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

Abstract

This invention presents a SAR image-based ship identification method based on multi-scale feature fusion and anisotropic perception, belonging to the field of synthetic aperture radar (SAR) target identification. SAR image-based ship identification technology plays a crucial role in maritime monitoring and can be used for tasks such as trade management, maritime traffic monitoring, and maritime search and rescue. This invention first designs a multi-scale attention weighting module to adaptively select effective feature scales and assign weights while suppressing interference from irrelevant scales. Second, it constructs a cross-level feature fusion module to fuse semantic and detail features from high and low layers of the backbone network, enriching the full-scale representation of ship targets. Finally, it develops an anisotropic feature extraction module to overcome the limitations of traditional square convolutional receptive fields, generating rectangular receptive fields that match the ship's shape. This method significantly improves the accuracy of ship identification, effectively meeting the high-precision requirements of maritime monitoring for ship identification technology.

Description

Technical Field

[0001] This 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 anisotropic perception. Background Technology

[0002] Synthetic Aperture Radar (SAR), with its unique advantages of all-weather operation and high resolution, has become a key piece of equipment in fields such as marine monitoring and maritime management. Currently, traditional ship identification methods mainly rely on manually designed features, such as geometric and textural 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 its adaptability to different data is poor, making it difficult to meet the demands of the complex and ever-changing marine environment. Second, these methods are susceptible to noise interference during feature extraction, leading to unclear features and thus limiting the improvement of identification accuracy. Furthermore, traditional methods typically require complex mathematical theory support and have weak transferability, making it difficult to maintain stable identification performance under different scenarios.

[0003] In recent years, with the development of deep learning technology, methods based on convolutional neural networks (CNNs) have gradually become the mainstream approach for SAR ship identification. Although these methods have improved identification accuracy to some extent, they still face several 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 identification performance when dealing with ships of different sizes. 2) Ships in SAR images have unique shape prior features, and traditional square convolutional kernels struggle to effectively match the slender shape and arbitrary orientation of ships, thus affecting the accuracy and discriminative power of feature extraction. 3) Existing CNN classifiers still suffer from large intra-class distances and small inter-class distances when dealing with intra-class diversity and inter-class similarity, further limiting the improvement of identification performance. Therefore, developing an identification 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 identification in SAR images. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a SAR image ship recognition method based on multi-scale feature fusion and anisotropic perception. It utilizes a mission-specific Ship Enhancement Attention Network (SA2Net) to fully leverage multi-scale features and integrate prior knowledge of ship shape into an end-to-end network. SA2Net comprises three core modules: a Cross-Level Feature Fusion (CFFM) module for generating rich and reliable multi-scale representations; a Multi-Scale Attention Weighted (MSAWM) module responsible for selecting feature scales and assigning weights; and an Anisotropic Feature Extraction (AFEM) module that enhances feature representation by incorporating prior knowledge of ship shape. Comprehensive experiments demonstrate that SA2Net outperforms several cutting-edge methods, ensuring the accuracy of SAR image ship recognition.

[0005] This invention discloses a ship recognition method for SAR images based on multi-scale feature fusion and anisotropic perception, addressing the problem of insufficient ship recognition accuracy in SAR images. The method mainly comprises eight parts: dataset preparation, ship CNN feature extraction, multi-scale attention weighting module, cross-level feature fusion module, anisotropic feature extraction module, ship recognition model establishment, ship recognition model testing, and ship recognition model evaluation. To generate rich and reliable multi-scale representations, CFFM is introduced. CFFM fuses high-level semantics and low-level details through a weaving-like mechanism, generating features with rich representational power at each scale. To fully and rationally utilize multi-scale features, this invention introduces MSAWM. MSAWM fuses high-level semantics and low-level details through a weaving-like mechanism, generating features with rich representational power at each scale. To utilize prior knowledge of ship morphology, this invention proposes AFEM. Considering the generally slender shape of ships, this invention breaks through the limitations of traditional square convolutional kernels and innovatively uses rectangular convolutional kernels for feature extraction. This design can generate a rectangular receptive field matching the ship's shape, solving the inherent defects of square convolutional kernels in ship feature extraction. Meanwhile, experimental results on two open SAR ship datasets (i.e., the Open-SARShip dataset and the FUSAR-Ship dataset) jointly demonstrate that SA2Net outperforms the three-class comparison method in recognition performance, and this invention achieves state-of-the-art SAR ship classification accuracy.

