WO2024082374A1

WO2024082374A1 - Few-shot radar target recognition method based on hierarchical meta transfer

Info

Publication number: WO2024082374A1
Application number: PCT/CN2022/133980
Authority: WO
Inventors: 郭贤生; 张玉坤; 李林; 司皓楠; 钱博诚; 钟科; 黄健
Original assignee: 电子科技大学长三角研究院(衢州)
Priority date: 2022-10-19
Filing date: 2022-11-24
Publication date: 2024-04-25
Also published as: CN115345322B; CN115345322A

Abstract

The present invention belongs to the technical field of target recognition, and particularly relates to a few-shot radar target recognition method based on hierarchical meta transfer. In the present invention, features are extracted on the basis of an attention mechanism and hierarchical deep knowledge transfer at a feature level, a sample level and a task level is performed in order to search for an embedding space, such that a sample is close to a category atom of targets of the same category and is away from a category atom of targets of other categories. At the feature level, a feature encoder based on the attention mechanism is designed and global domain-invariant representations of samples are fully mined, such that the problem of domain difference in data distribution of the samples is solved; at the sample level, an atom encoder is designed, and more stable category atoms are generated, such that the influence of outlier samples is avoided; and at the task level, a meta learner is designed, learning experience of training tasks is accumulated and is transferred to a new task, and the capability of cross-task knowledge transfer of a model is developed, such that target recognition based on meta transfer is realized. The target recognition method of the present invention is an intelligent target recognition method.

Description

A small sample radar target recognition method based on hierarchical meta-transfer

Technical Field

The invention belongs to the technical field of radar target recognition, and in particular relates to a small sample radar target recognition method based on hierarchical element migration.

Background technique

Radar target recognition technology refers to the technology of using radar to detect targets and determine the type, model and other attributes of the target by analyzing the captured information. It shows great application potential in fields such as terrain exploration and battlefield reconnaissance. With the development of artificial intelligence technology, deep learning methods have attracted widespread attention from researchers due to their automatic and powerful feature extraction capabilities, which has promoted the emergence and advancement of intelligent radar target recognition technology. However, deep learning model training often relies on a large number of labeled samples. Due to timeliness constraints and resource limitations, obtaining a large number of labeled samples consumes huge manpower, material resources and time costs. Therefore, using meta-learning to share knowledge in small sample scenarios to improve target recognition performance is one of the current research hotspots in the field of radar target recognition technology.

The paper "Guo J, Wang L, Zhu D, et al. SAR Target Recognition With Limited Samples Based on Meta Knowledge Transferring Using Relation Network [C] // 2020 International Symposium on Antennas and Propagation (ISAP). IEEE, 2021: 377-378" proposes a small sample radar target recognition method based on contrastive learning. It constructs a neural network to calculate the distance between two input samples to analyze the degree of matching, so as to determine whether they belong to the same category. When classifying unlabeled samples, the label of the nearest labeled sample is used as the predicted label. However, this method requires comparing the sample to be tested with each labeled sample, which is cumbersome and complex to calculate. To solve the above problems, the paper "Cai J, Zhang Y, Guo J, et al. ST-PN: A Spatial Transformed Prototypical Network for Few-Shot SAR Image Classification [J]. Remote Sensing, 2022, 14(9): 2019" proposed a small sample radar target recognition method based on category atoms. The features of each type of labeled samples are averaged as category atoms. When classifying unlabeled samples, only their features need to be compared with category atoms, thereby reducing the computational complexity. However, this method does not conduct in-depth exploration of sample features and target features. The generated category atoms are easily affected by outlier samples in small sample scenarios, and have problems of poor quality and poor robustness, which in turn affects the target recognition performance. At the same time, considering that these meta-learning methods only seek the similarity relationship between samples, when faced with new tasks that are different from the training tasks, the model cannot optimize the cross-task transfer of knowledge. Therefore, studying small sample target recognition methods based on hierarchical meta-transfer is expected to further improve target recognition performance.

