SAR ship detection method based on edge-guided multi-scale frequency domain perception

Abstract

The application discloses a SAR ship detection method based on edge-guided multi-scale frequency domain perception and belongs to the technical field of target detection. First, SAR ship image data sets are acquired and divided into a training set, a verification set and a test set; then, a SAR ship detection network model is built, wherein a backbone network is used to extract a multi-scale feature map with direction perception, including a double-path edge-guided fusion module, a multi-scale group interaction module and a double-attention residual module, a neck network feature pyramid network processes the multi-scale feature map, and an adaptive high-frequency screening module is integrated to generate a multi-scale fusion feature map used for detection; then, the training set and the verification set are input into the detection model for training and verification, the network parameters are updated by calculating a loss function, and an optimal parameter model is acquired; finally, the test set is input into the optimal parameter model, and a detection result is output. The application can significantly improve the detection precision and recall rate of multi-scale ships.

Description

Technical Field

[0001] This invention relates to SAR ship detection methods, specifically to SAR ship detection methods based on edge-guided multi-scale frequency domain sensing, belonging to the field of target detection technology. Background Technology

[0002] Synthetic Aperture Radar (SAR), with its all-weather, all-day, and high-resolution imaging capabilities, has significant research value and application prospects in fields such as maritime traffic monitoring, national defense security, and resource exploration. In particular, it provides key technical support for tasks such as maritime rescue, border patrol, and waterway safety, and has therefore become a research hotspot both domestically and internationally.

[0003] With the rise of deep learning technology, it has gradually become the mainstream in the field of SAR target detection. In the field of SAR ship detection, some existing detection models adopt dynamic deformable convolution and large receptive field attention mechanisms to improve the model's ability to perceive local details and global context. Others design adaptive large kernel selection modules and multi-scale attention mechanisms combined with structural reparameterization technology to improve the detection accuracy and edge computing efficiency of small ships in SAR images. However, in SAR ship target detection tasks, ship targets are often affected by speckle noise and complex sea conditions, resulting in blurred target edges and difficulty in capturing directional scattering structures. The scale of ships varies drastically, and small target features are easily lost in deep networks. Low-frequency background noise such as sea clutter severely masks the high-frequency details of the target. These problems lead to low detection accuracy and high false positive and false negative rates in existing detection models.

[0004] To address this, a SAR ship detection method based on edge-guided multi-scale frequency domain sensing is proposed. Summary of the Invention

[0005] To address the problems in existing SAR ship detection methods, such as speckle noise and interference from complex sea conditions leading to blurred ship target edges and difficulty in extracting directional scattering features, large variations in ship target scale causing the loss of small target features during feature extraction, and low-frequency background interference such as sea clutter masking high-frequency details of ship targets, resulting in missed detections and false detections, this invention proposes a SAR ship detection method based on edge-guided multi-scale frequency domain perception. This method extracts directional scattering information and edge contour features of ship targets through a dual-path edge-guided fusion module, extracts and fuses multi-scale target features from different receptive fields through a multi-scale inter-group interaction module, suppresses low-frequency background interference and enhances the high-frequency discriminative information of ship targets through an adaptive high-frequency filtering module, and recalibrates deep features from spatial and channel dimensions through a dual-attention residual module. This improves the detection accuracy of ship targets at different scales under complex sea conditions and reduces the missed detection rate and false detection rate.

[0006] To achieve the above objectives, this invention provides a SAR ship detection method based on edge-guided multi-scale frequency domain sensing, comprising the following steps:

[0007] Step 1: Obtain the SAR ship image dataset and divide it into training set, validation set and test set;

[0008] Step 2: Build a SAR ship detection network model based on edge-guided multi-scale frequency domain sensing, including: backbone network, neck network and detection head;

[0009] The backbone network is used to extract direction-aware multi-scale feature maps, including a dual-path edge-guided fusion module, a multi-scale inter-group interaction module, and a dual-attention residual module; the neck network uses a bidirectional feature fusion path and integrates an adaptive high-frequency filtering module to process the multi-scale feature maps and generate multi-scale fused feature maps; the detection head receives the multi-scale fused feature maps and detects ship targets at different scales;

[0010] Step 3: Input the SAR ship images from the training set and the validation set into the SAR ship detection network model for training and validation, respectively. Calculate the loss function and perform backpropagation to update the network parameters to obtain the optimal parameter model.

[0011] Step 4: Input the test set into the optimal parameter model and output the SAR ship detection results.

[0012] Furthermore, in step 2, the process of the backbone network extracting multi-scale feature maps includes:

[0013] The input image is processed sequentially using a convolutional module and a dual-path edge-guided fusion module to obtain a shallow enhanced feature map;

[0014] The shallow enhanced feature map is input into the first-stage feature extraction module, and then processed sequentially by the multi-scale inter-group interaction module and the convolution module to obtain the first-stage output feature map.

[0015] The first-stage output feature map is input into the second-stage feature extraction module, processed by the multi-scale inter-group interaction module to obtain the first-scale feature map, and then processed by the convolution module to obtain the second-stage output feature map.

[0016] The second-stage output feature map is input into the third-stage feature extraction module, processed by the multi-scale inter-group interaction module to obtain the second-scale feature map, and then processed by the convolution module to obtain the third-stage output feature map.

[0017] The third-stage output feature map is input into the fourth-stage feature extraction module, and then processed sequentially through the multi-scale inter-group interaction module, the spatial pyramid pooling module, and the dual attention residual module to obtain the fourth-stage output feature map, i.e., the third-scale feature map.

