SAR image small sample target detection system and method based on context perception and Gaussian flow representation

Abstract

The invention discloses a SAR image small sample target detection system and method based on context perception and Gaussian flow representation, belonging to the field of target recognition technology. The system comprises: an image acquisition module, a context perception enhancement module, a manifold class distribution estimation module, a class balanced difference aggregation module and a cosine decoupling module; the image acquisition module is used to obtain a SAR image to be queried; the context perception enhancement module extracts the most semantically representative support features into support class prototypes and obtains original query features; the manifold class distribution estimation module converts the support features into complex class distributions; the class balanced difference aggregation module performs deep semantic feature aggregation on original query features and support features of different categories based on class distribution to obtain aggregated query features; the cosine decoupling module processes the original query features and aggregated query features using a classification branch that has been processed with feature normalization, cosine similarity measurement and a learnable scaling factor to obtain a target recognition result.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and in particular relates to a SAR image small sample target detection system and method based on context perception and Gaussian flow characterization. Background Art

[0002] In the past decade, with the substantial improvement of computer hardware computing power and storage capacity, deep neural networks have developed rapidly, promoting technological changes in the field of target detection and recognition. In the early stage, there were mainly two-stage detectors that first generated target candidate regions and then performed refined classification and regression tasks. The two-stage detector achieved good detection and recognition accuracy through a rough-to-fine processing method, but at the same time, it had complex network structure, low operating efficiency and insufficient real-time performance. In order to be applied to real-time scenarios, researchers abandoned the method of generating target candidate regions and switched to a single-stage detector that directly predicted the target box and target category. In addition to the above two mainstream detectors, there are also anchor-free detectors that reduce the complexity of the model by removing the pre-set anchor box size. Benefiting from the excellent detection and recognition performance in natural scenes, these general detectors have also been applied by researchers to complex SAR scenes, and many improvements have emerged. However, the existing detection models are designed with data-driven as the core, and the detection accuracy of the model is heavily dependent on the number of labeled samples. When the amount of labeled sample data is insufficient, the model will fall into an overfitting state because the sparse labeled samples cannot fit the high-dimensional function of the deep neural network.

[0003] Inspired by the fact that humans can recognize a new object from a very small number of samples, researchers hope that machine learning models can also have the ability to learn quickly, so small sample learning came into being. Small sample learning is to extract some category-independent prior knowledge by pre-learning massive data of basic categories and transfer it to the recognition task of a very small number of new categories, solving the overfitting problem of the model caused by the small number of training samples. Currently, small sample learning is mainly used for classification tasks in the field of computer vision, and has achieved many groundbreaking results. Benefiting from the excellent recognition performance in natural scenes, these small sample learnings have also been applied by researchers to complex remote sensing scenes, and many improvements specific to the remote sensing field have emerged. However, most research work focuses on classification tasks, and there are serious deficiencies in the research methods that combine more complex target detection tasks with small sample learning.

[0004] Small sample target detection: Small sample target detection technology has become a hot research direction because of its ability to quickly identify new categories under the condition of a small number of labeled samples. At present, it is mainly divided into research based on metric learning, transfer learning and meta-learning: 1) Small sample target detection based on metric learning. This detection method adopts the idea of ​​learning to compare category similarity and achieves good classification performance for sparsely labeled new category targets. The detector first obtains the features of the target to be detected, and maps it and the supporting image features to a latent feature space, and classifies and detects the target by measuring the class prototype similarity distance. The model structure of this type of detector is relatively simple, and it is easy to add new categories, but there is a problem of poor positioning accuracy for new category targets. 2) Small sample target detection based on transfer learning. The transfer learning method is also called the fine-tuning method. The core idea is to first pre-train on richly labeled basic class data, and then realize transfer learning on sparse new class data by freezing some parameters of the trained model. TFA first achieved good detection and recognition results for new class data by adding a cosine classifier to the Faster R-CNN architecture and only fine-tuning the last layer of the detector. FSCE reduces the similarity between different categories of objects by introducing a contrastive learning branch on the architecture of TFA, thereby improving the detection and recognition accuracy of the model for new categories. DeFRCN alleviates the potential contradictions caused by the application of the Faster R-CNN architecture to sparse data scenarios by designing a gradient decoupling module GDL and a prototype calibration module PCB. SSR-FSD combines the semantic relationship and visual information between categories, and introduces class relationship reasoning into the detection model to improve the detection and recognition accuracy on sparse new class data. The core of this type of detector is the data-driven training strategy, which has the advantages of simple learning strategy and relatively high detection accuracy, but the fine-tuning method makes the model prone to domain shift problems. 3) Small sample object detection based on meta-learning. The meta-learning method is a mainstream method in the field of small sample learning. Its core idea is to train the model by simulating small sample tasks as units, and only a small number of iterations of very little data are needed based on the trained parameters to quickly generalize to new tasks. FSRW adds a meta-feature extractor and a reweighting module to the single-stage detection framework, and realizes the detection and recognition of new class objects by supporting the reweighting of meta-features and query features. Meta R-CNN obtains the class prototype vector of the support category by adding a branch that shares the parameters of the backbone network, and then fuses each vector with the features of the query image to detect the object in the query image. FsDetView focuses on the deep fusion of support and query features, and improves the recognition accuracy of new categories by changing the original dot product aggregation method to a new hybrid method.

[0005] However, most of the above models are designed for natural scenes. When they are introduced into the remote sensing field of ultra-long-distance observation, especially when applied to SAR images of complex and sparse scenes, the detection and recognition performance of the models is greatly reduced. Summary of the invention

[0006] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0007] A SAR image small sample target detection system based on context perception and Gaussian flow characterization, including: image acquisition module, context perception enhancement module, manifold class distribution estimation module, class balanced difference aggregation module and cosine decoupling module;

[0008] The image acquisition module is used to obtain the SAR image to be queried;

[0009] The context-aware enhancement module extracts the most semantically representative support features into support class prototypes, and enhances the query features of the SAR image to be queried based on the support class prototypes to obtain original query features;

[0010] The manifold class distribution estimation module converts the support features into complex class distributions using a normalized flow-based variational autoencoder;

[0011] The class-balanced differential aggregation module performs deep semantic feature aggregation on the original query features and the supporting features of different categories based on the class distribution, learns the feature representation between classes, and obtains the aggregated query features;

[0012] The cosine decoupling module processes the original query features and the aggregated query features using a classification branch that has been processed with feature normalization, cosine similarity measurement, and a learnable scaling factor to obtain a target recognition result.

