Satellite synthetic aperture radar marine ship monitoring method and system based on three-layer sparsity perception, storage medium and product

Abstract

The invention discloses a satellite synthetic aperture radar marine ship monitoring method and system based on three-layer sparsity perception, a storage medium and a product, belongs to the technical field of computer vision and remote sensing image processing, and solves the problem that SAR imaging space sparsity caused by Sinc point diffusion response is not fully considered in the prior art. A to-be-monitored synthetic aperture radar marine scene image is obtained and divided into a plurality of image blocks, and the image blocks are combined with position codes to form original input features; and based on the three-layer sparsity perception model, monitoring the category information of the ship target and the corresponding spatial position information of the ship target on the original input features. The method is used for marine ship monitoring.

Description

Technical Field

[0001] A method, system, storage medium, and product for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing are disclosed. This invention belongs to the field of computer vision and remote sensing image processing technology. Background Technology

[0002] Automated maritime vessel surveillance technology plays a vital role in various fields, including transportation automation, intelligent fisheries production, and national defense modernization. Real-time monitoring and identification of vessels at sea can effectively improve shipping scheduling efficiency, ensure maritime traffic safety, and provide crucial technical support for marine fisheries supervision, illegal fishing monitoring, and maritime emergency rescue.

[0003] Synthetic Aperture Radar (SAR) is an active microwave remote sensing imaging technology with all-weather, day-and-night imaging capabilities, demonstrating good applicability in maritime vessel monitoring. Existing methods for maritime vessel detection based on SAR remote sensing images mainly include traditional signal processing methods and deep learning-based target detection methods. Traditional methods typically rely on statistical models such as constant false alarm rate (CFAR), modeling local sea clutter characteristics and adaptively setting detection thresholds to achieve target extraction. These methods have clear implementation paths and rich engineering experience, but their performance is highly dependent on background statistical assumptions, making them difficult to adapt to complex and changing marine environments. They are prone to increased false alarm rates or severe missed detections in scenarios with strong clutter, non-uniform backgrounds, or weakly scattering targets. Currently, deep learning methods are gradually becoming the mainstream technology for SAR vessel detection. Through an end-to-end network structure, they automatically learn feature representations, improving detection accuracy and robustness to a certain extent. However, existing deep learning models are mostly derived from optical image detection tasks. Their network structure design and feature modeling strategies are often not specifically optimized for SAR imaging characteristics. They typically employ dense feature extraction and dense prediction mechanisms across the entire image, resulting in high computational resource consumption. Furthermore, their adaptability to complex backgrounds and weak targets remains insufficient, limiting their widespread adoption in large-scale, high-real-time engineering applications. Existing methods still face the following technical challenges in maritime ship detection: First, existing methods fail to adequately consider the spatial sparsity of SAR imaging caused by the Sinc point spread response, resulting in low feature modeling efficiency and weak feature representation. In SAR imaging, the energy response of point targets is affected by the system's point spread function, mainly concentrated in a finite region, while most image areas are low-energy background. This imaging characteristic determines that SAR images exhibit significant sparsity in spatial distribution. However, many existing detection methods still employ global dense convolution or global self-attention mechanisms, repeatedly calculating large amounts of background regions. This not only increases computational complexity but also easily introduces background noise interference, weakening key ship features during feature representation.

[0004] Second, existing methods fail to adequately consider the sparsity of ship spatial distribution in large-scale marine scenes, which can easily lead to an increase in false alarms and affect the overall detection stability and reliability. Existing methods often employ a dense scanning strategy in the candidate region generation stage, generating a large number of candidate boxes across the entire image. This dense candidate generation method produces a large number of redundant candidates in scenes with relatively few ships, significantly increasing the computational burden of subsequent classification and regression stages. Simultaneously, the introduction of a large number of low-quality candidates can easily distract the model, leading to an increase in false alarms and affecting the overall detection stability and reliability.

[0005] Third, existing methods fail to fully consider the significant non-uniformity and discreteness of the scattering characteristics of SAR ship targets, leading to issues such as missed detections and insufficient positioning accuracy. Due to the complexity of the ship's structure and the influence of various electromagnetic scattering mechanisms, ships in SAR images typically appear as several discrete strong scattering points rather than a continuous and complete target outline. This scattering characteristic results in a discontinuous spatial distribution of ship pixels, posing a significant challenge to feature extraction and target localization. Existing methods often employ regular grids or continuous region modeling, which struggles to effectively characterize the spatial topological relationships between scattering points, thus easily leading to missed detections and insufficient positioning accuracy under conditions of small ships, weakly scattering ships, or complex backgrounds.

[0006] Fourth, existing methods suffer from low computational efficiency and insufficient monitoring accuracy. In Transformer-based detection frameworks, the decoding stage typically introduces a large number of markers for feature encoding and decoding. However, a significant portion of these markers correspond to background regions or low-value features. This intensive marker processing not only increases the computational burden on the model's inference stage but may also lead to attention resources being occupied by irrelevant information, reducing the model's focus on key ship features and consequently affecting detection accuracy and real-time performance. Summary of the Invention

[0007] The purpose of this invention is to provide a satellite synthetic aperture radar (SAR) method, system, storage medium, and product for monitoring ships at sea based on three-layer sparsity sensing. This addresses the problems of existing technologies failing to adequately consider the spatial sparsity of SAR imaging caused by the Sinc point diffusion response, resulting in low feature modeling efficiency and weak feature representation; failing to adequately consider the sparsity of the spatial distribution of ships in large-scale marine scenes, leading to an increase in false alarms and affecting overall detection stability and reliability; fully considering the significant non-uniformity and discreteness of the scattering characteristics of SAR ship targets, resulting in missed detections and insufficient positioning accuracy; and low computational efficiency and insufficient monitoring accuracy.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A satellite synthetic aperture radar method for monitoring ships at sea based on three-layer sparsity sensing includes the following steps: Step 1: Obtain the synthetic aperture radar image of the marine scene to be monitored, divide it into several image blocks, and combine it with the position coding to form the original input features; Step 2: Based on the three-layer sparse perception model, monitor the category information of ship targets and their corresponding spatial location information of the original input features; The three-layer sparsity perception model includes a feature layer sparsity perception module that extracts sparse features from the original input features and models sparse features to obtain a sparse feature map; a candidate layer sparsity perception module that generates candidate boxes based on the sparse feature map; and a label layer sparsity perception module that performs sparse labeling on the features corresponding to the candidate boxes, decodes the sparse labels, and outputs the category information of the ship target and its corresponding spatial location information.

[0009] Furthermore, the combination of image blocks and positional encoding in step 1 is obtained through sine and cosine positional embedding, which is defined as: in, This indicates the position number of the input image patch in the spatial sequence. Indicates the frequency index parameter. , Indicates the embedding dimension of the position encoding. Indicates location number In the even-numbered dimensions of the embedded image patch Position code on Indicates location number In the odd-numbered dimensions of the embedded image patch Position encoding on.

[0010] Furthermore, the feature layer sparsity perception module in step 2 includes four feature sparsity perception layers connected in sequence. Each feature sparsity perception layer includes a sparse distance-azimuth multi-head self-attention layer that extracts sparse features with directional decoupling by performing a distance-azimuth decoupling self-attention mechanism on the input features, a sparse scattering point hypergraph topology perception layer that models the high-order topological relationship between ship scattering points based on sparse features and a sparse scattering point hypergraph topology perception mechanism to obtain a sparse feature map, and a downsampling layer. The input features of the sparse distance-azimuth multi-head self-attention layer in the first feature sparsity perception layer are the original input features, the input features of the second feature sparsity perception layer are the output of the first feature sparsity perception layer, the input features of the third feature sparsity perception layer are the output of the second feature sparsity perception layer, and the input features of the fourth feature sparsity perception layer are the output of the third feature sparsity perception layer.

