and data-driven, the clutter suppression is performed through image adaptive projection, the subspace basis vector close to the real target is provided, at the same time the irrelevant and noisy response generated by the skip connection is eliminated, the information loss is reduced, and the high signal-to-clutter ratio is improved; the input of the network is the data after median cancellation preprocessing, the label is the target data under ideal clutter-free, which is obtained by cancellation with the empty scene without target, and the output is the clutter suppression result containing only the target.

Images