SAR image ship identification method based on multi-scale feature fusion and anisotropic perception

Abstract

The invention discloses an SAR image ship identification method based on multi-scale feature fusion and anisotropic perception, and belongs to the field of SAR target identification. The SAR image ship identification technology plays a key role in the field of maritime affair monitoring and can be used for trade management, maritime traffic monitoring, maritime search and rescue and other tasks. According to the method, firstly, a multi-scale attention weighting module is designed to adaptively select effective feature scales and distribute weights, and meanwhile irrelevant scale interference is restrained; secondly, constructing a cross-level feature fusion module to fuse semantic features and detail features output by high and low layers of the backbone network, and enriching full-scale representation of the ship target; and finally, developing an anisotropic feature extraction module to break through the limitation of a traditional square convolution receptive field, and generating a rectangular receptive field matched with the ship shape. According to the method, the ship identification precision is remarkably improved, and the high-precision requirement of the maritime affair monitoring field on the ship identification technology is effectively met.

Description

Technical Field

[0001] The present invention belongs to the field of Synthetic Aperture Radar (SAR) target recognition and relates to a SAR image ship recognition method based on multi-scale feature fusion and anisotropy perception. Background Art

[0002] Synthetic Aperture Radar (SAR), with its unique advantages of all-weather and high resolution, has become a key piece of equipment in fields such as ocean monitoring and maritime management. Currently, traditional ship identification methods rely primarily on manually designed features, such as geometric and texture features, combined with machine learning algorithms for classification. However, these methods have significant limitations. First, the process of manually designing features is time-consuming and labor-intensive, and has poor adaptability to diverse data, making it difficult to meet the demands of complex and changing maritime environments. Second, these methods are susceptible to noise interference during feature extraction, resulting in unclear features and limiting improvements in recognition accuracy. Furthermore, traditional methods typically require complex mathematical theory support and have weak transferability, making it difficult to maintain stable recognition performance in different scenarios.

[0003] In recent years, with the advancement of deep learning technology, convolutional neural network (CNN)-based methods have gradually become the mainstream solution for SAR ship recognition. Although these methods have improved recognition accuracy to a certain extent, they still face many challenges: 1) Ship targets have multi-scale characteristics, and existing methods often fail to fully explore and utilize these multi-scale features, resulting in poor recognition performance when dealing with ships of different sizes. 2) Ships in SAR images have unique shape priors, and traditional square convolution kernels have difficulty effectively matching the slender shapes and arbitrary orientations of ships, thus affecting the accuracy and discriminability of feature extraction. 3) Existing CNN classifiers still suffer from large intra-class distances and small inter-class distances when dealing with the diversity within ship categories and the similarity between categories, further restricting recognition performance. Therefore, developing a recognition method that can fully utilize multi-scale features and incorporate prior knowledge of ship shape is of great significance for improving the accuracy and robustness of ship recognition in SAR images. Summary of the Invention

[0004] To address these challenges, this paper proposes a method for ship recognition in SAR images based on multiscale feature fusion and anisotropy perception. This method utilizes a task-specific Ship Enhanced Attention Network (SA2Net) to fully leverage multiscale features and integrate prior knowledge of ship shape into an end-to-end network. SA2Net comprises three core modules: a Cross-Level Feature Fusion Module (CFFM) for generating rich and reliable multiscale representations; a Multi-Scale Attention Weighting Module (MSAWM) for filtering feature scales and assigning weights; and an Anisotropic Feature Extraction Module (AFEM) for enhancing feature representation by incorporating prior knowledge of ship shape. Comprehensive experiments demonstrate that SA2Net outperforms several cutting-edge methods, ensuring accurate ship recognition in SAR images.

