A neural network-based ISAR three-dimensional point cloud target recognition method

Abstract

The present application relates to a kind of neural network-based ISAR three-dimensional point cloud target identification method, belong to inverse synthetic aperture radar target identification technical field, solve the problem of poor algorithm generalization ability in the prior art in target identification, complex model.Specifically including: constructing three-dimensional point cloud dataset, the three-dimensional point cloud dataset includes ISAR three-dimensional point cloud and corresponding target model;Three-dimensional point cloud recognition neural network is constructed, and training is carried out, and target model identification neural network is obtained;Unknown model target is identified by trained identification neural network, and target model is obtained;It also includes: the radar echo data of different model targets received by distributed array antenna system is generated and saved n channel ISAR image;Image registration, amplitude and phase correction are carried out to the image, and after pitch direction super-resolution imaging, the ISAR three-dimensional point cloud corresponding to target model is obtained.The high-precision identification for air target (aircraft) type is realized.

Description

Technical Field

[0001] This invention relates to the field of inverse synthetic aperture radar (ISAR) target recognition technology, and in particular to an ISAR three-dimensional point cloud target recognition method based on neural networks. Background Technology

[0002] Compared to optical imaging technology, radar imaging offers the advantage of all-weather, all-day operation. With the continuous development and application of broadband radar technology, radar imaging has achieved significant applications and advancements in both defense and civilian fields. Radar imaging includes Synthetic Aperture Radar (SAR) and Inverse Synthetic Aperture Radar (ISAR). SAR refers to a situation where the radar moves while the target is stationary relative to the ground; ISAR refers to a situation where the radar is stationary while the target is moving. Both operate on the same basic imaging principle. Because ISAR can image non-cooperative targets with unknown motion states, target identification based on two-dimensional images obtained from ISAR imaging is an extremely effective method, such as identifying aircraft models during flight.

[0003] Traditional radar target recognition mainly includes two types of methods: model-based recognition and template-based recognition. The main steps of model-based recognition methods include: pre-training a classifier, target signal preprocessing, extracting target feature signals, feature mapping transformation, classifying the target using the classifier, and outputting the classification result. Model-based recognition methods require constructing a physical model of the target, such as its shape, structure, and surface texture features. The universality and robustness of this algorithm largely depend on the quality of the constructed physical model. The main steps of template-based recognition methods include: manually extracting features and using these features for classification. Commonly used features include: features based on invariant moments, features based on target affine projection moments, motion features, target position, radar cross section (RCS) features, signal time-domain waveform entropy, frequency-domain waveform entropy, average time-spectrum waveform entropy, and features extracted using principal component analysis and manifold learning methods. After extracting the target features, the extracted features are fed into a classifier for classification. Classic classification algorithms include support vector machines, nearest neighbor classification, and Bayesian classifiers.

[0004] The above analysis shows that traditional radar target recognition methods suffer from numerous problems, including poor algorithm generalization ability, complex models, and the need for large datasets to construct databases or model libraries. Deep learning-based target recognition methods have attracted increasing attention from scholars in recent years. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a neural network-based ISAR three-dimensional point cloud target recognition method to solve the problems of poor generalization ability and complex models in existing target recognition algorithms.

[0006] On one hand, embodiments of the present invention provide a neural network-based ISAR three-dimensional point cloud target recognition method, specifically including:

[0007] Construct a 3D point cloud dataset, which includes ISAR 3D point clouds and corresponding target models;

[0008] A three-dimensional point cloud recognition neural network is constructed and trained based on the three-dimensional point cloud dataset to obtain a target model recognition neural network;

[0009] Obtain the ISAR three-dimensional point cloud of the unknown target model, and use the target model recognition neural network to identify the target model;

[0010] The construction of the three-dimensional point cloud dataset includes: acquiring radar echo data of different target types received by the distributed array antenna system, generating and saving n-channel ISAR images; performing image registration and amplitude-phase correction on the n-channel ISAR images; and obtaining the ISAR three-dimensional point cloud corresponding to the target type through elevation super-resolution imaging based on the amplitude-phase corrected ISAR images.

