Ship target recognition device based on multi-scale map attention

Abstract

The invention discloses a ship target recognition device based on multi-scale map attention. The ship target recognition device comprises a multi-band SAR radar, a database and an upper computer. The multi-band SAR radar, the database and the upper computer are sequentially connected, the multi-band SAR radar scans a monitored sea area and stores SAR images of the SAR radar into the database, and the upper computer comprises a data preprocessing module, a multi-scale image data reconstruction module, a feature vector aggregation module and a target recognition module. According to the multi-scale SAR radar marine ship target recognition device and method based on graph structure data reconstruction and the graph attention network, the characteristics of multiple frequency bands and multiple angles are fused, the advantages of high precision and high intelligence are achieved, and the accuracy and efficiency of marine ship target recognition are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of SAR radar data processing, and in particular, to a ship target recognition device based on multi-scale graph attention. Background Art

[0002] SAR radar plays a key role in the all-weather, all-sea area monitoring and recognition of ship targets at sea, and is a core technology for various countries in the fields of marine security, maritime search and rescue, sea area monitoring, situation assessment, etc. This technology has important strategic value for safeguarding China's marine rights and interests and protecting marine security, and has become one of the key links in the informatization construction of China's marine field.

[0003] SAR radar realizes the discovery and spatial positioning of targets by detecting the reflection of electromagnetic waves by targets. Specifically, the SAR radar emits high-power electromagnetic pulse signals and radiates them to the monitored sea area through an antenna. When the electromagnetic wave encounters a target object, part of the signal is reflected back to the receiving device, thereby generating a high-resolution SAR image. By analyzing these images, the recognition and tracking of sea targets can be achieved. However, there are significant differences in the performance and application scenarios of SAR radars with different frequency bands: high-frequency SAR radars have higher resolution and can identify the detailed features of targets, and are suitable for fine recognition tasks. However, its penetration ability is weak and it is easily affected by environmental factors such as atmospheric disturbances, rain, snow, and fog; low-frequency SAR radars have strong penetration ability and can detect hidden sea targets or low-visibility ships, but their resolution is low and it is difficult to accurately identify target details.

[0004] Since it is difficult for a single-frequency-band SAR radar to balance high resolution and strong penetration, fusing multi-frequency-band SAR radars for target recognition has become an effective technical approach. Multi-frequency-band SAR radars can integrate the advantages of each frequency band and fuse SAR image data of different frequency bands, thereby improving the accuracy, reliability, and anti-interference ability of target recognition. This multi-frequency-band fusion technology can scan the monitored sea area in all directions, at multiple angles, and at multiple scales, not only expanding the coverage range of target detection, but also increasing the probability of target detection, and significantly reducing the risks of target escape and interference.

[0005] Although multi - band SAR radars have demonstrated excellent advantages in target monitoring and recognition, they still face numerous technical challenges in practical applications. Firstly, there is the problem of processing massive heterogeneous data. SAR data in different frequency bands is diverse and complex, and how to efficiently store, manage, and analyze this data has become a major difficulty. Secondly, the effective fusion of cross - band features is crucial. SAR radars in different frequency bands have their own characteristics. The high - frequency band provides fine resolution, while the low - frequency band has strong penetration. How to extract key features from multi - band data and achieve deep fusion directly affects the accuracy and reliability of target recognition. Finally, the collaborative optimization of multiple SAR radar stations is also an important challenge. Data sharing and cooperation among multiple SAR radar stations need to ensure real - time performance and consistency while improving system efficiency. Solving these problems is of great significance for further enhancing the practical application effect and performance of multi - band SAR radars.

[0006] Therefore, realizing multi - scale detection of ship targets based on multi - band SAR radars not only has important theoretical research value but also has broad application prospects in practice. This direction is not only a technical focus in the field of SAR radar detection in our country but also an important technical support for enhancing the ability of maritime situation awareness and ensuring maritime safety in the future. Summary of the Invention

[0007] In order to overcome the deficiency that it is difficult to have both high resolution and strong penetration in a single - band SAR radar, the purpose of the present invention is to provide a ship target recognition device based on multi - scale graph attention, which integrates the characteristics of multi - bands and multi - angles and has the advantages of high precision and high intelligence.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] A ship target recognition device based on multi - scale graph attention includes a multi - band SAR radar, a database, and a host computer. Among them, the multi - band SAR radar, the database, and the host computer are connected in sequence;

[0010] The multi - band SAR radar is used to scan the monitored sea area and store the SAR images of each frequency band into the database for processing by the host computer;

[0011] The database is used to store SAR images;

[0012] The host computer includes:

[0013] A data pre - processing module: used for pre - processing multi - band SAR image data;

[0014] A multi - scale graph data reconstruction module: used to construct the SAR images of the multi - band SAR radar into multi - scale graph - structured data;

[0015] Feature vector aggregation module: used to aggregate each node of the constructed multi-scale graph data to obtain feature vectors;

[0016] Target recognition module: used to perform target recognition on the aggregated feature vectors of each node.

