Few-shot ISAR target recognition method based on TA-SCSB

Abstract

This invention belongs to the field of target detection and classification technology, specifically relating to a few-shot ISAR target recognition method based on TA-SCSB. The method includes the following steps: Step 1: Structural prior feature extraction stage; Step 2: Task-adaptive structural statistical modeling stage; Step 3: Structural channel selection bottleneck stage; Step 4: Determination of structure-aware manifold feature inference mechanism stage; Step 5: Adaptive classification and stability joint optimization stage. This method introduces a task-level structural statistical modeling and channel selection coupling mechanism into few-shot ISAR recognition for the first time. It dynamically generates channel weight distributions based on the structural statistical features (variance, information entropy, gradient strength) of supporting samples, achieving task-adaptive selection and weighting of feature channels.

Description

Technical Field

[0001] This invention belongs to the field of target detection and classification technology, specifically involving a few-sample ISAR target recognition method based on TA-SCSB (Task-Adaptive Structural Channel-Selection Bottleneck, TA-SCSB). It can be widely applied to ISAR ship target recognition, maritime surveillance and defense reconnaissance, UAV remote sensing intelligent analysis, target detection and classification under complex sea conditions, etc. It is particularly suitable for high-precision ISAR intelligent recognition tasks under few-sample conditions, providing technical support for scenarios such as national defense security, maritime law enforcement and maritime management. Background Technology

[0002] Inverse Synthetic Aperture Radar (ISAR) imaging technology can acquire two-dimensional structural information of moving targets without active cooperation, and has the advantage of all-weather and all-time observation. It is widely used in the identification and monitoring of maritime and air targets such as ships and aircraft.

[0003] In ISAR image recognition tasks, traditional methods mainly rely on the target's outline, scattering center distribution, and local structural features to distinguish different ship types. However, due to the influence of factors such as target attitude changes, sea surface disturbances, defocusing, and signal-to-noise ratio fluctuations during the ISAR imaging process, image quality and structural stability vary greatly, making it difficult to guarantee the recognition accuracy and stability of the model in complex environments.

[0004] In recent years, with the development of deep learning technology, researchers have attempted to apply convolutional neural networks (CNNs), Transformer structures, and multi-scale feature fusion mechanisms to ISAR target recognition, aiming to automatically learn the mapping relationship between scattering features and structural semantics. These methods show good recognition performance when there is sufficient data, but when the number of samples is limited or the sample categories are unevenly distributed, the networks often suffer from feature overfitting and weak generalization ability, failing to meet the needs of actual maritime monitoring systems for high-reliability recognition with few samples.

[0005] Furthermore, ISAR images exhibit significant structure-dominant characteristics (such as contour morphology, component distribution, and concentrated scattering intensity areas), while the feature extraction process of most existing depth models is fixed and does not dynamically adjust the feature channel structure according to task differences, resulting in significant fluctuations in recognition performance under different tasks. Therefore, how to adaptively select the most discriminative structural feature channels under limited sample conditions has become an important research direction in the field of ISAR intelligent recognition.

[0006] Currently, the main types of related methods in the field of ISAR target recognition and few-shot learning are as follows:

[0007] (1) Static recognition method based on convolutional features

[0008] Early studies often used convolutional neural networks (CNNs) or residual networks (ResNets) to directly extract feature vectors from ISAR images and then achieved target classification through fully connected layers. These methods, such as ISAR-CNN and DeepISAR, achieve high recognition accuracy when there are sufficient samples. However, because their convolutional kernel channel weights are fixed, they cannot adaptively optimize features to accommodate task differences or changes in image quality. Therefore, they are prone to overfitting under conditions of few samples, noise interference, and cross-viewpoint scenarios.

[0009] (2) Improved model based on multi-scale feature fusion

[0010] To enhance the discriminative power and robustness of models, some studies have introduced multi-scale convolutional fusion, attention mechanisms, or feature pyramid structures. For example, multi-scale feature pyramids (MSF-ISAR) are used to jointly represent local features and global semantics, or spatial attention is superimposed on feature layers to strengthen salient scattering regions. However, the channel selection strategies of these methods remain fixed, lacking dynamic selection mechanisms between channels, and are still susceptible to feature redundancy in the identification of different tasks or ship types.

[0011] (3) Structure-Semantic Joint Modeling Method Based on Manifold Reasoning (REMI)

[0012] REMI (Representation Enhancement via Manifold Inference) is a representative ISAR image recognition method in recent years. This method simultaneously represents the structural features of the target at both the global and local levels through multi-scale image transformation, feature fusion, and manifold space modeling, achieving high recognition accuracy. The core idea of ​​REMI is to maintain the geometric similarity between samples through manifold constraints in a high-dimensional feature space, and to enhance structural semantic information during the fusion stage.

[0013] However, the REMI framework has the following shortcomings:

[0014] A. The feature channel structure is fixed during the training phase, lacking task-level adaptability, and the feature expression is inconsistent under different tasks;

[0015] B. The use of full-channel features in the manifold modeling stage introduces redundant information and noise interference, affecting the learning stability of small sample tasks;

[0016] C. The dynamic adjustment mechanism of the structural channel under different viewing angles or noise conditions is not considered, and the identification results are less stable in complex sea conditions.

[0017] (4) ISAR identification method based on few-shot metric learning

[0018] In recent years, the concept of Few-shot Learning (FSL) has been introduced into ISAR scenarios, with typical methods including ProtoNet, MatchingNet, and RelationNet. These methods achieve inter-class differentiation under few-shot conditions through metric space learning, but most do not incorporate the structural information specific to ISAR, relying only on texture or semantic differences, thus limiting recognition accuracy in practical applications.

[0019] The shortcomings and need for improvement of existing technologies

[0020] Based on the above research, existing ISAR target identification methods mainly have the following shortcomings:

[0021] (1) Lack of task adaptability: The existing network structure adopts fixed channel settings in the feature extraction stage, which cannot dynamically select the most discriminative structural channel for different tasks and imaging conditions.

[0022] (2) Feature redundancy and channel interference problem: During the manifold inference or multi-scale fusion stage, a large number of irrelevant or noisy channels are activated at the same time, which makes it difficult for training with few samples to converge and the model has poor stability.

[0023] (3) Insufficient structural consistency: The structure of ISAR images varies significantly under different viewpoints and noise conditions. Existing methods have failed to maintain the consistency of structural feature selection, resulting in large fluctuations in recognition results.

[0024] (4) Generalization difficulties in small sample scenarios: The lack of targeted optimization mechanisms makes it difficult for the model to effectively learn the structural commonalities within the task when the training samples are limited.

[0025] The closest existing technology to this invention is the REMI (Representation Enhancement via Manifold Inference) method, which achieves structure-semantic fusion through manifold modeling, but does not include a task-adaptive structure optimization mechanism. This invention innovatively introduces a Task-Adaptive Structure Channel Selection Bottleneck Network (TA-SCSB), which utilizes the structural statistical features of supporting samples to generate dynamic gating weights. This achieves a combination of sparse selection of structural channels and manifold optimization at the bottleneck layer, significantly improving recognition stability and generalization performance under conditions of few samples.

[0026] Despite significant progress in few-shot ISAR target recognition technology in recent years, with related research proposing various innovative approaches in feature reconstruction, manifold reasoning, graph structure modeling, attention enhancement, and multi-view representation, the following common problems still exist overall:

[0027] (1) Lack of task adaptability in feature extraction: Most methods (such as REMI, Few-Shot Multi-StaticISAR, Patch Graph Transformer, etc.) use fixed feature channels or static fusion methods in the network structure, and fail to dynamically adjust the channel selection according to the feature distribution differences of different tasks, resulting in a significant decrease in recognition performance when the model is transferred or the sample distribution changes.