[0011] Optionally, the image registration and amplitude-phase correction of the n-channel ISAR image includes:

[0012] In the n-channel ISAR image, find n matching pixel units and extract the amplitude and phase values ​​of each pixel unit to obtain the amplitude and phase correction values:

[0013]

[0014] Among them, C ref To match the amplitude and phase values ​​of pixel units in the reference channel ISAR image, C i For the amplitude and phase values ​​of the matching pixel unit in the i-th channel ISAR image, C am,ph This is the calculated amplitude and phase correction value, C of the reference channel. am,ph =1;

[0015] Based on the calculated C am,ph Multiplying the image with the ISAR images from n channels yields the amplitude-phase corrected ISAR image.

[0016] Optionally, the ISAR three-dimensional point cloud obtained from the amplitude-phase corrected ISAR image through elevation super-resolution imaging includes:

[0017] Based on the amplitude-phase corrected ISAR image, the radar image S of the antenna phase center formed by the nth radar transmitting antenna and the receiving antenna. n (x',r') is represented as:

[0018]

[0019] Where λ is the wavelength, x' and r' represent the coordinates of the distance Doppler (RD) cell, f(x,r) is the two-site scatter function, and R n (r,s) represents the distance to the target center in the nth antenna phase;

[0020] The radar image obtained by discretizing the phase center of the antenna formed by the nth radar transmitting and receiving antennas is as follows:

[0021] S=Aγ+β

[0022] Where S = [S1, S2, ..., S n ] T S represents n discrete ISAR two-dimensional images. n Let γ represent the nth ISAR 2D image, where γ = [γ1, γ2, ..., γ]. L ] T Represents the reconstructed ISAR 3D point cloud, γ L This represents the scattering intensity at the Lth point, where L is the number of points in the ISAR 3D point cloud. b ⊥ni Let s be the distance between the i-th scattering point of the target and the origin in the direction perpendicular to the radar line of sight. j denoted as , where is the distance of the j-th array radar antenna from the array center, and β is Gaussian white noise.

[0023] Optionally, the three-dimensional point cloud recognition neural network includes: an input layer, a max pooling layer, and a multilayer perceptron;

[0024] The input layer is used to input n point data, where each point data is represented by three-dimensional coordinates;

[0025] The max pooling layer is used to find the maximum feature point in each dimension of n data points.

[0026] The multilayer perceptron is used to obtain feature vectors of multiple dimensions and to obtain class vectors of k dimensions; the class vectors of k dimensions represent the probabilities that the three-dimensional point cloud input to the neural network belongs to k classes respectively.

[0027] Optionally, based on the three-dimensional coordinates [batch,n,3] of the n point data, the second and third dimensions of the three-dimensional coordinates are swapped to obtain the swapped dimension coordinates: [batch,3,n].

[0028] Then, the multilayer perceptron is used to transform the data to obtain a feature vector with dimensions [batch, 1024, n].

[0029] Based on the feature vectors of [batch, 1024, n], a multilayer perceptron is used to obtain the k-dimensional category vectors for each.

[0030] Optionally, the step of obtaining the ISAR three-dimensional point cloud of the unknown target model and identifying the target model using the target model recognition neural network includes:

[0031] The obtained n-channel ISAR 3D point cloud is input into the trained 3D point cloud recognition network for feature extraction, and the extracted feature vector is obtained.

[0032] Based on the extracted feature vectors, identification is performed using a multilayer perceptron to obtain the probabilities of k target types, which in turn yields the target model.

[0033] Optionally, the step of inputting the obtained n-channel ISAR 3D point cloud into a trained 3D point cloud recognition network for feature extraction, resulting in an extracted feature vector, includes:

[0034] ISAR 3D point cloud is represented as an n×3 tensor, where n represents the number of points and 3 corresponds to the xyz coordinates;

[0035] The input ISAR 3D point cloud is then multiplied with the transformation matrix in the point cloud recognition network, and features are extracted from the point cloud data based on multiple multilayer perceptrons and max pooling layers to obtain the final global features.

[0036] Optionally, the radar image of the antenna phase center obtained by discretization includes:

[0037] The iterative residual is decomposed into residuals, which are expressed as follows:

[0038]

[0039] Where, ε k The residual of the k-th atom, s is the measurement vector, A k The k-th row of the coefficient matrix The k-th element of the result vector obtained by the OMP algorithm;

[0040] Based on the components after residual iteration, the optimal range of error is obtained, thereby determining the minimum value of error.

