Small Sample SAR Interference Identification Method Based on Two-Stage Dual Calibration Network

Abstract

The present invention provides a small sample SAR interference recognition method based on a two-stage dual calibration network, which relates to the field of radar target recognition technology, including obtaining interference time-frequency images to be identified and time-frequency images carrying labels in a database from a radar system, and inputting them into a trained two-stage dual calibration network to obtain a recognition result, wherein the recognition result includes the category of the interference object contained in the interference time-frequency image; since the trained two-stage dual calibration network of the present invention uses the basic prototype features of the pre-training stage to guide the prototype features in the meta-learning stage, the network can learn calibration to reduce the serious distribution offset under low SNR. And the present invention jointly calibrates support and query instances by calculating cross-attention scores, thereby realizing dual calibration from support to query and from query to support. The present invention improves the dual cross-attention mechanism by decoupling the embedded spatial and channel dimensions to further focus on reliable features and eliminate unreliable features.

Description

Technical Field

[0001] The invention belongs to the technical field of SAR image interference recognition, and in particular relates to a small sample SAR interference recognition method based on a two-stage dual calibration network. Background Art

[0002] SAR (Synthetic Aperture Radar), as a broadband microwave imaging remote sensing sensor, is inevitably subject to unintentional or intentional interference, which seriously restricts its imaging performance. Therefore, effective interference suppression is the guarantee for the normal operation of the SAR system, and interference identification is the prerequisite for solving this problem. However, the traditional SAR interference identification algorithm based on deep learning has poor data efficiency and requires a large number of labeled training data samples. However, preparing a large number of labeled samples for training in interference identification is usually laborious, expensive, or even unrealistic. Therefore, small sample learning algorithms have gradually attracted the attention of researchers.

[0003] Existing FSL (Few-Shot Learning) algorithms can be roughly divided into two categories: optimization-based methods and metric-based methods. Specifically, optimization-based methods aim to efficiently adapt model parameters to new tasks through a small number of gradient descent steps. The idea of ​​metric-based methods is to learn good embeddings and appropriate comparison metrics.

[0004] There are two main types of existing attention-based FSL algorithm research, namely single-set attention and cross-set attention. FSL methods based on single-set attention aim to calibrate the features of support or query samples separately. However, due to ignoring the inherent cross-set correlation between support and query instances, the improvement effect of these methods is minimal. In contrast, FSL methods based on cross-set attention aim to calibrate the features between support samples and query samples.

[0005] Traditional deep learning-based SAR interference recognition algorithms have poor data efficiency and require a large number of labeled training data samples. However, preparing a large number of labeled samples for training in interference recognition is usually laborious, expensive, or even unrealistic. Existing SAR interference recognition algorithms based on "small sample learning" have invested a lot of effort to carefully design feature extractors or learning strategies. However, they all ignore the ability of the model to dynamically adapt to new classes at low SNR (Signal-to-noise Ratio) during the meta-test phase, i.e., learning calibration, which is crucial for severe distribution shifts between new classes and base classes, which can lead to spurious correlations between reliable and unreliable features. Specifically, such methods can only rely on prior knowledge learned from base classes to complete the prediction of new classes in the meta-stage. However, in the presence of severe distribution shifts, this feature is fatal to the SAR interference recognition task. Especially at low SNR, stationary prior knowledge causes the model to fail to fully extract features of interest or to mistakenly focus on meaningless ambient noise, both of which lead to incorrect predictions. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a small sample SAR interference identification method based on a two-stage dual calibration network. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] The present invention provides a small sample SAR interference identification method based on a two-stage dual calibration network, comprising:

[0008] S100, acquiring a time-frequency image of interference to be identified and a time-frequency image carrying a label in a database from a radar system;

[0009] S200, inputting the interference time-frequency image and the time-frequency image carrying the label into the trained two-stage dual calibration network to obtain a recognition result, wherein the recognition result includes the category of the interference object contained in the interference time-frequency image; wherein the trained two-stage dual calibration network is obtained by training a predefined two-stage dual calibration network in a pre-training stage and a meta-learning stage.

[0010] Beneficial effects:

[0011] 1. We propose a novel two-stage training framework for few-shot SAR interference identification, called DAN, which provides base prototypes to guide meta-prototypes during meta-training and enables the model to learn calibration to mitigate severe distribution shift at low SNR.