[0013] Preferably, the workflow of the context awareness enhancement module includes:

[0014] In the subspace of the support features that represents the most semantic information, assume that a support feature , and the support features Perform a global average pooling operation to obtain the global feature query matrix Q:

[0015] ,

[0016] Among them, GAP( ) represents global average pooling, represents the real number space, d represents the feature dimension;

[0017] The support features are transformed through a linear layer Project to the query key space and generate the key matrix K:

[0018] ,

[0019] in, represents the weight matrix of the linear layer, , h represents the height of the feature, w represents the width of the feature, and k represents the expected dimension of the feature after the change;

[0020] Support features for Apply a convolutional layer with a kernel size of 1, and get the value matrix V:

[0021] ,

[0022] Among them, Conv 1×1 ( ) indicates a convolution operation with a kernel size of 1;

[0023] Based on the global feature query matrix Q, the key matrix K and the value matrix V, a support class prototype C is extracted:

[0024] ,

[0025] Among them, softmax() represents the normalized exponential function, and T represents the transpose operation of the matrix;

[0026] Given a query feature , based on the support class prototype C, we get the affinity matrix A:

[0027] ,

[0028] in, express and the shared linear projection of C;

[0029] Based on the affinity matrix A, the query feature of the SAR image to be queried is enhanced to obtain the original query feature :

[0030] ,

[0031] Among them, α represents a learnable parameter.

[0032] Preferably, the workflow of the manifold class distribution estimation module includes:

[0033] Support for image collections After feature extraction, we get the supporting feature set ,in, are mutually independent characteristic variables;

[0034] The supporting feature set After the encoder maps the value into the latent space, the support class distribution is obtained The mean μ and variance σ of , the mean μ and the variance σ are reparameterized to generate the initial simple latent variable z0:

[0035] ,

[0036] Among them, ε represents the random noise of standard normal distribution;

[0037] The latent variable z0 is transformed into a popular variable z with more actual class distribution information. k :

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] Among them, z' represents the latent variable obtained after manifold transformation, z represents the original latent variable, represents the manifold transformation function, u, w and b represent learnable parameters, represents the Jacobian of the popular transformation, represents the gradient calculation during back propagation, f i represents a reversible epidemic change, i=1,2,...,k;

[0043] Based on the latent variable z0 the latent variable z k , by maximizing the lower bound of evidence Transformed to a complex class distribution.

[0044] Preferably, the workflow of the class-balanced differential aggregation module includes:

[0045] Based on the original query features , let the query features of query class t be ;

[0046] Sample the potential support features supporting class j from the complex class distribution , and the potential support features are aggregated by the dot product method With the query characteristics Aggregate to get common features :

[0047] ,

[0048] in, Indicates channel multiplication, sig indicates sigmoid operation;

[0049] The potential supporting features With the query characteristics Perform differential operation to obtain differential features :

[0050] ;

[0051] The query feature , the common characteristics and the differential characteristics Input into the attention encoding module for fusion to obtain the aggregated query features :

[0052] ,

[0053] Among them, Attention represents the attention operation, d k express feature dimension.

[0054] Preferably, the working process of the cosine decoupling module includes:

[0055] The query features are processed using a classification branch processed with feature normalization, cosine similarity metric and learnable scaling factor And the query features after aggregation , and get the target recognition result:

[0056] ,

[0057] ,

[0058] Where N represents the classification score, represents a classification network, represents the feature extractor for classification, scale represents the learnable scaling factor, θ i Represents the angle between class vectors, W i represents the weight vector of each class, B represents the predicted bounding box, represents the regression network, Represents a feature extractor for regression.

[0059] The present invention also provides a SAR image small sample target detection method based on context perception and Gaussian flow characterization, the method is applied to any of the above-mentioned systems, comprising the following steps:

[0060] S1. Obtain the SAR image to be queried;

[0061] S2. extracting the most semantically representative support features into support class prototypes, and enhancing the query features of the SAR image to be queried based on the support class prototypes to obtain original query features;

[0062] S3. using a normalized streaming variational autoencoder to convert the support features into a complex class distribution;

[0063] S4. performing deep semantic feature aggregation on the original query features and the supporting features of different categories based on the class distribution, learning feature representations between classes, and obtaining aggregated query features;

[0064] S5. Process the original query features and the aggregated query features using a classification branch that has been processed with feature normalization, cosine similarity measurement, and a learnable scaling factor to obtain a target recognition result.

[0065] Preferably, S2 includes:

[0066] In the subspace of the support features that represents the most semantic information, assume that a support feature , and the support features Perform a global average pooling operation to obtain the global feature query matrix Q:

[0067] ,

[0068] Among them, GAP( ) represents global average pooling, represents the real number space, d represents the feature dimension;

[0069] The support features are transformed through a linear layer Project to the query key space and generate the key matrix K:

[0070] ,

[0071] in, represents the weight matrix of the linear layer, , h represents the height of the feature, w represents the width of the feature, and k represents the expected dimension of the feature after the change;

[0072] Support features for Apply a convolutional layer with a kernel size of 1, and get the value matrix V:

[0073] ,

[0074] Among them, Conv 1×1 ( ) indicates a convolution operation with a kernel size of 1;

[0075] Based on the global feature query matrix Q, the key matrix K and the value matrix V, a support class prototype C is extracted:

[0076] ,

[0077] Among them, softmax() represents the normalized exponential function, and T represents the transpose operation of the matrix;

[0078] Given a query feature , based on the support class prototype C, we get the affinity matrix A:

[0079] ,

[0080] in, express and the shared linear projection of C;

[0081] Based on the affinity matrix A, the query feature of the SAR image to be queried is enhanced to obtain the original query feature :

[0082] ,

[0083] Among them, α represents a learnable parameter.

[0084] Preferably, S3 includes:

[0085] Support for image collections After feature extraction, we get the supporting feature set ,in, are mutually independent characteristic variables;

[0086] The supporting feature set After the encoder maps the value into the latent space, the support class distribution is obtained The mean μ and variance σ of , the mean μ and the variance σ are reparameterized to generate the initial simple latent variable z0:

[0087] ,

[0088] Among them, ε represents the random noise of standard normal distribution;

[0089] The latent variable z0 is transformed into a popular variable z with more actual class distribution information. k :

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] Among them, z' represents the latent variable obtained after manifold transformation, z represents the original latent variable, represents the manifold transformation function, u, w and b represent learnable parameters, represents the Jacobian of the popular transformation, represents the gradient calculation during back propagation, f i represents a reversible epidemic change, i=1,2,...,k;

[0095] Based on the latent variable z0 the latent variable z k , by maximizing the lower bound of evidence Transformed to a complex class distribution.