[0041] Optionally, the step of acquiring radar echo data of different aircraft models received by the distributed array antenna system to generate and save n-channel ISAR images includes:

[0042] Acquire radar echo data from all scattering points received by the distributed array antenna system, and generate and save an ISAR two-dimensional image with n channels using the radar echo data from all scattering points received.

[0043] Based on the obtained n-channel ISAR two-dimensional images, a two-dimensional ISAR image is obtained using coordinate transformation and the Doppler algorithm.

[0044] Optionally, the step of acquiring radar echo data from all scattering points received by the distributed array antenna system, and generating and saving an ISAR two-dimensional image with n channels using the received radar echo data from all scattering points, includes:

[0045] At time t, the distance R(t) from the p-th scattering point on the target to the radar is expressed as:

[0046]

[0047] Among them, R m Let x' be the distance from the radar to the center of the target. p ,y′ p Let be the coordinates of the p-th scattering point on the target in the xO'y coordinate system;

[0048] Based on the coordinate transformation, the total echo signal from the L scattering points on the target is obtained. for:

[0049]

[0050] Where rect() represents the rectangular window function, The value represents the time within a radar pulse, where j is the imaginary unit, λ represents the wavelength, and t represents the time.

[0051] Two-dimensional ISAR images are obtained based on echo signals.

[0052] On the other hand, compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0053] 1. This invention expands the dimensions of ISAR imaging by combining MIMO (Multiple Input Multiple Output) real aperture and synthetic aperture technologies to obtain ISAR three-dimensional point cloud results. For the first time, neural network training is directly performed on ISAR three-dimensional point cloud data, and the recognition results have high accuracy and robustness.

[0054] 2. This invention performs synthetic aperture processing on the original data to generate a single-channel range-azimuth two-dimensional image, and performs MIMO real aperture super-resolution processing on the pitch dimension to generate a range-azimuth-pitch three-dimensional point cloud.

[0055] 3. For ISAR 3D point clouds, the present invention performs better than traditional methods for ISAR 2D image recognition, and the recognition performance is more robust after processing by the present invention.

[0056] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0057] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0058] Figure 1 This is a flowchart of a neural network-based ISAR three-dimensional point cloud target recognition method in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the ISAR three-dimensional point cloud imaging principle proposed in an embodiment of the present invention;

[0060] Figure 3 This is a detailed flowchart illustrating the target recognition method for ISAR 3D point clouds based on neural networks proposed in this embodiment of the invention.

[0061] Figure 4 It is an image obtained by ISAR two-dimensional imaging of the selected channel measured data proposed in the embodiments of the present invention;

[0062] Figure 5 This is a two-dimensional image of a certain measured data in 16 channels after registration, as proposed in an embodiment of the present invention.

[0063] Figure 6 This is a point cloud obtained through ISAR three-dimensional imaging, as proposed in this embodiment of the invention;

[0064] Figure 7 This is a schematic diagram of the principle for training a point cloud recognition network proposed in an embodiment of the present invention. Detailed Implementation

[0065] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0066] A specific embodiment of the present invention discloses a neural network-based ISAR three-dimensional point cloud target recognition method, such as... Figure 1 As shown, it includes:

[0067] Step S1: Construct a 3D point cloud dataset, which includes ISAR 3D point clouds and corresponding target models; for example, the target refers to an aircraft, and the target model refers to the aircraft model.

[0068] Step S2: Construct a three-dimensional point cloud recognition neural network and train it based on the three-dimensional point cloud dataset to obtain a target model recognition neural network;

[0069] Step S3: Obtain the ISAR 3D point cloud of the unknown target model, and use the target model recognition neural network to identify the target model.

[0070] Constructing a 3D point cloud dataset specifically includes:

[0071] Step S11: Obtain radar echo data of different aircraft models received by the distributed array antenna system, generate and save n (n greater than or equal to 2) channel ISAR images;

[0072] Specifically, the radar echo data of all scattering points received by the distributed array antenna system is acquired, and an ISAR two-dimensional image of n channels is generated and saved using the received radar echo data of all scattering points; such as Figure 2 As shown, in Figure 2 In the distributed array radar system, T represents the radar transmitting antenna, R represents the radar receiving antenna, the xOy coordinate system is the fixed coordinate system of the radar, and the radar is located at the origin. The xO'y coordinate system is the reference frame of the target, with the target center located at O'. The x-axis is the line connecting the radar to the target center, which changes as the target moves. According to the far-field assumption of electromagnetic wave radiation, at time t, the distance R(t) from the p-th scattering point on the target to the radar can be expressed as:

[0073]

[0074] Among them, R m Let x' be the distance from the radar to the center of the target. p ,y′ p ) represents the coordinates of the p-th scattering point on the target in the xO'y coordinate system.