[0028] (2) Insufficient utilization of structural features: Existing models generally focus on enhancing global semantics or texture details, but do not make sufficient use of structural features in ISAR images that reflect geometric shape and scattering center layout. This makes it difficult to effectively express the morphological differences and component relationships of targets, affecting the feature discrimination under small sample conditions.

[0029] (3) Poor stability under cross-viewpoint and noise conditions: Under different observation angles, defocus or large fluctuations in signal-to-noise ratio, the structural response of existing methods is unstable and the model output is sensitive to changes in viewpoint. Especially in low-sample tasks, recognition fluctuations and performance degradation are likely to occur.

[0030] (4) Feature redundancy and channel interference are prominent issues: Most baseline methods do not introduce channel-level screening mechanisms in the feature fusion stage. A large number of redundant or inefficient features participate in the calculation at the same time, which not only increases the complexity of the model, but also weakens the expressive power of the effective channels, affecting the convergence and generalization of few-sample tasks.

[0031] (5) Insufficient stability and interpretability: Most current few-sample recognition methods focus on overall performance improvement and lack constraints on structural channel selection and task stability. They are difficult to explain the reasons for the differences in channel activation between different tasks, which limits their application in engineering deployment and reliability verification. Summary of the Invention

[0032] (a) Technical problems to be solved

[0033] The technical problem to be solved by this invention is:

[0034] (1) Insufficient recognition generalization ability under conditions of few samples:

[0035] In the existing field of ISAR target recognition, representative methods include REMI (Representation Enhancement via Manifold Inference). [1]While exhibiting good performance in multi-scale feature fusion and manifold modeling, its fixed feature channel structure struggles to adapt to the feature distributions of different tasks in scenarios with limited sample sizes, leading to problems such as overfitting, decreased discriminative ability, and unstable cross-view recognition performance. This invention proposes a Task-Adaptive Structural Channel-Selection Bottleneck (TA-SCSB) method. By introducing a task-aware structural channel selection mechanism, the model dynamically adjusts feature channel weights based on the statistical characteristics of a small number of supporting samples. This enables adaptive extraction and optimization of structural information under limited sample conditions, effectively improving recognition accuracy and cross-task generalization ability.

[0036] (2) The problem of poor adaptability of fixed structure feature extraction to multi-task applications:

[0037] The existing REMI framework uses static structural channels for feature extraction, which cannot differentiate the modeling of scattering distribution, morphological features, and structural complexity differences among different ship targets during imaging, resulting in non-targeted feature representation across tasks. This invention designs a task-adaptive structural channel selection bottleneck module based on REMI. It generates channel weight distributions using the structural statistical features of a small number of supporting samples, enabling dynamic selection and sparse activation of feature channels. This allows the model to automatically focus on the most discriminative structural features under different tasks, significantly improving the model's feature utilization efficiency and task adaptability.

[0038] (3) The stability of structural features under complex perspectives and noise interference:

[0039] In real ISAR imaging environments, factors such as target attitude changes, defocusing, and sea surface background noise can cause structural features to become unstable, thus affecting recognition accuracy. Traditional methods struggle to guarantee consistency in structural feature extraction under different viewpoints or noise conditions. This invention introduces stability regularization and structural consistency constraints into the bottleneck layer. While maintaining the stability of key structural channel selection, it enhances the model's robustness to scattering center drift, image blurring, and local noise disturbances, ensuring stable and reliable recognition results.

[0040] (II) Technical Solution

[0041] To address the aforementioned technical problems, this invention provides a few-sample ISAR target identification method based on TA-SCSB, the method comprising the following steps:

[0042] Step 1: Structural prior feature extraction stage;

[0043] Step 2: Task-adaptive structural statistical modeling stage;

[0044] Step 3: Bottleneck selection stage of structural channel selection;

[0045] Step 4: Determining the structure-aware manifold feature reasoning mechanism;

[0046] Step 5: Adaptive classification and stability joint optimization stage.

[0047] (III) Beneficial Effects

[0048] Compared with existing technologies, the Task-Adaptive Structural Channel-Selection Bottleneck (TA-SCSB) proposed in this invention systematically and innovatively improves upon existing few-shot ISAR target recognition methods, addressing issues such as fixed feature channels, poor task adaptability, unstable structural representation, and performance degradation across viewpoints. Its key improvements and areas for protection are mainly reflected in the following five aspects:

[0049] (1) Task-adaptive structure channel selection mechanism (core innovation)

[0050] For the first time, a task-level structural statistical modeling and channel selection coupling mechanism is introduced in few-sample ISAR identification. By dynamically generating channel weight distribution through the structural statistical features (variance, information entropy, gradient strength) of the supporting samples, the task adaptive selection and weighting of feature channels can be achieved.

[0051] The points to be protected are: the construction method of the task description vector, the Gumbel-Softmax gating channel selection mechanism, the temperature annealing scheduling strategy and its joint optimization method with task statistical modeling.

[0052] (2) Differentiable sparse activation structure of the Structural Channel Selection Bottleneck (SCSB) module

[0053] A bottleneck layer structure based on differentiable gating is proposed, which can dynamically filter key structural channels in the feature flow, taking into account both sparsity and continuous trainability, and achieving a balance between information compression and structure preservation.

[0054] Points to be protected: Design of the differentiable gated function of the bottleneck layer, and the joint form of the channel sparse constraint term and the stability regularization term.

[0055] (3) Structure-Aware Manifold Feature Inference (SAMI) Mechanism

[0056] In the bottleneck feature subspace, a structure-weighted manifold adjacency relationship is established, and cross-sample geometric consistency and structural self-alignment are achieved through density self-correction propagation, thereby significantly enhancing the stability and discriminativeness of feature distribution under the condition of few samples.

[0057] To protect the points: the construction method of the structure-aware adjacency matrix, the design of the density self-correcting weight mapping function and the structural consistency constraint loss function.

[0058] (4) Joint regularization optimization mechanism for cross-perspective stability and channel sparsity

[0059] In the model training phase, cross-view stability regularization and channel sparsity constraints are introduced to maintain the consistency of gated channel distribution and suppress redundancy under multi-view imaging conditions, thereby improving the model's generalization and physical interpretability.

[0060] The points to be protected are: cross-view gating weight difference constraint formula and multinomial balanced optimization structure of joint loss function.

[0061] (5) End-to-end adaptive classification and stability joint optimization framework (ACSC)

[0062] A unified optimization framework integrating task-adaptive classification, manifold preservation, and channel stability constraints is constructed to achieve closed-loop adaptive learning from structural feature extraction to classification inference.

[0063] The points to be protected are: a collaborative training method for prototype metric learning and task adaptive gating mechanism, a dynamic weight scheduling strategy, and a full-process end-to-end optimization framework design.

[0064] Compared with the closest prior art (REMI, 2024), the advantages and technical effects of the present invention are as follows:

[0065] Existing REMI methods rely on fixed multi-scale embedding structures and static manifold constraints in few-sample ISAR identification. While possessing some discriminative ability, their feature representations are prone to instability and insufficient task adaptability under conditions of scarce samples, large viewpoint variations, or strong noise interference. To overcome these problems, the TA-SCSB (Task-Adaptive Structural Channel-Selection Bottleneck) proposed in this invention achieves significant improvements in structural modeling, feature selection, and optimization mechanisms. Specific technical advantages are as follows:

[0066] (1) Task-adaptive structural feature selection mechanism

[0067] Compared to REMI's fixed channel allocation method, TA-SCSB adaptively allocates channel weights through mission-level structural statistical modeling, enabling the network to dynamically focus on the most discriminative structural information for different ship types, attitudes, and mission conditions.

