Cross-modal sea surface target recognition method and system for high-dynamic unmanned platform

Abstract

This application relates to the field of image recognition technology, specifically providing a method and system for cross-modal sea surface target recognition for highly dynamic unmanned platforms. The method includes: constructing a cross-modal transfer recognition network based on the DenseNet Siamese network, comprising a feature extraction module, a cross-modal feature alignment module, and a classification decision module; performing adaptive shared general feature extraction and modal feature extraction for different branches in the feature extraction module to obtain a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch; minimizing the maximum mean difference between the visible light feature vector set and the SAR feature vector set to obtain aligned features and classify them to determine the target recognition category. This application effectively solves the key problems of scarce samples, difficult cross-modal transfer, and limited computational resources in sea surface target recognition by constructing a cross-modal transfer recognition network based on the DenseNet Siamese network.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a cross-modal sea surface target recognition method and system for highly dynamic unmanned platforms. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) platforms' ocean perception capabilities and the widespread application of unmanned systems, the rapid and accurate identification of sea surface targets by highly dynamic unmanned platforms has become a key technological requirement. Traditional sea surface target identification methods are mainly based on manual feature extraction and classical machine learning algorithms. These methods typically rely on the geometric, textural, and statistical features of the target, using classifiers such as support vector machines and random forests for target identification. However, manual feature extraction methods perform poorly in complex marine environments, especially under conditions of rough seas, target occlusion, or low signal-to-noise ratio, where the identification accuracy drops sharply.

[0003] Transfer learning improves the learning performance in the target domain by leveraging knowledge from the source domain, making it particularly suitable for scenarios where the target domain has few samples. Current major transfer learning methods include instance-based transfer, feature-based transfer, parameter-based transfer, and relation-based transfer. However, existing transfer learning methods still have limitations in cross-modal scenarios, especially in cross-modal transfer from visible light images to Synthetic Aperture Radar (SAR) images. Due to the significant differences in the imaging mechanisms of the two modalities, directly applying traditional transfer learning methods yields limited results. Furthermore, in practical applications, highly dynamic unmanned platforms also face strict resource constraints. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a cross-modal sea surface target recognition method and system for highly dynamic unmanned platforms. By constructing a cross-modal transfer recognition network based on the DenseNet Siamese network, it effectively solves the key problems of scarce samples, difficult cross-modal transfer, and limited computing resources in sea surface target recognition.

[0005] To achieve the objectives of this application, the following technical solution is provided:

[0006] Firstly, this application proposes a cross-modal sea surface target identification method for highly dynamic unmanned platforms, including:

[0007] A cross-modal transfer recognition network based on the DenseNet Siamese network is constructed. The cross-modal transfer recognition network includes a feature extraction module, a cross-modal feature alignment module, and a classification decision module.

[0008] First, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to the pre-trained feature extraction module via visible light image input and SAR image input, respectively. The feature extraction module performs adaptive shared general feature extraction and modal feature extraction for different branches, resulting in a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Second, the cross-modal feature alignment module minimizes the maximum mean difference between the visible light feature vector set and the SAR feature vector set to obtain aligned features. Finally, the classification decision module performs target classification on the aligned features to determine the target identification category.

[0009] Construct visible light-SAR image pairs, and perform cross-modal feature alignment training on the cross-modal transfer recognition network based on the visible light-SAR image pairs to obtain the trained cross-modal transfer recognition network;

[0010] The trained cross-modal migration recognition network is used to perform migration recognition on the airborne SAR images acquired in real time by the high-dynamic unmanned platform.

[0011] A further improvement of this application is that the feature extraction module includes a shared general feature extraction layer and a branch feature extraction layer. The shared general feature extraction layer includes four DenseNet block structures, each of which employs a bottleneck structure. The branch feature extraction layer includes a visible light branch and a SAR branch. The visible light branch and the SAR branch share the DenseNet block structures and network parameters of the shared general feature extraction layer. The adaptive shared general feature extraction and modal feature extraction for different branches in the feature extraction module, to obtain a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch, includes:

[0012] The shared general feature extraction layer is used to generate general features for images of both modalities. Based on the shared general feature extraction layer, DenseNet-121 structure is used for visible light modal feature extraction in the visible light branch, and DenseNet-69 structure is used for SAR modal feature extraction in the SAR branch. Specifically:

[0013] ;

[0014] ;

[0015] In the formula, For visible light feature vector set, For SAR feature vector set; The DenseNet-121 network is used for visible light modal feature extraction. The DenseNet-69 network is used for SAR modal feature extraction; To input the visible light mode sea surface target image of the visible light branch, The input is the SAR mode sea surface target image of the SAR branch; These are network parameters shared by the visible light branch and the SAR branch; For the mode-specific parameters of the visible light branch, Modality-specific parameters for the SAR branch;

[0016] In this embodiment, adaptive kernel convolution modules are integrated into the visible light branch and the SAR branch respectively. The adaptive kernel convolution modules adaptively adjust the convolution kernel parameters of the corresponding branches according to the local statistical characteristics of the input feature map.

[0017] A further improvement of this application is that the adaptive kernel convolution module adaptively adjusts the convolution kernel parameters of the corresponding branch according to the local statistical characteristics of the input feature map, including:

[0018] For any branch, an initial sampling position is generated for the convolutional kernel using a coordinate generation algorithm; the statistical properties of the local region of the input feature map are calculated based on the offset prediction network, specifically:

[0019] ;

[0020] ;

[0021] In the formula, This refers to the size of a local window. and Positions Local mean and variance at point; This refers to the input feature map. In spatial location The feature value at that location, i.e., the pixel / feature response value of the corresponding sampling point within that local window). It refers to the offset index of the local window in the first spatial dimension (row / vertical direction); This refers to the offset index of the local window in the second spatial dimension (column / horizontal direction);

[0022] Based on local statistical characteristics, the AKConv module generates adaptive convolutional kernel weights that fit the current region, specifically:

[0023] ;

[0024] In the formula, For position The corresponding adaptive convolution kernel weights, Based on the convolutional kernel weights, and For learnable parameters, This represents element-wise multiplication. It is the hyperbolic tangent activation function.

[0025] A further improvement of this application is that the adaptive kernel convolution module determines the local feature noise score of the input feature map, and adjusts the corresponding convolution kernel weights based on the local feature noise score to obtain the final convolution kernel weights, specifically as follows:

[0026] ;

[0027] ;

[0028] In the formula, For position The corresponding noise score, For position The local region features extracted from the input feature map, centered on the input feature map; , For high-frequency thresholds, The index of the discrete frequency component is used to indicate the number of the discrete frequency component obtained after performing a Fast Fourier Transform on the local feature block. One frequency component; The highest frequency index is the maximum frequency component of the local region's characteristics, corresponding to the Nyquist frequency or the upper limit of the discrete spectrum. For position The corresponding final convolutional kernel weights, For position The corresponding adaptive convolution kernel weights; This is the noise suppression coefficient.