[0011] Furthermore, the specific implementation steps of the sparse distance-azimuth multi-head self-attention layer are as follows: The input features are used to form feature vector sequences in the range and azimuth directions, respectively. A gated recurrent unit (GRU) or LSTM is then used to describe the pixel dependencies in the same direction. The formula for the GRU is as follows: in, and These represent the feature vectors in the range and azimuth directions, respectively. Subsequently, the query vector and key vector for the corresponding directions are obtained through linear embedding, as shown in the formula: in, , These represent the query vectors in the distance and azimuth directions, respectively. , These represent the key vectors for the range and azimuth directions, respectively. , , and They represent , , and The corresponding embedding weights; After linear embedding, the query vectors and key vectors in the range and azimuth directions are multiplied separately to obtain the attention maps in the range and azimuth directions, respectively. These are then processed using the Softmax function, and finally, the Manhattan distance decay function is introduced. The final attention map is obtained by weighted suppression of spatial correlation, and the formula is: in, , These represent the final attention maps in the distance and azimuth directions, respectively. and These represent the Manhattan distance decay functions introduced in the range and azimuth directions, respectively. and They represent and transpose, This represents element-wise multiplication. and The Middle The Manhattan distance decay function for each location index is: and The formula is: in, This indicates the center of Manhattan, which is the reference center location used for calculating distances to Manhattan. The first part represents the attention map in the distance direction and the attention map in the azimuth direction. Location index, The attenuation coefficient; The final decoupled sparse attention feature is obtained by multiplying the distance attention pair of the input feature with the final attention map of the azimuth direction, and then multiplying the result with the final attention map of the distance direction. The formula is as follows: in, , representing the distance to the attention pair value, Represents the projection weight matrix. Indicates input features, This indicates transpose.

[0012] Furthermore, the sparse scattering point hypergraph topology sensing layer includes an adaptive hyperedge generation module and a hypergraph convolution operation module connected in sequence. The specific implementation steps of the adaptive hyperedge generation module are as follows: Sparsity features As input, it is flattened to obtain the set of vertex features. Let the hypergraph be represented as... ,in, This represents the set of vertex features, corresponding to the potential scattering point features in synthetic aperture radar (SAR) images of maritime scenes. This represents a set of hyperedges, where each hyperedge connects multiple vertex features to characterize their higher-order relationships. A continuously differentiable vertex participation mechanism is introduced, using a continuous participation matrix. The hypergraph is metric, where, and These represent the number of vertex features and the number of hyperedges, respectively. Representing vertex features Participating in the super-border To what extent, This represents an identifier and has no specific meaning. The specific steps for measuring a hypergraph are as follows: Let the vertex features be Vertex features Perform global average pooling and global max pooling, and then concatenate the two to obtain the global vertex. Its expression is: in, This represents averaging the features of all vertices over the spatial dimension. This represents taking the maximum value of all vertex features in the spatial dimension. For channel dimension, Indicates length is A real-valued vector; Subsequently, through the projection layer Generate global vertex representation global offset and in the global prototype Building on A dynamic hyperedge prototype Specifically, it is expressed as: in, That is, mapping the global vertex representation to A hyperedge prototype; Simultaneously, the correlation between vertex features and hyperedges is calculated, i.e., another projection layer is introduced. The formula for mapping vertex features to query vectors is as follows: in, Represents the query vector. , Indicates length is The real-valued vector, This represents the query projection weight matrix, used to weight the input vertex features. Mapped to query vector ; Subsequently, the query vector conduct Subspace decomposition yields , ,in, Indicates the first query vectors The decomposition yielded the first Subspace; For dynamic hyperedge prototypes Perform the same subspace decomposition to obtain , Indicates the first A dynamic hyperedge prototype The decomposition yielded the first Subspace; In the In each subspace, the query vector of vertex features With the hyper-edge prototype The similarity is calculated using the scaled dot product, with the following formula: The overall similarity is obtained by averaging the subspace results based on the similarity score, using the following formula: Next, regarding Normalization is performed to obtain the participation of vertex features in the hyperedge. The formula is: The specific implementation steps of the hypergraph convolution operation module are as follows: Based on participation, the features of the vertices connected to the hyperedge are aggregated using hyperedges to achieve feature mapping from vertex features to hyperedges. Then, the mapped hyperedge-level features are backpropagated back to the vertex features to update them. The aggregation process uses Choquet integrals, and this bidirectional aggregation process can be represented as: in, and Represented as a learnable weight matrix, This represents the Sigmoid activation function. Indicates the first Vertex eigenvectors.

[0013] Furthermore, the specific implementation steps of the candidate layer sparsity sensing module in step 2 are as follows: Step 3.1: Sparse feature map output by the feature layer sparsity perception module. Inputting the sparse candidate generation network yields the basic features of the candidate regions. The specific steps are as follows: Step 3.11: Sparse feature map output by the feature layer sparsity perception module. ,pass Convolution is used for position encoding to obtain position response features. The formula is: in, and They represent Convolution weights and biases; Position response features are obtained by using the Sigmoid activation function. Mapped to the probability value of each spatial location belonging to the candidate region. And based on a preset threshold Determine candidate activation regions The formula is: Step 3.12: Based on the set of candidate activation locations and position response characteristics Generate a set of candidate activation locations The set of candidate boxes within the specified sparse region ; Step 3.13: Utilize position response characteristics Guided deformable convolution on sparse feature maps Perform adaptive feature adjustment to obtain the aligned feature map. Aligned feature maps The first in Feature maps of candidate activation locations The calculation formula is: in, Indicates the number of convolution sampling points. For the first Fixed sampling positions for regular convolution with a number of convolution sampling points Indicates the first The scientific system offset of each sampling point Indicates the first The modulation coefficients corresponding to each sampling point Indicates the first The weights corresponding to each sampling point; Step 3.14: Set the candidate boxes Projected onto the aligned feature map Based on the corresponding coordinate system, high-dimensional feature representations of each candidate box are extracted using a dynamic header, and the RoIAlign operation is performed to obtain the basic features of the corresponding candidate regions. The formula is: in, The RoIAlign function represents the feature extraction function for candidate regions. Step 3.2: Set the candidate boxes According to the scaling factor respectively and Scale expansion is performed to obtain multi-scale context candidate boxes. and and in the feature map Perform RoIAlign operations on each of the above to obtain multi-scale contextual features. and The formula is: Step 3.3: Identify the basic features of the candidate regions. and multi-scale contextual features , The input context-coordinated shuffle attention module coordinates and fuses contextual information at different scales, and outputs the fused candidate contextual features. The specific steps are as follows: Basic features of candidate regions and multi-scale contextual features , Apply Convolution is used to refine local features, and these features are then concatenated along the channel dimension to obtain a comprehensive contextual feature map. The formula is: in, , and Indicates correspondence The weight parameters of the convolutional layer, This indicates a splicing operation along the channel dimension; Channel shuffling attention is introduced to achieve context reconciliation, i.e., to synthesize the context feature map. Perform a channel shuffling operation, and then compress the spatial dimension using global average pooling to generate channel descriptors. The calculation method is as follows: in, and These represent the spatial height and width of the integrated context feature map, respectively. Represents the pixel coordinates of the integrated context feature map; And based on channel descriptors The channel weights are learned using a two-layer fully connected network, which involves first reducing the dimensionality and then increasing it, and then using a non-linear activation function to map the weights to obtain the channel weights. Its expression is: in, and Let represent the weight matrices of the dimensionality-reduced and dimensionality-increasing fully connected networks, respectively. This refers to the channel compression ratio. , for Activation function, and , Channel descriptor The number of channels; Learned channel weights Applying a channel-by-channel, element-by-element approach to the comprehensive context feature map The context features are relabeled to obtain the final output features of the context-coordinated shuffle attention module. The formula is: in, This indicates a channel-by-channel multiplication operation.