Summary of the invention

The purpose of the present invention is to provide a small sample radar target recognition method based on hierarchical meta-transfer to overcome the above-mentioned shortcomings. The present invention extracts features based on the attention mechanism, and hierarchical deep knowledge transfer at the feature level, sample level, and task level to seek an embedding space that makes the sample close to the category atoms of the same type of target and far away from the category atoms of other types of targets. Among them, a feature encoder based on the attention mechanism is designed at the feature level to fully exploit the global domain-invariant features of the sample to overcome the domain difference problem of the sample in the data distribution; an atom encoder is designed at the sample level to generate more stable category atoms to avoid the influence of outlier samples; at the task level, a meta-learner is designed to accumulate the learning experience of the training task and transfer it to the new task, cultivate the model's ability to transfer knowledge across tasks, and realize meta-transfer target recognition. Therefore, the small sample radar target recognition method based on hierarchical meta-transfer proposed in the present invention is an intelligent target recognition method.

The technical solution of the present invention is:

A small sample radar target recognition method based on hierarchical element migration includes the following steps:

S1. Obtaining original images of each target in the source domain and the target domain when the target is static through radar, and cutting the images obtained by observing the target at different azimuth angles to obtain samples;

S2. Use samples to build training tasks

Where P is the total number of tasks,

It includes support set and query set, where the support set is composed of labeled samples extracted from the source domain, and the query set is composed of labeled samples extracted from the target domain;

S3. Training and learning through hierarchical meta-transfer models

Update, specifically:

S31. Constructing a feature encoder based on the attention mechanism at the feature level

Initialize the feature encoder using a meta-learner

After that, extract

Deep global features of the support set and query set;

S32. Constructing a category atom encoder based on attention mechanism at sample level

Initialize the category atom encoder using a meta-learner

Afterwards, based on the obtained

Deep global feature calculation of support set samples

The category atoms of the support set are obtained, and the probability of the corresponding samples belonging to different categories is obtained according to the distance between the support set samples and the different category atoms. Then, the category atom loss function is designed and minimized according to the probability to update the category atom encoder and category atoms.

S33. Accumulate the learning experience of the current training task at the task level and update the meta-learner:

according to

The deep global features of the query set samples and the distances between atoms of different categories are used to obtain the probability that the corresponding samples belong to different categories. The meta-learner loss function is designed based on the probability, and the meta-learner is updated by minimizing the loss function to obtain the updated meta-learner.

S4. Complete all training tasks by repeating step S3 to obtain the meta-learner trained by all meta-training tasks. The trained meta-learner is recorded as

S5: The labeled samples of the task to be tested are the support set, and the unlabeled samples to be tested are the query set; the meta-learner obtained in S4 is used for initialization

A feature encoder for target recognition and a category atom encoder are obtained, and the feature encoder for target recognition is used to extract deep global features for the support set and query set samples. The category atom encoder for target recognition is used to calculate and update the category atoms based on the deep global features of the support set, and the distance function dist(·) is used to calculate the distance between the deep global features of the sample to be tested in the query set and the atoms of different categories, and the label of the category atom with the closest distance is selected as the predicted label of the sample to be tested to obtain the recognition result.

Furthermore, in step S2, the support set is constructed by extracting labeled samples in the source domain in the form of K way N shot, which is defined as

K way N shot means randomly extracting N labeled training samples from each category of K types of targets.

is the nth sample of the kth class target; the query set is composed of labeled samples extracted in the target domain in the form of K way M shot, defined as

in,

is the mth sample of the kth class target; the samples in the support set and query set are samples of the same class target in different domains, and the corresponding class labels are defined as

in,

Furthermore, in step S31, the feature encoder

It includes a neural network module and an attention mechanism module. The specific method of extracting deep global features is as follows:

Extract generalized features from samples through neural network modules;

The generalized features are divided into blocks and straightened into vectors. The dimension of each vector is d ₁ , denoted as [b ₁ , b ₂ , …, b _R ] ^T , where R is the number of blocks. A learnable vector b ₀ of the same dimension is added to represent the global features of the entire sample. The features after embedding the learnable information are denoted as B = [b ₀ , b ₁ , b ₂ , …, b _R ] ^T .

Transform feature B and reduce its dimension to different d-dimensional embedding subspaces:

E＝BW ^e

U＝ ^BWu

V＝BW ^v

Among them, ^We , ^Wu , ^Wv are different transformation matrices, E, U, V are the transformation features in different embedding subspaces, and the attention mechanism is used to obtain global features.

