Enhanced identification method for radar target cross-domain random perturbation imaging

Abstract

The application discloses an enhanced identification method for radar target cross-domain random disturbance imaging, and mainly solves the problem that direct application of a general data enhancement method to radar data expansion results in loss of target key characteristics.The implementation scheme is as follows: a training sample set and a test sample set are constructed; a radar complex data cross-domain expression framework is constructed; the training sample data are projected cross-domain by using the radar complex data cross-domain expression framework; the projected results are disturbed by using generated random masks with different shapes and sizes; enhanced results are obtained by re-imaging the random disturbed data according to the radar imaging principle; a classifier network model is constructed; the classifier network model is iteratively trained by using the enhanced results; and target identification results are obtained by inputting the test sample set into the trained classifier network model.The application can keep the target key characteristics complete in the enhanced results, effectively improves the target identification precision, and can be used for target identification, detection or image segmentation and other interpretation tasks.

Description

Technical Field

[0001] This invention belongs to the field of signal and information processing technology, specifically relating to a radar target enhancement and recognition method that can be used for interpretation tasks such as target recognition, detection, or image segmentation. Background Technology

[0002] Synthetic Aperture Radar (SAR) can operate continuously under various weather conditions, providing unique and useful information. It has received widespread attention and application in fundamental radar signal processing tasks such as image fusion, near-shore vessel detection, and target recognition. Early research often relied on handcrafted features, such as target shadows, statistical distributions, and time-frequency coefficients. These features were preprocessed to extract the target's scattering center and encode the target features. However, the identifiability of these handcrafted features significantly impacted the final detection performance. In recent years, the development of deep learning has brought widespread attention to data-driven algorithms. Many deep learning-based radar-related algorithms have been proposed, and their effectiveness in SAR has been validated. However, the learning efficiency of deep models is severely affected by the quantity of high-quality data with labeled information. Furthermore, the operational space in SAR data collection is highly variable, making it nearly impossible to collect enough measurement data to cover the entire radar sensor space. This limited sample size causes the training process to fall into local optima, hindering the improvement of deep model fitting capabilities.

[0003] Currently, a common solution is data augmentation. Many researchers have combined it with downstream tasks, achieving effective results improvements. Patent application CN202111251808.3 discloses a "mask-based data augmentation method," which continuously collects flawed original images using big data and other technologies. The flawed regions are manually marked on the original images to form mask areas. The mask bounding boxes are calculated based on the coordinates of each point in the mask area, forming flaw boxes. An image mask library is established. Image masks are randomly selected from the library for calculation to ensure the randomness of the generated flaws, simulating flaws found in real-world applications as much as possible. Image transformation operations are used to generate flaws on the target image by replacing regions. The target region of the existing limited data is extracted, transformed, and added to the target region of the specified image. Although this method can generate large sample data from small sample data, it is based on the amplitude information of optical images. Since radar data not only contains small target characteristics but also complex phase information, directly applying this method to radar data will result in the loss of important features in radar targets and will not be able to utilize the inherent complex characteristics to improve learning accuracy.

[0004] Patent document CN202311442013.X discloses a "Radar Dataset Augmentation Method Based on Generative Adversarial Neural Networks," which uses a generative adversarial neural network to transform randomly sampled noise into radar simulation data for the desired scenario through game-like training between the generator and the decision maker, thereby achieving the purpose of radar data augmentation. However, the training process of this method is unstable and prone to model crashes. Furthermore, the training time and computational resource requirements are too high, resulting in high costs and making it unsuitable as a lightweight component for downstream tasks. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing an enhanced identification method for radar target cross-domain random disturbance imaging. This method utilizes cross-domain information of radar signals to stably enhance targets in radar complex data, thereby avoiding the loss of important characteristics of the enhancement results and improving target identification accuracy without significantly increasing computational resources.

[0006] The technical approach to achieving the objective of this invention is as follows: by utilizing the radar imaging principle to perform cross-domain representation enhancement on radar targets to avoid the loss of important characteristics of the enhancement results; by simulating unpredictable disturbances in real-world scenarios through masking to increase the diversity of enhancement results, thereby further improving target recognition accuracy without significantly increasing computational resources.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following:

[0008] (1) Obtain the training sample set and the test sample set:

[0009] Obtain A complex radar data points X from a publicly available dataset, including C target categories, where C ≥ 4 and A ≥ 2000;

[0010] The radar complex data with elevation angles P and Q in the radar complex data X, along with their corresponding labels, are used to form a pre-training sample set and a test sample set, respectively.