[0014] Furthermore, the specific implementation steps of the marker layer sparsity sensing module in step 2 are as follows: Step 4.1: Combine the fused candidate context features output by the candidate layer sparsity awareness module. Rearrangement and expansion are performed to construct the corresponding encoded tag sequence. Each coded tag is used to characterize local or contextual feature information within the candidate region; Step 4.2: Introduce a masking mechanism to exclude encoded marker sequences. The zero-padding regions and invalid candidate regions are labeled to obtain the encoded label sequence. ; Step 4.3: Process the encoded tag sequence obtained in Step 4.2. Input a sparse label prediction network, evaluate the contribution of each label, predict its activation state, and perform operations based on the activation state. Choose a strategy to generate a binary mask. By filtering the encoded tag sequence element by element, a sparse encoded tag sequence is obtained. ; The sparse label prediction network comprises four normally connected linear units. Each normally connected linear unit includes a normalized layer and a linear layer connected in sequence. A Gaussian error linear unit is placed after the linear layer of the last normally connected linear unit. The input of the first normally connected linear unit is the encoded label sequence. The output of the first normalized linear unit (NRMU) is globally average-pooled and then skip-connected to the second and third NRMUs, respectively, before being used as the input to the third and fourth NRMUs. The formula for the sparse label prediction network is: in, , , and These represent the weight parameters of the 1st to 4th normalized linear layers, This indicates normalization processing. Indicates the activation status of each marker; The activation function is defined as: in, The Gaussian error function is... This represents the output through the last normalized linear unit; Step 4.4: Encode the sparse coding sequence The input encoder encodes a sparse encoded tag sequence through a self-attention mechanism. Model the relationships between them to obtain the encoded labeled features. ; Step 4.5: Encode the marked features In a deformable cross-attention network, the relationship between query tags and encoded features is modeled to obtain decoded query features. The specific steps are as follows: Centered on decoding queries, each query is labeled with features. Set a reference sampling point This reference sampling point is typically determined by the center of the candidate box or the result of the previous decoding, for a given query. ,key Sum Learn a set of sampling offsets through deformable convolution. ; Linear mapping is performed on the query, key, and value respectively, and the corresponding features are extracted at the sampling positions determined based on the reference sampling point and the sampling offset. Its formal representation is as follows: in, This indicates the coordinates corresponding to the reference sampling point. This represents the feature sampling offset predicted by deformable convolution. for The linear embedding weights corresponding to the variables, Represents the encoded tag features The first in The input features corresponding to each sampling point This indicates that bilinear interpolation sampling is performed on the feature map at the sampling location; Simultaneously, the labeled features output by the encoder Perform a linear mapping to generate query features used in the decoding stage. Its expression is: in, For learnable query projection weight matrix, This represents the decoding query features used to guide attention computation during the decoding phase; Based on features Feature fusion is performed using the standard scaled dot product attention mechanism, and its calculation form is as follows: in, The dimension of the key vector. This indicates the query features extracted at the sampling location. This represents the key features extracted at the sampling location and This represents the value feature extracted at the sampling location; Feature sampling offset Adaptive learning via deformable convolution is calculated as follows: in, and These are the weights and bias parameters of the deformable convolution, respectively. Under multi-head attention settings, for the first Each attention point, with fused features Decode query features as input Guided by reference sampling points, the attention results from different sampling points are weighted and converged to obtain the final decoded features. Its expression is: in, The query feature vector is used in the decoding stage to guide the calculation of attention weights and determine reference sampling points. , Indicates the number of attention heads. This indicates the number of sampling points for each attention head. Indicates the first The first one in the attention. The sampling offset of each sampling point, including , , The corresponding sampling offset, For the corresponding attention weights, and satisfying... Its characteristics are determined by the query features. The similarity to the key features at the sampling location is obtained by Softmax normalization. This represents the feature vector obtained through bilinear interpolation; Step 4.6: Process the final decoded features using a feedforward network. The system performs classification and location regression processing to output monitoring results for ships at sea. This includes information on the type of ship target and its corresponding spatial location.

[0015] A satellite synthetic aperture radar (SAR) system for monitoring ships at sea based on three-layer sparse sensing includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the satellite SAR system for monitoring ships at sea based on three-layer sparse sensing.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the satellite synthetic aperture radar maritime vessel monitoring method based on three-layer sparsity sensing.

[0017] A computer program product includes a computer program that, when executed by a processor, implements the steps of a satellite synthetic aperture radar maritime vessel monitoring method based on three-layer sparsity sensing.

[0018] Compared with the prior art, the advantages of the present invention are as follows: I. Sparsity feature modeling capability that fully matches the SAR imaging mechanism: In the feature extraction stage, this invention introduces a feature layer sparsity sensing module to address the sparsity of SAR imaging. By decoupling the traditional two-dimensional global self-attention into two one-dimensional attention processes, range attention and azimuth attention, it enables separate modeling of different scattering mechanisms of SAR images in the range and azimuth directions.

[0019] Because the SAR point spread function exhibits different energy distribution characteristics in the range and azimuth directions, its local response shows significant anisotropy. Directly applying two-dimensional global attention can easily introduce feature associations inconsistent with the physical imaging mechanism, leading to low modeling efficiency and limited generalization ability. This invention further introduces a spatial attenuation prior based on controlled distance during the attention weight calculation process, ensuring that the attention response spatially conforms to the local energy concentration and rapid attenuation law reflected by the SAR point spread function (Sinc), thereby effectively suppressing interference from distant, irrelevant features. Through this mechanism, this invention reduces the computational complexity of attention while making feature associations more consistent with the physical laws of SAR imaging, enabling more accurate differentiation of spatial differences between targets and the background, and fundamentally improving the stability and reliability of target detection in sparse scattering scenarios.

[0020] II. High-order topology sensing capabilities for ship scattering characteristics: Addressing the characteristic that ships in SAR images typically appear as discrete strong scattering points with highly non-uniform scattering responses, the sparse scattering point hypergraph topological sensing layer in this invention introduces a hypergraph structure to model the high-order relationships between multiple scattering points. Through an adaptive hyperedge generation mechanism and a continuously differentiable vertex feature participation modeling method, it can dynamically characterize the collaborative relationships of different scattering points within the same ship structure. This overcomes the limitations of traditional regular grid or local neighborhood modeling, which can only express low-order relationships. This topological sensing mechanism significantly enhances the expressive ability for small ships, weakly scattering ships, and structurally complex ships, effectively improving the stability and robustness of target localization.