[0012] 2. This paper proposes a novel dual calibration module based on a dual cross attention mechanism, which jointly calibrates support and query instances by calculating cross attention scores, thereby achieving dual calibration from support to query and query to support at the instance level. It is worth noting that the dual cross attention mechanism is improved by decoupling the spatial and channel dimensions of the embedding to further focus on reliable features and eliminate unreliable features.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of a small sample SAR interference identification method based on a two-stage dual calibration network provided by the present invention;

[0015] Figure 2 This is a structural diagram of the two-stage dual calibration network proposed in the present invention;

[0016] Figure 3 It is a structural diagram of the dual calibration module of the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0018] like Figure 1 As shown, the present invention provides a small sample SAR interference identification method based on a two-stage dual calibration network, comprising:

[0019] S100, acquiring a time-frequency image of interference to be identified and a time-frequency image carrying a label in a database from a radar system;

[0020] S200, inputting the interference time-frequency image and the time-frequency image carrying the label into the trained two-stage dual calibration network to obtain a recognition result, wherein the recognition result includes the category of the interference object contained in the interference time-frequency image; wherein the trained two-stage dual calibration network is obtained by training a predefined two-stage dual calibration network in a pre-training stage and a meta-learning stage.

[0021] The predefined two-stage dual calibration network of the present invention comprises an input layer, a feature extractor, a dual calibration module and a prototype classifier connected in sequence;

[0022] Among them, the input layer is used to input time-frequency images, the feature extractor is used to extract sample features of the input time-frequency images, the dual calibration module is used to collaboratively calibrate the sample features in the spatial dimension and the channel dimension to obtain calibration features, and the prototype classifier is used to classify according to the calibration features to obtain the category of the object contained in the input time-frequency image.

[0023] In a specific embodiment of the present invention, the feature extractor is obtained by pre-construction, and the construction process includes:

[0024] a1, build and train a deep CNN classifier based on supervised learning with the entire base class;

[0025] b1, remove the fully connected layer of the deep CNN classifier and use the classifier composed of the remaining parts as the feature extractor.

[0026] The present invention constructs and trains a deep CNN classifier with the entire base class based on supervised learning. Then, the last fully connected layer is discarded and the remaining part is composed of a feature extractor. The feature extractor is well initialized as a meta-learning stage. The feature extractor includes channel parameters and weight parameters, which are expressed as:

[0027] (1);

[0028] In the formula, represents the parameters of the feature extractor, represents the parameters of the fully connected layer of the feature extractor, is the cross entropy loss, represents a fully connected layer with weight parameters, represents the feature extractor, represents the auxiliary dataset, Include Class sample images.

[0029] Among them, the first support set and the first query set All included Class sample images, first support set Each class of sample images in The first query set Each class of sample images has The first support set The sample images in all carry labels. The first query set None of the sample images in carry labels.

[0030] Combination Figure 2 and Figure 3 In a specific embodiment of the present invention, the process of training a predefined two-stage calibration network in the pre-training stage includes:

[0031] a2, get auxiliary data set , and divide it into the first support set and the first query set ;

[0032] It is worth noting that the present invention models the small sample interference recognition algorithm with the traditional meta-learning paradigm. The main purpose of the general meta-learning algorithm is to learn transferable knowledge across tasks, that is, the knowledge learned in the training task can be used to solve new tasks with insufficient known information.

[0033] The present invention selects an auxiliary data set containing a large number of labeled sample images , which is divided into the first support set and the first query set , used to train the dual calibration module in the meta-training phase, where for The total number of sample images included, It is the auxiliary data set sample images, , yes The corresponding label, for The corresponding label set, and The goal of the meta-learning phase is to Training a model that can be effectively generalized to new datasets Model, for The corresponding label set.

[0034] b2, from the first support set and the first query set A sample image is extracted from each of the two images to form a first support query pair, and the first support query pair is input into the feature extractor to generate the first Base prototype features of a class and Basic query features .

[0035] Since the base prototype features are more robust than the support features, the base prototype features and the base query features are applied to perform a double calibration operation. Specifically, for each meta-task, A randomly selected support-query pair is fed into the feature extractor , to generate basic prototype features and basic query features, which can be written in the following form:

[0036] (2);

[0037] (3);

[0038] in, Indicates The base prototype characteristics of the class, The first support set Middle sample images, express The corresponding label, Indicates Basic query features, Represents the first query set Middle Sample images.

