A tail track detection model construction method based on frequency domain multi-scale perception

Abstract

The application discloses a tail trace detection model construction method based on frequency domain multi-scale perception, relates to the technical field of ocean observation, and is used for capturing weak tail trace features to perform sea surface internal wave tail trace detection.The tail trace image dataset is constructed, an RT-DETR model is taken as a baseline model to construct a tail trace detection model, a multi-scale texture information selection network is constructed to replace a backbone network of the baseline model, a frequency domain multi-scale attention perception module is constructed to replace an AIFI module of the baseline model, a gold-cane-rod convolution downsampling module is adopted to replace a standard convolution downsampling module of the baseline model, the tail trace detection model is trained and optimized based on the tail trace image dataset, and a target detection result is acquired.The tail trace detection model AUVRT-DETR constructed by the application keeps the lightweight advantage, improves the detection precision, and achieves the best balance between the detection precision and the calculation efficiency, and while realizing the highest detection precision, the model parameter quantity and the calculation cost are significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of marine observation technology, and more particularly to the field of sea surface wake detection technology, specifically a wake detection model construction method based on frequency domain multi-scale sensing. Background Technology

[0002] Fluid disturbances such as internal waves, turbulence, and Kelvin waves generated by AUVs during navigation propagate upwards through dynamic mechanisms and modulate sea surface microscale roughness, forming wavefront wakes observable by SAR. Due to SAR's advantages of all-weather, all-day imaging and high sensitivity to changes in sea surface microstructure, detecting seafront wakes using SAR images has become a key method for identifying non-cooperative underwater targets. However, AUV wakes often exhibit weak features and low contrast, and are easily obscured by complex sea conditions and environmental interference, leading to the submersion of target characteristics.

[0003] In early research, SAR wake detection relied heavily on traditional methods such as the Radon transform. These methods typically model low-level features designed manually, such as texture, scale, and orientation gradients. However, this feature engineering, which depends on prior experience, has limited applicability. Once environmental parameters change, the model needs to be redesigned, resulting in weak generalization ability, low computational efficiency, and a high false alarm rate.

[0004] With the rise of deep learning, object detection technology based on convolutional neural networks (CNNs) has made groundbreaking progress. Current CNN detectors mainly encompass two-stage methods, represented by Fast R-CNN and Faster R-CNN, and single-stage methods, represented by the YOLO series. The former ensures accuracy through candidate region generation, while the latter achieves real-time performance through end-to-end regression. For SAR wake detection, single-stage detectors rely on non-maximum suppression (NMS) post-processing, which easily leads to truncation or missed detections for long and discontinuous wake targets. Furthermore, existing lightweight models lack the ability to preserve texture details in complex sea conditions, and existing public datasets mostly focus on large-scale ship wakes, lacking high-quality SAR image data specifically for AUV wakes, severely restricting the development of this technology.

[0005] Therefore, there is an urgent need to develop a whet detection model and method that can robustly capture weak whet features in a strong noise background, so as to improve the capture effect of fine whet texture information. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a wake detection model based on frequency domain multi-scale perception, so as to solve the problem of low wake detection accuracy caused by blurred edges and loss of texture in the background of complex sea surface in the prior art.

[0007] To address the above objectives, this invention provides a method for constructing a wake detection model based on frequency domain multi-scale sensing, comprising:

[0008] S1. Construct a composite model that integrates the sea surface background and internal wave wakes, perform electromagnetic scattering simulation to generate simulated SAR images, obtain real SAR images, and fuse them with simulated SAR images to construct a wake image dataset;

[0009] S2. Construct a trail detection model using the RT-DETR model as the baseline model. Train and optimize the trail detection model based on the trail image dataset. This includes constructing a multi-scale texture information selection network to replace the backbone network of the baseline model for multi-scale feature extraction of the model input; constructing a frequency domain multi-scale attention perception module to replace the AIFI module of the baseline model for attention-based weighted processing of the output of the multi-scale texture information selection network; and using a "golden cudgel" convolutional downsampling module to replace the standard convolutional downsampling module of the baseline model to obtain target detection results.

[0010] In S1, constructing the wake image dataset includes:

[0011] S1.1 Construct a sea surface background model, generate a two-dimensional wave direction spectrum based on a unified sea surface spectrum model, sample in the two-dimensional wavenumber domain and introduce random phase, generate a sea surface height field through two-dimensional inverse Fourier transform, and generate a multi-scale sea surface structure including steady-state sea state to complex sea state based on the parameter combination of wind speed and wind direction.

[0012] S1.2, Construct an internal wave wake model, regard the underwater robot as a disturbance point source in a layered fluid system, use the point source method, solve the dispersion relation according to the linear internal wave theory, generate the spatial distribution of the internal wave field excited by the underwater robot, and simulate the internal wave wake under various navigation conditions based on the parameter combination of speed and depth.

[0013] S1.3: Obtain sea surface wave data based on the sea surface height field of S1.1, and obtain wake wave height data based on the internal wave field spatial distribution of S1.2. Linearly superimpose the wake wave height and sea surface wave to obtain a composite model that integrates sea surface features and wake features.

[0014] S1.4, the total scattering field of the composite sea surface is constructed by using a dual-scale model combined with the Kirchhoff tangent plane approximation method and the perturbation method, and the radar cross section is calculated.

[0015] S1.5 uses the radar cross section as the imaging data of the synthetic aperture radar SAR image to generate an echo signal. The range Doppler imaging algorithm is used to focus and process the echo signal to generate a simulated SAR image that fuses ideal sea conditions and internal wave wakes.

[0016] S1.6 Acquire real sea surface SAR images and perform image data fusion with simulated SAR images, including target area normalization, amplitude dynamic range adjustment and spatial geometric alignment, to obtain wake scene data with real background statistical characteristics, construct a wake image dataset and label the wakes of the image data in the dataset.

[0017] In S2, the multi-scale texture information selection network MTISNET includes an input module, at least one multi-scale texture information selection MTIS module based on cross-stage partial connectivity (CSP), and an output module.

[0018] The input module includes at least one convolutional layer for performing convolution operations on the model input image and outputting an initial feature map;

[0019] The input end of the multi-scale texture information selection module is connected to the output end of the input module, and is used to perform cross-stage connection processing and multi-scale texture information selection on the initial feature map, and output a multi-scale feature map.

[0020] The output module includes at least one convolutional layer. The input of the output module is connected to the output of the multi-scale texture information selection module, which is used to map the multi-scale feature map to the output dimension required by the target task and output the feature extraction result.

[0021] The multi-scale texture information selection module includes a segmentation module, an MTIS sub-module, and a stitching module. The segmentation module is used to segment the initial feature map into a first sub-feature map and a second sub-feature map along the channel dimension.

[0022] At least one MTIS submodule is set up, and multiple MTIS submodules are connected in series. Each MTIS submodule is set with multiple adaptive branches. By setting the local edge enhancement module HLEE in each adaptive branch, multi-scale texture feature extraction is performed on the first sub-feature map. The features extracted by each branch are spliced ​​together to obtain semantic aggregation features.

