Remote sensing small sample target detection method based on generative meta-learning strategy

Abstract

The invention belongs to the technical field of image recognition, and discloses a remote sensing small sample target detection method based on a generative meta-learning strategy, and the method comprises the steps: inputting an image into a dynamic relation double-flow collaborative network through constructing a K-way N-shot meta-learning task, and carrying out the feature extraction and alignment enhancement through an improved non-causal Mama unit and a gating fusion mechanism; representing a support set category as probability distribution by using a generative meta-learner, and calculating a Wasserstein distance between the support set distribution after enhancement and the query feature distribution; inputting the query features into an orthogonal frequency domain decomposition detection head for frequency domain analysis and reconstruction; and outputting a target category and a bounding box based on the reconstructed features, adopting non-maximum suppression post-processing, and calculating an IoU optimization parameter during training. According to the method, the classification accuracy and generalization performance in a small sample scene are effectively improved, the calculation complexity overfitting risk is remarkably reduced, the detection efficiency is high, and the method is easy to deploy in hardware equipment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image recognition, and particularly relates to a remote sensing small sample target detection method based on a generative meta-learning strategy. BACKGROUND

[0002] Remote sensing small sample target detection plays an important role in disaster emergency, environmental monitoring and military reconnaissance. For target detection technology in conventional scenes, existing methods have achieved good detection performance. However, there are significant differences between remote sensing images and natural images. Due to the unique perspective of remote sensing, the variable scale of ground objects, the complex background and the sparse characteristics of target features, the difficulty of remote sensing small sample detection is increased. In order to improve the rapid response and recognition ability of the earth observation system to rare or sudden targets, the research on remote sensing small sample target detection is particularly important.

[0003] For remote sensing small sample target detection, it can be generally divided into research paradigms based on transfer learning and meta-learning. The method based on transfer learning aims to utilize the general knowledge learned from large-scale basic data, and fine-tunes the model through a small number of samples to adapt to new categories. The method based on meta-learning designs and solves a large number of simulated small sample tasks to make the model learn "how to learn", so as to quickly generalize to new categories with few samples. Among them, the transfer learning method depends on the quality of the pre-training model of the basic category and the fine-tuning strategy, and the characteristics of large intra-class difference and high inter-class similarity in remote sensing images are easy to cause negative transfer or overfitting when the model is fine-tuned, which limits the application of the transfer learning method. In contrast, the meta-learning method simulates the few sample conditions when testing under the task-driven, and its generalization mechanism to new categories is more direct, so it has attracted widespread attention in remote sensing small sample detection. It can be roughly divided into meta-learning based on measurement, meta-learning based on optimization and hybrid methods combining meta-learning and transfer learning.

[0004] Specifically, existing technologies most similar to this approach include Meta-RCNN, the instance-based pruning mean prototype method, spatial alignment mechanisms, attention-based bounding box generation networks, and the Meta-Det framework. These methods address small-sample object detection tasks through strategies such as channel-level feature fusion, mean prototype aggregation, pixel-level fine-grained matching, attention-guided candidate box generation, and task-decoupled parameter learning. The Meta-RCNN method suffers from sensitivity to support set sampling noise, and background interference easily degrades the quality of meta-knowledge. This is because the PRN network fails to effectively distinguish between target features and background noise during the fusion process, causing noise to be passed on to the detection process along with the meta-knowledge. The instance-based pruning mean prototype method suffers from the loss of detailed information during the averaging operation and is highly dependent on instance pruning accuracy. This is because mean pooling smooths local features, and pruning errors directly affect the spatial representativeness of the prototype. The spatial alignment mechanism method suffers from high computational complexity and requires the support prototype to have a complete spatial structure. This is because pixel-wise similarity measurement leads to a surge in computation, and a lack of spatial structure destroys the alignment effectiveness. The drawbacks of attention-based bounding box generation networks lie in the additional computational burden introduced by the attention mechanism and their unstable noise suppression capabilities. This is because attention weight calculation relies on global interactions, and noise interference can affect the reliability of weight allocation. The Meta-Det framework method suffers from the potential weakening of multi-task collaboration due to task decoupling, and its limited improvement for long-tail categories. This is because the separation of classification and localization parameters may hinder feature sharing between tasks, making it difficult to fully utilize small sample information.

[0005] Therefore, to address the issues of noise sensitivity and background interference, this scheme generates robust feature prototypes through generative probabilistic modeling, explicitly filters out noise, and enhances the discriminative power of target features. To address the issues of feature detail loss and low computational efficiency, this scheme combines optimal transmission metric and gated dual-stream fusion to achieve efficient feature interaction while preserving fine-grained information. To address the issues of insufficient task collaboration and weak long-tail generalization, this scheme enhances feature complementarity between tasks through a dynamic multi-scale fusion mechanism, improving the model's adaptability to rare categories.

[0006] Meanwhile, metric-based meta-learning methods typically learn an embedding space where similarity between query samples and support samples is measured using distance functions (such as cosine distance and Euclidean distance) to achieve classification. A common drawback of these methods is that when the number of support samples is extremely small, the constructed class prototypes often lack representativeness and struggle to handle intra-class differences in remote sensing targets caused by variations in illumination, angle, and scale, leading to inaccurate measurements. With the development of generative models and attention mechanisms, remote sensing few-sample detection methods based on deep meta-learning are constantly emerging. These methods, by introducing generative data augmentation, relation modeling, and feature alignment mechanisms, can still construct highly discriminative feature spaces under conditions of scarce samples, continuously improving detection performance. Therefore, a remote sensing few-sample target detection method based on deep meta-learning is urgently needed.

[0007] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0008] (1) Existing technologies have low target detection accuracy under conditions of extremely small samples (such as 1-shot and 5-shot) of remote sensing images, and the generalization ability of the model is insufficient.

[0009] (2) When faced with the problem of large intra-class differences and high inter-class similarity in remote sensing images due to illumination and viewing angle, the existing technology has difficulty extracting discriminative class representations, which easily leads to false detections and false negatives.

[0010] (3) Existing technologies fail to effectively utilize generative models to enhance the distribution of support sets in the latent space, making it difficult to eliminate feature distribution bias under small sample conditions and limiting the model's ability to estimate the true distribution of categories.