[0018] The first-scale feature map, the second-scale feature map, and the third-scale feature map are combined into a multi-scale feature map, and the multi-scale feature map is input into the neck network for processing.

[0019] Furthermore, the implementation process of the dual-path edge-guided fusion module for processing features includes:

[0020] The input features are processed using a direction-aware convolution path. Basic features are extracted through a 3×3 standard convolution, and then horizontal and vertical scattering features are captured by 1×3 and 3×1 asymmetric convolutions. Finally, channel information is fused through a 1×1 convolution to obtain the direction-aware features.

[0021] The input features are processed using an edge enhancement path. The initial edge response is extracted using the Scharr edge detection operator. The edge response map is obtained by calculating the gradient magnitude. After normalization and activation function processing, it is added to the input features and then input to the feature enhancement block for depth extraction to obtain the edge enhancement features. The feature enhancement block includes 1×1 convolution, 3×3 convolution and 1×1 convolution.

[0022] The orientation-aware features and the edge enhancement features are adaptively fused using learnable weight parameters. After fusion, the features are integrated with local context information through a 3×3 convolution. Then, the weights of each channel are adjusted by a channel attention module. Finally, the features are added to the input features to output the final dual-path edge-guided fusion features.

[0023] Furthermore, the implementation process of the multi-scale inter-group interaction module for processing features includes:

[0024] The input features are processed by three parallel branches: 1×1 convolution, 3×3 convolution, and 5×5 convolution. The outputs of the three branches are then concatenated along the channel dimension to generate multi-scale concatenated features.

[0025] The multi-scale splicing features are input into the feature enhancement unit and processed sequentially through 1×1 convolution, 3×3 depthwise separable convolution, batch normalization and ReLU activation function to obtain multi-scale enhanced splicing features;

[0026] The multi-scale enhanced splicing features are input to the channel attention branch and the spatial attention branch respectively for processing. The outputs of the channel attention branch and the spatial attention branch are weighted and fused. After fusion, the channels are integrated by 1×1 convolution. Finally, the residual connection is added to the input features to output the final multi-scale inter-group interaction features.

[0027] Furthermore, the specific process of the dual-attention residual module processing features includes:

[0028] The input features are processed sequentially through a 3×3 convolution, a batch normalization layer, a ReLU activation function, another 3×3 convolution, and a batch normalization layer to obtain the basic convolutional features.

[0029] The basic convolutional features are input into the spatial attention submodule for processing to generate spatial attention weights. The spatial attention weights are then used to weight the basic convolutional features in terms of spatial dimensions to obtain spatially enhanced features.

[0030] The spatial enhancement features are input into the channel attention submodule for processing to generate channel attention weights. The channel attention weights are then used to weight the spatial enhancement features according to the channel dimension to obtain dual attention enhancement features.

[0031] The dual attention enhancement features are added to the input features through a residual connection to output the final dual attention residual features.

[0032] Furthermore, in step 2, the specific process of the adaptive high-frequency screening module processing features includes:

[0033] The input features are transformed to the frequency domain using discrete cosine transform to obtain frequency domain features;

[0034] The frequency domain features are filtered using an adaptive high-pass filter to obtain the filtered frequency domain features.

[0035] The filtered frequency domain features are input into the channel path for processing. Global average pooling and global max pooling are applied to the filtered frequency domain features to compress spatial information. Then, after ReLU activation, 1×1 convolution and channel concatenation, 1×1 convolution and Sigmoid activation function are used to generate channel attention weights. Then, inverse discrete cosine transform is applied back to the spatial domain, and then multiplied element-wise with the input features to obtain the output of the channel path.

[0036] The filtered frequency domain features are input into the spatial enhancement path for processing. By performing an inverse discrete cosine transform (ICC) on the filtered frequency domain features back to the spatial domain, high-frequency response features are obtained. These high-frequency spatial response features are then processed sequentially through a 3×3 depthwise separable convolution and a 1×1 convolution to generate spatial mask features. These spatial mask features are then normalized using a Sigmoid activation function to generate spatial attention weights, which are then multiplied element-wise with the input features to obtain the output of the spatial enhancement path.

[0037] After adding the output of the channel path and the output of the spatial enhancement path, the two paths are further integrated by 3×3 convolution. Finally, the residual connection is used to sum the input features to output the final adaptive high-frequency filtering features.

[0038] Furthermore, in step 3, the loss function includes classification loss and regression loss; wherein the classification loss uses the binary cross-entropy loss function, and the regression loss uses the CIoU loss function.

[0039] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0040] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] This invention designs a dual-path edge-guided fusion module in the shallow layer of the backbone network. It extracts features in parallel through a direction-aware convolutional path and an edge enhancement path. The direction-aware convolutional path explicitly models the directional scattering structure of the ship using an asymmetric convolutional architecture, while the edge enhancement path uses the rotation-invariant Scharr edge detection operator to extract robust edge responses. The two paths are adaptively fused through learnable weights, which effectively enhances the geometric contour and directional scattering features of the ship target and solves the edge blurring problem caused by speckle noise.

[0043] This invention constructs a multi-scale inter-group interaction module, which extracts multi-scale features of different receptive fields through parallel multi-branch convolution, and integrates channel attention and spatial attention branches to achieve cross-dimensional feature interaction, thereby enhancing the network's sensitivity to targets of different scales and significantly improving the model's adaptability to drastic scale changes and small targets.

