Synthetic Aperture Radar Ship Detection Method Based on Dynamic Feature Enhancement

Abstract

This invention relates to a synthetic aperture radar (SAR) ship detection method based on dynamic feature enhancement, belonging to the field of image analysis technology. It includes acquiring radar images and inputting them into a feature extraction network. The invention achieves high-precision real-time detection of SAR ship images through a deep learning model architecture based on convolutional neural networks and self-attention mechanisms. The dynamic feature enhancement module combines dynamic convolutional weighting mechanisms with multi-scale deep feature extraction methods to achieve adaptive perception of ship image context information. The hierarchical feature selection module introduces perceptual features with different gradients, improving the deep learning detection model's ability to represent local and global features. The global semantic information mining module achieves global context information modeling through frequency domain adaptive enhancement and optimization of the shortcomings of traditional linear self-attention mechanisms, suppressing background regions and enhancing the ability to identify ship targets.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a synthetic aperture radar (SAR) ship detection method based on dynamic feature enhancement. Background Technology

[0002] Synthetic Aperture Radar (SAR) offers all-weather, all-day imaging capabilities, but ship detection in SAR images still faces several challenges: First, the scale of ships varies greatly, ranging from small speedboats to large cargo ships, resulting in extremely uneven pixel distribution, and small target features are easily submerged by background noise. Second, in near-shore port areas, ships are densely packed and mutually obscure each other, leading to blurred target boundaries and feature confusion. Third, strong scattering objects such as ocean clutter, port metal structures, and breakwaters form bright spots similar to ships in the image, creating strong background interference and easily causing false detections. Fourth, practical applications, such as real-time monitoring via UAVs, require detection algorithms to meet high frame rates while ensuring high accuracy, typically requiring a processing speed of ≥30 frames per second.

[0003] With the development of deep learning technology, detection methods based on convolutional neural networks have gradually become mainstream. These methods can be divided into two-stage detection algorithms and one-stage detection algorithms. Although the two-stage method has higher detection accuracy, it has a complex process, slow inference speed, and difficulty in meeting real-time requirements, thus having obvious shortcomings.

[0004] Most methods use simple splicing and addition to fuse multi-scale features, without fully considering the semantic conflicts and complementary relationships between features from different levels and sources. This results in insufficient feature representation capabilities in complex scenarios. They mainly learn to suppress background interference from training data and lack prior knowledge guidance on the imaging mechanism of synthetic aperture radar and the scattering characteristics of typical interference objects. Their generalization ability is limited in unknown scenarios and strong interference environments.

[0005] To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to design a synthetic aperture radar ship detection algorithm that can balance accuracy and speed, so as to be able to detect ships stably and meet real-time requirements.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a synthetic aperture radar ship detection method based on dynamic feature enhancement, comprising the following steps:

[0008] Step 1: The radar image is input into the feature extraction network. Through the dynamic feature enhancement module, the dynamic convolution weight mechanism and cross-scale depth features are used to extract features from the radar image. According to the extracted feature context information, the weights are dynamically generated and assigned to multiple parallel directional depth convolution branches to adaptively fuse ship outline and texture information from different directions, and output a feature map set, which includes shallow features, mid-level features and deep features.

[0009] Step 2: Extract deep features of semantic information based on the feature map set, input them into the global semantic information mining module for modeling, and use the polarity-aware attention mechanism to enhance the feature discrimination between the ship target and background interference by analyzing the interaction between positive and negative feature components in the modeled features. Combined with adaptive filtering of features in the frequency domain, enhance the low-frequency contour information related to the ship body, suppress high-frequency background noise, and output enhanced global semantic features.

[0010] Step 3: Combine shallow and mid-level features with enhanced global semantic features and input them into the hierarchical feature selection module. Calculate the similarity between the feature region and the learnable cue vector, dynamically select and fuse features that contribute to ship detection, and suppress regions judged as background to obtain optimized multi-scale fusion features.