[0039] Continue to refer Figure 2 and Figure 3 In a specific embodiment of the present invention, the process of training a predefined two-stage dual calibration network in the meta-learning stage includes:

[0040] a3, get a new data set , and divide it into the second support set and the second queryset ;

[0041] Among them, the second support set and the second queryset All included Class sample images, each class of sample images in the second support set has The second query set Each class of sample images has The second support set The sample images in the second query set all carry labels. None of the sample images in carry labels;

[0042] The present invention first obtains a new data set , and divide it into two sub-datasets, namely the second support set and the second query set. The second support set of sample images with labels is expressed as And the second query set consists of unlabeled sample images, expressed as ,in express The number of samples in express Middle sample images, for The corresponding label, The second support set , the second query set and The number of categories included, represents the number of sample images of each category in the second support set, Represents the second query set The total number of sample images included, and .in, , Represent the second support set and the second queryset Selected sample images, , yes , The corresponding label.

[0043] b3, from the second support set and the second queryset Select one sample image from each to form a sample image pair;

[0044] c3, using feature extractor Extract the sample features of the sample image pairs respectively and get the corresponding Prototype characteristics of a class and The query features ;

[0045] d3, the prototype feature and query features Input to the dual calibration module included in the two-stage dual calibration network, so that the dual calibration module can calibrate the prototype features and query features Co-calibrate in the spatial dimension and channel dimension in turn to obtain the dual-calibration prototype feature and dual calibration query features ;

[0046] The dual calibration module includes a spatial dual calibration module and a channel dual calibration module. The spatial dual calibration module adopts a single-head spatial cross-attention mechanism or a multi-head spatial cross-attention mechanism; the channel dual calibration module adopts a multi-head channel cross-attention mechanism.

[0047] e3, the dual calibration prototype features and dual calibration query features Input different classifiers, obtain the corresponding classification probabilities, and calculate the total loss function based on the classification probabilities;

[0048] f3, judge whether the training meets the cutoff condition according to the total loss function. If the cutoff condition is not met, adjust the internal parameters of the dual calibration module and return to b3; if the cutoff condition is met, obtain the trained two-stage dual calibration network.

[0049] In order to enable the model to fully learn calibration in the meta-learning stage, the present invention constructs a dual calibration module Specifically, the dual calibration module consists of uncoupled spatial and channel dual calibration modules based on a dual cross-attention mechanism, which utilizes the dual cross-attention mechanism to eliminate the spurious correlations between support samples and query samples at the instance level, that is, to achieve dual calibration from support to query and from query to support.

[0050] In the meta-learning phase, the feature extractor is trained with a smaller learning rate. Fine-tune and train the dual calibration module from scratch with a larger learning rate According to the episodic training method, Many kind Specifically, in each training session, we first Tag library Random selection categories, and then randomly select The second support set , and randomly select The second query set , to construct the meta-task Then, by minimizing the three loss functions and and The respective learning rates are used to jointly train the feature extractors and dual calibration modules .

[0051] In a specific embodiment of the present invention, d3 comprises:

[0052] d31, the prototype feature and query features Input into the spatial dual calibration module to convert them into Query, Key and Value matrices respectively;

[0053] Under low SNR conditions, similar reliable features will be correlated with each other regardless of how the distribution changes, while other unreliable features will not. The present invention utilizes the self-attention mechanism to calculate the double cross attention scores between the prototype features and the query features.

[0054] In form, given Class query characteristics ,in , , , and Respectively represent the channel height, channel width and channel depth. The present invention firstly transforms the prototype feature and query features Reshape to ,in , and then transform the prototype features into and query features Convert to Query, Key and Value matrices respectively. Most importantly, The Query vector is obtained from the query features. and query features The converted Query, Key and Value matrices are expressed as:

[0055] (4);

[0056] In the formula, and Represents the prototype characteristics and query features Query, Key, Value matrix, Is a prototype feature and query features The shared embedding matrix ensures that they are embedded in the same feature space. It is worth noting that each embedding matrix is ​​based on It is implemented in the form of convolution. It is a prototype feature The characteristics of Represents query features The matrix of represents the spatial dimension of the matrix, express The set of real numbers.

