A SAR ship target recognition method and system based on component information enhancement

Abstract

This invention belongs to the field of SAR image target recognition technology, and discloses a method and system for SAR ship target recognition based on component information enhancement. The method includes the following steps: acquiring a SAR ship image to be identified; extracting key points from the SAR ship image to obtain key points of the ship image; performing K-means clustering on the key points of the ship image to obtain bow and stern component information of the ship target, and generating a component Gaussian heatmap; constructing a component information enhancement network, and inputting the component Gaussian heatmap into the component information enhancement network for recognition, to obtain the ship recognition result. This invention extracts and clusters key points of the ship in SAR images, constructs a component information enhancement classification network PIANet, uses component information for driving guidance, and improves the network's learning ability for similarity features, thereby improving the recognition effect.

Description

Technical Field

[0001] This invention belongs to the field of SAR image target recognition technology, specifically relating to a SAR ship target recognition method and system based on component information enhancement. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses all-weather, all-day imaging capabilities, making it invaluable in maritime traffic control, fisheries management, and maritime emergency rescue. SAR vessel identification is one of the most practical applications of SAR automatic target identification, and it forms the foundation and is a key research area for maritime surveillance tasks such as integrated maritime decision-making and maritime traffic planning.

[0003] Due to the unique imaging mechanism of SAR, images contain numerous speckle noise points, making it difficult to distinguish targets from the background and noise. Therefore, SAR images are more challenging to process than optical remote sensing images. To achieve better ship recognition performance, the following challenging issues exist: SAR ship targets have complex textures, shapes, and structures, and defocusing and cross-bright spot phenomena may exist, potentially leading to poor sample image quality. Furthermore, differences in resolution and imaging angle can produce significant feature differences in SAR ship target slices. Additionally, SAR ship data exhibits significant intra-class diversity and inter-class similarity, requiring recognition algorithms to acquire more fine-grained features to achieve higher recognition performance.

[0004] To address these challenges, numerous recognition algorithms targeting the mechanisms of SAR images have been proposed in recent years. Component information contains the physical features of ship targets, accurately representing their structural characteristics and functional attributes. Constructing a recognition framework based on the synergistic fusion of ship component physical features and deep learning representation capabilities is a key path to improving the accuracy and interpretability of SAR ship target recognition. This approach enhances the model's ability to analyze the physical structure of ship targets through component-level features and leverages deep learning to uncover complex inter-component relationships. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0006] A SAR ship target identification method based on component information enhancement includes the following steps:

[0007] Acquire a SAR ship image to be identified, and extract key points from the SAR ship image to obtain the key points of the ship image;

[0008] K-means clustering is performed on the key points of the ship image to obtain the bow and stern component information of the ship target, and a Gaussian heat map of the components is generated.

[0009] A component information enhancement network is constructed, and the Gaussian heat map of the component is input into the component information enhancement network for identification to obtain the ship identification result.

[0010] Preferably, the method for obtaining the key points of the ship image includes:

[0011] Calculate the gradient value I of pixel (x,y) in the horizontal direction in the SAR ship image to be identified. x and the gradient value I in the vertical direction y ;

[0012] Based on the gradient value I in the horizontal direction x and the gradient value I in the vertical direction y We use a Gaussian function to perform weighted calculations to obtain the covariance matrix:

[0013] ,

[0014] Where ω(x, y) represents the Gaussian function value at the pixel (x, y) to be calculated, I x I x I represents the product in the horizontal direction. y I y I represents the product in the vertical direction. x I y Represents the product of the horizontal and vertical directions;

[0015] Traverse the image matrix, calculate the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculate the corner response function R:

[0016] ,

[0017] Where λ1 and λ2 represent two eigenvalues ​​of the covariance matrix, and k represents a constant;

[0018] A threshold is set for the function. If the corner response value obtained by the corner response function R is greater than the threshold and is a local maximum value, then the corresponding pixel is marked as a key point of the ship image.

