Target feature extraction method combining GPR physical mechanism and CNN

Abstract

A target feature extraction method combining a GPR physical mechanism and a CNN comprises the following steps: establishing a scattered field equation of a GPR underground semi-infinite space to obtain parameter information used for representing target scattering characteristics in GPR echo signals; factors influencing target characteristic intensity in GPR echo signals are analyzed, and common characteristics, including intensity characteristics and geometric structure characteristics, represented by a target in a GPR B-SCAN image are analyzed; designing a convolutional neural network model to extract the features of the B-SCAN image, and completing the recognition of a target feature hyperbolic curve through a classifier network. The method can reduce the situation of target feature hyperbola misjudgment to a great extent, and has practical application significance.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition processing, and in particular to a target feature extraction method combining GPR physical mechanism and CNN. Background Art

[0002] Since there are various randomly distributed interference sources in real underground scenes, in order to obtain more information, ground penetrating radar (GPR) generally uses the method of transmitting ultra-wideband electromagnetic pulse signals to detect targets. The echo data obtained will contain more interference information, which puts higher requirements on GPR data processing methods. At present, both at home and abroad, ground penetrating radar application technology is vigorously developed. GPR data application technology is mainly reflected in cross-disciplinary subjects such as earth, microwave, antenna, computer, information, and intelligent processing. The effective interpretation method of GPR data is mainly to obtain information such as the distribution characteristics of the physical parameters of the underground medium and the geometric structure parameters of the buried target. This method is a bottleneck problem that restricts the application of GPR detection technology. In particular, the ability to obtain GPR data application in actual measurement scenes is still relatively weak at this stage. It mainly relies on the qualitative interpretation of personnel with rich engineering experience, which is time-consuming, labor-intensive, and inefficient.

[0003] Since there are usually some differences in electrical characteristic parameters between underground targets and background media, the greater the difference in electrical characteristic parameters (mainly the dielectric constant of the medium), the more obvious the intensity characteristics of the target in the echo signal. At the same time, considering that in the reflection measurement mode, the ground penetrating radar is affected by the coupling electromagnetic waves between the receiving and transmitting antennas and the electromagnetic waves reflected from the earth's surface, there will be strong direct waves in the echo signal that interfere with the signal. For this reason, it is urgent to mine the important feature information in the GPR data from the perspective of machine vision and conduct quantitative analysis and processing to improve the readability of the results.

[0004] In recent years, in view of the breakthroughs of DCNN in the fields of optical image segmentation, object detection and recognition, GPR data has been widely used by more and more researchers in the fields of GPR nondestructive testing for military and public works. DCNN models the network by constructing positive and negative samples, and automatically extracts the features of the target from the training samples. CNN is a feature extraction structure with more hidden layers, which improves the robustness of the detection technology and creates opportunities for its use in complex GPR environments. Regarding the problem of GPR underground target detection, existing data have proved that CNN is superior to traditional artificially designed CNN in terms of the accuracy of obtaining object targets.

[0005] The model-driven GPR data inversion method usually takes the MAXWELL equations as the theoretical basis and combines the GREEN function to solve the parameters. The shape, size, position and dielectric parameters of the underground buried target can be obtained by inverting the GPR inverse scattering field. At present, the methods for extracting B-SCAN image feature information and inverting GPR parameters are mainly based on the physical mechanism of GPR and computer vision technology, striving to interpret GPR information intelligently. By simulating the human vision and analysis process through computers, the intelligent processing of GPR image information is realized, that is, the feature extraction process is used to simulate the human perception of underground buried targets, and the learning of the classifier network is used to simulate the human recognition process. From the application level, the human-computer interaction GPR data interpretation system currently still requires the operator to have a certain engineering background and experience, which greatly hinders the development of applications. The data-driven GPR B-SCAN image hyperbola detection method only uses the structural characteristics of the target, and does not fully utilize the characteristics of the B-SCAN image itself under the constraints of the GPR physical mechanism.

[0006] Based on the in-depth study of GPR physical model drive, combining data-driven methods based on deep learning technology to mine important information in B-SCAN images, improve the level of automation of GPR information interpretation, and reduce human participation factors in the human-computer interaction process is an effective way to achieve intelligent processing application bottlenecks. Summary of the invention

[0007] The present invention proposes a target feature extraction method combining the physical mechanism of GPR and CNN. Under the physical mechanism constraint of GPR, a two-level cascade structure DCNN is designed, which takes advantage of the characteristics of large amplitude and horizontal linear structure of direct wave. First, the first-level CNN is used to locate the area where the direct wave is located in the B-SCAN image, and then the zeroing method is adopted to eliminate it; then, based on the first-level CNN, the second-level CNN is used to extract deep semantic features of the B-SCAN image, and the classifier network is used to identify the characteristic hyperbola and interference signal in the image. This method can greatly reduce the misjudgment of the target feature hyperbola, and has more practical application significance.