[0017] Furthermore, the multi-band SAR radar is composed of M SAR radars distributed at different locations, each SAR radar is used to send signals of different frequency bands B m (B m =[f L ,f H ]) And receive the corresponding frequency band B m SAR images Where m is the corresponding SAR radar number (m=1,…,M, and M≥2), f L It is frequency band B m The lower cut-off frequency, f H It is frequency band B m The upper cutoff frequency, and are respectively the modulation signal sent by the m-th SAR radar at the i-th time and the SAR image corresponding to the i-th received SAR radar, where i = 1, 2, 3, …, N.

[0018] Furthermore, the database includes: a database for aggregating and storing SAR images of M SAR radars distributed at different locations The SAR images of each SAR radar are time aligned to ensure that the SAR images corresponding to the same subscript i of different SAR radars are in the same time period.

[0019] Furthermore, the data preprocessing module includes: for performing SAR radar image preprocessing, specifically:

[0020] Collect SAR images of M SAR radars in the same time period i from the database For the SAR images collected by each SAR radar m Perform normalization processing to obtain a normalized SAR image

[0021]

[0022] in, and Respectively represent The minimum and maximum values ​​of

[0023] For each SAR radar normalized SAR image Perform Fourier transform to obtain the frequency domain signal

[0024]

[0025] Among them, FFT represents the Fourier transform, and k represents the frequency serial number;

[0026] For the frequency-domain signal of each SAR radar Perform power spectral density estimation to obtain the power spectrum

[0027]

[0028] Among them, N represents the number of sampling points; the power spectral density estimation is used to quantify the power level of the SAR image at each frequency to identify the main frequency components in the SAR images of different frequency bands B m ;

[0029] For the power spectrum of each SAR radar Perform filtering processing to obtain the filtered power spectrum

[0030]

[0031] Among them, H m (k) represents the frequency response function of the filter corresponding to the m-th SAR radar. For each H m (k), the following is obtained:

[0032]

[0033] Among them, is the center frequency of the filter Q is the quality factor of the filter:

[0034]

[0035] Among them, f s is the sampling rate, k L and k H are the frequency variables corresponding to the lower cut-off frequency and the upper cut-off frequency of the Butterworth filter, and B is the -3dB bandwidth;

[0036] For the filtered power spectrum of each SAR radar Perform inverse Fourier transform to obtain the SAR radar SAR image after frequency-selective denoising:

[0037]

[0038] Among them, IFFT represents the inverse Fourier transform.

[0039] Furthermore, the multi-scale graph data reconstruction module includes: constructing the multi-band SAR images into graph-structured data, specifically:

[0040] For each preprocessed and denoised SAR image of the SAR radar stations Represent it as SAR radar graph data G i As the nodes in The constructed set of nodes is represented as:

[0041]

[0042] Construct a set of nodes V containing M nodes i , and each node is represented as a vector Where m represents the SAR radar station number and i represents the sampling point number in the time series.

[0043] Calculate the cosine similarity between each node as:

[0044]

[0045] Construct the edge set E between nodes i :

[0046]

[0047] Where σ is the similarity threshold, σ > 0, σ < 1. When the cosine similarity of the SAR image data between two SAR radar stations m and n is greater than the similarity threshold σ, the relevance of the SAR radar stations m and n is high, and an edge is constructed between these two nodes in the corresponding SAR radar graph data G i On it.

[0048] Based on the constructed set of nodes and the edge set, construct the SAR radar graph data corresponding to the i-th time period:

[0049] G i =(V i , E i ) (13).