[0011] Step 4: Input the optimized multi-scale fusion features into the detection head based on the deep learning architecture. Through the interaction between the learnable target query vector and the multi-scale fusion features, decode and generate the class label and bounding box coordinates of the ship target, realizing real-time end-to-end detection without anchor boxes and non-maximum suppression.

[0012] Furthermore, it also includes the following:

[0013] Steps one through four are executed sequentially by a pre-trained deep learning detection model. The deep learning detection model includes a dynamic feature enhancement module, a global semantic information mining module, a hierarchical feature selection module, and a detection head connected in sequence. The output of each module serves as the input of the next module, together forming a complete network for ship detection in synthetic aperture radar images.

[0014] The dynamic enhancement module in step one also includes:

[0015] The dynamic hybrid convolution module is used to perform dynamic convolution weight allocation, followed by batch normalization and activation function processing to output enhanced features;

[0016] Convolutional gated linear units are used as nonlinear activation units to receive the enhanced features output by the dynamic hybrid convolution module, and the enhanced features are nonlinearly transformed through a gating mechanism.

[0017] Furthermore, based on the extracted feature context information, weights are dynamically generated and assigned to multiple parallel directional depthwise convolutional branches. The specific process is as follows:

[0018] The input is processed through three parallel deep convolution branches, namely a regular branch, a horizontal branch, and a vertical branch, which are used to extract local neighborhood information, long-range horizontal dependencies, and spatial information in the vertical direction, respectively.

[0019] The input features are adaptively averaged pooled to obtain global channel information. Then, the dynamic weights of each branch are generated through convolutional layers. The original weight parameters corresponding to the three parallel deep convolutional branches are output and normalized using the Softmax function to obtain the dynamic weights of each branch.

[0020] The output features extracted from the regular branch, horizontal branch, and vertical branch are multiplied by the corresponding generated dynamic weights for weighted fusion. The three weighted feature maps are then added element-wise to merge them into a single feature map. After batch normalization and deactivation function processing, the enhanced features are output.

[0021] Furthermore, by employing a polarity-aware attention mechanism, the interaction between positive and negative components in the modeled features is analyzed to enhance the feature differentiation between the ship target and background interference. The specific process is as follows:

[0022] The input deep feature map is split into positive and negative feature components along the channel dimension. A sign function is used to multiply the matrix to extract all elements greater than zero in the original features to form positive feature components, and all elements less than zero are extracted and their absolute values ​​are taken to form negative feature components.

[0023] Spatial global average pooling is performed on the positive and negative feature components respectively to obtain two one-dimensional channel descriptor vectors representing global context information.

[0024] Two one-dimensional channel descriptor vectors are concatenated and fed into a fully connected neural network for interactive modeling. The network learns the non-linear relationship between global information of positive and negative values ​​and outputs a fused channel attention weight vector.

[0025] The channel attention weight vectors are applied back to the original positive and negative feature components, respectively, and the channels are recalibrated again.

[0026] The recalibrated positive and negative eigenvalues ​​are re-merged according to their original spatial locations, and then... The convolutional layers perform fusion and dimension adjustment, ultimately outputting a feature map enhanced with polarity awareness.

[0027] Furthermore, by adaptively filtering the features in the frequency domain, the low-frequency contour information related to the ship's main body is enhanced, while high-frequency background noise is suppressed. The specific process is as follows:

[0028] The feature map after polarity-aware enhancement is subjected to a fast Fourier transform in the spatial dimension to transform the features from the spatial domain to the frequency domain, and the corresponding complex spectrum is obtained, including the amplitude spectrum and the phase spectrum. The amplitude spectrum is decomposed into low-frequency amplitude components and high-frequency amplitude components.

[0029] Based on the input global context semantic information, weight parameters matching the scale of the feature map are generated. Using the generated weight parameters, a two-dimensional filtering weight matrix of the same size as the amplitude spectrum is constructed. The matrix is ​​designed to have a value greater than 1 in the low-frequency region to enhance the low-frequency information related to the ship's main outline and structure, and a value less than 1 in the high-frequency region to suppress high-frequency background interference caused by sea clutter and speckle noise.