[0075] By comparing the changes in the target coordinate system before and after the rotation, the following coordinate transformation relationship can be obtained:

[0076]

[0077] Among them, (x p ,y p Let (x′) be the coordinates of the p-th scattering point on the target in the xOy coordinate system. p ,y′ p Let θ be the coordinates of the p-th scattering point on the target in the xO'y coordinate system, and let θ be the rotation angle of the target around its own center.

[0078] Based on the coordinate transformation relationship, we obtain:

[0079]

[0080] Among them, (x p ,y p ) represents the coordinates of the p-th scattering point on the target in the xOy coordinate system.

[0081] Therefore, the total echo signal from the L scattering points on the target is for

[0082]

[0083] Where rect() represents the rectangular window function, The value represents the time within a radar pulse, where j is the imaginary unit, λ represents the wavelength, and t represents the time.

[0084] After obtaining the echo signal shown in formula (4), it can be found that there is a translational component R. m Quadratic terms Intersection These distortions can cause ISAR image distortion, so phase compensation of the received echo signal is necessary by constructing a matched signal. After phase compensation, the variables t and... Perform a Fourier transform (range-Doppler algorithm) to obtain a two-dimensional ISAR image.

[0085] Step S12: Based on the n-channel ISAR two-dimensional images, perform image registration and amplitude / phase correction on the n-channel ISAR images;

[0086] Specifically, the phase error between the n-channel ISAR images can lead to significant errors in the pitch-to-3D reconstruction results, thus affecting the accuracy of subsequent target recognition. Therefore, it is necessary to perform image registration and amplitude-phase correction on the acquired n-channel ISAR 2D images. To perform image registration and amplitude-phase correction, it is necessary to find the corresponding n matching pixel units in each of the n-channel ISAR images and extract the amplitude and phase values ​​of the pixel units.

[0087] For example, for the ISAR image of the i-th channel in an n-channel ISAR image, the amplitude and phase values ​​of the matched pixel unit are C. i Then, one of the n channels is designated as a reference, and the amplitude and phase values ​​of the pixel units in the reference channel are C. ref The amplitude and phase correction values ​​can be constructed as follows:

[0088]

[0089] Among them, C ref To match the amplitude and phase values ​​of pixel units in the reference channel ISAR image, C iFor the amplitude and phase values ​​of the matching pixel unit in the i-th channel ISAR image, C am,ph This is the calculated amplitude and phase correction value, C of the reference channel. am,ph Set C to 1. am,ph Amplitude and phase correction can be achieved by multiplying the images with n channels of ISAR.

[0090] S13: Based on the amplitude-phase corrected ISAR image, the ISAR three-dimensional point cloud is obtained by elevation super-resolution imaging;

[0091] Specifically, after image registration and amplitude-phase correction are completed, the radar image S at the antenna phase center formed by the nth radar transmitting antenna and receiving antenna is... n (x',r') can be represented as

[0092]

[0093] Where λ is the wavelength, x' and r' represent the coordinates of the distance Doppler (RD) cell, f(x,r) is the two-site scatter function, and R n (r,s) represents the distance to the target center in the nth antenna phase.

[0094] The scatter function is expressed as

[0095]

[0096] Here, sinc() represents the sinc function, and Δx and Δr represent the resolution. It is generally assumed that the scatter function is a two-dimensional Dirac function.

[0097] To make the subsequent signal processing steps more efficient, equation (6) needs to be simplified.

[0098] Based on the assumptions in equation (7), equation (6) can be simplified to:

[0099]

[0100] Where γ(x,r,s) represents the backscattering coefficient of the target point, s represents the distance between the array radar antenna and the center of the array, and a represents half the length of the radar array antenna.

[0101] Under the far-field approximation, the distance expression R n (r,s) can be approximated as

[0102]

[0103] Among them, b / / n b represents the distance of the target from the origin in the direction parallel to the radar line of sight. ⊥n This indicates the distance between the target and the origin in the direction perpendicular to the radar line of sight.

