SAR (Synthetic Aperture Radar) long-tail image target classification method based on topological structure

Abstract

The invention relates to an SAR long-tail image target classification method based on a topological structure. The method comprises the following steps: extracting a backbone network feature map represented by an SAR image; after initial features are given, feature decomposition of frequency perception is executed; then, the structural branches extract stable geometric priori; the texture branch learning has a noise perception capability and texture semantics with a reliability weight; finally, the dynamic tail class balance module predicts fusion weights related to the classes for the classes; and inputting the fused representation into a classifier for target classification of the SAR image. According to the method, explicit decoupling and information retention of the structure and the texture are realized through the structure branch constrained by the frequency domain and the noise sensing texture branch; according to the category density and the representation stability, the fusion proportion of the two branches is adjusted in a self-adaptive mode; a tail class-oriented structure-driven prototype system is provided, and a stable semantic structure can be formed, so that the recognition performance of few-sample tail classes is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, and particularly relates to a SAR long-tail image target classification method based on topology. Background Technology

[0002] Synthetic Aperture Radar (SAR) target recognition has long been the cornerstone of achieving all-weather, day-and-night intelligent sensing systems. With the rapid deployment of SAR sensors in surveillance, mapping, and autonomous platforms, the demand for reliable Automatic Target Recognition (ATR) has increased significantly. Recently, deep learning-based SAR ATR frameworks have made significant progress. However, their effectiveness in real-world environments remains severely limited by the long-tailed class distribution: a few head classes account for the vast majority of samples, while many classes crucial for practical tasks appear only sparsely.

[0003] Traditional long-tail learning methods typically address the data imbalance problem through class balancing and reweighting, decoupling representations from classifiers, bi-branch balancing, boundary-based decision adjustment, or distribution-aware modeling. However, these techniques are primarily designed for natural images and implicitly assume that the minority class only faces a lack of samples. This assumption does not hold true in SAR imaging. Unlike optical images, SAR imaging characteristics arise from the combined effects of complex factors such as geometry, shape, electromagnetic scattering, and speckle noise. As is widely acknowledged in the SAR field, SAR data naturally contains two distinct layers of information: structural information (physically stable) and textural information (highly sensitive to noise and imaging conditions).

[0004] This asymmetry leads to a crucial but long-neglected phenomenon: in long-tailed SAR data, reliable texture information decays faster than structural information as category frequency decreases. Tail categories typically exhibit unstable, noise-dominated textures, failing to form compact and meaningful semantic clusters. Furthermore, mainstream SAR ATR algorithms often learn holistic features, failing to respect the hierarchical nature of SAR signals. Meanwhile, structure-texture decoupling, frequency-domain-based modeling, and feature stabilization strategies—methods proven effective in natural images—remain largely unexplored in SAR. Summary of the Invention

[0005] Long-tailed synthetic aperture radar (SAR) target recognition is not only affected by data scarcity, but more importantly, by the long-neglected information imbalance between structural stability and texture fragility. Existing resampling and reweighting strategies distort data distribution, risking the destruction of SAR physical realism; while a unified feature learning pipeline mixes inherently heterogeneous structural and texture cues, causing noise-driven texture to dominate the tail-class feature representation. To address this fundamental bottleneck, this application argues that long-tailed SAR recognition is not only a data imbalance problem, but also an information imbalance between robust structural information and fragile texture information. This perspective suggests a paradigm shift: SAR representation must be explicitly deconstructed based on its physical characteristics, selectively preserved, and adaptively balanced.

[0006] To address this issue, this application proposes the STDB structure-texture dual-branch network, a novel information-preserving paradigm that explicitly decouples and coordinates the modeling of structure and texture. A frequency-domain-constrained structure branch extracts robust geometric topological information, while a noise-aware texture branch suppresses unreliable scattering texture patterns through a self-discriminative learning mechanism. Furthermore, a Dynamic Tail Class Balancing Module (DTBM) adaptively fuses the outputs of the two branches based on class density and the stability of feature representations.

[0007] This application also introduces the first structure-driven prototyping system for tail classes, which can construct stable category prototypes based on geometric information, thereby alleviating the problem of long-term semantic instability of tail classes. Extensive experiments on multiple long-tailed SAR benchmarks show that this application can significantly improve tail class recognition accuracy and overall classification balance, and brings higher interpretability through cross-branch structure-texture collaboration.