[0019] Preferably, the method for generating the Gaussian heat map of the component includes:

[0020] Determine the number of clusters K, and select K key points from the ship image as initial cluster centers;

[0021] Calculate the Euclidean distance from each of the ship image keypoints to the cluster center, and assign the corresponding ship image keypoint to the category with the smallest distance;

[0022] Calculate the mean of the key points of the ship images in the same category as the new cluster centers;

[0023] Repeat the operation until the new cluster centers are consistent with the original cluster centers to complete the K-means clustering and obtain the bow component information and stern component information of the ship target;

[0024] For the bow component information and the stern component information, generate the Gaussian heat map of the components:

[0025] ,

[0026] Where x* represents the true x-coordinate of the cluster center, y* represents the true y-coordinate of the cluster center, and σ represents the adaptive Gaussian kernel standard deviation.

[0027] Preferably, the component information enhancement network includes: cosine similarity joint attention (CSCA) and three stacked component information attention enhancement modules (PIAM);

[0028] The cosine similarity joint attention (CSCA) is as follows:

[0029] ,

[0030] ,

[0031] ,

[0032] Where N represents the number of channels, Q i and K i Let CS(Q) represent the feature vector of the i-th channel. i ,K i ) represents the input feature channel similarity based on the cosine similarity function, Q mn and K mn The eigenvector representing position (n, m), SS(Q) mn ,K mn () represents spatial similarity based on the cosine similarity function, W represents the width of the feature map, H represents the height of the feature map, and V represents the features aggregated by attention weighting. This indicates element-wise multiplication;

[0033] The component information attention enhancement module PIAM performs cosine similarity joint attention (CSCA) calculations on the key vectors K1 and K2 obtained by linearly transforming the component weighted features P1 and P2, respectively, and then aggregates them with the query vectors Q1 and Q2 of the original image feature F to obtain the first feature:

[0034] ,

[0035] Then, the calculated CSCA value is aggregated with the bow component enhancement feature F1 and the stern component enhancement feature F2, respectively, to obtain the second feature:

[0036] ,

[0037] Finally, the first feature and the second feature are concatenated and fused to obtain the output of the Component Information Attention Enhancement Module (PIAM).

[0038] The present invention also provides a SAR ship target recognition system based on component information enhancement. The recognition system applies the above-mentioned recognition method and includes: a key point extraction module, an information acquisition module, and a recognition module.

[0039] The key point extraction module is used to acquire the SAR ship image to be identified, extract key points from the SAR ship image to obtain the key points of the ship image.

[0040] The information acquisition module is used to perform K-means clustering on the key points of the ship image to obtain the bow and stern component information of the ship target, and generate a Gaussian heat map of the components.

[0041] The identification module is used to construct a component information enhancement network and input the Gaussian heat map of the component into the component information enhancement network for identification, thereby obtaining the ship identification result.

[0042] Preferably, the workflow of the key point extraction module includes:

[0043] Calculate the gradient value I of pixel (x,y) in the horizontal direction in the SAR ship image to be identified. x and the gradient value I in the vertical direction y ;

[0044] Based on the gradient value I in the horizontal direction x and the gradient value I in the vertical direction y We use a Gaussian function to perform weighted calculations to obtain the covariance matrix:

[0045] ,

[0046] Where ω(x, y) represents the Gaussian function value at the pixel (x, y) to be calculated, I x I x I represents the product in the horizontal direction. y I y I represents the product in the vertical direction. x I y Represents the product of the horizontal and vertical directions;

[0047] Traverse the image matrix, calculate the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculate the corner response function R:

[0048] ,

[0049] Where λ1 and λ2 represent two eigenvalues ​​of the covariance matrix, and k represents a constant;

[0050] A threshold is set for the function. If the corner response value obtained by the corner response function R is greater than the threshold and is a local maximum value, then the corresponding pixel is marked as a key point of the ship image.