[0104] To process large-scale data more efficiently using matrix analysis, equation (8) needs to be discretized. Substituting equation (9) into equation (8) and discretizing it, we get:

[0105] S=Aγ+β (10)

[0106] Where S = [S1, S2, ..., S n ] T S represents n discrete ISAR two-dimensional images. n Let γ represent the nth ISAR 2D image, where γ = [γ1, γ2, ..., γ]. L ] T Represents the reconstructed ISAR 3D point cloud, γ L This represents the scattering intensity at the Lth point, where L is the number of points in the ISAR 3D point cloud. b ⊥ni Let s be the distance between the i-th scattering point of the target and the origin in the direction perpendicular to the radar line of sight. j denoted as , where is the distance of the j-th array radar antenna from the array center, and β is Gaussian white noise.

[0107] At this point, the problem of reconstructing the third dimension is transformed into solving equation (10).

[0108] Super-resolution algorithms are used for pitch-axis 3D reconstruction. Commonly used super-resolution reconstruction methods include greedy tracking algorithms, convex relaxation algorithms, and combined algorithms. Greedy tracking algorithms are widely used due to their simple structure and low computational cost. This invention primarily employs the orthogonal matching pursuit algorithm (OMP) for pitch-axis 3D reconstruction. The reason for this is that the OMP algorithm, through a greedy approach, selects a local optimum in each iteration, gradually approximating the original signal. This algorithm has high computational efficiency and is more suitable for pitch-axis 3D reconstruction of ISAR images with large datasets. The OMP algorithm mainly utilizes the correlation principle to select the atoms with the best matching iteration edge; secondly, it performs Schmitt orthogonalization on the selected atoms; and thirdly, it projects the signal onto the space composed of these orthogonal atoms to obtain the signal components and iteration residuals on the selected atoms. Errors are generated during the algorithm iteration process. To analyze these errors, residual decomposition is needed to obtain the signal components and iteration residuals on the selected atoms. The residuals are expressed as:

[0109]

[0110] Where, ε k The residual of the k-th atom, s is the measurement vector, A k For the k-th row (k-th atom) of the coefficient matrix, This is the k-th element of the result vector obtained by the OMP algorithm.

[0111] Based on the components after residual iteration, the optimal range of error is obtained, thereby determining the minimum value of error.

[0112] By repeatedly collecting echo signals from different types of aerial targets (specifically, different models of aircraft) and performing the aforementioned ISAR three-dimensional imaging on the echo signals, corresponding ISAR three-dimensional point clouds can be obtained. This allows the construction of an ISAR three-dimensional point cloud dataset for known aircraft models, which can then be used for the training of the subsequent ISAR three-dimensional point cloud recognition network.

[0113] Specifically, step S2 includes: constructing a 3D point cloud recognition neural network and training it based on the 3D point cloud dataset to obtain a target model recognition neural network. For example... Figure 7 As shown.

[0114] Specifically, as a concrete example, the 3D point cloud recognition neural network includes: an input layer, a max pooling layer, and a multilayer perceptron.

[0115] In the input layer, a batch of point cloud images are input, each containing n data points, and each data point is represented by three-dimensional coordinates. Therefore, the input dimension is represented by a coordinate system of [batch, n, 3]. Then, based on the input dimension, the second and third dimensions are swapped, resulting in a swapped dimension of [batch, 3, n]. This is then transformed using a multilayer perceptron to obtain a feature vector with a dimension of [batch, 1024, n].

[0116] The maximum feature point in each dimension of these n points is obtained through a max pooling layer. Based on the feature vector of [batch, 1024, n], each point cloud map corresponds to a 1024-dimensional feature vector. After passing through a multilayer perceptron, k-dimensional category vectors are obtained for each point cloud. These k-dimensional category vectors represent the probabilities that the input 3D point cloud belongs to k categories. By selecting the highest probability, the target category is obtained, thus completing the construction of the ISAR 3D point cloud recognition network model.

[0117] Based on the target model recognition neural network, it is necessary to determine the parameters in the target model recognition neural network, that is, to obtain the parameters inside the multilayer perceptron, which are obtained by training the ISAR 3D point cloud dataset.