[0050] Furthermore, the feature vector aggregation module is used to aggregate the constructed graph data for each node to obtain a feature vector, specifically:

[0051] For a certain node v on the graph m , calculate the importance e n Of its neighbor node v mn Relative to this node:

[0052] e mn = Leaky ReLU(a T [Wvm ||Wv n ) (14)

[0053] Among them, W is the trainable weight parameter for the node feature transformation of this layer, Leaky ReLU(·) is the LeakyRelu activation function, α is the weight coefficient. To ensure the weight coefficient, e mn It is applicable to nodes with different numbers of edges and performs normalization processing:

[0054]

[0055] Among them, N m represents all the single-hop neighbor nodes of node v m ;

[0056] Based on the attention mechanism, the features of the central node and its neighbor nodes are weighted and summed to aggregate the node features, obtaining the central node v m Concatenate the aggregated feature vectors:

[0057]

[0058] Among them, ‖ represents the concatenation of the output features, is the weight coefficient calculated by the attention mechanism a k (·) of the k-th layer, and W k is the weight parameter for the linear transformation of the node features corresponding to the k-th layer;

[0059] At the final (prediction) layer, the attention mechanism is adopted, and the average summation method is used to achieve node feature aggregation, obtaining the feature vector corresponding to the central node, realizing the multi-scale aggregation of the ship target feature vector:

[0060]

[0061] Furthermore, the target recognition module includes: using the multi-scale feature vector v ′ m of the central node as the multi-scale feature vector v i of the ship target corresponding to the i-th sea area monitoring, which is used to classify and identify the ship target. The ship target classification and recognition model SVNClassifier is trained through the following optimization problem:

[0062]

[0063] Among them, v i is the multi-scale aggregation feature vector of the ship target, y i is the corresponding label of the ship target, and α iis the Lagrange multiplier, and C is the regularization parameter used to control the penalty degree of misclassification; K(v i , v j ) = φ(v i ) · φ(v j ) is the kernel function used to calculate the inner product in the high-dimensional space;

[0064] Input the multi-scale aggregation feature vectors v i corresponding to each ship target into the classification model SVMClassifier to obtain the classification and recognition results y i of each ship target:

[0065] y i = SVMClassifier(v i ), i = 1, 2, 3, …, N (20)

[0066] Realize the classification and recognition of ship targets in the sea area monitored by the multi-band SAR radar.

[0067] The beneficial effects of the present invention are mainly manifested in:

[0068] 1. Multi-band SAR radar collaboration: By combining the high resolution of the high-frequency SAR radar and the strong penetration ability of the low-frequency SAR radar, realize the comprehensive and multi-dimensional monitoring of complex sea areas, and significantly improve the coverage range and recognition accuracy of target detection;

[0069] 2. Graph-based data processing and feature fusion: Utilize graph data reconstruction technology and graph neural network to effectively fuse multi-band SAR data, extract multi-scale feature vectors, improve data processing efficiency and feature expression ability, and enhance the accuracy and reliability of recognition;

[0070] 3. High-precision and highly intelligent target recognition: Use support vector machines for ship target recognition, which have high generalization and anti-interference capabilities for multi-source data, ensure the recognition accuracy and consistency in real-time sharing, and significantly improve the multi-radar detection efficiency and the overall performance of the system. Brief Description of the Drawings

[0071] Figure 1 is the schematic diagram of the device of the present invention. Detailed Embodiments

[0072] The present invention will be further described below with reference to the drawings. The embodiments of the present invention are used to explain the present invention rather than limit the present invention. Any modification and change made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

[0073] As Figure 1As shown in the figure, an embodiment of the present invention provides a ship target recognition device based on multi-scale graph attention, including a multi-band SAR radar 1, a database 2, and a host computer 3. The multi-band SAR radar 1, the database 2, and the host computer 3 are connected in sequence. The multi-band SAR radar is used to scan the monitored sea area and store the SAR image data of each band in the database.

[0074] The multi-band SAR radar is composed of M SAR radars distributed at different positions, and each SAR radar transmits different frequency bands B m (B m =[f L , f H ) modulated signals The frequency bands of the SAR radar stations cover from 300 MHz to 12 GHz. The resolution of each SAR radar gradually increases from 300 MHz to 12 GHz, and the penetrability decreases accordingly. Therefore, the frequency band bandwidth and the center frequency of the frequency band of each SAR radar can be adjusted according to the actual number of SAR radars and monitoring needs. The corresponding receiving frequency band of each SAR radar is B m SAR images where m refers to the corresponding SAR radar number (m = 1,..., M, and M ≥ 2), and f L is the lower cut-off frequency of the frequency band B m , and f H is the upper cut-off frequency of the frequency band B m . and respectively refer to the modulated signal transmitted by the m-th SAR radar at the i-th time and the SAR image of the SAR radar received corresponding to the i-th time, where i = 1, 2, 3,..., N;

[0075] The database aggregates and stores the SAR images of M SAR radars distributed at different positions and performs time alignment on the SAR images of each SAR radar to ensure that the SAR images corresponding to the same subscript i of different SAR radars are in the same time period, which is beneficial for subsequent feature fusion of SAR images.