[0008] The technical solution adopted by the present invention is:

[0009] A target feature extraction method combining GPR physical mechanism and CNN includes the following steps:

[0010] Step 1: Establish the scattering field equation of the semi-infinite space under the GPR to obtain the parameter information in the GPR echo signal used to characterize the scattering characteristics of the target;

[0011] Step 2: Analyze the factors that affect the intensity of target features in the GPR echo signal, and analyze the common features of the target in the GPR B-SCAN image, including intensity features and geometric structure features;

[0012] Step 3: Design a convolutional neural network model to extract the features of the B-SCAN image and use the classifier network to complete the recognition of the target feature hyperbola.

[0013] In step 1, the scattering field equation of the semi-infinite space under the GPR includes equation (9) and equation (10):

[0014]

[0015] k 2 =-εμω 2 +iσμω (10);

[0016] In formula (9), E0 represents the amplitude of the electric field strength, ω is the angular frequency, r is the displacement vector, E is the electric field strength, e refers to a natural constant, i is an imaginary unit, and k=α+iβ represents the wave number.

[0017] In formula (10), k is the wave number, ε is the dielectric constant, μ is the magnetic permeability of the medium, ω is the angular frequency, i is the imaginary unit, and σ is the conductivity.

[0018] The parameter information used to characterize the target scattering characteristics in the GPR echo signal is specifically: attenuation factor α, phase shift coefficient β, tangent value tanδ of the dielectric loss angle δ, and electromagnetic wave propagation phase velocity v.

[0019] In step 2, the reflected wave is received by the GPR receiving antenna, and the refracted wave is transmitted to the next layer of the medium. The attenuation coefficient of the electromagnetic wave in the lossy medium is set to the attenuation coefficient of the electromagnetic wave in the i-th layer of the medium. i for:

[0020]

[0021] In formula (19), η i is the attenuation coefficient of electromagnetic waves; e is the natural logarithm; j is the imaginary unit; ω k is the kth frequency component of the electromagnetic wave pulse signal; d i represents the depth of the i-th layer of medium; c represents the speed of light in vacuum; ε i and σ i represent the relative permittivity and conductivity of the i-th layer medium respectively.

[0022] Let the scattering point (x p ,y p ) and the measuring point (x q ,yq ) is r pq , at the measuring point (x q ,y q ) The target signal measured at It is expressed as:

[0023]

[0024] In formula (20), is the target signal; x q ,y q are the horizontal and vertical coordinates of the measuring point respectively; α is the coefficient; γ p For point P (x p ,y p ) at ; ζ is the delay operator; e is the emitted electromagnetic pulse signal.

[0025] Round trip delay τ pq for:

[0026]

[0027] In formula (21), is rounded down; ε r is the dielectric constant of the medium; c is the speed of light in a vacuum; r pq is (x p ,y p ) and (x q ,y q ) between the two.

[0028] Combining formula (19) and formula (20), it can be seen that the GPR echo signal is affected by many factors. When the reflection characteristic hyperbola of the target can be observed in the B-SCAN image, the dielectric constant is the most important factor compared with other parameters. When the difference in dielectric constant between underground media increases, the reflection phenomenon at the interface will become more intense. As the difference in dielectric constant between media increases, the reflection at the interface will also become more intense, making the hyperbolic characteristic information on the B-SCAN image more obvious.

[0029] In step 3, a convolutional neural network model is used to extract features of the B-SCAN image, which specifically includes the following steps:

[0030] S3.1: By extracting the feature edge information of the object in the B-SCAN image, the feature information corresponding to the target object in the B-SCAN image is obtained;

[0031] In the B-SCAN image, the second-order derivative corresponding to the area with small or uniform intensity change is very small and close to 0; while the second-order inverse of the area with large intensity change corresponds to a larger value. This process can be described by the following formula:

[0032]

[0033] In formula (1), is the second-order derivative of the image position function; f(x,y) is the coordinate position corresponding to the target; x represents the horizontal coordinate corresponding to the target; y represents the vertical coordinate corresponding to the target; f(x+1,y+1) means that the image coordinate position is (x+1,y+1); f(x+1,y-1) means that the image coordinate position is (x+1,y-1); f(x-1,y+1) means that the image coordinate position is (x-1,y+1); f(x-1,y-1) means that the image coordinate position is (x-1,y-1); f(x+1,y) means that the image coordinate position is; f(x-1,y) means that the image coordinate position is (x-1,y); f(x,y+1) means that the image coordinate position is (x,y+1); f(x,y-1) means that the image coordinate position is (x,y-1).

[0034] S3.2: When the difference between the target and the background in the B-SCAN image is greater, the gradient change will be greater. In the convolutional neural network model, the corresponding feature information is obtained by convolution of the input image. Let the value of the neuron in the jth feature map of the i-th layer be for:

[0035]

[0036] In formula (2), g(·) is the activation function; is the weight of connecting the jth feature of the i-th layer to the mth feature map of the i-1th layer, b (i) is the bias term of the i-th layer; Represents the value of the neuron in the mth feature map of the i-1th layer; M represents the number of feature map rows of the input target image; N represents the number of feature map columns of the input target image.

[0037] The convolutional neural network model CNN uses convolution kernels of different scales to perform convolution calculations on the input f(x, y) to obtain the corresponding feature map.