[0044] This invention constructs an adaptive high-frequency screening module, which designs an adaptive high-pass filter with learnable parameters in the frequency domain. It can dynamically adjust the cutoff frequency and filter band smoothness according to the input data. Through parallel processing of channel path and spatial enhancement path, it achieves dynamic suppression of low-frequency background interference such as sea clutter and precise enhancement of high-frequency discriminative details of ship targets, effectively reducing false detection and missed detection in complex sea conditions.

[0045] This invention integrates a dual-attention residual module at the end of the backbone network and adopts a residual learning framework that uses spatial attention and channel attention to perform dual refinement of deep features in terms of spatial location and semantic channels, thereby further focusing on the target region and improving the feature discrimination ability in complex backgrounds. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow provided by the present invention;

[0047] Figure 2 This is a schematic diagram of the overall network model structure provided by the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of the dual-path edge-guided fusion module provided by the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of the multi-scale inter-group interaction module provided by the present invention;

[0050] Figure 5 This is a schematic diagram of the adaptive high-frequency screening module provided by the present invention;

[0051] Figure 6 This is a schematic diagram of the structure of the dual attention residual module provided by the present invention;

[0052] Figure 7 This is a schematic diagram comparing the detection results of different methods provided by the present invention in a near-shore interference scenario;

[0053] Figure 8 This is a schematic diagram comparing the detection results of different methods provided by the present invention in a dense ship and vessel scenario.

[0054] Figure 9 This is a schematic diagram comparing the detection results of different methods provided by the present invention in a multi-scale ship and vessel scenario.

[0055] Figure 10 This diagram illustrates a comparison of detection results of different methods provided by the present invention in a multi-target scenario. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0057] Example 1:

[0058] The SAR ship detection method based on edge-guided multi-scale frequency domain sensing provided in this embodiment, such as Figure 1 As shown, it includes the following steps:

[0059] Step 1: Obtain the SAR ship image dataset, preprocess the dataset, and divide it into training set, validation set, and test set according to the set ratio;

[0060] Furthermore, the steps for obtaining the dataset, preprocessing the dataset, and partitioning the dataset are as follows:

[0061] The datasets to be acquired are: the Synthetic Aperture Radar Ship Detection Dataset (SSDD), the High-Resolution Synthetic Aperture Radar Image Dataset (HRSID), and the SAR-Ship-Dataset, which include SAR images and their corresponding annotation files.

[0062] The SSDD dataset was first proposed by Li et al. in 2017, and the official version was released by Zhang et al. in 2021. It contains 1160 SAR images with a total of 2456 ship target instances, with varying image sizes and a wide range of target scales. The HRSID dataset was released by Wei et al. in 2020, containing 5604 high-resolution SAR images with a total of 16951 ship target instances. The images have high resolution and contain a large number of small targets. The SAR-Ship-Dataset dataset contains 43819 SAR images with a total of 59535 ship targets.

[0063] The dataset preprocessing specifically involves performing basic preprocessing operations such as unifying the size, converting the data type, and normalizing the original images in the dataset, and performing data augmentation, including mosaic enhancement, mirroring, and flipping, to avoid overfitting of the model.

[0064] The dataset is divided as follows: the enhanced ship image dataset is randomly divided into three parts, with the training set accounting for 70%, the validation set accounting for 20%, and the test set accounting for 10%.

[0065] Step 2: Build a SAR ship detection network model based on edge-guided multi-scale frequency domain sensing, including an input backbone network, a neck network, and a detection head;

[0066] Model structure reference Figure 2The backbone network is used to extract orientation-aware multi-scale feature maps. It mainly consists of convolutional modules, dual path edge-guided fusion (DEF) modules, multi-scale group interaction (MSGI) modules, spatial pyramid pooling-fast (SPPF) modules, and dual attention residual blocks (DARB) modules.

[0067] The spatial pyramid pooling module consists of a convolutional module and three consecutive 5×5 max-pooling layers connected in series. After the input features are processed by the convolutional module, they are passed through the three max-pooling layers in sequence. The original features are then concatenated with the features obtained after the three pooling steps, and finally, a convolutional module is used to output multi-scale fused features.

[0068] Furthermore, in step 2, the specific feature extraction process of the backbone network includes:

[0069] First, the input image is processed sequentially through two convolutional modules and a dual-path edge-guided fusion module to obtain a shallow enhanced feature map;

[0070] Subsequently, the shallow enhanced feature map is processed by the first-stage multi-scale inter-group interaction module and convolution module to obtain the first-stage output feature map;

[0071] The first-stage output feature map enters the second stage, and after processing by the multi-scale inter-group interaction module, the first-scale feature map P3 is obtained. P3 is downsampled by the convolution module and then enters the third stage. The multi-scale inter-group interaction module in the third stage outputs the second-scale feature map P4. P4 is downsampled by the convolution module and then enters the fourth stage. The fourth stage is processed sequentially by the multi-scale inter-group interaction module, the spatial pyramid pooling module, and the dual attention residual module, and outputs the third-scale feature map P5.

[0072] The convolutional module has a downsampling function, including standard convolution, batch normalization layers, and the SiLU activation function. The stride of the standard convolution is set to 2, which downsamples the feature map size to half of the input scale. All convolutional modules in the backbone network use the same downsampling ratio, that is, the height and width of the feature map are halved after each downsampling.

