SAR ship target classification method based on CNN-Mama network and scattering topology features
By constructing a dual-branch network based on CNN-Mamba network and scattering topological features, the problem of insufficient utilization of target scattering features in SAR images is solved, and the accuracy and robustness of SAR ship target classification are improved.
Patent Information
- Application Number
- CN202510465863.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-07-29
AI Technical Summary
The insufficient utilization of scattering features of targets in existing SAR images and the limited ability to extract features of single CNN models, resulting in unsatisfactory classification results of SAR ship targets.
A dual-branch network based on CNN-Mamba network and scattering topological features is constructed, including CNN-Mamba feature extraction network and scattering topological feature extraction network. Local features are extracted through the residual convolution module and the state space modeling module, and global features are extracted in combination with scattering point extraction and topological feature construction module, and visual and topological features are fused for classification.
It significantly improves the accuracy and robustness of SAR ship target classification, making full use of the scattering characteristics and topological structure information of SAR images.
Smart Images
Figure CN120388224A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of SAR remote sensing target recognition, and particularly to a SAR ship target classification method based on a CNN-Mamba network and scattering topological features. Background Art
[0002] Synthetic Aperture Radar (SAR) is an advanced active microwave radar that can perform high-precision imaging of the ground or targets all day and all weather, and has played an important role in fields such as marine surveillance, environmental monitoring, aerospace, etc. SAR ship classification is a technology for fine-grained recognition of ship targets in SAR images and is an important task of SAR remote sensing.
[0003] Deep learning algorithms have powerful feature extraction capabilities and good robustness, and have demonstrated excellent performance in SAR ship target recognition tasks. However, models based on Convolutional Neural Networks (CNNs) are limited by local receptive fields, which usually leads to insufficient feature extraction and thus unsatisfactory classification results; models based on Transformers perform well in global modeling, but due to their self-attention mechanism having quadratic complexity, the processing of high-dimensional data such as images will bring a large computational burden, restricting the application scenarios of the models. In addition, different from optical images, SAR images have characteristics such as scattering, polarization, and speckle noise. And SAR images are sensitive to imaging angles, and the SAR images of the same ship target show obvious differences at different angles, which further increases the difficulty of classification.
[0004] Therefore, it is necessary to develop a new SAR ship target classification method to solve the problems of insufficient utilization of the scattering characteristics of targets in SAR images and limited feature extraction capabilities of a single CNN model. Summary of the Invention
[0005] Aiming at the problems of insufficient utilization of the scattering characteristics of targets in SAR images and limited feature extraction capabilities of a single CNN model in the existing methods, the present invention proposes a SAR ship target classification method based on a CNN-Mamba network and scattering topological features. The Mamba model based on State Space Modeling (SSM) is introduced into the SAR ship classification problem, and a CNN-Mamba feature extraction network is constructed to fully utilize the local feature extraction ability of the convolutional layer and the long-distance modeling ability of SSM. A scattering topological feature extraction network is constructed to obtain scattering topological features to represent information such as the structural form of ship targets, so as to enrich the feature representation of ship targets in SAR images. Tests on public datasets show that the proposed method can effectively extract the features of ship targets in SAR images, significantly improving the accuracy and robustness of SAR ship target classification.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] The present invention provides a SAR ship target classification method based on a CNN-Mamba network and scattering topological features. The method constructs a two-branch network, including a CNN-Mamba feature extraction network and a scattering topological feature extraction network;
[0008] The CNN-Mamba feature extraction network includes four stages. Each stage is composed of several Conv-SSM modules. Each Conv-SSM module is composed of a residual convolution module and an SSM module in parallel. The stages are connected by a downsampling module to extract the visual features of the SAR ship image;
[0009] The scattering topological feature extraction network includes a scattering point extraction module and a topological feature construction module to extract the scattering topological features of the SAR ship image;
[0010] The visual features obtained by the CNN-Mamba feature extraction network and the scattering topological features obtained by the scattering topological feature extraction network are fused and input into the classification module to obtain the final target classification result.