The global features are transformed back to _d1 dimensions through linear mapping LN(·), and the residual structure is combined with feature B to obtain

Feature B is first mapped to a high-dimensional space through a fully connected layer, and the dimension of the high-dimensional space is recorded as _d2 , and then mapped back to a low-dimensional space of _d1 to obtain the deep feature

With features

Use residual structure to combine and obtain deep global features

The learnable vector

Take out the deep global features as the corresponding samples

For the task

Using feature encoder

To the task

The support set and query set are feature encoded to obtain:

in,

The tasks are

Deep global features of the support and query sets, and

Furthermore, in step S32, the specific method of updating the category atom encoder and the category atom is:

Using the deep global features of the support set extracted from S31

Will

Transform and reduce the dimensions to different d-dimensional embedding subspaces:

in,

and

are different transformation matrices,

and

It is the transformation feature in different embedding subspaces, and the sample-level global feature is obtained by using the attention mechanism

The sample-level global features are transformed back to _d1 dimensions through linear mapping LN(·), and the residual structure and deep global features are combined to obtain

Through the fully connected layer, the features are first

Map to a high-dimensional space of _d2 , and then map back to a low-dimensional space of _d1 to obtain deep features

With features

The residual structure is used to combine and obtain sample-level deep global features

The sample-level deep global features are averaged to obtain the sample-level category atoms

Using the category atomic encoder to calculate the deep global features of the support set samples

All the class atoms in are represented as

in

According to the tasks obtained

The deep global features of the support set samples and the distances of atoms of different categories obtained to obtain the samples

The probability of being judged as category k is:

Where dist(·) is the distance function;

Design and minimize the category atom loss function based on probability:

Update the category atom encoder and record the updated model as

The updated category atom is

in,

Furthermore, in step S33, the specific method for updating the meta-learner is:

According to the tasks obtained

The deep global features of the query set samples and the distances of atoms of different categories are obtained to obtain the samples

The probability of being judged as category k is:

Design the meta-learner loss function based on probability:

Among them, margin is the set threshold, γ is the balance parameter, and the meta-learner is updated by minimizing the loss function to obtain the updated meta-learner

The beneficial effects of the present invention are as follows: for small sample target recognition scenarios, the present invention fully mines the global features of samples at the feature level, fully explores the robustness features of different samples of the same target at the sample level, and designs a meta-learner at the task level to effectively accumulate learning experience of different tasks. Through hierarchical learning at the feature level, sample level, and task level, the quality of feature information is improved, the negative impact of outlier samples is reduced, the autonomous learning ability of the model is cultivated, and the robustness of small sample target recognition technology is improved. The small sample radar target recognition method based on hierarchical meta-transfer proposed by the present invention is an intelligent radar target recognition method.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow chart of an algorithm of the present invention.

FIG. 2 is a comparison chart of the recognition accuracy of the background technology method and the method of the present invention.

Detailed ways

The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments:

As shown in Figure 1, the present invention designs a small sample radar target recognition method based on hierarchical meta-transfer, including feature level, sample level and task level. For each meta-training task, at the feature level, an attention mechanism is used to construct a feature encoder to extract more important features in a single sample; at the sample level, an attention mechanism is used to construct an atom encoder, and high-quality category atoms are generated as representative information of the corresponding category by integrating the information of different samples of the same type of target. At the task level, a meta-learner is constructed to acquire autonomous learning ability by accumulating learning experience of different meta-training tasks. When facing a new task to be tested, the trained meta-learner is further optimized based on a small number of labeled samples to generate high-quality category atoms for target recognition. The sample to be tested is compared with the category atom, and the category of the category atom with the highest similarity is selected as the predicted category of the test sample to complete the recognition of the test sample.

Example:

This example is a practical application of the method according to the present invention. In practical applications, synchronous initialization is performed when establishing the feature encoder and the category atom encoder so that they can be processed faster.

Step 1. Collect and preprocess original image samples in the source domain and target domain respectively, and preliminarily filter out redundant information of the target background to prepare for training the model.

The radar obtains the original images of each target at different pitch angles when it is static. At each fixed pitch angle, the target is observed at different azimuth angles. The acquired images are recorded as source domain and target domain according to the different pitch angles, and they are cut and preprocessed.

Step 2. Use samples to build training tasks

Each task includes a support set and a query set to train an object recognition model with autonomous learning capabilities.

Let a K classification task be

Construct all meta-training tasks and record them as

Where P is the total number of tasks.

The support set is composed of labeled samples extracted from the source domain in the form of K way N shot and recorded as

Among them, K way N shot means randomly extracting N labeled training samples from each category of K types of targets.

is the nth sample of the kth class target; the labeled samples are extracted in the target domain in the form of K way M shot to form a query set and recorded as

in,

is the mth sample of the kth class target. The samples in the support set and query set should be samples of the same class target in different domains. The corresponding class labels are

in,

Step 3: In order to accumulate learning experience from different tasks and cultivate the model's ability to learn autonomously, the meta-learner is trained and learned through the hierarchical meta-transfer model.