[0011] (2) Construct a radar complex data cross-domain representation framework D, which consists of a cascaded domain transformation module and a data processing module;

[0012] (3) Input the training sample set into the radar complex data cross-domain representation framework D for data reprojection, and randomly freeze the data obtained after projection to simulate unpredictable disturbances in real-world scenarios, thereby obtaining the data X after adding the disturbance. * ;

[0013] (4) Add perturbation to the data X * As input to the inverse execution of the radar complex data cross-domain representation framework, the image is re-imagined to obtain the enhanced result X corresponding to the radar target projected back into the two-dimensional time domain. new ;

[0014] (5) Construct a classifier network model S that includes sequentially connected depthwise separable convolutional modules and a classification module;

[0015] (6) Using Enhanced Results X new The classifier network model S is iteratively trained through backpropagation to obtain the trained classifier network model S. * ;

[0016] (7) Input the test sample set into the trained radar target classification network model S * This yields the classification results for radar targets.

[0017] The present invention has the following advantages over the prior art:

[0018] 1. This invention introduces radar imaging mechanism and reprojects two-dimensional time domain data to other domains for target enhancement. Its effect is far superior to the enhancement result of simple rotation interpolation of amplitude data, effectively improving the problem of local data scarcity and improving target recognition accuracy.

[0019] 2. Because the domain transformation module constructed in this invention contains three different selectable Fourier transforms, the cross-domain information is not limited to the two-dimensional frequency domain, but can also be extended to the range frequency domain and azimuth frequency domain, which can make full use of the complex information of the four domains of radar in different scenarios.

[0020] 3. Because the present invention uses freezing to simulate unpredictable interference in real-world scenarios during the target enhancement process, it has high flexibility in practical applications. It can adopt various operation forms such as cutting and hiding according to specific needs, which improves the diversity of enhancement results and further enhances the target recognition accuracy.

[0021] 4. Because the method proposed in this invention is for radar complex data rather than two-dimensional images, it can be extended to data in various different forms of expression compared with the prior art;

[0022] 5. Because the classifier network model constructed in this invention can process the input data in real time during iterative training, the data augmentation mode can autonomously select online augmentation and offline augmentation. In practical applications, the augmentation mode can be selected by comprehensively considering factors such as model performance, training speed and resource utilization efficiency in specific scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0024] Figure 2 This is a schematic diagram of the structure of the classification network model constructed in this invention;

[0025] Figure 3 The present invention and existing comparative methods show the reconstruction results of radar complex data in the publicly available MSTAR dataset;

[0026] Figure 4 This invention compares the recognition accuracy of the MSTAR public dataset with existing methods under various limited environments. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Reference Figure 1 The present invention includes the following steps:

[0029] Step 1: Obtain the training sample set and the test sample set.

[0030] Obtain A complex radar data points X from a publicly available dataset, including C target categories, where C ≥ 4 and A ≥ 2000;

[0031] The radar complex data with elevation angles P and Q in the radar complex data X, along with their corresponding labels, are used to form a pre-training sample set and a test sample set, respectively.

[0032] In this example, C = 10, A = 2747, P = 17°, Q = 15°; the radar data is a vehicle target acquired by SAR, with an operating band of X-band, a center frequency of 9.6 GHz, a signal bandwidth of 0.591 GHz, a spotlight imaging mode, a multiplicative noise of -10 dB, an additive noise range of -32 to 34 dB, a dynamic range of 64 dB, an azimuth beamwidth of 8.8 degrees, a tilt beamwidth of 6.8 degrees, a resolution of 0.3 × 0.3 meters, a pixel pitch of 0.2 × 0.2 meters, and a data size of 128 × 128.

[0033] Step 2: Construct a cross-domain representation framework for radar complex data.