[0008] This application consists of two cooperating branches: a frequency-domain-constrained structure branch for extracting stable geometric structures, and a noise-aware texture branch for filtering unreliable scattering textures. Furthermore, a Dynamic Tail Class Balancing Module (DTBM) adaptively adjusts the contributions of the two branches based on class density and the stability of feature representations. Finally, this application proposes the first structure-driven prototype mechanism for few-sample tail classes, capable of constructing stable geometric-semantic anchors for tail classes even under extremely data-scarce conditions.

[0009] To achieve the above objectives, this application discloses a SAR long-tail image target classification method based on topology, comprising the following steps: The SAR image is represented as: The corresponding backbone network feature map is denoted as: in, These represent the image space height, width, and channel dimensions, respectively. Represents the feature map of the backbone network; Given initial features Then, perform frequency-aware feature decomposition: in and These represent the feature responses that are structure-dominant and texture-dominant, respectively. This indicates the operation of performing frequency domain feature decomposition, which decomposes the input features. It is broken down into two parts: structure and texture. Subsequently, structural branches extract stable geometric priors. : Indicates structural branches; Texture branch learning incorporates noise-aware texture semantics with reliability weights. : Indicates texture branching; Ultimately, the dynamic tail class balancing module balances the categories. Predicting category-related fusion weights ,form: The fused representation is then input into a classifier for target classification in SAR images.

[0010] Furthermore, the structural branching extracts a structural representation constrained by the frequency domain, specifically including the following steps: Initial features The two-dimensional discrete Fourier transform yields A learnable low-pass mask We obtain it from the following formula: in It is a lightweight CNN. For the sigmoid function; The characteristic of structural dominance is: Where ⊙ represents element-wise multiplication. This means performing an inverse transformation on the frequency domain data to bring the structural information in the frequency domain back to the spatial domain, thus obtaining the final structural dominance feature; Use a regular expression to encourage consistent low-frequency filtering behavior in the low-pass mask: in Used to suppress high-frequency leakage, thereby preventing erroneous texture components from entering structural branches; Extract the strongest scattering centers And construct a topology graph : The set of nodes Spatial coordinates of the scattering center, adjacency matrix Defined as: They are the scattering centers and spatial coordinates, It is a hyperparameter used to control the decay rate of similarity in the adjacency matrix; Using a graph neural network (GNN) to encode global structural relationships: in Pyramids representing multi-scale structural features; To address the issue of unstable structural semantics in tail categories due to insufficient samples, an exponential moving average is incorporated as a consistency constraint loss: These are the structural features extracted by structural branches, representing the geometric information extracted by the model through structural branches. It is an operator used to freeze gradients, preventing structural features from changing during training; This loss is used to stabilize the representation of structural features during training.

[0011] Furthermore, the texture branch performs noise-aware self-discriminative texture learning, specifically including: Use a complementary, learnable bandpass mask Bandpass texture decomposition is performed, and its form is as follows: in For lightweight CNNs, For the sigmoid function; This yields the texture-dominated feature. : To ensure the bandpass filter has an effective passband, a regularization term is added: This feature is used to enhance the bandpass characteristics of the filter, thereby distinguishing texture details; Encoding is performed using a dual-path residual encoder, where one path models local high-frequency features through depthwise separable convolutions, and the other path captures global scattering correlations through nonlocal attention. The overall texture encoding is represented as: Indicates local high-frequency features, This refers to non-local attention, which focuses on the relationships between distant pixels in an image to enhance global perception. To mitigate the noise response, a reliability score is calculated for each token. : This represents a multilayer perceptron. It is the feature representation of each token in the texture branch, and each token corresponds to a local region in the image; The final texture representation is as follows: The loss function is as follows: Positive sample pairs are constructed from reliable tokens, and negative sample pairs are constructed from unreliable or cross-category tokens. These are samples and Texture features, It's a temperature parameter used to control the smoothness of the similarity score. It is a similarity function. It is the index of the negative sample, representing a sample that is related to... Dissimilar tokens; This loss function is used to prevent the model from mimicking tail samples dominated by noise.