[0038] S3.3: From the above prior conditions, it can be seen that in the B-SCAN image, the intensity of the target object is greater than the intensity of the background medium. In terms of amplitude, the main value component of the feature map corresponds to the target characteristic information. The process of extracting the characteristic information of the input target through L feature mappings of an input B-SCAN image is expressed as:

[0039]

[0040] In formula (3), is the feature image of the target object; f Σ [·] is the feature fusion implementation module; is the sth feature map of the last layer (L layer) of convolution operation; s is the number of convolution operation layers; To perform L feature mappings; is the l-th feature image; w is the width of the input image Ι; h is the height of the input image Ι; M is the number of channels of the input image Ι; C is the convolution kernel in the convolution layer; k is the size of C; N is the number of C; D is the number of channels of C, the number of channels of the convolution kernel is equal to the number of channels of the input image. Since the GPR B-SCAN image is a single channel, D=M=1 is taken here; For Feature image obtained after pooling; To perform pooling operation on the l-th feature image. S3.4: After performing continuous convolution and pooling operations on B-SCAN to obtain the corresponding feature map, The objects and noise therein are further identified. Since the target in the B-SCAN image under the reflection measurement mode usually has a hyperbolic structural feature and the noise is generally random, the target feature hyperbola can be further identified through the classifier network.

[0041] In step 3, a two-stage cascaded CNN is designed, as follows:

[0042] The cascade network structure of the first layer is a direct wave feature extraction network. For the direct wave feature extraction network, convolution operation is first used to extract features, and then feature dimension reduction is performed through pooling. The size of the convolution output feature map is calculated as:

[0043]

[0044] In formula (22), The width of the input feature map for the convolutional layer; The height of the input feature map of the convolution layer; is the width of the convolution output feature map; is the height of the output feature map, and the size of the convolution kernel matrix is ​​F conv ×F conv ; P0 represents the number of zero paddings at the boundary of the feature map matrix; S conv Indicates the stride of the convolution operation.

[0045] The size of the output feature map of the Pooling layer is calculated as:

[0046]

[0047] In formula (23), Respectively represent the width and height of the input feature map of the pooling layer; Respectively represent the width and height of the pooled output feature map; S pooling represents the pooling step size, and the size of the pooling kernel matrix is ​​F pooling ×F pooling .

[0048] In the direct wave feature extraction network, the output feature map contains horizontal linear structural feature information with large amplitude, that is, the global structural information corresponding to the direct wave in the B-SCAN image is extracted to determine its spatial position.

[0049] The cascade network structure of the second layer is the target feature extraction and classification network. First, the features of the target object are extracted. The cascade network adds 3 convolutional layers and reduces 1 pooling layer, and uses the deconvolution network to sample and calculate the output feature map. The corresponding category identification is constructed using the obtained target feature image and used to train the classification recognition network.

[0050] The present invention provides a target feature extraction method combining the physical mechanism of GPR and CNN, and the technical effects are as follows:

[0051] 1) The method of target characteristic hyperbolic detection combining the physical mechanism of GPR and CNN proposed in the present invention is based on the intensity characteristics and geometric structure characteristics of the target in the echo signal, and a CNN network model with a corresponding two-level cascade structure is designed to improve the accuracy of target characteristic hyperbolic detection.

[0052] 2) The present invention proposes a method that combines GPR physical model-driven and data-driven. According to the intensity characteristics and geometric structure characteristics of the target in the echo signal, a corresponding two-level cascade structure CNN network model is developed to improve the accuracy of detecting the target feature hyperbola.

[0053] 3) The method of the present invention can greatly reduce the misjudgment of the target feature hyperbola and has great application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0055] Figure 1 Schematic diagram of GPR electromagnetic scattering phenomenon.

[0056] Figure 2(a) is a schematic diagram of the reflection measurement model;

[0057] Figure 2(b) is a schematic diagram of the hyperbolic features of the B-SCAN image.

[0058] Figure 3This is a simulation diagram of the dielectric attenuation model.

[0059] FIG4( a ) is a model diagram of a GPR detecting underground scene using a reflection measurement method;

[0060] Figure 4(b) is a GPR B-SCAN simulation diagram.

[0061] Figure 5 For different r Schematic diagram of the target scattering intensity under .

[0062] Figure 6 The figure is a flow chart of the B-SCAN image feature extraction method based on cascaded CNN.

[0063] Figure 7 To extract the loss curve of the network training model. DETAILED DESCRIPTION

[0064] A target feature extraction method combining the physical mechanism of ground penetrating radar (GPR) and convolutional neural network (CNN). The present invention intends to use GPR B-SCAN as a research object, carry out research on a data-driven GPR B-SCAN image feature hyperbola detection algorithm, and solve its problems such as false detection and missed detection. An electromagnetic scattering model of the GPR underground semi-infinite space is established, and based on the MAXWELL equation, the scattering field equation of the GPR underground semi-infinite space is established to extract parameter information reflecting the surface scattering characteristics. At the same time, the main factors affecting the intensity of target features in the echo signal are analyzed and discussed, and the common features of the target in the GPR B-SCAN image are summarized, such as intensity features and geometric structure features. On this basis, a corresponding convolutional neural network model is designed to extract the feature map of the B-SCAN image, and the classifier network is used to complete the recognition of the target feature hyperbola.