[0011] (4) Existing technologies lack a dynamic collaborative alignment mechanism for the features of the support set and the query set, making it difficult to achieve efficient feature interaction and enhancement under task-driven conditions, which restricts the improvement of query feature representation capabilities. Summary of the Invention

[0012] To overcome the problems existing in related technologies, the present invention discloses an embodiment of a remote sensing small-sample target detection method based on a generative meta-learning strategy, the technical solution of which is as follows:

[0013] This invention is implemented as follows: a remote sensing small-sample target detection method based on a generative meta-learning strategy, comprising the following steps:

[0014] S1. Prepare training samples of satellite remote sensing images containing support sets and query sets, construct a K-way N-shot meta-learning training task, and simulate a small sample detection environment in a real scene.

[0015] S2 will support the dynamic relationship between the image input of the support set and the query set in a dual-stream collaborative network. By improving the non-causal Mamba unit, two-dimensional position coding and gating fusion mechanism, dual-stream feature extraction and spatial semantic fusion are performed to achieve dynamic alignment and enhancement of support and query features.

[0016] S3, the generative meta-learner based on optimal transmission represents the support set categories as probability distributions, and enhances the distribution of the support set in the latent space by integrating generative models;

[0017] S4. The Wasserstein distance between the enhanced support set category distribution and the query sample feature distribution is calculated using optimal transport theory to establish the feature correspondence between the support set and the query set.

[0018] S5 inputs the query features enhanced by optimal transmission into the orthogonal frequency domain decomposition detection head for frequency domain analysis and feature reconstruction, and adaptively decomposes the features into low-frequency contour components and high-frequency detail components.

[0019] S6, based on the reconstructed features, outputs the target's class probability distribution and accurate bounding box coordinate regression values ​​through a multi-scale prediction head. Non-maximum suppression post-processing is used to generate the final predicted bounding box. During training, the intersection-union ratio (IoU) between the predicted box and the ground truth label is calculated to optimize the model parameters.

[0020] In step S1, the support set contains K target categories, with N labeled samples for each category; the query set contains unlabeled image sequences from the same category.

[0021] In step S2, the dynamic relational dual-stream collaborative network consists of three parts: a non-causal Mamba unit, a two-dimensional multi-head self-attention mechanism, and an adaptive gating fusion mechanism. The spatial perception mechanism is improved based on the original Mamba structure, and the gating mechanism is introduced to adaptively adjust the features.

[0022] Input feature tensor Remodeling into a sequence Represented as:

[0023]

[0024] In the formula, For the reconstructed feature sequence, For shape reshaping calculations, To reshape the sequence, For batch size, These represent the number of channels, height, and width of the position embedding tensor, respectively.

[0025] Learnable relative position embeddings are introduced along the height and width dimensions, respectively. The expression is:

[0026]

[0027] In the formula, For learnable relative position embeddings, These are trainable parameters, used to capture the positional biases in the vertical and horizontal directions, respectively.

[0028] Injecting location information into sequence features, the expression is:

[0029]

[0030] In the formula, To incorporate sequence features that combine location information, Transpose of relative position embedding;

[0031] A spatial context-aware gating system is constructed using one-dimensional convolution and the Sigmoid activation function to adaptively adjust the feature response intensity; the sequence is routed via a non-causal Mamba block. Processing to capture long-range context dependencies, Mamba output is coupled with a gated signal. To perform element-by-element fusion, the expression is:

[0032]

[0033] In the formula, For output features, For the output characteristics of Mamba blocks, For activation function, For convolution calculation, This is the transpose of the sequence features;

[0034] Finally, the output sequence is reshaped back into spatial dimensions. In the form of.

[0035] Furthermore, based on the Mamba structure, a two-dimensional multi-head self-attention mechanism is introduced to form a two-stream collaborative structure together with Mamba; by explicitly modeling the interaction relationships between all spatial locations, the complete global context of features is captured.

[0036] Input features Convert to a query Q, key K, and value V, where, The dimension of the K vector is used to calculate the attention weights. ;

[0037]

[0038] In the formula, Calculate for the softmax activation function. It is the transpose of the K vector;

[0039] Introducing two-dimensional relative position encoding enables the attention mechanism to perceive the relative geometric relationship between positions;

[0040] A gated adaptive fusion mechanism is designed to equally divide the input features along the channel dimension, resulting in evenly distributed features. and The input consists of two branches: Mamba and self-attention, which are learned gating systems. and The expression for combining the outputs of the two branches is:

[0041]

[0042]

[0043] In the formula, As a feature of fusion, As a characteristic of being evenly distributed to the Mamba network, Features are evenly distributed to the self-attention network. For learnable gated outputs of the Mamba branch, For learnable gated outputs of self-attention branches, All are learnable projection weights.

[0044] In step S3, the generative meta-learner based on optimal transmission represents the support set categories as probability distributions, including:

[0045] Prior distribution Set as a standard multivariate Gaussian distribution Approximate posterior distribution It is then modeled as a multivariate Gaussian distribution with a diagonal covariance matrix. ; distribution parameter mean and variance Through feature encoder The expression for generation is:

[0046]

[0047] In the formula, The distribution of supporting features is calculated for the feature encoder;

[0048] Variational features generated using reparameterization techniques The calculation process is as follows:

[0049]

[0050] In the formula, For a bias that follows a standard Gaussian distribution;

[0051] The lower bound of evidence is decomposed into two loss terms with clear physical meaning: reconstruction loss and KL divergence loss; reconstruction loss Defined as input features With the decoder Reconstructed features L2 distance between them;

[0052]

[0053] KL divergence loss Represented as:

[0054]

[0055] In the formula, For KL divergence calculation, Given an observation sequence S, what is the posterior distribution of the latent variable z? Let z be the prior distribution of the latent variable z;

[0056] Introducing consistency loss The loss function is defined as the reconstructed features. After a lightweight linear classifier The output results are compared with the true category labels. Cross-entropy between ;

[0057]

[0058] In the variational feature aggregation stage, variational features sampled from the distribution are... By deeply integrating with query features, the meta-learner employs a category-independent aggregation strategy.

[0059] Also includes: given category Query features and categories Support features Accurately estimate the corresponding category distribution Variational features are obtained through sampling. The aggregate expression is:

[0060]

[0061] In the formula, The aggregated features of the output This is an aggregation operation between two features. The mean of the category distribution. For the variance of the class distribution, For bias, It is the sigmoid activation function.

[0062] In step S4, a distance metric based on optimal transmission is introduced on the basis of the generative meta-learner. The class similarity is measured by calculating the Wasserstein distance between the query sample and the supporting class distribution.

