Method and apparatus for few-shot sar target recognition, and medium

Abstract

This application relates to a method, apparatus, device, and medium for few-sample SAR target recognition. The method includes: dividing SAR samples into support and query sets based on a meta-learning framework; converting the original SAR image into a semantic map and an attribute scattering center topology map; extracting three types of features and calculating class prototypes through three parallel feature branches; obtaining the confidence score of each branch using the Softmax function; assigning teacher and student branches according to the confidence score; minimizing KL divergence to achieve dynamic knowledge transfer; weighted fusion of predicted log odds followed by Softmax output; and backpropagation to optimize the network. This method, through multi-prototype fusion and interactive distillation, mitigates the problems of speckle noise and high intra-class variance in SAR images, maintaining high robustness and generalization ability even in extreme data-scarce scenarios.

Description

Technical Field

[0001] This application relates to the field of SAR target recognition technology, and in particular to a method, apparatus, device and medium for few-sample SAR target recognition. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave remote sensing technology, possesses all-weather, all-day operation capabilities and can penetrate interference such as clouds, playing a crucial role in Automatic Target Recognition (ATR) scenarios such as civilian vehicle monitoring. SAR ATR is essentially a fine-grained recognition task, and achieving accurate target recognition remains challenging due to the limited resolution of SAR images and their unique electromagnetic scattering characteristics. While deep learning-based SAR target recognition methods have made significant progress, they rely on large-scale labeled datasets. However, SAR image acquisition is difficult, labeling is costly, and sample scarcity is a prominent issue, leading to models prone to overfitting and insufficient generalization ability. Therefore, Few-shot learning (FSL) has become a key technological direction in this field.

[0003] Existing SAR FSL methods mainly fall into three categories: transfer learning, data augmentation, and meta-learning. Transfer learning extracts general features through cross-domain pre-training, data augmentation expands the dataset by generating synthetic samples, and meta-learning constructs a task-level training framework to quickly adapt to new classes. However, most of these methods focus on model architecture design and do not fully integrate the characteristics of the SAR domain. The inherent speckle noise in SAR images blurs target boundaries, and changes in imaging parameters (azimuth, viewing angle, etc.) easily lead to high intra-class variance of targets, making it difficult for traditional FSL methods to learn robust discriminative features. Some multi-feature fusion methods attempt to combine data from different modalities, but they fail to solve the problem of semantic inconsistency between features, making it difficult to fully leverage complementary advantages.

[0004] While the combination of domain knowledge and deep learning has shown potential in SAR target recognition, such as using azimuth and attribute scattering center (ASC) information to improve performance, the application of domain knowledge in few-shot learning is still in its early stages. Existing solutions fail to effectively balance data scarcity, imaging interference, and model generalization requirements. There is an urgent need for a technical solution that deeply integrates SAR domain knowledge, efficiently utilizes limited samples, and resolves semantic conflicts among multiple features to improve the accuracy and robustness of SAR target recognition under few-shot conditions. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, device, and medium for few-sample SAR target identification that can effectively improve the accuracy and robustness of target identification, addressing the aforementioned technical problems.

[0006] A method for few-sample SAR target identification, the method comprising:

[0007] Obtain a SAR sample dataset with category labels, and divide the SAR sample dataset into a support set and a query set based on a meta-learning framework;

[0008] The original SAR images in the SAR sample dataset are converted into semantic maps and attribute scattering center topology maps. Using a SAR target recognition network with three parallel feature branches, features of the original SAR images, the corresponding semantic maps, and the attribute scattering center topology maps are extracted to obtain the original image features, semantic map features, and attribute scattering center features.

[0009] Using the three types of features corresponding to each original SAR image in the support set, calculate the category prototype of each target category on the three feature branches;

[0010] Using the three types of features and corresponding category prototypes of each original SAR image in the query set, the predicted probability distribution of the three feature branches is calculated by the Softmax function, and the maximum value of the probability distribution on each feature branch is taken as the confidence of the corresponding feature branch. The three feature branches are assigned as teacher branches and student branches according to the relative confidence values. Dynamic knowledge transfer between feature branches is achieved by minimizing KL divergence.

[0011] The predicted log odds of the three feature branch networks are normalized and weighted by confidence, and then the target recognition results of the original SAR images in the query set are obtained by passing the Softmax function.

[0012] The loss function is calculated based on the target recognition results and the corresponding category labels. The parameters of the SAR target recognition network are optimized through backpropagation until convergence, thus obtaining the trained SAR target recognition network.

[0013] The SAR image to be identified is acquired, and after converting the SAR image into a semantic map and an attribute scattering center topology map, they are input together into a trained SAR target recognition network to obtain the SAR target recognition result.

[0014] In one embodiment, the semantic map is automatically generated through intensity-driven binarization and morphological thinning processes, specifically including:

[0015] The pixels with the highest and lowest gray values ​​in the original SAR image are extracted. After removing discrete points by a counting filter, a semantic map containing the target, shadow, and background is obtained through morphological processing and maximum connected component processing.

[0016] In one embodiment, the semantic map is automatically generated using an intensity-driven binarization and morphological thinning method, specifically including:

[0017] The attribute set of the attribute scattering center includes the azimuth position, distance position, amplitude, frequency dependence factor, length, attitude angle, and attitude dependence parameter of the scattering center.