[0073] Furthermore, the model structure reference for the dual-path edge-guided fusion module. Figure 3 It is embedded in the shallow layers of the backbone network and adopts a parallel architecture of direction-aware convolutional paths and edge enhancement paths. The specific implementation process includes:

[0074] The input features are processed using a direction-aware convolutional path. Basic features are extracted using a standard 3×3 convolution, followed by the capture of horizontal and vertical scattering features using 1×3 and 3×1 asymmetric convolutions. Finally, channel information is fused using a 1×1 convolution to obtain the direction-aware features. This process is illustrated in the following equation:

[0075] ,

[0076] Among them, F in For the input feature map, The output feature map of the direction-aware convolution path. This represents a convolution operation with a kernel size of 3×l. This indicates a convolution operation with a kernel size of 1×3. This indicates a convolution operation with a kernel size of 1×l.

[0077] The input features are processed using an edge enhancement path. The rotation-invariant Scharr edge detection operator is used to extract the initial edge response, and the edge response map is obtained by calculating the gradient magnitude. The horizontal and vertical convolution kernels of the Scharr operator are shown in the following equations: ,

[0078] Where Sx and Sy represent the parameters of their respective convolutional filters. The gradient magnitude is calculated as follows:

[0079] ,

[0080] in, Represents the initial edge response map; and These represent the input feature maps respectively. The gradient response obtained after processing by Scharr convolution kernels in the horizontal and vertical directions;

[0081] Subsequently, the initial edge response map After normalization and activation function processing, the enhanced edge response map E is obtained. edge This is added to the input features and then fed into a feature enhancement block consisting of 1×1 convolutions, 3×3 convolutions, and 1×1 convolutions for depth extraction to obtain edge enhancement features. The process is shown in the following formula:

[0082] ,

[0083] Direction-aware features and edge-enhancing features are adaptively fused using learnable weight parameters. The fused features are then integrated using a 3×3 convolution to capture local contextual information. Next, a channel attention module adaptively adjusts the weights of each channel. Finally, the features output by the channel attention module are compared with the input features. The elements are added one by one to output the final dual-path edge-guided fusion feature. This design simulates the directional scattering generated by structures such as ship hulls and decks in SAR images, enhancing the network's ability to perceive the geometric characteristics of targets while strengthening the salience of edge details.

[0084] Furthermore, the model structure reference for the multi-scale inter-group interaction module. Figure 4 It adopts a multi-branch parallel architecture, and its specific implementation process includes:

[0085] The input features are processed through three parallel branches: 1×1 convolution, 3×3 convolution, and 5×5 convolution. These branches are used for cross-channel information fusion, local structure extraction, and large receptive field context capture, respectively. The outputs of the three branches are concatenated along the channel dimension to generate multi-scale fused features. The process is shown in the following formula:

[0086] ,

[0087] in This indicates feature concatenation along the channel dimension. This indicates a convolution operation with a kernel size of 3×3. This indicates a convolution operation with a kernel size of 5×5, which fuses features across multiple scales. First, channel information is integrated through a 1×1 convolution to obtain channel integrated features. Then, the channel integrated features are input into the feature enhancement unit, and successively processed by a 3×3 depthwise separable convolution to enhance spatial details, batch normalization, and ReLU activation function to obtain enhanced multi-scale fusion features. The process is shown in the following formula:

[0088] ,

[0089] in, This indicates a 3×3 depth separable convolution operation. This indicates a batch normalization operation. This indicates the ReLU activation operation.

[0090] The enhanced multi-scale fusion features are input into the channel attention branch and the spatial attention branch for processing, respectively. The channel attention branch performs global average pooling on the enhanced multi-scale fusion features to obtain channel statistics. These channel statistics are then input into a multi-layer perceptron (MLP) to model channel dependencies, and finally, a sigmoid activation function is used to generate channel weights. The spatial attention branch uses 7×7 convolutional kernels to capture large-scale spatial dependencies, and outputs a spatial weight map after batch normalization layers and a sigmoid activation function, highlighting areas where targets may exist. Large-size convolutional kernels can effectively model the spatial relationship between ship targets and surrounding sea clutter, enhancing the perception capability of target edges and local regions.

[0091] The channel attention weights are multiplied element-wise with the multi-scale augmented features to obtain the channel augmented features; the spatial attention weights are multiplied element-wise with the input features to obtain the spatial augmented features; the channel augmented features and spatial augmented features are added element-wise, and then channel-integrated through a 1×1 convolution. Finally, the residuals are added to the module input features to output the multi-scale inter-group interaction features. The residual structure ensures stable gradient propagation while preserving the original feature information.

[0092] Furthermore, the model structure of the dual-attention residual module is referenced. Figure 6 Deployed at the end of the backbone network, it adopts a residual learning framework, and its specific process includes:

[0093] The input features are processed sequentially through a 3×3 convolution, a batch normalization layer, a ReLU activation function, another 3×3 convolution, and a batch normalization layer to obtain the basic convolutional features.

[0094] Subsequently, the basic convolutional features are input into the spatial attention submodule. The spatial attention submodule comprises two parallel paths: the first path performs global average pooling and global max pooling respectively, concatenates the resulting features, and then processes them sequentially through a 3×3 convolution, a batch normalization layer, and another 3×3 convolution; the second path processes them sequentially through global max pooling, a 3×3 convolution, a batch normalization layer, a ReLU activation function, and another 3×3 convolution. The outputs of the two paths are summed element-wise, and the spatial attention weights are generated using a sigmoid activation function. These weights are then multiplied element-wise with the input features of the spatial attention submodule to obtain the spatially enhanced features.