[0076] The host computer includes a data preprocessing module, a multi-scale graph data reconstruction module, a feature vector aggregation module, and a target recognition module.

[0077] The data preprocessing module is used to perform SAR radar image preprocessing, including the following steps:

[0078] 1.1. Collect the SAR images of M SAR radars at the same time period i from the database Subsequently, for the SAR images collected by each SAR radar m respectively Perform normalization to obtain a normalized SAR image:

[0079]

[0080] Among them, and respectively represent the minimum and maximum values.

[0081] 1.2. For the normalized SAR image of each SAR radar Perform Fourier transform to obtain a frequency-domain signal

[0082]

[0083] Among them, FFT represents the Fourier transform, k represents the frequency index. Performing Fourier transform on the SAR image to analyze it from the time domain to the frequency domain is beneficial for segmenting and processing information in each frequency band.

[0084] 1.3. For the frequency-domain signal of each SAR radar Perform power spectral density estimation to obtain a power spectrum:

[0085]

[0086] Among them, N represents the number of sampling points. The purpose of power spectral density estimation is to quantify the power level of the SAR image at each frequency. Since each SAR radar station corresponds to a specific frequency band B m , therefore, performing power spectral density estimation in the frequency domain can more conveniently identify and segment the main frequency components in different frequency bands B m in the SAR radar SAR image.

[0087] 1.4. For the power spectrum of each SAR radar Perform filtering to obtain a filtered power spectrum:

[0088]

[0089] Among them, H m (k) represents the frequency response function of the filter corresponding to the m-th SAR radar. For each H m (k), there is:

[0090]

[0091] Among them, is the center frequency of the filter Q is the quality factor of the filter:

[0092]

[0093] where f s is the sampling rate, and k L and k H are the frequency variables corresponding to the lower cut-off frequency and the upper cut-off frequency of the Butterworth filter. Here, for each SAR radar, k L and k H will be determined according to the corresponding SAR radar band B m , that is is the center of the corresponding SAR radar band, B is the -3dB bandwidth. The Butterworth filter has a flat amplitude-frequency response and higher smoothness in the passband, which means that all frequencies in the passband can pass through the filter with a relatively consistent gain. This helps to maintain the original amplitude characteristics of the SAR radar SAR image, avoid introducing additional amplitude variations, and also reduces the ripples and discontinuities of the frequency response, ensuring that the filtered SAR image is distortion-free to the greatest extent possible.

[0094] 1.5. For the power spectrum of each filtered SAR radar perform an inverse Fourier transform (IFFT) to obtain the SAR radar SAR image after frequency-selective denoising:

[0095]

[0096] where IFFT represents the inverse Fourier transform.

[0097] The multi-scale graph data reconstruction module is used to construct the multi-band SAR radar SAR image into graph structure data and is completed by the following process:

[0098] 2.1. For the preprocessed and denoised data of each SAR radar station represent it as the nodes in the SAR radar graph data G i The constructed node set is represented as:

[0099] V_i = {v_i^m┤|(m = 1, 2, …, M)} (10)

[0100] In this way, a node set V containing M nodes is constructed i , and each node can be represented as a vector, that is where m represents the SAR radar station number and i represents the sampling point number in the time series.

[0101] 2.2. Calculate the cosine similarity between each node as:

[0102]

[0103] ​2.3. Construct the edge set E between nodes i :

[0104]

[0105] Among them, σ is the similarity threshold, σ > 0, σ < 1. The selection of the similarity threshold σ can be adjusted according to the actual situation. For example, it can be dynamically adjusted according to factors such as target size, distance, and motion state. Here, the present invention selects σ as 0.8. When the cosine similarity of the SAR radar data between two SAR radar stations m and n is greater than the similarity threshold σ, it is considered that these two SAR radar stations m and n have a strong correlation. An edge is constructed between these two nodes in the corresponding SAR radar map data G i . The edge of the constructed graph is undirected, indicating that the correlation between the two nodes is equal.