[0118] The specific operation is as follows: First, assign random values ​​to each parameter to be determined in the target model recognition neural network; then, use the ISAR 3D point cloud in the dataset as the input of the target model recognition neural network, and use the aircraft category corresponding to the ISAR 3D point cloud as the expected output of the target model recognition neural network. At this time, the target model recognition neural network will also calculate an actual output based on the given input; then, compare the expected output result with the actual output result, and use the difference between the expected output result and the actual output result to adjust the parameters in the target model recognition neural network; repeat the above process until the difference between the actual output and the expected output of the network is less than a given threshold, which completes the training process and yields the target model recognition neural network training model.

[0119] Step S3 specifically includes: acquiring the ISAR three-dimensional point cloud of the unknown target model, and using the target model recognition neural network to identify the target model.

[0120] Specifically, after training the target model recognition neural network using the constructed ISAR 3D point cloud dataset, the trained neural network can be used to predict the category of ISAR 3D point clouds of unknown aircraft models. The specific process of target recognition is as follows: First, the ISAR 3D point cloud is input into the trained target model recognition neural network. The ISAR 3D point cloud can be represented as an n×3 tensor, where n represents the number of points and 3 corresponds to the xyz coordinates. Then, the input ISAR 3D point cloud is multiplied by the transformation matrix in the point cloud recognition network to ensure the invariance of the input 3D point cloud to specific spatial transformations. The point cloud data is then subjected to feature extraction through multiple multilayer perceptron and max pooling operations to obtain the final global features. Specifically, this process extracts a feature vector consisting of 1024 feature values ​​for type recognition. Finally, the global features are used by a multilayer perceptron to identify the final classification result, that is, the feature vector consisting of 1024 feature values ​​extracted in the previous step is converted into the probability that the ISAR 3D point cloud belongs to k aircraft models, thus completing the aircraft model recognition task.

[0121] The following examples provide a detailed description of the target recognition method for ISAR 3D point clouds based on neural networks proposed in this invention.

[0122] Example 1, as Figure 2 As shown in the figure, this invention proposes a target recognition method for ISAR 3D point clouds based on neural networks, which includes the following steps:

[0123] 1. The input is a set of all point cloud data in a frame, represented as an n×3 tensor, where n represents the number of point clouds and 3 corresponds to the xyz coordinates.

[0124] 2. The input data is first aligned by multiplying it with the transformation matrix learned by a feature transformation module, which ensures the model's invariance to specific spatial transformations.

[0125] 3. After extracting features from each point cloud data through multiple multilayer perceptrons, feature alignment is performed through feature transformation.

[0126] 4. Perform max pooling operations on each dimension of the feature to obtain the final global feature.

[0127] 5. The global features are used to identify the final aircraft model through a multilayer perceptron.

[0128] For example, for aircraft type recognition based on ISAR 3D point clouds, firstly, through operations such as multilayer perceptron, feature transformation and max pooling in the neural network, a feature vector consisting of 1024 feature values ​​is extracted for type recognition. Then, through multilayer perceptron operation, the feature vector consisting of these 1024 feature values ​​is converted into the probability that the ISAR 3D point cloud belongs to k aircraft types, and finally the aircraft type recognition task is completed.

[0129] Example 2 provides a detailed description of the target recognition method for ISAR 3D point clouds based on neural networks proposed in this invention.

[0130] like Figure 2 As shown in the figure, this embodiment proposes a target recognition method for ISAR 3D point clouds based on neural networks, which includes the following steps:

[0131] Step 1: Generate and save multi-channel ISAR two-dimensional images based on the raw data received by the distributed array antenna system;

[0132] Step 2: Perform image registration and amplitude / phase correction on the multi-channel ISAR two-dimensional images;

[0133] Step 3: Based on the amplitude-phase corrected ISAR image, obtain the ISAR three-dimensional point cloud through elevation super-resolution imaging;

[0134] Step 4: Pre-training of the target model recognition neural network.

[0135] Step S5: Based on the pre-trained model, identify the aircraft model using ISAR 3D point cloud.

[0136] A detailed flowchart of this method is shown below. Figure 3 As shown:

[0137] First, matched filtering is performed. During the radar coherence processing time, the echo signal is multiplied by the conjugate of the reference signal to obtain a single-frequency pulse signal as shown in formula (4). Motion compensation is then performed on the matched-filtered echo signal. Envelope alignment is the first step in motion compensation. Envelope alignment eliminates the misalignment of adjacent echoes in the range direction caused by the translation of the moving target relative to the radar. The core idea is to adjust the range direction of the echo signal during the accumulation time to the data relative to the turntable target by adjusting the time delay. Higher-order terms or random terms in the echo signal are compensated by phase autofocus. The ISAR two-dimensional image is obtained after azimuth-to-Fourier transformation as shown in formula (4). Figure 4 As shown.