[0095] The spatial enhancement features are input into the channel attention submodule and weighted by channel dimension to obtain dual attention enhancement features;

[0096] Finally, the dual-attention enhanced features are added element-wise to the original input features of the dual-attention residual module to output the final dual-attention residual features.

[0097] The neck network uses a bidirectional feature fusion path of the feature pyramid network to process multi-scale feature maps and generate multi-scale fused feature maps. The adaptive high-frequency filtering module is integrated into the feature pyramid structure and performs frequency domain adaptive filtering and spatial enhancement on the feature maps in the top-down feature fusion path. The neck network mainly consists of an adaptive high-frequency filtering module (AHFS), upsampling (Resize), concatenation (Concat), and convolution module.

[0098] The neck network receives multi-scale feature maps P3, P4, and P5 output from the backbone network and processes them through a bidirectional feature fusion path from top to bottom and bottom to top. First, P5 is upsampled and concatenated with P4 along the channel dimension, then processed by an adaptive high-frequency filtering module to obtain a mesoscale fusion feature. Next, the mesoscale fusion feature is upsampled and concatenated with P3, then processed by the adaptive high-frequency filtering module to obtain the first fusion feature N3. Subsequently, N3 is downsampled through a convolutional module and concatenated with the mesoscale fusion feature, then processed by the adaptive high-frequency filtering module to obtain the second fusion feature N4. N4 is then downsampled through a convolutional module and concatenated with P5, then processed by the adaptive high-frequency filtering module to obtain the third fusion feature N5. Finally, N3, N4, and N5 are input into their respective detection heads for detecting ship targets at different scales.

[0099] The bidirectional feature fusion path can fully integrate edge and position information in shallow features with semantic information in deep features; the adaptive high-frequency filtering module is set after feature fusion at each scale to suppress low-frequency background interference such as sea clutter, enhance high-frequency discrimination features such as edge and contour of ship targets, thereby improving the detection capability of ship targets at different scales.

[0100] Furthermore, the model structure of the Adaptive High Frequency Selection (AHFS) module is referenced. Figure 5 The specific process includes:

[0101] The input features are transformed to the frequency domain using discrete cosine transform.

[0102] Design an adaptive high-pass filter whose frequency mask M is derived from learnable parameters. and Control. The process is shown in the following formula:

[0103] ,

[0104] Where M(u,v) represents the high-pass filter weight at the frequency coordinate (u,v); u and v represent the horizontal and vertical coordinates in the frequency domain, respectively; Let represent the Sigmoid activation function; D(u,v) represents the normalized distance from the frequency coordinate (u,v) to the center of the frequency domain. The position of the cutoff frequency determines the degree of suppression of low-frequency components; By controlling the smoothness of the filter band, the frequency selection part becomes continuously differentiable. This adaptive high-pass filter dynamically adjusts according to the input data during training, thereby effectively filtering out low-frequency background while preserving high-frequency details of ship targets.

[0105] The frequency domain features after adaptive high-pass filtering are input into the channel path and spatial enhancement path for parallel processing. The channel path applies global average pooling and global max pooling to the frequency domain features after adaptive high-pass filtering to compress spatial information. Then, after ReLU activation, 1×1 convolution and channel concatenation, channel attention weights are generated by sequentially applying 1×1 convolution and Sigmoid activation function. Finally, the channel attention weights are multiplied element-wise with the input features through inverse discrete cosine transform back to the spatial domain to obtain the output of the channel path.

[0106] This spatial enhancement path directly transforms the frequency domain features after adaptive high-pass filtering back to the spatial domain through inverse discrete cosine transform to obtain high-frequency response features. The high-frequency spatial response features are then processed sequentially through 3×3 depthwise separable convolution and 1×1 convolution to obtain spatial mask features. The spatial mask features are then normalized using the Sigmoid activation function to generate spatial attention weights. The generated spatial weights are then multiplied element-wise with the original input features to obtain the output of the spatial enhancement path.

[0107] The output of the channel path is added to the output of the spatial enhancement path, further integrated by 3×3 convolution, and finally summed with the input features through residual connection to output the final adaptive high-frequency filtering features.

[0108] Through the above design, the adaptive high-frequency filtering module achieves a more comprehensive and adaptive enhancement of the features of ship targets in SAR images, effectively improving feature saliency and distinguishability.

[0109] The detection head receives the multi-scale fused feature map output by the neck network, which is used to detect ship targets at different scales;

[0110] Furthermore, the multi-scale fused feature maps N3, N4, and N5 output from the FPN are input into a dedicated multi-detector head: This head comprises three detectors, corresponding to small, medium, and large target scales respectively. Each detector head has the same structure, consisting of two consecutive CBS modules and a final prediction convolutional layer. Feature map N3 serves as the input to the first detector head, specifically for detecting small-scale ship targets; feature map N4 serves as the input to the second detector head, specifically for detecting medium-scale ship targets; and feature map N5 serves as the input to the third detector head, specifically for detecting large-scale ship targets.

[0111] Furthermore, the three detection heads perform forward inference on their respective feature maps N3, N4, and N5, independently predicting candidate bounding boxes and their confidence scores. Finally, the prediction results from all scales are aggregated, and non-maximum suppression post-processing is used to filter out highly overlapping redundant boxes, outputting the final ship detection results, thus completing the end-to-end process from SAR image to target localization.