[0011] Further, the process of the CNN-Mamba feature extraction network extracting the visual features of the SAR ship image includes the following steps:
[0012] S2.1 The input image X is first processed by a 4×4 convolutional layer with a stride of 4 to obtain a feature map F with a size of H×W and a channel dimension of C; F is evenly divided into F1 and F2 in the channel dimension and input into the residual convolution module and the SSM module respectively to extract local features and global features;
[0013] S2.2 The residual convolution module is composed of a 3×3 convolutional layer, a BatchNorm layer and a ReLU activation function; F1 passes through the convolutional layer, the BatchNorm layer and the ReLU activation layer in sequence and then is convolved again with the BatchNorm layer to obtain the feature F′1; F′1 is added element-wise to F1 and activated by ReLU to obtain the feature F c ;
[0014] S2.3 The SSM module consists of a layer normalization layer, a depthwise separable convolution layer, an SS2D module, a linear connection layer, and a SiLU activation function. Among them, SS2D is the core module in the Vision Mamba model for modeling long-range pixel dependencies in images. The feature F2 in step S2.1 first passes through the layer normalization layer to obtain the feature F'2, and then F'2 passes through a 3×3 depthwise separable convolution and a SiLU activation layer and is input into the SS2D module to obtain the feature F''2. At the same time, F'2 passes through the linear connection layer and the SiLU activation layer to obtain the feature F'''2. After F''2 and F'''2 are multiplied element by element and passed through the linear connection layer, the feature F is obtained. m ;
[0015] S2.4 Concatenate the feature F obtained in step S2.2 c and the feature F obtained in step S2.3 m in the channel dimension, and perform a channel shuffle operation to obtain the fused feature F'. a ;
[0016] S2.5 The downsampling module divides the feature map F' obtained in step S2.4 a into four submaps M0 - M3 along the H and W directions with a stride of 2, and the size of each submap is H / 2×W / 2×C. Concatenate the submaps M0 - M3 in the channel dimension to obtain the feature map M', and after passing through the layer normalization layer, compress the number of channels from 4C to 2C through an unbiased linear layer to obtain the first-stage feature F. a ;
[0017] S2.6 Input the feature F a into the next-stage Conv-SSM module, repeat the above operations, and finally obtain the visual feature F after four stages. vis .
[0018] Furthermore, the process of the scattering topology feature extraction network extracting the scattering topology features of the SAR ship image is as follows: First, extract the main scattering points of the SAR image based on the energy threshold and the clustering algorithm, then construct the scattering point topological structure and extract the topological features reflecting the ship's structural form through the graph convolution layer, and further enhance the discriminative ability of the scattering topology features by combining with the multi-layer perceptron.
[0019] Furthermore, the scattering topology feature extraction network extracting the scattering topology features of the SAR ship image includes the following steps:
[0020] S3.1 The scattering point extraction module first extracts key scattering points from the input SAR image X based on the energy threshold strategy; uses the K-Means clustering algorithm to screen and aggregate the scattering points, retains a fixed number N of scattering point nodes, and constructs a scattering point set.
[0021] S3.2 The topological feature construction module constructs a topological structure diagram based on the spatial information between scattering points to characterize the topological morphological features of the ship; first, calculate the Euclidean distance d(p i , p j ):
[0022]
[0023] For the scattering point p i , determine its K nearest scattering points based on the Euclidean distance, and construct an adjacency matrix where A(i, j) = 1 indicates that the scattering point p j is one of the K nearest neighbors of p i , otherwise A(i, j) = 0. This adjacency matrix is used to define the connectivity between nodes, thus establishing a preliminary graph structure; further use the Dijkstra shortest path algorithm to calculate the shortest path distance from each scattering point to other scattering points, and construct a distance matrix where D(i, j) represents the shortest path length from the scattering point p i to the scattering point p j , as the edge weight information in the topological structure diagram, quantifying the structural connection between scattering points from a topological perspective;
[0024] S3.3 Map the scattering point p i to the feature map F' a obtained in step 2.4, and extract the local perception feature from its corresponding position as the initial feature of the node in the graph structure ;
[0025] S3.4 Input the topological structure diagram constructed in step S3.2 and the initial features of the nodes in step S3.3 into the graph convolutional network together; after passing through the first graph convolutional layer and ReLU activation, and then passing through the second graph convolutional layer, output the feature F' spt , whose dimension is N×C, then perform layer normalization on F' spt , and input it into the multi-layer perceptron layer to enhance the feature interaction between nodes, and finally obtain the scattering topological feature F spt .