[0063] Furthermore, the feature representation of the query sample Considering it as a Dirac distribution, the first The support categories are represented as a uniform distribution. The corresponding number Each support category sample is represented as , The formula for calculating the Wasserstein distance between the two is:

[0064]

[0065] In the formula, This is the feature representation obtained by mapping the query samples. For the first One supporting sample, The output is the Wasserstein distance between the two. To support the sample size, The distance is Euclidean.

[0066] The predicted class of the query sample is represented by the minimum Wasserstein distance:

[0067]

[0068] In the formula, To query the predicted class label of a sample, The goal is to select the index that minimizes the squared Wasserstein distance from among N candidate category indices.

[0069] In step S5, the frequency domain analysis and feature reconstruction include:

[0070] The channel dimension is expanded by point convolution, and spatial features are extracted using depthwise separable convolution;

[0071] The feature map is equally divided into two branches in the channel dimension. One branch is activated by the GELU function, and the other branch retains the linear features. The two branches are multiplied element-wise, and the number of channels is restored by point convolution.

[0072] The spatial feature map is divided into 8×8 pixel blocks. A fast Fourier transform is performed on each image block to convert it to the frequency domain. A learnable frequency domain filter is used to adaptively weight the spectrum, and selective enhancement and suppression of frequency components are achieved by multiplying element by element.

[0073] Perform an inverse Fourier transform on the filtered spectrum to convert the processed frequency domain information back into spatial domain image patches; reassemble all processed image patches into a complete feature map, remove the filled boundary parts, and finally output the enhanced feature with the same size as the input.

[0074] In step S6, during training, the Intersection over Union (IoU) between the predicted bounding boxes and the ground truth labels is calculated to optimize the model parameters. This includes: the target box score calculated using the IoU is defined as the ratio between the intersection and union of two boxes, and the formula for calculating IoU is:

[0075]

[0076] In the formula, There are two boxes respectively. The region is the union of the two boxes. The intersection area of ​​the two boxes

[0077] Combining all the above technical solutions, the beneficial effects of this invention are as follows:

[0078] First, the remote sensing few-shot target detection method based on a generative meta-learning strategy provided by this invention sets the initial parameters of a dynamic relational dual-stream collaborative network, inputs satellite remote sensing images containing support and query sets into the network for training, and obtains a few-shot target detection model based on meta-learning. During detection, this invention first extracts features from support and query samples respectively through the dynamic relational dual-stream collaborative network, and achieves dynamic alignment and enhancement between the two; subsequently, a generative meta-learner based on optimal transmission uses the aligned features for meta-learning, and eliminates intra-class differences through generative enhancement in the latent space; finally, an orthogonal frequency domain decomposition detection head decomposes and reconstructs the enhanced query features, outputting the target's category, location, and detection box, thereby intuitively displaying the prediction results in the image. The modules of this invention are highly efficient in collaboration, significantly improving detection accuracy and generalization ability under few-shot conditions, and effectively reducing the training cost of the model; the strong synergy among the modules demonstrates excellent generalization ability in scenarios with extremely few samples.

[0079] Secondly, this invention constructs a parallel two-stream network model for remote sensing small-sample target detection based on a generative meta-learning strategy. It enhances spatial representation through improved non-causal Mamba units and two-dimensional location encoding, and integrates the state-space model and self-attention advantage using a gated fusion mechanism. Remote sensing training set images are input into this model for training, resulting in a small-sample target detection model. The remote sensing image to be detected is input into the trained model, and spatial dependency modeling and global semantic inference are performed separately via the parallel two-stream structure. After fusion, a feature map containing rich target information is obtained. Then, the support set features are enhanced through probabilistic modeling and variational feature generation. The similarity between samples and categories is queried using a distributed prototype based on optimal transmission and the Wasserstein distance metric, ultimately outputting the target's category and location information. This invention enhances feature discriminativeness and suppresses noise through a spatial and frequency domain collaborative mechanism, exhibiting excellent adaptability and detection performance for multi-scale and irregular remote sensing ground objects, effectively improving classification accuracy and generalization performance in small-sample scenarios. The lightweight model design significantly reduces computational complexity and the risk of overfitting, resulting in high detection efficiency and easy deployment on hardware devices. The model is simple to operate and has strong generalization ability.

[0080] Third, this invention, with a generative meta-learner at its core, innovatively combines state-space models, probabilistic generation theory, and optimal transport theory to construct a parallel dual-stream collaborative detection framework. Firstly, addressing the fundamental challenge of few-shot learning, this invention proposes representing support set categories as structured distributions through probabilistic modeling, replacing traditional single-point prototype estimation. This fundamentally enhances the model's robustness to intra-class differences and its discriminative power in the feature space. This idea is modularized into variational feature generation and distribution similarity measurement modules that can be flexibly embedded into the meta-learning process. Furthermore, based on the principle of co-optimization of global context modeling and computational efficiency, this invention constructs a parallel dual-stream backbone structure integrating improved Mamba units and self-attention. It adaptively fuses spatial details and global semantics through a gating mechanism and performs lightweight filtering enhancement at the frequency domain level, ensuring accurate representation of multi-scale, irregular features in complex remote sensing scenarios. Finally, extensive experiments on multiple remote sensing few-shot datasets validate the effectiveness of the proposed method.

[0081] Fourth, the detection method based on generative meta-learning strategy proposed in this invention fundamentally enhances the model's discriminative ability under extremely small sample conditions (such as 1-shot and 5-shot) through probabilistic modeling and variational feature generation. Experiments on typical remote sensing small sample benchmarks show that this method can significantly improve detection accuracy and recall, effectively solving the core problems of low accuracy and weak generalization in existing technologies, and providing a reliable guarantee for achieving high-precision and high-robust small-sample identification.

[0082] Fifth, by modeling categories as probability distributions and using optimal transmission theory for measurement, this invention effectively addresses the problems of large intra-class differences and high inter-class similarities caused by illumination and viewing angle. Simultaneously, the synergistic enhancement mechanism in the spatial and frequency domains further improves the model's ability to discriminate features of irregularly shaped objects at different scales, significantly reducing the risk of false positives and false negatives in complex scenarios, and greatly expanding the application scope of this invention.

[0083] Sixth, the generative feature enhancement module proposed in this invention effectively constructs a highly discriminative feature space, providing a solid foundation for small-sample detection. This module maps support features to a structured latent distribution through probabilistic modeling and generates rich variational features through sampling. This process not only learns the underlying distribution characteristics of small-sample data but also effectively eliminates feature distribution bias caused by sample scarcity, generating highly discriminative feature prototype representations, providing crucial technical support for subsequent detection and classification tasks. Furthermore, this module is flexibly designed and can be integrated into various meta-learning frameworks as a plug-and-play unit.

