Underground target classification method based on MobileNetV3 and FP-GPR

Abstract

The invention relates to the technical field of target classification, in particular to an underground target classification method based on MobileNetV3 and FP-GPR, and mainly solves the technical problem that an existing GPR target recognition method is low in accuracy and low in efficiency. According to the method, data is subjected to H-Alpha decomposition, Freeman decomposition and Pauli decomposition to obtain eight polarization characteristic parameters, the eight polarization characteristic parameters are fused to form a multi-dimensional characteristic matrix to serve as a data set, then the data set is input into an improved MobileNetV3 network added with an SE module to be trained to obtain a target network, and finally underground target recognition is conducted through the target network. On one hand, the parameter quantity required by the convolutional neural network can be greatly reduced, and light weight and high efficiency are realized in a data input stage; and on the other hand, the key feature information is better captured while the light weight is kept, so that the target recognition performance of the network is improved, and the accuracy is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of target classification, and particularly to an underground target classification method based on MobileNetV3 and FP-GPR. Background Art

[0002] With the acceleration of the urbanization process, the development and utilization of urban underground space have attracted more and more attention. To ensure the safety of urban underground space, it is crucial to accurately detect the types of underground targets. Ground Penetrating Radar (GPR), as a technical tool integrating high efficiency, non-destructiveness, and deep penetration ability, has been widely used in the field of underground target recognition.

[0003] There are mainly two existing GPR target recognition methods: The first is to manually extract features and combine them with a classifier for classification. This feature extraction method has subjectivity and instability, and at the same time, it will lose the polarization attributes of underground targets, resulting in a low accuracy rate of target classification. The second is to combine GPR images with deep learning to achieve underground target classification. However, this method has a high misjudgment rate when distinguishing underground pipelines and spherical targets with similar cross-sectional views, and the convolutional neural network based on image recognition has large parameters, which will affect the efficiency and accuracy of underground classification. Therefore, this application proposes an underground target classification method based on MobileNetV3 and FP-GPR. Summary of the Invention

[0004] To overcome the technical defects of the existing GPR target recognition methods, which have low accuracy and low efficiency, the present invention provides an underground target classification method based on MobileNetV3 and FP-GPR.

[0005] The underground target classification method based on MobileNetV3 and FP-GPR provided by the present invention includes the following steps:

[0006] S10. Build FP-GPR, and use the FP-GPR to obtain B-scan images of three polarization modes, namely HH, VH, and VV, of underground targets;

[0007] S20. Perform H-Alpha decomposition on the B-scan image to obtain entropy H and average scattering angle α, perform Freeman decomposition on the B-scan image to obtain surface scattering power P s , volume scattering power P v and secondary scattering power P d , perform Pauli decomposition on the B-scan image to obtain odd-order scattering intensity |a| 2 , even-order scattering intensity |b| 2and the cross-polarization scattering intensity |c| 2 , so as to obtain eight polarization characteristic parameters corresponding to each signal point on the B-scan image;

[0008] S30. Fuse the eight polarization characteristic parameters to obtain a multi-dimensional feature matrix, use the multi-dimensional feature matrix as a data set, and randomly divide it into a training set, a validation set, and a test set;

[0009] S40. Add an SE module to the MobileNetV3 network to form an improved MobileNetV3 network, and input the data set into the improved MobileNetV3 network for training to obtain a target network;

[0010] S50. Use the multi-dimensional feature matrix obtained from the B-scan image as an input, and obtain the underground target classification result through the target network.

[0011] Optionally, in step S10, the FP-GPR uses a vector network analyzer as a radar signal source, uses a pair of horn antennas as transmitting / receiving antennas, and obtains the full polarization information of the underground target by changing the positions of the transmitting / receiving antennas.

[0012] Optionally, in step S20, the formula for the H-Alpha decomposition is:

[0013] ;

[0014] where, , represents the eigenvalue of the polarization coherence matrix;

[0015] ;

[0016] where, ), represents the first element of the th eigenvector .

[0017] Optionally, in step S20, the formula for the Freeman decomposition is:

[0018] ;

[0019] ;

[0020] where, , , respectively represent the weight coefficients of surface scattering, volume scattering, and secondary scattering, represents the complex scattering coefficient of transmitting horizontal polarization and receiving horizontal polarization, represents the complex scattering coefficient for transmitting horizontally polarized waves and receiving vertically polarized waves, represents the complex scattering coefficient for transmitting vertically polarized waves and receiving vertically polarized waves, and is a constant. When the real part of is non - negative, otherwise

[0021] Optionally, in step S20, the formula for Pauli decomposition is:

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] wherein, , and respectively represent the complex coefficients of odd - order scattering, even - order scattering, and cross - polarization scattering.