[0065] The specific steps include:

[0066] Step 1: Using the electromagnetic scattering model of the GPR underground semi-infinite space and based on the solution of the Maxwell equation, the scattering field equation of the GPR underground semi-infinite space is derived to obtain parameter information in the echo signal for characterizing the target scattering characteristics.

[0067] When electromagnetic waves propagate in the underground half-space, when the relative dielectric constant appears in a medium with different physical parameters, electromagnetic physical phenomena such as reflection, refraction and diffraction will occur. Figure 1In the different scattering paths shown, L1 indicates that the electromagnetic wave emitted by the GPR is directly reflected to the receiving antenna at the upper boundary of the target; L2 indicates that the refracted wave is reflected again to the receiving antenna at the lower boundary of the target; L3 indicates the creeping wave formed by the diffraction of the electromagnetic wave on the surface of the target object; L4 indicates the process in which the electromagnetic wave is received after multiple reflections between the upper and lower surfaces of the target. In the echo signal, the reflected wave signal at the upper boundary of the target shown in path L1 is the most obvious; affected by factors such as attenuation, the reflected signal at the lower boundary shown in L2 is relatively weak; the creeping wave shown in L3 and the signal after multiple reflections inside the target in L4 are usually submerged by noise and interference signals.

[0068] In order to further study the relationship between the scattered field and the medium parameter distribution, the following prerequisites are set in the GPR reflection measurement model structure shown in Figure 2(a):

[0069] (1) The underground target is a point scattering model;

[0070] (2) The underground transmission medium is uniform.

[0071] The GPR transmitting antenna Tx transmits a detection pulse underground. The reflected wave after encountering the target is received by the receiving antenna Rx. As the GPR antenna moves along the azimuth and passes through the target, the echo signal obtained is a function of time and position. The target in the corresponding B-SCAN image shows a hyperbolic geometric structure feature, as shown in Figure 2(b).

[0072] The electromagnetic waves emitted by GPR propagate in the medium in accordance with the MAXWELL equation. The differential formula is as follows:

[0073]

[0074] in, is the Hamiltonian operator, E is the electric field intensity, r is the displacement vector, t is the time, B is the magnetic induction intensity, H is the magnetic field intensity, D is the electric displacement vector, J is the current density vector, and ρ is the charge density. And, among these electrical parameters, their mutual relations satisfy the following relationship:

[0075]

[0076] In the above formula, D is the electric displacement vector, E is the electric field intensity, B is the magnetic induction intensity, H is the magnetic field intensity, J is the current density vector, ε is the dielectric constant, μ is the medium magnetic permeability, and σ is the electrical conductivity.

[0077] Recalculate the curl of the first curl formula in formula (4) and rearrange it as follows:

[0078]

[0079] The equation (6) is the wave equation. is the Hamiltonian operator, E is the electric field intensity, r is the displacement vector, t is the time, ε is the dielectric constant, μ is the magnetic permeability of the medium, is the second derivative of t, is the derivative of t, σ is the conductivity, and ρ is the charge density.

[0080] Without considering the induced charge density effect, equation (6) can be obtained:

[0081]

[0082] in, is the Hamiltonian operator, E is the electric field intensity, r is the displacement vector, t is the time, ε is the dielectric constant, μ is the magnetic permeability of the medium, is the second derivative of t, is the derivative of t, and σ is the conductivity. is the displacement current part, Belonging to the conduction current part, Fourier transform is performed on equation (7) to obtain:

[0083]

[0084] in, is the second derivative of r, ω is the angular frequency, r is the displacement vector, E is the electric field strength, ε is the dielectric constant, μ is the medium magnetic permeability, i is the imaginary unit, and σ is the conductivity. Solving equation (8) yields:

[0085]

[0086] Among them, E0 represents the amplitude of the electric field strength, ω is the angular frequency, r is the displacement vector, E is the electric field strength, e refers to the natural constant, i is the imaginary unit, k=α+iβ represents the wave number, and satisfies the following relationship:

[0087] k 2 =-εμω 2 +iσμω (10);

[0088] In formula (10), k is the wave number, ε is the dielectric constant, μ is the magnetic permeability of the medium, ω is the angular frequency, i is the imaginary unit, and σ is the conductivity.

[0089] α is the attenuation factor, β is the phase shift coefficient, satisfying:

[0090]

[0091] In formula (11), α is the attenuation factor, ω is the angular frequency, ε is the dielectric constant, μ is the medium magnetic permeability, σ is the conductivity, and β is the phase shift coefficient.

[0092] For a plane electromagnetic wave with an angular frequency of ω, the tangent value of the dielectric loss angle δ, tanδ, can be expressed as:

[0093]

[0094] In formula (12), tan is the tangent function, δ is the dielectric loss angle, σ is the conductivity, ω is the angular frequency, and ε is the dielectric constant.

[0095] The phase velocity of electromagnetic waves propagating in underground media is shown in formula (9):

[0096]

[0097] In formula (12), v is the phase velocity of electromagnetic wave propagation, ω is the angular frequency, and β is the phase shift coefficient.