[0005] This paper discloses a SAR image ship recognition method based on multi-scale feature fusion and anisotropic perception, aiming to address the issue of insufficient ship recognition accuracy in SAR images. The method primarily comprises eight steps: dataset preparation, ship CNN feature extraction, a multi-scale attention weighting module, a cross-level feature fusion module, an anisotropic feature extraction module, ship recognition model establishment, model testing, and model evaluation. To generate rich and reliable multi-scale representations, the CFFM is introduced. CFFM fuses high-level semantics with low-level details through a weaving-like mechanism, generating richly expressive features at all scales. To fully and rationally utilize multi-scale features, the present invention introduces the MSAWM. MSAWM fuses high-level semantics with low-level details through a weaving-like mechanism, generating richly expressive features at all scales. To leverage prior knowledge of ship morphology, the present invention proposes the AFEM. Given the generally elongated shape of ships, this method breaks through the limitations of traditional square convolution kernels and innovatively employs rectangular convolution kernels for feature extraction. This design produces a rectangular receptive field that matches the ship's shape, addressing the inherent limitations of square convolution kernels in ship feature extraction. At the same time, experimental results on two open SAR ship datasets (i.e., Open-SARShip dataset and FUSAR-Ship dataset) jointly show that SA2Net surpasses three types of comparative methods in recognition performance, and the present invention achieves the most advanced SAR ship classification accuracy.

[0006] The specific technical solution of the SAR image ship recognition method based on multi-scale feature fusion and anisotropy perception provided by the present invention is as follows:

[0007] Step 1: Prepare a synthetic aperture radar (SAR) image ship and ocean target recognition dataset, preprocess the data, and divide the dataset into a training set and a test set.

[0008] Step 2: Use a cross-level feature fusion module to obtain a multi-scale feature pyramid through the feature extraction network;

[0009] Step 3: Use the ship anisotropic feature extraction module to perform convolution in five directions on each scale feature of the multi-scale feature pyramid, and then fuse them to obtain the feature map of the current scale, that is, to obtain a set of multi-scale feature maps;

[0010] Step 4: Using a multi-scale attention weighted module, the multi-scale feature map set output by the previous module is globally pooled to obtain a multi-scale feature vector representing global information. This is then concatenated to obtain a total feature vector, which is then fused with the attention weight matrix to obtain a weight vector with the same dimension as the number of scales. Subsequently, based on this weight pre-vector, the multi-scale feature vector is weighted summed to obtain enhanced features. Finally, the enhanced features are sequentially passed through a fully connected layer and a softmax function to obtain the final target recognition result.

[0011] Step 5: A SAR image ship recognition model is established using the cross-level feature fusion module, the ship anisotropic feature extraction module, and the multi-scale attention weighted module. After training, testing, and evaluating the model, a trained SAR image ship recognition model is obtained, and the model is used to perform the SAR image ship recognition task.

[0012] The cross-level feature fusion module is specifically as follows:

[0013] The feature extraction network is specifically: Network, SAR images through After the network, high-level features, same-level features and low-level features are obtained;

[0014] The cross-level feature fusion module also includes bilinear interpolation upsampling, 1×1 convolution, and convolution downsampling; high-level features , same-level features and low-level features After feature weaving and fusion, a three-layer feature pyramid with hierarchical representation capabilities is obtained , , , the mathematical expression of this process is:

[0015]

[0016]

[0017]

[0018] in represents 1×1 convolution, represents bilinear interpolation upsampling, Represents convolution downsampling.

[0019] The anisotropic feature extraction module is specifically as follows:

[0020] For feature pyramid , where l is the reference of the pyramid feature, square convolution is used with rectangular convolution in four directions: horizontal, vertical, left oblique, and right oblique, to obtain 5 sets of convolution results, and then cascade features are generated through cascade operations ; Then batch normalization, linear rectification unit ReLU and 3×3 convolution are used to perform the cascade feature Processing is performed to obtain a multi-scale feature map set .

[0021] The enhanced features are specifically obtained through the following process:

[0022] For multi-scale feature map collection , firstly, they are converted into feature vectors representing global information through global average pooling 、 、 , and then concatenate them to get the total eigenvector ;

[0023] Then get the weight prediction value :

[0024]

[0025] in is the attention weight matrix, Represents the dimension; then the multi-scale feature vectors are weighted summed to obtain the enhanced features:

[0026]

[0027] in, To enhance the features, represents element-wise addition between vectors, Represents element-wise multiplication between vectors

[0028] The innovation of this invention lies in the Ship Enhanced Attention Network (SA2Net). This method uses a Multi-Scale Attention Weighted Module (MSAWM) to adaptively select effective feature scales and assign weights, thereby addressing the underutilization of multi-scale features. It also constructs a Cross-Level Feature Fusion Module (CFFM) to fuse semantic and detail features output by high- and low-level layers of the backbone network, enriching the full-scale representation of ship targets. Furthermore, it develops an Anisotropic Feature Extraction Module (AFEM) to overcome the limitations of the traditional square convolution receptive field and generate a rectangular receptive field that matches the ship's shape, thereby addressing the issue of ineffective utilization of prior knowledge of ship shape. This method achieves an accuracy of 82.91% on a three-category task on the OpenSARShip dataset, 61.10% on a six-category task, and 88.28% on a seven-category task on the FUSAR-Ship dataset.