[0034] Construct a cross-domain representation framework D for radar complex data, consisting of a cascaded domain transformation module and a data processing module;

[0035] 2.1) Establish a domain transformation module including range-to-Fourier transform, azimuth-to-Fourier transform, and two-dimensional Fourier transform, and select one of them for cross-domain representation of radar complex data;

[0036] 2.2) Establish a data processing module consisting of sequentially connected zero-filling operations and window function weighting elimination operations for data preprocessing;

[0037] 2.3) The domain transformation module and the data processing module are cascaded to form a radar complex data cross-domain expression framework D.

[0038] Step 3: Perform interference simulation on the radar complex data.

[0039] 3.1) Input the training sample set into the radar complex data cross-domain representation framework D for data reprojection. For each radar complex data image, select one of the three Fourier transforms (range-direction Fourier transform, azimuth-direction Fourier transform, and two-dimensional Fourier transform) for corresponding cross-domain projection to obtain A images of cross-domain represented two-dimensional data X′:

[0040]

[0041] Among them, F R (·) represents the distance-to-Fourier transform, F A (·) indicates the azimuth-to-Fourier transform; the two-dimensional data X′ is projected from top to bottom into the range frequency domain, azimuth frequency domain, and two-dimensional frequency domain, respectively;

[0042] 3.2) For the two-dimensional data X′ obtained after projection, retain the components of its frequency domain data center M×N unchanged, and delete the other remaining components of the frequency domain data to obtain the zero-removed two-dimensional data X′. Z :

[0043] X′ Z =Z -1 (X′)

[0044] Among them, Z -1 (·) represents the zeroing operation. In this example, the data projected to the distance frequency domain is set to, but is not limited to, M=100, N=128; the data projected to the azimuth frequency domain is set to, but is not limited to, M=128, N=100; and the data projected to the two-dimensional frequency domain is set to, but is not limited to, M=100, N=100.

[0045] 3.3) For the zero-removed two-dimensional data X′ Z Perform a windowing operation to obtain radar cross-domain complex data X′ with a frame size of M×N. W :

[0046] X′ W =W -1 (X′ Z )

[0047] Among them, W -1 (·) represents the windowing operation. In this example, the window function is the Taylor window function. The settings for the data projected to the range frequency domain are not limited to M=100, N=128. The settings for the data projected to the azimuth frequency domain are not limited to M=128, N=100. The settings for the data projected to the two-dimensional frequency domain are not limited to M=100, N=100.

[0048] 3.4) Generate a random mask of varying shape and size to simulate unpredictable disturbances during real-world data collection. Apply this mask to the frequency data, zeroing out the frequency components within each individual data mask while keeping the remaining frequency components outside the mask unchanged. This yields the simulation result of the unpredictable disturbance, which is the data X after the disturbance has been added. * :

[0049] X * =T(X′) W )

[0050] Here, T(·) represents the frequency freeze operation, which can perform any operation on the random frequency components, such as cutting or hiding. The goal of this freeze operation strategy is to simulate random perturbations so that any similar operation can achieve the same result. In this example, the specific operation of T(·) is to freeze the component to zero.

[0051] Step 4: Obtain the cross-domain reconstruction results of the radar target.

[0052] 4.1) For the data X after adding perturbation * Apply window function weighting to make the resulting image smoother.

[0053] 4.2) After applying the window function, the data is padded with zeros to a size of X×Y to eliminate the sidelobe effect. In this example, X = 128 and Y = 128.

[0054] 4.3) Using the inverse operation of the corresponding Fourier function selected in step 3.1), the cross-domain data is reprojected back into the two-dimensional time domain to obtain the enhanced result X of the radar target projected back into the two-dimensional time domain. new :

[0055] X new =F -1 (Z(W(X * )))

[0056] Where Z(·) is the zero-padding operation, W(·) is the windowing operation, and in this example, the window function is the Taylor window function, F -1 (·) represents the inverse Fourier transform operation.

[0057] Step 5: Construct a classifier network model S with depthwise separable convolutional modules and a classification module connected sequentially.

[0058] Reference Figure 2 The classifier network model S constructed in this step has the following structure:

[0059] The depthwise separable convolutional module contains N inverted residual structures, each of which includes sequentially connected extended convolutional layers, depthwise convolutional layers, and projective convolutional layers.