[0106] 2.4. Based on the node set and edge set constructed in the above steps, construct the SAR radar map data corresponding to the i-th time period:

[0107] G i =(V i , E i ) (13)

[0108] The feature vector aggregation module is used to aggregate the constructed graph data for each node to obtain a feature vector, which is completed by the following process:

[0109] 3.1. For a certain node v m on the graph, calculate the importance degree of its neighbor node v n relative to this node:

[0110] e mn =Leaky ReLU(a T [Wv m ||Wv n ) (14)

[0111] Among them, W is the trainable weight parameter for the feature transformation of the nodes in this layer, LeakyReLU(·) is the LeakyRelu activation function, and α is the weight coefficient. To ensure that the weight coefficient e mn is applicable to nodes with different numbers of edges, normalization is performed:

[0112]

[0113] Among them, represents all single-hop neighbor nodes of node v m .

[0114] 3.2 After obtaining the weight coefficients, based on the attention mechanism, the features of the central node and its neighbor nodes are weighted and summed to aggregate the node features, and the central node v m is obtained for the concatenated and aggregated feature vector:

[0115]

[0116] where, ‖ represents the concatenation of the output features, is the weight coefficient calculated by the attention mechanism (a k (·)) at the k-th layer, and W k is the weight parameter of the linear transformation of the node features corresponding to the k-th layer. Using the multi-head attention mechanism enhances the representation ability of the feature vector, and can focus on different parts of the input at different scales, maximizing the fusion of multi-band SAR information.

[0117] 3.3 On the final (prediction) layer, a multi-attention mechanism is adopted, and node aggregation is achieved using averaging instead of concatenation, so as to obtain the final aggregated node feature vector:

[0118]

[0119] The target recognition module uses the multi-scale feature vector v′ m of the central node as the multi-scale feature vector v i corresponding to the i-th sea area monitoring for classifying and identifying ship targets. By solving the following optimization problem, the ship target classification and recognition model SVMClassifier is trained:

[0120]

[0121]

[0122] where, v i is the multi-scale aggregation feature vector of the ship target, y i is the corresponding label of the ship target, α i is the Lagrange multiplier, C is the regularization parameter, which controls the penalty degree of misclassification. K(v i , v j ) = φ(v i ) · φ(v j ) is the kernel function, which allows calculating the inner product in a high-dimensional space without explicit mapping.

[0123] Thus, by inputting the multi-scale aggregation feature vector v i corresponding to each ship target into the classification model SVMCDlassifier, the classification and recognition results y i of each ship target can be obtained:

[0124] y i = SVMClassifier(v i ), i = 1, 2, 3, …, N (20)

[0125] Thus, the accurate classification and recognition of ship targets in the sea area monitored by the multi-band SAR radar are realized, providing a high-precision reference for the monitoring of ships in the sea area.

Claims

1. A ship target recognition device based on multi-scale graph attention, characterized in that: It includes a multi-band SAR radar, a database and a host computer, wherein the multi-band SAR radar, the database and the host computer are connected in sequence; The multi-band SAR radar is used to scan the monitored sea area and store the SAR images of each frequency band in the database for processing by the host computer; The database is used to store SAR images; The host computer comprises: Data preprocessing module: used for preprocessing multi-band SAR image data; Multi-scale graph data reconstruction module: used to construct multi-band SAR radar SAR images into multi-scale graph structure data; Feature vector aggregation module: used to aggregate each node of the constructed multi-scale graph data to obtain feature vectors; Target recognition module: used to perform target recognition on the aggregated feature vectors of each node.

2. According to claim 1, a ship target recognition device based on multi-scale graph attention is characterized in that: The multi-band SAR radar is composed of M SAR radars distributed at different locations, each SAR radar is used to send different frequency bands B m (B m =[f L ,f H ]) And receive the corresponding frequency band B m SAR images Where m is the corresponding SAR radar number (m=1,…,M, and M≥2), f L It is frequency band B m The lower cut-off frequency, f H It is frequency band B m The upper cutoff frequency, and are respectively the modulation signal sent by the m-th SAR radar at the i-th time and the SAR image corresponding to the i-th received SAR radar, where i = 1, 2, 3, …, N.

3. According to claim 1, a ship target recognition device based on multi-scale graph attention is characterized in that: The database includes: a database for aggregating and storing SAR images of M SAR radars distributed at different locations The SAR images of each SAR radar are time aligned to ensure that the SAR images corresponding to different SAR radars with the same subscript i are in the same time period.