[0057] d32, using prototype features and query features Query, Key and Value matrices, corresponding to the calculation of their respective spatial cross attention and ;

[0058] The matrix and are sent to the spatial dual calibration module to calculate the spatial cross attention and , the formula is as follows:

[0059] (5);

[0060] In the formula, represents matrix multiplication, is a row-wise softmax function to normalize the attention scores. Represents the transpose of the matrix. It is worth noting that the cross use of the Query matrix constitutes an interactive Query-Key pair, which makes the spatial cross attention and There is an interactive perception receptive field between prototype features and query features, express The set of real numbers.

[0061] d33, using spatial cross attention and , Co-calibration prototype features and query features , obtain the prototype features of spatial calibration and the query features of spatial calibration;

[0062] Since the spatial dual calibration module adopts a single-head spatial criss-cross attention mechanism or a multi-head spatial criss-cross attention mechanism, there are outputs in two cases.

[0063] Among them, the single-head spatial cross attention mechanism can be modeled as:

[0064] (6);

[0065] In the formula, represents a single-head spatial cross attention mechanism, Indicates matrix multiplication, superscript Indicates the transposition of the matrix; express The set of real numbers.

[0066] Therefore, the prototype features are jointly calibrated in the spatial dimension under the single-head spatial cross attention mechanism. and query features The formula is:

[0067] (7);

[0068] In the formula, and It is an intermediate parameter, which is only used for the convenience of presentation and does not represent any meaning.

[0069] Finally, the output of the spatial dual calibration module under the single-head spatial cross attention mechanism is:

[0070] (8);

[0071] In the formula, and They represent the prototype features and query features of spatial calibration under the multi-head spatial cross attention mechanism, Represents matrix addition.

[0072] It is worth noting that the prototype feature is added to the spatial dual calibration module and query features As residual connections to achieve stable training convergence.

[0073] In order to jointly learn the global information from different subspaces of the same input, the present invention can also adopt a multi-head spatial cross attention mechanism, which can be modeled as:

[0074] (9);

[0075] in, represents the spatial multi-head attention mechanism, Representing prototype features Query, Key, Value matrix, Represents query features Query, Key, Value matrix, Indicates a connection operation. Indicates that the input object is From 1 to The spatial attention head of Indicates that the input object is From 1 to The spatial attention head of represents the spatial linear projection matrix, , represents the number of spatial attention heads, represents the spatial attention mechanism, express and Shared The embedding matrix of the spatial attention head, , , Indicates the sequence number of the spatial attention head.

[0076] Therefore, the output of the spatial dual calibration module under the multi-head spatial cross attention mechanism is expressed as:

[0077] (10);

[0078] Here, and They respectively represent the prototype features and query features of spatial calibration under the multi-head spatial cross-attention mechanism.

[0079] d34, inputting the spatially calibrated prototype features and the spatially calibrated query features into the channel dual calibration module to convert them into Query, Key, and Value matrices, respectively;

[0080] Intuitively, each channel feature of higher layers contains more discriminative and category-specific representations. However, while emphasizing the coupled spatial correlation between support and query instances to calibrate reliable features is effective, their coupled architecture fails to learn the channel correlation between support and query instances, which is crucial for few-shot image classification. Therefore, we propose a channel dual calibration module based on a dual cross-attention mechanism to improve the representation of reliable features and remove spurious correlations between support and query instances in the channel dimension.

[0081] d35, using the Query, Key and Value matrices of the spatially calibrated prototype features and query features, the respective channel cross attentions are calculated;

[0082] The processing process of the channel dual calibration module is exactly the same as that of the spatial dual calibration module. The channel dual calibration module of the present invention adopts a multi-channel cross attention mechanism. Therefore, the multi-channel cross attention can be calculated using formulas (4)-(5), (9)-(10), which is expressed as:

[0083] (11);

[0084] In the formula, represents the channel multi-head attention mechanism, Indicates that the input object is From 1 to The channel attention head of Indicates that the input object is From 1 to The channel attention head of represents the channel linear projection matrix, , represents the channel attention mechanism, express and Shared The embedding matrix of the channel attention head, , , Indicates the number of channel space headers, Indicates the sequence number of the channel space header.