[0096] Preferably, S4 includes:

[0097] Based on the original query features , let the query features of query class t be ;

[0098] Sample the potential support features supporting class j from the complex class distribution , and the potential support features are aggregated by the dot product method With the query characteristics Aggregate to get common features :

[0099] ,

[0100] in, Indicates channel multiplication, sig indicates sigmoid operation;

[0101] The potential supporting features With the query characteristics Perform differential operation to obtain differential features :

[0102] ;

[0103] The query feature , the common characteristics and the differential characteristics Input into the attention encoding module for fusion to obtain the aggregated query features :

[0104] ,

[0105] Among them, Attention represents the attention operation, d k express feature dimension.

[0106] Preferably, S5 includes:

[0107] The query features are processed using a classification branch processed with feature normalization, cosine similarity metric and learnable scaling factor And the query features after aggregation , and get the target recognition result:

[0108] ,

[0109] ,

[0110] Where N represents the classification score, represents a classification network, represents the feature extractor for classification, scale represents the learnable scaling factor, θ i Represents the angle between class vectors, W i represents the weight vector of each class, B represents the predicted bounding box, represents the regression network, Represents a feature extractor for regression.

[0111] Compared with the prior art, the present invention has the following beneficial effects:

[0112] The present invention solves the problem that the model is susceptible to degradation in new class detection performance due to uneven target size distribution and significant changes in target attitude and orientation in complex SAR scenes. Secondly, the present invention accurately estimates the distribution of new classes of sparse samples through variational learning, avoids the model's sensitivity to the variance of new class examples, and solves the problem that the model is prone to confusion between basic classes and new classes, thereby improving the detection model's recognition accuracy for new classes. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0114] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;

[0115] Figure 2 is an overall architecture diagram of an embodiment of the present invention;

[0116] Figure 3 A schematic diagram of a context information perception enhancement module according to an embodiment of the present invention;

[0117] Figure 4 A schematic diagram of a manifold distribution estimation module according to an embodiment of the present invention;

[0118] Figure 5 Schematic diagram of a class-balanced differential feature aggregation module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0119] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0120] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0121] Embodiment 1

[0122] In this embodiment, if Figure 1 , Figure 2 As shown, the SAR image small sample target detection system based on context perception and Gaussian flow characterization includes: an image acquisition module, a context perception enhancement module, a manifold class distribution estimation module, a class balanced differential aggregation module and a cosine decoupling module.

[0123] The image acquisition module is used to obtain the SAR image to be queried.

[0124] The context-aware enhancement module extracts the most semantically representative support features into support class prototypes, and enhances the query features of the SAR image to be queried based on the support class prototypes to obtain the original query features.

[0125] In this embodiment, due to ultra-long-distance observation, the target size distribution in the SAR image is uneven, and the same target exhibits different characteristics under different imaging conditions. In addition, the scarcity of new class samples makes it difficult for the detector to obtain sufficient positive anchor boxes, further exacerbating the difficulty of detecting and identifying foreground targets in SAR images under complex background interference. Previous studies such as ICPE and IMFA explored the information coupling between query and support images, but these studies only enhanced the query perception information of support features and ignored the problem of low-quality positive anchor boxes of foreground targets in the query image due to complex background interference. To this end, this embodiment designs a context information perception module, such as Figure 3 As shown, the invention aims to use the low-dimensional features of the support image to guide the model to generate high-quality query image bounding boxes, thereby solving the foreground-background imbalance problem in SAR images under complex electromagnetic interference. Inspired by FPD, this embodiment extracts the most representative support features into support class prototypes, and uses them as meta-knowledge to dynamically adjust the feature representation of the query image, so that the support features can guide and enhance the query features. When the enhanced query features are sent to the RPN, the generation quality of the bounding boxes can be improved, more bounding boxes related to the query image can be generated, and most background boxes and suggestion boxes of mismatched categories can be filtered out, thereby reducing the complex noise interference in the SAR image.

[0126] Specifically, the workflow of the context-aware enhancement module includes: in the subspace of the most semantically representative supporting features, assume a supporting feature , and for supporting features Perform a global average pooling operation to obtain the global feature query matrix Q:

[0127] ,

[0128] Among them, GAP( ) represents global average pooling, represents the real number space, d represents the feature dimension, and the dimension of the sub-support feature space is reduced by calculating the average value of each channel, which effectively extracts its semantic information and reduces the introduction of complex noise in the image; the support feature is transformed into Project to the query key space and generate the key matrix K:

[0129] ,

[0130] in, represents the weight matrix of the linear layer, , h represents the height of the feature, w represents the width of the feature, and k represents the expected dimension of the feature after the change; for supporting features Apply a convolutional layer with a kernel size of 1, and get the value matrix V:

[0131] ,

[0132] Among them, Conv 1×1 ( ) represents a convolution operation with a convolution kernel size of 1; the above process effectively captures and enhances the spatial context information and inter-channel feature dependencies in the support feature map. These condensed class features are then used to provide unique prototype semantic information for query image feature enhancement; next, we use the refined class feature representation C as meta-knowledge to dynamically adjust the feature representation of the query image, and refine it into the support class prototype C based on the global feature query matrix Q, key matrix K and value matrix V:

[0133] ,

[0134] Among them, softmax( ) represents the normalized exponential function, T represents the matrix transpose operation; given the query feature , based on the supporting class prototype C, we get the affinity matrix A:

[0135] ,

[0136] in, express and C are shared linear projections, which project them into the same latent space; based on the affinity matrix A, the query features of the SAR image to be queried are enhanced to obtain the original query features :

[0137] ,

[0138] Among them, α represents a learnable parameter, which is initialized to zero to help stabilize the training. Therefore, based on the affinity matrix A, the support class prototype can be used as meta-knowledge to dynamically adjust the feature representation of the query image, thereby enhancing the semantic information of the target in the query image.

[0139] The manifold class distribution estimation module utilizes a normalized manifold variational autoencoder to transform support features into complex class distributions.

[0140] Previous studies only obtained the support class vector through a simple average pooling operation and fused it with the query image features to detect targets in the query image. However, the support class vector obtained by this method is easily affected by the scarce and significantly changing support sample instances, and it is difficult to accurately capture the support category feature information. To solve this problem, VFA converts the support features into category distributions. Furthermore, IMFA introduces MMD on the basis of VFA to measure the distance between the prior distribution and the potential distribution, and applies it to SAR images. However, whether it is VFA or IMFA, the prior distribution and potential distribution of the support features are simply modeled as standard normal distributions. This cannot accurately reflect the actual complex high-dimensional distribution of the support data features, and MMD hinders the maximum coupling and exchange of information between the support and query features. Therefore, this embodiment designs a manifold class distribution estimation module to accurately model the complex high-dimensional distribution of the support features, thereby maximizing the accurate utilization of the support feature information. The detailed structure is as follows Figure 4 shown.