[0029] The advantages of this method lie in its ability to fully leverage multi-scale features to improve the accuracy and robustness of ship identification; its integration of prior knowledge of ship shape enhances the pertinence and effectiveness of feature extraction; and its collaborative approach using the MSAWM, CFFM, and AFEM modules to achieve optimal recognition performance, effectively overcoming the core challenges of SAR ship identification. This method provides an effective method for ship identification in SAR images, significantly improving its accuracy and effectively meeting the high-precision requirements for ship identification technology in the maritime surveillance field. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of the process of providing a method for identifying ship targets in SAR images in the present invention;

[0031] Figure 2 This is a schematic diagram of the structure of the CFFM module proposed in the present invention;

[0032] Figure 3 This is a schematic diagram of the structure of the AFEM module proposed in the present invention;

[0033] Figure 4 This is a structural diagram of the MSAWM module proposed in the present invention. DETAILED DESCRIPTION

[0034] In order to better understand the purpose, structure and function of the present invention, the present invention is further described below with reference to the accompanying drawings.

[0035] A ship recognition method for SAR images based on multi-scale feature fusion and anisotropy perception, such as Figure 1 As shown, the following steps are included:

[0036] Step 1: Prepare the dataset;

[0037] Step 1.1: Three categories of Open-SARShip dataset;

[0038] Based on the well-known Open-SARShip dataset, a small three-category OpenSARShip dataset was established for experiments, which includes bulk carriers, container ships and tankers; a new dataset was obtained and the dataset was divided into two parts in a ratio of 7:3 to obtain the training set and test set .

[0039] Open-SARShip is an open benchmark dataset for SAR ship and ocean target recognition and identification. It has a ship resolution of approximately 20 meters in range and azimuth, covering various ocean, land, coastal, river and island scenarios, and mainly includes three major ship categories, namely bulk carriers, container ships and tankers.

[0040] Step 1.2: Six-category Open-SARShip dataset;

[0041] Based on the well-known Open-SARShip dataset, a larger challenging six-category dataset was established, which includes bulk carriers, cargo, container ships, fisheries, general cargo, and tankers. According to the minimum number of samples in the six-category dataset, the training-test ratio was set to 8:2, and the training set was obtained. and test set .

[0042] Step 1.3: FUSAR-Ship Dataset

[0043] Based on the well-known FUSAR-Ship dataset, the order of SAR images in the FUSAR-Ship dataset is adjusted randomly, and the dataset is divided into two parts in a ratio of 7:3 to obtain the training set and test set .

[0044] FUSAR-Ship is an open benchmark dataset for SAR ship and ocean target recognition and identification, covering various ocean, land, coastal, river and island scenes, mainly including 15 major ship categories, 98 subcategories and many non-ship targets.

[0045] Step 2: Ship feature extraction;

[0046] Step 2.1: Image preprocessing;

[0047] The training set obtained in step 1.1 is resampled using classic image resampling, and each image is resized to 224×224 to obtain the input training set, which is denoted as .

[0048] The training set obtained in step 1.2 is resampled using classic image resampling, and each image is resized to 224×224 to obtain the input training set, which is denoted as .

[0049] The training set obtained in step 1.3 is resampled using classic image resampling, and each image is resized to 224×224 to obtain the input training set, which is denoted as .

[0050] For the convenience of description, the above training sets are collectively referred to as .

[0051] Step 2.2: Establish a cross-level feature fusion module CFFM;

[0052] use , first the training set enter Backbone network, get the output features of each layer of the network . Specifically, it consists of a five-stage fully convolutional network. The algorithm proposes a residual structure to ensure accuracy under the condition of network deepening. The algorithm is often used as a feature extraction network for a two-stage detector and is currently the method with the best detection accuracy in the existing technology.

