A cross-domain small sample radar one-dimensional image target recognition method based on feature decoupling

Abstract

The present application belongs to the technical field of radar target recognition, and particularly relates to a cross-domain small sample radar one-dimensional image target recognition method based on feature decoupling. The present application uses a simulation data set and a measured data set as sample data, and the network constructed includes a feature extraction module, a decoupling module, a data enhancement module and a classification module. The decoupling module decouples the original features into domain-independent features and domain-dependent features, and the classification module includes a domain classifier and a category classifier, which discriminates the domain-dependent features and the domain-independent features extracted from the simulation data, the mixed data and the measured data. The present application combines the simulation data with the measured data, uses the simulation data to assist the training of the measured data, and improves the recognition rate of the model for the measured target.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a cross-domain small-sample radar one-dimensional image target recognition method based on feature decoupling. Background Technology

[0002] Target recognition based on high-resolution one-dimensional range profiles of radar is a key research area in the field of automatic radar target recognition. The rise of deep learning methods has brought numerous conveniences to automatic radar target recognition and significantly improved target recognition accuracy. However, in real-world environments, it is very difficult to obtain a sufficient number of non-cooperative target samples with complete attitude angles. Often, the target's flight trajectory is simple, the attitude angle changes are small, the diversity of the collected samples is insufficient, and the recognition performance of the trained model is poor.

[0003] Although real-world samples of non-cooperative targets are difficult to obtain, simulation software can generate sufficient simulation data. While these simulation data differ significantly from real-world data, and even the target types may differ, the simulation data can be used to assist in training with the real-world data. The model can learn characteristics relevant to the real-world data from the simulation data, and transferring this knowledge from the simulation data to the real-world data can effectively alleviate the training difficulties caused by the lack of real-world data. Summary of the Invention

[0004] This invention addresses the small sample size problem in HRRP recognition in real-world testing environments by proposing a cross-domain small sample HRRP recognition method based on feature decoupling. This method combines simulation data with real-world data, using simulation data to assist in training the model on real-world targets, thereby improving the model's recognition rate for real-world targets.

[0005] The technical solution of this invention is as follows:

[0006] A cross-domain small-sample radar one-dimensional image target recognition method based on feature decoupling includes the following steps:

[0007] S1. Constructing the dataset:

[0008] The one-dimensional image simulation dataset and the measured dataset of the target are obtained as sample data respectively. The simulation dataset is used as the source domain S and the measured dataset is used as the target domain T.

[0009] S2, Data Preprocessing:

[0010] All sample data in S1 are preprocessed, including center alignment and energy normalization. Center alignment is used to solve translation sensitivity, and energy normalization is used to solve amplitude sensitivity. Each sample has a dimension of 1×N, where N is the number of sampling points.

[0011] S3. Construct a network model, such as Figure 1As shown, it includes a feature extraction module, a decoupling module, a data augmentation module, and a classification module, specifically:

[0012] The feature extraction module uses a one-dimensional Transformer network as the feature extractor, which shares parameters across three different stages: source domain data pre-training, mixed data meta-training, and target domain data fine-tuning. The input for the pre-training stage is the source domain data; the input for the meta-training stage is the mixed data augmented by the data augmentation module (source and target domain data); and the input for the fine-tuning stage is the target domain data. The data augmentation module employs the CutMix data augmentation method, whose goal is to combine two training samples (x...)... A y A ) and (x B y B Generate new training samples The merging operation, which trains the model using new training samples, is represented as follows:

[0013]

[0014]

[0015] Where M∈{0,1} L represents a binary mask used to cover a portion of the region so that the corresponding portion of another sample can be filled; ⊙ represents pointwise multiplication, and α follows a uniform distribution of (0, 1); to sample the binary mask M, the boundaries of the cropping region must first be sampled, indicating the area where the two samples need to be mixed. The cropping box is represented as:

[0016] r x ~U(0,L),r l =L(1-α)

[0017] Where, r x r represents the starting position of the cropping region. l L is the length of the clipping region, and L is the sample dimension. After determining the clipping region, set the clipping region in the mask region to 0 and the other regions to 1. This will remove the clipping region in one sample and fill the clipping region in another sample into the sample, thus completing the sample mixing operation.