[0029] A further improvement of this application is that the step of minimizing the maximum mean difference between the visible light feature vector set and the SAR feature vector set through the cross-modal feature alignment module to obtain the aligned features is specifically as follows:

[0030] ;

[0031] In the formula, For cross-modal alignment loss, Visible light feature vector set With SAR feature vector set The maximum mean difference between them; The square norm in the regenerated Hilbert space; Visible light feature vector set The total number of feature vectors contained therein; Visible light feature vector set The first in A high-dimensional feature representation of visible light feature vectors after kernel function mapping; SAR feature vector set The total number of feature vectors contained therein; SAR feature vector set The first in A high-dimensional feature representation of a SAR feature vector after kernel function mapping.

[0032] A further improvement of this application is that, during the training process of the cross-modal transfer recognition network, a composite loss function is calculated, specifically as follows:

[0033] ;

[0034] In the formula, This is a composite loss; For classifying losses, ,in, For category weights, To focus parameters, To predict probabilities; MMD loss is used for cross-modal feature alignment. CORAL loss is used to match the second-order statistical properties of different modalities. ,in, Let covariance be the feature of the source domain. Let covariance be the feature matrix of the target domain. For feature dimension, It is the Frobenius norm; For entropy regularization loss, This is a sparsity loss; , , and These are the weight coefficients corresponding to the MMD loss, the CORAL loss, the entropy regularization loss, and the sparsity loss, respectively.

[0035] A further improvement of this application is that, during the cross-modal feature alignment training of the cross-modal transfer recognition network based on the visible light-SAR image pair, a loss function is calculated;

[0036] The loss function consists of three weighted parts, as shown in the following formula:

[0037] ;

[0038] In the formula, The classification loss is used to constrain the class separability of aligned features. This is a feature alignment loss used to constrain the consistency of the distribution of visible light features and SAR features in a unified semantic space. These are the weight coefficients corresponding to the feature loss function; To compare learning loss, To compare the weight coefficients corresponding to the learning loss;

[0039] ;

[0040] In the formula, For cosine similarity functions, For temperature parameters, It is a natural exponential function. Indicates the first The feature vectors obtained from the visible light images by the feature extraction module This represents the feature vector obtained from the SAR image paired with it.

[0041] A further improvement of this application is that, after training the cross-modal transfer recognition network on cross-modal feature alignment based on the visible light-SAR image, the method further includes: performing end-to-end fine-tuning of the cross-modal transfer recognition network trained on cross-modal feature alignment using labeled data from the target domain dataset; the total loss function in the fine-tuning stage is:

[0042] ;

[0043] In the formula, This is the total loss function during the fine-tuning phase. For classifying losses, For regularization loss, The weighting coefficients for the regularization loss.

[0044] A further improvement of this application is that it further includes: scaling the pixel values ​​of the input initial visible light image to the [0,1] interval using standard normalization operations to obtain the preprocessed visible light image; and processing the input initial SAR image using logarithmic transformation and adaptive histogram equalization to obtain the preprocessed SAR image; specifically: assuming the input initial SAR image is... The preprocessed SAR image for:

[0045] ;

[0046] In the formula, The input is the initial SAR image. This is the preprocessed SAR image. Scaling factor and These are the mean and standard deviation of the image after logarithmic transformation, respectively.

[0047] Secondly, this application provides a cross-modal sea surface target recognition system for highly dynamic unmanned platforms, used to implement the aforementioned cross-modal sea surface target recognition method for highly dynamic unmanned platforms, comprising:

[0048] A cross-modal transfer recognition network construction unit is used to construct a cross-modal transfer recognition network based on the DenseNet Siamese network. The cross-modal transfer recognition network includes a feature extraction module, a cross-modal feature alignment module, and a classification decision module.

[0049] First, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input into the pre-trained feature extraction module via a first image input terminal and a second image input terminal, respectively. The feature extraction module performs adaptive shared general feature extraction and modal feature extraction for different branches, resulting in a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Second, the cross-modal feature alignment module minimizes the maximum mean difference between the visible light feature vector set and the SAR feature vector set to obtain aligned features. Finally, the classification decision module performs target classification on the aligned features to determine the target identification category.

[0050] A cross-modal transfer recognition network training unit is used to construct visible light-SAR image pairs and perform cross-modal feature alignment training on the cross-modal transfer recognition network based on the visible light-SAR image pairs to obtain the trained cross-modal transfer recognition network.

[0051] A cross-modal migration recognition unit is used to perform migration recognition on airborne SAR images acquired in real time by the highly dynamic unmanned platform using the trained cross-modal migration recognition network.

[0052] Compared with the prior art, this application has the following beneficial effects:

[0053] The cross-modal sea surface target recognition method and system for highly dynamic unmanned platforms provided in this application constructs a cross-modal transfer recognition network based on the DenseNet Siamese network. In this network, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to a pre-trained feature extraction module via visible light image input and SAR image input, respectively. Different network depths and connection methods are used in the feature extraction module to adaptively extract shared general features and modal features for different branches, based on the different characteristics of the visible light and SAR images. This yields a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Then, based on the cross-modal feature alignment module, the maximum mean difference between the visible light and SAR feature vector sets is minimized. A learnable mapping function projects the features of different modalities onto a unified semantic space, achieving effective cross-modal transfer. This effectively solves the key problems of scarce samples, difficult cross-modal transfer, and limited computational resources, providing an efficient, accurate, and robust sea surface target recognition solution for highly dynamic unmanned platforms. Attached Figure Description

[0054] The accompanying drawings are provided to further understand this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof.

[0055] Figure 1 A schematic diagram of an optional process for a cross-modal sea surface target recognition method for a highly dynamic unmanned platform provided in an embodiment of this application;

[0056] Figure 2 A schematic diagram of the model architecture of the cross-modal transfer recognition network provided in an embodiment of this application;

[0057] Figure 3 This is a schematic diagram illustrating the structure and function of the AKConv module provided in an embodiment of this application;

[0058] Figure 4 A schematic diagram of the training and testing loss curves of the cross-modal transfer recognition network provided in the embodiments of this application;

[0059] Figure 5 A schematic diagram of the training and testing accuracy curves of the cross-modal transfer recognition network provided in the embodiments of this application;

[0060] Figure 6 Example 1 image of a simulated UAV-borne SAR Cargo target provided in this application embodiment;

[0061] Figure 7Example 2 image of a simulated UAV-borne SAR Cargo target provided in this application embodiment;

[0062] Figure 8 Example 3 image of a simulated UAV-borne SAR Cargo target provided in this application embodiment;

[0063] Figure 9 Example 1 image of a Tanker target simulating an unmanned aerial vehicle (UAV) SAR provided in this application embodiment;

[0064] Figure 10 Example 2 image of a Tanker target simulating an unmanned aerial vehicle (UAV) SAR provided in this application embodiment;

[0065] Figure 11 Example 3 image of a Tanker target simulating an unmanned aerial vehicle (UAV) SAR provided in this application embodiment;

[0066] Figure 12 Example 1 image of another target from a simulated UAV-borne SAR provided in this application embodiment;

[0067] Figure 13 Example 2 image of another target from a simulated UAV-borne SAR provided in this application embodiment;

[0068] Figure 14 Example 3 image of another target from a simulated UAV-borne SAR provided in this application embodiment;

[0069] Figure 15 The result curve of the cross-modal transfer recognition network model provided in the embodiments of this application with a training sample size of 1200;

[0070] Figure 16 The test accuracy curve of the cross-modal transfer recognition network model provided in the embodiments of this application with a training sample size of 1200. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0072] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature; in the description of this application, unless otherwise stated, "multiple" means two or more.