[0068] Technical results: In the 5-shot setting, the average classification accuracy improved from 80.5% in REMI to 86.2%, and the cross-task generalization accuracy improved by approximately 6.3%.

[0069] (2) Structural preservation and geometric consistency are significantly enhanced.

[0070] The Structure-Aware Manifold Inference (SAMI) of this invention introduces structure-weighted adjacency relations within the bottleneck subspace, enabling the model to maintain structural topological stability under different perspectives and signal-to-noise conditions.

[0071] Technical results: The Structural Consistency Index (SCI) is improved by approximately 12% compared to the REMI, and the identification stability variance is reduced by 28% at low signal-to-noise ratios (SNR = 5dB).

[0072] (3) Improved robustness of cross-perspective recognition

[0073] By introducing cross-view stability regularization in the optimization, the model can still maintain a consistent channel activation distribution under multi-angle imaging conditions.

[0074] Technical results: With a viewing angle shift of ±15°, the decrease in recognition rate was reduced from 10.8% in REMI to 5.7%.

[0075] (4) The model has stronger physical interpretability.

[0076] The channel selection weights are directly related to the structural statistical characteristics and can be mapped to physical scattering regions such as the ship's hull and island, giving the model's decision-making process a clear physical meaning.

[0077] Technical effect: The overlap (IoU) between the salient regions of the model and the manually annotated key parts was improved from 0.61 in REMI to 0.77.

[0078] (5) Optimized convergence for greater stability and computational efficiency

[0079] By using Adaptive Classification and Stability Joint Optimization (ACSC), we can achieve collaborative updates of multiple loss terms in a single end-to-end framework, thus avoiding the instability of multi-stage training.

[0080] Technical results: Training convergence speed is improved by about 25%, accuracy variance in the validation phase is reduced by 32%, and overall inference time is reduced by about 18%. Attached Figure Description

[0081] Figure 1 This is a schematic diagram illustrating the principle of the technical solution of the present invention. Detailed Implementation

[0082] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0083] This invention provides a few-sample ISAR target recognition method based on TA-SCSB (Task-Adaptive Structural Channel-Selection Bottleneck, abbreviated as TA-SCSB). The method is based on the REMI framework and introduces a task-adaptive structural channel selection mechanism in the process of structural feature extraction and manifold inference. It can dynamically adjust the channel weights according to the structural statistical information of a small number of support samples, so as to achieve highly robust recognition of multi-view, few-sample, and high-noise ISAR images.

[0084] The method includes the following steps:

[0085] Step 1: Structure Prior Extraction

[0086] In few-shot ISAR target recognition tasks, traditional deep feature extraction methods often rely on large-scale labeled data to learn the salient features of the target. However, when the sample size is limited, the network is prone to overfitting or texture dependence. Unlike natural images, the discriminative power of ISAR images mainly comes from the geometric structure of the target and the distribution of scattering centers. Therefore, in few-shot scenarios, extracting multi-scale structural features with high structural saliency and low noise sensitivity is particularly important.

[0087] Based on the above understanding, this step aims to extract a set of candidate structural feature channels containing rich morphological information from the original ISAR image. These channels retain structural information that is highly correlated with the target morphology in the spatial, frequency, and scattering levels, providing structural prior support for task-adaptive channel selection and manifold feature inference in subsequent stages.

[0088] The implementation of the structural prior feature extraction stage involves three sub-steps:

[0089] Step 1.1: Multi-scale structural deconstruction;

[0090] Step 1.2: Physical-Statistical Dual Feature Fusion;

[0091] Step 1.3: Channel Standardization and Response Enhancement;

[0092] The input for this stage is a few-sample ISAR image I∈R H×W The output is a set of candidate channels for the structure, C = {C1, C2, ..., C}. M}; Each channel C i This corresponds to a salient structural feature in a physical or statistical sense.

[0093] Step 1.1: Multi-scale structural deconstruction

[0094] To characterize the structural patterns of ISAR images at different scales, this step designs a multi-scale structure deconstruction module (MSD); specifically, the input image I is processed through three spatial frequency filtering kernels at different scales. low ,f mid ,f high The target is decomposed and its low-frequency morphology, main outline, and high-frequency scattering structure are extracted separately.

[0095]

[0096] in, S1 represents a convolution or bandpass filtering operation at the k-th scale; S2 preserves large-scale geometric information (low frequency), S2 captures the mid-scale structural outline, and S3 focuses on high-frequency details and scattering center features.

[0097] The feature maps from the three scales are then concatenated and spatially normalized to form the basic multi-scale structural representation:

[0098]

[0099] Step 1.2: Physics-Statistics Dual Feature Fusion

[0100] ISAR images not only reflect the geometric contours of the target, but also implicitly contain physical scattering features (such as echo intensity, phase distribution, energy concentration, etc.). In order to fully express the structural information, this step introduces a physical-statistical dual feature fusion unit (PSDF) after the MSD output to fuse the target's physical scattering response and statistical texture index.

[0101] Specifically, we define the local response vector for each spatial location (x, y):

[0102] r(x,y)=[E(x,y),G(x,y),T(x,y)]

[0103] Where E(x,y) is the scattering energy density; G(x,y) is the local gradient magnitude; T(x,y) is the texture variance; then global channel aggregation is calculated:

[0104]

[0105] Among them, Ω i Let w(x,y) be the receptive field region of channel i, and w(x,y) be the weight mapping. Energy conservation is ensured by softmax normalization.

[0106]

[0107] In this way, local scattering modes can be converged according to energy significance to obtain channel features with physical interpretability;

[0108] Step 1.3: Channel Standardization and Response Enhancement

[0109] To prevent numerical differences between features at different scales or in different regions from causing training instability, this step introduces a two-step enhancement strategy after the fusion result:

[0110] Step 1.3.1: Channel Standardization

[0111]

[0112] Wherein, μ(C i ) and σ(C i ) represent the channel mean and standard deviation, respectively, and ∈ is the stability constant;

[0113] Step 1.3.2: Structural Response Enhancement

[0114] To enhance the response intensity of critical structural regions (such as hull edges, masts, and island superstructures), a weighted function based on structural attention is introduced:

[0115]

[0116] Among them, A i The structural saliency map is generated jointly by Sobel edge gradients and Laplacian second-order information, with β being the enhancement coefficient (empirically taken as 0.2–0.4); the final output is a set of candidate channels for the structure.

[0117] C = {C ' 1,C ' 2,…,C ' M}

[0118] In the structural prior feature extraction stage, a candidate structural channel set containing target morphology and scattering information is constructed through multi-scale structural deconstruction and physical-statistical fusion, providing a stable, interpretable and robust structural representation basis for the subsequent task adaptive channel selection bottleneck (SCSB).

[0119] Step 2: Task Adaptive Structural Statistical Modeling Stage

[0120] (Task-AdaptiveStatisticsModeling)

[0121] In few-sample ISAR target recognition, one of the core challenges faced by the model is the high heterogeneity of structural distribution among different tasks. For example, different ship types (such as aircraft carriers, destroyers, and cruisers) have significant differences in imaging conditions, attitude angles, scattering center layouts, and structural scales. If a fixed feature extraction channel or a uniform structural weighting method is still used, the model cannot make the optimal selection and representation of structural features for specific tasks, resulting in feature redundancy and a decrease in recognition performance.