[0053] like Figure 2 As shown, according to the CFFM module, the output features in the above structure are extracted As an object woven with features. Shallow features Contains more details and less semantics. In contrast, and It is a deep feature with less details and more semantics. The whole process of CFFM is defined as: bilinear interpolation upsampling, 1×1 convolution and convolution downsampling are used for high-level features, same-level features and low-level features respectively. After all feature channel dimensions are unified to 256, feature fusion is achieved by element-by-element addition. Multi-layer feature weaving and fusion obtains three layers of feature pyramid with hierarchical representation capabilities , , , the mathematical expression of this process is:

[0054]

[0055]

[0056]

[0057] in represents a 1×1 convolution operation that normalizes the channel dimension to 256. represents bilinear interpolation upsampling, represents 3×3 convolution downsampling with a stride of 2.

[0058] At this point, the CFFM cross-level feature fusion module has been established.

[0059] Step 3: Establish the ship anisotropic feature extraction module AFEM;

[0060] like Figure 3 As shown, the anisotropic feature extraction module is used to extract the pyramid features output by the previous module. (where l refers to the pyramid feature), and Convolution kernel to obtain pyramid features The original features and four convolution kernels in different directions are used to fit the shape of the SAR ship, and then the output is obtained. Five sets of feature maps are then generated through a cascade operation Specifically, AFEM adopts a multi-branch combination structure of square convolution and rectangular convolution in four directions (horizontal, vertical, left oblique, and right oblique): the square convolution branch retains the original features, while the rectangular convolution in four directions enhances the detailed features by constructing differentiated receptive fields.

[0061] Then, three consecutive operations of batch normalization (BN), linear rectification unit (ReLU) and 3×3 convolution (conv) are used to perform the above features. Processing, record the above three operations as The final output of AFEM is .

[0062] The AFEM process is defined as: , where Indicates cascade operation, express The horizontal convolution kernel, express The vertical convolution kernel, Represents the rectangular convolution kernel of the main diagonal, The rectangular convolution kernel of the sub-diagonal line; represents the left diagonal convolution size, Represents the right diagonal convolution size; finally, the output multi-scale feature map set .

[0063] The horizontal convolution kernel has a length of 1 and a width of n. This makes it easier to obtain horizontal features and obtain a horizontal rectangular receptive field for the feature map. In this invention, a 1×3 convolution kernel is used, and to ensure that the size after convolution is consistent with a 3×3 convolution, a horizontal step size of 1 and a padding of 1 are used; a vertical step size of 1 and a padding of 0 are used.

[0064] The vertical convolution kernel has a length of n and a width of 1. This makes it easier to obtain features in the vertical direction and obtain a vertical rectangular receptive field for the feature map. In this invention, a 3×1 convolution kernel is used, and to ensure that the size after convolution is consistent with a 3×3 convolution, a horizontal step size of 1 and a padding of 0 are used; a vertical step size of 1 and a padding of 1 are used.

[0065] The main diagonal convolution kernel is an n×n convolution kernel in which all parameters except those on the main diagonal are masked, making it easy to obtain features in the upper left to lower right direction and obtain a slanted rectangular receptive field for the feature map. In the present invention, the main diagonal convolution kernel in the 3×3 convolution kernel is used, and to ensure that the size after convolution is consistent with the 3×3 convolution, the horizontal step size is 1 and the padding is 1; the vertical step size is 1 and the padding is 1.

[0066] The sub-diagonal convolution kernel is an n×n convolution kernel in which all parameters except those on the sub-diagonal are masked, making it easy to obtain features in the upper right to lower left direction and obtain a slanted rectangular receptive field for the feature map. In the present invention, the sub-diagonal convolution kernel in the 3×3 convolution kernel is used, and to ensure that the size after convolution is consistent with the 3×3 convolution, the horizontal step size is 1 and the padding is 1; the vertical step size is 1 and the padding is 1.

[0067] Step 4: Establish a multi-scale attention weighting module MSAWM. The multi-scale attention weighting module (MSAWM) adaptively weights the effective feature scale and suppresses irrelevant scales by evaluating the correlation score between each scale feature and the final recognition probability.

[0068] like Figure 4 As shown, the MSAWM module is used to output the multi-scale feature map set of the previous module. , firstly, they are converted into feature vectors representing global information through global average pooling , and then concatenate them to get the total eigenvector In order to enable the module to automatically select the optimal feature scale to obtain better recognition results, and the attention weight matrix Do the related operation, where Represents the dimension of the feature vector; that is, multi-scale features are fused into a 3D weight vector to obtain the feature scale selection weight for a specific recognition task , .