[0018] The decoupling module consists of several fully connected layers and batch normalization layers, such as Figure 2 As shown, fully connected layers 2-1 and 2-2 constitute the first decoupling module, used to learn the mean and variance as domain-independent features. Fully connected layers 3-1 and 3-2 constitute the second decoupling module, used to extract domain-dependent features. The input to the decoupling module is the original features output by the feature extractor. The decoupling module decouples the original features into domain-independent features and domain-dependent features.

[0019] The classification module includes a domain classifier and a category classifier. The domain classifier performs binary classification on the samples, determining whether a sample belongs to the simulated domain or the measured domain. The domain classifier includes a domain-independent classifier C. di Domain-related classifier C ds Domain-independent classifiers are used to perform domain discrimination on domain-independent features extracted from simulation data, mixed data, and measured data; domain-related classifiers are used to perform domain discrimination on domain-related features extracted from simulation data, mixed data, and measured data.

[0020] This invention measures the correlation between two separated features by calculating the KL divergence loss between the output of the neural network and the true label. The KL divergence loss can be defined as:

[0021]

[0022] Where y is the real label. This is the output of the neural network. To ensure that the separated features retain as much information as possible from the original features, it is desirable that the separated features can be reconstructed back into the original features.

[0023] Domain-independent discriminant loss is represented as:

[0024]

[0025] Where KL(·) is the divergence loss, The output of the simulation data after passing through the domain-independent classifier, The output of the domain-independent classifier is the measured data. The output of the domain-independent classifier is the mixed data; y s y t y m The purpose of setting all values ​​to 0.5 is to prevent the domain-independent classifier from distinguishing whether a sample belongs to the measured domain or the simulated domain.

[0026] Domain-related discriminant loss can be expressed as:

[0027]

[0028] Where λ is the confidence score of the category to which the mixed data belongs, i.e., the mixing ratio of the mixed data, and CE(·) is the cross-entropy loss. The output of the simulation data after passing through the domain correlation classifier, The output of the domain correlation classifier is the measured data. The output of the domain correlation classifier is the mixed data; y2 s y2 m1 Set 1, then set y2 t y2m2 Setting it to 0 is intended to enable the domain-related classifier to distinguish the category of the domain to which the sample belongs;

[0029] The classifiers consist of a first-class classifier S and a second-class classifier T. The inputs to the first-class classifier S and the second-class classifier T are domain-independent features. The target class classification loss is expressed as:

[0030] L cls =λ·L cls-s +(1-λ)L cls-t

[0031] Where L cls-s For the classification loss of simulation data, L cls-t The classification loss is based on the measured data; the total loss function is expressed as:

[0032] L = L cls +L di +L ds +L rec

[0033] Where L rec For reconstruction loss:

[0034]

[0035] Where X represents the original features extracted by the first feature extractor, x1 and x2 are the features separated by the decoupling module, and f con (·) is a reconstructed network;

[0036] S4. Using simulation data as the source domain and measured data as the target domain, train the system using samples from both domains. The specific training steps are as follows:

[0037] S41. Pre-training stage: Use the HRRP simulation dataset to pre-train the feature extraction module and the decoupling module;

[0038] S42. Meta-training Phase: The input to the meta-training phase consists of simulation data and experimental data. The input data includes three branches: the support set of the simulation data, the support set of the experimental data, and the mixed data after data augmentation of the simulation data and the experimental data. Each branch undergoes feature extraction and decoupling. Decoupling separates domain-independent features and domain-dependent features. The purpose of extracting domain-independent features is to enable the simulation data and the experimental data to be represented in the same feature space. The features after decoupling will pass through two domain classifiers and two classifiers. The domain classifiers are used to distinguish between domain-dependent features and domain-independent features, and the classifiers are used for target classification.