[0141] The workflow of the manifold class distribution estimation module includes: After feature extraction, we get the supporting feature set In this embodiment, a picture set is supported to assist the SAR image to be queried, wherein: are mutually independent feature variables; will support feature sets After the encoder maps the value into the latent space, the support class distribution is obtained The mean μ and variance σ of , the mean μ and variance σ are reparameterized to generate the initial simple latent variable z0:

[0142] ,

[0143] Among them, ε represents the random noise of standard normal distribution; the latent variable z0 is transformed into , to obtain a latent variable z with more realistic class distribution information k , whose popular transformation is defined as:

[0144] ,

[0145] Among them, z' represents the latent variable obtained after manifold transformation, z represents the original latent variable, represents the manifold transformation function, u, w and b represent learnable parameters; the Jacobian determinant of the popular transformation of the manifold transformation is:

[0146] ,

[0147] ,

[0148] in, Represents the gradient calculation during back propagation. We combine several reversible popular transformations to achieve a simple initial distribution Transformed to more complex releases that are closer to actual supported features :

[0149] ,

[0150] f i Represents a reversible popular change, i=1,2,...,k; This embodiment learns the potential probability density function through these popular transformations, so the sample can be remapped from the latent space back to the original supporting feature space through the reverse transformation. This can increase the flexibility and expressiveness of the model, so that the model can better learn the potential structural information of the supporting features. Moreover, these transformation functions are usually differentiable and can be trained using the back propagation algorithm. By learning the parameters of these transformation functions, the model can gradually transform the distribution of the hidden layer to be closer to the real data distribution. Finally, based on the latent variable z0 latent variable z k , by maximizing the evidence lower bound will support the feature set Converted into a complex class distribution; specifically, following the variational inference principle, the likelihood boundary is obtained, called the maximum evidence lower bound (ELBO), referred to as L b . So we maximize L b It is equivalent to the maximum likelihood function, which is:

[0151] ,

[0152] in, Indicates that the distance between two distributions is measured using the divergence KL. represents the true posterior distribution, represents the distribution of the latent variable, represents the distance metric of the reconstructed distribution, represents the reconstructed distribution, represents the distance measure of the true distribution, represents the reconstructed prior distribution, represents the distance measure of the original latent variable distribution, represents the posterior distribution of the final latent variable, represents the prior distribution of the final latent variable, represents the posterior distribution of the original latent variable, represents the kth learnable parameter, represents the gradient calculation during the back propagation process, K represents the number of final manifold transformations, k represents the number of specific manifold transformations, Z represents the estimated potential distribution, and x represents the original data. Further, the above formula can be simplified to the reconstruction loss and KL divergence , where the reconstruction loss Defined as input and reconstruction goals The L2 distance between them, similarly, the KL divergence Decomposed into the KL divergence of the Gaussian distribution and the Jacobian of the popular transformation.

[0153] In addition, this embodiment also introduces the L2 regularization constraint , which can prevent overfitting. Because the number of samples of the new class and the basic class is unbalanced, the distribution estimated by the manifold class distribution estimation module is biased towards the basic class with a larger number of samples. By controlling the size of the parameter, L2 regularization can help the model better capture the common features of the data, rather than focusing too much on the noise or outliers in the training data. This embodiment optimizes the above-mentioned loss objectives to support the features. Transformed to a complex class distribution.

[0154] The class-balanced differential aggregation module performs deep semantic feature aggregation on the original query features and supporting features of different categories based on class distribution, learns the feature representation between classes, and obtains the aggregated query features.

[0155] Previous studies (such as Meta R-CNN, FSRW, and FPD) achieved feature aggregation by re-weighting support features with query features of specific categories. For SAR images with complex interference, FSODS uses a transformer encoder to map support features and query features to the same feature space to achieve feature aggregation of specific categories. The above research works all re-weight features between queries and supports of specific (same) categories. This aggregation method severely limits the relationship modeling between base classes and new classes. VFA is the first to alleviate the information exchange barriers between new classes and base classes by interactively aggregating features between different categories. However, VFA simply uses a dot product (multiplication) method to directly aggregate features, which is obviously not suitable for sparse SAR targets with significantly variable features. Therefore, this embodiment designs a class-balanced differential aggregation module, such as Figure 5 The class-balanced differential aggregation module encourages the model to learn feature representations between classes by allowing query and support features of different categories to be aggregated for deep semantic features, thereby reducing the model's bias towards the base class and reducing its sensitivity to the variance of new class SAR target samples.

[0156] Specifically, first, assume that the query feature of a query class t is , the supporting features of a certain supporting class j . Supported features After being processed by the manifold class distribution estimation module, it is transformed into a class distribution, from which more robust potential features are sampled. With query features Why choose the latent support features sampled from the distribution , rather than the supporting features reconstructed by MCDE With original support features This is due to the latent variable representing the data distribution Contains more representative common features of the class, which are robust to the variance of the supporting instances. This can better balance and reduce the impact of variance on aggregation, thereby improving the robustness and generalization ability of the model. Secondly, we adopt a class-agnostic aggregation scheme, that is, the query class t and the support class j are not one-to-one corresponding, allowing feature aggregation between different categories of query and support features. Most previous works (such as VFA, etc.) use dot product operations to perform feature aggregation. The dot product aggregation method makes it more difficult for the detector to distinguish the aggregated features. Because the potential features sampled from the distribution It already contains more common feature information of this category, which will introduce redundant common information when aggregated with the query features and cannot reflect the unique semantic characteristics of the query class t.

[0157] In this embodiment, the workflow of the class-balanced differential aggregation module includes: based on the original query features , let the query features of query class t be ; Sample potential support features supporting class j from complex class distribution , and the potential support features are aggregated by the dot product method With query features Aggregate to get common features :

[0158] ,

[0159] in, Indicates channel multiplication, sig indicates sigmoid operation; in order to capture the differences between features of different categories, this embodiment uses the potential support feature With query features The differential operation can reflect the unique characteristics of the query class t and increase the distance between classes. In this way, the detector can more clearly distinguish the feature differences between different categories. This method is particularly effective when processing complex SAR images because it can better adapt to targets with significantly variable features, improve the accuracy of detection and recognition, and obtain differential features. The formula is as follows:

[0160] .