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

[0007] Step 1: Prepare a synthetic aperture radar (SAR) image dataset for ship and marine target identification, preprocess the data, and divide the dataset into training and testing sets;

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

[0009] Step 3: Using the ship anisotropic feature extraction module, 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, thus obtaining the multi-scale feature map set.

[0010] Step 4: Employ a multi-scale attention weighting module. For the multi-scale feature map set output from the previous module, global pooling is used to obtain multi-scale feature vectors representing global information. These vectors are then concatenated to obtain a total feature vector. This 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 pre-vector, the multi-scale feature vectors are weighted and 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: Establish a SAR image ship recognition model using a cross-level feature fusion module, a ship anisotropic feature extraction module, and a multi-scale attention weighting module. After training, testing, and evaluating the model, a trained SAR image ship recognition model is obtained. Use this model to perform SAR image ship recognition tasks.

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

[0013] The feature extraction network is specifically as follows: Network, SAR image after The network yields high-level features, same-level features, and low-level features;

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

[0015]

[0016]

[0017]

[0018] in Represents a 1×1 convolution. This indicates bilinear interpolation upsampling. This indicates convolutional downsampling.

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

[0020] For feature pyramid Here, l represents the pyramid feature. Square convolutions are used, along with rectangular convolutions in four directions: horizontal, vertical, left-slant, and right-slant, resulting in five sets of convolutional results. These results are then concatenated to generate cascaded features. Then, batch normalization, ReLU linear rectified units, and 3×3 convolution are sequentially applied to the cascaded features obtained above. The process yields a set of multi-scale feature maps. .

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

[0022] For multi-scale feature map sets First, global average pooling is used to transform them into feature vectors representing global information. , , Then concatenate them to obtain the total eigenvector. ;

[0023] Then obtain the weight prediction value. :

[0024]

[0025] in Here is the attention weight matrix. The dimension is then represented; subsequently, the multi-scale feature vectors are weighted and summed to obtain the enhanced features:

[0026]

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

[0028] The innovation of this invention lies in proposing a Ship Enhanced Attention Network (SA2Net). It addresses the problem of insufficient utilization of multi-scale features by adaptively selecting effective feature scales and assigning weights through a Multi-Scale Attention Weighted Module (MSAWM). A Cross-Level Feature Fusion Module (CFFM) is constructed to fuse semantic and detail features from high and low layers of the backbone network, thereby enriching the full-scale representation of ship targets. An Anisotropic Feature Extraction Module (AFEM) is developed to overcome the limitations of traditional square convolutional receptive fields, generating rectangular receptive fields that match the ship's shape, thus solving the problem of ineffective application of prior knowledge of ship shape. This invention achieves an accuracy of 82.91% on the three-class classification task of the OpenSARShip dataset, 61.10% on the six-class classification task, and 88.28% on the seven-class classification task of the FUSAR-Ship dataset.

[0029] The advantages of this invention lie in its full utilization of multi-scale features, which improves the accuracy and robustness of ship identification; its organic integration of prior knowledge of ship shape enhances the targeting and effectiveness of feature extraction; and the synergistic interaction of the MSAWM, CFFM, and AFEM modules to achieve optimal identification performance, effectively overcoming the core challenges of SAR ship identification. This invention provides an effective SAR image ship identification method, significantly improving the accuracy of ship identification and effectively meeting the high-precision requirements of the maritime monitoring field for ship identification technology. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the method for identifying ship targets in SAR images according to the present invention.

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

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

[0033] Figure 4 This is a schematic diagram of the MSAWM module proposed in this invention. Detailed Implementation

[0034] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, further elaborates on the invention.

[0035] A method for ship identification based on SAR images using multi-scale feature fusion and anisotropic perception, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: Prepare the dataset;

[0037] Step 1.1: Three types of Open-SARShip datasets;

[0038] Based on the well-known Open-SARShip dataset, a small three-class Open-SARShip dataset was built for experiments, including bulk carriers, container ships, and tankers. The new dataset was then divided into two parts in a 7:3 ratio to obtain the training set. and test set .