[0073] With the rapid development of maritime perception modes by unmanned aerial vehicle (UAV) platforms and the widespread application of unmanned systems, the rapid and accurate identification of sea surface targets by highly dynamic unmanned platforms has become a key technological requirement. When performing maritime missions, UAVs and other highly dynamic unmanned platforms need to quickly identify and classify various sea surface targets in complex marine environments, including different types of ships, buoys, and marine platforms. However, due to limitations in platform payload capacity, cost control, and environmental factors, existing sea surface target identification technologies face numerous challenges.

[0074] Traditional methods for identifying sea surface targets primarily rely on manual feature extraction and classical machine learning algorithms. These methods typically depend on the target's geometric, textural, and statistical features, using classifiers such as support vector machines and random forests for target recognition. However, manual feature extraction methods perform poorly in complex marine environments, especially under conditions of rough seas, target occlusion, or low signal-to-noise ratio, where recognition accuracy drops sharply. In recent years, deep learning technology has made breakthroughs in target recognition. Convolutional Neural Networks (CNNs), with their powerful feature learning capabilities, have performed exceptionally well in various visual tasks. However, effective training of deep learning models requires a large amount of labeled data, and obtaining real training samples for highly dynamic unmanned platforms faces significant challenges: on the one hand, actual flight tests are costly, and the number of samples obtained is extremely limited; on the other hand, images acquired by highly dynamic platforms often have low resolution, unclear target features, and are difficult and error-prone to manual annotation. To address the problem of insufficient training data, transfer learning techniques have emerged. Transfer learning improves the learning effect in the target domain by utilizing knowledge from the source domain, making it particularly suitable for scenarios where target domain samples are scarce. Current major transfer learning methods include instance-based transfer, feature-based transfer, parameter-based transfer, and relation-based transfer. However, existing transfer learning methods still have limitations in the application of cross-modal scenarios, especially in cross-modal transfer from visible light images to SAR images. Due to the significant differences in the imaging mechanisms of the two modalities, the direct application of traditional transfer learning methods has limited effectiveness.

[0075] Meanwhile, in practical applications, highly dynamic unmanned platforms also face strict resource constraints. These platforms typically have limited computing resources, requiring target recognition algorithms to have low computational complexity and memory footprint. Traditional deep learning models often have a large number of parameters, making it difficult to run in real time in resource-constrained environments. Although lightweight techniques such as model compression and knowledge distillation can reduce model size to some extent, they often come at the cost of sacrificing recognition accuracy.

[0076] To address the aforementioned technical problems, this application proposes the following technical solutions and corresponding embodiments.

[0077] The following is combined with Figures 1 to 16 The embodiments shown illustrate the technical solutions of this application:

[0078] Example 1

[0079] This application provides a cross-modal sea surface target recognition method for highly dynamic unmanned platforms, applied to UAVs (unmanned aerial vehicles), with reference to... Figure 1 As shown, the steps include S101 to S103 as follows:

[0080] Step S101: Construct a cross-modal transfer recognition network based on the DenseNet Siamese network. The cross-modal transfer recognition network includes a feature extraction module, a cross-modal feature alignment module, and a classification decision module.

[0081] First, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to a pre-trained feature extraction module via visible light image input and SAR image input, respectively. The feature extraction module performs adaptive shared general feature extraction and modal feature extraction for different branches, resulting in a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Second, a cross-modal feature alignment module minimizes the maximum mean difference between the visible light feature vector set and the SAR feature vector set to obtain aligned features. Finally, a classification decision module performs target classification on the aligned features to determine the target identification category.

[0082] In this embodiment, the network architecture of the cross-modal transfer recognition network is a cross-modal transfer learning model based on the improved DenseNet (Dense Convolutional Network) Siamese network architecture. It is mainly applied to the automatic identification system of sea surface targets of highly dynamic unmanned platforms such as UAVs, and realizes cross-modal target identification from visible light images to Synthetic Aperture Radar (SAR) images.

[0083] Reference Figure 2 As shown, the cross-modal migration recognition network of this embodiment further includes a data preprocessing module and a post-processing module. The data preprocessing module is used to standardize the visible light image and SAR image input to the cross-modal migration recognition network, respectively. In this embodiment, after inputting the visible light image and SAR image through two input terminals, for the visible light image, a standard normalization operation is used to scale the pixel values ​​of the visible light image to the [0,1] range; for the SAR image, logarithmic transformation and adaptive histogram equalization are used to solve the problems of weak target detail loss due to the large dynamic range of the SAR image and insufficient contrast under speckle noise interference. Specifically, let the input SAR image be... The preprocessed image is:

[0084] ;

[0085] In the formula, The input SAR image, The image after preprocessing. Scaling factor and These are the mean and standard deviation of the image after logarithmic transformation, respectively.

[0086] In this embodiment, the feature extraction module comprises two parts: shared general feature extraction and modal feature extraction for different branches. In the shared general feature extraction part, a weight-sharing approach is used for the Siamese network. Specifically, for the SAR image and visible light image processed by the data preprocessing module, the same DenseNet block structure is used in the shared general feature extraction part to extract general grayscale, target shape structure, and other features of the SAR image and visible light image. As a feasible implementation, the shared general feature extraction part includes four identical DenseNet block (dense block) structures. Each dense block employs a bottleneck structure, i.e., a combination of 1×1 convolution → 3×3 convolution. In each dense block, a modality-aware feature selection mechanism is introduced. This mechanism adjusts the feature extraction process for different modal data through learnable weight parameters, specifically:

[0087] ;

[0088] In the formula, This represents any modality type, which includes visible light image type and SAR image type; These are the weight parameters for the modality-aware convolutional layer; Indicates the first One convolutional layer, This represents the feature links of all preceding layers; Modal type The final output features of the lower dense block.