[0161] The query feature , common characteristics and differential features Input into the attention encoding module for fusion to obtain the aggregated query features , the encoding module makes the query feature It not only uses the self-attention mechanism to deeply explore its own feature representation potential, but also combines common features with more common information. and differential features of differential feature information , which significantly enhances the ability to represent global common information and local difference information:

[0162] ,

[0163] Among them, Attention represents the attention operation, d k express The feature dimension of the query feature is It can deeply tap its own feature representation potential, enabling it to extract richer and more detailed feature information in a complex data environment. At the same time, the encoding module introduces common features and differential characteristics , which not only provides global information about the commonalities between different data samples, but also contains local information about the differences between different samples. The combination of global and local information enables the model to understand and characterize the data more comprehensively, improving the quality and accuracy of feature representation. In addition, the feature interaction mechanism between different categories in the encoding module also plays a key role. By allowing features of different categories to interact, the model can reduce the bias towards the base class, which is particularly important for processing the relationship between the base class and the new class. This feature interaction mechanism enables the model to better focus on the class correlation between the base class and the new class, reducing the model's sensitivity to the variance of new class examples, thereby effectively enhancing the model's generalization ability and making it more robust and accurate when processing new class samples.

[0164] The cosine decoupling module uses a classification branch that has been processed with feature normalization, cosine similarity metric and learnable scaling factor to process the original query features and aggregated query features to obtain the target recognition result.

[0165] At present, most detectors use aggregated query features to not only identify the target category in the image, but also regress the specific location of the target in the image, which leads to the problem of coupling between classification and regression features. This coupling is not conducive to the simultaneous fine-grained prediction of target category and location, especially in SAR images. When the target size is small or the targets are clustered together, the prediction box of the model regression will be inaccurate, thus affecting the detection performance. Therefore, we achieve accurate recognition and positioning of the target by decoupling the features used for classification and regression tasks in the detection head. VFA combines query features And query features after aggregation The two are sent to the classification branch and regression branch of the detection head respectively to achieve the decoupling of classification and regression tasks. However, this processing is easily affected by the imbalance of the number of samples in the category and may cause incorrect recognition of new categories. Inspired by TFA, this embodiment aggregates the query features It is fed into a classification branch that has been processed with feature normalization, cosine similarity metric, and learnable scaling factor, which significantly reduces the interference between categories and improves the generalization ability and classification accuracy of the model.

[0166] Specifically, the workflow of the cosine decoupling module includes: processing query features using a classification branch that has been processed with feature normalization, cosine similarity metric, and learnable scaling factor And query features after aggregation , and get the target recognition result:

[0167] ,

[0168] ,

[0169] Where N represents the classification score, represents a classification network, represents the feature extractor for classification, scale represents the learnable scaling factor, θ i Represents the angle between class vectors, W i represents the weight vector of each class, B represents the predicted bounding box, represents the regression network, Represents the feature extractor for regression. After being processed by the cosine decoupling module, the aggregated feature F with translation invariance is A It will not harm the detector regression task. Moreover, the classifier based on cosine similarity can reduce the interference caused by class imbalance between the base class and the new class to a certain extent through vector normalization and cosine similarity calculation, thereby reducing the discovery of overfitting and improving the overall classification accuracy.

[0170] Embodiment 2

[0171] In this embodiment, the designed small sample target detection system is compared with the existing model to evaluate its detection performance and effectiveness. To this end, this embodiment selects some classic methods based on meta-learning, such as Meta-RCNN, VFA and PFD. In addition, some classic methods based on fine-tuning paradigm are also selected, such as TFA, FSCE and CIR-FSD. The experimental results on the fine-grained aircraft detection dataset SAR-AIRcraft-1.0 and the cross-domain dataset MSAR-AIR are shown in Tables 1 and 2. The results show that the detection and recognition performance of SAR images of this application is significantly better than the existing methods.

[0172] (1) Experimental results on the SAR-AIRcraft-1.0 dataset:

[0173] Table 1 shows the experimental results, which compares the performance of existing methods on the SAR-AIRcraft-1.0 dataset. The present invention significantly outperforms existing detectors in detecting and identifying new categories. Specifically, when ARJ21 is a new category, the present invention improves by 4.38%, 6.55%, 9.07% and 9.28% over the existing best new category detection and recognition methods in 3, 5, 10 and 50-shot sample settings, respectively. In addition, compared with the baseline, the present invention improves by 6.72%, 11.36%, 13.43%, 14.66% and 20.04% in 1, 3, 5, 10 and 50-shot sample settings, respectively.

[0174] Table 1

[0175] .

[0176] When Boeing787 is used as a new category, SCGF improves by 1.19%, 1.73%, 1.27%, and 1.65% in 3, 5, 10, and 50-shot sample settings, respectively, compared to the existing best methods. In addition, compared to the baseline, SCGF improves by 6%, 4.03%, 3.99%, 4.47%, and 7.05% in 1, 3, 5, 10, and 50-shot sample settings, respectively. From this analysis, it is clear that the present invention has the most advanced detection and recognition performance.

[0177] As new classes are added, the detection accuracy of the present invention continues to improve and is significantly better than existing methods. This is because most previous works use the average paradigm to estimate the class prototype of new classes, which causes the model to misjudge the class center of the new class. In other words, most previous works estimate the class center of unstable new classes, which severely limits the performance of the model in detecting and identifying new classes. However, our SCGF is achieved by estimating the class distribution of supporting features and then sampling more generalized features from the estimated class distribution as class prototypes, which is robust to the variance of some novel class examples. Therefore, SCGF has significant detection and recognition capabilities on sparse new class examples.

[0178] In addition to the above quantitative analysis, this embodiment also visualizes the detection and recognition results when the number of sample instances is set to 10. From the visualization results, it can be concluded that the present invention not only has high detection and recognition performance for new categories, but also maintains the detection performance for basic categories, which is attributed to the freezing of the parameters of the backbone and shared head in the fine-tuning stage.

[0179] (2) Experimental results on MSAR-SAR:

[0180] In order to fully demonstrate the advantages of the present invention, this embodiment is experimented on the cross-domain dataset MSAR-AIR. As shown in Table 2, the experimental results show that the present invention is significantly better than existing detectors in detecting and identifying new categories. Specifically, when airplane is a new category, compared with the existing best methods, the detection and recognition accuracy of the present invention in 1, 3, 5, 10 and 50-shot sample settings is improved by 2.78%, 1.44%, 1.71%, 2.25% and 3.7%, respectively. In addition, compared with the baseline, our method improves the detection and recognition accuracy of new classes by 7.43%, 5.01%, 5.29%, 7.09% and 9.24% under the same small sample task setting.