[0012] Furthermore, the dynamic tail-class balancing module performs the following steps: category The number of samples is Its normalized density for: It is the number of samples in each category of the dataset; Calculate the stability of each category in terms of structural and textural features: It is the variance of structural features. It is the variance of texture features; Map category statistics to fusion weights: These are the weights of control category c in the structural branch features and the weights of control category c in the texture branch features, respectively. Tail categories rely more on structural features, while head categories rely more on texture features. By fusing texture representation and structural relationships, a task-adaptive category representation is obtained. for: The category is used to directly mitigate the feature collapse problem of tail categories.

[0013] Furthermore, the dynamic tail-class balancing module performs the following steps: For category First, calculate its sample size. ; Through batch-level variance measurement of structure and the inherent stability of texture features: It is a category The variance of structural features, It is the variance of texture features; Calculate the fusion weights: Where g(⋅) is implemented as a two-layer MLP that includes softmax normalization; By fusing texture representation and structural relationships, the final fused feature representation is as follows: The fused feature representations are input into the classifier.

[0014] Furthermore, structural features guide the construction of category prototypes, which are then refined through a multi-stage geometric aggregation and diffusion process, specifically including: and Representing categories The Middle Structural and textural features of each sample; Variance-weighted structured pooling: categorical The initial structural prototype is defined as follows: in It is the stability weight, calculated as follows: and Representing categories The structural characteristics, mean and variance; As a category Structural anchor points; Within each category c, a kNN graph is constructed based on structural features: It is the set of structural features of all samples in category c, representing the nodes in the graph. It is the similarity matrix between samples within category c, representing the adjacency matrix of the graph, constructed using similarity metrics; Encode geometric similarity using edge weights: It is a hyperparameter that controls the smoothness of Gaussian similarity and affects the calculation of similarity. They are the first in category c and the Structural characteristics of each sample; Detailed structural prototype Represented as: in These are learnable parameters; Calculate the structural affinity between category prototypes: in, It is the structural prototype of category c after refinement; category k with similar geometric structure obtains higher affinity weight; Tail categories borrow structural information from geometrically similar categories: Where γ' is the diffusion coefficient. For normalized cross-class affinity weights; Given the final structural prototype The fused features Alignment to its prototype enhances representation learning; For each sample , It is a sample The corresponding category label indicates the category c to which the sample belongs. Calculate the prototype alignment loss: The classifier's output is further adjusted based on prototype similarity: in Controlling prototype-based reinforcement, It is the weight vector of the classifier, used to weight and fuse features, and affects the output of the classifier; The complete loss function is: in , , These are the coefficients for balancing structural consistency, texture differentiation, and prototype alignment, respectively.

[0015] The beneficial effects of this application are as follows: Identifying the information imbalance between structure and texture is the core bottleneck of long-tailed SAR recognition, and revealing that the performance degradation of tail classes is mainly due to texture instability, rather than simply insufficient samples. A dual-branch SAR characterization framework is proposed, which achieves explicit decoupling and information preservation of structure and texture through a frequency-domain-constrained structure branch and a noise-aware texture branch. A Dynamic Tail Class Balancing Module (DTBM) was designed, which can adaptively adjust the fusion ratio of the two branches based on class density and representation stability. The first structure-driven prototype system for tail classes is proposed, which can form a stable semantic structure, thereby significantly improving the recognition performance of tail classes with few samples. Attached Figure Description

[0016] Figure 1 This is a flowchart of the target classification method provided in the embodiments of this application; Figure 2 These are the results of the comparative tests in this application. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] The Structure-Texture Dual Branch Network (STDB) proposed in this application is a physically motivated framework with information-preserving properties designed to address the long-tail imbalance problem in SAR target recognition. Unlike previous methods, which typically manipulate the data distribution at the image level or mix differential scattering behaviors in a unified feature space, STDB explicitly models the inherent hierarchical characteristics of SAR images: it decomposes the features of the backbone network into structural and texture branches. Subsequently, a Dynamic Tail Class Balancing Module (DTBM) adaptively fuses these two branches based on class density and the stability of feature representations.

[0019] The SAR image is represented as: The corresponding backbone network feature map is denoted as... in, These represent the spatial height, width, and passage dimension, respectively. This is used to represent the features extracted from the original image x, and this process is usually accomplished by the backbone network of a convolutional neural network (CNN).