[0122] To address this issue, this step proposes a Task-Adaptive Structural Statistical Modeling (TASSM) mechanism. Its core idea is to automatically learn and represent the structural distribution pattern of the current recognition task using statistical information from a small number of support samples. This is achieved by analyzing the candidate structural channel set C = {C1, C2, ..., C...}. M Perform structural statistical modeling to generate task-level description vectors s task This vector can quantitatively reflect the importance and distinctiveness of different channels in the current task, and provide adaptive guidance signals for subsequent Channel Selection Bottleneck (SCSB);

[0123] The task-adaptive structural statistical modeling and solution phase involves three sub-steps:

[0124] Step 2.1: Calculation of local structural indicators: Extract multidimensional statistical indicators that reflect the significance of structural characteristics from each channel;

[0125] Step 2.2: Task-level statistical aggregation: Based on the supporting sample set, the statistical indicators of each channel are weighted and aggregated to form a stable task representation;

[0126] Step 2.3: Feature Normalization and Vector Encoding: Normalize and embed the statistical results to obtain the task description vector s. task ;

[0127] Step 2.1: Calculation of local structural indices

[0128] For each channel C i ∈C, firstly, three types of indicators that can characterize the stability and morphological complexity of the channel structure are extracted:

[0129]

[0130] in, The pixel variance within the channel is used to reflect the intensity fluctuations of the structure distribution; Information entropy describes the uncertainty and complexity of the structure within a channel; The average gradient intensity is used to quantify the sharpness of edges and contours;

[0131] To synthesize these characteristics, a structural significance index for weighted combinations is defined:

[0132]

[0133] Where α1, α2, and α3 are empirical weights (usually taken as 0.4, 0.3, and 0.3), which can be adaptively adjusted during training;

[0134] Step 2.2: Task-level statistical aggregation

[0135] Let the supporting sample set be S = {I1, I2, ..., I...} K} contains K few-sample ISAR images; each sample, after undergoing a structural prior feature extraction stage, yields a set of channel features C. k ={C k,1 ,…,C k,M For each channel, calculate the cross-sample statistic to obtain the task-level average structure index:

[0136]

[0137] To improve the separability between different channels, normalization and relative weight constraints are introduced:

[0138]

[0139] The final task description vector is obtained as follows:

[0140]

[0141] This vector reflects the importance distribution of each channel under the current task and serves as the input condition for the Channel Selection Bottleneck Module (SCSB).

[0142] Step 2.3: Feature Normalization and Vector Encoding

[0143] To further enhance the expressive power of the task description vector, this step designs a lightweight statistical embedding encoder (SEE); the encoding process is as follows:

[0144] h task =σ(W2ReLU(+b2)

[0145] Where W is the learnable weight matrix; b is the bias term; h task The final task structure description is embedded to generate subsequent channel gating distributions;

[0146] In the task adaptive structural statistical modeling stage, by constructing local statistical indicators and task-level structural description vectors, the adaptive capture of task-level structural patterns is realized. This provides a quantitative, interpretable, and differentiable task structure prior for Channel Selection Bottleneck (SCSB), and realizes the key leap from "static channel" to "task-driven channel".

[0147] Step 3: Structural Channel Selection Bottleneck Stage

[0148] (StructuralChannel-SelectionBottleneck, SCSB)

[0149] After completing the task-level structural statistical modeling, the model has a global understanding of the importance of each channel structure in the current identification task. However, how to transform this statistical prior into an effective channel selection mechanism in the feature extraction network remains a key challenge for few-shot ISAR identification. Traditional feature channel processing methods often use fixed convolution kernels or static weighting operations. Once the weights are determined during the training phase, they are difficult to dynamically adjust with changes in task features. This static structure leads to significant feature redundancy and information noise in multi-task, few-shot scenarios, making it impossible for the network to fully focus on the most discriminative structural channels in the current task.

[0150] To address this, the present invention proposes a structural channel selection bottleneck module.

[0151] (Structural Channel-Selection Bottleneck, SCSB) is used to achieve task-driven dynamic selection and compression of feature channels, which improves the generalization and interpretability of the model while ensuring structural expressiveness.

[0152] The core idea of ​​this module is to utilize the task description embedding vector h generated in the previous stage. task As a priori signal, the set of structural feature channels C = {C1, C2, ..., C...} is subjected to a set of learnable gating functions (Gumbel-Softmaxgating). M Dynamic selection and weighting are performed to form a structural feature subspace F at the bottleneck layer that is adaptive to the task. bottle This bottleneck mechanism differs from traditional dimensionality reduction modules (such as PCA or 1×1 convolutional compression). It introduces sparsity constraints while maintaining the continuity of feature distribution through a probabilistic channel selection strategy, thereby avoiding the destruction or loss of structural information. Specifically, it first maps the task description embedding to the weight vector space corresponding to the number of channels M through a linear transformation.

[0153] π = W g h task +b g ,

[0154] Where π = [π1, π2, ..., π] M ] represents the activation prior for each channel, W g and b g These are learnable parameters; next, a Gumbel-Softmax sampling strategy is introduced to generate differentiable gated weights:

[0155]

[0156] in, For random perturbation, τ is a temperature parameter used to control the smoothness of the gating distribution; a higher τ makes the selection more uniform, while a lower τ value makes the gating weights sparser and more selective; in the early stages of training, τ takes a larger value (e.g., 2.0) to maintain a stable gradient; as training progresses, τ gradually anneals to a lower value (e.g., 0.5), making the model tend to have a clear channel selection.

[0157] After channel selection is completed, the bottleneck output feature is obtained by weighted fusion of all candidate channels:

[0158]

[0159] Among them, g i This reflects the importance of the i-th structural channel in the current task, F bottle This represents the bottleneck feature mapping after task adaptive filtering; unlike traditional channel attention (such as SE or CBAM), SCSB does not rely on explicit global average pooling or convolution kernel mapping, but establishes a direct causal relationship between the task description vector and the channel weight distribution, thereby achieving true task-level structured channel selection.

[0160] To prevent information loss due to excessive channel sparsity in the later stages of network training during optimization, this step introduces sparsity regularization and stability constraints; sparsity regularization constrains the gate weights in the form of L1 norm.

[0161] L sparse =∥g∥1,

[0162] This is used to suppress redundant channels and improve feature concentration; stability constraints maintain structural consistency through gating similarity from different perspectives:

[0163]

[0164] in, and These are the channel weights for two imaging perspectives under the same task; by jointly optimizing the classification loss, manifold consistency loss, sparsity constraints and stability constraints, the model can gradually form an adaptive selection and robust representation of the task feature structure during training.

[0165] It is worth noting that the SCSB module not only possesses task adaptability at the algorithm level, but also has strong physical interpretability; the gate weight g i This can be directly viewed as the degree of response of each channel to the target shape, scattering distribution, or energy center in different tasks. In practical engineering applications of ISAR, this interpretability helps to analyze the model's criteria for distinguishing different ship categories. For example, in aircraft carrier identification tasks dominated by hull features, low-frequency channels usually have higher weights, while in destroyer identification tasks with dense scattering points, the weights of mid- and high-frequency channels increase significantly. This visualization of structure-semantic association not only enhances the reliability of the model but also provides theoretical support for subsequent task transfer and knowledge distillation.