[0060] This extended convolutional layer has a kernel size of 1×1, which is used to increase the number of channels;

[0061] This deep convolutional layer has a depthwise separable convolution size of 3×3. It performs spatial convolution directly on each channel of the input feature map without changing the number of channels.

[0062] This projective convolutional layer, with a kernel size of 1×1, is used to reduce the number of channels to the required number of output channels;

[0063] The stride of each of the above convolutional kernels is 1, the padding is 1, and the non-linear activation function layers are all ReLU functions. In this example, N=7.

[0064] The classification module employs a two-dimensional convolutional neural network comprising M sequentially connected two-dimensional convolutional layers, an average pooling layer, a fully connected layer, and a Softmax activation function.

[0065] This two-dimensional convolutional layer, with a kernel size of 1×1 and a stride of 1, is used for feature extraction to generate a feature map.

[0066] This average pooling layer has a 7×7 kernel size, which is used to shrink the feature map and reduce the amount of computation.

[0067] This fully connected layer has a 1×1 kernel size, which is used to expand the feature map into a one-dimensional vector and provide input to the Softmax activation function.

[0068] The Softmax activation function is used to normalize a numerical vector into a probability distribution vector, and selects the target with the highest probability as the output of the entire classifier network model S. In this example, M=1.

[0069] Step 6: Iteratively train the classifier network model.

[0070] 6.1) Set the initial iteration count to m = 0, and set the maximum iteration count to M. max If the value is ≥100, the network parameters of the classifier network model S are randomly initialized to θ0;

[0071] 6.2) The enhanced result X obtained in step 4.3) new The pre-training samples are input into the classifier network model S, where the depthwise separable convolutional module extracts features, generating multiple sets of mapping features. The classification module then performs target classification on each set of mapping features, outputting multiple probability vectors p of length C. i ;

[0072] 6.3) Based on the probability vector p i Using the cross-entropy loss function L CECalculate the loss value of the classifier network. m :

[0073] Loss m =L CE (y i ,p i )

[0074] Among them, y i This is the one-hot encoded vector of the label corresponding to the input sample;

[0075] 6.4) Using the backpropagation algorithm, based on the loss value... m Calculate network parameters θ m gradient

[0076] 6.5) Employ the gradient descent algorithm, based on the calculated gradient... Update network parameters θ m The updated network parameters are obtained: Where α represents the learning rate, Indicates the differentiation operation;

[0077] 6.6) Determine if the current iteration count has reached the maximum iteration count:

[0078] If the target is reached, training ends, and the trained classifier network model S is obtained. * ;

[0079] Otherwise, let m = m + 1 and return to step 6.2).

[0080] Step 7, Obtain the trained classifier network model S * The result of target identification.

[0081] The test sample set is input into the trained radar target classification network model S. * The radar target identification results are obtained. The technical effects of this invention can be further illustrated by the following simulation experiments.

[0082] I. Simulation Conditions

[0083] The dataset used is MSTAR.

[0084] The relevant imaging parameters of the dataset used are shown in Table 1:

[0085] Table 1. Relevant imaging parameters of the MSTAR dataset.

[0086]

[0087] The simulation environment is shown in Table 2:

[0088] Table 2 Basic Experimental Environment

[0089]

[0090] II. Simulation Content

[0091] Simulation 1: Under the imaging parameters of the MSTAR dataset listed in Table 1 above, target augmentation of the publicly available MSTAR dataset was performed using the present invention and existing methods for target augmentation based on amplitude information. The results are as follows: Figure 3 ,in:

[0092] Figure 3 (a) is the original radar image.

[0093] Figure 3 (b) are random masks of varying sizes and shapes.

[0094] Figure 3 (c) is a diagram showing the result of target enhancement based on the amplitude information of radar data.

[0095] Figure 3 (d) is an enhanced image of cross-domain random disturbance imaging of radar data in the range-frequency domain.

[0096] Figure 3 (e) is an enhanced image of cross-domain random disturbance imaging of radar data in the azimuth frequency domain.

[0097] Figure 3 (f) is an enhanced image of cross-domain random disturbance imaging of radar data in the two-dimensional frequency domain.