[0018] In one embodiment, among the three parallel feature branches:

[0019] The feature branches of the original SAR image and semantic map features are extracted using a ResNet-12 encoder;

[0020] A 3-layer GNN-FILM encoder is used to extract features from the topological map of the attribute scattering centers.

[0021] In one embodiment, when calculating the category prototype for each target category on each of the three feature branches:

[0022] For each target category, the mean vector of all sample features in the support set under the corresponding feature branch is calculated, and the mean vector is used as the category prototype of that category in that feature branch.

[0023] In one embodiment, the predicted log odds are calculated based on the negative squared Euclidean distance between the feature vectors of each feature branch and the corresponding class prototype.

[0024] In one embodiment, the loss function is a joint loss function of confidence-weighted classification loss and distillation loss.

[0025] This application also provides a device for few-sample SAR target identification, the device comprising:

[0026] The SAR sample dataset acquisition module is used to acquire SAR sample datasets with category labels and divide the SAR sample data into support sets and query sets based on a meta-learning framework.

[0027] The feature extraction module is used to convert the original SAR image in the SAR sample data into a semantic map and an attribute scattering center topology map. Using a SAR target recognition network containing three parallel feature branches, the features of the original SAR image, the corresponding semantic map, and the attribute scattering center topology map are extracted respectively to obtain the original image features, semantic map features, and attribute scattering center features.

[0028] The category prototype calculation module is used to calculate the category prototype of each target category on the three feature branches by utilizing the three types of features corresponding to each original SAR image in the support set;

[0029] The dynamic knowledge transfer module is used to utilize the three types of features corresponding to each original SAR image in the query set and the corresponding category prototypes, calculate the predicted probability distribution of the three feature branches through the Softmax function, and take the maximum value of the probability distribution on each feature branch as the confidence of the corresponding feature branch. According to the relative value of the confidence, the three feature branches are assigned as teacher branches and student branches. Dynamic knowledge transfer between feature branches is achieved by minimizing the KL divergence.

[0030] The query set identification result calculation module is used to normalize the predicted log odds of the three feature branch networks, perform confidence weighted fusion, and then obtain the target identification result of the original SAR image in the query set through the Softmax function;

[0031] The network training module is used to calculate the loss function based on the target recognition result and the corresponding category label, optimize the SAR target recognition network parameters through backpropagation until convergence, and obtain the trained SAR target recognition network.

[0032] The SAR target recognition module is used to acquire the SAR image to be recognized, convert the SAR image into a semantic map and an attribute scattering center topology map, and then input them into the trained SAR target recognition network to obtain the SAR target recognition result.

[0033] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps described above in the method for few-sample SAR target identification.

[0034] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps described above in the method for few-sample SAR target identification.

[0035] The aforementioned method, apparatus, device, and medium for few-sample SAR target recognition divide the SAR sample dataset into a support set and a query set based on a meta-learning framework. It converts the original SAR images in the dataset into semantic maps and attribute scattering center topology maps. Using a SAR target recognition network with three parallel feature branches, it extracts features from the original SAR images, corresponding semantic maps, and attribute scattering center topology maps, obtaining original image features, semantic map features, and attribute scattering center features. Using the three types of features corresponding to each original SAR image in the support set, it calculates the class prototypes of each target category on the three feature branches. Using the three types of features and corresponding class prototypes corresponding to each original SAR image in the query set, it calculates the predicted probability distributions of the three feature branches using a Softmax function. The maximum value of the probability distribution on each feature branch is used as the confidence level of the corresponding feature branch. Based on the confidence level, the three feature branches are assigned as teacher branches and student branches. Dynamic knowledge transfer between branches is achieved by minimizing the KL divergence. The predicted log probabilities of the three feature branch networks are normalized, weighted, and fused using confidence levels, and then processed by Softmax. The function obtains the target recognition results of the original SAR images in the query set. Based on the target recognition results and corresponding category labels, a loss function is calculated. The parameters of the SAR target recognition network are optimized through backpropagation until convergence, resulting in a trained SAR target recognition network. The SAR image to be recognized is converted into a semantic map and an attribute scattering center topology map, which are then input into the trained SAR target recognition network to obtain the SAR target recognition result. This method effectively resists the problems of speckle noise and high intra-class variance in SAR images through multi-prototype feature fusion and dynamic interactive distillation, maintaining high robustness even in extreme data-scarce scenarios. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a few-sample SAR target identification method in one embodiment;

[0037] Figure 2 This is a schematic diagram of a multi-prototype strategy in one embodiment;

[0038] Figure 3 This is a schematic diagram of the architecture of the method in one embodiment;

[0039] Figure 4 This is a schematic diagram illustrating an example of a SAR multimodal view in one embodiment, wherein, Figure 4 (a) represents the original SAR image. Figure 4 (b) represents a segmented semantic image. Figure 4 (c) represents the topological image of the attribute scattering centers;

[0040] Figure 5This is a structural block diagram of a few-sample SAR target identification device in one embodiment;

[0041] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] Existing SAR target recognition methods with few samples suffer from insufficient integration of SAR domain knowledge, difficulty in handling high intra-class variability caused by coherent field noise and imaging parameter variations, and semantic inconsistencies that easily arise from multi-feature fusion. Figure 1 As shown, a method for few-sample SAR target identification is provided, which specifically includes the following steps:

[0044] Step S100: Obtain the SAR sample dataset with category labels, and divide the SAR sample dataset into a support set and a query set based on the meta-learning framework.