[0089] In the modal feature extraction part for different branches, the Siamese network in this embodiment sets different branch depths for different modal types. Specifically, the visible light branch adopts the complete DenseNet-121 structure, while the SAR branch adopts the DenseNet-69 structure to reduce the amount of computation, considering that the features of SAR images are relatively simple. The feature extraction process for the two branches is as follows:

[0090] ;

[0091] ;

[0092] In the formula, The visible light feature vector set (visible light feature vector set) is the output of the visible light image after feature extraction. The feature vector set (SAR feature vector set) is the output of a SAR image after feature extraction for the SAR branch. The DenseNet-121 network is used for feature extraction from visible light images. The DenseNet-69 network is used for feature extraction in SAR images; To input a visible light branch of the visible light mode image of the sea surface target, The input is a SAR mode image of the sea surface target from the SAR branch; Network parameters shared by the visible light branch and the SAR branch; For the mode-specific parameters of the visible light branch, These are mode-specific parameters for the SAR branch.

[0093] Here, in the different branch structures of the cross-modal transfer recognition network in this embodiment, an Adaptive Kernel Convolution (AKConv) module is integrated. In different branch paths (shared general feature extraction + branch modal feature extraction), the convolution kernel parameters are adaptively adjusted according to the local characteristics of the input features. Specifically, refer to... Figure 3 As shown, the AKConv module takes the input feature map as input. As input, the following steps are executed sequentially: coordinate generation algorithm → offset prediction network → adaptive sampling position calculation → feature resampling → irregular convolution operation, to obtain the output feature map. .

[0094] Specifically, the AKConv module is for any output position The standard rule convolution kernel The sampling grid is used as the initial sampling template. Let the convolution kernel's... Each sampling point corresponds to a two-dimensional index. ,in:

[0095] ;

[0096] ;

[0097] The initial sampling position at this location is defined as:

[0098] ;

[0099] in, The expansion coefficient is the coefficient used when the convolutional layer has a stride. hour, It can be equivalently represented in the coordinate system of the input feature map. The coordinate generation algorithm described above can generate a set of regular sampling points at each spatial location, consistent with that of regular convolution, providing an initial sampling reference for subsequent offset prediction and irregular convolution.

[0100] After generating initial sampling positions for the convolutional kernels using the aforementioned coordinate generation algorithm, an offset prediction network is first defined to predict the two-dimensional offset to be applied to each initial sampling point based on the local structure and texture changes of the input feature map, thereby achieving adaptive sampling. The offset prediction network employs a lightweight convolutional sub-network, with a structure consisting of two convolutional layers combined with batch normalization (BN) and ReLU (Rectified Linear Unit) non-linear activation layers. The offset prediction network calculates the statistical characteristics of local regions of the input feature map.

[0101] ;

[0102] ;

[0103] In the formula, This refers to the size of a local window. and Positions Local mean and variance at point; This refers to the input feature map. In spatial location The feature value at that location, i.e., the pixel / feature response value of the corresponding sampling point within that local window). It refers to the offset index of the local window in the first spatial dimension (row / vertical direction); This refers to the offset index of the local window in the second spatial dimension (column / horizontal direction);

[0104] Based on local statistical characteristics, the AKConv module generates adaptive convolutional kernel weights that fit the current region, specifically:

[0105] ;

[0106] In the formula, For position The corresponding adaptive convolution kernel weights, Based on the convolutional kernel weights, and For learnable parameters, This indicates element-wise multiplication.

[0107] Meanwhile, to enhance the noise immunity of the AKConv module, this embodiment identifies and suppresses noise components by analyzing the frequency domain characteristics of local features, specifically:

[0108] ;

[0109] In the formula, For position The corresponding noise score, This represents the Fast Fourier Transform. For high-frequency thresholds, The frequency index, i.e., the sequence number of the discrete frequency component, is used to represent the number obtained after performing a Fast Fourier Transform on the local feature block. One frequency component; For the highest frequency index, The index of the highest frequency component in the spectrum, corresponding to the upper limit of the Nyquist frequency; It refers to the location The local region features are extracted from the input feature map with the center as the center.

[0110] Based on the above noise score, the response of the convolution kernel is adjusted in the AKConv module, specifically as follows:

[0111] ;

[0112] In the formula, For position The corresponding final convolutional kernel weights, For position The corresponding adaptive convolution kernel weights; This is the noise suppression coefficient; For position The corresponding noise score.

[0113] Therefore, based on the adaptively adjusted convolutional kernel weights, the weights of the noisy region are selectively attenuated by controlling the noise suppression intensity according to the noise score of the current region, thereby meeting the noise suppression requirements for cross-modal feature extraction.

[0114] Furthermore, considering the computational resource limitations of highly dynamic unmanned platforms, this embodiment designs a lightweight variant of AKConv (AKConv-Lite). This variant reduces computational overhead through the following strategies: first, it uses grouped convolutions and depthwise separable convolutions to reduce the number of parameters; second, it uses low-rank decomposition to approximate adaptive weight calculation; and third, it introduces quantized perceptual training to support fixed-point inference. The lightweight version reduces the number of parameters and computational cost by approximately 60% and 70%, respectively, while the recognition accuracy decreases by no more than 2%.

[0115] In this embodiment, a dual attention mechanism combining channel attention and spatial attention is designed in the branch feature extraction layer to adaptively select important features. Specifically, the importance weight of each channel is calculated through the channel attention module:

[0116] ;

[0117] In the formula, Here is the channel attention weight vector. and These represent the input feature maps respectively. Perform global average pooling and global max pooling. It is a multilayer perceptron (MLP). This is the Sigmoid activation function.

[0118] The importance weights of spatial locations are calculated using the spatial attention module.

[0119] ;

[0120] In the formula, This is the spatial attention weight vector. and These represent the input feature maps respectively. Perform average pooling and max pooling along the channel dimension; for Convolution operations of different sizes; This is the Sigmoid activation function.

[0121] The final feature is represented as: In the formula, As the final feature, Here is the channel attention weight vector. This is the spatial attention weight vector. This is the input feature map.

[0122] In this embodiment, a cross-modal feature alignment module is designed. This module achieves feature alignment by minimizing the maximum mean discrepancy (MMD) of different modal feature distributions, thus realizing effective cross-modal transfer. Specifically:

[0123] ;

[0124] In the formula, For cross-modal alignment loss, Visible light feature vector set With SAR feature vector set The maximum mean difference between them; The square norm in the regenerated Hilbert space; Visible light feature vector set The total number of feature vectors contained therein; Visible light feature vector set The first in A high-dimensional feature representation of visible light feature vectors after kernel function mapping; SAR feature vector set The total number of feature vectors contained therein; SAR feature vector set The first in A high-dimensional feature representation of a SAR feature vector after kernel function mapping.