[0023] The stitching module includes at least one convolutional layer for connecting the last MTIS sub-module and the segmentation module. It stitches the second sub-feature map along the channel dimension with the first sub-feature map after processing by the MTIS sub-module to obtain a stitched feature map.

[0024] Multi-scale texture feature extraction is performed on the first sub-feature map based on the local edge enhancement module. First, the first sub-feature map is smoothed by local average pooling to obtain high-frequency residual features.

[0025] ;

[0026] In the formula, Indicates input features, Indicates high-frequency residual characteristics, Indicates to Perform average pooling;

[0027] To each Perform 3×3, 5×5 and 7×7 convolution operations, concatenate the features obtained after the three convolutions along the channel dimension, and perform channel fusion through 1×1 convolution to obtain edge fusion features;

[0028] For edge fusion features, a lightweight attention mechanism based on local importance (LIA) is introduced to calculate the pixel-level local importance of the model input:

[0029] ;

[0030] In the formula, Indicates the center pixel, Represented in pixels Weighted response within the center's neighborhood, Represents the neighborhood radius. Indicates Centered on, with The first in the neighborhood of radius A local area, , They represent The , 1 pixel, Indicates the first Weight coefficients for each local region, based on pixel fusion Obtain the local importance of the model input. ;

[0031] Introduce a gating mechanism to The first channel feature As adaptive gate calibration To obtain attention features:

[0032] ;

[0033] In the formula, For activation function, This indicates that an upsampling operation is performed using bilinear interpolation. This indicates element-wise multiplication.

[0034] By replacing the multi-head self-attention mechanism MHSA of the attention-based intra-scale feature interaction AIFI module of the target detection model RT-DETR with a polarity-aware attention mechanism that introduces linear complexity, a multi-scale attention-aware SMAP module is constructed, which includes a linear attention module, a numerical stabilization module DyT, a structure modulation module Mona, and a spectral spatial enhancement module SSEFN.

[0035] The linear attention module is used to deconstruct attention computation based on positive and negative kernel functions;

[0036] The numerical stabilization module is used to dynamically and nonlinearly transform the features;

[0037] The structure modulation module is used to perform residual modulation on features in both channel and spatial dimensions to obtain structure-modulated features.

[0038] The spectral spatial enhancement module is used to decouple the structural modulation features into two orthogonal evolution paths: a spatial sensing path and a spectral modulation path. The outputs of the two paths are fused through a hybrid domain gating mechanism to obtain the spectral spatial enhancement features.

[0039] Based on dynamic hyperbolic tangent function A dynamic activation structure is constructed to replace the normalized activation structure of the AIFI module, and a numerically stable module is built by introducing a scaling factor. Dynamically adjust the non-linear slope of attention:

[0040] ;

[0041] In the formula, and For affine parameters, Used to scale input features. Used to offset input features It is the hyperbolic tangent function. This represents a scalar used to control the shape of the Tanh function;

[0042] The output features of the numerical stabilization module are normalized to obtain stable normalized features. Residual modulation is then applied to these stable normalized features through dimensionality reduction, dimensionality increase, multi-scale depthwise convolution, and pointwise convolution to obtain modulation terms. Structural modulation features are then obtained based on these modulation terms.

[0043] ;

[0044] In the formula, Indicates structural modulation characteristics, Indicates stable normalization characteristics, This represents a dimension reduction projection. Indicates up-dimensional projection. This represents a multi-scale depthwise convolution operation. This indicates a pointwise convolution operation.

[0045] Spatial perception path first Feature dimension compression and average pooling are performed, and then the data is extracted through two consecutive spatial awareness modules. The spatial perception information is then used to restore the spatial resolution by upsampling, obtain spatial context features, and generate a spatial context feature map. Each spatial perception module consists of a convolutional layer, a normalization layer, and an activation function layer connected in sequence.

[0046] The spectral modulation path passes through 1×1 convolution and 3×3 depthwise convolution pairs in sequence. Feature processing is performed to obtain intermediate spectral modulation features. These intermediate features are then transformed to the frequency domain using a Fast Fourier Transform (FFT). In the frequency domain, the amplitude and phase of the intermediate spectral modulation features are globally modulated using learnable complex weights to obtain frequency domain features. Finally, these frequency domain features are transformed to the spatial domain using an Inverse Fast Fourier Transform (IFFT) to obtain frequency domain enhancement features, generating a frequency domain enhancement feature map.

[0047] A hybrid domain gating mechanism is introduced to concatenate spatial context features and intermediate spectral modulation features in the channel dimension to obtain gated fusion features. The gated fusion features are then transformed by 1×1 convolution and 3×3 depth convolution, and gating weights are generated by activation functions. The gating weights are then multiplied element-wise with the frequency domain enhancement features, and the feature response of the spectral modulation path is dynamically modulated based on the gating weights.

[0048] The Golden Cudgel Convolution GCConv downsampling module includes multiple parallel convolution branches for multi-path convolution processing of the input feature map;

[0049] During the training phase, each convolutional branch independently performs a convolution operation on the input feature map, and the output feature maps of each branch are concatenated along the channel dimension to generate the output feature map of the Golden Cudgel convolutional downsampling module; during the inference phase, the multiple parallel convolutional branches are fused into an equivalent single convolutional layer through convolutional reparameterization technology.

[0050] The convolutional branches sequentially include convolutional layers, batch normalization layers, and activation function layers, and the multiple parallel convolutional branches have different convolutional kernel sizes or dilatations.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] This invention proposes a multi-scale texture information selection network, which effectively extracts the corresponding edge texture features by dividing the wake features into multiple scales, thus solving the problems of blurred edges and texture loss of weak wakes against complex sea surface backgrounds.

[0053] This invention constructs a frequency domain multi-scale attention perception module, introduces a linear complexity Pola attention mechanism, and achieves multi-scale local perception and feature calibration through the Mona lightweight adapter and Dynamic Tanh (DyT). Combined with frequency domain filtering and gating mechanisms, it significantly improves the ability to extract high-frequency textures and fine-grained structures in complex scenes while maintaining real-time performance.

[0054] This invention employs a golden cudgel convolution downsampling optimization strategy. Through a multi-branch convolution structure, it can retain more key information during downsampling, enhance the feature representation ability of small and blurred targets, and effectively alleviate the information loss caused by downsampling.

[0055] The target detection model AUVRT-DETR, specifically designed for AUV internal wave wakes, provided by this invention, achieves the best balance between detection accuracy and computational efficiency. Compared to RT-DETR, the target detection model provided by this invention improves mAP@0.5 by 1.2%, mAP@0.5:0.95 by 3%, reduces the number of parameters by 22%, and reduces the computational cost by 12%. While achieving the highest detection accuracy, it significantly reduces the number of model parameters and computational overhead. Attached Figure Description

[0056] Figure 1 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 3 meters per second.

[0057] Figure 2 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 5 meters per second.

[0058] Figure 3 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 10 meters per second.