[0069] Then, based on the weighted prediction , for the characteristic scale Perform weighted summation to obtain enhanced features , thereby improving SAR ship recognition performance:

[0070]

[0071] in, represents element-wise addition between vectors, Represents element-wise multiplication between vectors.

[0072] Finally, the enhanced features The final target recognition result can be obtained by passing through a fully connected layer and a softmax function in sequence.

[0073] Step 5: Build a ship recognition model;

[0074] A SAR image ship recognition model is established using the CFFM cross-level feature fusion module constructed in step 2, the AFEM ship anisotropic feature extraction module constructed in step 3, and the MSAWM multi-scale attention weighted module constructed in step 4 to obtain the ship recognition model of the method of the present invention.

[0075] Step 6: Train the ship recognition model;

[0076] The classic SGD algorithm initializes the stochastic gradient descent optimizer: setting the learning rate, training rounds, and batch value. Stochastic Gradient Descent (SGD) is an iterative optimization algorithm. Its core mechanism is to approximate the gradient direction of the loss function by randomly selecting a single sample or a small batch of samples, and then update the model parameters accordingly. The specific process is: in each iteration, a sample (or small batch) is randomly sampled from the training set, the loss gradient under the current parameters is calculated, and then the parameters are updated along the negative gradient direction (the update formula is ), where η is the learning rate, denoting the gradient). Compared to traditional batch gradient descent, SGD does not require traversing the entire data set, significantly reducing memory overhead and computational costs, making it particularly suitable for training large datasets. However, its gradient estimate is noisy due to random sampling, leading to oscillations in the parameter update path, potentially converging to a local optimum or slowly approaching the global optimum. To alleviate this problem, momentum methods (which accelerate convergence and suppress oscillations) or learning rate decay strategies (such as exponential decay and cosine annealing) are often combined to improve stability while maintaining efficient computation. This makes SGD one of the most fundamental optimizers in deep learning model training, and it is widely used in scenarios such as online learning and non-convex optimization.

[0077] The classification network and softmax loss function method in step 4 are used to construct the loss function.

[0078] Using the training set obtained in step 1, the SGD optimizer and loss function are used to train the model parameters on the model established in step 5 to obtain a trained ship recognition model.

[0079] Step 7: Test the ship recognition model;

[0080] The test set obtained in step 1 is resampled using the classic image resampling method, and each image is resized to 224×224.

[0081] The obtained test set is used to test the ship recognition model obtained in step 6 using the standard classification network test method in Definition 26, and the test results of the three test sets on the ship recognition model are obtained (i.e. 、 and ).

[0082] Step 8: Evaluate the ship identification model;

[0083] The ship identification model obtained in step 6 As input, the standard evaluation index calculation method is used to calculate the accuracy, recall, precision, three-level indicators and confusion matrix, as shown in Table 1.

[0084] Table 1:

[0085]

[0086] Accuracy Refers to the proportion of correct results among the results predicted as positive examples. Defined as , is the number of positive samples correctly identified, is the number of negative samples correctly identified, is the number of positive samples correctly identified, is the number of missed positive samples, and the correct number of ship classifications is , the total number of all ships is ;

[0087] Recall Refers to the number of correct predictions among all positive samples. Defined as ,in, is the number of positive samples correctly identified, is the number of positive samples that are missed;

[0088] Accuracy Refers to the number of correct predictions among all the positive samples predicted. Defined as , in, is the number of positive samples correctly identified, is the number of positive samples correctly identified;

[0089] Level 3 indicators At the same time, the accuracy of the recognition model is taken into account and recall . Level 3 indicators Defined as ,in represents the recall rate, Indicates accuracy;

[0090] The confusion matrix is ​​the most basic, intuitive, and computationally simple method for measuring the accuracy of a recognition model. It counts the number of observations that the recognition model incorrectly and correctly classifies, and then displays the results in a table.

[0091] The ship identification model obtained in step 6 As input, the standard evaluation index calculation method in Definition 27 is used to calculate the accuracy, recall, precision, three-level indicators and confusion matrix. The results are shown in Table 2.