[0181] Table 2

[0182] .

[0183] When ship is a new category, compared with the best existing methods, the detection and recognition accuracy of the present invention is improved by 2.69%, 4.41%, 4.62%, 2.26% and 3.09% under 1, 3, 5, 10 and 50-shot sample settings. In addition, compared with the baseline, SCGF improves the detection and recognition accuracy of new classes by 9.25%, 12.17%, 14.13%, 8.15% and 10.4% under the same few-shot task setting. The above analysis shows that the present invention has good generalization and can meet the cross-domain detection capabilities in practical scenarios. Specifically, for the cross-domain dataset MSAR-AIR, the present invention shows significant detection and recognition performance. The present invention achieves better results on data from different fields, reduces the cost of data re-annotation, and improves data utilization and efficiency. This is because in actual SAR applications, it is usually necessary to process data from multiple sensors. Models with good cross-domain detection capabilities can be more widely used in scenarios in different fields, thereby enhancing the practicality and applicability of the model. Obviously, the present invention has state-of-the-art detection performance and good cross-domain detection capabilities. This capability enables the model to maintain efficient detection performance in domains or data distributions where few samples are needed for training.

[0184] Through the above quantitative analysis, in addition to the actual application requirements, this embodiment visualizes the detection and recognition results obtained by training the detector, and the number of labeled samples is 10. It can be seen from the visualization results that the present invention not only has high detection and recognition performance for new categories, but also maintains the detection and recognition performance for basic categories, which is attributed to the freezing of the parameters of the backbone network and the shared header in the fine-tuning stage.

[0185] Embodiment 3

[0186] To verify the effectiveness of the designed SCGF detector and its constituent modules, we conducted comprehensive ablation experiments on two SAR target detection datasets:

[0187] (1) Strong Baseline Meta R-CNN+. In order to verify the method of determining a strong baseline, this embodiment freezes the relevant components of the detector during fine-tuning. The experimental results are shown in Table 3, where B and R represent the backbone and S represents the shared head. Compared with the original baseline, excellent results can be obtained by freezing the parameters of our modules in the second stage. Specifically, the parameters of the backbone network and the shared head are frozen during fine-tuning, while the other components of the detector are unfrozen during fine-tuning. For example, by fine-tuning the RPN in the fine-tuning stage, the imbalance between the foreground and the background can be alleviated, thereby providing more proposal boxes for new examples.

[0188] Table 3

[0189] .

[0190] (2) Analysis of the effectiveness of the context-aware enhancement module: In order to explore the effectiveness of the designed context-aware enhancement module of the present invention, this embodiment integrates the context-aware enhancement module into the original baseline during the training phase. The experimental results on the aircraft fine-grained dataset SAR-AIRcraft-1.0 are shown in Table 4. For data split 1 (Split1), compared with the strong baseline, the addition of the context-aware enhancement module significantly improves the detection and recognition accuracy of the detector by 4.25%, 3.13%, 3.94% and 5.54% under the 3, 5, 10 and 50-shot sample settings, respectively. For data split 2 (Split2), the context-aware enhancement module improves the detection accuracy of Boeing 787 by 1.16% and 1.65% under the 5 and 50-shot sample settings. In addition, the results of the present invention on the cross-domain dataset MSAR-AIR are shown in Table 5. The detection accuracy of the context-aware enhancement module under the 3, 5, 10 and 50-shot sample settings is improved by 0.79%, 0.29%, 0.64% and 2.28%, respectively. For data split 2 (Split2), under the same sample settings, the detection accuracy of the Ship category increased by 2.28%, 2.16%, 0.80% and 2.77% respectively, indicating that the model performance has been significantly improved. These improvements are due to the strong feature enhancement capability of the context-aware enhancement module for the query branch, which enables the model to pay more attention to the target information, reduce the influence of complex background and noise, and provide a solid foundation for subsequent model operations.

[0191] (3) Effectiveness analysis of the manifold class distribution estimation module: Sparse SAR sample data often leads to inaccurate prototype estimation of new categories by traditional target detectors. In contrast, the manifold class distribution estimation module can achieve accurate class prototypes by modeling the distribution of sparse data. The experimental results on the fine-grained aircraft dataset SAR-AIRcraft-1.0 are shown in Table 4. For data split 1 (Split1), the detection accuracy of the aircraft model ARJ21 is improved by 2.81%, 2.91%, 3.14% and 2.63% under the 3-, 5-, 10- and 50-shot sample settings, respectively. For data split 2 (Split2), under the same sample number setting, the detection accuracy of the aircraft model Boeing787 is improved by 0.59%, 0.53%, 1.25% and 1.22%, respectively. In addition, the results on the cross-domain dataset MSAR-AIR are shown in Table 5. The results show that for data split 1 (Split1), under the 3-, 5-, 10- and 50-shot sample settings, the manifold class distribution estimation module improves the detection accuracy of the aircraft category by 0.59%, 1.46%, 0.78% and 2.08%, respectively. For data split 2 (Split2), under the same sample number setting, the detection accuracy of the ship category is improved by 3.03%, 0.78%, 0.91% and 2.23%, respectively. In summary, the integration of the manifold class distribution estimation module significantly improves the detection and recognition performance of new categories. This improvement is attributed to the fact that the manifold class distribution estimation module uses multiple variational autoencoders to transform support features into the distribution of support classes, making the estimated class distribution robust to the absence or change of a single support sample and allowing the center of the support class to be accurately estimated. The latent features sampled from the estimated class distribution are robust to the support samples and are insensitive to their variance. In summary, the manifold class distribution estimation module ensures accurate estimation of class prototypes, thereby providing robust latent features for subsequent aggregation with query features.