[0059] Figure 4 This is a diagram of the sea surface structure at a wind angle of 30 degrees and a wind speed of 6 meters per second.

[0060] Figure 5 This is a diagram of the sea surface structure at a wind angle of 60 degrees and a wind speed of 6 meters per second.

[0061] Figure 6 This is a diagram of the sea surface structure at a wind angle of 90 degrees and a wind speed of 6 meters per second.

[0062] Figure 7 A real aerial photograph of the wake of an inner wave on the sea surface;

[0063] Figure 8 Simulation image of the sea surface internal wave wake generated based on the composite model of S1.3;

[0064] Figure 9 This is a diagram of the AUVRT-DETR network structure provided by the present invention;

[0065] Figure 10 This is a structural diagram of the MTISNet network provided by the present invention;

[0066] Figure 11 for Figure 10 Structure diagram of the HLEE module;

[0067] Figure 12 This is a structural diagram of the SMAP module provided by the present invention;

[0068] Figure 13 for Figure 12 Structure diagram of the Mona module;

[0069] Figure 14 This is a structural diagram of the GCConv module provided by the present invention;

[0070] Figure 15 The original image of the trail to be detected provided by this invention;

[0071] Figure 16 For YOLOv8m-based models Figure 14 Image showing the results of the wake detection;

[0072] Figure 17 For the RT-DETR model Figure 14 Results of wake detection

[0073] Figure 18 For the AUVRT-DETR model to Figure 14 The result of the wake detection. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0075] A method for constructing a wake detection model based on frequency domain multi-scale sensing includes:

[0076] S1. Construct a composite model that integrates the sea surface background and internal wave wakes, perform electromagnetic scattering simulation to generate simulated SAR images, obtain real SAR images, and fuse them with simulated SAR images to construct a wake image dataset;

[0077] S2. Construct a trail detection model using the RT-DETR model as the baseline model. Train and optimize the trail detection model based on the trail image dataset. This includes constructing a multi-scale texture information selection network to replace the backbone network of the baseline model for multi-scale feature extraction of the model input; constructing a frequency domain multi-scale attention perception module to replace the AIFI module of the baseline model for attention-based weighted processing of the output of the multi-scale texture information selection network; and using a "golden cudgel" convolutional downsampling module to replace the standard convolutional downsampling module of the baseline model to obtain target detection results.

[0078] In S1, constructing the wake image dataset includes:

[0079] S1.1 Construct a sea surface background model, generate a two-dimensional wave direction spectrum based on a unified sea surface spectrum model, sample in the two-dimensional wavenumber domain and introduce random phase, generate a sea surface height field through two-dimensional inverse Fourier transform, and generate a multi-scale sea surface structure including steady-state sea state to complex sea state based on the parameter combination of wind speed and wind direction.

[0080] S1.2, Construct an internal wave wake model, regard the underwater robot as a disturbance point source in a layered fluid system, use the point source method, solve the dispersion relation according to the linear internal wave theory, generate the spatial distribution of the internal wave field excited by the underwater robot, and simulate the internal wave wake under various navigation conditions based on the parameter combination of speed and depth.

[0081] S1.3: Obtain sea surface wave data based on the sea surface height field of S1.1, and obtain wake wave height data based on the internal wave field spatial distribution of S1.2. Linearly superimpose the wake wave height and sea surface wave to obtain a composite model that integrates sea surface features and wake features.

[0082] S1.4, the total scattering field of the composite sea surface is constructed by using a dual-scale model combined with the Kirchhoff tangent plane approximation method and the perturbation method, and the radar cross section is calculated.

[0083] S1.5 uses the radar cross section as the imaging data of the synthetic aperture radar SAR image to generate an echo signal. The range Doppler imaging algorithm is used to focus and process the echo signal to generate a simulated SAR image that fuses ideal sea conditions and internal wave wakes.

[0084] S1.6 Acquire real sea surface SAR images and perform image data fusion with simulated SAR images, including target area normalization, amplitude dynamic range adjustment and spatial geometric alignment, to obtain wake scene data with real background statistical characteristics, construct a wake image dataset and label the wakes of the image data in the dataset.

[0085] In S2, the multi-scale texture information selection network MTISNET includes an input module, at least one multi-scale texture information selection MTIS module based on cross-stage partial connectivity (CSP), and an output module.

[0086] The input module includes at least one convolutional layer for performing convolution operations on the model input image and outputting an initial feature map;

[0087] The input end of the multi-scale texture information selection module is connected to the output end of the input module, and is used to perform cross-stage connection processing and multi-scale texture information selection on the initial feature map, and output a multi-scale feature map.

[0088] The output module includes at least one convolutional layer. The input of the output module is connected to the output of the multi-scale texture information selection module, which is used to map the multi-scale feature map to the output dimension required by the target task and output the feature extraction result.

[0089] The multi-scale texture information selection module includes a segmentation module, an MTIS sub-module, and a stitching module. The segmentation module is used to segment the initial feature map into a first sub-feature map and a second sub-feature map along the channel dimension.

[0090] At least one MTIS submodule is set up, and multiple MTIS submodules are connected in series. Each MTIS submodule is set with multiple adaptive branches. By setting the local edge enhancement module HLEE in each adaptive branch, multi-scale texture feature extraction is performed on the first sub-feature map. The features extracted by each branch are spliced ​​together to obtain semantic aggregation features.

[0091] The stitching module includes at least one convolutional layer for connecting the last MTIS sub-module and the segmentation module. It stitches the second sub-feature map along the channel dimension with the first sub-feature map after processing by the MTIS sub-module to obtain a stitched feature map.

[0092] Multi-scale texture feature extraction is performed on the first sub-feature map based on the local edge enhancement module. First, the first sub-feature map is smoothed by local average pooling to obtain high-frequency residual features.

[0093] ;

[0094] In the formula, Indicates input features, Indicates high-frequency residual characteristics, Indicates to Perform average pooling;

[0095] To each Perform 3×3, 5×5 and 7×7 convolution operations, concatenate the features obtained after the three convolutions along the channel dimension, and perform channel fusion through 1×1 convolution to obtain edge fusion features;

[0096] For edge fusion features, a lightweight attention mechanism based on local importance (LIA) is introduced to calculate the pixel-level local importance of the model input:

[0097] ;

[0098] In the formula, Indicates the center pixel, Represented in pixels Weighted response within the center's neighborhood, Represents the neighborhood radius. Indicated by Centered on, with The first in the neighborhood of radius A local area, , They represent The , 1 pixel, Indicates the first Weight coefficients for each local region, based on pixel fusion Obtain the local importance of the model input. ;

[0099] Introduce a gating mechanism to The first channel feature As adaptive gate calibration To obtain attention features:

[0100] ;

[0101] In the formula, For activation function, This indicates that an upsampling operation is performed using bilinear interpolation. This indicates element-wise multiplication.