[0021] III. Efficient tag-level sparsity modeling capability for the decoding stage: To address the issue of numerous redundant labels involved in the decoding phase of Transformer-based detection frameworks, this invention introduces a sparsity-aware mechanism at the label layer. A sparse label prediction network evaluates the importance of encoded labels, retaining only key labels that contribute significantly to ship monitoring results for subsequent decoding. Combined with a deformable cross-attention network, the decoding process performs attention calculations only at a limited number of sampling locations near the reference point, effectively reducing unnecessary computations and gradient interference. This design significantly reduces computational complexity while improving the targeting of feature decoding, allowing the model to focus more intently on potential ship regions, thereby enhancing detection accuracy and inference efficiency. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the structure of the present invention; Figure 3 This is a schematic diagram of the sparse distance-azimuth multi-head self-attention layer in this invention; Figure 4 This is a schematic diagram of the topological sensing layer of the sparse scattering point hypergraph in this invention; Figure 5 This is a schematic diagram of the sparse candidate generation network of the present invention; Figure 6 This is a schematic diagram of the context-coordinated shuffling attention module of the present invention; Figure 7 This is a schematic diagram of the sparse label prediction network of the present invention; Figure 8 This is a schematic diagram of the deformable cross-attention network of the present invention; Figure 9 This is a schematic diagram of the structure of a computer device according to the present invention; Figure 10 This is a table comparing the accuracy of the present invention with that of the prior art. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, a satellite synthetic aperture radar method for monitoring ships at sea based on three-layer sparsity sensing includes the following steps: Step 1: Obtain the synthetic aperture radar image of the marine scene to be monitored, divide it into several image blocks, and combine it with the position coding to form the original input features; The combination of image patches and positional coding is obtained through sine and cosine positional embedding, which is defined as: in, This indicates the position number of the input image patch in the spatial sequence. Indicates the frequency index parameter. , Indicates the embedding dimension of the position encoding. Indicates location number In the even-numbered dimensions of the embedded image patch Position code on Indicates location number In the odd-numbered dimensions of the embedded image patch Position encoding on.

[0026] Step 2: Based on the three-layer sparse perception model, monitor the category information of ship targets and their corresponding spatial location information of the original input features; The three-layer sparsity perception model includes a feature layer sparsity perception module that extracts sparse features from the original input features and models sparse features to obtain a sparse feature map; a candidate layer sparsity perception module that generates candidate boxes based on the sparse feature map; and a label layer sparsity perception module that performs sparse labeling on the features corresponding to the candidate boxes, decodes the sparse labels, and outputs the category information of the ship target and its corresponding spatial location information. Figure 2 As shown.

[0027] like Figure 3As shown, the feature layer sparsity perception module includes four feature sparsity perception layers connected in sequence. Each feature sparsity perception layer includes a sparse distance-azimuth multi-head self-attention layer that extracts sparse features with directional decoupling by performing a distance-azimuth decoupling self-attention mechanism on the input features; a sparse scattering point hypergraph topology perception layer that models the high-order topological relationship between ship scattering points based on sparse features and a sparse scattering point hypergraph topology perception mechanism to obtain a sparse feature map; and a downsampling layer. The input features of the sparse distance-azimuth multi-head self-attention layer in the first feature sparsity perception layer are the original input features; the input features of the second feature sparsity perception layer are the output of the first feature sparsity perception layer; the input features of the third feature sparsity perception layer are the output of the second feature sparsity perception layer; and the input features of the fourth feature sparsity perception layer are the output of the third feature sparsity perception layer. The Sparse Range-Azimuth Multi-Head Self-Attention (SRA-MSA) layer utilizes the sparse imaging characteristics of SAR in the range and azimuth directions to decouple global self-attention into row attention in the range direction and column attention in the azimuth direction.

[0028] The specific implementation steps of the sparse distance-azimuth multi-head self-attention layer are as follows: The input features are used to form feature vector sequences in the range and azimuth directions, respectively. A gated recurrent unit (GRU) or LSTM is then used to describe the pixel dependencies in the same direction. The formula for the GRU is as follows: in, and These represent the feature vectors in the range and azimuth directions, respectively. Subsequently, the query vector and key vector for the corresponding directions are obtained through linear embedding, as shown in the formula: in, , These represent the query vectors in the distance and azimuth directions, respectively. , These represent the key vectors for the range and azimuth directions, respectively. , , and They represent , , and The corresponding embedding weights; After linear embedding, the query vectors and key vectors in the range and azimuth directions are multiplied separately to obtain the attention maps in the range and azimuth directions, respectively. These are then processed using the Softmax function, and finally, the Manhattan distance decay function is introduced. The final attention map is obtained by weighted suppression of spatial correlation, and the formula is: in, , These represent the final attention maps in the distance and azimuth directions, respectively. and These represent the Manhattan distance decay functions introduced in the range and azimuth directions, respectively. and They represent and transpose, This represents element-wise multiplication. and The Middle The Manhattan distance decay function for each location index is: and The formula is: in, This indicates the center of Manhattan, which is the reference center location used for calculating distances to Manhattan. The first part represents the attention map in the distance direction and the attention map in the azimuth direction. Location index, The attenuation coefficient; Attention maps are calculated separately for the range and azimuth directions, and a Manhattan distance decay function is introduced after Softmax. Spatial correlation is weighted and suppressed, thus making attention more focused on key response regions near the reference location. The final decoupled sparse attention feature is obtained by multiplying the distance attention pair of the input feature with the final attention map of the azimuth direction, and then multiplying the result with the final attention map of the distance direction. The formula is as follows: in, , representing the distance to the attention pair value, Represents the projection weight matrix. Indicates input features, This indicates transpose.

[0029] The Sparse Scattering Point Hyper-Graph Topology Perception (SSP-HGTP) layer comprises an adaptive hyperedge generation module and a hypergraph convolution operation module connected sequentially. Addressing the discrete, non-uniform, and high-order structural relationships of SAR ship target scattering points, the SSP-HGTP layer introduces a hypergraph topology modeling approach to explicitly characterize the multi-point relationships between sparse scattering points. SSP-HGTP consists of two sequential parts: Adaptive Hyperedge Generation (AHG) and Hypergraph Convolution Operation (HCO). Figure 4 As shown.