[0027] Optionally, in step S40, the improved MobileNetV3 network uses H - swish as the activation function.

[0028] Optionally, after the dataset is input into the improved MobileNetV3 network in step S40, the following operations are performed:

[0029] S41. Perform a dimension - increasing operation through a convolutional layer;

[0030] S42. Perform a normalization operation through a normalization layer;

[0031] S43. Perform an activation function operation through H - swish;

[0032] S44. Perform a splitting operation on the spatial dimension and channel dimension through a depth - separable convolutional layer;

[0033] S45. Perform symmetric normalization and activation function operations;

[0034] S46. Perform global average pooling and fully - connected operations through the SE module;

[0035] S47. Perform a dimension - decreasing operation through a convolutional layer;

[0036] S48. Output through a normalization layer.

[0037] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0038] The underground target classification method based on MobileNetV3 and FP-GPR provided by the present invention decomposes the data collected by FP-GPR through H-Alpha decomposition, Freeman decomposition, and Pauli decomposition to obtain eight polarization characteristic parameters, and fuses them to form a multi-dimensional feature matrix as a data set. Then, the data set is input into an improved MobileNetV3 network with an added SE module for training to obtain a target network. Finally, underground target recognition is performed through this target network:

[0039] First, this method uses the multi-dimensional feature matrix as the input of the subsequent improved MobileNetV3 network. While retaining all polarization information, the input network data is streamlined through the polarization decomposition method, which can significantly reduce the number of parameters required by the convolutional neural network, achieving lightweight and high efficiency in the data input stage;

[0040] Second, this method adds an SE module to the existing MobileNetV3 network to form an improved MobileNetV3 network, which can enable the MobileNetV3 network to better capture key feature information while maintaining lightweight, thereby improving the target recognition performance of the network and ensuring accuracy. Brief Description of the Drawings

[0041] The accompanying drawings herein are incorporated into and constitute a part of this specification, showing embodiments consistent with the present invention and used together with the specification to explain the principles of the present invention.

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 It represents the flowchart of the underground target classification method in the embodiments of the present invention;

[0044] Figure 2 It represents the B-scan image collected in step S10 in the embodiments of the present invention. Detailed Embodiments

[0045] In order to more clearly understand the above objects, features, and advantages of the present invention, the following will further describe the solution of the present invention. It should be noted that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0046] In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0047] The following combines Figures 1 to 2 to detail the specific embodiments of the present invention.

[0048] This embodiment provides an underground target classification method based on MobileNetV3 and FP-GPR, including steps S10 to S50.

[0049] S10. Set up FP-GPR, and use FP-GPR to obtain B-scan images of an underground target in three polarization modes of HH, VH, and VV.

[0050] It is easy to understand that FP-GPR is a full-polarization ground penetrating radar, HH means the radar emits horizontally polarized electromagnetic waves and receives horizontally polarized echoes, VH means the radar emits vertically polarized electromagnetic waves and receives horizontally polarized echoes, VV means the radar emits vertically polarized electromagnetic waves and receives vertically polarized echoes, and the B-scan image is a B-mode ultrasonic scan image.

[0051] Specifically, in step S10, FP-GPR uses a vector network analyzer as the radar signal source, uses a pair of horn antennas as the receiving / transmitting antennas, and obtains the full-polarization information of the underground target by changing the positions of the receiving / transmitting antennas.

[0052] More specifically, the working frequency band in the vector network analyzer is 1.8 GHz - 5 GHz, and the step frequency is 8 MHz. The distance between the receiving / transmitting antennas is 2 cm, and a measuring point is set every 5 cm along the survey line. The target to be measured is buried in a sandbox with dimensions of 2.0 m × 1.2 m × 0.8 m (length × width × height), and the antennas are set 5 cm above the dry sand surface.

[0053] It should be noted that the four typical underground targets for classification are a metal ball, a metal pipe, a metal dihedral angle, and a metal multi-branch. Among them, the diameter of the metal ball is 15 cm, the length of the metal pipe is 55 cm and the diameter is 13 cm. The metal dihedral angle is composed of metal plates with an included angle of 90 degrees, a length of 35 cm and a width of 30 cm. The main part of the metal multi-branch is 40 cm long. In this embodiment, FP-GPR collects a total of 1200 original data images, among which, along different survey lines, there are 100 images of each target in three polarization states.

[0054] It is easy to understand that the polarization data of the underground target under three polarization conditions can be obtained from the B-scan images collected by FP-GPR, such as Figure 2As shown, the HH, VV, and VH echo numerical intensities corresponding to each measurement point are written as the corresponding 2×2 polarization scattering matrix , and the full polarization scattering matrix is constructed according to the measurement point distribution.