[0030] The constructed adaptive frequency domain filter weight matrix is ​​multiplied element-wise with the decomposed low-frequency amplitude component and high-frequency amplitude component respectively to achieve differentiated and adaptive modulation of different frequency components.

[0031] The modulated low-frequency and high-frequency amplitude components are recombined and combined with the original phase spectrum to form a new filtered spectrum. An inverse fast Fourier transform is performed on the new spectrum to convert the features from the frequency domain back to the spatial domain, thereby outputting a feature map that preserves key contours while suppressing noise enhancement.

[0032] Furthermore, the output is enhanced global semantic features, and the specific output process is as follows:

[0033] The enhanced feature map, which preserves key contours while suppressing noise, is combined with the polarity-aware enhanced feature map and added element-wise to form a residual connection path. The channel dimension is then adjusted and integrated to obtain the recombined feature map.

[0034] Batch normalization is performed on the recombined feature map, and then a non-linear activation function is applied to the activated feature map. Global average pooling is then performed on the spatial dimension to generate a one-dimensional global feature vector.

[0035] Furthermore, the similarity between the feature regions and the learnable cue vectors is calculated, features contributing to ship detection are dynamically selected and fused, and regions judged as background are suppressed to obtain optimized multi-scale fused features. The specific process is as follows:

[0036] Shallow features, mid-level features, and enhanced global semantic features are aligned by channel and resolution to form a unified multi-scale feature set;

[0037] A local attention weight map is generated through a convolutional layer. The attention weight map is used to initially evaluate the local importance of each spatial location on the feature map for ship detection.

[0038] Construct a query set containing multiple learnable cue vectors, each cue vector representing a typical ship component and background pattern;

[0039] Calculate the similarity between the feature vector at each spatial location in the multi-scale feature set and all learnable cue vectors, and generate a feature selection mask;

[0040] The mask is applied to the preprocessed shallow features, mid-level features, and upsampled global semantic features respectively, to enhance features that are related to the height of the ship target and to suppress features that match the background interference pattern.

[0041] The weighted shallow features, mid-level features, and global semantic features are element-wise summed and then fused and refined using a lightweight convolutional block, ultimately outputting optimized multi-scale fused features.

[0042] Furthermore, the class labels and bounding box coordinates of the ship targets are decoded and generated, enabling real-time end-to-end detection without anchor boxes and non-maximum suppression. The specific process is as follows:

[0043] The optimized multi-scale fusion features are encoded and injected with location information to form an encoded feature sequence; a set of learnable target query vectors are initialized, and the query vectors are interacted with the encoded feature sequence through the Transformer decoder to aggregate target-related information;

[0044] Each decoded query vector is input into the classification head and regression head in parallel, and the class confidence and bounding box coordinates of the ship target are directly output.

[0045] Training and inference are performed through a one-to-one matching mechanism, enabling end-to-end real-time detection without anchor box design or non-maximum suppression post-processing.

[0046] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0047] This synthetic aperture radar (SAR) ship detection method based on dynamic feature enhancement achieves high-precision real-time detection of SAR ship images through a deep learning detection model architecture based on convolutional neural networks and self-attention mechanisms. The dynamic feature enhancement module combines dynamic convolutional weighting mechanisms with multi-scale deep feature extraction methods to achieve adaptive perception of ship image context information, thereby enhancing the ability to recognize ship target edges and texture information. The hierarchical feature selection module introduces perceptual features with different gradients and, combined with the cascaded structure of the reparameterization module, achieves adaptive fusion of local and global features, improving the deep learning detection model's ability to represent local and global features. The global semantic information mining module achieves efficient global context information modeling by frequency domain adaptive enhancement and optimization of the shortcomings of traditional linear self-attention mechanisms, suppressing background regions and enhancing the ability to recognize ship targets. Attached Figure Description