[0020] This application follows the following information flow: given initial features Then, frequency-aware feature decomposition is performed first: in and These represent the feature responses that are structure-dominant and texture-dominant, respectively.

[0021] Subsequently, structural branching extracts stable geometric priors: The texture branch learns texture semantics that are noise-aware and have reliability weights: Finally, the Dynamic Tail Class Balancing Module (DTBM) balances the categories. Predicting category-related fusion weights ,form: The fused representation is then input into a classifier for classification. Further details on how structural paths can be extended through a structure-driven prototype system will be provided later.

[0022] In one embodiment, structural branching extracts structural representations constrained by the frequency domain.

[0023] SAR images consist of physically distinct scattering components: structural information (target outline, main scattering centers, coarse-scale topology) and textural information (fine-scale micro-scattering and speckle noise). Structural cues are primarily distributed in the low-frequency band and remain highly stable across different viewing angles. Therefore, isolating these structural components from the high-frequency components can prevent high-frequency perturbations from disrupting the semantic representation of tail categories.

[0024] Frequency domain constrained structural decomposition: Initial features The two-dimensional discrete Fourier transform (DFT) is denoted as A learnable low-pass mask We obtain it from the following formula: in It is a lightweight CNN. This refers to the sigmoid function.

[0025] The characteristic of structural dominance is defined as follows: Where ⊙ represents element-wise multiplication.

[0026] A regular expression term is used to encourage consistent low-frequency filtering behavior of the low-pass mask: in Used to suppress high-frequency leakage, thereby preventing erroneous texture components from entering structural branches.

[0027] Structural Topology Encoder (STE): To capture multi-scale topology, a two-module STE is constructed: (1) Local topology module: Through multi-scale depth separable convolution, this module can extract stable target contours and the neighborhood of the scattering center.

[0028] Extract the strongest scattering centers And construct a topology graph. The set of nodes Spatial coordinates of the scattering center, adjacency matrix Defined as: It is a hyperparameter used to control the decay rate of similarity in the adjacency matrix. Specifically, It controls how the distance between two nodes affects their connection strength. Smaller This results in strong connections only between very close scattering centers, while larger ones... It will maintain the connection between scattering centers at a distance.

[0029] 2) Global topology module: A graph neural network (GNN) layer is used to encode global structural relationships. in A pyramid representing multi-scale structural features.

[0030] Structural Stability Loss: Due to insufficient sample size, tail categories often exhibit unstable structural semantics. Therefore, this application employs EMA (Exponential Moving Average) consistency constraints: This loss can stabilize the representation of structural features during training.

[0031] In one embodiment, the texture branch performs noise-aware self-discriminative texture learning: Textures carry high-frequency characteristics of the target material and its micro-geometry, but they are also highly sensitive to speckle noise, imaging artifacts, and class imbalance. Without proper modeling, textures of tail classes can easily degenerate into noise-dominated responses.

[0032] Bandpass texture decomposition: Define a complementary, learnable bandpass mask. Its form is: in For lightweight CNNs, This refers to the sigmoid function.

[0033] This leads to the texture-dominant feature: To ensure the bandpass filter has a clear effective passband, a regularization term is added: This feature enhances the bandpass characteristics of the filter, making the filtering transition sharper and thus more clearly distinguishing texture details.

[0034] Self-Discriminating Texture Encoder (SDTL): SDTL consists of the following components: (1) Dual-path residual encoder: one path models local high-frequency features through depthwise separable convolution; the other path captures global scattering correlations through nonlocal attention.

[0035] The overall texture encoding is represented as: (2) Reliability estimation: To reduce noise response, a reliability score is calculated for each token: The final texture representation is as follows: Self-Discriminative Contrastive Loss: The loss function is as follows: Positive sample pairs are constructed from highly reliable tokens, and negative sample pairs are constructed from unreliable or cross-category tokens. This loss prevents the model from mimicking tail samples dominated by noise.

[0036] In one embodiment, the Dynamic Tail Class Balancing Module (DTBM) includes: Different categories contribute varying proportions to structure and texture. Head categories typically contain rich textures, while tail categories rely more on stable structural information. Therefore, using a fixed fusion strategy is not ideal. In this embodiment, the goal of DTBM is to automatically learn the specific fusion coefficients for each category.