To update, the hierarchical meta-transfer model is composed of feature level, sample level and task level, specifically:

Step 31. Design feature encoder at feature level

For the training task obtained in step 2

The support set and query set are respectively extracted with features to explore the deep information of the sample for identification. Further, the specific steps of step 31 are:

Step 31-1. Design feature encoder at feature level

The feature encoder consists of a neural network module and an attention mechanism module. The neural network module has a strong feature extraction capability and can mine the deep features of the sample. The attention mechanism module is to enable the model to selectively focus on the important information in the sample and improve the efficiency of the model's information processing.

Feature extractor in

Initialize it:

Step 31-2. Use the neural network module and attention mechanism to extract the deep global features of the sample. The specific steps are as follows:

Step 31-2-1. Use the convolutional neural network module conv(·) to train the support set samples

Extract generalized features. For the sake of clarity, the support set sample representation symbol is abbreviated as S. The feature extraction process is as follows:

Step 31-2-2. Divide the sample generalization features obtained in step 31-2-1 into blocks and straighten them into vectors. The dimension of each vector is d ₁ . All vectors are recorded as [b ₁ , b ₂ , …, b _R ] ^T , where R is the number of blocks. In order to effectively integrate the information in the block features, add a learnable vector b ₀ of the same dimension to represent the global features of the entire sample. The feature that embeds the learnable information is recorded as B = [b ₀ , b ₁ , b ₂ , …, b _R ] ^T ,

Step 31-2-3. To further filter out redundant information, the feature B obtained in step 31-2-2 is transformed and reduced to different d-dimensional embedding subspaces:

E＝BW ^e (2)

U＝ ^BWu (3)

V＝BW ^v (4)

Among them, ^We , ^Wu , ^Wv are different transformation matrices, E, U, V are the transformation features in different embedding subspaces, and the attention mechanism is used to obtain global features:

Step 31-2-4. To alleviate the gradient disappearance, the global features obtained in step 31-2-3 are transformed back to d ₁ dimensions through linear mapping LN(·), and the residual structure is used to combine with the features obtained in step 31-2-2:

Step 31-3. Since the information in high-dimensional space is richer, a layer of fully connected network is used to map the features obtained in step 31-2 to the high-dimensional space. Note that the dimension of the high-dimensional space is _d2 , and then a layer of fully connected network is used to map it back to the original dimension _d1 . Each fully connected layer is processed with an activation function to learn and obtain more abstract deep features.

Enhance the expressiveness of information. To avoid the gradient vanishing problem, combine it with the features obtained in step 3-2 using a residual structure to obtain a deep global feature:

in,

The corresponding learnable vector

Take out the deep global features as the corresponding samples

Step 31-4. Task

The support set and query set are feature encoded:

in,

The tasks are

Deep global features of the support and query sets, and

Step 32. Design an attention-based category atom encoder at the sample level

And in the current training task

The updated category atoms are calculated to provide reliable representative information for target recognition. Further, the specific steps of step 4 are as follows:

Step 32-1. For the task

Designing category atom encoders at the sample level

And using the current meta-learner

Class Atom Encoder in

Initialize it:

Step 32-2. Use the category atom encoder obtained in step 32-1

The deep global feature calculation task for the support set samples obtained in step 31

The specific steps are as follows:

Step 32-2-1. To remove redundant information, explore the deep features of samples in different embedding subspaces and analyze the deep global features of the support set samples.

Transform and reduce the dimension to d dimension respectively:

in,

are different transformation matrices,

It is the transformation feature in different embedding subspaces, and the attention mechanism is used to explore the sample-level global features:

Step 32-2-2. To alleviate the gradient vanishing, the sample-level global features obtained in step 32-2-1 are transformed back to d ₁ dimensions through linear mapping LN(·), and the residual structure is combined with the support set deep global features obtained in step 31:

Step 32-2-3. Since the information in high-dimensional space is richer, a layer of fully connected network is used to map the features obtained in step 32-2-2 to a high-dimensional space of _d2 dimensions, and then a layer of fully connected network is used to map it back to the original dimension of _d1 dimensions. Each layer of fully connected layer is processed with an activation function to learn and obtain more abstract deep features.