[0045] Step S110: Convert the original SAR images in the SAR sample dataset into semantic maps and attribute scattering center topology maps. Use a SAR target recognition network with three parallel feature branches to extract features from the original SAR images, the corresponding semantic maps, and the attribute scattering center topology maps, respectively, to obtain the original image features, semantic map features, and attribute scattering center features.

[0046] Step S120: Using the three types of features corresponding to each original SAR image in the support set, calculate the category prototype of each target category on the three feature branches.

[0047] Step S130: Using the three types of features and corresponding category prototypes of each original SAR image in the query set, the predicted probability distribution of the three feature branches is calculated by the Softmax function, and the maximum value of the probability distribution on each feature branch is taken as the confidence of the corresponding feature branch. The three feature branches are assigned as teacher branches and student branches according to the relative confidence values, and dynamic knowledge transfer between feature branches is achieved by minimizing the KL divergence.

[0048] Step S140: Normalize the predicted log odds of the three feature branch networks and perform confidence-weighted fusion, then use the Softmax function to obtain the target recognition result of the query set SAR image.

[0049] Step S150: Calculate the loss function based on the target recognition result and the corresponding category label, optimize the parameters of the SAR target recognition network through backpropagation until convergence, and obtain the trained SAR target recognition network.

[0050] Step S160: Obtain the SAR image to be identified, convert the SAR image into a semantic map and an attribute scattering center topology map, and input them together into the trained SAR target recognition network to obtain the SAR target recognition result.

[0051] In this method, by leveraging SAR domain knowledge, semantic maps and attribute scattering centers (ASCs) are used to represent the geometric and electromagnetic scattering features of the target, thus forming a comprehensive multi-prototype representation. The semantic map encodes the target's geometry, providing a more stable geometric feature representation compared to the original SAR image, and is less susceptible to speckle noise and subtle variations in imaging conditions. Meanwhile, the ASC provides a representation of the intrinsic scattering mechanism, offering discriminative information that is generally more stable than the original image. This multi-prototype strategy fundamentally enriches the feature representation, laying a solid foundation for few-shot learning, as shown in Figure 2.

[0052] However, directly fusing these heterogeneous features can suffer from semantic inconsistencies. To address this, an interactive distillation mechanism is introduced in this method to facilitate collaborative knowledge transfer across feature branches—SAR images, semantic maps, and ASCs. A key innovation within this framework is the sample-adaptive confidence weighting strategy. Since the discriminative power of each feature branch differs for different samples, confidence scores are dynamically calculated in this method to quantify the model's certainty in predicting each branch. First, these scores guide the distillation process, forcing feature branches with lower confidence to learn from their corresponding branches with higher confidence. This creates a dynamic "teaching" process that improves the quality of the overall feature representation. In the final classification stage, confidence scores are used as adaptive weights to fuse the metric distances from each branch, enabling the model to dynamically prioritize the most reliable information for the final decision. This integrated design ensures that this method not only leverages the complementary advantages of different representations but also adaptively adjusts its learning and inference processes based on the unique characteristics of each input sample.

[0053] like Figure 3The diagram shows the architecture of the proposed method. The proposed architecture runs on three parallel feature extraction streams. The input SAR image is first converted into two additional feature branches: a semantic map and attribute scattering centers (ASCs). These three representations are then fed into independent branches to extract features and generate specific prototypes. Next, the confidence weight of each sample's feature branch is calculated. A confidence-based interactive distillation mechanism performs knowledge transfer between branches to collaboratively enhance shared discriminative patterns and suppress feature branch-specific noise. For the final classification, the metric distance between the sample and the prototype is dynamically weighted according to the confidence weight of each branch to obtain a fused metric score, thereby effectively maximizing inter-class separability. The entire framework is trained end-to-end through joint optimization of the cross-entropy loss function and the knowledge distillation loss function.

[0054] In step S100, a training and testing data paradigm adapted to low-sample scenarios is constructed using a meta-learning framework, providing high-quality data support for subsequent feature extraction, prototype learning, and knowledge distillation processes. The SAR sample dataset covers images of multiple categories of a single target, such as civilian vehicles like cars, pickup trucks, SUVs, flatbed trucks, and excavators, with each sample clearly labeled to ensure the accuracy of subsequent model training and performance evaluation. The dataset collection needs to cover diverse imaging conditions, including different depression angles (e.g., 15°, 17°, 22°, 30°), azimuth angles, and imaging bands, to simulate the complexity of actual SAR imaging scenarios and ensure data diversity and representativeness. Simultaneously, preprocessing operations are performed on the acquired raw SAR images, including center cropping to a uniform size (e.g., 84×84 pixels) and grayscale normalization, to eliminate the impact of image size differences and uneven grayscale distribution on model training, ensuring the consistency and stability of data input.