[0037] Class density estimation: Let the classes be... The number of samples is Its normalized density is defined as: Representation Stability Estimation: Calculates the stability of each category across structural and textural features. Larger variance implies semantic instability.

[0038] Adaptive Stream Gate: Maps class statistics to fusion weights. Tail categories will get a larger (That is, it relies more on structural features), while the head category will get a larger [status]. (That is, it relies more on texture features).

[0039] Final Fusion: The category representation of task adaptation is as follows: This balanced representation can directly alleviate the feature collapse problem of tail categories.

[0040] In one embodiment, the Dynamic Tail Class Balancing Module (DTBM) includes: Different categories exhibit significant differences in reliability distribution across structure and texture. Head categories typically possess rich texture information, while tail categories rely more on stable structural features. Therefore, static fusion strategies cannot adapt to this variation. DTBM aims to learn a category-adaptive fusion strategy.

[0041] Class density estimator: For category First, calculate its sample size. This metric reflects the expected reliability of texture feature learning.

[0042] Representation stability estimator: Through batch-level variance measurement of structure and the inherent stability of texture features: A larger variance indicates higher semantic instability.

[0043] Adaptive stream gate: DTBM calculates fusion weights: g(⋅) is implemented as a two-layer MLP that includes softmax normalization.

[0044] Tail categories will naturally receive a larger (More dependent on structure), while the head category will gain greater [rewards / advantages]. (More dependent on texture).

[0045] Final feature fusion: The final fused features are represented as follows: This feature will be fed into the classifier. This design ensures that each category achieves an optimal balance between structure and texture, resulting in a stable and discriminative representation.

[0046] While the bi-branch representations in the aforementioned embodiments effectively decouple stable structures from fragile textures, tail categories still face the problem of semantic incompleteness: their limited sample size cannot cover the inherent geometric manifold of the category. This typically leads to weak decision boundaries, classifier output bias, and a lack of reliable category anchors.

[0047] To address this fundamental problem, one embodiment proposes a Structure-Driven Prototyping System (SDPS). This system primarily guides the construction of category prototypes based on structural features and further refines the prototypes through a multi-stage geometric aggregation and diffusion process.

[0048] SDPS is designed based on a key observation: for long-tailed categories, structural features have much lower intra-class variance than texture features. Therefore, the prototypes generated from structural features can provide consistent and stable semantic anchors, allowing tailed categories to build semantic spaces based on these anchors; while textures are only used to supplement details without dominating the overall embedding space.

[0049] and Representing categories The Middle Structural and textural features of each sample.

[0050] Structural Prototype Initialization: Taking the average value directly from structural features can still introduce bias when the sample size is very small. Therefore, this embodiment uses variance-aware aggregation to construct the prototype, in order to highlight stable structural components and suppress sample-specific noise.

[0051] Variance-weighted structural pooling: This method groups categories... The initial structural prototype is defined as follows: in It is the stability weight, calculated as follows: and Representing categories The structural characteristics, mean and standard deviation.

[0052] The benefits of this design include: reducing the weight of unstable structural samples; reducing tail category prototype initialization bias; and maintaining structural semantic consistency even with very few samples.

[0053] therefore, Can be used as a category Structural anchor.

[0054] Geometric Prototype Refinement: Tail categories typically only contain incomplete or noisy samples of their true shape manifold. Therefore, relying solely on the obtained initial prototype is still insufficient. To address this, this embodiment proposes a geometry refinement module for aligning the prototype with structural relationships across samples.

[0055] Neighborhood Graph Construction: Within each category c, a kNN graph is constructed based on structural features: Edge weights encode geometric similarity: Prototype Geometric Alignment: The refined structural prototype is represented as follows: in These are learnable parameters.

[0056] This refinement step pulls the prototype toward the center of the geometric neighborhood, preventing it from collapsing to outliers (a common problem with tail classes), and expands the prototype distribution along a reasonable intra-class direction, resulting in smoother and more stable class anchors.

[0057] Prototype Diffusion Across Classes: Even after geometric refinement, the prototypes of tail categories are still limited by data scarcity. To enrich their representation, this embodiment proposes a structure-driven cross-category diffusion mechanism.