[0030] The specific implementation steps of the adaptive hyperedge generation module are as follows: Sparsity features As input, it is flattened to obtain the set of vertex features. Let the hypergraph be represented as... ,in, This represents the set of vertex features, corresponding to the potential scattering point features in synthetic aperture radar (SAR) images of maritime scenes. This represents a set of hyperedges, where each hyperedge connects multiple vertex features to characterize their higher-order relationships. A continuously differentiable vertex participation mechanism is introduced, using a continuous participation matrix. The hypergraph is metric, where, and These represent the number of vertex features and the number of hyperedges, respectively. Representing vertex features Participating in the super-border To what extent, This represents an identifier and has no specific meaning. The specific steps for measuring a hypergraph are as follows: Let the vertex features be Vertex features Perform global average pooling and global max pooling, and then concatenate the two to obtain the global vertex. Its expression is: in, This represents averaging the features of all vertices over the spatial dimension. This represents taking the maximum value of all vertex features in the spatial dimension. For channel dimension, Indicates length is A real-valued vector; Subsequently, through the projection layer Generate global vertex representation global offset and in the global prototype Building on A dynamic hyperedge prototype Specifically, it is expressed as: in, That is, mapping the global vertex representation to A hyperedge prototype; Simultaneously, the correlation between vertex features and hyperedges is calculated, i.e., another projection layer is introduced. The formula for mapping vertex features to query vectors is as follows: in, Represents the query vector. , Indicates length is The real-valued vector, This represents the query projection weight matrix, used to weight the input vertex features. Mapped to query vector ; Subsequently, the query vector conduct Subspace decomposition yields , ,in, Indicates the first query vectors The decomposition yielded the first Subspace; For dynamic hyperedge prototypes Perform the same subspace decomposition to obtain , Indicates the first A dynamic hyperedge prototype The decomposition yielded the first Subspace; In the In each subspace, the query vector of vertex features With the hyper-edge prototype The similarity is calculated using the scaled dot product, with the following formula: The overall similarity is obtained by averaging the subspace results based on the similarity score, using the following formula: Next, regarding Normalization is performed to obtain the participation of vertex features in the hyperedge. The formula is: The specific implementation steps of the hypergraph convolution operation module are as follows: Based on participation, the features of the vertices connected to the hyperedge are aggregated using hyperedges to achieve feature mapping from vertex features to hyperedges. Then, the mapped hyperedge-level features are backpropagated back to the vertex features to update them. The aggregation process uses Choquet integrals, and this bidirectional aggregation process can be represented as: in, and Represented as a learnable weight matrix, This represents the Sigmoid activation function. , This indicates the parameter or formula to be substituted into the parentheses. Indicates the first Vertex eigenvectors.

[0031] Through the above process, SSP-HGTP can effectively model the high-order topological relationships between sparse scattering points of ships, making features more discriminative at the spatial structure level, thereby providing more structure-aware feature representations for subsequent candidate generation and target decoding.

[0032] The Sparse Proposal Generation Network (SPGN) addresses the highly sparse spatial distribution of ships in large-scale SAR images. In the candidate generation stage, it avoids traditional dense sliding window or dense anchor box strategies. Instead, it generates a limited number of candidate boxes only in high-confidence regions through location sparsity and quantity constraints, thereby reducing computational redundancy and improving candidate quality. Figure 5 As shown, the specific implementation steps of the candidate layer sparsity sensing module in step 2 are as follows: Step 3.1: Sparse feature map output by the feature layer sparsity perception module. Inputting the sparse candidate generation network yields the basic features of the candidate regions. The specific steps are as follows: Step 3.11: Sparse feature map output by the feature layer sparsity perception module. Through channel number 1 Convolution is used for position encoding to obtain position response features. The formula is: in, and They represent Convolution weights and biases; Position response features are obtained by using the Sigmoid activation function. Mapped to the probability value of each spatial location belonging to the candidate region. And based on a preset threshold Determine candidate activation regions The formula is: Step 3.12: Based on the set of candidate activation locations and position response characteristics Generate a set of candidate activation locations The set of candidate boxes within the specified sparse region ; Step 3.13: Introduce a feature adaptation mechanism to dynamically adjust the original features through deformable convolution: specifically, utilize positional response features. Guided deformable convolution on sparse feature maps Perform adaptive feature adjustment to obtain the aligned feature map. Aligned feature maps The first in Feature maps of candidate activation locations The calculation formula is: in, Indicates the number of convolution sampling points. For the first Fixed sampling positions for regular convolution with a number of convolution sampling points Indicates the first The scientific system offset of each sampling point Indicates the first The modulation coefficients corresponding to each sampling point Indicates the first The weights corresponding to each sampling point; Step 3.14: Set the candidate boxes Projected onto the aligned feature map Based on the corresponding coordinate system, high-dimensional feature representations of each candidate box are extracted using a dynamic header, and the RoIAlign operation is performed to obtain the basic features of the corresponding candidate regions. The formula is: in, The RoIAlign function represents the feature extraction function for candidate regions. Each candidate box is extracted using projection operations and a dynamic head, resulting in a high-dimensional feature representation. Each candidate box corresponds to one dimension. The feature vectors form a set The candidate feature set provides high-quality, low-redundancy input for subsequent context coordination modeling and sparse decoding of the label layer.

[0033] Step 3.2: Set the candidate boxes According to the scaling factor respectively and Scale expansion is performed to obtain multi-scale context candidate boxes. and and in the feature map Perform RoIAlign operations on each of the above to obtain multi-scale contextual features. and The formula is: The Context Harmony Shuffle Attention (CHSA) module addresses the problem of insufficient contextual information after sparse candidate generation due to the spatial constraints of candidate regions. In maritime ship SAR monitoring missions, a single candidate region often only covers a portion of the ship's scattering features. Directly modeling features based on this region can easily overlook surrounding environmental information crucial for discrimination. Therefore, CHSA employs a multi-scale context generation + channel shuffle attention approach to introduce contextual compensation while suppressing low-value interference information, such as… Figure 6 As shown.

[0034] Step 3.3: Identify the basic features of the candidate regions. and multi-scale contextual features , The input context-coordinated shuffle attention module coordinates and fuses contextual information at different scales, and outputs the fused candidate contextual features. The specific steps are as follows: Basic features of candidate regions and multi-scale contextual features , Apply the corresponding Convolution is used to refine local features, and these features are then concatenated along the channel dimension to obtain a comprehensive contextual feature map. The formula is: in, , and Indicates correspondence The weight parameters of the convolutional layer, This indicates a splicing operation along the channel dimension; Channel shuffling attention is introduced to achieve context reconciliation, i.e., to synthesize the context feature map. Perform a channel shuffling operation, and then compress the spatial dimension using global average pooling to generate channel descriptors. The calculation method is as follows: in, and These represent the spatial height and width of the integrated context feature map, respectively. Represents the pixel coordinates of the integrated context feature map; And based on channel descriptors The channel weights are learned using a two-layer fully connected network, which involves first reducing the dimensionality and then increasing it, and then using a non-linear activation function to map the weights to obtain the channel weights. Its expression is: in, and Let represent the weight matrices of the dimensionality-reduced and dimensionality-increasing fully connected networks, respectively. This refers to the channel compression ratio. , for Activation function, and , Channel descriptor The number of channels; Learned channel weights Applying a channel-by-channel, element-by-element approach to the comprehensive context feature map The context features are relabeled to obtain the final output features of the context-coordinated shuffle attention module. The formula is: in, This indicates a channel-by-channel multiplication operation.

[0035] The feature labels corresponding to the candidate boxes are evaluated and screened for value, and only sparse labels that contribute to ship monitoring are retained.

[0036] The specific implementation steps of the marker layer sparsity awareness module are as follows: Step 4.1: Combine the fused candidate context features output by the candidate layer sparsity awareness module. Rearrangement and expansion are performed to construct the corresponding encoded tag sequence. Each coded tag is used to characterize local or contextual feature information within the candidate region; Step 4.2: Introduce a masking mechanism to exclude encoded marker sequences. The zero-padding regions and invalid candidate regions are labeled to obtain the encoded label sequence. To prevent these invalid labels from participating in the subsequent Hungarian matching and gradient propagation process, thereby avoiding meaningless gradients interfering with model training; Step 4.3: Process the encoded label sequence obtained in Step 4.2. Input a sparse label prediction network, evaluate the contribution of each label, predict its activation state, and perform operations based on the activation state. Choose a strategy to generate a binary mask. By filtering the encoded tag sequence element by element, a sparse encoded tag sequence is obtained. ; The Sparse Token Prediction Network (STPN) addresses the problems of low feature decoding efficiency and insufficient target feature representation caused by a large number of redundant tokens participating in the Transformer decoding stage. In SAR maritime vessel monitoring missions, only a small number of tokens are highly correlated with the actual vessel targets, while the remaining tokens mostly originate from the background or invalid regions, such as... Figure 7 As shown.