[0084] Seventh, the parallel dual-stream fusion structure and gating mechanism of this invention achieve efficient dynamic interaction and global context modeling of support set and query set features. This structure performs fine-grained spatial dependency modeling on one path and global semantic reasoning on the other, adaptively integrating the advantages of both paths under task-driven conditions using improved non-causal Mamba units and a gating fusion mechanism. This not only achieves efficient modeling of long-range dependencies and significantly improves the feature representation quality of the query set, but also greatly reduces the computational complexity of traditional self-attention mechanisms through collaborative computation, achieving a balance between accuracy and efficiency.

[0085] Eighth, this invention, by constructing a generative meta-learner, can achieve high-precision target recognition with very few labeled samples, directly addressing the core pain points of high data labeling costs and a scarcity of rare target samples in the remote sensing industry. Its commercial value lies in its ability to quickly adapt to scenarios such as land monitoring, disaster assessment, and military reconnaissance using limited data, significantly reducing data preparation cycles and manual labeling costs, greatly lowering the commercial cost of acquiring remote sensing image data. Simultaneously, it effectively overcomes the problems of false positives and false negatives caused by changes in illumination, scale differences, and terrain interference in remote sensing images, providing remote sensing data service providers with a cost-effective detection application solution.

[0086] Ninth, this invention proposes a novel paradigm that integrates generative meta-learning with probabilistic feature enhancement, fundamentally correcting the bias in small sample feature distribution. It innovatively introduces optimal transport theory as a cross-domain metric, solving the matching challenge of large intra-class differences and high inter-class similarity in complex scenarios. Furthermore, it designs a gated dual-stream architecture, achieving dynamic fusion of global and local features while maintaining efficient computation. These breakthroughs provide the industry with the first system-level solution for small sample detection that balances high accuracy, strong generalization, and engineerability.

[0087] Tenth, remote sensing image annotation is a highly technical and costly skill, while traditional deep learning methods heavily rely on a large number of labeled samples. This invention achieves high-precision target detection with very few labeled samples through generative meta-learning and variational feature generation, while also improving performance with small sample sizes. It addresses the difficulty in obtaining high-quality remote sensing images and annotations. Furthermore, previous methods often introduced complex structures, leading to heavy computational burdens and hindering practical deployment. This invention, however, significantly improves accuracy while maintaining reasonable computational complexity, providing a systematic solution for the further deployment of target detection models with small sample sizes, thus solving the problem of the difficulty in deploying such models.

[0088] Eleventh, this invention not only avoids falling into the trap of existing technological biases, but also systematically overcomes several typical technological biases that have long existed in the field of remote sensing small-sample target detection, from its theoretical foundation to its architectural design. The industry often simply attributes the small-sample problem to the limited number of samples, ignoring the underlying complex issues such as feature distribution shifts and domain differences. This invention, through probability distribution modeling and optimal transport theory, deepens the problem into modeling and aligning data distribution and domain differences, fundamentally providing a more systematic solution. Attached Figure Description

[0089] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure;

[0090] Figure 1 This is a flowchart of a remote sensing small-sample target detection method based on a generative meta-learning strategy provided in an embodiment of the present invention;

[0091] Figure 2 This is a roadmap of a remote sensing small-sample target detection method based on a generative meta-learning strategy provided in this embodiment of the invention;

[0092] Figure 3 This is a diagram of a dual-stream collaborative network structure provided in an embodiment of the present invention;

[0093] Figure 4 This is a structural diagram of a generative meta-learner based on optimal transmission provided in an embodiment of the present invention;

[0094] Figure 5 This is a structural diagram of the orthogonal frequency domain decomposition detection head provided in an embodiment of the present invention;

[0095] Figure 6 This is a comparison chart of detection performance on the DIOR dataset provided in this embodiment of the invention;

[0096] Figure 7 This is a comparison chart of the detection performance on the NWPU VHR-10 and HRRSD datasets provided in this embodiment of the invention. Detailed Implementation

[0097] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0098] The innovations of this invention are as follows: It proposes a generative meta-learning paradigm that fundamentally solves the problem of fragile feature representation in small samples through probabilistic modeling and variational feature generation. Theoretically, it innovatively introduces optimal transport theory for cross-domain robustness measurement, effectively overcoming the challenges of large intra-class differences and high inter-class similarities in remote sensing images. Architecturally, this invention designs a gated dual-stream dynamic fusion network to achieve efficient modeling of long-range contexts with linear computational complexity, providing the industry with a groundbreaking, practical, and engineeringable systematic solution.

[0099] Example 1, such as Figure 1 As shown, the remote sensing small-sample target detection method based on a generative meta-learning strategy provided in this embodiment of the invention includes the following steps:

[0100] S1. Prepare training samples of satellite remote sensing images containing support sets and query sets, construct a K-way N-shot meta-learning training task, and simulate a small sample detection environment in a real scene.

[0101] A systematic approach is taken to prepare training samples of satellite remote sensing images suitable for few-shot learning, clearly defining the support set and query set. The support set strictly follows the K-way N-shot meta-learning paradigm, containing K target categories, with each category providing N samples precisely labeled with bounding boxes. The query set contains unlabeled image sequences from the same category. This task construction mechanism effectively simulates the extremely scarce sample detection environment in real-world remote sensing scenarios, laying the foundation for evaluating the model's generalization ability.

[0102] S2 will support the dynamic relationship between the image input of the support set and the query set in a dual-stream collaborative network. By improving the non-causal Mamba unit, two-dimensional position coding and gating fusion mechanism, dual-stream feature extraction and spatial semantic fusion are performed to achieve dynamic alignment and enhancement of support and query features.

[0103] This network supports a dynamic dual-stream collaborative network that allows parallel input of images from the support set and query set into dynamic relationships. One path employs an improved non-causal Mamba unit combined with 2D location encoding, focusing on modeling long-range spatial dependencies of multi-scale, irregular features; the other path utilizes a self-attention mechanism to capture the global semantic context of the images. The two feature paths are adaptively weighted and integrated through a gated fusion mechanism, preserving the computational efficiency advantages of the state-space model while fully incorporating the global modeling capabilities of self-attention, significantly enhancing the feature representation power for complex remote sensing scenes. Furthermore, through generative feature enhancement and distributed prototype representation techniques, a more discriminative unified feature space is constructed based on the dual-stream features, achieving dynamic cross-sample alignment and semantic enhancement of features from the support set and query set, fundamentally improving the model's generalization performance under few-sample conditions.