Enhance the expressiveness of information. To avoid the gradient vanishing problem, combine it with the features obtained in step 32-2-2 using a residual structure to obtain sample-level deep global features:

in,

Step 32-2-4. Average the sample-level deep global features obtained in step 32-2-3 to obtain the category atoms after the sample-level attention mechanism exploration

Step 32-2-5. Calculation task

All the class atoms in are represented as

in

Corresponding to the processing flow of step 32-2-1 to step 32-2-4.

Step 32-3. Calculate the task obtained in step 31

The deep global features of the support set samples and the distances of different types of atoms obtained in step 32-2 further obtain the samples

The probability of being judged as category k is:

Here, dist(·) is the distance function.

Step 32-4. Design and minimize the category atom loss function according to probability to update the category atom encoder and category atoms. The specific steps are as follows:

Step 32-4-1. Design the following loss function so that the sample

The probability of being judged as category k is as large as possible to obtain a model with recognition ability. Minimize the loss function and update the category atom encoder:

Step 32-4-2. Note that the updated model is

The updated category atom is

in,

Step 33. Accumulate the learning experience of the current training task at the task level and update the meta-learner to

In order to enable the meta-learner to have autonomous learning capabilities to cope with new target recognition tasks, further, the specific steps of step 33 are:

Step 33-1. Calculate the task obtained in step 31

The deep global features of the query set samples and the distances of atoms of different categories obtained in step 32 are further obtained.

The probability of being judged as category k is:

Here, dist(·) is the distance function.

Step 33-2. Design a meta-learner loss function based on probability, minimize the loss function to update the meta-learner, and obtain

The specific steps are:

Step 33-2-1. Design the meta-learner classification loss function based on the classification probability obtained in step 33-1:

Step 33-2-2. In order to improve the separability of samples and enhance the recognition performance of the model, the model training also uses contrast loss as the loss function, which is defined as follows:

in,

Among them, margin is the set threshold. This constraint can reduce the sample characteristics

With the corresponding class atom

The distance between them increases with the distance between atoms of other types.

The distance between them should be as large as possible.

Step 33-2-3. Combine the loss functions of step 33-2-1 and step 33-2-2 to obtain the total meta-learner loss function:

Among them, γ is a balance parameter. Minimize the meta-learner loss function to update the meta-learner and obtain

The updated meta-learner

Thus accumulating in the task

learning experience.

Step 4. Update i=i+1 and repeat step 3 until all training tasks are completed and multiple trainings are completed to obtain the meta-learner trained by all meta-training tasks.

Step 5. The labeled samples of the task to be tested are called the support set, and the unlabeled samples to be tested are called the query set.

To identify the sample to be tested, further, the specific steps of step 5 are:

Step 5-1. Process the task to be tested based on the learning experience accumulated on the training task, and initialize the task model to be tested according to step 31

And extract deep global features for support set and query set samples.

Step 5-2. Initialize the task model to be tested according to step 32

Calculate and update the category atoms using the support set;

Step 5-3. Use the distance function dist(·) to calculate the deep global features of the query set sample and the distances between atoms of different categories, select the label of the category atom with the closest distance as the predicted label of the sample to be tested, and obtain the recognition result.

Simulation Example

The implementation model is used to experiment with the MSTAR dataset for acquisition and recognition of moving and stationary targets. The sensor of this dataset uses a high-resolution spotlight synthetic aperture radar, adopts HH polarization mode, works in the X-band, and has a resolution of 0.3m×0.3m. Most of the data are SAR slice images of stationary vehicles, which contain a total of ten types of targets, namely BMP2, T72, BTR70, 2S1, BRDM2, BTR60, D7, T62, ZIL131, ZSU234 and T72. Seven types of targets are taken to form the meta-training task, and the remaining three types of targets are used to construct the task to be tested. The sample data observed at a pitch angle of 17° is used as the source domain sample, and the sample data observed at a pitch angle of 15° is used as the target domain sample. The specific number of samples in the experiment is shown in Table 1.