[0048] Figure 1 A schematic diagram of the network structure of the present invention is shown;

[0049] Figure 2 A schematic diagram of the detection process structure of the present invention is shown;

[0050] Figure 3 A schematic diagram of the dynamic feature enhancement module structure of the present invention is shown;

[0051] Figure 4 A schematic diagram of the network structure of the global semantic information mining module of the present invention is shown;

[0052] Figure 5 A schematic diagram of the hierarchical feature selection module structure of the present invention is shown. Detailed Implementation

[0053] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0054] Example 1:

[0055] like Figure 1-5 As shown, the synthetic aperture radar ship detection method based on dynamic feature enhancement has the following overall detection process: input synthetic aperture radar ship image, extract multi-scale features through dynamic feature enhancement module, input deep feature S5 into global semantic information mining module GSIMM for intra-scale interaction and semantic enhancement, and output enhanced global semantic features.

[0056] Shallow features S3, mid-level features S4, and enhanced global semantic features are input into the hierarchical feature selection module to perform adaptive fusion of local and global features, resulting in optimized multi-scale fusion features.

[0057] The detection head is constructed by inputting multi-scale fused features into a deep learning model architecture based on a self-attention mechanism. Through learnable query vector interaction, the system directly outputs the category and bounding box coordinates of the ship target to achieve detection.

[0058] Step 1: Multi-scale feature extraction based on dynamic feature enhancement module

[0059] The dynamic feature enhancement module, as the core of the improved backbone network, is responsible for extracting multi-scale ship features with detailed and contextual information;

[0060] Define the input ship image as Here, H and W represent the height and width of the input ship image, respectively. A convolutional module adjusts the size and number of channels of the input image. The Dynamic Feature Enhancement (DFEM) module, through dynamic convolutional weighting and cross-scale feature extraction, enables the model to capture detailed information about ships at different scales, outputting feature maps of three different scales. C represents the number of image channels, and r represents the percentage decrease in image resolution. The dynamic feature enhancement module introduces convolutional gated linear units to enhance local feature extraction capabilities and robustness while maintaining computational efficiency. For a given ship input feature map... First, a convolutional module is used to recalibrate the channel dimensions and integrate information to obtain... Output features The process is divided into two branches for parallel processing using the idea of ​​cross-stage partial connections. and The processing procedure is as follows:

[0061]

[0062] Where CNL represents the convolutional block ConvNormLayer, This indicates that the channel is divided into 1 / 2 segments. In the first branch, the feature... First pass Features are obtained after convolution and batch normalization (BatchNorm). Then, the features are processed through the Dynamic Hybrid Convolutional Module (DMCM). Dynamic deep feature extraction is performed to obtain features The features are then passed through a convolutionally gated linear unit (CGLU) to obtain the features. The processed features can enhance the focus on the contour information of the ship target and improve the training convergence ability of the model. The processing formula is shown below:

[0063]

[0064] in, express Convolution, DMCM stands for Dynamically Hybrid Convolutional Module, CGLU stands for Convolutionally Gated Linear Unit, BN stands for Batch Normalization. In the second branch, the features are processed... The convolution and SiLU activation functions are used to obtain the result. By introducing smooth nonlinearity to enhance the model's expressive power, the two branches are finally combined using matrix multiplication to obtain the enhanced output features. and through Convolution restores the output channel count to the original channel count. The Dynamic Feature Enhancement (DFEM) module provides the model with fine-grained feature selection capabilities. The processing procedure is shown in the following formula:

[0065]

[0066] Where SiLU represents a nonlinear activation function. This represents the characteristic matrix multiplication operation;