[0039] S43. Fine-tuning stage: Only input the actual test data, and use the actual test data to fine-tune the feature extraction network and decoupling module to obtain the trained network;

[0040] S5. Input the acquired one-dimensional radar image of the target into the trained network for identification.

[0041] The beneficial effects of this invention are: by combining simulation data with measured data, and using simulation data to assist in training measured data, the model's recognition rate of measured targets is improved. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall network model structure of the present invention.

[0043] Figure 2 This is a schematic diagram of the decoupling module.

[0044] Figure 3 Dimensionality reduction visualization of t-sne, where (a) is CNN; (b) is FCDNN. Detailed Implementation

[0045] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0046] Example

[0047] S1. Constructing the dataset:

[0048] The simulation data includes six target classes: Ah-64, ALCM, An-26, B-1B, B-52, and F-15, totaling 6×9×1800 samples, which are used as the source domain data. The measured data includes four types of civil aircraft: A319, A320, A321, and B738, which are used as the target domain data. From each type of civil aircraft, a flight path with a significant change in attitude angle is selected as the target for database construction. From all samples with a signal-to-noise ratio greater than 23dB, samples are extracted using a method of selecting 1 every N samples to form the training set. The remaining samples of this flight path are used as the validation set, and the remaining flight path samples are used as the test set.

[0049] S2, Data Preprocessing:

[0050] Center alignment is used for all samples in S1 to address translation sensitivity, and energy normalization is used to address amplitude sensitivity. Each simulated sample has a dimension of 1×300, and each measured sample has a dimension of 1×600. To unify the dimensions of the measured and simulated data input models, each measured sample is downsampled. Since the data length of the simulated sample is 300, downsampling can be achieved by taking every other sample point from each measured sample, resulting in a data length of 300 for each measured sample. Gaussian white noise is randomly added to each simulated data set.

[0051] S3. Constructing the network model:

[0052] The overall network architecture of this invention is as follows: Figure 1 As shown, it is mainly divided into a feature extraction module, a decoupling module, a data augmentation module, and a classification module.

[0053] S31, Feature Decoupling Module. Minimizing mutual information allows the original features to be separated into two completely independent features. This study measures the correlation between the two separated features by calculating the KL divergence loss between the neural network output and the true label. The KL divergence loss can be defined as:

[0054]

[0055] Where y is the real label. This is the output of the neural network. To ensure that the separated features retain as much information as possible from the original features, it is desirable that the separated features can be reconstructed back to the original features. Therefore, a reconstruction loss is defined:

[0056]

[0057] Where X represents the original features extracted by the first feature extractor, x1 and x2 are the features separated by the decoupling module, and f con (·) is a reconstructed network.

[0058] S32, Data Augmentation Module. To fully blend images from the two domains and enhance the discriminative power of the domain classifier, data augmentation is employed. This study uses the CutMix data augmentation method. The goal of CutMix is ​​to combine two training samples (x...)... A y A ) and (x B y B Generate new training samples The model is trained using the new training samples. The merging operation can be represented as:

[0059]

[0060]

[0061] Where M∈{0,1} L This represents a binary mask used to cover a portion of the region so that it can be filled with that portion of another sample. ⊙ represents pointwise multiplication, and α follows a uniform distribution of (0, 1). To sample the binary mask M, the boundaries of the cropping region must first be sampled, indicating the area where the two samples need to be blended. The cropping box can be represented as:

[0062] r x~U(0,L),r l =L(1-α)

[0063] Where, r x r represents the starting position of the cropping region. l L represents the length of the clipping region, and L represents the sample dimension. After determining the clipping region, set the clipping region in the mask area to 0 and the other regions to 1. This will remove the clipping region from one sample and fill the clipping region from another sample into that sample, thus completing the sample blending operation.