Table 1 Specific number of experimental data

To remove the influence of background clutter, the sample image size is cut into 64×64 at the center. This case uses a 3-classification task, that is, each meta-training task and the task to be tested contains three types of targets. For the meta-training task, 3 of the 7 types of targets are randomly selected to form the meta-training task. For small sample target recognition, samples are randomly extracted from the source domain in the form of 3way 5shot to form the support set of the task, that is, 5 samples are randomly extracted from each of the 3 types of targets in the source domain for this task; the query set is composed of samples randomly extracted from the target domain in the form of 3way 15shot, that is, 15 samples are randomly extracted from each of the 3 types of targets in this task. For the meta-training task, the samples in the support set and the query set are all labeled samples. In a similar way, samples of the target category to be tested are randomly extracted to form the task to be tested, where the support set comes from the source domain and is the labeled samples observed at a pitch angle of 17°, and the query set comes from the target domain and is the samples to be tested observed at a pitch angle of 15°. In addition, this case also simulates target domain samples under different noise environments. A certain percentage of pixels are randomly selected from the test samples of the query set in the test task, and the pixels are destroyed by replacing the intensity of their pixels with independent samples that obey the uniform distribution. The added random noise obeys the uniform distribution of [0, μ _max ], where μ _max is the maximum value of the pixels in the image. The selected pixel ratios are 0%, 5%, and 15%, respectively, representing the target domains under different noise environments, where 0% represents the test samples constructed from the 15° pitch angle observation samples in the original data set.

The present invention designs experiments in different noise environments for small sample target recognition to verify the superiority of the proposed algorithm, and compares the recognition results of the background technology method and the method of the present invention on the task to be tested. In the experiment, the neural network module of the feature encoder consists of four convolutional layers, and the maximum pooling operation is used after each convolutional layer to reduce the size of the model and improve the calculation speed. Table 2 shows the detailed parameters of each convolutional layer and pooling operation, including the size of the convolution kernel, the step size during convolution, the padding size, and the size of the pooling window. In addition, other parameters in the experiment are specifically set to: R = 3, d ₁ = 252, d = 64, d ₂ = 128, γ = 0.01, and margin = 200. 200 meta-training tasks are used for training, and the average recognition rate of 1000 tasks to be tested is used as a quantitative indicator of the algorithm performance. As the noise level in the target domain increases, the background technology methods all show a significant decline to varying degrees. Among them, the recognition accuracy of background technology method 1 in 0% and 15% noise environments is 77.43% and 71.66% respectively, and the recognition accuracy of the background technology method is 71.67% and 68.1%, while the method of the present invention can still maintain a high recognition rate, with recognition accuracy rates of 83.86%, 82.24%, and 81.92% in 0%, 5%, and 15% noise environments, respectively, which has obvious advantages. In summary, the experimental results prove that the present invention effectively explores the deep global features of samples in small sample target recognition scenarios, cultivates the autonomous learning ability of the model, establishes a more stable meta-learning model, and improves target recognition performance.

Table 2 Experimental parameter settings

卷积层Convolutional Layer	卷积核尺寸Convolution kernel size	步长Step Length	填充尺寸Filling size	池化窗尺寸Pooling window size
第一层level one	5×55×5	11	00	2×22×2
第二层Second floor	3×33×3	11	00	2×22×2
第三层the third floor	3×33×3	11	11	2×22×2
第四层Fourth floor	3×33×3	11	11	2×22×2

Claims

A small sample radar target recognition method based on hierarchical element migration is characterized by comprising the following steps:

S1. Obtaining original images of each target in the source domain and the target domain when the target is static through radar, and cutting the images obtained by observing the target at different azimuth angles to obtain samples;

S2. Use samples to build training tasks
Where P is the total number of tasks,
It includes support set and query set, where the support set is composed of labeled samples extracted from the source domain, and the query set is composed of labeled samples extracted from the target domain;

S3. Training and learning through hierarchical meta-transfer models
Update, specifically:

S31. Constructing a feature encoder based on the attention mechanism at the feature level
Initialize the feature encoder using a meta-learner
After that, extract
Deep global features of the support set and query set;

S32. Constructing a category atom encoder based on attention mechanism at sample level
Initialize the category atom encoder using a meta-learner
Afterwards, based on the obtained
Deep global feature calculation of support set samples
The category atoms of the support set are obtained, and the probability of the corresponding samples belonging to different categories is obtained according to the distance between the support set samples and the different category atoms. Then, the category atom loss function is designed and minimized according to the probability to update the category atom encoder and category atoms.

S33. Accumulate the learning experience of the current training task at the task level and update the meta-learner:

according to
The deep global features of the query set samples and the distances between atoms of different categories are used to obtain the probability that the corresponding samples belong to different categories. The meta-learner loss function is designed based on the probability, and the meta-learner is updated by minimizing the loss function to obtain the updated meta-learner.