[0055] S20. Perform H-Alpha decomposition on the B-scan image to obtain entropy H and average scattering angle α, and perform Freeman decomposition on the B-scan image to obtain surface scattering power P s , volume scattering power P v and secondary scattering power P d , perform Pauli decomposition on the B-scan image to obtain odd-order scattering intensity |a| 2 , even-order scattering intensity |b| 2 and cross-polarization scattering intensity |c| 2 , so as to obtain eight polarization characteristic parameters corresponding to each signal point on the B-scan image.

[0056] It is easy to understand that the H-Alpha decomposition (H-α Decomposition) is a classic polarization decomposition method based on polarimetric SAR (Synthetic Aperture Radar) data, mainly used to analyze the target scattering mechanism, and realizes target classification and physical interpretation through two parameters: entropy (Entropy, H) and average scattering angle (Alpha Angle, α); Freeman decomposition is a classic incoherent target decomposition method in polarimetric SAR (Synthetic Aperture Radar) data analysis, proposed by Freeman and Durden in 1998, aiming to extract the scattering mechanism characteristics of surface targets by decomposing the polarization covariance matrix (such as the T3 or C3 matrix); Pauli decomposition is a classic method in polarimetric synthetic aperture radar (PolSAR) data processing. By decomposing the polarization scattering matrix into a linear combination of Pauli basis vectors, it extracts target scattering characteristics and is widely used in ground object classification and visualization.

[0057] Specifically, the formula for H-Alpha decomposition is:

[0058] ;

[0059] where, , represents the eigenvalue of the polarization coherence matrix;

[0060] ;

[0061] where, ), represents the first element of the th eigenvector .

[0062] It is easy to understand that the entropy H is used to describe the disorder of different scattering types, and the average scattering angle α is used to describe the physical characteristics of the target.

[0063] Specifically, the formula for Freeman decomposition is:

[0064] ;

[0065] ;

[0066] where 、 、 represent the weight coefficients of surface scattering, volume scattering, and second-order scattering respectively, represents the complex scattering coefficient for transmitting horizontally polarized waves and receiving horizontally polarized waves, represents the complex scattering coefficient for transmitting horizontally polarized waves and receiving vertically polarized waves, represents the complex scattering coefficient for transmitting vertically polarized waves and receiving vertically polarized waves, and are constants. When the real part of is non-negative , otherwise .

[0067] Specifically, the formula for Pauli decomposition is:

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] where 、 and represent the complex coefficients of odd-order scattering, even-order scattering, and cross-polarization scattering respectively.

[0073] It is easy to understand that Pauli decomposition decomposes the polarization scattering matrix into three scattering regions and represents the polarization scattering matrix as a weighted sum of three matrices. Through Pauli decomposition, three polarization characteristic parameters |a| 2 、|b| 2 、|c| 2 can be obtained, and |a| 2 、|b| 2 、|c| 2 correspond to the plane single-scattering power, the dihedral angle scattering power with a directional angle of 0 degrees, and the dihedral angle scattering power with a directional angle of 45 degrees respectively.

[0074] S30. Fuse the eight polarization feature parameters to obtain a multi-dimensional feature matrix, use the multi-dimensional feature matrix as a data set, and randomly divide it into a training set, a validation set, and a test set.

[0075] S40. Add an SE module to the MobileNetV3 network to form an improved MobileNetV3 network, and input the data set into the improved MobileNetV3 network for training to obtain the target network.

[0076] It is easy to understand that MobileNetV3 is a lightweight convolutional neural network architecture developed by Google, mainly used for image recognition and other computer vision tasks in resource-constrained environments such as mobile devices and embedded devices.

[0077] It is easy to understand that the improved MobileNetV3 network consists of several Bneck structures. Compared with the existing MobileNetV3 network, it adds an attention mechanism, that is, an SE module, which can enable the MobileNetV3 network to better capture key feature information while maintaining light weight, thereby improving the target recognition performance of the network.

[0078] Specifically, the Bneck structure consists of a backbone part and a residual part.

[0079] Specifically, the improved MobileNetV3 network uses H-swish as the activation function. H-swish has the characteristics of no upper bound, lower bound, smoothness, and non-monotonicity, and is superior to the common ReLU function in the model.

[0080] Specifically, after the data set is input into the improved MobileNetV3 network, the following operations are performed:

[0081] S41. The backbone part performs a dimension-increasing operation on the input features through a convolutional layer and expands the channels of the input feature layer;

[0082] S42. Perform a normalization operation through a normalization layer to prevent the problem of gradient explosion caused by the too large difference in the order of magnitude of each polarization parameter;

[0083] S43. Perform an activation function operation through H-swish;

[0084] S44. Perform a splitting operation on the spatial dimension and channel dimension through a depthwise separable convolutional layer, so as to reduce the parameters and computational amount of the network;

[0085] S45. Perform symmetric normalization and activation function operations;

[0086] S46. Through the SE module, global average pooling and fully connected operations are performed to obtain the channel weights. The length of the channel weights is the same as the number of channels of the feature layer passing through the depthwise separable convolutional layer. Multiply the channel weights by the features passing only through the depthwise separable convolutional layer to obtain the corresponding feature layer;

[0087] S47. Perform a dimensionality reduction operation through a convolutional layer;

[0088] S48. Output through a normalization layer.