[0098] According to formula (6), for good conductors, such as metals, the displacement current part can usually be ignored. It degenerates into a diffusion equation. When the conductivity of the medium is small, the displacement current term will play a dominant role in the propagation of the GPR high-frequency electromagnetic pulse signal, which is expressed as a wave equation. When tanδ is very small, such as tanδ<<1, substitute it into formula (11); the approximate calculation formula of the phase shift coefficient β can be obtained:

[0099]

[0100] In formula (14), β is the phase shift coefficient, ω is the angular frequency, ε is the dielectric constant, and μ is the magnetic permeability of the medium.

[0101] According to the relationship ε=ε0ε r , μ=μ0, Substituting formula (14) into formula (13), the phase velocity calculation formula can be rewritten as:

[0102]

[0103] In formula (15), v is the phase velocity of electromagnetic wave propagation, ε r represents the relative dielectric constant of the background medium, and c represents the speed of light in a vacuum.

[0104] At the same time, the approximate calculation formula for the attenuation factor α can be expressed as:

[0105]

[0106] In formula (16), α is the attenuation factor, ε r Represents the relative dielectric constant of the background medium.

[0107] According to formula (9), when the GPR propagation depth r = 1 / α, the intensity of the scattered field will decay to The corresponding depth at this time is usually called the "skin depth". The simulation was carried out for the buried cavity and water-filled medium targets in the soil medium layer. Figure 3 It is the statistical result of the scattering intensity change trend of the cavity and water-filled medium targets at different burial depths. The relative dielectric constants are 1 and 81, and the conductivity is 0 and 0.03 (unit: S / m), respectively. The burial depth range of the target is set to 0.5-1.45 (unit m). The relative dielectric constant of the soil is 11, and the conductivity is 0.01 (unit: mS / m).

[0108] As the burial depth of the cavity and water-filled medium targets increases, the attenuation will also increase, and the amplitude of the scattered electric field in the echo signal will gradually decrease.

[0109] The above discussion process is mainly analyzed from the perspective of dielectric attenuation, and the propagation depth of electromagnetic waves in the underground half space is related to the frequency. At the same time, the transmission power is also the main factor affecting the GPR detection depth. In fact, when the GPR works in the frequency range of 10M-1000MHz, for low-loss media, its dielectric constant and conductivity are less dependent on frequency and can generally be regarded as constants. Under this premise, the present invention will further carry out the research on the extraction and application method of GPR echo signal feature information.

[0110] Step 2: Analyze the main factors that affect the intensity of target features in the echo signal, and summarize the common characteristics of the target in the GPR B-SCAN image, such as intensity characteristics and geometric structure characteristics, and other characteristics of the GPR echo signal under the reflection measurement mode.

[0111] In the GPR radar echo response data, the target amplitude characteristic information in the GPR B-SCAN diagram, including the maximum amplitude, is used to measure the corresponding characteristic hyperbola to describe the shape and characteristics of the target. In practical applications, it is generally based on the fact that the intensity of the target in the echo response is higher than the intensity of the measured medium. Figure 4(a) is a model of GPR detection of underground scenes using a reflection measurement method (assuming a layered structure). The transmitting antenna of the GPR radar transmits the ultra-wideband electromagnetic pulse signal to the surface, and its interface will produce various electromagnetic physical phenomena such as reflection, refraction, and scattering.

[0112] Due to the influence of factors such as medium non-uniformity, multi-layer structure, target overlap, propagation path changes and noise interference, the GPR B-SCAN image generally has more complex structural characteristics. A GPR B-SCAN simulation image is shown in Figure 4(b).

[0113] Below, the main electrical parameters that affect the intensity of object characteristic information in GPR echo response data are analyzed. According to Snell's law, the reflection coefficient γ and the refraction coefficient T can be expressed as:

[0114]

[0115] In formula (17), γ is the reflection coefficient, ε i (i=1,2) is the relative dielectric constant of medium i.

[0116]

[0117] In formula (18), T is the refractive index, γ is the reflection coefficient, and ε i (i=1,2) is the relative dielectric constant of medium i.

[0118] ε1 and ε2 are the relative dielectric constants of medium 1 and medium 2. On this basis, the GPR receiving antenna is used to receive the reflected wave, and the refracted wave is transmitted to the next layer of medium, and the attenuation coefficient of the electromagnetic wave in the lossy medium is set in the i-th layer of medium, and the attenuation coefficient of the electromagnetic wave is η i for:

[0119]

[0120] In formula (19), η i is the attenuation coefficient of electromagnetic waves, e is the natural logarithm, j is the imaginary unit, ω k is the kth frequency component of the electromagnetic wave pulse signal, d i represents the depth of the i-th layer of medium, c represents the speed of light in vacuum, ε i and σ i represent the relative permittivity and conductivity of the i-th layer medium respectively.

[0121] Let the scattering point (x p ,y p ) and the measuring point (x q ,y q ) is r pq , in (x q ,y q ) The target signal measured at It can be expressed as:

[0122]

[0123] In formula (20), is the target signal, x q ,y q is the horizontal and vertical coordinates of the measurement point, α is the coefficient, γ p For point P (xp ,y p ), ζ is the delay operator, and e is the emitted electromagnetic pulse signal.