S4. Complete all training tasks by repeating step S3 to obtain the meta-learner trained by all meta-training tasks. The trained meta-learner is recorded as

S5: The labeled samples of the task to be tested are the support set, and the unlabeled samples to be tested are the query set; the meta-learner obtained in S4 is used for initialization
A feature encoder for target recognition and a category atom encoder are obtained, and the feature encoder for target recognition is used to extract deep global features for the support set and query set samples. The category atom encoder for target recognition is used to calculate and update the category atoms based on the deep global features of the support set, and the distance function dist(·) is used to calculate the distance between the deep global features of the sample to be tested in the query set and the atoms of different categories, and the label of the category atom with the closest distance is selected as the predicted label of the sample to be tested to obtain the recognition result.
The small sample radar target recognition method based on hierarchical meta-transfer according to claim 1 is characterized in that in step S2, the support set is formed by extracting labeled samples in the source domain in the form of K way N shot, defined as
K way N shot means randomly extracting N labeled training samples from each category of K target.
is the nth sample of the kth class target; the query set is composed of labeled samples extracted in the target domain in the form of K way N shot, defined as
in,
is the mth sample of the kth class target; the samples in the support set and query set are samples of the same class target in different domains, and the corresponding class labels are defined as
in,
The small sample radar target recognition method based on hierarchical element migration according to claim 2 is characterized in that in step S31, the feature encoder
It includes a neural network module and an attention mechanism module. The specific method of extracting deep global features is as follows:

Extract generalized features from samples through neural network modules;

The generalized features are divided into blocks and straightened into vectors. The dimension of each vector is d 1 , denoted as [b 1 , b 2 , …, b R ] T , where R is the number of blocks. A learnable vector b 0 of the same dimension is added to represent the global features of the entire sample. The features after embedding the learnable information are denoted as

Transform feature B and reduce its dimension to different d-dimensional embedding subspaces:

E＝BW e

U＝ BWu

V＝BW v

Among them, We , Wu , Wv are different transformation matrices, E, U, V are the transformation features in different embedding subspaces, and the attention mechanism is used to obtain global features.

The global features are transformed back to d1 dimensions through linear mapping LN(·), and the residual structure is combined with feature B to obtain

Feature B is first mapped to a high-dimensional space through a fully connected layer, and the dimension of the high-dimensional space is recorded as d2 , and then mapped back to a low-dimensional space of d1 to obtain the deep feature
With features
Use residual structure to combine and obtain deep global features

The learnable vector
Take out the deep global features as the corresponding samples

For the task
Using feature encoder
To the task
The support set and query set are feature encoded to obtain:
in,

The tasks are
Deep global features of the support and query sets, and
The small sample radar target recognition method based on hierarchical element migration according to claim 3 is characterized in that in step S32, the specific method of updating the category atom encoder and the category atom is:

Using the deep global features of the support set extracted from S31
Will
Transform and reduce the dimensions to different d-dimensional embedding subspaces:

in,
and
are different transformation matrices,
and
It is the transformation feature in different embedding subspaces, and the sample-level global feature is obtained by using the attention mechanism

The sample-level global features are transformed back to d1 dimensions through linear mapping LN(·), and the residual structure and deep global features are combined to obtain

Through the fully connected layer, the features are first
Map to a high-dimensional space of d2 , and then map back to a low-dimensional space of d1 to obtain deep features
With features
The residual structure is used to combine and obtain sample-level deep global features

The sample-level deep global features are averaged to obtain the sample-level category atoms

Using the category atomic encoder to calculate the deep global features of the support set samples
All the class atoms in are represented as
in

According to the tasks obtained
The deep global features of the support set samples and the distances of atoms of different categories obtained to obtain the samples
The probability of being judged as category k is:

Where dist(·) is the distance function;

Design and minimize the category atom loss function based on probability:

Update the category atom encoder and record the updated model as
The updated category atom is
in,
The small sample radar target recognition method based on hierarchical meta-transfer according to claim 4 is characterized in that in step S33, the specific method of updating the meta-learner is:

According to the tasks obtained
The deep global features of the query set samples and the distances of atoms of different categories are obtained to obtain the samples
The probability of being judged as category k is:

Design the meta-learner loss function based on probability:

Among them, margin is the set threshold, γ is the balance parameter, and the meta-learner is updated by minimizing the loss function to obtain the updated meta-learner