[0055] In this embodiment, the preprocessed SAR sample dataset is divided into a support set and a query set based on the meta-learning framework. The division process follows the N-way K-shot few-shot task configuration rule: first, the entire SAR sample dataset is divided into non-overlapping training sets ( ) and test set ( The training set contains categories that are base classes ( ), The test set contains categories of the new class ( ), and satisfy ∩ = This ensures that the meta-testing phase can effectively verify the model's ability to generalize to unseen categories.

[0056] Specifically, in the meta-training phase, from the training set ( In a training task (Episode), N base classes are randomly selected. For each selected base class, K samples are randomly chosen to form the support set for that task, used to learn the category prototype. N -way K In the -shot task, select N There are *k* categories, and each category has *k* samples forming a support set, represented as: S = {( x i , y i )} N×K in x i and y i These represent the original SAR images supported and their corresponding class labels, respectively. Additionally, M samples are extracted from each selected base class to form the query set for this task, represented as: Q = {( x j , y j )} N×M ,in x i and y i This indicates that the dataset supports a set of raw SAR images and their corresponding class labels, used to verify the model's learning performance on this task and to calculate the loss. Here, N represents the number of classes for each task (i.e., "N-way"), K represents the number of samples used to build prototypes for each class (i.e., "K-shot"), and M represents the number of samples used for testing each class. Typically, a reasonable ratio is set based on the dataset size (e.g., K=1, 5, 10, M is 2-5 times K) to balance training efficiency and evaluation reliability.

[0057] Specifically, in the meta-testing phase, the N-way K-shot configuration, which is completely consistent with that in the meta-training phase, is used from the test set ( Samples are randomly drawn from the new class to construct the support set and query set. At this point, only the class labels of the samples in the support set are used to generate the class prototype of the new class, and the class labels of the query set are only used for model performance evaluation. The model needs to utilize the general learning ability learned in the meta-training phase to quickly adapt to the new class based on a small number of samples in the support set, so as to achieve accurate classification of the samples in the query set.

[0058] In step S110, SAR segmentation semantic map and ASC topology map are used to represent the geometric features and structural characteristics of the target, respectively. Examples of SAR semantic map and ASC topology map are shown in Figure 4.

[0059] Since target-shadow pairs jointly constrain the geometry of objects under given imaging conditions, dividing the scene into target, shadow, and background can produce a stable target representation. To avoid manual annotation, in this embodiment, a semantic map is automatically generated through binarization and morphological thinning. The process first extracts a set of pixels with the highest and lowest gray values ​​from the image, removes discrete points using a counting filter, and obtains smooth target and shadow regions through morphological processing and maximum connected component processing.

[0060] Furthermore, the Attribute Scattering Center (ASCM) model clarifies the dependence of the scatterer on frequency and angle, suggesting that the electromagnetic scattering response of a target can be viewed as... p The superposition of independent scattering centers is represented as:

[0061]

[0062] in,

[0063]

[0064] In formulas (1) and (2), This represents the total scattered field of the target. θ ={[ , , , , , , ]} represents the attribute set of the scattering center, where, , These represent the azimuth and distance positions of the scattering center, respectively. Indicates amplitude, Indicates frequency-dependent factor, Indicates the length of the scattering center. Indicates the azimuth angle of the scattering center; Indicates the azimuth of the scattering center. Posture dependence.

[0065] In this embodiment, the attribute scattering center topology map is extracted and constructed based on sparse representation using the orthogonal matching pursuit algorithm.

[0066] In this embodiment, three different inputs—the original SAR image, the semantic map, and the attribute scattering center topology map—are all embedded into a one-dimensional vector for subsequent prototype learning.

[0067] In this embodiment, among the three parallel feature branches: a feature branch using a ResNet-12 encoder to extract features from the original SAR image and semantic map; and a feature branch using a 3-layer GNN-FILM encoder to extract features from the attribute scattering center topology map.

[0068] Specifically, for the original SAR image and semantic map, feature extraction is directly performed using a CNN-based embedding model. This utilizes two different encoders. and To generate their respective embeddings:

[0069]

[0070] In the above formula, These represent the original SAR image and the semantic map, respectively. , These are respectively represented as features extracted from the original SAR image and the semantic map.

[0071] In contrast, the attribute scattering center consists of a set of N scattering points for each target. To effectively learn the structural information of the target, a graph-based representation method is employed in this embodiment. This involves first constructing a graph from these points, and then using an embedding model based on a graph neural network to generate feature vectors.

[0072] Specifically, the attribute scattering centers are first represented as a topological graph structure, as shown below. A node set consists of a set of nodes and edges connecting them. A node set is defined as follows: Where m represents the total number of nodes in the graph structure, and each node ∈ V It corresponds to a scattering center. The edge set is defined as E. { = , ∈ V} represents the relationship between nodes. Edge ∈ E Represents a node and There is a direct relationship between them. Node V and edge E are both used as features for initializing the input of subsequent network layers. Specifically, each node... ∈ V The initial features are represented as The aim is to encode the fundamental properties of its corresponding scattering center:

[0073]

[0074] In formula (4), ( , The coordinates () represent the coordinates of the scattering center relative to its spatial location. Indicates its scattering intensity (amplitude) and scattering type. It is based on frequency-dependent factors. and scattering center length Based on Table 1.