[0026] Furthermore, fuse the visual features obtained by the CNN-Mamba feature extraction network and the scattering topological features obtained by the scattering topological feature extraction network and input them into the classification module. The specific process is as follows: Pass the visual feature F vis obtained by the CNN-Mamba feature extraction network through the global average pooling layer, and then flatten it into a one-dimensional feature vector V vis ; Flatten the scattering topological feature F spt obtained by the scattering topological feature extraction network into a one-dimensional feature vector Vspt ; Concatenate the feature vectors V vis and V spt , and map them to an n-dimensional classification space through a fully connected layer, where n is the number of ship classes, and output the final classification result.
[0027] Furthermore, the total loss function in the training process of the double-branch network is composed of cross-entropy loss and island loss, expressed as:
[0028] L = L CE + λL IL
[0029] where: L CE is the cross-entropy loss, L IL is the island loss, and λ is the parameter for balancing the cross-entropy loss and the island loss.
[0030] Furthermore, the cross-entropy loss measures the difference between the predicted probability distribution and the true label distribution, defined as:
[0031]
[0032] where: y i is the class label of the i-th sample, z i is the prediction result of the model, and m is the number of samples in the batch.
[0033] Furthermore, the island loss is extended from the center loss, which can expand the inter-class distance and shrink the intra-class distance in the feature space. The expression is:
[0034]
[0035] where: N is the set of class labels, c j and c k are the center points of the j-th class and the k-th class respectively, and λ1 is the parameter; L c is the center loss, which can cluster the feature vectors of different samples of the same class towards the class center during training, defined as:
[0036]
[0037] where: y i is the class of the i-th sample, x i is the feature vector of the i-th sample obtained from the fully connected layer, represents the center of all samples with the same class label as y i , and m is the number of samples in the batch.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] The present invention provides a SAR ship target classification method based on a CNN-Mamba network and scattering topological features, aiming to solve the problems of insufficient utilization of the scattering features of targets in SAR images and limited feature extraction ability of a single CNN model in existing methods. The proposed method is based on deep learning and constructs a dual-branch network, mainly including a CNN-Mamba feature extraction network and a scattering topological feature extraction network. Among them, the CNN-Mamba feature extraction network combines a residual convolution module and a state space modeling (SSM) module to give full play to the advantages of the convolutional layer in local feature extraction and the ability of SSM in long-distance dependence modeling, thereby enhancing the expression ability of the significant features of ship targets. The scattering topological feature extraction network includes a scattering point extraction module and a topological feature construction module. First, the main scattering points of the SAR image are extracted based on an energy threshold and a clustering algorithm. Subsequently, the scattering point topological structure is constructed, and the topological features reflecting the ship structure form are extracted through a graph convolutional layer, and the discriminative ability of the scattering topological features is further enhanced by combining a multi-layer perceptron. Finally, the visual features obtained by the CNN-Mamba feature extraction network and the scattering topological features obtained by the scattering topological feature extraction network are fused and input into the classification module to obtain the final target classification result. The method of the present invention fully integrates the advantages of CNN and Mamba in feature extraction and the scattering topological structure information of SAR ship targets, significantly improving the accuracy and robustness of SAR ship target classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.