[0067] The Dynamic Hybrid Convolutional Module (DMCM) employs a dynamic convolutional weight mechanism that adaptively adjusts the receptive field distribution to capture complex features at different scales. DMCM uses depthwise convolution to extract spatial features meticulously while reducing the number of parameters and computational complexity. The DMCM consists of three stages. In the first stage, the input feature X is processed through three parallel depthwise convolutional branches: a regular branch, a horizontal branch, and a vertical branch. Each branch uses convolutional kernels of different shapes to extract different types of features. The regular branch captures local, isotropic neighborhood information in the input feature map, the horizontal branch captures long-distance dependencies of features in the horizontal direction, and the vertical branch extracts spatial information in the vertical direction. The processing is illustrated in the following formula:

[0068]

[0069] in, In the second stage of depthwise convolution, the input feature map X is compressed into its height and width using adaptive average pooling. This yields a feature vector containing global channel information, enabling the aggregation of the global context through a single... The convolutional layer expands the number of channels and generates weights for each branch. The expanded features are then rearranged dimensionally, splitting the channel dimensions into three groups, each corresponding to a branch. The weights of the three branches are normalized using the Softmax activation function, thus obtaining the dynamic importance weights of each branch across different samples and channels. The contribution of each branch is automatically adjusted based on the input features, highlighting key spatial regions and improving the flexibility of feature representation. The processing procedure is shown in the following formula:

[0070]

[0071] Where g is the global channel feature vector. The pixel value at position (i, j) in channel c of the input feature map forms the basic elements of the original feature space. The shape of each channel is real tensors, and , Let be the weight coefficients of the i-th convolutional kernel. This represents the transpose of the three branch weights, used for matrix calculations. Intermediate weights generated for convolutional layers It is shaped as The real tensor, Rearrange, represents feature rearrangement used to implement dynamic convolutional kernel weight allocation. G is the number of branches parameter; G=3 corresponds to the weights of 3 sets of dynamic convolutional kernels, determining the parallelism of the multi-branch architecture. B represents the training batch, C represents the number of image channels, and H and W represent the height and width of the feature map. This is the rearranged weight vector.

[0072] In the third stage, the output feature maps of the three branches are multiplied by the weights of the corresponding branches. Then, the three weighted features are summed to adaptively fuse feature information from three different receptive fields and shapes. The fused feature maps are then processed by batch normalization and the SiLU activation function. The processing steps are shown in the following formula:

[0073]

[0074] in, This indicates an element-wise multiplication operation. The weighted summed feature map is used as the input for batch normalization. The i-th branch represents the feature processing result of the input X. BN is batch normalization used to standardize the feature distribution and accelerate training. This is the final output feature map;

[0075] Step 2: The global semantic information mining module enhances the algorithm's ability to understand global contextual information.

[0076] The sequence with positional encoding is input into the polar linear attention module. The module splits the query vector and key vector into positive and negative signs based on element polarity, and calculates the interaction between same-sign and opposite-sign pairs. The polar linear attention module introduces a channel-level learnable power function to rescale the attention weights to reduce entropy, thereby improving the deep learning detection model's focus on ships. The processing procedure of the polar linear attention module is shown in the following formula:

[0077]

[0078] Where Q, K, and v represent the query vector, key vector, and value vector, respectively. This indicates a channel splitting operation. and These represent 1 / 2 of the v-value vector, where Used to calculate responses with the same sign. Used to calculate opposite-sign responses. This indicates a learnable power function. Matmul represents polarity splitting, and Matmul represents matrix multiplication. This indicates an element-wise multiplication operation. This indicates a channel splicing operation. A learnable polarity-aware coefficient matrix with positive signs. This represents the learnable polarity-aware coefficient matrix for negative signs, where SS is the matrix obtained through matrix multiplication. The calculated feature map, OS, is obtained through matrix multiplication. The calculated feature map, y represents the final output;

[0079] The polar linear attention module explicitly models the interaction of positive and negative values ​​through polarity perception and a learnable power function, enabling better differentiation between ships and interference objects in complex backgrounds. The output features are then enhanced by a spectrum enhancement feedforward module, which employs two parallel paths to capture spatial features at different scales. After transforming the features to the frequency domain, the module uses learnable weights and biases as adaptive filters to enhance and suppress different frequency components, distinguishing low-frequency ship contour features from high-frequency background noise. The SiLU activation function is used to control the output of one branch as the gating signal for the other branch, achieving more flexible feature fusion. The processing procedure of the spectrum enhancement feedforward module is shown in the following formula:

[0080]

[0081] in, This represents dilated depthwise convolution. Represents depthwise convolution. and This represents the learnable weight matrix. and This represents the learnable bias parameters. and These represent the Fast Fourier Transform and the Inverse Fast Fourier Transform, respectively. Synthetic Aperture Radar (SAR) ship images contain complex texture and spatial information. The global semantic information mining module helps the deep learning detection model understand the global contextual information of SAR ship images, improving the training speed of the algorithm while ensuring accuracy.

[0082] Step 3: The hierarchical feature selection module enhances the fusion capability of the deep learning detection model.

[0083] This invention proposes a novel hierarchical feature selection module, which takes input features as input. It contains a wealth of spatial details. It contains rich semantic information;

[0084] The hierarchical feature selection module will input features and Divide into three branches respectively, and use Convolution will and Adjusted to the same number of channels, the first two branches obtain attention maps containing local and global features through local and global attention modules of different block sizes. The smaller scale is used to capture fine-grained local feature interactions, which is beneficial for the hierarchical feature selection module to focus on strong scattered echoes and local contour information in synthetic aperture radar ship images. The larger scale is used to capture long-range dependencies, which helps the deep learning detection model understand the global context information of the ship and the entire sea surface, port metal equipment, and buildings. The local-global attention module uses a learnable cue vector to calculate the cosine similarity with each spatial block. The lower the cosine similarity, the more likely the spatial block is to be represented as background. The enhanced features are fused with the features of the third branch through deep convolution and ReLU activation function, and then concatenated to obtain an implicit feature map containing global context information. and ,pass Convolution reduces the number of feature map channels from 3C to C, thus reducing the computational cost of deep learning detection models. This is achieved through reparameterization modules and... Convolution yields output features During the training phase of the deep learning detection model, the reparameterization module employs a multi-branch structure to enrich the gradient flow of feature extraction, thereby enhancing the deep learning detection model's focus on the ship contours of synthetic aperture radar images. The process of the hierarchical feature selection module is shown in the following formula:

[0085]

[0086] In the above formula, This is a local-global attention module, where P=2 represents the block size, dividing the feature map into... The block, This is a local-global attention module, where P=4 represents the block size, dividing the feature map into... The block, RepConv, represents the reparameterized convolution. This indicates a channel splicing operation. This represents the activation function.

[0087] To better integrate local details with global contextual information and improve the perception ability of deep learning detection models in complex scenes, this invention proposes a novel hierarchical feature selection module.

[0088] In this scheme, the synthetic aperture radar (SAR) ship detection method based on dynamic feature enhancement achieves high-precision real-time detection of SAR ship images through a deep learning detection model architecture based on convolutional neural networks and self-attention mechanisms. The dynamic feature enhancement module combines dynamic convolutional weighting mechanisms with multi-scale deep feature extraction methods to achieve adaptive perception of ship image context information, thereby enhancing the ability to recognize ship target edges and texture information. The hierarchical feature selection module introduces perceptual features with different gradients and, combined with the cascaded structure of the reparameterization module, achieves adaptive fusion of local and global features, improving the deep learning detection model's ability to represent local and global features. The global semantic information mining module achieves efficient global context information modeling, suppresses background regions, and enhances the ability to recognize ship targets by adaptively enhancing and optimizing the traditional linear self-attention mechanism in the frequency domain.