[0102] By replacing the multi-head self-attention mechanism MHSA of the attention-based intra-scale feature interaction AIFI module of the target detection model RT-DETR with a polarity-aware attention mechanism that introduces linear complexity, a multi-scale attention-aware SMAP module is constructed, which includes a linear attention module, a numerical stabilization module DyT, a structure modulation module Mona, and a spectral spatial enhancement module SSEFN.

[0103] The linear attention module is used to deconstruct attention computation based on positive and negative kernel functions;

[0104] The numerical stabilization module is used to dynamically and nonlinearly transform the features;

[0105] The structure modulation module is used to perform residual modulation on features in both channel and spatial dimensions to obtain structure-modulated features.

[0106] The spectral spatial enhancement module is used to decouple the structural modulation features into two orthogonal evolution paths: a spatial sensing path and a spectral modulation path. The outputs of the two paths are fused through a hybrid domain gating mechanism to obtain the spectral spatial enhancement features.

[0107] Based on dynamic hyperbolic tangent function A dynamic activation structure is constructed to replace the normalized activation structure of the AIFI module, and a numerically stable module is built by introducing a scaling factor. Dynamically adjust the non-linear slope of attention:

[0108] ;

[0109] In the formula, and For affine parameters, Used to scale input features. Used to offset input features It is the hyperbolic tangent function. This represents a scalar used to control the shape of the Tanh function;

[0110] The output features of the numerical stabilization module are normalized to obtain stable normalized features. Residual modulation is then applied to these stable normalized features through dimensionality reduction, dimensionality increase, multi-scale depthwise convolution, and pointwise convolution to obtain modulation terms. Structural modulation features are then obtained based on these modulation terms.

[0111] ;

[0112] In the formula, Indicates structural modulation characteristics, Indicates stable normalization characteristics, This represents a dimension reduction projection. Indicates up-dimensional projection. This represents a multi-scale depthwise convolution operation. This indicates a pointwise convolution operation.

[0113] Spatial perception path first Feature dimension compression and average pooling are performed, and then the data is extracted through two consecutive spatial awareness modules. The spatial perception information is then used to restore the spatial resolution by upsampling, obtain spatial context features, and generate a spatial context feature map. Each spatial perception module consists of a convolutional layer, a normalization layer, and an activation function layer connected in sequence.

[0114] The spectral modulation path passes through 1×1 convolution and 3×3 depthwise convolution pairs in sequence. Feature processing is performed to obtain intermediate spectral modulation features. These intermediate features are then transformed to the frequency domain using a Fast Fourier Transform (FFT). In the frequency domain, the amplitude and phase of the intermediate spectral modulation features are globally modulated using learnable complex weights to obtain frequency domain features. Finally, these frequency domain features are transformed to the spatial domain using an Inverse Fast Fourier Transform (IFFT) to obtain frequency domain enhancement features, generating a frequency domain enhancement feature map.

[0115] A hybrid domain gating mechanism is introduced to concatenate spatial context features and intermediate spectral modulation features in the channel dimension to obtain gated fusion features. The gated fusion features are then transformed by 1×1 convolution and 3×3 depth convolution, and gating weights are generated by activation functions. The gating weights are then multiplied element-wise with the frequency domain enhancement features, and the feature response of the spectral modulation path is dynamically modulated based on the gating weights.

[0116] The entire process of obtaining spectral space enhancement features based on the spectral space enhancement module is as follows:

[0117] ;

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, The input characteristics of the spectral spatial enhancement module are equivalent to the structural modulation characteristics. , Represents spatially perceptible path input features. The input characteristics represent the spectrum modulation path. and express Depthwise convolution, and express convolution, Representation layer normalization, This represents two consecutive sequences of spatial sensing modules. This indicates Average Pooling. Indicates an upsampling operation. Represents spatial context features, This indicates that the GELU activation function is used. Indicates the gating weight, Represents the Fast Fourier Transform. Inverse Fast Fourier Transform Represents complex weights, This represents the bias feature map. Indicates frequency domain enhancement features, This indicates the output of the spectrum spatial enhancement module.

[0124] The Golden Cudgel Convolution GCConv downsampling module includes multiple parallel convolution branches for multi-path convolution processing of the input feature map;

[0125] During the training phase, each convolutional branch independently performs a convolution operation on the input feature map, and the output feature maps of each branch are concatenated along the channel dimension to generate the output feature map of the Golden Cudgel convolutional downsampling module; during the inference phase, the multiple parallel convolutional branches are fused into an equivalent single convolutional layer through convolutional reparameterization technology.

[0126] The convolutional branches sequentially include convolutional layers, batch normalization layers, and activation function layers, and the multiple parallel convolutional branches have different convolutional kernel sizes or dilatations.

[0127] In S1, sea surface roughness exhibits multi-scale characteristics, manifesting as large-scale waves and small-scale irregularities. Large-scale waves affect low-frequency scattering, while small-scale irregularities affect high-frequency scattering. A dual-scale model combining the Kirchhoff Approximation Method (KAM) and the Small Perturbation Method (SPM) can effectively address this problem. The sea surface is represented as an inclined, perturbed plane for electromagnetic scattering simulation, using appropriately sized discrete small elements, each considered as an inclined surface with minute undulations. The scattering effects of these small elements are categorized into coherent and incoherent scattering. Coherent scattering is calculated using KAM, while incoherent scattering is estimated using SPM. Using the calculated radar cross section (RCS) as SAR imaging data, a range-Doppler imaging algorithm is selected to focus the echo signal, yielding the final SAR image. The electromagnetic scattering simulation results are shown in Table 1.

[0128] Table 1 Simulation parameters;

[0129] .

[0130] As shown in Table 1, the signal parameters include a bandwidth of 120 MHz, corresponding to high range resolution; a pulse width of 5 microseconds, which affects the transmission energy and near-range blind zone; and a carrier frequency of 1 GHz. The platform parameters include a platform height of 5 km, a speed of 100 m / s, and a scene center distance of 11 km from the ground nadir point. These geometric and motion parameters determine the imaging swath, azimuth resolution, and Doppler parameters.

[0131] While simulated wake SAR images can effectively reflect the physical modulation characteristics of internal waves, their backgrounds are often idealized and differ from real ocean scenes in terms of scattering texture, shoreline structure, and target clutter, which hinders the generalization of wake detection models. To construct a detection dataset that more closely resembles the real-world mission environment, this invention fuses simulated wake images with publicly available real SAR background data to obtain wake scene data with realistic background statistical characteristics.

[0132] This invention selects the HRSID (High-Resolution SAR Images Dataset) as the background source. This dataset contains a large number of high-resolution marine SAR images, covering nearshore, offshore, islands and reefs, waterways, complex coastlines, and various ship distribution scenarios, providing rich and diverse realistic background textures. To maintain scene balance, this invention selects image subsets of different types, such as nearshore and offshore, and images with and without ships, according to the statistical proportions of scene categories in the HRSID dataset. This ensures that the final dataset covers typical marine monitoring scenarios and avoids detection performance deviations caused by category bias.