[0041] Figure 1 It is a flowchart of the SAR ship target classification method based on a CNN-Mamba network and scattering topological features provided by an embodiment of the present invention;
[0042] Figure 2 It is a network structure diagram of the SAR ship target classification method based on a CNN-Mamba network and scattering topological features provided by an embodiment of the present invention;
[0043] Figure 3 It is a structural schematic diagram of the residual convolution module of the SAR ship target classification method based on a CNN-Mamba network and scattering topological features provided by an embodiment of the present invention;
[0044] Figure 4Schematic diagram of the SSM module of the SAR ship target classification method based on the CNN-Mamba network and scattering topology features provided by the embodiments of the present invention;
[0045] Figure 5 Schematic diagram of the scattering points and topological structure of ship targets in a SAR image of the SAR ship target classification method based on the CNN-Mamba network and scattering topology features provided by the embodiments of the present invention. Detailed implementation manners
[0046] To better understand the technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.
[0047] As Figure 1 and Figure 2 shown, the present invention provides a SAR ship target classification method based on the CNN-Mamba network and scattering topology features, including the following steps:
[0048] S1 Select a SAR ship image dataset, randomly divide it into a training set and a test set according to a ratio of 7:3; uniformly adjust the image size to 224×224, and perform data augmentation operations on the SAR images during the training process, including random horizontal flipping.
[0049] S2 Construct a CNN-Mamba feature extraction network:
[0050] This network includes four stages. Each stage consists of several Conv-SSM modules. Each Conv-SSM module is composed of a residual convolution module and an SSM module in parallel. The stages are connected by downsampling modules to reduce the spatial size, increase the channel dimension, and enhance the feature expression ability. The specific steps are as follows:
[0051] S2.1 The input image X is first processed by a 4×4 convolutional layer with a stride of 4 to obtain a feature map F with a size of H×W and a channel dimension of C; F is evenly divided into F1 and F2 in the channel dimension, and is respectively input into the residual convolution module and the SSM module to extract local features and global features;
[0052] S2.2 As Figure 3 shown, the residual convolution module consists of a 3×3 convolutional layer, a BatchNorm (BN) layer, and a ReLU activation function; F1 passes through the convolutional layer, the BN layer, and the ReLU activation layer in sequence and then is convolved and BN again to obtain the feature F′1; F′1 and F1 are added element-wise and then passed through the ReLU activation to obtain the feature F c , and this residual connection can alleviate the degradation problem in deep networks;
[0053] S2.3 As Figure 4As shown in the figure, the SSM module consists of a layer normalization (LN) layer, a depthwise separable convolution layer, an SS2D module, a linear connection layer, and a SiLU activation function. Among them, SS2D is the core module in the Visual Mamba model for modeling long-range pixel dependencies in images. The feature F2 in step S2.1 first passes through the LN layer to obtain the feature F′2. Then, F′2 passes through a 3×3 depthwise separable convolution and a SiLU activation layer and is input into the SS2D module to obtain the feature F″2. At the same time, F′2 passes through a linear connection layer and a SiLU activation to obtain the feature F″′2. After F″2 and F″′2 are multiplied element by element and passed through a linear connection layer, the feature F is obtained. m ;
[0054] S2.4 Concatenate the feature F c obtained in step S2.2 and the feature F m obtained in step S2.3 in the channel dimension, and perform a channel shuffle operation to prevent information loss between channels, obtaining the fused feature F′ a ;
[0055] S2.5 The downsampling module divides the feature map F′ a obtained in step S2.4 into four submaps M0 - M3 along the H and W directions with a stride of 2, each with a size of H / 2×W / 2×C; concatenate the submaps M0 - M3 in the channel dimension to form the feature map M′, and after passing through the LN layer, compress the number of channels from 4C to 2C through an unbiased linear layer, obtaining the first-stage feature F a ;
[0056] S2.6 Input the feature F a into the next-stage Conv-SSM module, repeat the above operations, and finally obtain the visual feature F vis .