[0124] Round trip delay τ pq for:

[0125]

[0126] In formula (21), To round down, ε r is the dielectric constant of the medium, c is the speed of light in a vacuum, r pq is (x p ,y p ) and (x q ,y q ) between the two.

[0127] Combining formula (19) and formula (20), it can be seen that the GPR echo signal is affected by many factors. When the reflection characteristic hyperbola of the target can be observed in the B-SCAN image, the dielectric constant is the most important factor compared with other parameters. When the difference in dielectric constant between underground media increases, the reflection phenomenon at the interface will become more intense. This is because the difference in dielectric constant will cause the electromagnetic wave to be refracted and reflected more at the interface of the medium, thereby generating a stronger reflection signal. On the B-SCAN image, these reflection signals will show more obvious hyperbolic characteristics. Therefore, as the difference in dielectric constant between the medium increases, the reflection at the interface will become more intense, making the hyperbolic characteristic information on the B-SCAN image more obvious. The simulation of the buried cavity and water-filled medium target in the soil medium layer was carried out.

[0128] Figure 5 It is the statistical result of the scattering intensity of the cavity and water-filled medium targets under different geological environments, where the relative dielectric constant of the underground soil background medium is set to range from 4 to 40, and the conductivity range is: 1.4×10 -4 ~5×10 -2 (mS / m), the relative permittivity of the void and water-filled medium targets are 1 and 81, and the conductivity is 0 and 0.03 (S / m), respectively.

[0129] To facilitate comparative analysis, Figure 5 In the results shown, the amplitude of the scattered electric field of the target is positive, and the amplitude corresponding to the hollow target is inverted. According to the definition of reflection coefficient (17), as the background medium ε r As ε increases, the difference between the hollow target and the background medium gradually increases, while the difference between the water-filled medium target is just the opposite. rWhen ε <9, the scattered electric field intensity of the water-filled medium target is greater, and r When the attenuation is constant, the dielectric constant is the key parameter affecting the detection of underground targets.

[0130] Step 3: A two-stage cascaded CNN model structure was designed to extract the feature map of the B-SCAN image, and the classifier network was used to complete the recognition of the target feature hyperbola.

[0131] Due to the direct coupling between GPR receiving antennas and the ground surface (or near the receiving antenna), strong echoes are often generated, making the direct wave (i.e., interference signal) in the GPR B-SCAN imaging system much stronger than the target feature signal. Therefore, the semantic feature network can be used to filter out the regional position in the GPR image, and on this basis, the target feature information can be detected to improve the accuracy of the detection result. A two-stage cascade CNN is designed, and its network structure is as follows: Figure 6 shown.

[0132] The cascade network structure of the first layer is shown in Figure 6 The upper part is the direct wave feature extraction network. The direct wave intensity of the GPR radar echo is very high, which is often represented by an obvious horizontal structure on the B-SCAN image, such as Figure 6 As shown in A. For the direct wave feature extraction network, convolution operation is first used to extract features, and feature dimension reduction is performed through pooling. The size of the convolution output feature map can be expressed as:

[0133]

[0134] In formula (22), is the width of the convolutional layer input feature map, is the height of the convolutional layer input feature map, is the width of the convolution output feature map, is the height of the output feature map, and the size of the convolution kernel matrix is ​​F conv ×F conv , P0 represents the number of zero paddings at the boundary of the feature map matrix, S conv Indicates the stride of the convolution operation.

[0135] The size of the output feature map of the Pooling layer can be expressed as:

[0136]

[0137] In formula (23), Respectively represent the width and height of the pooling layer input feature map and the width and height of the pooling output feature map, S pooling represents the pooling step size, and the size of the pooling kernel matrix is ​​F pooling ×Fpooling .

[0138] For a B-SCAN image with an input size of 2016×480, in order to keep the size of the convolutional layer input and output feature maps consistent, the convolution kernel size F is taken conv =1, P0=1, S conv =1; the pooling layer takes the pooling kernel size F pooling =2, S pooling =1. In the direct wave feature extraction network, the feature map content corresponding to its output layer should mainly contain horizontal linear structural feature information with large amplitude. That is, the global structural information corresponding to the direct wave in the B-SCAN image is extracted to determine its spatial position. To this end, according to the characteristics of direct wave intensity and lateral distribution, combined with the convolution and pooling operations of the convolutional neural network, the direct wave is modeled and its features are extracted:

[0139] (1) Because the amplitude of the direct wave is large, its boundary characteristics are very obvious in the B-SCAN image. On this basis, a method for image edge detection based on convolution operation is proposed. At the same time, the convolution kernel function in CNN extracts features by sliding, thereby maintaining the position information of the direct wave features of the B-SCAN image;

[0140] (2) Based on the translation invariance of pooling, the linear continuous structural characteristics of the direct wave are used to reduce the dimension of multiple pooling features to obtain the final feature map. For example, the feature map of the final pooling layer of CNN is fused to obtain a 53×15 feature map. It can well reflect the structural characteristics of the continuous linear shape of the direct wave, but the local features of the B-SCAN image are greatly weakened. The network structure of the direct semantic feature is constructed by fusing the output features of multiple convolution and pooling layers to construct a significant feature description, thereby realizing the efficient extraction of the overall structural features of the direct wave.