[0138] Then, the maximum spectrum method was used to register the 16 channels to the ISAR two-dimensional image, resulting in formula (5). The maximum spectrum method is a commonly used method in image registration. A two-dimensional Fourier transform was performed on the ISAR two-dimensional image, and the maximum value of the fringe pattern in the frequency was selected for registration. The region of a single prominent point in the reference image was selected for amplitude and phase correction. The images of each channel after image registration are as follows: Figure 5 As shown.

[0139] Based on the principle of compressed sensing, the 3D reconstruction problem is transformed into a sparse signal recovery problem using formula (10). The OMP algorithm is used for super-resolution reconstruction, and residual iteration is performed using formula (11). During the iteration process, orthogonalization is performed to ensure the optimality of each iteration. Statistical filtering is applied to the 3D point cloud from the pitch-side super-resolution imaging to remove outliers. The filtered result is as follows: Figure 6 As shown.

[0140] A sliding window 3D imaging method is used to obtain a dataset for target recognition from the echo signal with a relatively long integration time. This dataset is then input into... Figure 7 The neural network shown is trained. This neural network directly identifies 3D point clouds, uses symmetric functions to handle the permutation invariance of point clouds, and uses a feature transformation module to handle the rotation invariance of point clouds. By extracting high-dimensional abstract features of the target, it achieves higher recognition accuracy and stronger robustness.

[0141] This invention expands the dimensions of ISAR imaging by combining MIMO (Multiple Input Multiple Output) real aperture and synthetic aperture technologies to obtain ISAR three-dimensional point cloud results. For the first time, neural network training is directly performed on ISAR three-dimensional point cloud data, and the recognition results have high accuracy and robustness, and can effectively identify aircraft models.

[0142] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0143] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for ISAR 3D point cloud target recognition based on neural networks, characterized in that, include: Construct a 3D point cloud dataset, which includes ISAR 3D point clouds and corresponding target models; A three-dimensional point cloud recognition neural network is constructed and trained based on the three-dimensional point cloud dataset to obtain a target model recognition neural network; Obtain the ISAR three-dimensional point cloud of the unknown target model, and use the target model recognition neural network to identify the target model; The construction of the three-dimensional point cloud dataset includes: acquiring radar echo data of different target types received by the distributed array antenna system, generating and saving n-channel ISAR images; performing image registration and amplitude-phase correction on the n-channel ISAR images; and obtaining the ISAR three-dimensional point cloud corresponding to the target type through elevation super-resolution imaging based on the amplitude-phase corrected ISAR images. The image registration and amplitude-phase correction of the n-channel ISAR image includes: In the n-channel ISAR image, find the corresponding n image-matching pixel units, and extract the amplitude and phase values ​​of the pixel units to obtain the amplitude and phase correction values: Among them, C ref To match the amplitude and phase values ​​of pixel units in the reference channel ISAR image, C i For the amplitude and phase values ​​of the matching pixel unit in the i-th channel ISAR image, C am,ph This is the calculated amplitude and phase correction value, C of the reference channel. am,ph =1; Based on the calculated C am,ph Multiplying the image with the n-channel ISAR image yields the amplitude-phase corrected ISAR image.

2. The ISAR 3D point cloud target recognition method based on neural networks according to claim 1, characterized in that, The ISAR image based on amplitude and phase correction is used to obtain an ISAR three-dimensional point cloud through elevation super-resolution imaging, including: Based on the amplitude-phase corrected ISAR image, the radar image S of the antenna phase center formed by the nth radar transmitting antenna and the receiving antenna. n (x',r') is represented as: Where λ is the wavelength, x' and r' represent the coordinates of the distance-Doppler (RD) cell, f(x,r) is the two-site scatter function, and R n (r,s) represents the distance to the target center in the nth antenna phase; The radar image obtained by discretizing the phase center of the antenna formed by the nth radar transmitting and receiving antennas is as follows: S=Aγ+β Where S = [S1, S2, ..., S n ] T S represents n discrete ISAR two-dimensional images. n Let γ represent the nth ISAR 2D image, where γ = [γ1, γ2, ..., γ]. L ] T Represents the reconstructed ISAR 3D point cloud, γ L This represents the scattering intensity at the Lth point, where L is the number of points in the ISAR 3D point cloud. b ⊥ni Let s be the distance between the i-th scattering point of the target and the origin in the direction perpendicular to the radar line of sight. j denoted as , where is the distance of the j-th array radar antenna from the array center, and β is Gaussian white noise.