4. According to claim 1, a ship target recognition device based on multi-scale graph attention is characterized in that: The data preprocessing module includes: a module for performing SAR radar image preprocessing, specifically: Collect SAR images of M SAR radars in the same time period i from the database For the SAR images collected by each SAR radar m Perform normalization processing to obtain a normalized SAR image in, and Respectively represent The minimum and maximum values ​​of For each SAR radar normalized SAR image Perform Fourier transform to obtain the frequency domain signal Among them, FFT represents Fourier transform, k represents the frequency number; For each SAR radar frequency domain signal Perform power spectral density estimation and obtain the power spectrum Wherein, N represents the number of sampling points; the power spectrum density estimation is used to quantify the power level of the SAR image at each frequency to identify different frequency bands B m The main frequency components in the SAR radar SAR image; For each SAR radar power spectrum Perform filtering to obtain the filtered power spectrum Among them, H m (k) represents the frequency response function of the filter corresponding to the mth SAR radar. For each H m (k), we get: in, is the center frequency of the filter Q is the quality factor of the filter, and we get: Among them, f s is the sampling rate, k L and k H is the frequency variable corresponding to the lower cutoff frequency and upper cutoff frequency of the Butterworth filter, and B is the -3dB bandwidth; For each SAR radar filtered power spectrum Perform inverse Fourier transform to obtain the SAR radar image after frequency selection and denoising: Here, IFFT stands for Inverse Fourier Transform.

5. According to claim 1, a ship target recognition device based on multi-scale graph attention is characterized in that: The multi-scale graph data reconstruction module includes: a module for constructing a multi-band SAR image into graph structure data, specifically: For each SAR radar station, the SAR image after preprocessing and denoising It is represented as SAR radar map data G i Nodes in The constructed node set is represented as: Construct a node set V containing M nodes i , each node is represented by a vector Where m represents the SAR radar station number, and i represents the sampling point number in the time series; The cosine similarity between nodes is calculated as: Construct the edge set E between nodes i : Where σ is the similarity threshold, σ>0, σ<1. When the cosine similarity of the SAR image data between two SAR radar stations m and n is greater than the similarity threshold σ, the correlation between the SAR radar stations m and n is high. i Build an edge between these two nodes; Based on the constructed node set and edge set, construct the SAR radar map data corresponding to the i-th time period: G i =(V i ,E i ) (13).

6. The ship target recognition device based on multi-scale graph attention according to claim 1 is characterized in that: The feature vector aggregation module is used to aggregate each node of the constructed graph data to obtain a feature vector, specifically: For a node v on the graph m , calculate its neighbor node v n The importance of this node mn : have been mn =Leaky ReLU(a T [Wv m ||Wv n ]) (14) Among them, W is the trainable weight parameter of the node feature transformation of this layer, Leaky ReLU (·) is the LeakyRelu activation function, and α is the weight coefficient. In order to ensure that the weight coefficient e mn This is applicable to nodes with different numbers of edges. mn Perform normalization: in, Represented as node v m All one-hop neighbor nodes; Based on the attention mechanism, the features of the central node and its neighboring nodes are weighted and aggregated to obtain the central node v m Concatenate the aggregated feature vectors: Among them, ‖ represents the concatenation of output features, is the attention mechanism a of the kth layer k (·) The calculated weight coefficient, W k is the weight parameter of the linear transformation of the node features corresponding to the kth layer; The attention mechanism is used in the final layer, and the average summation method is used to realize node feature aggregation, obtain the feature vector corresponding to the central node, and realize the multi-scale aggregation of the ship target feature vector:

7. The ship target recognition device based on multi-scale graph attention according to claim 1 is characterized in that: The target recognition module includes: m As the multi-scale feature vector v of the ship target corresponding to the i-th sea area monitoring i , which is used to classify and identify ship targets. The ship target classification and identification model SVMClassifier is obtained through the following optimization problem training: Among them, v i is the multi-scale aggregated feature vector of the ship target, y i is the label of the corresponding ship target, α i is the Lagrange multiplier, C is the regularization parameter used to control the degree of penalty for misclassification; K(v i ,v j )=φ(v i )·φ(v j ) is a kernel function used to calculate inner products in high-dimensional space; The multi-scale aggregated feature vector v corresponding to each ship target i Input into the classification model SVMClassifier to obtain the classification and recognition results y of each ship target i : y i =SVMClassifier(v i ),i=1,2,3,…,N (20) Realize the classification and identification of ship targets in the sea area monitored by multi-band SAR radar.

Images