[0141] First, we use 5-layer convolution and pooling to extract high-level semantic features from the B-SCAN image (2016×480) to obtain 512 images with a size of 53×15. Then, we use methods such as Cubic interpolation sampling to achieve size matching with the B-SCAN image, such as Figure 6 As shown in part B of the figure, the position of the direct wave in the B-SCAN image is determined and removed. By comparing the B-SCAN image with the extracted direct wave feature image, it can be seen that Figure 6 The position of the direct wave in part A and part B is consistent, indicating that this algorithm can effectively extract the characteristic information of the direct wave. In addition, if the interference intensity of the direct wave in the echo far exceeds that of the target, it is difficult to observe the structural characteristics of the target from a subjective visual perspective in the corresponding B-SCAN. After eliminating the direct wave, the B-SCAN imaging results are shown in Figure 6 In part C, the structural characteristics of the object can be clearly seen.

[0142] The cascaded network structure of the second layer is shown in Figure 6 The lower part is the target feature extraction and classification network. Different from the network of the direct wave feature method, the features of the target object are extracted first. In order to improve the high-level semantics and richer low-level visual features in the extracted object feature image, the convolutional network adds 3 convolutional layers and reduces 1 pooling layer, and uses the deconvolution network to sample and calculate the feature map; on this basis, the corresponding category identification is constructed using the obtained target feature image, and is used to train the classification recognition network to achieve category identification of the target object and the interference object, thereby achieving efficient extraction of the target object characteristics and interference characteristics. The established model was trained on 3800 B-SCAN images. The main parameters of the model are set as follows: the initial learning rate is 0.0001, the maximum epoch is 20, and the total number of iterations is 76,000. The mutual entropy loss function is used, and a loss is output every 100 iterations. The results are optimized using stochastic gradient descent (SGDM). The loss curve of the extraction network training model is as follows Figure 7 As shown, after 40,000 iterations, the model becomes stable.

[0143] The cascade structure CNN is designed based on the GPR physical model as a priori condition. The main process of processing the direct wave and target feature signals respectively according to the amplitude and structural characteristics of the B-SCAN image can be summarized as follows:

[0144] (1) The direct wave signal has a horizontal linear structure with a large amplitude. The focus of processing is to locate its area in the B-SCAN image and remove it.

[0145] (2) In the B-SCAN image after removing the direct wave, the target presents a hyperbolic structural feature with a downward opening and a large amplitude. The semantic feature detection network is mainly used to extract the hyperbola with complete structural features.

Claims

1. A target feature extraction method combining GPR physical mechanism and CNN, characterized in that The following steps are involved: Step 1: Establish the scattering field equation of the semi-infinite space under the GPR to obtain the parameter information in the GPR echo signal used to characterize the scattering characteristics of the target; Step 2: Analyze the factors that affect the intensity of target features in the GPR echo signal, and analyze the common features of the target in the GPR B-SCAN image, including intensity features and geometric structure features; Step 3: Design a convolutional neural network model to extract the features of the B-SCAN image and use the classifier network to complete the recognition of the target feature hyperbola.

2. The target feature extraction method combining GPR physical mechanism and CNN according to claim 1, characterized in that: In step 1, the scattering field equation of the semi-infinite space under the GPR includes equation (9) and equation (10): k 2 =-me 2 +ismō (10); In formula (9), E0 represents the amplitude of the electric field intensity, ω is the angular frequency, r is the displacement vector, E is the electric field intensity, e refers to the natural constant, i is the imaginary unit, and k=α+iβ represents the wave number; In formula (10), k is the wave number, ε is the dielectric constant, μ is the magnetic permeability of the medium, ω is the angular frequency, i is the imaginary unit, and σ is the conductivity.

3. The target feature extraction method combining GPR physical mechanism and CNN according to claim 1, characterized in that: The parameter information used to characterize the target scattering characteristics in the GPR echo signal includes: attenuation factor α, phase shift coefficient β, tangent value tanδ of the dielectric loss angle δ, and electromagnetic wave propagation phase velocity v.

4. The target feature extraction method combining GPR physical mechanism and CNN according to claim 1, characterized in that: In step 2, the reflected wave is received by the GPR receiving antenna, and the refracted wave is transmitted to the next layer of the medium. The attenuation coefficient of the electromagnetic wave in the lossy medium is set to the attenuation coefficient of the electromagnetic wave in the i-th layer of the medium. i for: In formula (19), η i is the attenuation coefficient of electromagnetic waves; e is the natural logarithm; j is the imaginary unit; ω k is the kth frequency component of the electromagnetic wave pulse signal; d i represents the depth of the i-th layer of medium; c represents the speed of light in vacuum; ε i and σ i represent the relative permittivity and conductivity of the i-th layer medium respectively; Let the scattering point (x p ,y p ) and the measuring point (x q ,y q ) is r pq , at the measuring point (x q ,y q ) The target signal measured at It is expressed as: In formula (20), is the target signal; x q ,y q are the horizontal and vertical coordinates of the measuring point respectively; α is the coefficient; γ p For point P (x p ,y p ) at a point where the reflection coefficient is; ζ is the delay operator; e is the emitted electromagnetic pulse signal; Round trip delay τ pq for: In formula (21), is rounded down; ε r is the dielectric constant of the medium; c is the speed of light in a vacuum; r pq is (x p ,y p ) and (x q ,y q ) between the two sides; Combining formula (19) and formula (20), it can be seen that the GPR echo signal is affected by many factors. When the reflection characteristic hyperbola of the target can be observed in the B-SCAN image, the dielectric constant is the most important factor compared with other parameters. When the difference in dielectric constant between underground media increases, the reflection phenomenon at the interface will become more intense. As the difference in dielectric constant between media increases, the reflection at the interface will also become more intense, making the hyperbolic characteristic information on the B-SCAN image more obvious.