[0089] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0090] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0091] In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; furthermore, the coupling or direct coupling or communication connection between the shown or discussed mutuals can be through some interfaces, and the indirect coupling or communication connection between the apparatus or modules can be electrical, mechanical or other forms.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A synthetic aperture radar ship detection method based on dynamic feature enhancement, characterized in that, Includes the following steps: Step 1: The radar image is input into the feature extraction network. Through the dynamic feature enhancement module, the dynamic convolution weight mechanism and cross-scale depth features are used to extract features from the radar image. According to the extracted feature context information, the weights are dynamically generated and assigned to multiple parallel directional depth convolution branches to adaptively fuse ship outline and texture information from different directions, and output a feature map set, which includes shallow features, mid-level features and deep features. Based on the extracted feature context information, weights are dynamically generated and assigned to multiple parallel directional depthwise convolutional branches. The specific process is as follows: The input features are processed through three parallel deep convolutional branches, namely a regular branch, a horizontal branch, and a vertical branch, which are used to extract local neighborhood information, horizontal long-distance dependencies, and vertical spatial information, respectively. The input features are adaptively averaged pooled to obtain global channel information. Then, the dynamic weights of each branch are generated through convolutional layers. The original weight parameters corresponding to the three parallel deep convolutional branches are output and normalized using the Softmax function to obtain the dynamic weights of each branch. The output features extracted from the regular branch, horizontal branch, and vertical branch are multiplied by the corresponding generated dynamic weights and then weighted and fused. The three weighted feature maps are then added element by element to merge them into a single feature map. After batch normalization and non-linear activation function processing, the enhanced features are output. Step 2: Deep feature extraction of semantic information is performed based on the feature map set, and the data is input into the global semantic information mining module for modeling. The interaction between positive and negative components in the features is modeled through a polarity-aware attention mechanism to enhance the feature discrimination between the ship target and background interference. The features are adaptively filtered in the frequency domain to enhance the low-frequency contour information related to the ship body, suppress high-frequency background noise, and output enhanced global semantic features. By employing a polarity-aware attention mechanism, the interaction between positive and negative components in the modeled features is analyzed to enhance the feature differentiation between the ship target and background interference. The specific process is as follows: The input deep feature map is split into positive and negative feature components along the channel dimension. A sign function is used to multiply the matrix to extract all elements greater than zero in the original features to form positive feature components, and all elements less than zero are extracted and their absolute values ​​are taken to form negative feature components. Spatial global average pooling is performed on the positive and negative feature components respectively to obtain two one-dimensional channel descriptor vectors representing global context information. Two one-dimensional channel descriptor vectors are concatenated and fed into a fully connected neural network for interactive modeling. The network learns the non-linear relationship between global information of positive and negative values ​​and outputs a fused channel attention weight vector. The channel attention weight vectors are applied back to the original positive and negative feature components, respectively, and the channels are recalibrated again. The recalibrated positive and negative feature components are merged again according to their original spatial positions, and then fused and adjusted in dimension by a 1*1 convolutional layer, finally outputting a feature map enhanced by polarity awareness. By adaptively filtering the features in the frequency domain, the low-frequency contour information related to the ship's main body is enhanced while high-frequency background noise is suppressed. The specific process is as follows: The feature map after polarity-aware enhancement is subjected to a fast Fourier transform in the spatial dimension to transform the features from the spatial domain to the frequency domain, and the corresponding complex spectrum is obtained, including the amplitude spectrum and the phase spectrum. The amplitude spectrum is decomposed into low-frequency amplitude components and high-frequency amplitude components. Based on the input global context semantic information, weight parameters matching the scale of the feature map are generated. Using the generated weight parameters, a two-dimensional filtering weight matrix of the same size as the amplitude spectrum is constructed. The matrix is ​​designed to have a value greater than 1 in the low-frequency region to enhance the low-frequency information related to the ship's main outline and structure, and a value less than 1 in the high-frequency region to suppress high-frequency background interference caused by sea clutter and speckle noise. The constructed adaptive frequency domain filter weight matrix is ​​multiplied element-wise with the decomposed low-frequency amplitude component and high-frequency amplitude component respectively to achieve differentiated and adaptive modulation of different frequency components. The modulated low-frequency and high-frequency amplitude components are recombined and combined with the original phase spectrum to form a new filtered spectrum. An inverse fast Fourier transform is performed on the new spectrum to convert the features from the frequency domain back to the spatial domain, thereby outputting an enhanced feature map that preserves key contours while suppressing noise. Step 3: Combine shallow and mid-level features with enhanced global semantic features and input them into the hierarchical feature selection module. Calculate the similarity between the feature region and the learnable cue vector, dynamically select and fuse features that contribute to ship detection, and suppress regions judged as background to obtain optimized multi-scale fusion features. The similarity between the feature region and the learnable cue vector is calculated, features that contribute to ship detection are dynamically selected and fused, and regions judged as background are suppressed to obtain optimized multi-scale fused features. The specific process is as follows: Shallow features, mid-level features, and enhanced global semantic features are aligned by channel and resolution to form a unified multi-scale feature set; A local attention weight map is generated through a convolutional layer. The attention weight map is used to initially evaluate the local importance of each spatial location on the feature map for ship detection. Construct a query set containing multiple learnable cue vectors, each cue vector representing a typical ship component and background pattern; Calculate the similarity between the feature vector at each spatial location in the multi-scale feature set and all learnable cue vectors, and generate a feature selection mask; The mask is applied to the preprocessed shallow features, mid-level features, and upsampled global semantic features respectively, to enhance features that are related to the height of the ship target and to suppress features that match the background interference pattern. The weighted shallow features, mid-level features and global semantic features are added element-wise and then fused and refined through a lightweight convolutional block to finally output the optimized multi-scale fused features. Step 4: Input the optimized multi-scale fusion features into the detection head based on the deep learning architecture. Through the interaction between the learnable target query vector and the multi-scale fusion features, decode and generate the class label and bounding box coordinates of the ship target, realizing real-time end-to-end detection without anchor boxes and non-maximum suppression.