[0085] d36, using channel cross attention, jointly calibrates the spatially calibrated prototype features and the spatially calibrated query features to obtain the dual-calibrated prototype features and dual calibration query features .

[0086] The output result of the channel dual calibration module of the present invention under the multi-head channel cross attention mechanism is expressed as:

[0087] (12);

[0088] In the formula, and Respectively and .

[0089] The present invention will and Reshape into dual calibration prototype features respectively and query features , as the final output of the dual calibration module.

[0090] Then, DAM is used to co-calibrate prototype features and query features at the instance level, which is expressed as:

[0091] (13);

[0092] In the formula, Represents a dual calibration module, Represents the prototype feature and query features Perform calibration.

[0093] In a specific embodiment of the present invention, e3 includes:

[0094] e31, will double calibration prototype features and dual calibration query features Input a prototype classifier so that the prototype classifier outputs a first classification probability, and the first classification probability is used to calculate the meta-loss;

[0095] The prototype classifier computes the double calibration query features With dual calibration prototype features The normalized similarity between them is used to output the classification probability, that is, the first classification probability is expressed as:

[0096] (14);

[0097] In the formula, represents the first classification probability, , represents the parameters of the meta-learner, Represents the second query set Middle The labels of sample images, is the first temperature scaling parameter, In order to express the sum conveniently, it means taking from 1 to , represents the cosine distance operator, Indicates Quasi-dual calibration prototype features.

[0098] The meta-loss calculated based on the first classification probability can be expressed as:

[0099] (15);

[0100] In the formula, for kind A set of meta-tasks with sample images.

[0101] e32, will double the calibration prototype features and dual calibration query features Input into the calibration classifier to obtain the second classification probability, and use the second classification probability to calculate the calibration loss;

[0102] Randomly selecting sample images in the second support set may lead to significant deviations between prototype features and base prototype features, especially in low SNR conditions. In addition, the prototype features must not only calibrate reliable features after DAM, but also ensure that these features are semantically related to the class label, that is, to ensure the identifiability of the meta-prototype. To this end, the present invention proposes a novel calibration loss. Specifically, the calibration classifier calculates the dual calibration prototype features. and basic prototype features The normalized similarity between them is used to output the second classification probability, expressed as:

[0103] (16);

[0104] In the formula, , represents the second classification probability, The second support set Middle The labels of the sample images of the class, is the second temperature scaling parameter, represents the cosine distance operator, The first support set Middle Class sample images, In order to express the convenience of summation, from 1 to .

[0105] The calibration loss can be expressed as:

[0106] (17);

[0107] In the formula, represents the prototype calibration task set.

[0108] e33, will double calibration prototype features and dual calibration query features Input into the global classifier to obtain the third classification probability, and use the third classification probability to calculate the global loss;

[0109] The present invention constructs a global classifier to strengthen the supervision of the meta-training stage. Specifically, the global classifier uses the parameters The fully connected layer , and then use the softmax operator to classify each query sample into the auxiliary dataset The classification probability of each dual calibration query feature is calculated as follows:

[0110] (18);

[0111] In the formula, , represents global average pooling, , Represents the fully connected layer in the local classifier Parameters;

[0112] The calculated global loss can be expressed as:

[0113] (19);

[0114] d34, calculates the total loss function based on meta-loss, calibration loss and global loss.

[0115] The present invention constructs a total loss as the objective function to optimize the dual calibration training phase:

[0116] (20);

[0117] In the formula, and is a positive balancing scalar that weighs the importance of different losses.

[0118] In a specific implementation of the present invention, before obtaining a trained two-stage dual calibration network, a small sample SAR interference identification method based on a two-stage dual calibration network further includes:

[0119] The internal parameters of the two-stage dual calibration network obtained in each round of training are selected, and the two-stage dual calibration network with the highest accuracy is retained as the trained two-stage dual calibration network.

[0120] The effect of the method of the embodiment of the present invention is verified by simulation experiments below.

[0121] First, the present invention obtains the interference identification dataset JamSet and the benchmark dataset miniImageNet. JamSet is a dataset for small sample SAR interference identification, which is divided into 32 categories for training, 8 categories for verification, and 10 categories for testing. miniImageNet is a benchmark dataset, each of its 100 classes contains 600 images, and the dataset is divided into 64 categories, 16 categories, and 20 categories for training, verification, and testing, respectively. For the JamSet dataset, its input shape is scaled to 1×128×128, and for the miniImageNet dataset, its input shape is scaled to 3×84×84.