[0089] S50. Using the multi-dimensional feature matrix obtained from the B-scan image as the input, obtain the underground target classification result through the target network.

[0090] In this embodiment, the network is trained and tested on a high-performance server equipped with an NVIDIA GEFORECE RTX 3080Ti GPU. The batch size of each input to the network is set to 4, the initial learning rate is set to 0.001, and the maximum number of training epochs is set to 50.

[0091] Under the condition of using the same dataset, the method proposed by the present invention achieves the highest accuracy rate of 98.75% on the test set. In the method proposed by the present invention, due to the addition of the SE module to improve the target recognition ability of the model, both the number of parameters and the model size have increased; however, compared with the commonly used networks in the existing ground penetrating radar target automatic recognition methods, such as the VGG16 network and the ResNet18 network, the number of parameters and the model size of the improved MobileNetV3 network proposed by this method have decreased by about half or more, realizing the accurate recognition of targets under a lightweight network.

[0092]

[0093] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Although the above embodiments have been described in detail, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the above embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the above embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. An underground target classification method based on MobileNetV3 and FP-GPR, characterized in that, It includes the following steps: S10. Build an FP-GPR, and use the FP-GPR to obtain B-scan images of an underground target in three polarization modes of HH, VH, and VV; S20. Perform H-Alpha decomposition on the B-scan image to obtain entropy H and average scattering angle α, perform Freeman decomposition on the B-scan image to obtain surface scattering power P s , volume scattering power P v and secondary scattering power P d , perform Pauli decomposition on the B-scan image to obtain odd-order scattering intensity |a| 2 , even-order scattering intensity |b| 2 and cross-polarization scattering intensity |c| 2 , thereby obtaining eight polarization characteristic parameters corresponding to each signal point on the B-scan image; S30. Fuse the eight polarization feature parameters to obtain a multi-dimensional feature matrix, use the multi-dimensional feature matrix as a data set, and randomly divide it into a training set, a validation set, and a test set; S40. Add an SE module to the MobileNetV3 network to form an improved MobileNetV3 network, and input the data set into the improved MobileNetV3 network for training to obtain a target network; S50. Use the multi-dimensional feature matrix obtained from the B-scan image as input, and obtain the underground target classification result through the target network.

2. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 1, wherein In step S10, the FP-GPR uses a vector network analyzer as a radar signal source, uses a pair of horn antennas as receiving / transmitting antennas, and obtains the full polarization information of the underground target by changing the positions of the receiving / transmitting antennas.

3. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 1, characterized in that, In step S20, the formula for H-Alpha decomposition is: ； Among them, , represents the eigenvalue of the polarization coherence matrix; ； Among them, ) represents the first element of the ith eigenvector.

4. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 3, characterized in that, In step S20, the formula for Freeman decomposition is: ； ； Among them, , , represent the weight coefficients of surface scattering, volume scattering, and secondary scattering respectively, represents the complex scattering coefficient for transmitting horizontally polarized waves and receiving horizontally polarized waves, represents the complex scattering coefficient for transmitting horizontally polarized waves and receiving vertically polarized waves, represents the complex scattering coefficient for transmitting vertically polarized waves and receiving vertically polarized waves, and are constants. When the real part of is non - negative, .

5. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 4, characterized in that In step S20, the formula for Pauli decomposition is: ； ； ； ； Among them, , and respectively represent the complex coefficients of odd-order scattering, even-order scattering, and cross-polarization scattering.

6. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 1, wherein In step S40, the improved MobileNetV3 network uses H-swish as an activation function.

7. The underground target classification method based on MobileNetV3 and FP-GPR according to claim 6, wherein In step S40, after the data set is input into the improved MobileNetV3 network, the following operations are performed: S41. Perform a dimensionality increase operation through a convolutional layer; S42. Perform a normalization operation through a normalization layer; S43. Perform an activation function operation through H-swish; S44. Perform a splitting operation on the spatial dimension and the channel dimension through a depthwise separable convolutional layer; S45. Perform symmetric normalization and activation function operations; S46. Perform global average pooling and fully connected operations through the SE module; S47. Perform a dimensionality reduction operation through a convolutional layer; S48. Output through a normalization layer.

Images