[0075] Table 1: Differences and Geometric scattering types under combination

[0076]

[0077] Specifically, edge ∈ E The initial features are intended to encode and The scattering relationship between them. Since spatial proximity implies a stronger connection, Defined as a function of Euclidean distance, and only if... and When the distance between them is less than a threshold, the edges Only then does it exist. Its formula is expressed as:

[0078]

[0079] In formula (5), Represents a node and The Euclidean distance between them. Parameters This is the standard deviation of the Gaussian kernel. (Value) This serves as a hard threshold to prune any connections between nodes that are too far apart.

[0080] In this embodiment, a GNN encoder is used. Topology-aware node representations are learned by iteratively aggregating neighborhood information. Finally, a single graph-level embedding is generated through an average pooling layer, representing the features of the attribute-scattering center topological graph, as follows:

[0081]

[0082] In formula (6), Represents the topological graph of attribute scattering centers. This indicates the characteristics of the scattering center.

[0083] In step S120, representative category prototypes are established for each target category on three feature branches, providing a discrimination benchmark for feature matching and classification of subsequent query set samples. A category prototype is a canonical representation of a target category in a specific feature space, comprehensively reflecting the core feature attributes of all samples in that category. Since this method uses three heterogeneous features—original image, semantic map, and attribute scattering centers (ASC)—it is necessary to construct category prototypes for each target category on three feature branches to ensure that subsequent target discrimination can be performed from multiple dimensions such as geometric structure, electromagnetic scattering, and appearance details.

[0084] In this embodiment, when calculating the category prototype of each target category on the three feature branches: for each target category, the mean vector of all sample features of the support set under the corresponding feature branch is calculated, and the mean vector is used as the category prototype of the category on that feature branch.

[0085] Specifically, each category c and feature branches M Category prototypes ∈ {I,S,A} Calculated as from the corresponding support set The centroid of the feature embedding, i.e. the mean vector, is represented by the following process:

[0086]

[0087] In formula (7), It comes from the feature branch M The support set of the first i The embedding vector of each sample, This indicates that it includes all categories. c The support set of the sample index, Represents a set cardinality (i.e., number of samples), prototype As a feature branch M Categories in the embedded space c The canonical representation of a category is the prototype.

[0088] In step S130, the prediction reliability of each feature branch is quantified, and the knowledge transfer relationship between branches is dynamically established to solve the problem of semantic inconsistency among multiple heterogeneous features, while strengthening the discrimination ability of weaker feature branches. Utilizing the three types of features corresponding to each original SAR image in the query set—namely, original image features, semantic map features, and attribute scattering center features—as well as the category prototypes of each target category on the three feature branches calculated in step S120, dynamic knowledge transfer between branches is achieved through an interactive distillation (ID) framework that includes probability distribution calculation, confidence quantification, teacher-student branch allocation, and KL divergence optimization. This allows the discrimination experience of high-reliability branches to guide the learning of low-reliability branches.

[0089] In this embodiment, for each original SAR image in the query set First, by embedding its features Applying the softmax function to the negative squared Euclidean distance between all class prototypes, for each feature branch M (where M... {I,S,A} derives a class probability distribution (soft label). Original SAR image The probability of belonging to category c in the category feature branch M is defined as:

[0090]

[0091] In formula (8), T represents the temperature parameter controlling the sharpness of the distribution, and K represents the total number of categories. The confidence score of this feature branch. Then it is defined as the maximum value of this distribution:

[0092]

[0093] In formula (9), the confidence score quantifies the certainty of the classification. Subsequently, for the original SAR image... Any feature branch with a high confidence score acts as the "teacher," while any feature branch with a low score acts as the "student." Knowledge is derived by minimizing the number of teachers (…). ) and students ( The propagation is achieved through the Kullback-Leibler (KL) divergence between probability distributions. The distillation loss for a single pair is:

[0094]

[0095] This objective forces the student network to mimic the teacher's prediction logic. The final distillation loss, LDistillation, aggregates the KL divergence loss of all dynamically assigned teacher-student pairs for each sample in the batch:

[0096]

[0097] In the above formula, This indicates the number of samples in the query set. The interactive distillation framework described above ensures that knowledge is always passed from the feature branch with higher confidence to the branch with lower confidence.

[0098] In step S140, the prediction information of the three feature branches is fused through a dynamic weighting strategy to give full play to the complementary advantages of each branch and generate an accurate and robust target recognition result. The predicted log odds of the three feature branches calculated in step S130, namely the original image feature branch, the semantic map feature branch, and the attribute scattering center feature branch, as well as the normalized confidence of each branch, are calculated.

[0099] In this embodiment, the predicted log odds are calculated based on the negative squared Euclidean distance between the feature vectors of each branch and the corresponding class prototypes, directly reflecting the similarity between the sample and each class prototype (the smaller the distance, the larger the log odds value, and the higher the similarity). The normalized confidence score is obtained by normalizing the maximum value of the predicted probability distribution of each branch in step S130, and is used to quantify the reliability of each branch in classifying the current sample.