5. The target feature extraction method combining GPR physical mechanism and CNN according to claim 1, characterized in that: In step 3, a convolutional neural network model is used to extract features of the B-SCAN image, which specifically includes the following steps: S3.1: By extracting the feature edge information of the object in the B-SCAN image, the feature information corresponding to the target object in the B-SCAN image is obtained; In the B-SCAN image, the second-order derivative corresponding to the area with small or uniform intensity change is very small and close to 0; while the second-order inverse of the area with large intensity change corresponds to a larger value. This process can be described by the following formula: In formula (1), is the second-order derivative of the image position function; f(x,y) is the coordinate position corresponding to the target; x represents the horizontal coordinate corresponding to the target; y represents the vertical coordinate corresponding to the target; f(x+1,y+1) means that the image coordinate position is (x+1,y+1); f(x+1,y-1) means that the image coordinate position is (x+1,y-1); f(x-1,y+1) means that the image coordinate position is (x-1,y+1); f(x-1,y-1) means that the image coordinate position is (x-1,y-1); f(x+1,y) means that the image coordinate position is; f(x-1,y) means that the image coordinate position is (x-1,y); f(x,y+1) means that the image coordinate position is (x,y+1); f(x,y-1) means that the image coordinate position is (x,y-1); S3.2: When the difference between the target and the background in the B-SCAN image is greater, the gradient change will be greater; in the convolutional neural network model, the corresponding feature information is obtained by convolving the input image; let the value of the neuron in the jth feature map of the i-th layer be for: In formula (2), g(·) is the activation function; is the weight of connecting the jth feature of the i-th layer to the mth feature map of the i-1th layer, b (i) is the bias term of the i-th layer; Represents the value of the neuron in the mth feature map of the i-1th layer; M represents the number of rows of the feature map of the input target image; N represents the number of columns of the feature map of the input target image; The convolutional neural network model CNN uses convolution kernels of different scales to perform convolution calculations on the input f(x,y) to obtain the corresponding feature map; S3.3: The process of extracting the feature information of the input target from an input B-SCAN image through L feature mappings is expressed as: In formula (3), is the feature image of the target object; f Σ [·] is the feature fusion implementation module; is the sth feature map of the last layer (L layer) of convolution operation; s is the number of convolution operation layers; To perform L feature mappings; is the l-th feature image; w is the width of the input image Ι; h is the height of the input image Ι; M is the number of channels of the input image Ι; C is the convolution kernel in the convolution layer; k is the size of C; N is the number of C; D is the number of channels of C, the number of channels of the convolution kernel is equal to the number of channels of the input image. Since the GPR B-SCAN image is a single channel, D=M=1 is taken here; For Feature image obtained after pooling; To perform pooling operation on the l-time feature image; S3.4: After performing continuous convolution and pooling operations on B-SCAN to obtain the corresponding feature maps, The objects and noise therein are further identified; since the target in the B-SCAN image under the reflection measurement mode has a hyperbolic structural feature, the target feature hyperbola is identified through the classifier network.

6. The target feature extraction method combining GPR physical mechanism and CNN according to claim 1, characterized in that: In step 3, a two-stage cascaded CNN is designed, as follows: The cascade network structure of the first layer is a direct wave feature extraction network. For the direct wave feature extraction network, convolution operation is first used to extract features, and then feature dimension reduction is performed through pooling. The size of the convolution output feature map is calculated as: In formula (22), The width of the input feature map for the convolutional layer; The height of the input feature map of the convolution layer; is the width of the convolution output feature map; is the height of the output feature map, and the size of the convolution kernel matrix is ​​F conv ×F conv ; P0 represents the number of zero paddings at the boundary of the feature map matrix; S conv Indicates the step size of the convolution operation; The size of the output feature map of the Pooling layer is calculated as: In formula (23), Respectively represent the width and height of the input feature map of the pooling layer; Respectively represent the width and height of the pooled output feature map; S pooling represents the pooling step size, and the size of the pooling kernel matrix is ​​F pooling ×F pooling ; In the direct wave feature extraction network, the output feature map contains horizontal linear structural feature information with large amplitude, that is, the global structural information corresponding to the direct wave in the B-SCAN image is extracted to determine its spatial position; The cascade network structure of the second layer is the target feature extraction and classification network. First, the features of the target object are extracted. The cascade network adds 3 convolutional layers and reduces 1 pooling layer, and uses the deconvolution network to sample and calculate the output feature map. The corresponding category identification is constructed using the obtained target feature image and used to train the classification recognition network.

Images