[0058] Cross-Class Structural Affinity: Calculate the structural affinity between category prototypes: Among them, category k with similar geometric structure will receive higher affinity weight.

[0059] Semantic Prototype Diffusion: Tail categories borrow structural information from geometrically similar categories: Where γ is the diffusion coefficient. This represents the normalized cross-class affinity weights.

[0060] Tail categories inherit geometric structures from semantically similar head categories; structural features ensure compatibility (textures are not diffused to avoid noise propagation); prototypes thus become “imaginative” anchors that can cover missing geometry.

[0061] Prototype-Guided Training: Given the final structure-driven prototype By fusing the features Alignment to its prototype enhances representation learning.

[0062] Prototype Alignment Loss: For each sample Calculate the prototype alignment loss: Prototype-Regularized Classification: The classifier's output will be further adjusted based on prototype similarity: in Control based on prototype reinforcement.

[0063] The complete loss function is: in , , Used to balance structural consistency, texture differentiation, and prototype alignment.

[0064] This embodiment constructs a unified "structural anchor space", which significantly improves the semantic integrity of all categories and the stability of tail class embedding.

[0065] Figure 2 These are the results of a comparative experiment on the head, tail, and middle classes on the SARShip-LT dataset.

[0066] The beneficial effects of this application are as follows: Identifying the information imbalance between structure and texture is the core bottleneck of long-tailed SAR recognition, and revealing that the performance degradation of tail classes mainly stems from texture instability, rather than simply insufficient samples.

[0067] A dual-branch SAR characterization framework is proposed, which achieves explicit decoupling and information preservation of structure and texture through a frequency-domain-constrained structure branch and a noise-aware texture branch.

[0068] A Dynamic Tail Class Balancing Module (DTBM) was designed, which can adaptively adjust the fusion ratio of the two branches based on class density and representation stability.

[0069] The first structure-driven prototype system for tail classes is proposed, which can form a stable semantic structure, thereby significantly improving the recognition performance of tail classes with few samples.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A SAR long-tail image target classification method based on topology, characterized in that, Includes the following steps: The SAR image is represented as: The corresponding backbone network feature map is denoted as: in, These represent the image space height, width, and channel dimensions, respectively. Represents the feature map of the backbone network; Given initial features Then, perform frequency-aware feature decomposition: in and These represent the feature responses that are structure-dominant and texture-dominant, respectively. This indicates the operation of performing frequency domain feature decomposition, which decomposes the input features. It is broken down into two parts: structure and texture. Subsequently, structural branches extract stable geometric priors. : Indicates structural branches; Texture branch learning possesses noise perception capabilities and yields texture semantics with reliability weights. : Indicates texture branching; Ultimately, the dynamic tail class balancing module balances the categories. Predicting category-related fusion weights ,form: The fused representation is then input into a classifier for target classification in SAR images.

2. The SAR long-tail image target classification method based on topology structure according to claim 1, characterized in that, Performing frequency-aware feature decomposition specifically includes the following steps: Initial features The two-dimensional discrete Fourier transform yields A learnable low-pass mask We obtain it from the following formula: in It is a lightweight CNN. For the sigmoid function; The characteristic of structural dominance is: Where ⊙ represents element-wise multiplication. This means performing an inverse transformation on the frequency domain data to bring the structural information in the frequency domain back to the spatial domain, thus obtaining the final structural dominance feature; Use a regular expression to encourage consistent low-frequency filtering behavior in the low-pass mask: in Used to suppress high-frequency leakage, thereby preventing erroneous texture components from entering structural branches; Extract the strongest scattering centers And construct a topology graph : The set of nodes Spatial coordinates of the scattering center, adjacency matrix Defined as: They are the scattering centers and spatial coordinates, It is a hyperparameter used to control the decay rate of similarity in the adjacency matrix; Using a graph neural network (GNN) to encode global structural relationships: in Pyramids representing multi-scale structural features; To address the issue of unstable structural semantics in tail categories due to insufficient samples, an exponential moving average is incorporated as a consistency constraint loss: These are the structural features extracted by structural branches, representing the geometric information extracted by the model through structural branches. It is an operator used to freeze gradients, preventing structural features from changing during training; This loss is used to stabilize the representation of structural features during training.