[0122] The present invention adopts ResNet12 and ResNet18 as feature extractors. In the pre-training stage, the feature extractor is trained 200 times on the entire base class through Adam with an initial learning rate of 0.01 and a weight decay of 0.0005. In the meta-learning stage, the pre-trained ResNet12 / ResNet18 is used as a feature extractor, fine-tuned with a smaller learning rate, and the dual calibration module is trained from scratch with a larger learning rate. In each meta-training, 2000 meta-tasks are randomly selected. In addition, the present invention uses Adam with a weight decay of 0.0005 and a cosine learning rate decay to optimize the model. For the JamSet dataset, the initial learning rate of the dual calibration module is set to 0.1, and for the miniImageNet dataset, its initial learning rate is set to 0.05. The initial learning rate is set to 0.001 to fine-tune the backbone of all models. For all datasets, spatial and channel multi-head and Set to 8 and 4 respectively; the first temperature parameter and the second temperature parameter are set to 0.2 and 0.4 respectively, balancing the scalar and Set to 0.4 and 0.2 respectively.

[0123] The specific simulation experiment content and results are shown in Table 1 and Table 2 below:

[0124] Table 1 Performance comparison (%) of the proposed method and the most advanced small sample recognition method in single sample tasks

[0125]

[0126] Table 2 Comparison of the performance (%) of the proposed method and the most advanced small sample recognition method in 5 sample tasks

[0127]

[0128] Tables 1 and 2 show the comparison results between the proposed method and the state-of-the-art methods on the JamSet dataset. The proposed method based on ResNet12 and ResNet18 achieves more advanced performance on all sub-datasets of JamSet. In particular, the proposed method based on ResNet12 significantly outperforms the state-of-the-art methods on multiple tasks, such as JamSet_0dB: 4.70% (1 sample) and 3.66% (5 samples); JamSet_10dB: 4.02% (1 sample) and 3.35% (5 samples); JamSet_20dB: 3.84% (1 sample) and 3.29% (5 samples); JamSet_30dB: 3.62% (1 sample) and 3.32% (5 samples). Obviously, the performance improvement at low SNR (i.e., JamSet_0dB) is more significant than that at high SNR. This is because the problem of spurious correlation between reliable features and other unreliable features is more prominent at low SNR than at high SNR. Most importantly, the proposed method gives the model the ability to learn calibration, which enables the model to focus on reliable features that are semantically related to the new class label and abandon other unreliable features. In addition, the performance improvement of the 1-sample task is more obvious than that of the 5-sample task because the problem of prototype estimation error is more serious in the 1-sample task.

[0129] The present invention also conducts several comparative experiments on the JamSet_0dB and miniImageNet meta-test datasets. The results show that the model without DAM has difficulty focusing on the target object occluded by the ambient noise, while the present invention can solve this problem and generate better feature maps. This is because the present invention enables the model to co-calibrate the support and query instances by calculating the double cross attention score, thereby significantly emphasizing reliable features and abandoning unreliable features.