[0051] Preferably, the workflow of the information acquisition module includes:

[0052] Determine the number of clusters K, and select K key points from the ship image as initial cluster centers;

[0053] Calculate the Euclidean distance from each of the ship image keypoints to the cluster center, and assign the corresponding ship image keypoint to the category with the smallest distance;

[0054] Calculate the mean of the key points of the ship images in the same category as the new cluster centers;

[0055] Repeat the operation until the new cluster centers are consistent with the original cluster centers to complete the K-means clustering and obtain the bow component information and stern component information of the ship target;

[0056] For the bow component information and the stern component information, generate the Gaussian heat map of the components:

[0057] ,

[0058] Where x* represents the true x-coordinate of the cluster center, y* represents the true y-coordinate of the cluster center, and σ represents the adaptive Gaussian kernel standard deviation.

[0059] Preferably, the component information enhancement network includes: cosine similarity joint attention (CSCA) and three stacked component information attention enhancement modules (PIAM);

[0060] The cosine similarity joint attention (CSCA) is as follows:

[0061] ,

[0062] ,

[0063] ,

[0064] Where N represents the number of channels, Q i and K i Let CS(Q) represent the feature vector of the i-th channel. i ,K i ) represents the input feature channel similarity based on the cosine similarity function, Q mn and K mn The eigenvector representing position (n, m), SS(Q) mn ,K mn () represents spatial similarity based on the cosine similarity function, W represents the width of the feature map, H represents the height of the feature map, and V represents the features aggregated by attention weighting. This indicates element-wise multiplication;

[0065] The component information attention enhancement module PIAM performs cosine similarity joint attention (CSCA) calculations on the key vectors K1 and K2 obtained by linearly transforming the component weighted features P1 and P2, respectively, and then aggregates them with the query vectors Q1 and Q2 of the original image feature F to obtain the first feature:

[0066] ,

[0067] Then, the calculated CSCA value is aggregated with the bow component enhancement feature F1 and the stern component enhancement feature F2, respectively, to obtain the second feature:

[0068] ,

[0069] Finally, the first feature and the second feature are concatenated and fused to obtain the output of the Component Information Attention Enhancement Module (PIAM).

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] This invention extracts key points of ships from SAR images and clusters them to construct a component information enhancement classification network (PIANet). It utilizes component information for driving and guiding the network, and improves the network's ability to learn similarity features, thereby enhancing recognition performance. A novel component information attention enhancement module (PIAM) is designed to better integrate component information with the original image features. The cosine similarity joint attention calculation method can better acquire channel and spatial similarity features between features, enabling the network to learn more similar features of the same type. An SPD-conv convolution module replaces the ordinary convolution downsampling operation, allowing component information to adapt to the image size and channels within the network, preventing information loss due to heatmap blurring after convolution. Attached Figure Description

[0072] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0074] 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, and 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.

[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] Example 1:

[0077] In this embodiment, as Figure 1 As shown, a SAR ship target identification method based on component information enhancement includes the following steps:

[0078] S1. Obtain the SAR ship image to be identified, extract key points from the SAR ship image to obtain the key points of the ship image.

[0079] In this embodiment, the method for obtaining key points of a ship image includes: the method for obtaining key point features of a ship is the Shi-Tomasi corner detection algorithm, specifically implemented as follows:

[0080] Calculate the gradient value I of pixel (x,y) in the horizontal direction in the SAR ship image to be identified. x and the gradient value I in the vertical direction y That is, the first-order partial derivatives of the gray value at that point in the x and y directions; based on the gradient value I in the horizontal direction. x and the gradient value I in the vertical direction y Calculate the product I in the horizontal and vertical directions respectively. x I x I y I y and I x I y Using the Gaussian function to apply I respectively x I x I y I y and I x I y We perform a weighted calculation to obtain the covariance matrix:

[0081] ,

[0082] Where ω(x, y) represents the Gaussian function value (with a variance of 1) at the pixel (x, y) to be calculated, I x I x I represents the product in the horizontal direction. y I y I represents the product in the vertical direction. x I y This represents the product of the horizontal and vertical directions; iterates through the image matrix, calculates the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculates the corner response function R:

[0083] ,

[0084] Where λ1 and λ2 represent two eigenvalues ​​of the covariance matrix, and k represents a constant between 0.04 and 0.06; a function threshold is set, and if the corner response value obtained by the corner response function R is greater than the threshold and is a local maximum, then the corresponding pixel is marked as a key point of the ship image.