[0098] from Figure 3 As can be seen, the target key features are significantly missing in the result image when the amplitude information is enhanced, while the target key features obtained by the enhancement of this invention are more complete, which is conducive to the network extracting and learning higher quality features, thereby improving the recognition accuracy.

[0099] Simulation 2, under the basic experimental conditions outlined in Table 2 above, gradually reduced the training sample size from 100% to 50%, 40%, 30%, 20%, and 10% of the original size. The enhancement results of the present invention and existing target enhancement methods based on amplitude information in Simulation 1 were compared with those in Simulation 1. Figure 2 The classifier network model shown performs target recognition, and its recognition accuracy is as follows: Figure 4 ,in:

[0100] Curve ① represents the recognition result without data augmentation;

[0101] Curve ② is the recognition result after target enhancement based on existing amplitude information;

[0102] Curve ③ represents the target recognition result after target enhancement in the range frequency domain according to this invention;

[0103] Curve ④ represents the target recognition result after target enhancement in the azimuth frequency domain according to this invention;

[0104] Curve ⑤ represents the target recognition result after target enhancement in the two-dimensional frequency domain according to this invention;

[0105] from Figure 4 As can be seen, under various few-sample environments, the present invention significantly improves accuracy, with a maximum gain of 18.93%. Compared to target augmentation using amplitude information, the present invention delivers a higher performance improvement to the classifier network model. Furthermore, as the number of training samples gradually decreases, the accuracy improvement of the present invention gradually increases, indicating that the present invention has significant advantages in few-sample environments.

Claims

1. An enhanced identification method for radar target cross-domain random disturbance imaging, characterized in that, Includes the following steps: (1) Obtain the training sample set and the test sample set: Obtain A complex radar data points X from a publicly available dataset, including C target categories, where C ≥ 4 and A ≥ 2000; The radar complex data with elevation angles P and Q in the radar complex data X, along with their corresponding labels, are used to form a pre-training sample set and a test sample set, respectively. (2) Construct a radar complex data cross-domain representation framework D, which consists of a cascaded domain transformation module and a data processing module; (3) Input the training sample set into the radar complex data cross-domain representation framework D for data reprojection, and randomly freeze the data obtained after projection to simulate unpredictable disturbances in real-world scenarios, thereby obtaining the data X after adding the disturbance. * ; (4) Add perturbation to the data X * As input to the inverse execution of the radar complex data cross-domain representation framework, the image is re-imagined to obtain the enhanced result X corresponding to the radar target projected back into the two-dimensional time domain. new ; (5) Construct a classifier network model S that includes sequentially connected depthwise separable convolutional modules and a classification module; (6) Using Enhanced Results X new The classifier network model S is iteratively trained through backpropagation to obtain the trained classifier network model S. * ; (7) Input the test sample set into the trained radar target classification network model S * The classification results of radar targets are obtained.

2. The method according to claim 1, characterized in that, Step (2) Construct the domain transformation module and data processing module in the radar complex data cross-domain representation framework D, the structure of which is as follows; The domain transformation module includes range-to-Fourier transform, azimuth-to-Fourier transform, and two-dimensional Fourier transform, one of which is selected for cross-domain representation of radar complex data; The data processing module includes sequentially connected zero-padding and window function weighting operations for data preprocessing.

3. The method according to claim 1, characterized in that, In step (3), the training sample set is input into the radar complex data cross-domain representation framework D for data reprojection. The data obtained after projection is randomly frozen to simulate unpredictable disturbances in real-world scenarios, as follows: (3a) For each radar complex data image, select one of the three Fourier transforms—range Fourier transform, azimuth Fourier transform, and two-dimensional Fourier transform—and perform the corresponding cross-domain projection to obtain A cross-domain representations of two-dimensional data X′: Among them, F R (·) represents the distance-to-Fourier transform, F A (·) indicates the azimuth-to-Fourier transform; from top to bottom, the data are projected into the range frequency domain, azimuth frequency domain, and two-dimensional frequency domain, respectively; (3b) For each two-dimensional data X′, retain the components of its frequency domain data center M×N part unchanged, delete the other remaining components of the frequency domain data, and obtain the zero-removed two-dimensional data X′. Z : X′ Z =Z -1 (X′) Among them, Z -1 (·) represents a zero-removal operation; (3c) The zero-removed two-dimensional data X′ Z Perform a windowing operation to obtain radar cross-domain complex data X′ with a frame size of M×N. W : X′ W =W -1 (X′ Z ) Among them, W -1 (·) represents the window clearing operation; (3d) Generate a random mask of varying size and shape to simulate unpredictable disturbances during real-world data collection. Apply this mask to the frequency data, zeroing out the frequency components within a single image mask while leaving the remaining frequency components outside the mask unchanged. This yields the simulation result of the unpredictable disturbance, which is the data X after adding the disturbance. * : X * =T(X′ W ) Where T(·) represents the frequency freeze operation.