[0125] In this embodiment, the features extracted and aligned across modalities are input into a classifier for target classification, outputting target class probabilities (classification decision module). This embodiment uses a fully connected layer to make decisions on the extracted features, outputting a classification label for each target. The classifier's output is normalized using a Softmax layer to ensure that the sum of probabilities for each class is 1. In this embodiment, the classification decision module is responsible for classifying the aligned features output by the cross-modal feature alignment module; specifically:

[0126] Input: Aligned features output by the cross-modal feature alignment module;

[0127] Network structure:

[0128] (a) Fully Connected Layer: Maps features to the number of categories The dimension is used to output logits.

[0129] (b) Softmax activation function: transforms logits into a probability distribution, as shown in the formula:

[0130] ;

[0131] In the formula, Input features; Given input features At that time, the input feature The sample category label belongs to the category The probability of; For the first Logits for each category; It is a natural exponential function; This is the sum of the Natural Index scores for all categories;

[0132] Output: The probability value for each category, in the following format: , This represents the confidence level that the input image belongs to each category.

[0133] To improve the robustness and accuracy of the recognition results, a post-processing module is included in the cross-modal transfer recognition network of this embodiment, especially for the complex marine environment of highly dynamic unmanned platforms. The specific design of the post-processing module in this embodiment is as follows:

[0134] After obtaining the probability vector (probability value) output by the classification decision module, the post-processing module performs confidence filtering, temporal integration, and contextual consistency checks. Specifically, by setting a confidence threshold, only categories with a maximum probability higher than the confidence threshold are retained; otherwise, they are marked as "unknown" or "background." For example, the confidence threshold is 0.8.

[0135] For sequential images (multiple consecutive frames), a sliding window average or weighted average is used to smooth the classification results and reduce instantaneous errors; the formula is:

[0136]

[0137] In the formula, The smoothed probability. The probability of the current frame. For historical integration probability, The attenuation factor is, for example, 0.7.

[0138] In the training process of the cross-modal transfer recognition network, this embodiment designs a loss function and optimization strategy for the cross-modal transfer task. Taking into account classification accuracy, cross-modal consistency, and model complexity, a composite loss function is designed, specifically as follows:

[0139] ;

[0140] In the formula, This is a composite loss; For classification loss, for a single sample Its real label is The model makes the class type The predicted probability is The predicted probability for the true category is For a small batch containing For each sample, the overall classification loss is the average of the samples. ,in, For the first The classification loss for each sample. For category The weighting coefficients, This is a focusing factor used to reduce the impact of easily separable samples on the gradient. For numerically stable terms, take the smallest positive number to avoid... ; MMD loss is used for cross-modal feature alignment. The feature set extracted from the source domain (visible light) is... Each The feature set extracted from the target domain (SAR) is Each The MMD loss is defined as: ,in For kernel function, This represents the batch sample size. CORAL loss is used to match the second-order statistical properties of different modalities. ,in, Let covariance be the feature of the source domain. Let covariance be the feature matrix of the target domain. For feature dimension, It is the Frobenius norm; For entropy regularization loss, This is a sparsity loss; , , and These are the weight coefficients corresponding to the MMD loss, the CORAL loss, the entropy regularization loss, and the sparsity loss, respectively.

[0141] Specifically, Focal Loss is used to reduce the influence of easily classified samples in the classification loss. To address the class imbalance problem in sea surface target identification, this application employs an improved Focal Loss. ,in, For category weights, To focus parameters, To predict probabilities.

[0142] This embodiment employs a cosine annealing learning rate scheduling strategy, combined with a warmup mechanism, specifically as follows:

[0143] ;

[0144] In the formula, for, for Number of rounds This represents the total number of training rounds. The baseline learning rate is linearly increased during the preheating phase, and in the subsequent cosine annealing phase, it is usually correlated with the cosine formula. Equivalent; This is the minimum learning rate for the cosine annealing phase; This is the maximum learning rate during the cosine annealing phase. This is a count of the current training progress.

[0145] Step S102: Construct visible light-SAR image pairs, and perform cross-modal feature alignment training on the cross-modal transfer recognition network based on the visible light-SAR image pairs to obtain the trained cross-modal transfer recognition network.

[0146] In this embodiment, the visible light-SAR image pair is constructed as follows: Preferably, under the condition of clock / location information or synchronous acquisition, the two modal images are first subjected to geographic / temporal pre-screening and geometric registration, and then target detection is performed on the two modalities respectively, and matching is performed according to the intersection over union (IoU) or center distance of the detection boxes; when If the center distance is below a threshold, the two detection boxes are considered to be the same target instance and cropped into a pair of images. If precise alignment information is lacking, a weak pairing strategy based on category labels or a pseudo-pairing strategy based on feature retrieval is used to construct training pairs. The cropped images are uniformly scaled, and corresponding scaling, logarithmic transformation, speckle removal, normalization, and data augmentation strategies are applied to visible light and SAR images respectively to ensure the stability of cross-modal training. During training, a sampling strategy combining intra-batch multiple positive sample mining and hard negative sample mining is used to improve the convergence of the contrast / alignment loss.

[0147] In this embodiment, a multi-stage progressive training strategy is designed for cross-modal transfer tasks. Here, a small number of visible-SAR image pairs are used for cross-modal feature alignment training; the core of this stage is learning to map features from different modalities to a unified semantic space. The training loss consists of three parts:

[0148] ;

[0149] In the formula, For classifying losses, For feature alignment loss, These are the weight coefficients corresponding to the feature loss function; To compare learning loss, The weight coefficients corresponding to the contrastive learning loss are used to compare the features of similar targets across different modalities and to distance the features of dissimilar targets.

[0150] ;

[0151] In the formula, For cosine similarity functions, For temperature parameters, It is a natural exponential function. Indicates the first The feature vectors obtained from the visible light images by the feature extraction module This represents the feature vector obtained from the SAR image paired with it.

[0152] It should be noted that the aforementioned cross-modal feature alignment training is the training content of the second stage of the multi-stage progressive training strategy in this embodiment; in the first stage, this embodiment uses a large-scale natural image dataset (such as ImageNet) to pre-train the DenseNet backbone network (feature extraction module). The goal of this stage is to enable the network to learn general visual feature representation capabilities. The pre-training loss function is the standard cross-entropy loss, specifically:

[0153] ;

[0154] In the formula, For pre-training loss, For the sample size, For the number of categories, For tags, To predict probabilities.

[0155] In this embodiment, after the second stage of cross-modal alignment training, a third stage of target domain fine-tuning is performed. Here, a small amount of labeled data from the target domain dataset (SAR images) is used to perform end-to-end fine-tuning of the entire cross-modal transfer recognition network. To prevent overfitting, this application employs various regularization techniques, including: (a) early stopping strategy: monitoring validation set performance to prevent overfitting; (b) weight decay: adding an L2 regularization term to the loss function; (c) data augmentation: including random rotation, flipping, scaling, etc.; (d) Dropout: using Dropout in fully connected layers to prevent overfitting. Here, the total loss function for the fine-tuning stage is:

[0156] .

[0157] In the formula, This is the total loss function during the fine-tuning phase. For classifying losses, For regularization loss, The weighting coefficients for the regularization loss.