[0112] Step 3: Input the preprocessed SAR images from the training and validation sets into the edge-guided multi-scale frequency domain sensing network in Step 2, calculate the loss function and perform backpropagation, update the network parameters, and obtain the optimal parameter model.

[0113] Step 3 specifically includes:

[0114] Step 3.1: Randomly initialize the parameters of the SAR ship detection network, input the preprocessed training set and validation set data from Step 1 into the edge-guided multi-scale frequency domain sensing SAR ship detection network in Step 2, perform forward propagation, generate the predicted bounding boxes and class probabilities of ship targets, and calculate the loss.

[0115] Step 3.2: Backpropagate the loss and update the network parameters. Use minimizing the loss function as the optimization objective to obtain and save the optimal parameter model.

[0116] The loss function in step 3.2 will consist of both classification loss and regression loss. For classification, a binary cross-entropy loss function will be used, while for regression, a complete intersection over union (CIoU) loss function will be employed. CIoU loss, building upon IoU loss, considers the distance between the center points of the bounding boxes and their aspect ratio, providing more accurate bounding box regression supervision and thus improving the model's accuracy in locating ship targets.

[0117] Step 3.1 trains the network by inputting the ship images and their annotations from Step 1. The model performance is monitored on the validation set, and the network model's precision (P), recall (R), and mean average precision (mAP) are finally evaluated on the test set. The core formulas are as follows:

[0118] ,

[0119] ,

[0120] ,

[0121] In the formula, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively.

[0122] Step 4: Input the preprocessed test set into the optimal parameter model trained in Step 3, and output the detection result map of SAR ship image.

[0123] This example uses three publicly available SAR ship detection datasets: the Synthetic Aperture Radar Ship Detection Dataset (SSDD), the High-Resolution SAR Images Dataset (HRSID), and the SAR-Ship-Dataset. For data preprocessing, all images are first scaled to 640×640 pixels and their data types are converted. Then, normalization is applied to normalize pixel values ​​to the [0,1] range. To enhance the model's generalization ability and avoid overfitting, data augmentation techniques such as mosaic enhancement, random mirroring, and random rotation are used to expand the training set. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The parameter configuration for model training is shown in Table 1.

[0124] Table 1 Training parameter configuration

[0125]

[0126] To fully evaluate the performance of the model proposed in this invention, a comprehensive comparison was conducted with several mainstream target detection models. The models compared included RT-DETR, YOLOv10, YOLOv11, YOLOv12, YOLOv26, and recently proposed advanced models for SAR image or small target detection, such as AC-YOLO, EMFENet, and DEPDet. AC-YOLO was published by He et al. in *PLoS One* in 2025; EMFENet was published by Shan et al. in *IEEE Transactions on Aerospace and Electronic Systems* in 2024; and DEPDet was published by Zhang et al. in *IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing* in 2024. YOLOv11 and YOLOv26 were officially released by Ultralytics, which will launch them in 2026, while YOLOv10 and YOLOv12 were proposed by Tsinghua University and the State University of New York, respectively. Detailed comparison results are listed in Tables 2, 3 and 4.

[0127] As shown in Table 2, the model constructed in this invention achieved excellent results on multiple metrics on the SSDD dataset. In terms of accuracy, EMFA-Net outperformed comparative models such as YOLOv10, YOLOv11, YOLOv12, YOLOv26, AC-YOLO, and EMFENet with a score of 96.2%. This indicates that the DEF module, through the parallel design of the direction-aware convolution path and the Scharr edge enhancement path, effectively enhanced the extraction capability of the geometric contours and directional scattering features of ship targets, while suppressing speckle noise interference, thus effectively reducing the false detection rate in the prediction stage. Regarding the overall detection accuracy mAP0.5:0.95, EMFA-Net achieved 64.8%, superior to DEPDet's 64.4% and all YOLO series models. This demonstrates that the multi-scale inter-group interaction mechanism of the MSGI module and the frequency-domain adaptive filtering of the AHFS module work synergistically, enabling the model to maintain stable detection accuracy even under stricter IoU thresholds. In terms of mAP0.5, EMFA-Net achieved 97.0%, second only to EMFENet's 98.4%. However, EMFENet's mAP0.5:0.95 was only 59.9%, indicating a significant deficiency in its localization accuracy under different IoU thresholds. Furthermore, compared to the recently outstanding YOLOv26, this invention outperformed it by 0.5 and 2.2 percentage points in mAP0.5 and mAP0.5:0.95, respectively. In summary, on the SSDD dataset, this invention achieved excellent precision and overall detection accuracy while maintaining high recall, demonstrating its comprehensive advantages.

[0128] Table 2. Comparison of the proposed model with mainstream models on SSDD.

[0129]

[0130] The comparison results on the HRSID dataset are shown in Table 3. The model of this invention ranks first in both R and mAP0.5:0.95 metrics, and also performs excellently in P and mAP0.5 metrics. In terms of recall, EMFA-Net achieves 86.8%, significantly better than YOLOv11 (84.8%), YOLOv12 (83.9%), DEPDet (84.3%), and EMFENet (83.6%). This indicates that the edge enhancement mechanism of the DEF module and the dual attention collaborative calibration of the DARB module effectively reduce the false negative rate of small targets in complex sea conditions. In the mAP0.5:0.95 metric, EMFA-Net achieves 69.2%, an improvement of 1.2 percentage points compared to the second-best performing DEPDet model. This indicates that the AHFS module suppresses sea clutter interference through frequency-domain adaptive high-pass filtering, enabling the model to achieve more stable and accurate localization under different IoU thresholds. These results verify the effectiveness and robustness of the model of this invention in handling multi-scale targets in complex scenes.