[0130] It is worth noting that the terms "first" and "second" in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0131] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0132] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A small sample SAR interference identification method based on a two-stage dual calibration network, characterized in that: include: S100, acquiring a time-frequency image of interference to be identified and a time-frequency image carrying a label in a database from a radar system; S200, inputting the interference time-frequency image and the time-frequency image carrying the label into a trained two-stage dual calibration network to obtain a recognition result, wherein the recognition result includes the category of the interference object contained in the interference time-frequency image; wherein the trained two-stage dual calibration network is obtained by training a predefined two-stage dual calibration network in a pre-training stage and a meta-learning stage; The predefined two-stage dual calibration network includes an input layer, a feature extractor, a dual calibration module and a prototype classifier connected in sequence; The input layer is used to input a time-frequency image, the feature extractor is used to extract sample features of the input time-frequency image, the dual calibration module is used to collaboratively calibrate the sample features in the spatial dimension and the channel dimension to obtain calibration features, and the prototype classifier is used to classify according to the calibration features to obtain the category of the object contained in the input time-frequency image; The process of training the predefined two-stage dual calibration network in the pre-training stage includes: a2, get auxiliary data set , and divide it into the first support set and the first query set ; Among them, the first support set and the first query set All included Class sample images, the first support set Each class of sample images in The first query set Each class of sample images has The first support set The sample images in all carry labels. None of the sample images in carry labels; b2, from the first support set and the first query set A sample image is extracted from each of the two images to form a first support query pair, and the first support query pair is input into the feature extractor to generate a first Base prototype features of a class and Basic query characteristics of the class ; The process of training a predefined two-stage dual calibration network in the meta-learning phase includes: a3, get a new data set , and divide it into the second support set and the second queryset ; Among them, the second support set and the second queryset All included Class sample images, each class of sample images in the second support set has The second query set Each class of sample images has The second support set The sample images in the second query set all carry labels. None of the sample images in carry labels; b3, from the second support set and the second queryset Select one sample image from each to form a sample image pair; c3, using the feature extractor to extract the sample features of the sample image pair respectively, and obtaining the corresponding Prototype characteristics of a class and The query features ; d3, the prototype feature and query features The dual calibration module included in the two-stage dual calibration network is input to enable the dual calibration module to calibrate the prototype feature and query features Co-calibrate in the spatial dimension and channel dimension in turn to obtain the dual-calibration prototype feature and dual calibration query features ; e3, the dual calibration prototype features and dual calibration query features Input different classifiers to obtain corresponding classification probabilities, and calculate the total loss function based on the classification probabilities; f3, judging whether the training meets the cut-off condition according to the total loss function, if the cut-off condition is not met, adjusting the internal parameters of the dual calibration module, and returning to b3; if the cut-off condition is met, obtaining a trained two-stage dual calibration network.

2. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 1 is characterized in that: The feature extractor is pre-built, and the building process includes: a1, build and train a deep CNN classifier based on supervised learning with the entire base class; b1, remove the fully connected layer of the deep CNN classifier, and use the classifier composed of the remaining parts as a feature extractor.

3. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 1 is characterized in that: The dual calibration module includes a spatial dual calibration module and a channel dual calibration module.

4. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 3 is characterized in that: The spatial dual calibration module adopts a single-head spatial cross-attention mechanism or a multi-head spatial cross-attention mechanism; the channel dual calibration module adopts a multi-head channel cross-attention mechanism.

5. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 3 is characterized in that: d3 includes: d31, the prototype feature and query features Input into the spatial dual calibration module to convert them into Query, Key and Value matrices respectively; d32, using prototype features and query features Query, Key and Value matrices, corresponding to the calculation of their respective spatial cross attention and ; d33, using the spatial cross attention and , co-calibrate the prototype features and query features , obtain the prototype features of spatial calibration and the query features of spatial calibration; d34, inputting the spatially calibrated prototype features and the spatially calibrated query features into the channel dual calibration module to convert them into Query, Key and Value matrices respectively; d35, using the Query, Key and Value matrices of the spatially calibrated prototype features and query features, the respective channel cross attentions are calculated; d36, using the channel cross attention, collaboratively calibrate the spatially calibrated prototype features and the spatially calibrated query features to obtain a dual-calibrated prototype feature and dual calibration query features .

6. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 5 is characterized in that: e3 includes: e31, the dual calibration prototype features and the dual calibration query feature Input a prototype classifier so that the prototype classifier outputs a first classification probability, and calculates a meta-loss using the first classification probability; e32, the dual calibration prototype features and dual calibration query features Input into the calibration classifier to obtain a second classification probability, and use the second classification probability to calculate the calibration loss; e33, the dual calibration prototype features and dual calibration query features Input into the global classifier to obtain a third classification probability, and use the third classification probability to calculate the global loss; e34, calculate the total loss function according to the meta-loss, the calibration loss and the global loss.

7. The small sample SAR interference identification method based on a two-stage dual calibration network according to claim 1 is characterized in that: Before obtaining the trained two-stage dual calibration network, the small sample SAR interference identification method based on the two-stage dual calibration network further includes: The internal parameters of the two-stage dual calibration network obtained in each round of training are selected, and the two-stage dual calibration network with the highest accuracy is retained as the trained two-stage dual calibration network.

Images