[0192] (4) Effectiveness analysis of the class-balanced differential aggregation module. The aggregation of features from the support and query branches is crucial for detecting new classes. The class-balanced differential aggregation module improves detection capabilities by aggregating deep semantic features between support and query features across different classes. This approach not only promotes the learning of feature representations between different classes, but also alleviates the model's bias towards the base class and reduces sensitivity to the variance of new class samples. The experimental results on the fine-grained aircraft dataset SAR-AIRcraft1.0 are shown in Table 4. For data split 1 (Split1), CBDA improves the detection accuracy of the aircraft model ARJ21 by 2.52%, 3.28%, 1.62% and 4.03% under the 3-, 5-, 10- and 50-shot sample settings, respectively. For data split 2 (Split2), the detection accuracy of the aircraft model Boeing787 is improved by 0.76%, 0.42%, 0.45% and 1.61%, respectively. In addition, the experimental results on the cross-domain dataset MSAR-AIR are shown in Table 5. The experiment shows that for data split 1 (Split1), under the same sample number setting, the class-balanced differential aggregation module improves the detection accuracy of the aircraft category by 0.73%, 0.3%, 0.43% and 1.75% respectively. For data split 2 (Split2), the class-balanced differential aggregation module improves the detection accuracy of the ship category by 2.71%, 2.96%, 2.27% and 1.29% respectively. The integration of the class-balanced differential aggregation module greatly improves the performance of the model in detecting and identifying new categories. This improvement is attributed to the feature interaction mechanism of the class-balanced differential aggregation module, which enables the model to better capture the correlation between the basic class and the new class, while reducing the sensitivity to the variance of the new class samples. Therefore, the class-balanced differential aggregation module effectively enhances the generalization ability of the model and has higher robustness and accuracy when processing new category samples.

[0193] (5) Analysis of the effectiveness of the cosine decoupling module. The coupling of classification and regression features hinders the detector's ability to make fine-grained predictions of the category and position of SAR targets. The cosine decoupling module of the present invention alleviates this problem by decoupling the features used for classification and regression tasks in the detection head, thereby achieving accurate target recognition and positioning. The experimental results on the SAR-AIRcraft1.0 dataset are shown in Table 4. Under the 3-, 5-, 10-, and 50-shot sample settings, the cosine decoupling module improves the detection accuracy of the aircraft model ARJ21 by 0.33%, 1.6%, 2.79%, and 4.49%, respectively. For data split 2 (Split2), CDM improves the detection accuracy of Boeing 787 aircraft by 0.36%, 0.14%, 0.76%, and 0.93%, respectively. In addition, the experimental results on the cross-domain MSAR-AIR dataset are shown in Table 5, which shows that under the 3-, 5-, 10-, and 50-shot sample settings, for data split 1 (Split1), the cosine decoupling module improves the detection accuracy of the aircraft category by 0.55%, 0.17%, 1.34%, and 1.64%, respectively. For data split 2 (Split2), the cosine decoupling module improves the detection accuracy of the ship category by 2.58%, 2.91%, 1.45%, and 2.56%, respectively. Integrating the cosine decoupling module into the detector can significantly improve the performance of detecting and recognizing new classes. This is because the cosine decoupling module method only applies aggregated query features in the classification branch based on cosine similarity, thereby maintaining the accuracy of SAR target positioning and enhancing its applicability in SAR applications.

[0194] Table 4

[0195] ,

[0196] Table 5

[0197] .

[0198] (6) Analysis of the effectiveness of hyperparameters: This embodiment evaluates the hyperparameters in the present invention to evaluate their impact on the proposed method. First, this embodiment studies the impact of the number of reversible manifold transformations, and the results are shown in Table 6. When the number of reversible manifold transformations K = 4, the modeled latent variables can more accurately and robustly represent the class distribution, thereby having the highest accuracy when detecting new classes. Next, we explored the impact of regularization on SCGF based on the optimal parameter α = 2.5 × 10−4 determined in VFA. The experimental results are shown in Table 7. When the weight coefficient β of the regularization term = 0.001, the model can better capture the common features of high-dimensional data and avoid over-emphasizing noise or outliers in the training data. This can prevent the latent distribution from being biased towards the base class with a larger sample size, thereby reducing the risk of overfitting.

[0199] Table 6

[0200] ,

[0201] Table 7

[0202] .

[0203] In summary, the present invention not only effectively improves the detection performance of the model for new classes, but also effectively enhances the detection performance of the model for basic classes. The present invention solves the problem of improving the detection performance of new classes while avoiding the catastrophic forgetting of the model for basic classes.

[0204] Embodiment 4

[0205] In this embodiment, a SAR image small sample target detection method based on context perception and Gaussian flow characterization is provided, comprising the following steps:

[0206] S1. Obtain the SAR image to be queried.

[0207] S2. Extract the most semantically representative support features into support class prototypes, and enhance the query features of the SAR image to be queried based on the support class prototypes to obtain the original query features.

[0208] S3. Using normalized streaming variational autoencoders to transform support features into complex class distributions.

[0209] S4. Based on the class distribution, the original query features and supporting features of different categories are deeply semantically aggregated, the feature representation between classes is learned, and the aggregated query features are obtained.

[0210] S5. Use a classification branch that has been processed with feature normalization, cosine similarity measurement and learnable scaling factor to process the original query features and the aggregated query features to obtain the target recognition result.

[0211] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A SAR image small sample target detection system based on context perception and Gaussian flow characterization, characterized by: include: Image acquisition module, context-aware enhancement module, manifold class distribution estimation module, class-balanced differential aggregation module, and cosine decoupling module; The image acquisition module is used to obtain the SAR image to be queried; The context-aware enhancement module extracts the most semantically representative support features into support class prototypes, and enhances the query features of the SAR image to be queried based on the support class prototypes to obtain original query features; The manifold class distribution estimation module converts the support features into complex class distributions using a normalized flow-based variational autoencoder; The class-balanced differential aggregation module performs deep semantic feature aggregation on the original query features and the supporting features of different categories based on the class distribution, learns the feature representation between classes, and obtains the aggregated query features; The cosine decoupling module processes the original query features and the aggregated query features using a classification branch that has been processed with feature normalization, cosine similarity measurement, and a learnable scaling factor to obtain a target recognition result; The workflow of the context-aware enhancement module includes: In the subspace of the support features that represents the most semantic information, assume that a support feature , and the support features Perform a global average pooling operation to obtain the global feature query matrix Q: ， Among them, GAP( ) represents global average pooling, represents the real number space, d represents the feature dimension; The support features are transformed through a linear layer Project to the query key space and generate the key matrix K: ， in, represents the weight matrix of the linear layer, , h represents the height of the feature, w represents the width of the feature, and k represents the expected dimension of the feature after the change; Support features for Apply a convolutional layer with a kernel size of 1, and get the value matrix V: ， Among them, Conv 1×1 ( ) indicates a convolution operation with a kernel size of 1; Based on the global feature query matrix Q, the key matrix K and the value matrix V, a support class prototype C is extracted: ， Among them, softmax() represents the normalized exponential function, and T represents the transpose operation of the matrix; Given a query feature , based on the support class prototype C, we get the affinity matrix A: ， in, express and the shared linear projection of C; Based on the affinity matrix A, the query feature of the SAR image to be queried is enhanced to obtain the original query feature : ， Among them, α represents a learnable parameter; The workflow of the manifold class distribution estimation module includes: Support for image collections After feature extraction, we get the supporting feature set ,in, are mutually independent characteristic variables; The supporting feature set After the encoder maps the value into the latent space, the support class distribution is obtained The mean μ and variance σ of , the mean μ and the variance σ are reparameterized to generate the initial simple latent variable z0: ， Among them, ε represents the random noise of standard normal distribution; The latent variable z0 is transformed into a popular variable z with more actual class distribution information. k : ， ， ， ， Among them, z' represents the latent variable obtained after manifold transformation, z represents the original latent variable, represents the manifold transformation function, u, w and b represent learnable parameters, represents the Jacobian of the popular transformation, represents the gradient calculation during back propagation, f i represents a reversible epidemic change, i=1,2,...,k; Based on the latent variable z0 the latent variable z k , by maximizing the lower bound of evidence Transformed to a complex class distribution.