[0166] Overall, the Structure Channel Selection Bottleneck (SCSB) achieves a leap from "static channel" to "task-adaptive channel" in few-shot ISAR recognition by combining task description embedding with differentiable gating mechanism. This module constructs an adjustable bottleneck layer in the feature space, which not only ensures the effectiveness of feature compression, but also retains the information related to the target structure to the maximum extent. As a result, the recognition accuracy and cross-task generalization ability of the entire model are significantly improved without a significant increase in the number of parameters.

[0167] Step 4: Determining the Structure-Aware Manifold Feature Inference Mechanism

[0168] (Structure-AwareManifoldInference, SAMI)

[0169] After passing through the Structure Channel Selection Bottleneck (SCSB), the model obtains the task-adaptive bottleneck feature representation F. bottle While this feature exhibits high structural consistency in space, it still suffers from two shortcomings: firstly, the feature distribution across different samples lacks global geometric constraints; secondly, the feature projection is unstable under cross-viewpoint or noise perturbation. To address these issues, this step proposes a Structure-Aware Manifold Inference (SAMI) mechanism. Its core idea is to establish local manifold relationships within the bottleneck feature subspace and achieve self-correction of feature distribution through structural consistency constraints and geometric regularization, enabling the model to maintain robust discriminative ability under conditions of few samples and cross-task scenarios. Specifically, let the sample set under the same task be X = {F1, F2, ..., F...}. N}, where each Fi ∈R d This represents the bottleneck features output by SCSB; to characterize the local geometric relationships between samples, this step first calculates the structural similarity matrix between sample pairs:

[0170]

[0171] Where σ is the scale parameter, used to control the smoothness of the neighborhood; the similarity matrix A reflects the local geometric connectivity of samples in the bottleneck feature space; unlike the traditional Euclidean distance metric, this step further introduces a structure-aware factor w. ij To emphasize the strength of the connection between structurally similar samples:

[0172]

[0173] Sim struct The structure-aware adjacency matrix represents the structural correlation based on channel activation similarity, where γ is the balance coefficient; the final structure-aware adjacency matrix is ​​defined as:

[0174]

[0175] This approach, while preserving local geometric properties, explicitly embeds structural information, making the manifold construction sensitive to channel selection results and demonstrating the adaptability of task features. During the feature propagation stage, SAMI maps bottleneck features to the manifold subspace and implements structural consistency constraints through the normalized Laplacian operator. for Given the degree matrix, the normalized Laplace matrix is:

[0176]

[0177] By minimizing the following manifold preservation loss, the local neighborhood can be made geometrically consistent in the embedding space:

[0178]

[0179] Where Tr(·) represents the matrix trace operation; this loss term guides the features to maintain local geometric stability in the embedding space through explicit structural constraints, thereby significantly reducing the risk of feature overfitting under small sample conditions;

[0180] To further enhance adaptability to complex perspective changes and noise disturbances, this step introduces a self-corrective weighting (SCW) mechanism during manifold propagation. This mechanism dynamically adjusts the weights of samples based on the local density of features, making feature updates more biased towards high-confidence regions. The specific definition is as follows:

[0181]

[0182] in Represents the local density of sample i. The average density is κ, and the adjustment parameter is κ; feature updates are achieved through weighted propagation.

[0183]

[0184] in This is the normalized adjacency matrix; this process can be understood as a structurally consistent "soft propagation" that enhances the global consistency of feature distribution while maintaining the local geometric topology.

[0185] Within the overall optimization framework, SAMI and SCSB work synergistically: the former generates structured features based on task-adaptive channel selection, while the latter constrains geometric consistency between samples through manifold inference. Together, they enable the model to not only focus on the most discriminative structural channels but also maintain a stable topological structure in the feature space at the manifold level. The final optimization objective consists of classification loss, manifold preservation loss, and sparsity regularization.

[0186] L total =L cls +λ1L manifold +λ2L sparse ,

[0187] λ1 and λ2 are trade-off coefficients used to balance discriminability and structural sparsity. Through joint optimization, the model further improves the continuity and robustness of the feature space while ensuring discrimination accuracy.

[0188] Overall, Structure-Aware Manifold Feature Inference (SAMI) constructs interpretable, robust, and task-adaptive geometric constraints in the bottleneck feature space by introducing structure-weighted adjacency modeling and density self-correction propagation mechanisms. This elevates the distribution learning of few-shot ISAR images from "static representation" to "structure-driven manifold consistency inference." While ensuring the model's lightweight nature, this mechanism effectively enhances feature stability and generalization ability under cross-viewpoint and cross-task conditions, providing high-quality structural manifold representations for the final few-shot recognition module.

[0189] Step 5: Adaptive classification and stability joint optimization stage

[0190] (AdaptiveClassificationandStabilityCo-Optimization,ACSC)

[0191] After completing the structural channel selection and manifold feature inference, the model has obtained the bottleneck feature F with task-adaptive feature distribution and geometric consistency.bottle However, in the few-sample ISAR recognition task, how to simultaneously ensure classification discriminativeness and feature stability under limited sample conditions remains a key challenge. To address this issue, this paper proposes an adaptive classification and stability joint optimization mechanism (ACSC). Through unified multi-loss constraints, the adaptive feature discriminative learning and channel stability regularization are optimized in a coordinated manner, thereby achieving stable and reliable recognition performance under complex conditions such as low sample number, cross-viewpoint, and noise perturbation.

[0192] The core idea of ​​this stage is to combine prototype-based metric learning with structural consistency regularization, enabling the model to adaptively generate highly discriminative boundaries for different task categories while maintaining consistency in channel activation distribution under multi-view conditions. Specifically, for each category c within a task, the category prototype vector is calculated using the bottleneck features supporting the samples.

[0193]

[0194] Where S c This represents the set of supporting samples for category c; for the query sample F q The Euclidean distance between it and the prototypes of each category is defined as:

[0195]

[0196] And calculate the classification probability based on the distance distribution:

[0197]

[0198] The classification loss is expressed as a negative log-likelihood:

[0199]

[0200] Where y q To query the true class label of the sample; this loss term measures the clustering and separation process in the space, ensuring that the model can still generate structural features with clear class boundaries even with few samples;

[0201] To further enhance the robustness of the model across different tasks and viewpoints, this invention introduces a cross-view stability regularization constraint in addition to the classification loss. This constraint maintains the consistency of channel selection patterns by comparing the gating weight distributions of samples from different viewpoints under the same task, thereby suppressing the model's oversensitivity to changes in imaging angle and scattering. The stability regularization term is defined as follows:

[0202]

[0203] in and These represent the weights of the i-th channel corresponding to two different perspective inputs under the same task; the lower L stab The value indicates that the model responds more consistently to the same structural feature channel under different observation conditions, which helps to improve the cross-view robustness of the feature space;

[0204] Furthermore, to avoid the loss of important features due to excessive sparsity in the task-adaptive gating mechanism, this step introduces ChannelSparsityRegularization to control the sparsity of the gating vectors in the form of L1 norm:

[0205] L sparse =||g||1,

[0206] This factor promotes sparser channel selection in the early stages of training to improve the model's ability to focus on key structural channels; in the later stages, it balances with stability constraints to prevent feature dimension collapse; the final joint optimization objective function is defined as:

[0207] L total =L cls +λ1L manifold +λ2L stab +λ3L sparse ,

[0208] Among them, L cls To ensure discriminative ability, L manifold Maintaining geometric consistency, L stab Improve channel stability, L sparse The sparsity of the constraint features is defined by λ1, λ2, and λ3, which are the balance coefficients. Through joint optimization of multiple losses, the model achieves a dynamic trade-off between discriminability, robustness, and stability globally.