3. The SAR long-tail image target classification method based on topology structure according to claim 2, characterized in that, Texture branch learning possesses noise perception capabilities and yields texture semantics with reliability weights. Specifically, it includes: Use a complementary, learnable bandpass mask Bandpass texture decomposition is performed, and its form is as follows: in For lightweight CNNs, For the sigmoid function; This yields the texture-dominated feature. : To ensure the bandpass filter has an effective passband, a regularization term is added: This feature is used to enhance the bandpass characteristics of the filter, thereby distinguishing texture details; Encoding is performed using a dual-path residual encoder, where one path models local high-frequency features through depthwise separable convolutions, and the other path captures global scattering correlations through nonlocal attention. The overall texture encoding is represented as: Indicates local high-frequency features, This refers to non-local attention, which focuses on the relationships between distant pixels in an image to enhance global perception. To mitigate the noise response, a reliability score is calculated for each token. : This represents a multilayer perceptron. It is the feature representation of each token in the texture branch, and each token corresponds to a local region in the image; The final texture representation is as follows: The loss function is as follows: Positive sample pairs are constructed from reliable tokens, and negative sample pairs are constructed from unreliable or cross-category tokens. These are samples and Texture features, It's a temperature parameter used to control the smoothness of the similarity score. It is a similarity function. It is the index of the negative sample, representing a sample that is related to... Dissimilar tokens; This loss function is used to prevent the model from mimicking tail samples dominated by noise.

4. The SAR long-tail image target classification method based on topology structure according to claim 3, characterized in that, The dynamic tail-class balancing module performs the following steps: category The number of samples is Its normalized density for: It is the number of samples in each category of the dataset; Calculate the stability of each category in terms of structural and textural features: It is the variance of structural features. It is the variance of texture features; Map category statistics to fusion weights: These are the weights of control category c in the structural branch features and the weights of control category c in the texture branch features, respectively. Tail categories rely more on structural features, while head categories rely more on texture features. By fusing texture representation and structural relationships, a task-adaptive category representation is obtained. for: The category is used to directly mitigate the feature collapse problem of tail categories.

5. The SAR long-tail image target classification method based on topology structure according to claim 3, characterized in that, The dynamic tail-class balancing module performs the following steps: For category First, calculate its sample size. ; Through batch-level variance measurement of structure and the inherent stability of texture features: It is a category The variance of structural features, It is the variance of texture features; Calculate the fusion weights: Where g(⋅) is implemented as a two-layer MLP that includes softmax normalization; By fusing texture representation and structural relationships, the final fused feature representation is as follows: The fused feature representations are input into the classifier.

6. The SAR long-tail image target classification method based on topology structure according to claim 5, characterized in that, The construction of category prototypes is guided by structural features, and the prototypes are further refined through a multi-stage geometric aggregation and diffusion process, specifically including: and Representing categories The Middle Structural and textural features of each sample; Variance-weighted structured pooling: categorical The initial structural prototype is defined as follows: in It is the stability weight, calculated as follows: and Representing categories The structural characteristics, mean and variance; As a category Structural anchor points; Within each category c, a kNN graph is constructed based on structural features: It is the set of structural features of all samples in category c, representing the nodes in the graph. It is the similarity matrix between samples within category c, representing the adjacency matrix of the graph, constructed using similarity metrics; Encode geometric similarity using edge weights: It is a hyperparameter that controls the smoothness of Gaussian similarity and affects the calculation of similarity. They are the first in category c and the Structural characteristics of each sample; Detailed structural prototype Represented as: in These are learnable parameters; Calculate the structural affinity between category prototypes: in, It is the structural prototype of category c after refinement; category k with similar geometric structure obtains higher affinity weight; Tail categories borrow structural information from geometrically similar categories: Where γ' is the diffusion coefficient. For normalized cross-class affinity weights; Given the final structural prototype The fused features Alignment to its prototype enhances representation learning; For each sample , It is a sample The corresponding category label indicates the category c to which the sample belongs. Calculate the prototype alignment loss: The classifier's output is further adjusted based on prototype similarity: in Controlling prototype-based reinforcement, It is the weight vector of the classifier, used to weight and fuse features, and affects the output of the classifier; The complete loss function is: in , , These are the coefficients for balancing structural consistency, texture differentiation, and prototype alignment, respectively.

Images