[0039] Open-SARShip is an open benchmark dataset for SAR vessel and marine target identification and recognition. It has a vessel resolution of approximately 20 meters in both distance and bearing, and covers various marine, land, coastal, river, and island scenarios. It mainly includes three vessel categories: bulk carriers, container ships, and tankers.

[0040] Step 1.2: Six categories of Open-SARShip datasets;

[0041] Based on the well-known Open-SARShip dataset, a larger and more challenging six-class dataset was built, including bulk carriers, cargo ships, container ships, fisheries, general cargo, and tankers. The training-test ratio was set to 8:2 based on the minimum sample size for each of the six classes, resulting in the training set. 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 was randomly adjusted, and the dataset was divided into two parts in a 7:3 ratio to obtain the training set. and test set .

[0044] FUSAR-Ship is an open benchmark dataset for SAR ship and marine target identification and recognition, covering various marine, land, coastal, river and island scenarios, 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 techniques, resizing each image to a size of 224×224 to obtain the input training set, denoted as . .

[0048] The training set obtained in step 1.2 is resampled using classic image resampling techniques, resizing each image to a size of 224×224 to obtain the input training set, denoted as . .

[0049] The training set obtained in step 1.3 is resampled using classic image resampling techniques, resizing each image to a size of 224×224 to obtain the input training set, denoted as . .

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

[0051] Step 2.2: Establish the Cross-Level Feature Fusion Module (CFFM);

[0052] use First, the training set enter The backbone network is used to obtain the output features of each layer of the network. . Specifically, it consists of a 5-stage fully convolutional network. The algorithm proposes a residual structure to ensure accuracy under conditions of network deepening. The algorithm is often used as a feature extraction network for two-stage detectors, and it is currently the method with better detection accuracy among existing technologies.

[0053] like Figure 2 As shown, the output features of the above structure are extracted based on the CFFM module. As an object of feature weaving. Shallow features It contains more detail and less semantics. Conversely, and It involves deep features with fewer details and more semantics. The entire CFFM process is defined as follows: Bilinear interpolation upsampling, 1×1 convolution, and convolutional downsampling are applied to high-level features, same-level features, and low-level features respectively. After all feature channel dimensions are uniformly normalized to 256, feature fusion is achieved through element-wise addition, resulting in... Multi-layer feature weaving and fusion yields a three-layer feature pyramid with hierarchical representation capabilities. , , The mathematical expression for this process is:

[0054]

[0055]

[0056]

[0057] in This represents a 1×1 convolution operation that normalizes the channel dimension to 256. This indicates bilinear interpolation upsampling. This indicates a 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, an anisotropic feature extraction module is used to extract the pyramid features output by the previous module. (where l represents the pyramid feature), and simultaneously adopts Convolutional kernels to obtain pyramid features The original features and four different convolution kernels are used to fit the shape of the SAR ship, and then the output is obtained. Five sets of feature maps were then generated through cascading operations. Specifically, AFEM employs a multi-branch combination structure of square convolutions and rectangular convolutions in four directions (horizontal, vertical, left-slant, and right-slant): the square convolution branches retain the original features, while the rectangular convolutions in the four directions enhance detailed features by constructing differential receptive fields.

[0061] Then, the features obtained above are processed by three consecutive operations: batch normalization (BN), rectified linear unit (ReLU), and 3×3 convolution (conv). Perform the processing, and record the above three operations as follows: This yields the final output of AFEM. .

[0062] The AFEM process is defined as follows: In the formula, Indicates a cascading operation. express Horizontal convolution kernel, express Vertical convolution kernel, Represents a rectangular convolution kernel along the main diagonal. A rectangular convolution kernel along the secondary diagonal; where, Indicates the size of the left diagonal convolution. This indicates the size of the right diagonal convolution; ultimately, it outputs a set of multi-scale feature maps. .

[0063] The horizontal convolutional kernel is a convolutional kernel with a length of 1 and a width of n. It easily obtains features in the horizontal direction, providing a horizontal rectangular receptive field for the feature map. In this invention, a 1×3 convolutional kernel is used, and to ensure that the size after convolution is consistent with that of a 3×3 convolution, a horizontal stride of 1 and padding of 1 are adopted; a vertical stride of 1 and padding of 0 are adopted.