2. The SAR image small sample target detection system based on context perception and Gaussian flow characterization according to claim 1, characterized in that: The workflow of the class-balanced differential aggregation module includes: Based on the original query features , let the query features of query class t be ; Sample the potential support features supporting class j from the complex class distribution , and the potential support features are aggregated by the dot product method With the query characteristics Aggregate to get common features : ， in, Indicates channel multiplication, sig indicates sigmoid operation; The potential supporting features With the query characteristics Perform differential operation to obtain differential features : ； The query feature , the common characteristics and the differential characteristics Input into the attention encoding module for fusion to obtain the aggregated query features : ， Among them, Attention represents the attention operation, d k express feature dimension.

3. The SAR image small sample target detection system based on context perception and Gaussian flow characterization according to claim 2 is characterized in that: The workflow of the cosine decoupling module includes: The query features are processed using a classification branch processed with feature normalization, cosine similarity metric and learnable scaling factor And the query features after aggregation , and get the target recognition result: ， ， Where N represents the classification score, represents a classification network, represents the feature extractor for classification, scale represents the learnable scaling factor, θ i Represents the angle between class vectors, W i represents the weight vector of each class, B represents the predicted bounding box, represents the regression network, Represents a feature extractor for regression.

4. A method for detecting small sample targets in SAR images based on context perception and Gaussian flow characterization, the method being applied to the system according to any one of claims 1 to 3, characterized in that: The following steps are involved: S1. Obtain the SAR image to be queried; S2. extracting the most semantically representative support features into support class prototypes, and enhancing the query features of the SAR image to be queried based on the support class prototypes to obtain original query features; S3. using a normalized streaming variational autoencoder to convert the support features into a complex class distribution; S4. performing deep semantic feature aggregation on the original query features and the supporting features of different categories based on the class distribution, learning feature representations between classes, and obtaining aggregated query features; S5. Process the original query features and the aggregated query features using a classification branch that has been processed with feature normalization, cosine similarity measurement, and a learnable scaling factor to obtain a target recognition result.

5. The method for detecting small sample targets in SAR images based on context perception and Gaussian flow characterization according to claim 4, characterized in that: The S2 includes: In the subspace of the support features that represents the most semantic information, assume that a support feature , and the support features Perform a global average pooling operation to obtain the global feature query matrix Q: ， Among them, GAP( ) represents global average pooling, represents the real number space, d represents the feature dimension; The support features are transformed through a linear layer Project to the query key space and generate the key matrix K: ， in, represents the weight matrix of the linear layer, , h represents the height of the feature, w represents the width of the feature, and k represents the expected dimension of the feature after the change; Support features for Apply a convolutional layer with a kernel size of 1, and get the value matrix V: ， Among them, Conv 1×1 ( ) indicates a convolution operation with a kernel size of 1; Based on the global feature query matrix Q, the key matrix K and the value matrix V, a support class prototype C is extracted: ， Among them, softmax() represents the normalized exponential function, and T represents the transpose operation of the matrix; Given a query feature , based on the support class prototype C, we get the affinity matrix A: ， in, express and the shared linear projection of C; Based on the affinity matrix A, the query feature of the SAR image to be queried is enhanced to obtain the original query feature : ， Among them, α represents a learnable parameter.

6. The method for detecting small sample targets in SAR images based on context perception and Gaussian flow characterization according to claim 5, characterized in that: The S3 includes: Support for image collections After feature extraction, we get the supporting feature set ,in, are mutually independent characteristic variables; The supporting feature set After the encoder maps the value into the latent space, the support class distribution is obtained The mean μ and variance σ of , the mean μ and the variance σ are reparameterized to generate the initial simple latent variable z0: ， Among them, ε represents the random noise of standard normal distribution; The latent variable z0 is transformed into a popular variable z with more actual class distribution information. k : ， ， ， ， Among them, z' represents the latent variable obtained after manifold transformation, z represents the original latent variable, represents the manifold transformation function, u, w and b represent learnable parameters, represents the Jacobian of the popular transformation, represents the gradient calculation during back propagation, f i represents a reversible epidemic change, i=1,2,...,k; Based on the latent variable z0 the latent variable z k , by maximizing the lower bound of evidence Transformed to a complex class distribution.

7. The method for detecting small sample targets in SAR images based on context perception and Gaussian flow characterization according to claim 6, characterized in that: The S4 includes: Based on the original query features , let the query features of query class t be ; Sample the potential support features supporting class j from the complex class distribution , and the potential support features are aggregated by the dot product method With the query characteristics Aggregate to get common features : ， in, Indicates channel multiplication, sig indicates sigmoid operation; The potential supporting features With the query characteristics Perform differential operation to obtain differential features : ； The query feature , the common characteristics and the differential characteristics Input into the attention encoding module for fusion to obtain the aggregated query features : ， Among them, Attention represents the attention operation, d k express feature dimension.

8. The method for detecting small sample targets in SAR images based on context perception and Gaussian flow characterization according to claim 7, characterized in that: The S5 includes: The query features are processed using a classification branch processed with feature normalization, cosine similarity metric and learnable scaling factor And the query features after aggregation , and get the target recognition result: ， ， Where N represents the classification score, represents a classification network, represents the feature extractor for classification, scale represents the learnable scaling factor, θ i Represents the angle between class vectors, W i represents the weight vector of each class, B represents the predicted bounding box, represents the regression network, Represents a feature extractor for regression.

Images