[0133] After background selection, an image fusion method was used to embed the simulated wake SAR image into the real background. The fusion process includes key steps such as target region normalization, amplitude dynamic range adjustment, and spatial geometric alignment, ensuring that the scattering intensity and texture of the wake appear natural in the real background without obvious edge artifacts or intensity abrupt changes. To further improve the robustness of the dataset, noise was added to the obtained dataset, and different denoising algorithms were used to obtain wake images with different signal-to-noise ratios.

[0134] The final wake detection dataset contains 2,000 images and 4,753 wake targets, including various complex sea surface backgrounds and internal wave wakes of different intensities. The wake locations are accurately labeled using labelimg, which not only preserves the consistency of the physical mechanism of the simulated wakes but also has the texture features of real SAR images, providing a reliable data foundation for the training and generalization of the AUVRT-DETR model in complex sea conditions.

[0135] Model training and testing were performed on machines equipped with NVIDIA RTX 4060 Ti and AMD Ryzen 5 7500F processors running Windows 10. All network models were built using the PyTorch 2.2.2+cu118 framework, with a virtual environment configuration created using Anaconda3. The model input image size was 640×640 pixels. The experimental environment parameters are shown in Table 2, including Python 3.10, CUDA 11.8, and cuDNN 8.9.7.

[0136] Table 2 Experimental environment parameters;

[0137] .

[0138] As shown in Table 2, the model training consisted of 400 training epochs, with a batch size of 4, a patience threshold of 30 for the early stopping mechanism, a learning rate of 0.0001, a momentum of 0.9, and the AdamW optimizer.

[0139] Precision (P), recall (R), average precision (AP), and mean average precision (mAP) are used to evaluate the model's accuracy in detecting trailing targets. mAP@0.5 and mAP@0.5:0.95 are used as the main performance indicators. mAP@0.5 represents the mean average precision when the Intersection over Union (IoU) threshold is 0.5, and mAP@0.5:0.95 represents the average mAP calculated with a step size of 0.05 as the IoU threshold increases from 0.5 to 0.95. A higher mAP@0.5 indicates better target localization, and a higher mAP@0.5:0.95 indicates better detection accuracy across various scenarios. To comprehensively evaluate the model's computational efficiency, metrics such as total parameters (Param), floating-point operations (GFLOPs), and frames per second (FPS) are also introduced. Current mainstream target detection methods are selected as benchmarks, covering classic two-stage detectors, single-stage YOLO series models, and advanced detection algorithms based on the Transformer architecture. All models were retrained and evaluated under the same experimental configuration and dataset. The detailed comparison results are shown in Table 3.

[0140] Table 3 Comparison of detection performance of different models;

[0141] .

[0142] As shown in Table 3, compared with the YOLO series detectors known for their speed, although models such as YOLOv8-M and YOLOv11-M perform well in inference speed, they are still insufficient in capturing features of weak trail targets. Specifically, YOLOv11-M has 20.1M parameters and 68.0G FLOPs, with mAP@0.5 and mAP@0.5:0.95 of 91.5% and 59.4%, respectively. In contrast, AUVRT-DETR, with only 15.5M parameters and 50.1G FLOPs, improves mAP@0.5 to 93.4% and mAP@0.5:0.95 to 63.2%. This fully demonstrates its superior ability to extract complex features under resource-constrained conditions compared to pure CNN architectures. AUVRT-DETR achieves a significant improvement in accuracy while maintaining its lightweight advantage. Therefore, AUVRT-DETR achieves the best balance between detection accuracy and computational efficiency, significantly reducing the number of model parameters and computational overhead while achieving the highest detection accuracy.

[0143] Compared to high-precision Transformer-type detectors, AUVRT-DETR has an overwhelming advantage in model complexity and inference speed. Take the currently high-performance DINO-R50 as an example; although its mAP@0.5 reaches 93.1%, its parameter count is as high as 47.6M and its FLOPs are as high as 274.0G, sacrificing massive computational resources for accuracy. AUVRT-DETR, on the other hand, surpasses DINO-R50 in mAP@0.5 while having only one-third the number of parameters and less than one-fifth the computational cost. Furthermore, compared to the equally real-time-focused RT-DETR-R18, AUVRT-DETR improves mAP@0.5 by 3.0% compared to 0.95, and also has a faster inference speed. These results demonstrate that AUVRT-DETR successfully overcomes the computational redundancy problem of traditional Transformer detectors, achieving a balance between high accuracy and high efficiency.

[0144] To comprehensively evaluate the model's performance in real-world wake detection scenarios, tests were conducted in complex nearshore environments with strong clutter, non-uniform offshore backgrounds with ship interference, and ideally unobstructed uniform offshore backgrounds. Comparisons of the detection results from YOLOv8m, RT-DETR-R18, and AUVRT-DETR revealed that in nearshore scenarios, both YOLOv8m and RT-DETR-R18 exhibited false detections due to the influence of highly reflective land echoes, resulting in a discrepancy between the detected and actual wake counts. In scenarios containing ships, YOLOv8m also generated significant false alarms. AUVRT-DETR, however, demonstrated superior detection performance, accurately identifying wakes of different sizes in complex environments with significantly higher detection confidence than other models. These results further validate the effectiveness of AUVRT-DETR in suppressing background clutter and enhancing target features, showcasing the model's excellent robustness and generalization ability.

[0145] To verify the effectiveness of MTISNet, the FMABlock module of SRConvNet, the CFBlock module of SCTNet, and the LGLBBlock module of Mobile U-ViT were integrated, and a multi-dimensional performance comparison was conducted with the standard ResNet-18 and MTISNet. The comparison results are shown in Table 4.

[0146] Table 4. Comparison of backbone network detection performance;

[0147] .

[0148] As shown in Table 4, MTISNet effectively balances the model's lightweight nature with its feature representation capabilities, significantly reducing computational costs while improving detection accuracy, thus providing a more reliable solution for target detection in complex environments.

[0149] To further evaluate the effectiveness of different enhancement strategies, two additional variants of the AIFI module were developed: SHSA using Single-Head Self-Attention, and DML incorporating a Dynamic Mixing Layer. These improved modules were compared with the baseline AIFI module in a systematic comparative experiment to comprehensively evaluate their impact on sea surface wake target detection performance. The evaluation results are shown in Table 5.

[0150] Table 5. Comparison of detection performance of different attention networks;

[0151] .

[0152] As shown in Table 5, the SMAP module significantly improves the overall detection performance of the model with only a small increase in computational overhead, achieving an optimal balance between detection accuracy and computational efficiency, thus verifying the effectiveness of the improvement strategy.

[0153] To verify the effectiveness of the GCConv downsampling module, a comparative experiment was conducted with the current mainstream convolution modules and standard convolution downsampling. The comparison results are shown in Table 6.

[0154] Table 6. Comparison of detection performance of different downsampling modules;

[0155] .

[0156] As shown in Table 6, GCConv achieves optimal detection accuracy while maintaining the same computational efficiency.