[0092] Table 2:

[0093]

[0094] The ship identification model obtained in step 6 As input, the standard evaluation index calculation method in Definition 27 is used to calculate the accuracy, recall, precision, three-level indicators and confusion matrix. The results are shown in Table 3.

[0095] Table 3:

[0096]

[0097] As shown in Tables 1, 2, and 3, the accuracy achieved by the present invention on the FUSAR-Ship dataset is 88.28%, the accuracy on the three-category dataset on the Open-SARShip dataset is 82.91%, and the accuracy on the six-category dataset is 61.10%. As can be seen from the above experimental recognition results, the present invention achieves excellent detection accuracy, demonstrating that the present invention can achieve high-precision ship target recognition in SAR images.

[0098] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A SAR image ship recognition method based on multi-scale feature fusion and anisotropy perception, characterized by: The following steps are involved: Step 1: Prepare a synthetic aperture radar (SAR) image ship and ocean target recognition dataset, preprocess the data, and divide the dataset into a training set and a test set. Step 2: Use a cross-level feature fusion module to obtain a multi-scale feature pyramid through the feature extraction network; Step 3: Use the ship anisotropic feature extraction module to perform convolution in five directions on each scale feature of the multi-scale feature pyramid, and then fuse them to obtain the feature map of the current scale, that is, to obtain a set of multi-scale feature maps; Step 4: Using the multi-scale attention weighted module, for the multi-scale feature map set output by the previous module, a multi-scale feature vector representing global information is obtained through global pooling, which is then concatenated to obtain a total feature vector. The total feature vector is then fused with the attention weight matrix to obtain a weight vector with the same dimension as the number of scales. Subsequently, based on this weight vector, the multi-scale feature vectors are weighted summed to obtain enhanced features. Finally, the enhanced features are sequentially passed through a fully connected layer and a softmax function to obtain the final target recognition result; Step 5: A SAR image ship recognition model is established using the cross-level feature fusion module, the ship anisotropic feature extraction module, and the multi-scale attention weighted module. After training, testing, and evaluating the model, a trained SAR image ship recognition model is obtained, and the model is used to perform the SAR image ship recognition task.

2. The method for ship recognition in SAR images based on multi-scale feature fusion and anisotropy perception according to claim 1, characterized in that: The cross-level feature fusion module is specifically as follows: The feature extraction network is specifically: Network, SAR images through After the network, high-level features, same-level features and low-level features are obtained; The cross-level feature fusion module also includes bilinear interpolation upsampling, 1×1 convolution, and convolution downsampling; high-level features , same-level features and low-level features After feature weaving and fusion, a three-layer feature pyramid with hierarchical representation capabilities is obtained , , , the mathematical expression of this process is: ； ； ； in represents 1×1 convolution, represents bilinear interpolation upsampling, Represents convolution downsampling.

3. The method for ship recognition in SAR images based on multi-scale feature fusion and anisotropy perception according to claim 2, characterized in that: The anisotropic feature extraction module is specifically as follows: For feature pyramid , where l is the reference of the pyramid feature, square convolution is used with rectangular convolution in four directions: horizontal, vertical, left oblique, and right oblique, to obtain 5 sets of convolution results, and then cascade features are generated through cascade operations ; Then batch normalization, linear rectification unit ReLU and 3×3 convolution are used to perform the cascade feature Processing is performed to obtain a multi-scale feature map set .

4. The method for ship recognition in SAR images based on multi-scale feature fusion and anisotropy perception according to claim 3 is characterized in that: The enhanced features are specifically obtained through the following process: For multi-scale feature map collection , firstly, they are converted into feature vectors representing global information through global average pooling 、 、 , and then concatenate them to get the total eigenvector ; Then get the weight prediction value : ； in is the attention weight matrix, Represents the dimension; then the multi-scale feature vectors are weighted summed to obtain the enhanced features: ； in, To enhance the features, represents element-wise addition between vectors, Represents element-wise multiplication between vectors.

5. The method for ship recognition in SAR images based on multi-scale feature fusion and anisotropy perception according to claim 4, characterized in that: The square convolution is a 3×3 convolution, and the rectangular convolutions in the horizontal, vertical, left oblique, and right oblique directions are Horizontal convolution, Vertical convolution, rectangular convolution on the main diagonal, and rectangular convolution on the sub-diagonal.

Images