[0037] The sparse label prediction network comprises four normally connected linear units. Each normally connected linear unit includes a normalized layer and a linear layer connected in sequence. A Gaussian error linear unit is placed after the linear layer of the last normally connected linear unit. The input of the first normally connected linear unit is the encoded label sequence. The output of the first normalized linear unit (NRMU) is globally average-pooled and then skip-connected to the second and third NRMUs, respectively, before being used as the input to the third and fourth NRMUs. The formula for the sparse label prediction network is: in, , , and These represent the weight parameters of the 1st to 4th normalized linear layers, This indicates normalization processing. Indicates the activation status of each marker; The activation function is defined as: in, The Gaussian error function outputs a one-dimensional logit, used for subsequent binary cross-entropy (BCE) loss calculations to determine whether each label should be retained. This represents the output through the last normalized linear unit; Specifically, in the first layer of output features, the channel dimension is divided into two parts. One part extracts global features by performing global average pooling on all labels, while the other part retains local label features. Then, the global features are copied and concatenated back to each label position, thus ensuring that each label possesses both local semantic information and global contextual information without changing the feature dimension. Step 4.4: Sparsely encoded label sequence. The input encoder encodes a sparse encoded tag sequence through a self-attention mechanism. Model the relationships between them to obtain the encoded labeled features. ; Step 4.5: Encode the marked features In a deformable cross-attention network, the relationship between query tags and encoded features is modeled to obtain decoded query features. The Deformable Cross-Attention Network (DCAN) introduces a reference-point-based sparse sampling mechanism, enabling the decoding process to focus only on a limited number of key sampling points near the reference location, thereby achieving sparsity awareness during the decoding stage. For example... Figure 8 As shown.

[0038] The specific steps are as follows: Centered on decoding queries, each query is labeled with features. Set a reference sampling point This reference sampling point is typically determined by the center of the candidate box or the result of the previous decoding, for a given query. ,key Sum Learn a set of sampling offsets through deformable convolution. ; Linear mapping is performed on the query, key, and value respectively, and the corresponding features are extracted at the sampling positions determined based on the reference sampling point and the sampling offset. Its formal representation is as follows: in, This indicates the coordinates corresponding to the reference sampling point. This represents the feature sampling offset predicted by deformable convolution. for The linear embedding weights corresponding to the variables, Represents the encoded tag features The first in The input features corresponding to each sampling point This indicates that bilinear interpolation sampling is performed on the feature map at the sampling location; Simultaneously, the labeled features output by the encoder Perform a linear mapping to generate query features used in the decoding stage. Its expression is: in, For learnable query projection weight matrix, This represents the decoding query features used to guide attention computation during the decoding phase; Based on features Feature fusion is performed using the standard scaled dot product attention mechanism, and its calculation form is as follows: in, The dimension of the key vector. This indicates the query features extracted at the sampling location. This represents the key features extracted at the sampling location and This represents the value feature extracted at the sampling location; Feature sampling offset Adaptive learning via deformable convolution is calculated as follows: in, and These are the weights and bias parameters of the deformable convolution, respectively. Under multi-head attention settings, for the first Each attention point, with fused features Decode query features as input Guided by reference sampling points, the attention results from different sampling points are weighted and converged to obtain the final decoded features. Its expression is: in, The query feature vector is used in the decoding stage to guide the calculation of attention weights and determine reference sampling points. , Indicates the number of attention heads. This indicates the number of sampling points for each attention head. Indicates the first The first one in the attention. The sampling offset of each sampling point, including , , The corresponding sampling offset, For the corresponding attention weights, and satisfying... Its characteristics are determined by the query features. The similarity to the key features at the sampling location is obtained by Softmax normalization. This represents the feature vector obtained through bilinear interpolation; Step 4.6: Process the final decoded features using a feedforward network. The system performs classification and location regression processing to output monitoring results for ships at sea. This includes information on the type of ship target and its corresponding spatial location.

[0039] Experiments were conducted on two publicly available SAR ship datasets, SSDD and HRSID, with data sources including RadarSat, TerraSAR-X, and Sentinel-1, covering multi-resolution and multi-scenario maritime ship targets. The COCO evaluation metric was used, with the Average Precision (AP) defined as follows: in, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively. Evaluation metrics include... (Average accuracy at IoU=0.5) (Strict accuracy at IoU=0.75), and for targets of different scales. (Small goals) (Medium target) and (Major objective). When analyzing false alarms and missed detections, the following methods are used: The indicator is calculated as follows: like Figure 10 As shown, on the SSDD and HRSID datasets, the average accuracy of the Triple-Level Sparsity Awareness (TLSA) method is 3.1% and 3.7% higher than the second-best methods, respectively, demonstrating its state-of-the-art performance. Furthermore, TLSA achieves the highest results on all evaluation metrics, a feat unmatched by other methods. This method truly considers the key characteristics of SAR maritime vessel surveillance missions—namely, SAR imaging sparsity and ship scattering sparsity—characteristics often neglected in existing methods. Therefore, TLSA's performance is the most reasonable and effective.

[0040] This invention significantly improves the accuracy and stability of maritime vessel monitoring. By introducing a range-azimuth decoupled self-attention mechanism and a sparse scattering point hypergraph topological sensing mechanism at the feature layer, the method explicitly models the spread function characteristics of SAR imaging points and the discrete distribution of ship scattering points during the feature extraction stage, making the extracted features more consistent with the physical imaging mechanism. This feature has stronger discriminative power in spatial structure and semantic expression, effectively reducing false alarms and missed detections, and improving the overall detection accuracy of ship targets at different scales. This invention effectively reduces computational complexity and improves inference efficiency. By introducing a sparse candidate generation strategy at the candidate layer, a limited number of candidate boxes are generated only in high-confidence regions, avoiding dense scanning across the entire large-scale ocean scene, significantly reducing redundant candidates and subsequent computational burden. Simultaneously, at the labeling layer, sparse label prediction and a deformable cross-attention mechanism are used to perform attention calculations only on a small number of key labels and spatial locations highly correlated with the ship target, allowing for a more rational allocation of computational resources, thereby improving overall inference efficiency while maintaining accuracy. This invention exhibits stronger robustness and engineering applicability in complex marine environments. Through the synergistic effect of multi-layered sparsity sensing, the method maintains stable monitoring performance even in scenarios with complex backgrounds, few ship targets, or insignificant scattering characteristics, reducing false detections caused by sea clutter, shoreline structures, or strong scattering facilities. Furthermore, while maintaining performance advantages, the method has a low parameter scale and computational requirements, making it suitable for deployment in edge computing devices or real-time monitoring systems, and possessing significant engineering application value.

[0041] Based on the same inventive concept, this application also provides a satellite synthetic aperture radar maritime vessel monitoring system, device, or apparatus for implementing the aforementioned three-layer sparse sensing-based system. It may include a server or terminal, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory of the system, device, or apparatus includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating devices and computer programs stored in the non-volatile storage media. The database of the system, device, or apparatus stores target detection data for sensing, retrieving, and locating synthetic aperture radar (SAR) remote sensing images. The I / O interfaces of the system, device, or apparatus are used for exchanging information between the processor and external devices. The communication interface of the system, device, or apparatus is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a sensing, retrieving, and locating detection method for SAR remote sensing images.