3. The ISAR 3D point cloud target recognition method based on neural networks according to claim 1, characterized in that, The three-dimensional point cloud recognition neural network includes: an input layer, a max pooling layer, and a multilayer perceptron; The input layer is used to input n point data, where each point data is represented by three-dimensional coordinates; The max pooling layer is used to find the maximum feature point in each dimension of n data points. The multilayer perceptron is used to obtain feature vectors of multiple dimensions and to obtain class vectors of k dimensions; the class vectors of k dimensions represent the probabilities that the three-dimensional point cloud input to the neural network belongs to k classes respectively.

4. The ISAR 3D point cloud target recognition method based on neural networks according to claim 3, characterized in that, Based on the 3D coordinates [batch,n,3] of n points, the second and third dimensions of the 3D coordinates are swapped to obtain the swapped dimensional coordinates: [batch,3,n]. Then, the multilayer perceptron is used to transform the data to obtain a feature vector with dimensions [batch, 1024, n]. Based on the feature vectors of [batch, 1024, n], a multilayer perceptron is used to obtain the k-dimensional category vectors for each.

5. The ISAR 3D point cloud target recognition method based on neural networks according to claim 1, characterized in that, Obtaining an ISAR 3D point cloud of an unknown target model, and using the target model recognition neural network to identify the target model, includes: The obtained n-channel ISAR 3D point cloud is input into the trained 3D point cloud recognition network for feature extraction, and the extracted feature vector is obtained. Based on the extracted feature vectors, identification is performed using a multilayer perceptron to obtain the probabilities of k target types, which in turn yields the target model.

6. The ISAR 3D point cloud target recognition method based on neural networks according to claim 5, characterized in that, The step of inputting the obtained n-channel ISAR 3D point cloud into the trained 3D point cloud recognition network for feature extraction yields the following extracted feature vectors: ISAR 3D point cloud is represented as an n×3 tensor, where n represents the number of points and 3 corresponds to the xyz coordinates; The input ISAR 3D point cloud is then multiplied with the transformation matrix in the point cloud recognition network, and features are extracted from the point cloud data based on multiple multilayer perceptrons and max pooling layers to obtain the final global features.

7. The ISAR 3D point cloud target recognition method based on neural networks according to claim 2, characterized in that, The radar image of the antenna phase center obtained by discretization includes: The iterative residual is decomposed into residuals, which are expressed as follows: Where, ε k The residual of the k-th atom, s is the measurement vector, A k The k-th row of the coefficient matrix The k-th element of the result vector obtained by the OMP algorithm; Based on the components after residual iteration, the optimal range of error is obtained, thereby determining the minimum value of error.

8. The ISAR 3D point cloud target recognition method based on neural networks according to claim 1, characterized in that, The process of acquiring radar echo data from different aircraft models received by the distributed array antenna system to generate and save n-channel ISAR images includes: Acquire radar echo data from all scattering points received by the distributed array antenna system, and generate and save an ISAR two-dimensional image with n channels using the radar echo data from all scattering points received. Based on the obtained n-channel ISAR two-dimensional images, a two-dimensional ISAR image is obtained using coordinate transformation and the Doppler algorithm.

9. The ISAR three-dimensional point cloud target recognition method based on neural networks according to claim 8, characterized in that, The process of acquiring radar echo data from all scattering points received by the distributed array antenna system, and generating and saving an ISAR two-dimensional image with n channels using the received radar echo data from all scattering points, includes: At time t, the distance R(t) from the p-th scattering point on the target to the radar is expressed as: Among them, R m Let x' be the distance from the radar to the center of the target. p ,y′ p Let be the coordinates of the p-th scattering point on the target in the xO'y coordinate system; Based on the coordinate transformation, the total echo signal from the L scattering points on the target is obtained. for: Where rect() represents the rectangular window function, The time within a radar pulse is represented by j (imaginary unit), λ (wavelength), and t (time). Two-dimensional ISAR images are then obtained based on the echo signals.

Images