[0158] Step S103: Use the trained cross-modal transfer recognition network to perform transfer recognition on the airborne SAR images acquired in real time by the high-dynamic unmanned platform.

[0159] This embodiment provides a cross-modal sea surface target recognition method for highly dynamic unmanned platforms. By constructing a cross-modal transfer recognition network based on a DenseNet Siamese network, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to a pre-trained feature extraction module via visible light image input and SAR image input, respectively. Adaptive shared general feature extraction and modal feature extraction for different branches are performed in the feature extraction module, taking into account the different characteristics of the visible light and SAR images, using different network depths and connection methods. This yields a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Then, based on a cross-modal feature alignment module, the maximum mean difference between the visible light and SAR feature vector sets is minimized. A learnable mapping function projects the features of different modalities onto a unified semantic space, achieving effective cross-modal transfer. This effectively solves key problems such as scarce samples, difficult cross-modal transfer, limited computational resources, and insufficient noise resistance, providing a highly efficient, accurate, and robust sea surface target recognition solution for highly dynamic unmanned platforms.

[0160] Example 2

[0161] Based on the above embodiments, in order to evaluate the lightweight, robustness and recognition accuracy of the model in the application scenario of highly dynamic unmanned platforms, this embodiment verifies and analyzes the above-mentioned cross-modal sea surface target recognition method for highly dynamic unmanned platforms.

[0162] In this embodiment, the reported experiment was completed on a single experimental server with eight NVIDIA GeForce RTX 4090 Ti graphics cards. The specific experimental environment and key hyperparameters are as follows: Hardware consisted of eight RTX 4090 Ti cards, an Intel Xeon multi-core CPU, ≥128GB of RAM, and NVMe SSD storage. The software environment included Ubuntu 20.04, Python 3.10, CUDA 12.1, cuDNN 8.3, and PyTorch 2.0. Training was performed using mixed precision (AMP) and the AdamW optimizer. Single-card batch size = 16, weight decay = The number of model parameters and GFLOPs were measured using standard tools under the same input size. The recognition accuracy was calculated based on the Top-1 of the validation set. To ensure reproducibility, the random seed Seed=42 was set. The following test data is based on the average of 5 tests on a single card.

[0163] In terms of lightweight design trade-offs, this embodiment introduces AKConv locally instead of globally replacing all convolutions. This ensures the network's ability to model local deformations and noise while reducing the overall computational complexity, thus balancing performance improvement and computational efficiency. Experiments were conducted on the model after adding AKConv. Figure 4 This is a schematic diagram of the training and testing loss curves of the cross-modal transfer recognition network provided in the embodiments of this application. Figure 5 This is a schematic diagram illustrating the training and testing accuracy curves of the cross-modal transfer recognition network provided in an embodiment of this application. (Refer to...) Figure 4 , Figure 5 As shown, the model continues to converge during training. After adding AKConv, the overall convergence is more stable and the convergence effect is obvious. However, due to the lightweighting operation, the model's recognition accuracy on the test set is slightly reduced from 98.93% to 95.18%. Table 1 below shows the comparison of lightweighting metrics before and after introducing AKConv into the DenseNet network.

[0164] Table 1 Comparison of key performance indicators of DenseNet network before and after the introduction of AKConv

[0165]

[0166] As shown in the table, the introduction of AKConv (combined with DenseNet) significantly improves model lightweighting and inference acceleration compared to the original DenseNet across several key performance metrics: 1) Reduced parameter count: from 0.04M to 0.03M, a 25% reduction, contributing to a smaller and easier-to-deploy model. 2) Slightly increased computational complexity: FLOPs / MACs increased from 73.81M to 82.65M, a slight increase in computational complexity (~12%), but with minimal impact on overall efficiency. 3) Significantly reduced inference latency: from 30.25ms to 1.79ms, an inference speed improvement of over 16 times, greatly accelerating real-time inference capabilities. 4) Significantly increased throughput: from 33.06FPS to 557.82FPS, an approximately 16.9-fold increase in throughput, meaning more data can be processed on the same hardware. 5) Reduced GPU memory usage: From 28.49MB to 11.08MB, a reduction of approximately 61%, which is beneficial for deployment on resource-constrained devices (such as mobile and embedded devices). 6) With the introduction of AKConv, DenseNet achieves lower parameter count, lower GPU memory usage, and significantly improved inference speed and throughput while maintaining similar computational complexity. This demonstrates that AKConv plays a significant role in improving model efficiency and optimizing inference performance, and is particularly suitable for applications with strict requirements for real-time performance and resource consumption (such as edge computing and mobile AI inference).

[0167] Meanwhile, to test the robustness of the model to clutter noise, this application establishes a hybrid noise model for sea surface clutter and noise:

[0168] ;

[0169] In the formula, It is multiplicative speckle noise (its intensity follows a Gamma distribution). The noise is additive Gaussian, and this method is used to calculate the average signal-to-noise ratio. Then, noise with different average signal-to-noise ratios is added to the test sample set in the experiment, and target recognition experiments are carried out respectively. The experimental results are shown in Table 2.

[0170] Table 2. Recognition performance of the improved Akconv model under different noise intensities.

[0171]

[0172] Experimental results show that the introduction of AKConv effectively saves computing resources while ensuring the robustness of the model in complex noise and clutter scenarios. It has good practical application value and is particularly suitable for scenarios with high requirements for model robustness and moderate requirements for real-time performance.

[0173] In this embodiment of the application, for highly dynamic unmanned platforms, such as UAV-borne SAR target detection and identification background, the aforementioned DenseNet twin network is still used as the source model, and some category samples in OpenSAR with lower resolution are selected to simulate UAV-borne SAR images, construct a transfer learning target domain dataset, conduct transfer recognition experiments from visible light pre-trained models to UAV-borne SAR images, and perform adaptive analysis under different training set sizes and different noise interference.

[0174] Based on the above strategy, the DenseNet Siamese network model pre-trained on ImageNet is invoked, and samples from three classes—Cargo, Tanker, and Other—in OpenSAR are selected to construct the dataset. Sample examples are provided below. Figures 6 to 14 As shown ( Figures 6 to 8 These are the images corresponding to Cargo target example 1, Cargo target example 2, and Cargo target example 3, respectively. Figures 9 to 11 These are the images corresponding to Tanker target example 1, Tanker target example 2, and Tanker target example 3, respectively. Figures 12 to 14 The images are for Other Target Example 1, Other Target Example 2, and Other Target Example 3, respectively. All SAR images are uniformly set to 256×256. The model is re-tuned and trained, and a transfer recognition experiment of the visible light pre-trained model to UAV-borne SAR images is carried out. The adaptability of the algorithm is tested under different training set sizes and different noise interference.