[0085] S2. Perform K-means clustering on the key points of the ship image to obtain the bow and stern component information of the ship target, and generate Gaussian heatmaps of the components.

[0086] In this embodiment, the method for generating a Gaussian heatmap of a component includes:

[0087] Determine the number of clusters K=2, and select K ship image key points as initial cluster centers; calculate the Euclidean distance from each ship image key point to the cluster center, and assign the corresponding ship image key points to the category with the smallest distance; calculate the mean of ship image key points in the same category as the new cluster center; repeat the operation until the new cluster centers are consistent with the original cluster centers, or the maximum number of iterations is reached, to complete K-means clustering and obtain the bow and stern component information of the ship target; for the bow and stern component information, generate Gaussian heatmaps of the components:

[0088] ,

[0089] Where x* represents the true x-coordinate of the cluster center, y* represents the true y-coordinate of the cluster center, and σ represents the adaptive Gaussian kernel standard deviation.

[0090] S3. Construct a component information enhancement network and input the component Gaussian heat map into the component information enhancement network for identification to obtain the ship identification result.

[0091] The component information enhancement network includes: Cosine Similarity Joint Attention (CSCA) and three stacked Component Information Attention Enhancement (PIAM) modules. The SAR ship image to be identified, along with component heatmaps of the bow and stern, undergoes feature fusion through the three PIAM modules. The output of the third PIAM module is then passed through a fully connected layer and subjected to a Softmax operation to generate the recognition result.

[0092] Cosine similarity joint attention (CSCA) is:

[0093] ,

[0094] ,

[0095] ,

[0096] Where N represents the number of channels, Q i and K i Let CS(Q) represent the feature vector of the i-th channel. i ,K i ) represents the input feature channel similarity based on the cosine similarity function, Q mn and K mn The eigenvector representing position (n, m), SS(Q) mn ,K mn ) represents spatial similarity based on the cosine similarity function, W represents the width of the feature map, H represents the height of the feature map, and V represents the features aggregated by attention weighting.

[0097] In this embodiment, the SAR ship image to be identified and the thermal images of the bow and stern components are used as network inputs. After the SAR ship image to be identified is processed by the convolutional layer for feature extraction, the thermal images of the bow and stern components are converted into attention weights, which are then used as attention masks and applied to the feature maps of the SAR ship image to be identified.

[0098] First, the component heatmaps are normalized, and then the original image features of the SAR ship image to be identified are extracted through a network. Then the heatmap is used as an attention mask. Expanded to feature map channels via 1×1 convolution. And weighted feature maps respectively:

[0099] ,

[0100] ,

[0101] ,

[0102] ,

[0103] in, , A(1) and A(2) represent the attention mask generated from the Gaussian heatmap of the component, respectively. A(1) and A(2) represent the Gaussian heatmap of the component, and Sigmoid represents the activation function. F p1 F p2 This represents the attention-weighted feature map, where C represents the number of channels. To avoid over-suppressing non-critical regions and preserve the original feature information, residual connections are performed on the output features.

[0104] ,

[0105] ,

[0106] in, The expression represents element-wise multiplication. F1 represents the feature map output after attention weighting for the bow component, and F2 represents the feature map output after attention weighting for the stern component. The original image features F, the bow component enhanced features F1, and the stern component enhanced features F2 are used as inputs to the Component Information Attention Enhancement (PIAM) module, fusing features from the three branches. Query vectors Q1 and Q2 and value vectors F1 and F2 are generated from the original image features F, and key vectors K1 and K2 and value vectors V1 and V2 are generated from the component weighted features P1 and P2. The query vectors and key vectors are used to calculate similarity, while the value vectors are used as the vectors to be weighted and aggregated by the similarity scores. This gives each vector its own trainable weights, improving the network's fitting ability.