[0157] To verify the effectiveness of the proposed improvement strategy, ablation experiments were conducted on the RT-DETR baseline model based on the AUV-WDD dataset. The comparison results of the ablation experiments are shown in Table 7.

[0158] Table 7 Comparison results of ablation experiments;

[0159] .

[0160] As shown in Table 7, AUVRT-DETR achieved the best performance in precision, recall, mAP@0.5, and mAP@0.5:0.95, at 97.3%, 88.4%, 93.4%, and 63.2%, respectively. AUVRT-DETR's parameter count and GFLOPs are still significantly lower than the baseline model, and its inference speed fully meets the requirements for real-time detection. After introducing the MTISNet module separately, the model achieved a significant improvement in accuracy while greatly reducing computational costs. The number of parameters decreased from 19.9M to 14.9M, and the computational cost GFLOPs decreased from 56.9G to 49.6G, while mAP@0.5:0.95 actually increased by 2.3%. This verifies that the module can effectively model multi-scale texture information while reducing model redundancy. Its excellent performance stems from the module's efficient texture feature extraction strategy, which enhances feature representation capabilities while reducing unnecessary computation, achieving simultaneous improvement in model lightweighting and detection performance. After introducing the SMAP module, the FPS was significantly improved, increasing from 67.6 on the baseline to 91.9, while mAP@0.5:0.95 increased to 61.6%. This module replaces the traditional computationally intensive MHSA with a linear complexity attention mechanism, significantly reducing the memory pressure and latency caused by attention computation. This result shows that SMAP successfully optimizes global feature interactions within the scale, significantly improving the model's inference efficiency while maintaining the ability to capture long-range dependencies. After introducing the GCConv module, the model accuracy increased to 96.2%, with mAP@0.5 and mAP@0.5:0.95 at 92.5% and 60.7%, respectively. The FPS reached the experimental single-module peak of 98.9. This indicates that the module, through reparameterization design and multi-branch structure, effectively enhances the preservation of local details during the downsampling stage, while greatly optimizing the forward propagation computation flow, enabling it to maintain a very high frame rate even when processing high-resolution input.

[0161] In the S1 simulation, a two-dimensional wave direction spectrum was generated based on the Elfouhaily spectral model. Sampling was then performed in the two-dimensional wavenumber domain, and a random phase was introduced to satisfy the statistical characteristics of the sea surface. Finally, a two-dimensional inverse Fourier transform was used to generate the sea surface height field. To enhance data diversity, various sea surface structures, ranging from low-wind steady-state sea states to high-sea-state rough backgrounds, were generated by setting different combinations of parameters such as wind speed (1–10 m / s) and wind direction. Figure 1 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 3 meters per second. Figure 2 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 5 meters per second. Figure 3 This is a diagram of the sea surface structure at a wind angle of 10 degrees and a wind speed of 10 meters per second. Figure 4 This is a diagram of the sea surface structure at a wind angle of 30 degrees and a wind speed of 6 meters per second. Figure 5This is a diagram of the sea surface structure at a wind angle of 60 degrees and a wind speed of 6 meters per second. Figure 6 This is a diagram of the sea surface structure at a wind angle of 90 degrees and a wind speed of 6 meters per second. Figures 1 to 6 The gradient color scale represents wave height, equivalent to the distance of the sea surface relative to mean sea level, in meters; by comparison... Figure 1 , Figure 2 and Figure 3 It can be seen that as wind speed increases, sea surface roughness increases, wave amplitude increases, texture becomes more obvious, and wave shape becomes more complex; through comparison Figure 4 , Figure 5 and Figure 6 It can be seen that as the wind angle increases, the wave propagation direction becomes more inclined relative to the horizontal direction, the wave propagation direction becomes steeper, and the entire sea surface texture rotates with the wind angle.

[0162] Figure 7 This is a real aerial photograph of the wake of internal waves on the sea surface. Figure 8 The simulation image of the sea surface internal wave wake generated based on the composite model of S1.3 is compared with... Figure 7 and Figure 8 It can be seen that the internal wave wake simulation image generated based on the composite model of S1.3 shows a high degree of consistency between the texture features of the internal wave wake on the sea surface and the actual image, verifying the effectiveness of the composite model of S1.3.

[0163] Figure 9 The AUVRT-DETR network structure diagram provided by this invention is as follows: Figure 9As shown, the baseline model of AUVRT-DETR is RT-DETR. The overall structure of RT-DETR includes an encoder module, a minimum uncertainty query selection module, a decoder module, and a head network module. The encoder is used to extract multi-scale features and enhance contextual information, including a backbone network, an AIFI module, a multi-scale feature fusion FFusion module, and a cross-scale contextual feature fusion CCFF module. The minimum uncertainty query selection module is located between the encoder and the decoder and is used to improve detection accuracy and convergence speed. The decoder module is iteratively optimized through multiple Transformer decoding layers. The head network module consists of a series of convolutional layers used to obtain the target detection results. This invention improves the RT-DETR model by constructing the AUVRT-DETR wake detection model. This includes proposing a novel lightweight feature extraction network, MTISNet, to replace the backbone network of RT-DETR, enhancing the feature representation of weak edges and texture details in wakes; constructing a frequency-domain multi-scale attention-aware SMAP module to replace the AIFI module of RT-DETR, improving the model's attention to wakes; and employing the GCConv module to optimize the downsampling process of RT-DETR, suppressing detail loss and improving the feature preservation ability of small-scale wake fragments. This invention constructs the AUV-WDD wake image dataset, covering various sea conditions and wake morphologies. Based on the RT-DETR framework, it effectively integrates modules such as MTISNet, SMAP, and GCConv, achieving synergistic optimization of feature enhancement, frequency-domain modulation, and high-quality downsampling. A good balance is achieved between detection accuracy and computational cost, providing an efficient and robust solution for AUV wake detection under complex sea conditions.

[0164] Figure 10 The structural diagram of the MTISNet network provided by this invention is as follows: Figure 10As shown, the Multi-Scale Texture Information Selection Network (MTISNET) includes an input module, four CSP-MTIS modules based on cross-stage partial connectivity (CSP), and an output module. The four CSP-MTIS modules are sequentially connected by 3×3 convolutional layers. The input module includes two 3×3 convolutional layers, and the input of the first CSP-MTIS module is connected to the output of the input module. The output module includes at least one convolutional layer, and the input of the output module is connected to the output of the last CSP-MTIS module. This is used to map the multi-scale feature map to the output dimension required by the target task and output the feature extraction result. The multi-scale texture information selection module includes a segmentation module, an MTIS sub-module, and a stitching module. The segmentation module divides the initial feature map into a first sub-feature map and a second sub-feature map along the channel dimension. At least one MTIS sub-module is provided, with multiple MTIS sub-modules connected in series. Each MTIS sub-module contains multiple adaptive branches. A local edge enhancement module (HLEE) is used in each adaptive branch to extract multi-scale texture features from the first sub-feature map. The features extracted from each branch are then stitched together to obtain semantic aggregation features. The stitching module includes at least one convolutional layer connecting the last MTIS sub-module and the segmentation module. It stitches the second sub-feature map along the channel dimension with the first sub-feature map processed by the MTIS sub-module to obtain a stitched feature map. The MTIS module first constructs adaptive pooling branches at multiple scales, compressing the feature maps to different spatial sizes (3×3, 6×6, 9×9, 12×12) to obtain semantic aggregation features at different scales. Each scale branch extracts local structural information through a lightweight structure of 1×1 and 3×3 convolutions. Then, the multi-scale enhancement features are uniformly upsampled back to the original image size, enabling the network to simultaneously perceive the structural performance of the trail at different scales.