[0042] Those skilled in the art will understand that, Figure 9 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices or apparatus to which the present invention is applied. Specifically, they may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0043] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0044] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0045] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0046] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0047] The databases involved in the various embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided by this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic units, etc., and are not limited to these.

[0048] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing, characterized in that, Includes the following steps: Step 1: Obtain the synthetic aperture radar image of the marine scene to be monitored, divide it into several image blocks, and combine it with the position coding to form the original input features; Step 2: Based on the three-layer sparse perception model, monitor the category information of ship targets and their corresponding spatial location information of the original input features; The three-layer sparsity perception model includes a feature layer sparsity perception module that extracts sparse features from the original input features and models sparse features to obtain a sparse feature map; a candidate layer sparsity perception module that generates candidate boxes based on the sparse feature map; and a label layer sparsity perception module that performs sparse labeling on the features corresponding to the candidate boxes, decodes the sparse labels, and outputs the category information of the ship target and its corresponding spatial location information.

2. The method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing according to claim 1, characterized in that: The combination of image blocks and positional encoding in step 1 is obtained through sine and cosine positional embedding, which is defined as follows: in, This indicates the position number of the input image patch in the spatial sequence. Indicates the frequency index parameter. , Indicates the embedding dimension of the position encoding. Indicates location number In the even-numbered dimensions of the embedded image patch Position code on Indicates location number In the odd-numbered dimensions of the embedded image patch Position encoding on.

3. The method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing according to claim 2, characterized in that: The feature sparsity perception module in step 2 includes four sequentially connected feature sparsity perception layers. Each feature sparsity perception layer includes a sparse distance-azimuth multi-head self-attention layer that extracts sparse features with directional decoupling by performing a distance-azimuth decoupling self-attention mechanism on the input features, a sparse scattering point hypergraph topology perception layer that models the high-order topological relationship between ship scattering points based on sparse features and a sparse scattering point hypergraph topology perception mechanism to obtain a sparse feature map, and a downsampling layer. The input features of the sparse distance-azimuth multi-head self-attention layer in the first feature sparsity perception layer are the original input features, the input features of the second feature sparsity perception layer are the output of the first feature sparsity perception layer, the input features of the third feature sparsity perception layer are the output of the second feature sparsity perception layer, and the input features of the fourth feature sparsity perception layer are the output of the third feature sparsity perception layer.

4. The method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing according to claim 3, characterized in that: The specific implementation steps of the sparse distance-azimuth multi-head self-attention layer are as follows: The input features are used to form feature vector sequences in the range and azimuth directions, respectively. A gated recurrent unit (GRU) or LSTM is then used to describe the pixel dependencies in the same direction. The formula for the GRU is as follows: in, and These represent the feature vectors in the range and azimuth directions, respectively. Subsequently, the query vector and key vector for the corresponding directions are obtained through linear embedding, as shown in the formula: in, , These represent the query vectors in the distance and azimuth directions, respectively. , These represent the key vectors for the range and azimuth directions, respectively. , , and They represent , , and The corresponding embedding weights; After linear embedding, the query vectors and key vectors in the range and azimuth directions are multiplied separately to obtain the attention maps in the range and azimuth directions, respectively. These are then processed using the Softmax function, and finally, the Manhattan distance decay function is introduced. The final attention map is obtained by weighted suppression of spatial correlation, and the formula is: in, , These represent the final attention maps in the distance and azimuth directions, respectively. and These represent the Manhattan distance decay functions introduced in the range and azimuth directions, respectively. and They represent and transpose, This represents element-wise multiplication. and The Middle The Manhattan distance decay function for each location index is: and The formula is: in, This indicates the center of Manhattan, which is the reference center location used for calculating distances to Manhattan. The first part represents the attention map in the distance direction and the attention map in the azimuth direction. Location index, The attenuation coefficient; The final decoupled sparse attention feature is obtained by multiplying the distance attention pair of the input feature with the final attention map of the azimuth direction, and then multiplying the result with the final attention map of the distance direction. The formula is as follows: in, , representing the distance to the attention pair value, Represents the projection weight matrix. Indicates input features, This indicates transpose.

5. A method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing, as described in claim 4, is characterized in that: The sparse scattering point hypergraph topology sensing layer includes an adaptive hyperedge generation module and a hypergraph convolution operation module connected in sequence. The specific implementation steps of the adaptive hyperedge generation module are as follows: Sparsity features As input, it is flattened to obtain the set of vertex features. Let the hypergraph be represented as... ,in, This represents the set of vertex features, corresponding to the potential scattering point features in synthetic aperture radar (SAR) images of maritime scenes. This represents a set of hyperedges, where each hyperedge connects multiple vertex features to characterize their higher-order relationships. A continuously differentiable vertex participation mechanism is introduced, using a continuous participation matrix. The hypergraph is metric, where, and These represent the number of vertex features and the number of hyperedges, respectively. Representing vertex features Participating in the super-border To what extent, This represents an identifier and has no specific meaning. The specific steps for measuring a hypergraph are as follows: Let the vertex features be Vertex features Perform global average pooling and global max pooling, and then concatenate the two to obtain the global vertex. Its expression is: in, This represents averaging the features of all vertices over the spatial dimension. This represents taking the maximum value of all vertex features in the spatial dimension. For channel dimension, Indicates length is A real-valued vector; Subsequently, through the projection layer Generate global vertex representation global offset and in the global prototype Building on A dynamic hyperedge prototype Specifically, it is expressed as: in, That is, mapping the global vertex representation to A hyperedge prototype; Simultaneously, the correlation between vertex features and hyperedges is calculated, i.e., another projection layer is introduced. The formula for mapping vertex features to query vectors is as follows: in, Represents the query vector. , Indicates length is The real-valued vector, This represents the query projection weight matrix, used to weight the input vertex features. Mapped to query vector ; Subsequently, the query vector conduct Subspace decomposition yields , ,in, Indicates the first query vectors The decomposition yielded the first Subspace; For dynamic hyperedge prototypes Perform the same subspace decomposition to obtain , Indicates the first A dynamic hyperedge prototype The decomposition yielded the first Subspace; In the In each subspace, the query vector of vertex features With the hyper-edge prototype The similarity is calculated using the scaled dot product, with the following formula: The overall similarity is obtained by averaging the subspace results based on the similarity score, using the following formula: Next, regarding Normalization is performed to obtain the participation of vertex features in the hyperedge. The formula is: The specific implementation steps of the hypergraph convolution operation module are as follows: Based on participation, the features of the vertices connected to the hyperedge are aggregated using hyperedges to achieve feature mapping from vertex features to hyperedges. Then, the mapped hyperedge-level features are backpropagated back to the vertex features to update them. The aggregation process uses Choquet integrals, and this bidirectional aggregation process can be represented as: in, and Represented as a learnable weight matrix, This represents the Sigmoid activation function. Indicates the first Vertex eigenvectors.