[0104] The dynamic relationship dual-stream collaborative network consists of three parts: non-causal Mamba units, two-dimensional multi-head self-attention, and an adaptive gating fusion mechanism. Based on the original Mamba structure, the spatial perception mechanism is improved, and a gating mechanism is introduced to adaptively adjust features.

[0105] Input feature tensor Remodeling into a sequence Represented as:

[0106]

[0107] In the formula, For the reconstructed feature sequence, For shape reshaping calculations, To reshape the sequence, For batch size, These represent the number of channels, height, and width of the position embedding tensor, respectively.

[0108] Learnable relative position embeddings are introduced along the height and width dimensions, respectively. The expression is:

[0109]

[0110] In the formula, For learnable relative position embeddings, These are trainable parameters, used to capture the positional biases in the vertical and horizontal directions, respectively.

[0111] Injecting location information into sequence features, the expression is:

[0112]

[0113] In the formula, To incorporate sequence features that combine location information, Transpose of relative position embedding;

[0114] A spatial context-aware gating system is constructed using one-dimensional convolution and the Sigmoid activation function to adaptively adjust the feature response intensity; the sequence is routed via a non-causal Mamba block. Processing to capture long-range context dependencies, Mamba output is coupled with a gated signal. To perform element-by-element fusion, the expression is:

[0115]

[0116] In the formula, For output features, For the output characteristics of Mamba blocks, For activation function, For convolution calculation, This is the transpose of the sequence features;

[0117] Finally, the output sequence is reshaped back into spatial dimensions. In the form of.

[0118] Furthermore, based on the Mamba structure, a two-dimensional multi-head self-attention mechanism is introduced to form a two-stream collaborative structure together with Mamba; by explicitly modeling the interaction relationships between all spatial locations, the complete global context of features is captured.

[0119] Input features Convert to a query Q, key K, and value V, where, The dimension of the K vector is used to calculate the attention weights. ;

[0120]

[0121] In the formula, Calculate for the softmax activation function. It is the transpose of the K vector;

[0122] Introducing two-dimensional relative position encoding enables the attention mechanism to perceive the relative geometric relationship between positions;

[0123] A gated adaptive fusion mechanism is designed to equally divide the input features along the channel dimension, resulting in evenly distributed features. and The input consists of two branches: Mamba and self-attention, which are learned gating systems. and The expression for combining the outputs of the two branches is:

[0124]

[0125]

[0126] In the formula, As a feature of fusion, As a characteristic of being evenly distributed to the Mamba network, Features are evenly distributed to the self-attention network. For learnable gated outputs of the Mamba branch, For learnable gated outputs of self-attention branches, All are learnable projection weights.

[0127] In step S3, the generative meta-learner based on optimal transmission represents the support set categories as probability distributions, including:

[0128] Prior distribution Set as a standard multivariate Gaussian distribution Approximate posterior distribution It is then modeled as a multivariate Gaussian distribution with a diagonal covariance matrix. ; distribution parameter mean and variance Through feature encoder The expression for generation is:

[0129]

[0130] In the formula, The distribution of supporting features is calculated for the feature encoder;

[0131] Variational features generated using reparameterization techniques The calculation process is as follows:

[0132]

[0133] In the formula, For a bias that follows a standard Gaussian distribution;

[0134] The lower bound of evidence is decomposed into two loss terms with clear physical meaning: reconstruction loss and KL divergence loss; reconstruction loss Defined as input features With the decoder Reconstructed features L2 distance between them;

[0135]

[0136] KL divergence loss Represented as:

[0137]

[0138] In the formula, For KL divergence calculation, Given an observation sequence S, what is the posterior distribution of the latent variable z? Let z be the prior distribution of the latent variable z;

[0139] Introducing consistency loss The loss function is defined as the reconstructed features. After a lightweight linear classifier The output results are compared with the true category labels. Cross-entropy between ;

[0140]

[0141] In the variational feature aggregation stage, variational features sampled from the distribution are... By deeply integrating with query features, the meta-learner employs a category-independent aggregation strategy.

[0142] Also includes: given category Query features and categories Support features Accurately estimate the corresponding category distribution Variational features are obtained through sampling. The aggregate expression is:

[0143]

[0144] In the formula, The aggregated features of the output This is an aggregation operation between two features. The mean of the category distribution. For the variance of the class distribution, For bias, It is the sigmoid activation function.

[0145] In step S4, a distance metric based on optimal transmission is introduced on the basis of the generative meta-learner. The class similarity is measured by calculating the Wasserstein distance between the query sample and the supporting class distribution.

[0146] Furthermore, the feature representation of the query sample Considering it as a Dirac distribution, the first The support categories are represented as a uniform distribution. The corresponding number Each support category sample is represented as , The formula for calculating the Wasserstein distance between the two is:

[0147]

[0148] In the formula, This is the feature representation obtained by mapping the query samples. For the first One supporting sample, The output is the Wasserstein distance between the two. To support the sample size, The distance is Euclidean.

[0149] The predicted class of the query sample is represented by the minimum Wasserstein distance:

[0150]

[0151] In the formula, To query the predicted class label of a sample, The goal is to select the index that minimizes the squared Wasserstein distance from among N candidate category indices.

[0152] After obtaining the enhanced support set category probability distribution in the previous step, the semantic distance between it and the query sample feature distribution is accurately measured using optimal transport theory. Specifically, a robust cross-domain similarity criterion is constructed by calculating the Wasserstein distance between the two. This distance metric is insensitive to changes in distribution morphology and can more accurately reflect the true semantic relationships between categories, thus providing a more reliable theoretical basis for the classification of query samples and establishing a precise feature correspondence between the support set and the query set.

[0153] S5 inputs the query features enhanced by optimal transmission into the orthogonal frequency domain decomposition detection head for frequency domain analysis and feature reconstruction, and adaptively decomposes the features into low-frequency contour components and high-frequency detail components.

[0154] The query features, after optimal transmission alignment, are fed into a specially designed orthogonal frequency domain decomposition detection head. This module adaptively decomposes spatial features into low-frequency components carrying target contour information and high-frequency components containing detailed textures through lightweight and fast orthogonal transformation. Based on this, frequency domain filtering and key frequency point selection are implemented, effectively enhancing the amplification of discriminative features and significantly suppressing background clutter and high-frequency noise interference commonly found in remote sensing images. This process is completed in the frequency domain, making it computationally efficient and achieving the goal of further improving feature quality while maintaining spatial domain sensing capabilities.