[0057] S3 Construct a scattering topological feature extraction network:
[0058] This network consists of a scattering point extraction module and a topological feature construction module. The specific steps are as follows:
[0059] S3.1 Figure 5 is a SAR ship image, the scattering point extraction result, and the scattering point topological structure; the scattering point extraction module first extracts key scattering points from the input SAR image X based on an energy threshold strategy. The distribution of key scattering points can effectively reflect the geometric structure information of the ship target; to reduce the influence of noise interference and redundant information, the K-Means clustering algorithm is further used to screen and aggregate the scattering points, retaining a fixed number N of scattering point nodes, thereby constructing a scattering point set with a clear structure and relatively balanced distribution.
[0060] S3.2 The topological feature construction module constructs a graph structure based on the spatial information between scattering points to characterize the topological morphological features of the ship; first, calculate the Euclidean distance d(p i , p j ) between each pair of scattering points:
[0061]
[0062] For the scattering point p i , determine its K nearest scattering points based on the Euclidean distance, and construct an adjacency matrix where A(i, j) = 1 indicates that the scattering point p j is one of the K nearest neighbors of p i , otherwise A(i, j) = 0. This adjacency matrix is used to define the connectivity between nodes, thus establishing a preliminary graph structure; further, use the Dijkstra shortest path algorithm to calculate the shortest path distance from each scattering point to other scattering points, and construct a distance matrix where D(i, j) represents the shortest path length from node p i to node p j , as the edge weight information in the topological structure diagram, quantifying the structural connection between scattering points from a topological perspective;
[0063] S3.3 Map the scattering point p i to the feature map F' a obtained in step S2.4, and extract local perception features from its corresponding position as the initial features of the nodes in the graph structure
[0064] S3.4 Input the topological structure diagram constructed in step S3.2 (characterized by the adjacency matrix A and the distance matrix D) and the initial features of the nodes in step S3.3 into the graph convolutional network together; the graph convolutional layer controls the message passing path through the adjacency matrix A, guiding the effective propagation and aggregation of features in the graph structure; at the same time, the distance matrix D provides a topological relationship metric, which helps to guide the network to better model the structural features of the target; after passing through the first graph convolutional layer and ReLU activation, and then passing through the second graph convolutional layer, output the feature F' spt , whose dimension is N×C, then perform layer normalization on F' spt , and input it into the multi-layer perceptron layer to enhance the feature interaction between nodes, and finally obtain the scattering topological feature F spt .
[0065] S4 Feature fusion and classification:
[0066] Pass the visual feature F vis obtained in step S2 through the global average pooling layer, and then flatten it into a one-dimensional feature vector V vis; Flatten the scattering topological feature F obtained in step S3 spt into a one-dimensional feature vector V spt ; For the feature vector V vis and V spt perform a concatenation operation and map it to an n-dimensional classification space through a fully connected layer, where n is the number of ship classes, and output the final classification result;
[0067] Training of the S5 network:
[0068] The total loss function consists of cross-entropy loss and island loss, which are described by the following formulas respectively:
[0069] The cross-entropy loss measures the difference between the predicted probability distribution and the true label distribution, and is defined as:
[0070]
[0071] where: y i is the class label of the i-th sample, and z i is the prediction result of the model;
[0072] The center loss can cluster the features of different samples of the same class towards the class center during training, and is defined as:
[0073]
[0074] where: y i is the class of the i-th sample, x i is the feature vector of the i-th sample obtained from the fully connected layer, represents the center of all samples with the same class label as y i , and m is the number of samples in the batch;
[0075] The island loss is extended from the center loss, and can expand the inter-class distance and shrink the intra-class distance in the feature space:
[0076]
[0077] where: N is the set of class labels, c j and c k are the center points of the j-th class and the k-th class respectively, and λ1 is a parameter;
[0078] The total loss function can be expressed as:
[0079] L = L CE + λL IL
[0080] where: λ is a parameter that balances the cross-entropy loss and the island loss.