[0100] Specifically, the predicted log odds are expressed as: Then, normalized confidence scores are used. The predicted log odds of each feature branch are aggregated as weights to obtain the fused predicted log odds, expressed as:

[0101]

[0102] Furthermore, in order to obtain the final predicted probability distribution Applying the Softmax function to .sample Category The probability is given by the following formula:

[0103]

[0104] In this method, the standard negative log-likelihood is used as the classification loss. For a batch of query samples... Each sample The true label is represented as The classification loss is calculated as follows:

[0105]

[0106] In step S150, the loss function is a joint function of confidence-weighted cross-entropy loss and interactive distillation loss.

[0107] In this embodiment, a composite loss function is optimized. End-to-end training is performed, with supervision applied separately to each modality and weighted by its respective confidence score. The total loss is defined as:

[0108] (15)

[0109] In formula (14), the hyperparameters Control the intensity of knowledge distillation.

[0110] In this paper, the effectiveness of the proposed method is also demonstrated experimentally. The experiments used measured SAR image data of stationary ground targets. This dataset contains SAR images of 10 types of ground targets, acquired from a complete 360° azimuth range and multiple elevation angles. All target images were normalized to 128×128 pixels before the experiments.

[0111] In the experimental setup of this embodiment, the dataset is divided into a meta-training set and a meta-testing set. Seven categories from a 17° pitch angle are selected as base classes for meta-training; the remaining three categories are selected as novel classes for meta-testing. During the testing and evaluation phase, to verify the robustness of the model, the novel classes are evaluated at three different pitch angles (15°, 17°, and 30°).

[0112] Image preprocessing and training configuration: SAR image samples from each dataset are first preprocessed to 84×84 pixels by center cropping. This embodiment uses the standard N-way K-shot meta-learning framework to construct meta-training and meta-testing tasks. The model is trained for a total of 100 epochs, with each epoch containing 200 meta-training tasks. During the evaluation phase, the average accuracy of 500 test tasks is calculated to ensure the stability and reliability of performance metrics.

[0113] Network structure and optimization parameters: In this method model, the image feature embedding module ( ) and semantic graph feature embedding module ( All use the ResNet-12 network; the attribute embedding module ( The model is constructed using a 3-layer graph neural network (GNN-FILM). The entire model is trained end-to-end using the Adam optimizer, with an initial learning rate set to 5 × 10⁻⁶. -4 Momentum parameters ( Set it to 0.5.

[0114] Simulation Results Analysis: The proposed method (CMPIDN) and six advanced techniques were used to perform target recognition on the MSTAR dataset under different experimental settings, yielding recognition results for the test samples using various methods. The six advanced techniques are: FEAT method, RelationNet method, ProtoNet method, DN4 method, CAN method, and DEEPEMD method.

[0115] To evaluate the method of this invention compared to six prior art methods, the N-way K-shot accuracy (where N is the number of classes in each meta-task and K is the number of support set samples), commonly used in few-shot learning, was employed as the evaluation metric. Training was conducted at 17°, and 3-way 1-shot, 3-way 5-shot, and 3-way 10-shot tasks were performed at pitch angles of 15°, 17°, and 30°. A higher recognition rate indicates better recognition performance.

[0116] The recognition rates of the seven methods used in the simulation experiment were calculated and the results are listed in Table 2.

[0117] Table 2. Comparison of 3-WAY-K-SHOT recognition rates (%) under different experimental settings on the MSTAR dataset.

[0118]

[0119] As shown in Table 2, for the MSTAR dataset used in the simulation experiments of this method, the SAR target recognition method based on confidence-weighted multi-prototype interactive distillation network (CMPIDN) proposed in this method achieved the best recognition results under all experimental settings.

[0120] Although the performance of existing contrastive techniques (such as FEAT and RelationNet) fluctuates with changes in pitch angle and the number of supporting samples (shots), the proposed model consistently maintains a leading edge. The performance advantage of the proposed method is particularly pronounced in the most challenging 1-shot scenarios (i.e., when labeled data is extremely scarce). For example, in a 17-15 setting, the proposed method outperforms the second-best CAN method by 4.74%; in a 17-17 setting, it outperforms the second-best DN4 method by 4.00%. This demonstrates the superior robustness of the proposed method under extremely scarce data conditions. This result illustrates that existing techniques (such as ProtoNet and DN4) are prone to performance bottlenecks when relying solely on learning from raw image features under severely limited data conditions (such as 1-shot). In contrast, the proposed method constructs more stable and accurate category prototypes through multi-prototype fusion and interactive distillation mechanisms, effectively overcoming the limitations of single-modality feature extraction, thereby uncovering richer target information and significantly improving the accuracy of SAR target recognition.

[0121] The aforementioned method for few-shot SAR target recognition utilizes CNNs to extract feature representations from the original SAR image and semantic map, and employs graph neural networks (GNNs) to extract feature representations of the ASCII table. The final classification result is obtained by weighting the distance between multiple features and multiple prototypes. An interactive distillation architecture is developed to establish semantic consistency among multiple features and promote knowledge transfer between feature branches, thereby significantly enhancing the discriminative power of weaker feature branches. Furthermore, sample-specific confidence weights are used to dynamically guide the knowledge distillation process. These weights also adaptively adjust the contribution of each feature branch to the final classification fusion, thus optimizing performance at the sample adaptation level. Finally, experiments on example datasets demonstrate the effectiveness of the proposed method and show that its performance significantly outperforms existing state-of-the-art few-shot SAR target recognition methods.