[0064] S33. Classifier Module. The classifier module contains two types: domain classifiers and category classifiers. Domain classifiers perform binary classification on samples, determining whether a sample belongs to the simulated domain or the measured domain. There are two domain classifiers: one is a domain-independent classifier, which performs domain discrimination on domain-independent features extracted from simulated data, mixed data, and measured data; the other is a domain-dependent classifier, which performs domain discrimination on domain-dependent features extracted from simulated data, mixed data, and measured data. The domain-independent discrimination loss can be expressed as:

[0065]

[0066] in The output of the simulation data after passing through the domain-independent classifier, The output of the domain-independent classifier is the measured data. The output of the mixed data is processed by a domain-independent classifier. The domain-correlated discriminant loss can be expressed as:

[0067]

[0068] Where λ is the confidence score of the category to which the mixed data belongs, i.e., the mixing ratio of the mixed data, and CE(·) is the cross-entropy loss. The output of the simulation data after passing through the domain correlation classifier, The output of the domain correlation classifier is the measured data. The output of the mixed data is processed by a domain correlation classifier. The target class classification loss can be expressed as:

[0069] L cls =λ·L cls-s +(1-λ)L cls-t

[0070] Where L cls-s For the classification loss of simulation data, L cls-t The classification loss is for the measured data. The total loss function is expressed as:

[0071] L = L cls +L di +L ds+L rec

[0072] S4. Using simulation data as the source domain and measured data as the target domain, train the system using samples from both domains. The specific training steps are as follows:

[0073] S41. Pre-trained Feature Extraction Network and Decoupling Module. First, the feature extractor and decoupling module need to be pre-trained using the HRRP simulation dataset. The pre-training process employs a conventional training method. The feature extractor uses a one-dimensional Transformer network.

[0074] S42. Meta-training Phase. The input to the meta-training phase consists of simulation data and experimental data. The input data comprises three branches: the support set for the simulation data, the support set for the experimental data, and a hybrid dataset consisting of augmented simulation and experimental data. Each branch passes through a feature extractor and a decoupling module. The decoupling module separates domain-independent features from domain-dependent features. The purpose of extracting domain-independent features is to ensure that the simulation and experimental data can be represented in the same feature space. The features processed by the decoupling module then pass through two domain classifiers and two classifiers. The domain classifiers distinguish between domain-dependent and domain-independent features, while the classifiers classify the target data.

[0075] S43. Fine-tuning Stage. The fine-tuning stage only inputs measured data. The feature extraction network and decoupling module are fine-tuned using the measured data.

[0076] S44. Testing Phase. The input for the testing phase is a test set of other tracks. After the features are extracted by the feature extractor, they are input into the decoupling module to obtain domain-independent features, which are then identified and classified by the category classifier.

[0077] The dataset was divided into four categories by taking 1 out of every 100, resulting in 414 training data points for the four categories of actual targets. The various indicators of each method were calculated based on the confusion matrix of the recognition rate. The statistical results are shown in Table 1, where FCDNN is the method proposed in this invention.

[0078] Table 1 Performance evaluation of various methods

[0079]

[0080] To more intuitively demonstrate the separability of the extracted features and verify the performance of the feature extraction network, dimensionality reduction and visualization are performed on the high-dimensional features extracted by CNN and FCDNN. t-SNE dimensionality reduction is used to visualize the features extracted by both methods. The feature distribution after reduction to two dimensions is shown below. Figure 3 As shown.

[0081] from Figure 3As can be seen, the CNN method has a significantly higher degree of feature aliasing, while the clustering effect of the present invention is more obvious, the degree of feature aliasing is lower, and the distinguishability is stronger.