[0209] In terms of optimization strategy, ACSC adopts a joint training process based on phased weight scheduling. The initial stage (Warm-up) focuses on classification and manifold constraints to quickly form a separable feature space. The middle stage gradually introduces stability regularization to improve multi-view consistency. The later stage dynamically reduces the sparse constraint weights to ensure the sufficiency of feature representation. The entire optimization process uses the AdamW optimizer, with progressive annealing of the learning rate, and gradient clipping to prevent gradient explosion in channel selection.

[0210] It is worth emphasizing that the advantages of ACSC are not only reflected in performance improvement, but also in its physical rationality and engineering interpretability. The stability constraint of the gating weights mathematically corresponds to the consistency constraint of the scattering center response, reflecting the physical invariance of the structural response of the same target under different viewpoints. The channel sparsity constraint simulates the human focusing mechanism on significant structural features in target recognition. The combination of the two enables the model to adaptively select the most discriminative structural pattern when faced with complex noise and limited data, thereby achieving high-precision and low-variance recognition under small sample conditions.

[0211] In summary, Adaptive Classification and Stability Joint Optimization (ACSC), as the final stage of TA-SCSB, establishes a comprehensive optimization framework that integrates task-adaptive discriminative learning, geometric manifold consistency, and cross-perspective robustness. This stage effectively alleviates the overfitting and feature drift problems commonly found in few-shot ISAR recognition, enabling the entire model to maintain high-precision recognition performance while possessing interpretable, transferable, and deployable engineering features, thus providing the final optimization loop for task-adaptive structure recognition networks.

[0212] Main contributions of this invention

[0213] 1) A task-adaptive structure channel selection mechanism is proposed, which can dynamically generate channel weights based on the structural statistical features of supporting samples, realize sparse selection and optimization of structural feature channels, effectively reduce feature redundancy and enhance discrimination ability.

[0214] 2) A structural channel selection bottleneck (SCSB) and an improved manifold inference module were designed, which only performs manifold modeling in the bottleneck subspace, thereby improving the model's ability to express structural similarity and its generalization performance in small sample scenarios.

[0215] 3) A joint optimization mechanism with stability and sparsity constraints was constructed to ensure the consistency of channel activation and feature stability of the model under different perspectives and noise interference conditions, so as to achieve highly robust few-sample ISAR identification.

[0216] Example 1

[0217] To verify the feasibility and practical performance of the TA-SCSB (Task-Adaptive Structural Channel-Selection Bottleneck) of this invention, the research team conducted systematic experiments and simulation tests on a self-built simulated ISAR image dataset. This dataset, constructed based on typical ship structural characteristics and radar scattering models, covers six types of targets: aircraft carriers, cruisers, destroyers, frigates, and civilian vessels. It simulates imaging scenarios with different viewing angles (azimuth range 0°–180°), elevation angles (5°, 10°, 15°), and signal-to-noise ratios (3–15dB), generating approximately 12,000 ISAR images in total. The dataset is divided into training, validation, and test sets according to a 5-way 5-shot few-shot learning protocol, covering typical sea surface scenarios with diverse structures and complex noise.

[0218]

[0219]

[0220] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A few-sample ISAR target recognition method based on TA-SCSB, characterized in that, The proposed method is based on the REMI framework and introduces a task-adaptive structural channel selection mechanism in the process of structural feature extraction and manifold inference. It can dynamically adjust the channel weights based on the structural statistical information of a small number of supporting samples, thereby achieving highly robust recognition of multi-view, low-sample, and high-noise ISAR images.

2. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 1, characterized in that, The method includes the following steps: Step 1: Structural prior feature extraction stage; Step 2: Task-adaptive structural statistical modeling stage; Step 3: Bottleneck selection stage of structural channel selection; Step 4: Determining the structure-aware manifold feature reasoning mechanism; Step 5: Adaptive classification and stability joint optimization stage.

3. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 2, characterized in that, Step 1: Structural prior feature extraction stage; This step aims to extract a set of candidate structural feature channels containing rich morphological information from the original ISAR image; These channels retain structural information highly correlated with the target morphology at the spatial, frequency, and scattering levels, providing structural prior support for task-adaptive channel selection and manifold feature inference in subsequent stages. The implementation of the structural prior feature extraction stage involves three sub-steps: Step 1.1: Multi-scale structural deconstruction; Step 1.2: Physical-Statistical Dual Feature Fusion; Step 1.3: Channel Standardization and Response Enhancement; The input for this stage is a few-sample ISAR image I∈R H×W The output is a set of candidate channels for the structure, C = {C1, C2, ..., C}. M }; Each channel C i This corresponds to a salient structural feature in a physical or statistical sense.

4. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 3, characterized in that, Step 1.1: Multi-scale structural deconstruction; To characterize the structural patterns of ISAR images at different scales, this step designs a multi-scale structure deconstruction module; specifically, the input image I is processed through three spatial frequency filtering kernels f at different scales. low ,f mid ,f high The target is decomposed and its low-frequency morphology, main outline, and high-frequency scattering structure are extracted separately. in, S1 represents a convolution or bandpass filtering operation at the k-th scale; S2 preserves large-scale geometric information, S2 captures the mesoscale structural outline, and S3 focuses on high-frequency details and scattering center features. The feature maps from the three scales are then concatenated and spatially normalized to form a basic multi-scale structural representation: Step 1.2: Physics-Statistics Dual Feature Fusion ISAR images not only reflect the geometric contours of the target, but also imply physical scattering features. In order to fully express the structural information, this step introduces a physical-statistical dual feature fusion unit after the MSD output to fuse the target's physical scattering response and statistical texture index. Specifically, we define the local response vector for each spatial location (x, y): r(x,y)=[E(x,y),G(x,y),T(x,y)] Where E(x,y) is the scattering energy density; G(x,y) is the local gradient magnitude; T(x,y) is the texture variance; then global channel aggregation is calculated: Among them, Ω i Let w(x,y) be the receptive field region of channel i, and w(x,y) be the weight mapping. Energy conservation is ensured by softmax normalization. In this way, local scattering modes can be converged according to energy significance to obtain channel features with physical interpretability; Step 1.3: Channel Standardization and Response Enhancement To prevent numerical differences between features at different scales or in different regions from causing training instability, this step introduces a two-step enhancement strategy after the fusion result: Step 1.3.1: Channel Standardization Wherein, μ(C i ) and σ(C i ) represent the channel mean and standard deviation, respectively, and ∈ is the stability constant; Step 1.3.2: Structural Response Enhancement To enhance the response intensity of critical structural regions, a weighting function based on structural attention is introduced: Among them, A i The structural saliency map is generated jointly by Sobel edge gradients and Laplacian second-order information, with β being the enhancement coefficient (empirically taken as 0.2–0.4); the final output is a set of candidate channels for the structure. C={C'1,C'2,…,C' M } The structural prior feature extraction stage constructs a candidate structural channel set containing target morphology and scattering information through multi-scale structural deconstruction and physical-statistical fusion, providing a stable, interpretable and robust structural representation basis for the subsequent task adaptive channel selection bottleneck (SCSB).

5. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 4, characterized in that, Step 2: Task adaptive structural statistical modeling stage; In few-sample ISAR target recognition, one of the core challenges faced by the model is the high heterogeneity of structural distribution among different tasks. Different ship types have significant differences in imaging conditions, attitude angles, scattering center layouts, and structural scales. If a fixed feature extraction channel or a uniform structural weighting method is still used, the model cannot make the optimal selection and representation of structural features for specific tasks, resulting in feature redundancy and a decline in recognition performance. To address this issue, this step proposes a task-adaptive structural statistical modeling mechanism. Its core idea is to automatically learn and represent the structural distribution pattern of the current recognition task by utilizing statistical information from a small number of supporting samples; and to analyze the candidate structural channel set C = {C1, C2, ..., C...}. M Perform structural statistical modeling to generate task-level description vectors s task This vector can quantitatively reflect the importance and distinctiveness of different channels in the current task, and provide adaptive guidance signals for subsequent channel selection bottlenecks; The task-adaptive structural statistical modeling and solution phase involves three sub-steps: Step 2.1: Calculation of local structural indicators: Extract multidimensional statistical indicators that reflect the significance of structural characteristics from each channel; Step 2.2: Task-level statistical aggregation: Based on the supporting sample set, the statistical indicators of each channel are weighted and aggregated to form a stable task representation; Step 2.3: Feature Normalization and Vector Encoding: Normalize and embed the statistical results to obtain the task description vector s. task .

6. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 5, characterized in that, Step 2.1: Calculation of local structural indices; For each channel C i ∈C, firstly, extract three types of indicators that can characterize the stability and morphological complexity of the channel structure: in, This represents the pixel variance within the channel, used to reflect the intensity fluctuations in the structure distribution; Information entropy describes the uncertainty and complexity of the structure within a channel; The average gradient intensity is used to quantify the sharpness of edges and contours; To synthesize these characteristics, a structural significance index for weighted combinations is defined: Where α1, α2, and α3 are empirical weights (usually taken as 0.4, 0.3, and 0.3), which can be adaptively adjusted during training; Step 2.2: Task-level statistical aggregation; Let the supporting sample set be S = {I1, I2, ..., I...} K } contains K few-sample ISAR images; each sample, after undergoing a structural prior feature extraction stage, yields a set of channel features C. k ={C k,1 ,…,C k,M For each channel, calculate the cross-sample statistic to obtain the task-level average structure index: To improve the separability between different channels, normalization and relative weight constraints are introduced: The final task description vector is obtained as follows: This vector reflects the importance distribution of each channel under the current task and serves as the input condition for the channel selection bottleneck module; Step 2.3: Feature Normalization and Vector Encoding To further enhance the expressive power of the task description vector, this step designs a lightweight statistical embedding encoder; the encoding process is as follows: h task =σ(W2ReLU(+b2) Where W is the learnable weight matrix; b is the bias term; h task The final task structure description is embedded to generate subsequent channel gating distributions; The task-adaptive structural statistical modeling stage achieves adaptive capture of task-level structural patterns by constructing local statistical indicators and task-level structural description vectors. This provides quantitative, interpretable, and differentiable task structure priors for channel selection bottlenecks, realizing a key leap from "static channels" to "task-driven channels".

7. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 6, characterized in that, Step 3: Bottleneck selection stage of structural channel; After completing the task-level structural statistical modeling, the model has a global understanding of the importance of each channel structure in the current identification task. However, how to transform this statistical prior into an effective channel selection mechanism in the feature extraction network remains a key challenge for few-shot ISAR identification. Traditional feature channel processing methods often use fixed convolution kernels or static weighting operations. Once the weights are determined during the training phase, they are difficult to dynamically adjust with changes in task features. This static structure leads to significant feature redundancy and information noise in multi-task, few-shot scenarios, making it impossible for the network to fully focus on the most discriminative structural channels in the current task. To address this, this step proposes a bottleneck module for structural channel selection, which enables task-driven dynamic filtering and compression of feature channels, thereby improving model generalization and interpretability while ensuring structural expressiveness. The core idea of ​​this module is to utilize the task description embedding vector h generated in the previous stage. task As a priori signal, the set of structural feature channels C = {C1, C2, ..., C...} is subjected to a set of learnable gating functions. M Dynamic selection and weighting are performed to form a structural feature subspace F at the bottleneck layer that is adaptive to the task. bottle This bottleneck mechanism differs from traditional dimensionality reduction modules. It introduces sparsity constraints while maintaining the continuity of feature distribution through a probabilistic channel selection strategy, thereby avoiding the destruction or loss of structural information. Specifically, it first uses a linear transformation to embed and map the task description into a weight vector space corresponding to the number of channels M. π=W g h task +b g , Where π = [π1, π2, ..., π] M ] represents the activation prior for each channel, W g and b g These are learnable parameters; next, a Gumbel-Softmax sampling strategy is introduced to generate differentiable gated weights: in, For random perturbation, τ is a temperature parameter used to control the smoothness of the gating distribution; a higher τ makes the selection more uniform, while a lower τ value makes the gating weights sparser and more selective; in the early stages of training, τ takes a larger value to maintain a stable gradient; as training progresses, τ gradually anneals to a lower value, making the model tend to have a clear channel selection. After channel selection is completed, the bottleneck output feature is obtained by weighted fusion of all candidate channels: Among them, g i This reflects the importance of the i-th structural channel in the current task, F bottle This represents the bottleneck feature mapping after task adaptive filtering; SCSB does not rely on explicit global average pooling or convolution kernel mapping, but establishes a direct causal relationship between the task description vector and the channel weight distribution, thereby achieving true task-level structured channel selection. To prevent information loss due to excessive channel sparsity in the later stages of network training during optimization, this step introduces sparsity regularization and stability constraints; sparsity regularization constrains the gate weights in the form of L1 norm: L sparse =∥g∥1, This is used to suppress redundant channels and improve feature concentration; stability constraints maintain structural consistency through gating similarity from different perspectives: in, and These are the channel weights for two imaging perspectives under the same task; by jointly optimizing the classification loss, manifold consistency loss, sparsity constraints and stability constraints, the model can gradually form an adaptive selection and robust representation of the task feature structure during training. The SCSB module not only possesses task adaptability at the algorithm level but also exhibits strong physical interpretability; the gate weights g i This can be directly viewed as the degree of response of each channel to the target shape, scattering distribution, or energy center in different tasks. In practical engineering applications of ISAR, this interpretability helps to analyze the model's criteria for distinguishing different ship categories. For example, in aircraft carrier identification tasks dominated by hull features, low-frequency channels usually have higher weights, while in destroyer identification tasks with dense scattering points, the weights of mid- and high-frequency channels increase significantly. This visualization of structure-semantic association not only enhances the reliability of the model but also provides theoretical support for subsequent task transfer and knowledge distillation. Overall, the structure channel selection bottleneck, by combining task description embedding with differentiable gating mechanism, achieves a leap from "static channel" to "task-adaptive channel" in few-sample ISAR recognition. This module constructs an adjustable bottleneck layer in the feature space, which not only ensures the effectiveness of feature compression, but also retains the information related to the target structure to the maximum extent. This results in a significant improvement in recognition accuracy and cross-task generalization ability of the entire model without a significant increase in the number of parameters.

8. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 7, characterized in that, Step 4: Determination of the structure-aware manifold feature reasoning mechanism; After overcoming the bottleneck in structural channel selection, the model obtains the bottleneck feature representation F for task adaptation. bottle While this feature exhibits high structural consistency in space, it still suffers from two shortcomings: firstly, the feature distribution across different samples lacks global geometric constraints; secondly, the feature projection is unstable under cross-viewpoint or noise perturbation. To address these issues, this step proposes a structure-aware manifold feature inference mechanism. Its core idea is to establish local manifold relationships within the bottleneck feature subspace and achieve self-correction of feature distribution through structural consistency constraints and geometric regularization, enabling the model to maintain robust discriminative ability under conditions of few samples and cross-task scenarios. Specifically, let the sample set under the same task be X = {F1, F2, ..., F...}. N }, where each F i ∈R d This represents the bottleneck features output by SCSB; to characterize the local geometric relationships between samples, this step first calculates the structural similarity matrix between sample pairs: Where σ is the scale parameter, used to control the smoothness of the neighborhood; the similarity matrix A reflects the local geometric connectivity of samples in the bottleneck feature space; unlike the traditional Euclidean distance metric, this step further introduces a structure-aware factor w. ij To emphasize the strength of the connection between structurally similar samples: Sim struct The structure-aware adjacency matrix represents the structural correlation based on channel activation similarity, where γ is the balance coefficient; the final structure-aware adjacency matrix is ​​defined as: This approach, while preserving local geometric properties, explicitly embeds structural information, making the manifold construction sensitive to the channel selection results and demonstrating the adaptability of task features. During the feature propagation phase, SAMI maps bottleneck features to a manifold subspace and implements structural consistency constraints through a normalized Laplacian operator; let... for Given the degree matrix, the normalized Laplace matrix is: By minimizing the following manifold preservation loss, the local neighborhood can be made geometrically consistent in the embedding space: Where Tr(·) represents the matrix trace operation; this loss term guides the features to maintain local geometric stability in the embedding space through explicit structural constraints, thereby significantly reducing the risk of feature overfitting under small sample conditions; To further enhance adaptability to complex perspective changes and noise disturbances, this step introduces a self-correcting weight mapping mechanism during manifold propagation. This mechanism dynamically adjusts the weights of samples based on the local density of features, making feature updates more biased towards high-confidence regions. The specific definition is as follows: in Represents the local density of sample i. The average density is κ, and the adjustment parameter is κ; feature updates are achieved through weighted propagation. in This is the normalized adjacency matrix; this process can be understood as a structurally consistent "soft propagation" that enhances the global consistency of feature distribution while maintaining the local geometric topology. Within the overall optimization framework, SAMI and SCSB work synergistically: the former generates structured features based on task-adaptive channel selection, while the latter constrains geometric consistency between samples through manifold inference. Together, they enable the model to not only focus on the most discriminative structural channels but also maintain a stable topological structure in the feature space at the manifold level. The final optimization objective consists of classification loss, manifold preservation loss, and sparsity regularization. L total =L cls +λ1L manifold +λ2L sparse , λ1 and λ2 are trade-off coefficients used to balance discriminability and structural sparsity. Through joint optimization, the model further improves the continuity and robustness of the feature space while ensuring discrimination accuracy. Overall, structure-aware manifold feature inference, by introducing structure-weighted adjacency modeling and density self-correction propagation mechanism, constructs interpretable, robust and task-adaptive geometric constraints in the bottleneck feature space, thereby elevating the distribution learning of few-shot ISAR images from "static representation" to "structure-driven manifold consistency inference". This mechanism effectively enhances feature stability and generalization ability under cross-viewpoint and cross-task conditions while ensuring model lightweightness, providing high-quality structural manifold representation for the final few-shot recognition module.

9. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 8, characterized in that, Step 5: Adaptive classification and stability joint optimization stage; After completing the structural channel selection and manifold feature inference, the model has obtained the bottleneck feature F with task-adaptive feature distribution and geometric consistency. bottle However, in the few-sample ISAR recognition task, how to ensure both classification discriminativeness and feature stability under limited sample conditions remains a key challenge. To address this issue, this paper proposes an adaptive classification and stability joint optimization mechanism. Through unified multi-loss constraints, the adaptive feature discriminative learning and channel stability regularization are optimized in a coordinated manner, thereby achieving stable and reliable recognition performance under complex conditions such as low sample number, cross-view, and noise perturbation. The core idea of ​​this stage is to combine prototype-based metric learning strategies with structural consistency constraints, enabling the model to adaptively generate highly discriminative boundaries for different task categories while maintaining consistency in channel activation distribution under multi-view conditions. Specifically, for each category c within a task, the category prototype vector is calculated using the bottleneck features supporting the samples. Where S c This represents the set of supporting samples for category c; for the query sample F q The Euclidean distance between it and the prototypes of each category is defined as: And calculate the classification probability based on the distance distribution: The classification loss is expressed as a negative log-likelihood: Where y q To query the true class label of the sample; this loss term measures the clustering and separation process in the space, ensuring that the model can still generate structural features with clear class boundaries even with few samples; To further enhance the robustness of the model across different tasks and perspectives, this invention introduces a cross-perspective stability constraint in addition to the classification loss. This constraint maintains the consistency of the channel selection pattern by comparing the gating weight distribution of samples from different perspectives under the same task, thereby suppressing the model's oversensitivity to changes in imaging angle and scattering. The stability regularization term is defined as follows: in and These represent the weights of the i-th channel corresponding to two different perspective inputs under the same task; the lower L stab The value indicates that the model responds more consistently to the same structural feature channel under different observation conditions, which helps to improve the cross-view robustness of the feature space; Furthermore, to avoid the loss of important features due to excessive sparsity in the task-adaptive gating mechanism, this step introduces channel sparsity constraint regularization to control the sparsity of the gating vector in the form of L1 norm: L sparse =||g||1, This factor promotes sparser channel selection in the early stages of training to improve the model's ability to focus on key structural channels; in the later stages, it balances with stability constraints to prevent feature dimension collapse; the final joint optimization objective function is defined as: L total =L cls +λ1L manifold +λ2L stab +λ3L sparse , Among them, L cls To ensure discriminative ability, L manifold Maintaining geometric consistency, L stab Improve channel stability, L sparse The sparsity of the constraint features is defined by λ1, λ2, and λ3, which are the balance coefficients. Through joint optimization of multiple losses, the model achieves a dynamic trade-off between discriminability, robustness, and stability globally. In terms of optimization strategy, ACSC adopts a joint training process based on phased weight scheduling. In the initial stage, it focuses on classification and manifold constraints to quickly form a separable feature space. In the middle stage, stability regularization is gradually introduced to improve multi-view consistency. In the later stage, the sparse constraint weights are dynamically reduced to ensure the sufficiency of feature representation. The entire optimization process uses the AdamW optimizer, with progressive annealing of the learning rate, and gradient clipping to prevent channel selection gradient explosion. The advantages of ACSC are not only reflected in performance improvement, but also in its physical rationality and engineering interpretability. The stability constraint of the gating weights mathematically corresponds to the consistency constraint of the scattering center response, reflecting the physical invariance of the structural response of the same target under different perspectives. The channel sparsity constraint simulates the human focusing mechanism on significant structural features in target recognition. The combination of the two enables the model to adaptively select the most discriminative structural pattern when faced with complex noise and limited data, thereby achieving high-precision, low-variance recognition under small sample conditions.

10. The few-sample ISAR target recognition method based on TA-SCSB as described in claim 9, characterized in that, The joint optimization of adaptive classification and stability, as the final stage of TA-SCSB, establishes a comprehensive optimization framework that integrates task-adaptive discriminative learning, geometric manifold consistency, and cross-perspective robustness. This stage effectively alleviates the overfitting and feature drift problems commonly found in few-shot ISAR recognition, enabling the entire model to maintain high-precision recognition performance while possessing interpretable, transferable, and deployable engineering features, thus providing the final optimization loop for the task-adaptive structural recognition network.

Images