[0107] The Component Information Attention Enhancement (PIAM) module performs cosine similarity joint attention (CSCA) calculations on the key vectors K1 and K2 obtained by linearly transforming the weighted features P1 and P2 of the components, and then aggregates them with the query vectors Q1 and Q2 of the original image feature F to obtain the first feature. This effectively models the channel relationship similarity and spatial similarity between the two components and the target, and obtains the saliency information of different components of the same class, thereby enhancing the network's ability to learn intra-class similarity.

[0108] ,

[0109] Then, the calculated CSCA values ​​are aggregated with the bow component enhancement feature F1 and the stern component enhancement feature F2, respectively, to obtain the second feature. This enhances the original image network with the component information of the target, making the network pay more attention to the bow and stern components in terms of channels and space, enhancing the network's fine-grained feature learning ability, and thus improving the network's recognition performance.

[0110] ,

[0111] Finally, the first and second features are concatenated and fused to obtain the output of the Component Information Attention Enhancement (PIAM) module.

[0112] In this embodiment, the recognition accuracy on the OpenSARShip test set is shown in Table 1.

[0113] Table 1

[0114] .

[0115] Example 2:

[0116] In this embodiment, a SAR ship target recognition system based on component information enhancement includes: a key point extraction module, an information acquisition module, and a recognition module.

[0117] The key point extraction module is used to acquire SAR ship images to be identified, extract key points from the SAR ship images to obtain key points of the ship images.

[0118] The workflow of the key point extraction module includes: calculating the gradient value I of each pixel (x, y) in the horizontal direction in the SAR ship image to be identified. x and the gradient value I in the vertical direction y Based on the gradient value I in the horizontal direction x and the gradient value I in the vertical direction y We use a Gaussian function to perform weighted calculations to obtain the covariance matrix:

[0119] ,

[0120] Where ω(x, y) represents the Gaussian function value at the pixel (x, y) to be calculated, I x I x I represents the product in the horizontal direction. y I y I represents the product in the vertical direction. x I y This represents the product of the horizontal and vertical directions; iterates through the image matrix, calculates the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculates the corner response function R:

[0121] ,

[0122] Where λ1 and λ2 represent two eigenvalues ​​of the covariance matrix, and k represents a constant; a threshold is set for the function, and if the corner response value obtained by the corner response function R is greater than the threshold and is a local maximum, then the corresponding pixel is marked as a key point of the ship image.

[0123] The information acquisition module is used to perform K-means clustering on key points of ship images to obtain information on the bow and stern components of the ship target, and generate Gaussian heatmaps of the components.

[0124] The workflow of the information acquisition module includes: determining the number of clusters K, selecting K ship image key points as initial cluster centers; calculating the Euclidean distance from each ship image key point to the cluster center, and assigning the corresponding ship image key points to the category with the smallest distance; calculating the mean of ship image key points in the same category as the new cluster center; repeating the operation until the new cluster centers are consistent with the original cluster centers, completing K-means clustering, and obtaining the bow and stern component information of the ship target; for the bow and stern component information, generating Gaussian heatmaps of the components.

[0125] ,

[0126] Where x* represents the true x-coordinate of the cluster center, y* represents the true y-coordinate of the cluster center, and σ represents the adaptive Gaussian kernel standard deviation.

[0127] The identification module is used to construct a component information enhancement network and input the component Gaussian heat map into the component information enhancement network for identification, thereby obtaining the ship identification result.

[0128] The component information enhancement network includes: cosine similarity joint attention (CSCA) and three stacked component information attention enhancement modules (PIAM).

[0129] Cosine similarity joint attention (CSCA) is:

[0130] ,

[0131] ,

[0132] ,

[0133] Where N represents the number of channels, Q i and K i Let CS(Q) represent the feature vector of the i-th channel. i ,K i ) represents the input feature channel similarity based on the cosine similarity function, Q mn and K mn The eigenvector representing position (n, m), SS(Q) mn ,K mn () represents spatial similarity based on the cosine similarity function, W represents the width of the feature map, H represents the height of the feature map, and V represents the features aggregated by attention weighting. This indicates element-wise multiplication;

[0134] The Component Information Attention Enhancement (PIAM) module performs cosine similarity joint attention (CSCA) calculations on the key vectors K1 and K2 obtained by linearly transforming the component weighted features P1 and P2, respectively, and then aggregates them with the query vectors Q1 and Q2 of the main feature F to obtain the first feature.