[0122] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0123] In one embodiment, such as Figure 5 As shown, a device for few-sample SAR target recognition is provided, comprising: a SAR sample dataset acquisition module 200, a feature extraction module 210, a category prototype calculation module 220, a dynamic knowledge transfer module 230, a query set recognition result calculation module 240, a network training module 250, and a SAR target recognition module 260, wherein:

[0124] The SAR sample dataset acquisition module 200 is used to acquire SAR sample datasets with category labels and divide the SAR sample data into a support set and a query set based on a meta-learning framework.

[0125] The feature extraction module 210 is used to convert the original SAR image in the SAR sample data into a semantic map and an attribute scattering center topology map. Using a SAR target recognition network containing three parallel feature branches, the features of the original SAR image, the corresponding semantic map, and the attribute scattering center topology map are extracted respectively to obtain the original image features, semantic map features, and attribute scattering center features.

[0126] The category prototype calculation module 220 is used to calculate the category prototype of each target category on the three feature branches by utilizing the three types of features corresponding to each original SAR image in the support set.

[0127] The dynamic knowledge transfer module 230 is used to utilize the three types of features corresponding to each original SAR image in the query set and the corresponding category prototypes, calculate the predicted probability distribution of the three feature branches through the Softmax function, and take the maximum value of the probability distribution on each feature branch as the confidence of the corresponding feature branch. Based on the relative value of the confidence, the three feature branches are assigned as teacher branches and student branches, and dynamic knowledge transfer between feature branches is achieved by minimizing the KL divergence.

[0128] The query set identification result calculation module 240 is used to perform normalized confidence weighted fusion of the predicted log odds of the three feature branch networks, and then obtain the target identification result of the original SAR image in the query set through the Softmax function.

[0129] The network training module 250 is used to calculate the loss function based on the target recognition result and the corresponding category label, optimize the SAR target recognition network parameters through backpropagation until convergence, and obtain the trained SAR target recognition network.

[0130] The SAR target recognition module 260 is used to acquire the SAR image to be recognized, convert the SAR image into a semantic map and an attribute scattering center topology map, and then input them into the trained SAR target recognition network to obtain the SAR target recognition result.

[0131] Specific limitations regarding the few-sample SAR target recognition device can be found in the limitations regarding the few-sample SAR target recognition method described above, and will not be repeated here. Each module in the aforementioned few-sample SAR target recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0132] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for few-sample SAR target recognition. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0133] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0134] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0135] Obtain a SAR sample dataset with category labels, and divide the SAR sample data into a support set and a query set based on a meta-learning framework;

[0136] The original SAR image in the SAR sample data is converted into a semantic map and an attribute scattering center topology map. Using a SAR target recognition network with three parallel feature branches, the features of the original SAR image, the corresponding semantic map, and the attribute scattering center topology map are extracted respectively to obtain the original image features, semantic map features, and attribute scattering center features.

[0137] Using the three types of features corresponding to each original SAR image in the support set, calculate the category prototype of each target category on the three feature branches;

[0138] Using the three types of features and corresponding category prototypes of each original SAR image in the query set, the predicted probability distribution of the three feature branches is calculated by the Softmax function, and the maximum value of the probability distribution on each feature branch is taken as the confidence of the corresponding feature branch. The three feature branches are assigned as teacher branches and student branches according to the relative confidence values. Dynamic knowledge transfer between feature branches is achieved by minimizing KL divergence.

[0139] The predicted log odds of the three feature branch networks are normalized and weighted by confidence, and then the target recognition results of the original SAR images in the query set are obtained by passing the Softmax function.

[0140] Based on the target recognition results and the corresponding category labels, the loss function is calculated, and the parameters of the SAR target recognition network are optimized through backpropagation until convergence, thus obtaining the trained SAR target recognition network.

[0141] The SAR image to be identified is acquired, and after converting the SAR image into a semantic map and an attribute scattering center topology map, they are input together into a trained SAR target recognition network to obtain the SAR target recognition result.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0143] Obtain a SAR sample dataset with category labels, and divide the SAR sample data into a support set and a query set based on a meta-learning framework;

[0144] The original SAR image in the SAR sample data is converted into a semantic map and an attribute scattering center topology map. Using a SAR target recognition network with three parallel feature branches, the features of the original SAR image, the corresponding semantic map, and the attribute scattering center topology map are extracted respectively to obtain the original image features, semantic map features, and attribute scattering center features.

[0145] Using the three types of features corresponding to each original SAR image in the support set, calculate the category prototype of each target category on the three feature branches;

[0146] Using the three types of features and corresponding category prototypes of each original SAR image in the query set, the predicted probability distribution of the three feature branches is calculated by the temperature-controlled Softmax function, and the maximum value of the probability distribution on each feature branch is taken as the confidence of the corresponding feature branch. The three feature branches are assigned as teacher branches and student branches according to the relative value of the confidence. Dynamic knowledge transfer between branches is achieved by minimizing the KL divergence.

[0147] The predicted log odds of the three feature branch networks are normalized and weighted by confidence, and then the target recognition results of the original SAR images in the query set are obtained by passing the Softmax function.