4. The method according to claim 1, characterized in that, In step (4), the perturbed data X will be added. * As input to the inverse execution of the radar complex data cross-domain representation framework, imaging is performed again, achieving the following: (4a) The data X after adding perturbation simulation * Apply window function weighting to make the resulting image smoother. (4b) Pad the data after applying the window function with zeros to a size of X×Y to eliminate the sidelobe effect; (4c) Using the inverse operation of the corresponding Fourier function selected in step (3a), the cross-domain data is reprojected back into the two-dimensional time domain to obtain the enhanced result X corresponding to the two-dimensional time domain. new : X new =F -1 (Z(W(X * ))) Where Z(·) is the zero-padding operation, W(·) is the windowing function operation, and F -1 (·) represents the inverse Fourier transform operation.

5. The method according to claim 1, characterized in that, Step (5) Construct the depthwise separable convolutional module and classification module included in the classifier network model S, with the following structure: The depthwise separable convolutional module contains N inverted residual structures, each of which includes sequentially connected extended convolutional layers, depthwise convolutional layers, and projective convolutional layers. This extended convolutional layer has a kernel size of 1×1, which is used to increase the number of channels; This deep convolutional layer has a depthwise separable convolution size of 3×3. It performs spatial convolution directly on each channel of the input feature map without changing the number of channels. This projective convolutional layer, with a kernel size of 1×1, is used to reduce the number of channels to the required number of output channels; Each of the above convolutional kernels has a stride of 1 and padding of 1, and the non-linear activation function layers are all ReLU functions; The classification module includes one or more sequentially connected two-dimensional convolutional layers, average pooling layers, fully connected layers, and a Softmax activation function. This two-dimensional convolutional layer, with a kernel size of 1×1 and a stride of 1, is used for feature extraction to generate a feature map. This average pooling layer has a 7×7 kernel size, which is used to shrink the feature map and reduce the amount of computation. This fully connected layer has a 1×1 kernel size, which is used to expand the feature map into a one-dimensional vector and provide input to the Softmax activation function. The Softmax activation function is used to normalize a numerical vector into a probability distribution vector, and selects the target with the highest probability as the output of the entire classifier network model S.

6. The method according to claim 1, characterized in that, Step (6) uses the augmentation results to iteratively train the classifier network model S through backpropagation, as follows: (6a) Set the initial iteration count to m = 0, and set the maximum iteration count to M. max If the value is ≥100, the network parameters of the classifier network model S are randomly initialized to θ0; (6b) Input the pre-training sample set into the classifier network model S. The depthwise separable convolutional module in the classifier network model extracts features from the input pre-training samples and generates multiple sets of mapping features. The classification module performs target classification on each set of mapping features and outputs multiple probability vectors p of length C. i ; (6c) Based on the probability vector p i Using the cross-entropy loss function L CE Calculate the loss value of the classifier network. m : Loss m =L CE (y i ,p i ) Among them, y i This is the one-hot encoded vector of the label corresponding to the input sample; (6d) Using the backpropagation algorithm, based on the loss value... m Calculate network parameters θ m gradient (6e) Employ the gradient descent algorithm, based on the calculated gradient... Update network parameters θ m The updated network parameters θ are obtained. m+1 : Where α represents the learning rate, Indicates the differentiation operation; (6f) Determine if the current iteration count has reached the maximum iteration count: If the target is reached, training ends, and the trained classifier network model S is obtained. * ; Otherwise, let m = m + 1 and return to step (6b).

Images