Claims

1. A cross-domain small-sample radar one-dimensional image target recognition method based on feature decoupling, characterized in that, Includes the following steps: S1. Constructing the dataset: The one-dimensional image simulation dataset and the measured dataset of the target are obtained as sample data respectively. The simulation dataset is used as the source domain S and the measured dataset is used as the target domain T. S2, Data Preprocessing: Preprocess all sample data in S1, including center alignment and energy normalization. Each sample has a dimension of 1×N, where N is the number of sampling points. S3. Construct the network model, including a feature extraction module, a decoupling module, a data augmentation module, and a classification module, specifically: The feature extraction module uses a one-dimensional Transformer network as the feature extractor. The feature extractor shares parameters across three different stages: source domain data pre-training, mixed data meta-training, and target domain data fine-tuning. The input for the pre-training stage is the source domain data; the input for the meta-training stage is the mixed data augmented by the data augmentation module (source and target domain data); and the input for the fine-tuning stage is the target domain data. The data augmentation module uses the CutMix data augmentation method, which aims to combine two training samples... and Generate new training samples The model is trained using the new training samples, and the merging operation is represented as: ， ， in, This represents a binary mask, used to cover up a portion of the area so that the same portion of the area can be filled by another sample. To multiply point by point, obey Uniform distribution; for binary mask Sampling is performed on the boundaries of the cropping region to indicate the area where the two samples need to be blended. The cropping box is represented as follows: ， in, This is the starting point of the cropping area. The length of the cropping area, For the sample dimension, after determining the clipping region, set the clipping region in the mask region to 0 and the other regions to 1. This will remove the clipping region in one sample and fill the clipping region in another sample into that sample, thus completing the sample mixing operation. The decoupling module consists of several fully connected layers and batch normalization layers. The input to the decoupling module passes through fully connected layers, batch normalization layers, and activation functions, and then enters multiple parallel fully connected layers. These parallel fully connected layers constitute the first and second decoupling modules. The first decoupling module is used to learn the mean and variance as domain-independent features, while the second decoupling module is used to extract domain-dependent features. The input to the decoupling module is the original features output by the feature extractor, and the decoupling module decouples the original features into domain-independent features and domain-dependent features. The classification module includes a domain classifier and a category classifier. The domain classifier performs binary classification on the samples, determining whether a sample belongs to the simulated domain or the measured domain. The domain classifier includes a domain-independent classifier. Domain-related classifier Domain-independent classifiers are used to perform domain discrimination on domain-independent features extracted from simulation data, mixed data, and measured data; domain-related classifiers are used to perform domain discrimination on domain-related features extracted from simulation data, mixed data, and measured data. Domain-independent discriminant loss is represented as: ， in, For divergence loss, The output of the simulation data after passing through the domain-independent classifier, The output of the domain-independent classifier is the measured data. The output of the domain-independent classifier is used to process the mixed data; , , The purpose of setting all values ​​to 0.5 is to prevent the domain-independent classifier from distinguishing whether a sample belongs to the measured domain or the simulated domain. Domain-related discriminant loss is expressed as: ， in, The confidence score for the category to which the mixed data belongs, which is also the mixing ratio of the mixed data. For cross-entropy loss, The output of the simulation data after passing through the domain correlation classifier, The output of the domain correlation classifier is the measured data. The output of the domain correlation classifier is used to process the mixed data; , Set to 1, , Setting it to 0 is intended to enable the domain-related classifier to distinguish the category of the domain to which the sample belongs; The classifier consists of a source domain classifier S and a target domain classifier T. The inputs to the source domain classifier S and the target domain classifier T are domain-independent features. The class classification loss is expressed as: ， in For the classification loss of simulation data, The classification loss is based on the measured data; the total loss function is expressed as: ， in For reconstruction loss: ， in The raw features extracted by the feature extractor. and The features separated by the decoupling module It is a reconstructed network; S4. Using simulation data as the source domain and measured data as the target domain, train the system using samples from both domains. The specific training steps are as follows: S41. Pre-training stage: Use the HRRP simulation dataset to pre-train the feature extraction module and the decoupling module; S42. Meta-training Phase: The input to the meta-training phase consists of simulation data and experimental data. The input data includes three branches: the support set of the simulation data, the support set of the experimental data, and the mixed data after data augmentation of the simulation data and experimental data. Each branch undergoes feature extraction and decoupling to separate domain-independent features and domain-dependent features. The features after decoupling are processed by two domain classifiers and two classifiers. The domain classifiers are used to distinguish between domain-dependent features and domain-independent features, and the classifiers are used for target classification. S43, Fine-tuning stage: Only input the actual test data, and use the actual test data to fine-tune the feature extraction module and decoupling module to obtain the trained network; S5. Input the acquired one-dimensional radar image of the target into the trained network for identification.

Images