[0135] ,

[0136] Then, the calculated CSCA value is aggregated with the bow component enhancement feature F1 and the stern component enhancement feature F2, respectively, to obtain the second feature:

[0137] ,

[0138] Finally, the first and second features are concatenated and fused to obtain the output of the Component Information Attention Enhancement (PIAM) module.

[0139] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A SAR ship target identification method based on component information enhancement, characterized in that, Includes the following steps: Acquire a SAR ship image to be identified, and extract key points from the SAR ship image to obtain the key points of the ship image; K-means clustering is performed on the key points of the ship image to obtain the bow and stern component information of the ship target, and a Gaussian heat map of the components is generated. A component information enhancement network is constructed, and the Gaussian heat map of the component is input into the component information enhancement network for identification to obtain the ship identification result. The component information enhancement network includes: cosine similarity joint attention. CSCA And three stacked Component Information Attention Enhancement (PIAM) modules; The cosine similarity joint attention CSCA for: ， ， ， in, N Indicates the number of channels. Q i and K i Indicates the first i Feature vectors of each channel CS ( Q i , K i This represents the input feature channel similarity based on the cosine similarity function. Q mn and K mn Indicates position ( n , m ) eigenvectors, SS ( Q mn , K mn ) represents spatial similarity based on the cosine similarity function. W Indicates the width of the feature map. H Indicates the height of the feature map, V This represents the features that are aggregated using attention weighting. This indicates element-wise multiplication; The Component Information Attention Enhancement (PIAM) module uses component-weighted features. P 1. P 2. The key vector obtained by linear transformation K 1. K 2. Features of the original image respectively F query vector Q 1. Q 2. Perform cosine similarity joint attention CSCA Calculate and compare with the value vector V 1. V 2. Perform aggregation to obtain the first feature: ， Then calculate CSCA The values ​​are respectively related to the enhancement features of the bow component. F 1 and stern component reinforcement features F 2. Aggregate the features to obtain the second feature: ， Finally, the first feature and the second feature are concatenated and fused to obtain the output of the Component Information Attention Enhancement Module (PIAM).

2. The SAR ship target identification method based on component information enhancement according to claim 1, characterized in that, The method for obtaining the key points of the ship image includes: Calculate the number of pixels in the SAR ship image to be identified respectively. x , y gradient value in the horizontal direction I x and gradient values ​​in the vertical direction I y ; Based on the gradient value in the horizontal direction I x and the gradient value in the vertical direction I y We use a Gaussian function to perform weighted calculations to obtain the covariance matrix: ， in, ω ( x , y ) indicates the pixel point to be calculated ( x , y The Gaussian function value was obtained. I x I x Represents the product in the horizontal direction. I y I y This represents the product in the vertical direction. I x I y Represents the product of the horizontal and vertical directions; Traverse the image matrix, calculate the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculate the corner response function. R : ， in, λ 1. λ 2 represents the two eigenvalues ​​of the covariance matrix. k Represents a constant; Set a function threshold; if the corner response function... R If the obtained corner response value is greater than the threshold and is a local maximum, then the corresponding pixel is marked as a key point of the ship image.

3. The SAR ship target identification method based on component information enhancement according to claim 2, characterized in that, The method for generating the Gaussian heatmap of the component includes: Determine the number of clusters K, and select K key points from the ship image as initial cluster centers; Calculate the Euclidean distance from each of the ship image keypoints to the cluster center, and assign the corresponding ship image keypoint to the category with the smallest distance; Calculate the mean of the key points of the ship images in the same category as the new cluster centers; Repeat the operation until the new cluster centers are consistent with the original cluster centers to complete the K-means clustering and obtain the bow component information and stern component information of the ship target; For the bow component information and the stern component information, generate the Gaussian heat map of the components: ， in, x * Represents the true x-coordinate of the cluster center. y * Represents the actual ordinate of the cluster center. σ This represents the adaptive Gaussian kernel standard deviation.