[0064] The vertical convolution kernel is a convolution kernel with a length of n and a width of 1. It easily obtains features in the vertical direction, providing 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 that of a 3×3 convolution, a horizontal stride of 1 and padding of 0 are adopted; a vertical stride of 1 and padding of 1 are also 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 easier to obtain features from the top left to the bottom right, thus obtaining a diagonal rectangular receptive field for the feature map. In this invention, the main diagonal convolution kernel from a 3×3 convolution kernel is used, and to ensure that the size after convolution is consistent with that of a 3×3 convolution, a horizontal stride of 1 and a padding of 1 are used; a vertical stride of 1 and a padding of 1 are also used.

[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 easier to obtain features from the upper right to the lower left, thus obtaining a diagonal rectangular receptive field for the feature map. In this invention, a sub-diagonal convolution kernel from a 3×3 convolution kernel is used. To ensure that the size after convolution is consistent with that of a 3×3 convolution, a horizontal stride of 1 and a padding of 1 are used; a vertical stride of 1 and a padding of 1 are also used.

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

[0068] like Figure 4 As shown, the MSAWM module is used to analyze the multi-scale feature map set output by the previous module. First, global average pooling is used to transform them into feature vectors representing global information. Then concatenate them to obtain the total feature vector. To enable the module to automatically select the optimal feature scale for better recognition results, then... With attention weight matrix Perform relevant calculations, among which This represents the dimension of the feature vector; that is, fusing multi-scale features into a 3-dimensional weight vector to obtain the feature scale selection weights for a specific recognition task. , .

[0069] Subsequently, based on this weighted prediction value For feature scale The enhanced features can be obtained by performing a weighted summation. This improves SAR ship identification performance.

[0070]

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

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

[0073] Step 5: Construct a ship identification 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 weighting module constructed in step 4, thus obtaining the ship recognition model of the present invention.

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

[0076] The classic SGD algorithm initializes the stochastic gradient descent optimizer by setting the learning rate, training epochs, and batch size. Stochastic Gradient Descent (SGD) is an iterative optimization algorithm whose core mechanism is to approximate the gradient direction of the loss function by randomly selecting a single sample or a mini-batch of samples, and then update the model parameters accordingly. Specifically, in each iteration, a sample (or mini-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, (Represents the gradient). Compared to traditional batch gradient descent, SGD does not require traversing the entire dataset, significantly reducing memory overhead and computational costs, making it particularly suitable for training large-scale datasets. However, its gradient estimation is noisy due to random sampling, causing oscillations in the parameter update path, which may lead to convergence to a local optimum or slowly approach the global optimum. To alleviate this problem, momentum methods (to accelerate convergence and suppress oscillations) or learning rate decay strategies (such as exponential decay and cosine annealing) are often combined to improve stability while retaining efficient computational characteristics, making it one of the most basic optimizers in deep learning model training and widely used in online learning, non-convex optimization, and other scenarios.

[0077] The loss function is constructed using the classification network and softmax loss function method from step 4.

[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 was resampled using the classic image resampling method, and each image was 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 testing method in Definition 26, resulting in test results for the three test sets on the ship recognition model (i.e., , and ).

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

[0083] Based on the ship identification model obtained in step 6 As input, the accuracy, recall, precision, three-level metrics, and confusion matrix were calculated using standard evaluation index calculation methods, as shown in Table 1.

[0084] Table 1:

[0085]

[0086] Accuracy Accuracy refers to the percentage of correct predictions out of all predicted positive results. Defined as , This represents the number of positive samples that were correctly identified. This represents the number of negative samples that were correctly identified. This represents the number of positive samples that were correctly identified. The number of underreported positive samples is [value], and the number of correctly classified ships is [value]. The total number of all ships is ;

[0087] Recall rate Recall refers to the number of correctly predicted samples out of all positive samples. Defined as ,in, This represents the number of positive samples that were correctly identified. This represents the number of positive samples that were missed.

[0088] Accuracy Precision refers to the number of correctly predicted samples out of all predicted positive samples. Defined as , in, This represents the number of positive samples that were correctly identified. This represents the number of positive samples that were correctly identified.