[0148] Based on the target recognition results and the corresponding category labels, the loss function is calculated, and the parameters of the SAR target recognition network are optimized through backpropagation until convergence, thus obtaining the trained SAR target recognition network.

[0149] The SAR image to be identified is acquired, and after converting the SAR image into a semantic map and an attribute scattering center topology map, they are input together into a trained SAR target recognition network to obtain the SAR target recognition result.

[0150] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0151] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0152] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for few-shot SAR target recognition, characterized in that, The method includes: Obtain a SAR sample dataset with category labels, and divide the SAR sample dataset into a support set and a query set based on a meta-learning framework; The original SAR images in the SAR sample dataset are converted into semantic maps and attribute scattering center topology maps. Using a SAR target recognition network with three parallel feature branches, features of the original SAR images, the corresponding semantic maps, and the attribute scattering center topology maps are extracted to obtain the original image features, semantic map features, and attribute scattering center features. Using the three types of features corresponding to each original SAR image in the support set, calculate the category prototype of each target category on the three feature branches; Using the three types of features and corresponding category prototypes of each original SAR image in the query set, the predicted probability distribution of the three feature branches is calculated by the Softmax function, and the maximum value of the probability distribution on each feature branch is taken as the confidence of the corresponding feature branch. The three feature branches are assigned as teacher branches and student branches according to the relative confidence values. Dynamic knowledge transfer between feature branches is achieved by minimizing KL divergence. The predicted log odds of the three feature branch networks are normalized and weighted by confidence, and then the target recognition results of the original SAR images in the query set are obtained by passing the Softmax function. The loss function is calculated based on the target recognition results and the corresponding category labels. The parameters of the SAR target recognition network are optimized through backpropagation until convergence, thus obtaining the trained SAR target recognition network. The SAR image to be identified is acquired, and after converting the SAR image into a semantic map and an attribute scattering center topology map, they are input together into a trained SAR target recognition network to obtain the SAR target recognition result.

2. The method for few-shot SAR target recognition according to claim 1, characterized in that, The semantic map is automatically generated using an intensity-driven binarization and morphological thinning method, specifically including: The pixels with the highest and lowest gray values ​​in the original SAR image are extracted. After removing discrete points by a counting filter, a semantic map containing the target, shadow, and background regions is obtained through morphological processing and maximum connected component extraction.

3. The method for few-sample SAR target identification according to claim 1, characterized in that, The attribute scattering center topology map is extracted and constructed using an orthogonal matching pursuit algorithm; The attribute set of the attribute scattering center includes the azimuth position, distance position, amplitude, frequency dependence factor, length, attitude angle, and attitude dependence parameter of the scattering center.

4. The method for few-sample SAR target identification according to claim 1, characterized in that, Of the three parallel feature branches: The feature branches of the original SAR image and semantic map features are extracted using a ResNet-12 encoder; A 3-layer GNN-FILM encoder is used to extract features from the topological map of the attribute scattering centers.

5. The method for few-sample SAR target identification according to claim 1, characterized in that, When calculating the category prototype for each target category on each of the three feature branches: For each target category, the mean vector of all sample features in the support set under the corresponding feature branch is calculated, and the mean vector is used as the category prototype of that category in that feature branch.

6. The method for few-sample SAR target identification according to claim 1, characterized in that, The predicted log odds are calculated based on the negative squared Euclidean distance between the feature vectors of each feature branch and the corresponding class prototype.

7. The method for few-sample SAR target identification according to claim 1, characterized in that, The loss function is a joint loss function of confidence-weighted classification loss and distillation loss.

8. A device for identifying targets using few-sample SAR, characterized in that, The device includes: The SAR sample dataset acquisition module is used to acquire SAR sample datasets with category labels and divide the SAR sample datasets into support sets and query sets based on a meta-learning framework. The feature extraction module is used to convert the original SAR images in the SAR sample dataset into semantic maps and attribute scattering center topology maps. Using a SAR target recognition network containing three parallel feature branches, the module extracts features from the original SAR images, the corresponding semantic maps, and the attribute scattering center topology maps, respectively, to obtain original image features, semantic map features, and attribute scattering center features. The category prototype calculation module is used to calculate the category prototype of each target category on the three feature branches by utilizing the three types of features corresponding to each original SAR image in the support set; The dynamic knowledge transfer module is used to utilize the three types of features corresponding to each original SAR image in the query set and the corresponding category prototypes, calculate the predicted probability distribution of the three feature branches through the Softmax function, and take the maximum value of the probability distribution on each feature branch as the confidence of the corresponding feature branch. According to the relative value of the confidence, the three feature branches are assigned as teacher branches and student branches. Dynamic knowledge transfer between feature branches is achieved by minimizing the KL divergence. The query set identification result calculation module is used to normalize the predicted log odds of the three feature branch networks, perform confidence weighted fusion, and then obtain the target identification result of the original SAR image in the query set through the Softmax function; The network training module is used to calculate the loss function based on the target recognition result and the corresponding category label, optimize the parameters of the SAR target recognition network through backpropagation until convergence, and obtain the trained SAR target recognition network. The SAR target recognition module is used to acquire the SAR image to be recognized, convert the SAR image into a semantic map and an attribute scattering center topology map, and then input them into the trained SAR target recognition network to obtain the SAR target recognition result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images