2. The synthetic aperture radar ship detection method based on dynamic feature enhancement according to claim 1, characterized in that, Also includes the following: Steps one through four are executed sequentially by a pre-trained deep learning detection model. The deep learning detection model includes a dynamic feature enhancement module, a global semantic information mining module, a hierarchical feature selection module, and a detection head connected in sequence, which together constitute a complete network for ship detection in synthetic aperture radar images. The dynamic feature enhancement module in step one also includes: The dynamic hybrid convolution module is used to perform dynamic convolution weight allocation, followed by batch normalization and activation function processing to output enhanced features; Convolutional gated linear units are used as nonlinear activation units to receive the enhanced features output by the dynamic hybrid convolution module, and the enhanced features are nonlinearly transformed through a gating mechanism.

3. The synthetic aperture radar ship detection method based on dynamic feature enhancement according to claim 1, characterized in that, The output is enhanced global semantic features, and the specific output process is as follows: The enhanced feature map, which preserves key contours while suppressing noise, is combined with the polarity-aware enhanced feature map and added element-wise to form a residual connection path. The channel dimension is then adjusted and integrated to obtain the recombined feature map. Batch normalization is performed on the recombined feature map, and then a non-linear activation function is applied to the activated feature map. Global average pooling is then performed on the spatial dimension to generate a one-dimensional global feature vector.

4. The synthetic aperture radar ship detection method based on dynamic feature enhancement according to claim 1, characterized in that, The system decodes and generates the class labels and bounding box coordinates of ship targets, enabling real-time end-to-end detection without anchor boxes or non-maximum suppression. The specific process is as follows: The optimized multi-scale fusion features are encoded and injected with location information to form an encoded feature sequence; a set of learnable target query vectors are initialized, and the query vectors are interacted with the encoded feature sequence through the Transformer decoder to aggregate target-related information; Each decoded query vector is input into the classification head and regression head in parallel, and the class confidence and bounding box coordinates of the ship target are directly output. Training and inference are performed through a one-to-one matching mechanism, enabling end-to-end real-time detection without anchor box design or non-maximum suppression post-processing.

Images