6. The method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing according to claim 5, characterized in that: The specific implementation steps of the candidate layer sparsity sensing module in step 2 are as follows: Step 3.1: Sparse feature map output by the feature layer sparsity perception module. Inputting the sparse candidate generation network yields the basic features of the candidate regions. The specific steps are as follows: Step 3.11: Sparse feature map output by the feature layer sparsity perception module. ,pass Convolution is used for position encoding to obtain position response features. The formula is: in, and They represent Convolution weights and biases; Position response features are obtained by using the Sigmoid activation function. Mapped to the probability value of each spatial location belonging to the candidate region. And based on a preset threshold Determine candidate activation regions The formula is: Step 3.12: Based on the set of candidate activation locations and position response characteristics Generate a set of candidate activation locations The set of candidate boxes within the specified sparse region ; Step 3.13: Utilize position response characteristics Guided deformable convolution on sparse feature maps Perform adaptive feature adjustment to obtain the aligned feature map. Aligned feature maps The first in Feature maps of candidate activation locations The calculation formula is: in, Indicates the number of convolution sampling points. For the first Fixed sampling positions for regular convolution with a number of convolution sampling points Indicates the first The scientific system offset of each sampling point Indicates the first The modulation coefficients corresponding to each sampling point Indicates the first The weights corresponding to each sampling point; Step 3.14: Set the candidate boxes Projected onto the aligned feature map Based on the corresponding coordinate system, high-dimensional feature representations of each candidate box are extracted using a dynamic header, and the RoIAlign operation is performed to obtain the basic features of the corresponding candidate regions. The formula is: in, The RoIAlign function represents the feature extraction function for candidate regions. Step 3.2: Set the candidate boxes According to the scaling factor respectively and Scale expansion is performed to obtain multi-scale context candidate boxes. and and in the feature map Perform RoIAlign operations on each of the above to obtain multi-scale contextual features. and The formula is: Step 3.3: Identify the basic features of the candidate regions. and multi-scale contextual features , The input context-coordinated shuffle attention module coordinates and fuses contextual information at different scales, and outputs the fused candidate contextual features. The specific steps are as follows: Basic features of candidate regions and multi-scale contextual features , Apply Convolution is used to refine local features, and these features are then concatenated along the channel dimension to obtain a comprehensive contextual feature map. The formula is: in, , and Indicates correspondence The weight parameters of the convolutional layer, This indicates a splicing operation along the channel dimension; Channel shuffling attention is introduced to achieve context reconciliation, i.e., to synthesize the context feature map. Perform a channel shuffling operation, and then compress the spatial dimension using global average pooling to generate channel descriptors. The calculation method is as follows: in, and These represent the spatial height and width of the integrated context feature map, respectively. Represents the pixel coordinates of the integrated context feature map; And based on channel descriptors The channel weights are learned using a two-layer fully connected network, which involves first reducing the dimensionality and then increasing it, and then using a non-linear activation function to map the weights to obtain the channel weights. Its expression is: in, and Let represent the weight matrices of the dimensionality-reduced and dimensionality-increasing fully connected networks, respectively. This refers to the channel compression ratio. , for Activation function, and , Channel descriptor The number of channels; Learned channel weights Applying a channel-by-channel, element-by-element approach to the comprehensive context feature map The context features are relabeled to obtain the final output features of the context-coordinated shuffle attention module. The formula is: in, This indicates a channel-by-channel multiplication operation.

7. A method for monitoring ships at sea using satellite synthetic aperture radar based on three-layer sparsity sensing, as described in claim 6, is characterized in that: The specific implementation steps of the marker layer sparsity sensing module in step 2 are as follows: Step 4.1: Combine the fused candidate context features output by the candidate layer sparsity awareness module. Rearrangement and expansion are performed to construct the corresponding encoded tag sequence. Each coded tag is used to characterize local or contextual feature information within the candidate region; Step 4.2: Introduce a masking mechanism to exclude encoded marker sequences. The zero-padding regions and invalid candidate regions are labeled to obtain the encoded label sequence. ; Step 4.3: Process the encoded tag sequence obtained in Step 4.

2. Input a sparse label prediction network, evaluate the contribution of each label, predict its activation state, and perform operations based on the activation state. Choose a strategy to generate a binary mask. By filtering the encoded tag sequence element by element, a sparse encoded tag sequence is obtained. ; The sparse label prediction network comprises four normally connected linear units. Each normally connected linear unit includes a normalized layer and a linear layer connected in sequence. A Gaussian error linear unit is placed after the linear layer of the last normally connected linear unit. The input of the first normally connected linear unit is the encoded label sequence. The output of the first normalized linear unit (NRMU) is globally average-pooled and then skip-connected to the second and third NRMUs, respectively, before being used as the input to the third and fourth NRMUs. The formula for the sparse label prediction network is: in, , , and These represent the weight parameters of the 1st to 4th normalized linear layers, This indicates normalization processing. Indicates the activation status of each marker; The activation function is defined as: in, The Gaussian error function is... This represents the output through the last normalized linear unit; Step 4.4: Encode the sparse coding sequence The input encoder encodes a sparse encoded tag sequence through a self-attention mechanism. Model the relationships between them to obtain the encoded labeled features. ; Step 4.5: Encode the marked features In a deformable cross-attention network, the relationship between query tags and encoded features is modeled to obtain decoded query features. The specific steps are as follows: Centered on decoding queries, each query is labeled with features. Set a reference sampling point This reference sampling point is typically determined by the center of the candidate box or the result of the previous decoding, for a given query. ,key Sum Learn a set of sampling offsets through deformable convolution. ; Linear mapping is performed on the query, key, and value respectively, and the corresponding features are extracted at the sampling positions determined based on the reference sampling point and the sampling offset. Its formal representation is as follows: in, This indicates the coordinates corresponding to the reference sampling point. This represents the feature sampling offset predicted by deformable convolution. for The linear embedding weights corresponding to the variables, Represents the encoded tag features The first in The input features corresponding to each sampling point This indicates that bilinear interpolation sampling is performed on the feature map at the sampling location; Simultaneously, the labeled features output by the encoder Perform a linear mapping to generate query features used in the decoding stage. Its expression is: in, For learnable query projection weight matrix, This represents the decoding query features used to guide attention computation during the decoding phase; Based on features Feature fusion is performed using the standard scaled dot product attention mechanism, and its calculation form is as follows: in, The dimension of the key vector. This indicates the query features extracted at the sampling location. This represents the key features extracted at the sampling location and This represents the value feature extracted at the sampling location; Feature sampling offset Adaptive learning via deformable convolution is calculated as follows: in, and These are the weights and bias parameters of the deformable convolution, respectively. Under multi-head attention settings, for the first Each attention point, with fused features Decode query features as input Guided by reference sampling points, the attention results from different sampling points are weighted and converged to obtain the final decoded features. Its expression is: in, The query feature vector is used in the decoding stage to guide the calculation of attention weights and determine reference sampling points. , Indicates the number of attention heads. This indicates the number of sampling points for each attention head. Indicates the first The first one in the attention. The sampling offset of each sampling point, including , , The corresponding sampling offset, For the corresponding attention weights, and satisfying... Its characteristics are determined by the query features. The similarity to the key features at the sampling location is obtained by Softmax normalization. This represents the feature vector obtained through bilinear interpolation; Step 4.6: Process the final decoded features using a feedforward network. The system performs classification and location regression processing to output monitoring results for ships at sea. This includes information on the type of ship target and its corresponding spatial location.

8. A satellite synthetic aperture radar maritime vessel monitoring system based on three-layer sparsity sensing, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

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

10. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

Images