[0165] Figure 11 for Figure 10 The structural diagram of the HLEE module is as follows: Figure 11As shown, the HLEE module first smooths the input features through local average pooling, extracting stable low-frequency background components. It then subtracts these components from the original features to obtain high-frequency residual features, explicitly enhancing details such as wake edges and ripple perturbations. Next, these high-frequency residual features undergo multi-scale convolution operations (3×3, 5×5, and 7×7) to capture fine-scale perturbations, local texture changes, and larger-scale wake extension morphology, achieving multi-scale edge feature extraction from shallow to deep layers. Subsequently, the HLEE module concatenates the three multi-scale features along the channel dimension and fuses them through 1×1 convolution, compressing, recombining, and unifying cross-scale information to provide a structured edge representation for subsequent attention modeling. A lightweight LIA mechanism is introduced on the fused edge features, calculating local importance based on a region-based Softmax strategy. This calculation enhances high-frequency details, focusing on coherent and structurally consistent detail patterns in the wake region, effectively suppressing background noise and random sea clutter interference, and further amplifying the continuous perturbation structure and weak edge morphology of the wake. A lightweight LIA mechanism is introduced, which incorporates a gating mechanism, using the first channel of the input feature as an adaptive gate to calibrate local importance.

[0166] Figure 12 This is a structural diagram of the SMAP module provided by the present invention. Figure 13 for Figure 12 The diagram shows the structure of the Mona model. In the baseline model, the multi-head self-attention (MHSA) mechanism used by the AIFI module requires calculating pairwise similarities for all sequences. While effective, this process incurs significant resource overhead; computational complexity and memory requirements increase quadratically with sequence length. The quadratic complexity of MHSA severely restricts the model's deployment capability and real-time inference performance on edge platforms. Furthermore, AIFI's global interaction mode based on self-attention distributes attention evenly across all regions, further obscuring the already limited wake energy and resulting in insufficient representation of key linear perturbations. This makes it difficult to accurately model their local features. To address these limitations, a multi-scale attention-aware SMAP module is constructed, including a linear attention module (Pola), a numerical stabilization module (DyT), a structure modulation module (Mona), and a spectral spatial enhancement module (SSEFN). Figure 12 As shown, the SMAP module replaces the MHSA in the original AIFI with the PolaFormer attention mechanism (Pola) which has linear complexity. By deconstructing the attention calculation through positive and negative kernel functions, the computational and memory burden on high-resolution SAR images during the inference stage is significantly reduced. Based on this, DyT and Mona are responsible for numerical and structural compensation and stabilization of weak signals, respectively. Figure 13As shown, Mona, as a lightweight adaptation unit, achieves feature remodulation in both channel and spatial dimensions through operations such as dimensionality reduction, dimensionality increase, and multi-scale deep convolution.

[0167] Figure 14 The GCConv module structure diagram provided by this invention is as follows: Figure 14 As shown, GCConv achieves effective feature fusion through a multi-branch convolutional structure. During the training phase, GCConv expands into a multi-convolution, multi-path structure, with each convolutional branch performing convolution, batch normalization, and activation operations separately. The outputs are then superimposed along the channel dimension to serve as the module output. During the inference phase, the convolutional branches are fused into a single 3×3 convolution through convolutional reparameterization to improve computational efficiency. Compared to traditional convolutional downsampling, GCConv's multi-path design effectively expands the receptive field and enhances the ability to capture local structures in the wake, especially in long, curved wake regions, where it can retain more spatial contextual information. The module output uses the SiLU activation function to ensure non-linear expressiveness while maintaining the continuity and clarity of wake edge features, thus balancing fine-grained wake information and overall feature representation during downsampling.

[0168] Figure 15 This is the original image of the trail to be detected provided by the present invention. Figure 16 For YOLOv8m-based models Figure 15 Image showing the results of the wake detection. Figure 17 For the RT-DETR model Figure 15 Image showing the results of the wake detection. Figure 18 For the AUVRT-DETR model to Figure 15 The results of the wake detection are shown in the image. Figures 15 to 18 It can be seen that both YOLOv8m and RT-DETR-R18 have false detections, resulting in a discrepancy between the number of detected trails and the actual number of trails. AUVRT-DETR, on the other hand, shows superior detection performance, accurately identifying trails of different sizes in complex environments, and its detection confidence is much higher than that of other models.

[0169] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a wake detection model based on frequency domain multi-scale sensing, characterized in that, include: S1. Construct a composite model that integrates the sea surface background and internal wave wakes, perform electromagnetic scattering simulation to generate simulated SAR images, obtain real SAR images, and fuse them with simulated SAR images to construct a wake image dataset; S2. A trail detection model is constructed using the RT-DETR model as the baseline model. The trail detection model is trained and optimized based on the trail image dataset. This includes constructing a multi-scale texture information selection network to replace the backbone network of the baseline model for multi-scale feature extraction of the model input; constructing a frequency domain multi-scale attention perception module to replace the AIFI module of the baseline model for attention-based weighted processing of the output of the multi-scale texture information selection network; and using a "golden cudgel" convolutional downsampling module to replace the standard convolutional downsampling module of the baseline model to obtain the target detection results. By replacing the multi-head self-attention mechanism MHSA of the attention-based intra-scale feature interaction AIFI module of the target detection model RT-DETR with a polarity-aware attention mechanism that introduces linear complexity, a multi-scale attention-aware SMAP module is constructed, which includes a linear attention module, a numerical stabilization module DyT, a structure modulation module Mona, and a spectral spatial enhancement module SSEFN. The linear attention module is used to deconstruct attention computation based on positive and negative kernel functions; The numerical stabilization module is used to dynamically and nonlinearly transform the features; The structure modulation module is used to perform residual modulation on features in both channel and spatial dimensions to obtain structure-modulated features. The spectral spatial enhancement module is used to decouple the structural modulation features into two orthogonal evolution paths: a spatial sensing path and a spectral modulation path. The outputs of the two paths are fused through a hybrid domain gating mechanism to obtain the spectral spatial enhancement features. Based on dynamic hyperbolic tangent function A dynamic activation structure is constructed to replace the normalized activation structure of the AIFI module, and a numerically stable module is built by introducing a scaling factor. Dynamically adjust the non-linear slope of attention: ； In the formula, and For affine parameters, Used to scale input features. Used to offset input features It is the hyperbolic tangent function. Represents a scalar used to control the shape of the Tanh function; The output features of the numerical stabilization module are normalized to obtain stable normalized features. Residual modulation is then applied to these stable normalized features through dimensionality reduction, dimensionality increase, multi-scale depthwise convolution, and pointwise convolution to obtain modulation terms. Structural modulation features are then obtained based on these modulation terms. ； In the formula, Indicates structural modulation characteristics, Indicates stable normalization characteristics, This represents a dimension reduction projection. Indicates up-dimensional projection. This represents a multi-scale depthwise convolution operation. This indicates a pointwise convolution operation.

2. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 1, characterized in that, In S1, constructing the wake image dataset includes: S1.1 Construct a sea surface background model, generate a two-dimensional wave direction spectrum based on a unified sea surface spectrum model, sample in the two-dimensional wavenumber domain and introduce random phase, generate a sea surface height field through two-dimensional inverse Fourier transform, and generate a multi-scale sea surface structure including steady-state sea state to complex sea state based on the parameter combination of wind speed and wind direction. S1.2, Construct an internal wave wake model, regard the underwater robot as a disturbance point source in a layered fluid system, use the point source method, solve the dispersion relation according to the linear internal wave theory, generate the spatial distribution of the internal wave field excited by the underwater robot, and simulate the internal wave wake under various navigation conditions based on the parameter combination of speed and depth. S1.3: Obtain sea surface wave data based on the sea surface height field of S1.1, and obtain wake wave height data based on the internal wave field spatial distribution of S1.

2. Linearly superimpose the wake wave height and sea surface wave to obtain a composite model that integrates sea surface features and wake features. S1.4, the total scattering field of the composite sea surface is constructed by using a dual-scale model combined with the Kirchhoff tangent plane approximation method and the perturbation method, and the radar cross section is calculated. S1.5 uses the radar cross section as the imaging data of the synthetic aperture radar SAR image to generate an echo signal. The range Doppler imaging algorithm is used to focus and process the echo signal to generate a simulated SAR image that fuses ideal sea conditions and internal wave wakes. S1.6 Acquire real sea surface SAR images and perform image data fusion with simulated SAR images, including target area normalization, amplitude dynamic range adjustment and spatial geometric alignment, to obtain wake scene data with real background statistical characteristics, construct a wake image dataset and label the wakes of the image data in the dataset.

3. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 2, characterized in that, In S2, the multi-scale texture information selection network MTISNET includes an input module, at least one multi-scale texture information selection MTIS module based on cross-stage partial connectivity (CSP), and an output module. The input module includes at least one convolutional layer for performing convolution operations on the model input image and outputting an initial feature map; The input end of the multi-scale texture information selection module is connected to the output end of the input module, and is used to perform cross-stage connection processing and multi-scale texture information selection on the initial feature map, and output a multi-scale feature map. The output module includes at least one convolutional layer. The input of the output module is connected to the output of the multi-scale texture information selection module, which is used to map the multi-scale feature map to the output dimension required by the target task and output the feature extraction result.

4. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 3, characterized in that, The multi-scale texture information selection module includes a segmentation module, an MTIS sub-module, and a stitching module. The segmentation module is used to segment the initial feature map into a first sub-feature map and a second sub-feature map along the channel dimension. At least one MTIS submodule is set up, and multiple MTIS submodules are connected in series. Each MTIS submodule is set with multiple adaptive branches. By setting the local edge enhancement module HLEE in each adaptive branch, multi-scale texture feature extraction is performed on the first sub-feature map. The features extracted by each branch are spliced ​​together to obtain semantic aggregation features. The stitching module includes at least one convolutional layer for connecting the last MTIS sub-module and the segmentation module. It stitches the second sub-feature map along the channel dimension with the first sub-feature map after processing by the MTIS sub-module to obtain a stitched feature map.

5. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 4, characterized in that, Multi-scale texture feature extraction is performed on the first sub-feature map based on the local edge enhancement module. First, the first sub-feature map is smoothed by local average pooling to obtain high-frequency residual features. ； In the formula, Indicates input features, Indicates high-frequency residual characteristics, Indicates to Perform average pooling; To each Perform 3×3, 5×5 and 7×7 convolution operations, concatenate the features obtained after the three convolutions along the channel dimension, and perform channel fusion through 1×1 convolution to obtain edge fusion features; For edge fusion features, a lightweight attention mechanism based on local importance (LIA) is introduced to calculate the pixel-level local importance of the model input: ； In the formula, Indicates the center pixel, Represented in pixels Weighted response within the center's neighborhood, Represents the neighborhood radius. Indicated by Centered on, with The first in the neighborhood of radius A local area, , They represent The , 1 pixel, Indicates the first Weight coefficients for each local region, based on pixel fusion Obtain the local importance of the model input. ; Introduce a gating mechanism to The first channel feature As adaptive gate calibration To obtain attention features: ； In the formula, For activation function, This indicates that an upsampling operation is performed using bilinear interpolation. This indicates element-wise multiplication.

6. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 5, characterized in that, Spatial perception path first Feature dimension compression and average pooling are performed, and then the data is extracted through two consecutive spatial awareness modules. The spatial perception information is then used to restore the spatial resolution by upsampling, obtain spatial context features, and generate a spatial context feature map. Each spatial perception module consists of a convolutional layer, a normalization layer, and an activation function layer connected in sequence. The spectral modulation path passes through 1×1 convolution and 3×3 depthwise convolution pairs in sequence. Feature processing is performed to obtain intermediate spectral modulation features. These intermediate features are then transformed to the frequency domain using a Fast Fourier Transform (FFT). In the frequency domain, the amplitude and phase of the intermediate spectral modulation features are globally modulated using learnable complex weights to obtain frequency domain features. Finally, these frequency domain features are transformed to the spatial domain using an Inverse Fast Fourier Transform (IFFT) to obtain frequency domain enhancement features, generating a frequency domain enhancement feature map. A hybrid domain gating mechanism is introduced to concatenate spatial context features and intermediate spectral modulation features in the channel dimension to obtain gated fusion features. The gated fusion features are then transformed by 1×1 convolution and 3×3 depth convolution, and gating weights are generated by activation functions. The gating weights are then multiplied element-wise with the frequency domain enhancement features, and the feature response of the spectral modulation path is dynamically modulated based on the gating weights.

7. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 6, characterized in that, The Golden Cudgel Convolution GCConv downsampling module includes multiple parallel convolution branches for multi-path convolution processing of the input feature map; During the training phase, each convolutional branch independently performs a convolution operation on the input feature map, and the output feature maps of each branch are concatenated along the channel dimension to generate the output feature map of the Golden Cudgel convolutional downsampling module; during the inference phase, the multiple parallel convolutional branches are fused into an equivalent single convolutional layer through convolutional reparameterization technology.

8. The method for constructing a wake detection model based on frequency domain multi-scale sensing according to claim 7, characterized in that, The convolutional branches sequentially include convolutional layers, batch normalization layers, and activation function layers, and the multiple parallel convolutional branches have different convolutional kernel sizes or dilatations.

Images