4. A SAR ship target identification system based on component information enhancement, wherein the identification system applies the identification method according to any one of claims 1-3, characterized in that, include: Key point extraction module, information acquisition module, and recognition module; The key point extraction module is used to acquire the SAR ship image to be identified, extract key points from the SAR ship image to obtain the key points of the ship image. The information acquisition module is used to perform K-means clustering on the key points of the ship image to obtain the bow and stern component information of the ship target, and generate a Gaussian heat map of the components. The identification module is used to construct a component information enhancement network and input the Gaussian heat map of the component into the component information enhancement network for identification, thereby obtaining the ship identification result. The component information enhancement network includes: cosine similarity joint attention. CSCA And three stacked Component Information Attention Enhancement (PIAM) modules; The cosine similarity joint attention CSCA for: ， ， ， in, N Indicates the number of channels. Q i and K i Indicates the first i Feature vectors of each channel CS ( Q i , K i This represents the input feature channel similarity based on the cosine similarity function. Q mn and K mn Indicates position ( n , m ) eigenvectors, SS ( Q mn , K mn ) represents spatial similarity based on the cosine similarity function. W Indicates the width of the feature map. H Indicates the height of the feature map, V This represents the features that are aggregated using attention weighting. This indicates element-wise multiplication; The Component Information Attention Enhancement (PIAM) module uses component-weighted features. P 1. P 2. The key vector obtained by linear transformation K 1. K 2. Features of the original image respectively F query vector Q 1. Q 2. Perform cosine similarity joint attention (CSCA) calculation and combine it with the value vector. V 1. V 2. Perform aggregation to obtain the first feature: ， Then the calculated CSCA values ​​are compared with the enhancement features of the bow components. F 1 and stern component reinforcement features F 2. Aggregate the features to obtain the second feature: ， Finally, the first feature and the second feature are concatenated and fused to obtain the output of the Component Information Attention Enhancement Module (PIAM).

5. The SAR ship target identification system based on component information enhancement according to claim 4, characterized in that, The workflow of the key point extraction module includes: Calculate the number of pixels in the SAR ship image to be identified respectively. x , y gradient value in the horizontal direction I x and gradient values ​​in the vertical direction I y ; Based on the gradient value in the horizontal direction I x and the gradient value in the vertical direction I y We use a Gaussian function to perform weighted calculations to obtain the covariance matrix: ， in, ω ( x , y ) indicates the pixel point to be calculated ( x , y The Gaussian function value was obtained. I x I x Represents the product in the horizontal direction. I y I y This represents the product in the vertical direction. I x I y Represents the product of the horizontal and vertical directions; Traverse the image matrix, calculate the two eigenvalues ​​of the covariance matrix corresponding to each pixel, and calculate the corner response function. R : ， in, λ 1. λ 2 represents the two eigenvalues ​​of the covariance matrix. k Represents a constant; Set a function threshold; if the corner response function... R If the obtained corner response value is greater than the threshold and is a local maximum, then the corresponding pixel is marked as a key point of the ship image.

6. The SAR ship target identification system based on component information enhancement according to claim 5, characterized in that, The workflow of the information acquisition module includes: Determine the number of clusters K, and select K key points from the ship image as initial cluster centers; Calculate the Euclidean distance from each of the ship image keypoints to the cluster center, and assign the corresponding ship image keypoint to the category with the smallest distance; Calculate the mean of the key points of the ship images in the same category as the new cluster centers; Repeat the operation until the new cluster centers are consistent with the original cluster centers to complete the K-means clustering and obtain the bow component information and stern component information of the ship target; For the bow component information and the stern component information, generate the Gaussian heat map of the components: ， in, x * Represents the true x-coordinate of the cluster center. y * Represents the actual ordinate of the cluster center. σ This represents the adaptive Gaussian kernel standard deviation.

Images