[0081] The SAR ship target classification method based on the CNN-Mamba network and scattering topological features proposed by the present invention is based on deep learning and constructs a dual-branch network, which consists of a CNN-Mamba feature extraction network and a scattering topological feature extraction network. Among them, the CNN-Mamba feature extraction network combines a residual convolution module and a state space modeling (SSM) module to give full play to the advantages of the convolutional layer in local feature extraction and the ability of SSM in long-distance dependence modeling, thereby improving the expression ability of the significant features of ship targets. The scattering topological feature extraction network includes a scattering point extraction module and a topological feature construction module. First, the main scattering points of the SAR image are extracted based on the energy threshold and clustering algorithm, and then the scattering point topological structure is constructed and the topological features reflecting the ship structure form are extracted through the graph convolutional layer, and the discriminant ability of the scattering topological features is further enhanced by combining with a multi-layer perceptron. The method of the present invention fully integrates the advantages of CNN and Mamba in feature extraction and the scattering topological structure information of SAR ship targets, significantly improving the accuracy and robustness of SAR ship target classification.
[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or equivalently replace some of the technical features, but these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A SAR ship target classification method based on the CNN-Mamba network and scattering topological features, characterized in that, The method constructs a dual-branch network, including a CNN-Mamba feature extraction network and a scattering topology feature extraction network; The CNN-Mamba feature extraction network includes four stages, each stage consists of several Conv-SSM modules, each Conv-SSM module is composed of a residual convolution module and an SSM module in parallel, and the stages are connected by downsampling modules to extract visual features of SAR ship images; The scattering topology feature extraction network includes a scattering point extraction module and a topology feature construction module for extracting scattering topology features of SAR ship images; The visual features obtained by the CNN-Mamba feature extraction network and the scattering topology features obtained by the scattering topology feature extraction network are fused and input into the classification module to obtain the final target classification result.
2. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 1, wherein The process of the CNN-Mamba feature extraction network extracting visual features of SAR ship images includes the following steps: S2.1 The input image X is first processed by a 4×4 convolutional layer with a stride of 4 to obtain a feature map F with a size of H×W and a channel dimension of C; F is evenly divided into F1 and F2 in the channel dimension and input into the residual convolution module and the SSM module respectively to extract local features and global features; S2.2 The residual convolution module consists of a 3×3 convolutional layer, a BatchNorm layer, and a ReLU activation function; F1 passes through a convolutional layer, a BatchNorm layer, and a ReLU activation layer in sequence, and then undergoes another convolution and BatchNorm layer to obtain the feature F1 ′ ; Add F1 ′ element-wise to F1 and activate through ReLU to obtain the feature F c ; S2.3 The SSM module consists of a layer normalization layer, a depthwise separable convolution layer, an SS2D module, a linear connection layer, and a SiLU activation function. The feature F2 in step S2.1 is first passed through the layer normalization layer to obtain the feature F'2. Then, F'2 is input into the SS2D module after passing through a 3×3 depthwise separable convolution and a SiLU activation layer, resulting in the feature F''2. At the same time, F'2 passes through the linear connection layer and the SiLU activation layer to obtain the feature F'''2. After multiplying F''2 and F'''2 element-wise, it passes through the linear connection layer to obtain the feature F m ; S2.4 Concatenate the feature F obtained in step S2.2 c and the feature F obtained in step S2.3 m along the channel dimension, and perform a channel shuffle operation to obtain the fused feature F' a ; The S2.5 downsampling module divides the feature map F′ obtained in step S2.4 a into four sub-maps M0 to M3 along the H and W directions with a stride of 2, and the size of each sub-map is H / 2 × W / 2 × C; the sub-maps M0 to M3 are concatenated in the channel dimension to obtain the feature map M ′ , and after passing through the layer normalization layer, the number of channels is compressed from 4C to 2C through a bias-free linear layer to obtain the first-stage feature F a ; S2.6 Input the feature F a into the next-stage Conv-SSM module, repeat the above operations, and finally obtain the visual feature F after four stages vis .
3. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 2, wherein The process of the scattering topology feature extraction network extracting scattering topology features of SAR ship images is as follows: First, the main scattering points of the SAR image are extracted based on the energy threshold and the clustering algorithm, then the scattering point topological structure is constructed and the topological features reflecting the ship structure form are extracted through the graph convolutional layer, and the discriminative ability of the scattering topology features is further enhanced by combining the multi-layer perceptron.
4. The SAR ship target classification method based on the CNN-Mamba network and scattering topology features according to claim 3, wherein The process of the scattering topology feature extraction network extracting scattering topology features of SAR ship images includes the following steps: S3.1 The scattering point extraction module first extracts key scattering points from the input SAR image X based on the energy threshold strategy; the K-Means clustering algorithm is used to screen and aggregate the scattering points, and a fixed number N of scattering point nodes are retained to construct a scattering point set; S3.2 The topological feature construction module constructs a topological structure diagram based on the spatial information between scattering points to characterize the topological morphological features of the ship; first, calculate the Euclidean distance d(p i , p j ) between each pair of scattering points: For the scattering point p i , determine its K nearest neighbor scattering points based on the Euclidean distance, and construct an adjacency matrix accordingly where A(i,j)=1 indicates that the scattering point p j is one of the K nearest neighbor points of p i , otherwise A(i,j)=0. This adjacency matrix is used to define the connectivity between nodes, thus establishing a preliminary graph structure; further use the Dijkstra shortest path algorithm to calculate the shortest path distance from each scattering point to other scattering points, and construct a distance matrix where D(i,j) represents the shortest path length from the scattering point p i to the scattering point p j , as the edge weight information in the topological structure diagram, to quantitatively describe the structural relationship between scattering points from a topological perspective; S3.3 Map the scattering point p i to the feature map F' obtained in step 2.4 a and extract the local perception feature from its corresponding position as the initial feature of the node in the graph structure S3.4 Input the topological structure diagram constructed in step S3.2 and the initial node features in step S3.3 into the graph convolutional network together; after the first graph convolutional layer and ReLU activation, and then through the second graph convolutional layer, output the feature F' spt , whose dimension is N×C, and then perform layer normalization on F' spt and input it into the multi-layer perceptron layer to enhance the feature interaction between nodes, and finally obtain the scattering topological feature F spt .
5. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 4, characterized in that, The visual features obtained by the CNN-Mamba feature extraction network and the scattering topological features obtained by the scattering topological feature extraction network are fused and input into the classification module. The specific process is as follows: The visual features F obtained by the CNN-Mamba feature extraction network vis pass through the global average pooling layer and are then flattened into a one-dimensional feature vector V vis ; The scattering topological features F obtained by the scattering topological feature extraction network spt are flattened into a one-dimensional feature vector V spt ; The feature vectors V vis and V spt are concatenated, and are mapped to an n-dimensional classification space through a fully connected layer, where n is the number of ship classes, and the final classification result is output.
6. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 1, characterized in that, During the training process of the dual-branch network, the total loss function is composed of the cross-entropy loss and the island loss, which is expressed as: L = L CE + λL IL Where: L CE is the cross-entropy loss, L IL is the island loss, and λ is a parameter that balances the cross-entropy loss and the island loss.
7. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 6, wherein The cross-entropy loss measures the difference between the predicted probability distribution and the true label distribution, and is defined as: where: y i is the class label of the i-th sample, z i is the prediction result of the model, and m is the number of samples in the batch.
8. The SAR ship target classification method based on the CNN-Mamba network and scattering topological features according to claim 6, wherein, The island loss is extended from the center loss, which can expand the inter-class distance and shrink the intra-class distance in the feature space, and the expression is: Where: N is the set of class labels, c j and c k are the center points of the j-th class and the k-th class respectively, and λ1 is a parameter; L c is the center loss, and the center loss can cluster the feature vectors of different samples of the same class towards the class center during training, and is defined as: where: y i is the class of the i-th sample, x i is the feature vector of the i-th sample obtained from the fully connected layer, represents the center of all samples with the same class label as y i and m is the number of samples in the batch.