[0089] Level 3 indicators It also takes into account the accuracy of the recognition model. and recall rate Level 3 indicators Defined as ,in Indicates recall rate, Indicate precision;

[0090] The confusion matrix is ​​the most basic, intuitive, and simplest method for measuring the accuracy of a recognition model. The confusion matrix is ​​created by counting the number of observations misclassified and correctly classified by the recognition model, and then displaying the results in a table.

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

[0092] Table 2:

[0093]

[0094] Based on the ship identification model obtained in step 6 As input, the standard evaluation index calculation method in Definition 27 was used to calculate the accuracy, recall, precision, three-level index 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 this invention on the FUSAR-Ship dataset is 88.28%, on the three classes of the Open-SARShip dataset is 82.91%, and on the six classes is 61.10%. These experimental results demonstrate that this invention achieves excellent detection accuracy, indicating that it can realize high-precision ship target identification in SAR images.

[0098] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the 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 within the protection scope of the present invention.

Claims

1. A method for ship identification based on SAR images using multi-scale feature fusion and anisotropic perception, characterized in that, Includes the following steps: Step 1: Prepare a synthetic aperture radar (SAR) image dataset for ship and marine target identification, preprocess the data, and divide the dataset into training and testing sets; Step 2: Employ a cross-level feature fusion module to obtain a multi-scale feature pyramid through a feature extraction network; Step 3: Using the ship anisotropic feature extraction module, 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, thus obtaining the multi-scale feature map set. Step 4: Employ a multi-scale attention weighting module. For the multi-scale feature map set output by the previous module, global pooling is used to obtain multi-scale feature vectors representing global information. These vectors are 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 pre-vector, the multi-scale feature vectors are weighted and summed to obtain enhanced features. Finally, the enhanced features are passed through a fully connected layer and a softmax function to obtain the final target recognition result. Step 5: Establish a SAR image ship recognition model using a cross-level feature fusion module, a ship anisotropic feature extraction module, and a multi-scale attention weighting module. After training, testing, and evaluating the model, a trained SAR image ship recognition model is obtained. Use this model to perform the SAR image ship recognition task. The anisotropic feature extraction module is as follows: For feature pyramid Here, 'l' represents the pyramid feature. Square convolutions are used, along with rectangular convolutions in four directions: horizontal, vertical, left-slant, and right-slant, resulting in five sets of convolutional results. These results are then concatenated to generate cascaded features. Then, batch normalization, ReLU linear rectified units, and 3×3 convolution are sequentially applied to the cascaded features M′ obtained above. l The processing yields multi-scale feature map sets A3, A4, and A5. The enhanced features are obtained through the following process: For the multi-scale feature map sets A3, A4, and A5, they are first converted into feature vectors f1, f2, and f3 representing global information by global average pooling, and then concatenated to obtain the total feature vector f. Then obtain the weight prediction value w. i : in Here, d represents the attention weight matrix; subsequently, the multi-scale feature vectors are weighted and summed to obtain the enhanced features: in, To enhance features, This represents element-wise addition between vectors. This represents element-wise multiplication between vectors.

2. The SAR image ship recognition method based on multi-scale feature fusion and anisotropic perception according to claim 1, characterized in that, The cross-level feature fusion module is specifically as follows: The feature extraction network is specifically a ResNet-50 network. After the SAR image passes through the ResNet-50 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 convolutional downsampling; high-level feature C5, same-level feature C4, and low-level feature C3 are fused through feature weaving to obtain a three-layer feature pyramid M3, M4, and M5 with hierarchical representation capabilities. The mathematical expression of this process is as follows: M3=ConvU(C5)+ConvU(C4)+Conv(C3); M4=ConvU(C5)+Conv(C4)+ConvD(C3); M5=Conv(C5)+ConvD(C4)+ConvD(C3); Where Conv(·) represents 1×1 convolution, ConvU(·) represents bilinear interpolation upsampling, and ConvD(·) represents convolution downsampling.

3. The SAR image ship recognition method based on multi-scale feature fusion and anisotropic perception according to claim 2, characterized in that, The square convolution is a 3×3 convolution, and the rectangular convolutions in the four directions of horizontal, vertical, left diagonal, and right diagonal are a 1×3 horizontal convolution, a 3×1 vertical convolution, a rectangular convolution along the main diagonal, and a rectangular convolution along the secondary diagonal.

Images