[0155] Frequency domain analysis and feature reconstruction include:

[0156] The channel dimension is expanded by point convolution, and spatial features are extracted using depthwise separable convolution;

[0157] The feature map is equally divided into two branches in the channel dimension. One branch is activated by the GELU function, and the other branch retains the linear features. The two branches are multiplied element-wise, and the number of channels is restored by point convolution.

[0158] The spatial feature map is divided into 8×8 pixel blocks. A fast Fourier transform is performed on each image block to convert it to the frequency domain. A learnable frequency domain filter is used to adaptively weight the spectrum, and selective enhancement and suppression of frequency components are achieved by multiplying element by element.

[0159] Perform an inverse Fourier transform on the filtered spectrum to convert the processed frequency domain information back into spatial domain image patches; reassemble all processed image patches into a complete feature map, remove the filled boundary parts, and finally output the enhanced feature with the same size as the input.

[0160] S6, based on the reconstructed features, outputs the target's class probability distribution and accurate bounding box coordinate regression values ​​through a multi-scale prediction head. Non-maximum suppression post-processing is used to generate the final predicted bounding box. During training, the intersection-union ratio (IoU) between the predicted box and the ground truth label is calculated to optimize the model parameters.

[0161] During training, the Intersection over Union (IoU) between the predicted bounding boxes and the ground truth labels is calculated to optimize model parameters. This includes: the IoU score is defined as the ratio between the intersection and union of two boxes, and the formula for calculating IoU is:

[0162]

[0163] In the formula, There are two boxes respectively. The region is the union of the two boxes. This represents the intersection of the two boxes.

[0164] Based on the reconstructed enhanced features, a multi-scale prediction head is used to perform two core tasks in parallel: first, generating the class probability distribution of the target at each spatial location; and second, performing accurate bounding box coordinate regression. During the inference phase, a non-maximum suppression post-processing algorithm is employed to remove highly overlapping redundant prediction boxes, generating a final, concise, and accurate detection result. During the model training phase, the classification loss and regression loss are comprehensively calculated, and optimized by combining the intersection-over-union (IoU) ratio of the predicted and ground truth bounding boxes, driving the entire system to continuously improve the accuracy and robustness of small-sample target detection in an end-to-end manner.

[0165] Example 2, as Figure 2 As shown in the embodiment of the present invention, the remote sensing small-sample target detection method based on a generative meta-learning strategy is used to train a dynamic relational dual-stream collaborative network on remote sensing images containing support sets and query sets, thereby obtaining an improved generative meta-learning target detection model. The dynamic relational dual-stream collaborative network of this model extracts deep features from support samples and query samples respectively through a parallel dual-stream architecture, and uses a gated fusion mechanism to achieve feature enhancement and dynamic alignment. A generative meta-learner based on optimal transmission performs distribution enhancement and prototype reconstruction in the latent space, effectively improving the discriminative power of category representation. Finally, an orthogonal frequency domain decomposition detection head performs frequency domain analysis and reconstruction on the enhanced features, directly completing target classification and bounding box regression based on the reconstructed features. After generating preliminary detection results, the NMS method is used to filter and obtain the final detection box, and the target category, confidence score, and location information are output simultaneously. The specific steps are as follows:

[0166] like Figure 3 As shown, dual-stream feature extraction and spatial semantic fusion are performed. The dynamic relationship between the support set and query set images is input into the dual-stream collaborative network. Through its improved non-causal Mamba unit and two-dimensional location encoding, the spatial representation capability of multi-scale and irregular land features is enhanced. At the same time, the state space model and self-attention advantage are integrated by combining the gating fusion mechanism. While realizing long-range dependency and global context modeling, the computational complexity is reduced. On the basis of the dual-stream architecture, through feature enhancement and distributed prototype representation, a highly discriminative feature space is constructed to realize the dynamic alignment and enhancement of support features and query features, effectively improving the generalization performance of the model under the condition of scarce samples.

[0167] like Figure 4 As shown, by utilizing a generative meta-learner based on optimal transfer, the categories in the support set are represented as probability distributions rather than single-point estimates. By integrating generative models to enhance the distribution of the support set in the latent space, richer and more reliable category distribution estimates are provided for optimal transfer computation, avoiding bias on specific samples.

[0168] By precisely measuring the similarity between the category distribution of the augmented support set obtained in step 4 and the feature distribution of the query samples using optimal transport theory, the Wasserstein distance is calculated to provide a more robust and accurate distance metric for the query samples, thus establishing a feature correspondence between the support set and the query set.

[0169] like Figure 5 As shown, frequency domain analysis and feature reconstruction are performed. The query features enhanced by optimal transmission are input into the orthogonal frequency domain decomposition detection head. Orthogonal frequency domain analysis and screening are performed through a lightweight frequency domain processing step. The features are adaptively decomposed into low-frequency contour components and high-frequency detail components. While maintaining spatial domain computation efficiency, the discriminativeness of frequency domain features is significantly enhanced, and background noise interference in remote sensing images is effectively suppressed.

[0170] Based on the reconstructed enhanced features, the target's class probability distribution and accurate bounding box coordinate regression values ​​are output through a multi-scale prediction head. Non-maximum suppression post-processing is then used to generate the final predicted bounding box, such as... Figure 6 and Figure 7 As shown.

[0171] Example 3: The remote sensing small-sample target detection system based on a generative meta-learning strategy provided in this embodiment of the invention includes:

[0172] The meta-learning task construction module is used to prepare training samples of satellite remote sensing images containing support sets and query sets, and to construct K-way N-shot meta-learning tasks to simulate small sample detection environments in real-world scenarios, providing the model with a training data foundation that conforms to the meta-learning paradigm.

[0173] The feature map acquisition module for predictable targets is used to input the support set and query set into the dynamic relational dual-stream collaborative network. By utilizing the improved non-causal Mamba unit and gated fusion mechanism, dual-stream feature extraction and spatial semantic fusion are achieved, and finally the dynamic alignment and enhancement of support features and query features are completed, improving the initial feature quality of the model under the condition of few samples.

[0174] The generative support feature distribution estimation module is used to improve the category estimates in the support set from single-point estimates to probability distribution representations by using a generative meta-learner based on optimal transfer. It also performs distribution enhancement in the latent space through a generative model, thereby providing richer and more robust category prototype estimates for subsequent computations.

[0175] The query-support relationship measurement module is used to accurately calculate the Wasserstein distance between the enhanced class distribution and the feature distribution of the query sample using optimal transport theory, thereby establishing a more robust feature correspondence between the support set and the query set and providing a reliable distance metric for classification.