[0131] Table 3. Comparison of the proposed model with mainstream models on HRSID.

[0132]

[0133] The comparison results on the SAR-Ship-Dataset dataset are shown in Table 4. This dataset contains 43,819 SAR images and 59,535 ship targets, demonstrating a significant scale advantage. Its resolution ranges from 3 to 25 meters, and the scenes cover nearshore, offshore, and complex port environments, including a large number of ship targets of different scales and diverse background interference. Experimental results show that the model of this invention achieves the best overall performance among all comparison methods. In terms of precision, EMFA-Net achieves 93.3%, outperforming YOLOv11 (92.4%), YOLOv12 (92.4%), EMFENet (92.8%), and DEPDet (92.2%). In terms of recall, it reaches 92.7%, outperforming YOLOv12 (92.5%) and DEPDet (91.9%). It achieves 96.6% mAP0.5, outperforming YOLOv26 (96.3%). Furthermore, it achieves 69.2% mAP0.5:0.95, a significant improvement over YOLOv26 (68.9%) and DEPDet (67.1%). These results demonstrate that the model of this invention exhibits excellent generalization ability on large-scale, multi-scene SAR datasets. The edge enhancement of the DEF module, the multi-scale inter-group interaction of the MSGI module, the frequency-domain adaptive filtering of the AHFS module, and the dual-attention collaborative calibration of the DARB module work synergistically to maintain stable detection performance under different resolutions and scene conditions.

[0134] Table 4. Comparison results of the proposed model and mainstream models on the SAR-Ship-Dataset.

[0135]

[0136] This embodiment also provides experimental detection results for a SAR ship detection method based on edge-guided multi-scale frequency domain sensing, as shown below. Figure 7 , Figure 8 , Figure 9 and Figure 10 As shown, these correspond to near-shore interference scenarios, densely packed ship scenarios, multi-scale ship scenarios, and multi-target mixed scenarios, respectively. In the comparison of detection results under different complex scenarios, the detection bounding box in this embodiment is closest to the actual annotation. In the near-shore interference scenario, this invention can accurately distinguish the target from the shoreline background. This is due to the enhanced geometric contour of the target by the dual-path edge-guided fusion module and the precise focusing of the target area by the dual-attention residual module, avoiding false detection problems caused by strong scattering echoes from port facilities. In the densely packed ship scenario, this invention can stably distinguish closely adjacent targets, verifying the effective enhancement of the discriminative features of dense targets by the synergistic effect of the multi-scale inter-group interaction module and the adaptive high-frequency filtering module. In the multi-target mixed scenario, the false alarm rate of this invention is significantly lower than that of the comparison model, proving the suppression effect of the adaptive high-frequency filtering module on low-frequency background interference such as sea clutter. The above results verify the superiority of this invention in complex background suppression and multi-scale target detection.

[0137] Example 2:

[0138] This invention provides an electronic device, including a processor and a memory. The memory stores program instructions for implementing the SAR ship detection method based on edge-guided multi-scale frequency domain sensing of Embodiment 1. The processor executes the program instructions stored in the memory to perform SAR ship detection.

[0139] The processor can also be called a Central Processing Unit (CPU). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0140] The methods described in Embodiment 1 of this invention are applicable to this electronic device embodiment. The specific functions implemented by this electronic device embodiment are the same as those in the above method embodiment, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0141] Example 3:

[0142] This embodiment provides a storage medium storing a computer program or instructions capable of implementing the above-described SAR ship detection method based on edge-guided multi-scale frequency domain sensing. When the computer program or instructions are run on a computer, the steps of the above-described SAR ship detection method based on edge-guided multi-scale frequency domain sensing are executed.

[0143] The computer program or instructions can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0144] The methods described in Embodiment 1 of this invention are all applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0145] This invention provides a SAR ship detection method based on edge-guided multi-scale frequency domain sensing. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A SAR ship detection method based on edge-guided multi-scale frequency domain sensing, characterized in that, Includes the following steps: Step 1: Obtain the SAR ship image dataset and divide it into training set, validation set and test set; Step 2: Build a SAR ship detection network model based on edge-guided multi-scale frequency domain sensing, including: backbone network, neck network and detection head; The backbone network is used to extract direction-aware multi-scale feature maps, including a dual-path edge-guided fusion module, a multi-scale inter-group interaction module, and a dual-attention residual module; the neck network uses a bidirectional feature fusion path and integrates an adaptive high-frequency filtering module to process the multi-scale feature maps and generate multi-scale fused feature maps; the detection head receives the multi-scale fused feature maps and detects ship targets at different scales; Step 3: Input the SAR ship images from the training set and the validation set into the SAR ship detection network model for training and validation, respectively. Calculate the loss function and perform backpropagation to update the network parameters to obtain the optimal parameter model. Step 4: Input the test set into the optimal parameter model and output the SAR ship detection results.

2. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 1, characterized in that, In step 2, the process of extracting multi-scale feature maps from the backbone network includes: The input image is processed sequentially using a convolutional module and a dual-path edge-guided fusion module to obtain a shallow enhanced feature map; The shallow enhanced feature map is input into the first-stage feature extraction module, and then processed sequentially by the multi-scale inter-group interaction module and the convolution module to obtain the first-stage output feature map. The first-stage output feature map is input into the second-stage feature extraction module, processed by the multi-scale inter-group interaction module to obtain the first-scale feature map, and then processed by the convolution module to obtain the second-stage output feature map. The second-stage output feature map is input into the third-stage feature extraction module, processed by the multi-scale inter-group interaction module to obtain the second-scale feature map, and then processed by the convolution module to obtain the third-stage output feature map. The third-stage output feature map is input into the fourth-stage feature extraction module, and then processed sequentially through the multi-scale inter-group interaction module, the spatial pyramid pooling module, and the dual attention residual module to obtain the fourth-stage output feature map, i.e., the third-scale feature map. The first-scale feature map, the second-scale feature map, and the third-scale feature map are combined into a multi-scale feature map, and the scale feature map is input into the neck network for processing.

3. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 2, characterized in that, The implementation process of the dual-path edge-guided fusion module for processing features includes: The input features are processed using a direction-aware convolution path. Basic features are extracted through a 3×3 standard convolution, and then horizontal and vertical scattering features are captured by 1×3 and 3×1 asymmetric convolutions. Finally, channel information is fused through a 1×1 convolution to obtain the direction-aware features. The input features are processed using an edge enhancement path. The initial edge response is extracted using the Scharr edge detection operator. The edge response map is obtained by calculating the gradient magnitude. After normalization and activation function processing, it is added to the input features and then input to the feature enhancement block for depth extraction to obtain the edge enhancement features. The feature enhancement block includes 1×1 convolution, 3×3 convolution and 1×1 convolution. The orientation-aware features and the edge enhancement features are adaptively fused using learnable weight parameters. After fusion, the features are integrated with local context information through a 3×3 convolution. Then, the weights of each channel are adjusted by a channel attention module. Finally, the features are added to the input features to output the final dual-path edge-guided fusion features.

4. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 2, characterized in that, The implementation process of the multi-scale inter-group interaction module for processing features includes: The input features are processed by three parallel branches: 1×1 convolution, 3×3 convolution, and 5×5 convolution. The outputs of the three branches are then concatenated along the channel dimension to generate multi-scale concatenated features. The multi-scale splicing features are input into the feature enhancement unit and processed sequentially through 1×1 convolution, 3×3 depthwise separable convolution, batch normalization and ReLU activation function to obtain multi-scale enhanced splicing features; The multi-scale enhanced splicing features are input to the channel attention branch and the spatial attention branch respectively for processing. The outputs of the channel attention branch and the spatial attention branch are weighted and fused. After fusion, the channels are integrated by 1×1 convolution. Finally, the residual connection is added to the input features to output the final multi-scale inter-group interaction features.

5. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 2, characterized in that, The specific process of feature processing by the dual-attention residual module includes: The input features are processed sequentially through a 3×3 convolution, a batch normalization layer, a ReLU activation function, another 3×3 convolution, and a batch normalization layer to obtain the basic convolutional features. The basic convolutional features are input into the spatial attention submodule for processing to generate spatial attention weights. The spatial attention weights are then used to weight the basic convolutional features in terms of spatial dimensions to obtain spatially enhanced features. The spatial enhancement features are input into the channel attention submodule for processing to generate channel attention weights. The channel attention weights are then used to weight the spatial enhancement features according to the channel dimension to obtain dual attention enhancement features. The dual attention enhancement features are added to the input features through a residual connection to output the final dual attention residual features.

6. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 1, characterized in that, In step 2, the specific process of the adaptive high-frequency screening module processing features includes: The input features are transformed to the frequency domain using discrete cosine transform to obtain frequency domain features; The frequency domain features are filtered using an adaptive high-pass filter to obtain the filtered frequency domain features. The filtered frequency domain features are input into the channel path for processing. Global average pooling and global max pooling are applied to the filtered frequency domain features to compress spatial information. Then, after ReLU activation, 1×1 convolution and channel concatenation, 1×1 convolution and Sigmoid activation function are used to generate channel attention weights. Then, inverse discrete cosine transform is applied back to the spatial domain, and then multiplied element-wise with the input features to obtain the output of the channel path. The filtered frequency domain features are input into the spatial enhancement path for processing. By performing an inverse discrete cosine transform (ICC) on the filtered frequency domain features back to the spatial domain, high-frequency response features are obtained. These high-frequency spatial response features are then processed sequentially through a 3×3 depthwise separable convolution and a 1×1 convolution to generate spatial mask features. These spatial mask features are then normalized using a Sigmoid activation function to generate spatial attention weights, which are then multiplied element-wise with the input features to obtain the output of the spatial enhancement path. After adding the output of the channel path and the output of the spatial enhancement path, the two paths are further integrated by 3×3 convolution. Finally, the residual connection is used to sum the input features to output the final adaptive high-frequency filtering features.

7. The SAR ship detection method based on edge-guided multi-scale frequency domain sensing according to claim 1, characterized in that, In step 3, the loss function includes classification loss and regression loss; wherein the classification loss uses the binary cross-entropy loss function, and the regression loss uses the CIoU loss function.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the SAR ship detection method based on edge-guided multi-scale frequency domain sensing as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The device stores a computer program or instructions that, when executed on a computer, perform the steps of the SAR ship detection method based on edge-guided multi-scale frequency domain sensing as described in any one of claims 1 to 7.

Images