[0176] The frequency domain decomposition and feature reconstruction module is used to send the query features after relation measurement into the orthogonal frequency domain decomposition detection head, and adaptively decompose them into low-frequency contour and high-frequency detail components through lightweight frequency domain analysis, effectively suppressing complex background noise in remote sensing images while performing feature reconstruction.

[0177] The detection output and target bounding box acquisition module is used to process the reconstructed enhanced features using a multi-scale prediction head, output the target's class probability and precise bounding box coordinates, and generate the final detection result using non-maximum suppression. During the training phase, the model parameters are continuously optimized by calculating the intersection-over-union (IoU) ratio between the predicted bounding box and the ground truth label.

[0178] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solution. Three publicly available remote sensing datasets are used: the DIOR dataset (23,463 images, 20 categories), the NWPU VHR-10 dataset (800 images, 10 categories), and the HRRSD dataset (21,761 images, 8 categories selected from 13 categories). Each dataset is divided into base classes and new classes according to the few-shot learning paradigm. DIOR and HRRSD use two sets of partitioning schemes, while NWPU VHR-10 uses one set. Small-shot experimental environments are constructed by setting different sample sizes (shots) of 1, 3, 5, and 10 for each new class.

[0179] The training environment was built on the PyTorch deep learning framework and ran on an Ubuntu system. An NVIDIA Tesla A800 GPU was used for accelerated training, and the MMFewShot library and Python scientific computing libraries were employed to ensure experimental consistency. Model performance was evaluated using mAP50 (mean accuracy at an IoU threshold of 0.5) as the core metric. Additionally, ablation experiments included multiple evaluation metrics such as mAP75, base class mean accuracy (bAP), and novel class mean accuracy (nAP) to ensure fair comparison with existing technologies and comprehensive performance analysis.

[0180] The AdamW algorithm was used as the optimizer during the training phase of this model, and the first 100 iterations were set as a learning rate warm-up phase. Specific training parameters were set as follows: the batch size was set to 8 during the base class training phase, and to 16 during the new class fine-tuning phase. This enabled the model to quickly generalize to new classes. Different model optimization strategies were implemented for training on datasets of varying sizes.

[0181] For base class training of the DIOR and HRRSD datasets, each input batch was trained in one step, resulting in a total of 60,000 steps of model parameter optimization. Correspondingly, for the NWPU VHR-10 dataset, base class training underwent 110,000 optimization steps to cover all instances in the dataset. During the base class training phase, the initial learning rate was uniformly set to 0.02, with a learning rate scheduling factor of 0.1. Specifically, the learning rate for the DIOR and HRRSD datasets decreased by a factor of 10 at iterations 40,000 and 50,000, while for the NWPU VHR-10 dataset, the learning rate was adjusted at iterations 80,000 and 100,000.

[0182] During the meta-learner fine-tuning phase, the optimization process on the DIOR and HRRSD datasets used an initial learning rate of 0.001 for 40,000 steps of model fine-tuning, with the learning rate decreasing by a factor of 10 at the 36,000th optimization step. Correspondingly, on the NWPU VHR-10 dataset, an initial learning rate of 0.005 was used for 80,000 optimization steps, with the learning rate reduced at the 70,000th optimization step to achieve the optimal performance of the model.

[0183] Through the standardized experimental setup described above, the adaptability and effectiveness of this invention under different resolutions and scene complexities were verified, providing a reliable performance benchmark for remote sensing target detection with small sample sizes.

[0184] Furthermore, to quantitatively demonstrate the model's accuracy performance in different remote sensing application scenarios, this experiment conducts a horizontal accuracy comparison with other solutions in this field on the same dataset, providing a quantitative reference standard for this invention. This experiment uses the DIOR and NWPU VHR-10 processed small-sample optical remote sensing image datasets as evaluation benchmarks for the comparative experiment, and also reports the accuracy performance of models with different backbone network depths on the HRRSD dataset.

[0185] First, a precision comparison experiment was conducted on the DIOR dataset. In this experiment, mAP50 (average precision when the IoU threshold is 0.5) was used as the evaluation index for remote sensing target detection precision across all categories. The detection precision of the method of this invention was compared with that of some small sample target detection models based on meta-learning. The experimental comparison results are shown in Table 1.

[0186] Table 1 Comparison of detection model accuracy on the DIOR dataset

[0187] Subsequently, similar comparative experiments were conducted on the NWPU VHR-10 dataset. The experiments primarily used small sample settings of 10 shots and 30 shots on the NWPU VHR-10 dataset, with nAP50 and nAP75 (IoU thresholds of 0.5 and 0.75, respectively) for the new class as evaluation metrics. The results show that, compared to other meta-learning methods, the proposed method achieves a certain degree of improvement in average accuracy for small-sample new class detection. The experimental comparison results are shown in Table 2.

[0188] Table 2. Comparison of algorithm accuracy on the NWPU VHR-10 dataset.

[0189] Meanwhile, to systematically evaluate the generalization performance and robustness of model architectures under different feature extraction capabilities, and to explore the impact of backbone network depth on few-shot remote sensing detection tasks, this experiment further compared the overall accuracy performance of models employing backbone network architectures of various depths on the HRRSD dataset. The experiment started with two different backbone network depths and compared their detection accuracy from the perspective of detection accuracy. The experimental comparison results are shown in Table 3.

[0190] Table 3. Accuracy report of the detection model on the HRRSD dataset

[0191] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A remote sensing small-sample target detection method based on a generative meta-learning strategy, characterized in that, The method includes the following steps: S1. Prepare training samples of satellite remote sensing images containing support sets and query sets, construct a K-way N-shot meta-learning training task, and simulate a small sample detection environment in a real scene. S2 will support the dynamic relationship between the image input of the support set and the query set in a dual-stream collaborative network. By improving the non-causal Mamba unit, two-dimensional position coding and gating fusion mechanism, dual-stream feature extraction and spatial semantic fusion are performed to achieve dynamic alignment and enhancement of support and query features. S3, the generative meta-learner based on optimal transmission represents the support set categories as probability distributions, and enhances the distribution of the support set in the latent space by integrating generative models; S4. The Wasserstein distance between the enhanced support set category distribution and the query sample feature distribution is calculated using optimal transport theory to establish the feature correspondence between the support set and the query set. S5 inputs the query features enhanced by optimal transmission into the orthogonal frequency domain decomposition detection head for frequency domain analysis and feature reconstruction, and adaptively decomposes the features into low-frequency contour components and high-frequency detail components. S6, based on the reconstructed features, outputs the target's class probability distribution and accurate bounding box coordinate regression values ​​through a multi-scale prediction head. Non-maximum suppression post-processing is used to generate the final predicted bounding box. During training, the intersection-union ratio (IoU) between the predicted box and the ground truth label is calculated to optimize the model parameters.

2. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S1, the support set contains K target categories, with N labeled samples for each category; the query set contains unlabeled image sequences from the same category.

3. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S2, the dynamic relational dual-stream collaborative network consists of three parts: a non-causal Mamba unit, a two-dimensional multi-head self-attention mechanism, and an adaptive gating fusion mechanism. The spatial perception mechanism is improved based on the original Mamba structure, and the gating mechanism is introduced to adaptively adjust the features. Input feature tensor Remodeling into a sequence Represented as: ； In the formula, For the reconstructed feature sequence, For shape reshaping calculations, To reshape the sequence, For batch size, These represent the number of channels, height, and width of the position embedding tensor, respectively. Learnable relative position embeddings are introduced along the height and width dimensions, respectively. The expression is: ； In the formula, For learnable relative position embeddings, These are all trainable parameters, used to capture position biases in the vertical and horizontal directions, respectively; Injecting location information into sequence features, the expression is: ； In the formula, To incorporate sequence features that combine location information, Transpose of relative position embedding; By using one-dimensional convolution and the Sigmoid activation function, a spatial context-aware gating system is constructed to adaptively adjust the feature response intensity. Sequence via noncausal Mamba block Processing to capture long-range context dependencies, Mamba output is coupled with a gated signal. To perform element-by-element fusion, the expression is: ； In the formula, For output features, For the output characteristics of Mamba blocks, For activation function, For convolution calculation, This is the transpose of the sequence features; Finally, the output sequence is reshaped back into spatial dimensions. In the form of.

4. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 3, characterized in that, Based on the Mamba structure, a two-dimensional multi-head self-attention mechanism is introduced to form a two-stream collaborative structure together with Mamba; by explicitly modeling the interaction relationships between all spatial locations, the complete global context of features is captured. Input features Convert to a query Q, key K, and value V, where, The dimension of the K vector is used to calculate the attention weights. ; ； In the formula, Calculate for the softmax activation function. It is the transpose of the K vector; Introducing two-dimensional relative position encoding enables the attention mechanism to perceive the relative geometric relationship between positions; A gated adaptive fusion mechanism is designed to equally divide the input features along the channel dimension, resulting in evenly distributed features. and The input consists of two branches: Mamba and self-attention, which are learned gating systems. and The expression for combining the outputs of the two branches is: ； ； In the formula, As a feature of fusion, As a characteristic of being evenly distributed to the Mamba network, Features are evenly distributed to the self-attention network. For learnable gated outputs of the Mamba branch, For learnable gated outputs of self-attention branches, All are learnable projection weights.

5. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S3, the generative meta-learner based on optimal transmission represents the support set categories as probability distributions, including: Prior distribution Set as a standard multivariate Gaussian distribution Approximate posterior distribution It is then modeled as a multivariate Gaussian distribution with a diagonal covariance matrix. ; distribution parameter mean and variance Through feature encoder The expression for generation is: ； In the formula, The distribution of supporting features is calculated for the feature encoder; Variational features generated using reparameterization techniques The calculation process is as follows: ； In the formula, For a bias that follows a standard Gaussian distribution; The lower bound of evidence is decomposed into two loss terms with clear physical meaning: reconstruction loss and KL divergence loss; reconstruction loss Defined as input features With the decoder Reconstructed features L2 distance between them; ； KL divergence loss Represented as: ； In the formula, For KL divergence calculation, Given an observation sequence S, what is the posterior distribution of the latent variable z? Let z be the prior distribution of the latent variable z; Introducing consistency loss The loss function is defined as the reconstructed features. After a lightweight linear classifier The output results are compared with the true category labels. Cross-entropy between ; ； In the variational feature aggregation stage, variational features sampled from the distribution are... By deeply integrating with query features, the meta-learner employs a category-independent aggregation strategy.

6. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 5, characterized in that, Also includes: Given category Query features and categories Support features Accurately estimate the corresponding category distribution Variational features are obtained through sampling. The aggregate expression is: ； In the formula, The aggregated features of the output This is an aggregation operation between two features. The mean of the category distribution. For the variance of the class distribution, For bias, It is the sigmoid activation function.

7. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S4, a distance metric based on optimal transmission is introduced on the basis of the generative meta-learner. The class similarity is measured by calculating the Wasserstein distance between the query sample and the supporting class distribution.

8. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 7, characterized in that, Representing the features of the query sample Considering it as a Dirac distribution, the first The support categories are represented as a uniform distribution. The corresponding number Each support category sample is represented as , The formula for calculating the Wasserstein distance between the two is: ； In the formula, This is the feature representation obtained by mapping the query samples. For the first One supporting sample, The output is the Wasserstein distance between the two. To support the sample size, The distance is Euclidean. The predicted class of the query sample is represented by the minimum Wasserstein distance: ； In the formula, To query the predicted class label of a sample, In order to be in Among the candidate category indices, the index that minimizes the squared Wasserstein distance is selected.

9. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S5, the frequency domain analysis and feature reconstruction include: The channel dimension is expanded by point convolution, and spatial features are extracted using depthwise separable convolution; The feature map is equally divided into two branches in the channel dimension. One branch is activated by the GELU function, and the other branch retains the linear features. The two branches are multiplied element-wise, and the number of channels is restored by point convolution. The spatial feature map is divided into 8×8 pixel blocks. A fast Fourier transform is performed on each image block to convert it to the frequency domain. A learnable frequency domain filter is used to adaptively weight the spectrum, and selective enhancement and suppression of frequency components are achieved by multiplying element by element. Perform an inverse Fourier transform on the filtered spectrum to convert the processed frequency domain information back into spatial domain image patches; reassemble all processed image patches into a complete feature map, remove the filled boundary parts, and finally output the enhanced feature with the same size as the input.

10. The remote sensing small-sample target detection method based on generative meta-learning strategy according to claim 1, characterized in that, In step S6, during training, the Intersection over Union (IoU) between the predicted bounding boxes and the ground truth labels is calculated to optimize the model parameters. This includes: the target box score calculated using the IoU is defined as the ratio between the intersection and union of two boxes, and the formula for calculating IoU is: ； In the formula, There are two boxes respectively. The region is the union of the two boxes. This represents the intersection of the two boxes.

Images