[0175] Using the constructed UAV-borne SAR image dataset, the model was trained with different training set sizes, and then the accuracy was tested on the test set. The main focus was on the performance of the algorithm in the optimization learning process of the original DenseNet twin network model, as well as its adaptability under different training sample sizes. Figure 15 This is a graph showing the results of the cross-modal transfer recognition network model provided in this application embodiment with a training sample size of 1200. Figure 16 The test accuracy curve of the cross-modal transfer recognition network model provided in this application embodiment is shown in Table 3 with a training sample size of 1200. Detailed test data are shown in Table 3.

[0176] Table 3. Results of visible light to SAR image migration recognition based on DenseNet Siamese network

[0177]

[0178] As shown in the table above, regarding the adaptability to different training sample sizes, when the number of training samples is 1200, the accuracy is 99.47%, the recall is 99.68%, the precision is 98.44%, and the F1 score is 99.55%. Reducing the number of training samples to 800, the accuracy on the test set decreases to 99.37%, the recall to 99.68%, the precision to 99.44%, and the F1 score to 99.55%. Further reducing the number of training samples to 500, the accuracy on the test set decreases to 97.33%, the recall to 97.38%, the precision to 98.15%, and the F1 score to 97.74%. These data indicate that the visible light to SAR image migration recognition method based on the DenseNet Siamese network exhibits strong adaptability to different training set sizes, achieving high accuracy even with extremely small training set sizes.

[0179] The experimental results above show that the transfer recognition method using DenseNet twin network and its optimization learning method can complete the training and optimization of deep networks with a very low training sample size and achieve a high recognition accuracy. It also has good adaptability and stability under different training sample sizes and good noise robustness for SAR target domain images. It is particularly suitable for high dynamic unmanned platform application scenarios such as difficulty in obtaining samples, insufficient effective labeled samples, and imaging noise interference.

[0180] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method in any of the embodiments of this application. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU (Central Processing Unit) or MPU (Microprocessor Unit)) of the system or apparatus reads and executes the program code stored in the storage medium.

[0181] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined above in the system of this application.

[0182] It should be noted that the computer-readable storage medium shown in this application can be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. For example, a computer-readable storage medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. The transmitted data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF (Radio Frequency), etc., or any suitable combination thereof.

[0183] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0184] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0185] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0186] In the several embodiments provided in this application, it should be understood that the disclosed systems, modules, and methods can be implemented in other ways. For example, the module embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between modules or units, and may be electrical, mechanical, or other forms.

[0187] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. This application is not limited to the exact structures described above and illustrated in the accompanying drawings, and it should not be considered that the specific implementation of this application is limited to these descriptions. For those skilled in the art, various changes and modifications made without departing from the concept of this application should be considered to fall within the protection scope of this application.

Claims

1. A cross-modal sea surface target identification method for highly dynamic unmanned platforms, characterized in that, include: A cross-modal transfer recognition network based on the DenseNet Siamese network is constructed. The cross-modal transfer recognition network includes a feature extraction module, a cross-modal feature alignment module, and a classification decision module. First, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to the pre-trained feature extraction module via visible light image input and SAR image input, respectively. The feature extraction module performs adaptive shared general feature extraction and modal feature extraction for different branches, resulting in a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Second, the cross-modal feature alignment module minimizes the maximum mean difference between the visible light feature vector set and the SAR feature vector set, obtaining aligned features. Finally, the classification decision module performs target classification on the aligned features to determine target identification. Category; wherein, the feature extraction module includes a shared general feature extraction layer and a branch feature extraction layer, the shared general feature extraction layer includes 4 DenseNet block structures, each DenseNet block structure adopts a bottleneck structure, and the shared general feature extraction layer generates general features for images of two modalities; the visible light branch and the SAR branch share each DenseNet block structure and network parameters of the shared general feature extraction layer; based on the shared general feature extraction layer, the visible light branch uses a DenseNet-121 structure for visible light modal feature extraction, and the SAR branch uses a DenseNet-69 structure for SAR modal feature extraction; A visible light-SAR image pair is constructed, and the cross-modal transfer recognition network is trained by cross-modal feature alignment based on the visible light-SAR image pair to obtain the trained cross-modal transfer recognition network; Specifically, the DenseNet backbone network of the feature extraction module is pre-trained using a natural image dataset, with the standard cross-entropy loss function being used for pre-training. ； In the formula, For pre-training loss, For the sample size, For the number of categories, For tags, To predict probabilities; Furthermore, the cross-modal transfer recognition network is trained on visible light-SAR images to learn how to map features of different modalities to a unified semantic space, and the loss function is calculated. The loss function consists of three weighted parts, as shown in the following formula: ； In the formula, The classification loss is used to constrain the class separability of aligned features. This is a feature alignment loss used to constrain the consistency of the distribution of visible light features and SAR features in a unified semantic space. These are the weight coefficients corresponding to the feature loss function; To compare the learning loss, it is used to bring the cross-modal feature distance of the same image pair closer together and to widen the cross-modal feature distance of different image pairs. To compare the weight coefficients corresponding to the learning loss; ； In the formula, For cosine similarity functions, For temperature parameters, It is a natural exponential function. Indicates the first The feature vectors obtained from the visible light images by the feature extraction module Indicates the relationship with the first Feature vectors obtained from SAR images paired with visible light images; After training the cross-modal transfer recognition network with cross-modal feature alignment based on the visible-SAR image, the network is fine-tuned end-to-end using labeled data from the target domain dataset. Multiple regularization techniques are employed during this process to prevent overfitting. The total loss function during the fine-tuning stage is: ； In the formula, This is the total loss function during the fine-tuning phase. For classifying losses, For regularization loss, These are the weighting coefficients for the regularization loss; The trained cross-modal migration recognition network is used to perform migration recognition on the airborne SAR images acquired in real time by the high-dynamic unmanned platform.

2. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to claim 1, characterized in that, The branch feature extraction layer includes a visible light branch and a SAR branch; based on the shared general feature extraction layer, visible light modal features are extracted using a DenseNet-121 structure in the visible light branch, and SAR modal features are extracted using a DenseNet-69 structure in the SAR branch, specifically as follows: ； ； In the formula, For visible light feature vector set, For SAR feature vector set; The DenseNet-121 network is used for visible light modal feature extraction. The DenseNet-69 network is used for SAR modal feature extraction; To input the visible light mode sea surface target image of the visible light branch, The input is the SAR mode sea surface target image of the SAR branch; These are network parameters shared by the visible light branch and the SAR branch; The modal parameters of the visible light branch are... These are the modal parameters of the SAR branch; In this embodiment, adaptive kernel convolution modules are integrated into the visible light branch and the SAR branch respectively. The adaptive kernel convolution modules adaptively adjust the convolution kernel parameters of the corresponding branches according to the local statistical characteristics of the input feature map.

3. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to claim 2, characterized in that, The adaptive kernel convolution module adaptively adjusts the convolution kernel parameters of the corresponding branch based on the local statistical characteristics of the input feature map, including: For any branch, an initial sampling position is generated for the convolution kernel using a coordinate generation algorithm; The statistical properties of local regions of the input feature map are calculated based on the offset prediction network, specifically: ； ； In the formula, This refers to the size of a local window. and Positions Local mean and variance at point; This refers to the input feature map. In spatial location Eigenvalues ​​at; This refers to the offset index of the local window in the first spatial dimension; This refers to the offset index of the local window in the second spatial dimension; Based on local statistical characteristics, the AKConv module generates adaptive convolutional kernel weights that fit the current region, specifically: ； In the formula, For position The corresponding adaptive convolution kernel weights, Based on the convolutional kernel weights, and For learnable parameters, This represents element-wise multiplication. It is the hyperbolic tangent activation function.

4. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to claim 3, characterized in that, The adaptive kernel convolution module determines the local feature noise score of the input feature map and adjusts the corresponding convolution kernel weights based on the local feature noise score to obtain the final convolution kernel weights; specifically: ； ； In the formula, For position The corresponding noise score; It refers to the location The local region features extracted from the input feature map, centered on the input feature map; For local feature regions at frequency The absolute value of the Fourier transform amplitude at that point; High-frequency threshold; The maximum frequency component of the local region characteristics corresponds to the Nyquist frequency or the upper limit of the discrete spectrum. For position The corresponding final convolutional kernel weights, For position The corresponding adaptive convolution kernel weights; This is the noise suppression coefficient.

5. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to claim 4, characterized in that, The process of minimizing the maximum mean difference between the visible light feature vector set and the SAR feature vector set through the cross-modal feature alignment module to obtain the aligned features is specifically as follows: ； In the formula, For cross-modal alignment loss, Visible light feature vector set With SAR feature vector set The maximum mean difference between them; The square norm in the regenerated Hilbert space; Visible light feature vector set The total number of feature vectors contained therein; Visible light feature vector set The first in A high-dimensional feature representation of visible light feature vectors after kernel function mapping; SAR feature vector set The total number of feature vectors contained therein; SAR feature vector set The first in A high-dimensional feature representation of a SAR feature vector after kernel function mapping.

6. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to claim 5, characterized in that, During the training process of the cross-modal transfer recognition network, the composite loss function is calculated as follows: ； In the formula, This is a composite loss; For classifying losses, ,in, For category weights, To focus parameters, To predict probabilities; MMD loss is used for cross-modal feature alignment. CORAL loss is used to match the second-order statistical properties of different modalities. ,in, Let covariance be the feature of the source domain. Let covariance be the feature matrix of the target domain. For feature dimension, It is the Frobenius norm; For entropy regularization loss, This is a sparsity loss; , , and These are the weight coefficients corresponding to the MMD loss, the CORAL loss, the entropy regularization loss, and the sparsity loss, respectively.

7. The cross-modal sea surface target identification method for highly dynamic unmanned platforms according to any one of claims 1-6, characterized in that, Also includes: The pixel values ​​of the input initial visible light image are scaled to the [0,1] range using a standard normalization operation to obtain the preprocessed visible light image; The input initial SAR image is processed using logarithmic transformation and adaptive histogram equalization to obtain the preprocessed SAR image; specifically: Let the initial SAR image be... The preprocessed SAR image for: ； In the formula, The input is the initial SAR image. The SAR image after preprocessing. Scaling factor and These are the mean and standard deviation of the image after logarithmic transformation, respectively.

8. A cross-modal sea surface target recognition system for highly dynamic unmanned platforms, used to implement the cross-modal sea surface target recognition method for highly dynamic unmanned platforms as described in any one of claims 1-7, characterized in that, include: A cross-modal transfer recognition network construction unit is used to construct a cross-modal transfer recognition network based on the DenseNet Siamese network. The cross-modal transfer recognition network includes a feature extraction module, a cross-modal feature alignment module, and a classification decision module. First, preprocessed visible light images from the source domain dataset and preprocessed SAR images from the target domain dataset are input to the pre-trained feature extraction module via a first image input terminal and a second image input terminal, respectively. The feature extraction module performs adaptive shared general feature extraction and modal feature extraction for different branches, resulting in a visible light feature vector set from the visible light branch and a SAR feature vector set from the SAR branch. Second, the cross-modal feature alignment module minimizes the maximum mean difference between the visible light feature vector set and the SAR feature vector set to obtain aligned features. Finally, the classification decision module performs target classification on the aligned features to determine the target recognition class. The feature extraction module includes a shared general feature extraction layer and a branch feature extraction layer. The shared general feature extraction layer includes four DenseNet block structures, each of which employs a bottleneck structure. The shared general feature extraction layer generates general features for images of both modalities. The visible light branch and the SAR branch share the DenseNet block structures and network parameters of the shared general feature extraction layer. Based on the shared general feature extraction layer, the visible light branch uses a DenseNet-121 structure for visible light modal feature extraction, and the SAR branch uses a DenseNet-69 structure for SAR modal feature extraction. A cross-modal transfer recognition network training unit is used to construct visible light-SAR image pairs and perform cross-modal feature alignment training on the cross-modal transfer recognition network based on the visible light-SAR image pairs to obtain the trained cross-modal transfer recognition network. Specifically, the DenseNet backbone network of the feature extraction module is pre-trained using a natural image dataset, with the standard cross-entropy loss function being used for pre-training. ； In the formula, For pre-training loss, For the sample size, For the number of categories, For tags, To predict probabilities; Furthermore, the cross-modal transfer recognition network is trained on visible light-SAR images to learn how to map features of different modalities to a unified semantic space, and the loss function is calculated. The loss function consists of three weighted parts, as shown in the following formula: ； In the formula, The classification loss is used to constrain the class separability of aligned features. This is a feature alignment loss used to constrain the consistency of the distribution of visible light features and SAR features in a unified semantic space. These are the weight coefficients corresponding to the feature loss function; To compare the learning loss, it is used to bring the cross-modal feature distance of the same image pair closer together and to widen the cross-modal feature distance of different image pairs. To compare the weight coefficients corresponding to the learning loss; ； In the formula, For cosine similarity functions, For temperature parameters, It is a natural exponential function. Indicates the first The feature vectors obtained from the visible light images by the feature extraction module Indicates the relationship with the first Feature vectors obtained from SAR images paired with visible light images; After training the cross-modal transfer recognition network with cross-modal feature alignment based on the visible-SAR image, the network is fine-tuned end-to-end using labeled data from the target domain dataset. Multiple regularization techniques are employed during this process to prevent overfitting. The total loss function during the fine-tuning stage is: ； In the formula, This is the total loss function during the fine-tuning phase. For classifying losses, For regularization loss, These are the weighting coefficients for the regularization loss; A cross-modal migration recognition unit is used to perform migration recognition on airborne SAR images acquired in real time by the highly dynamic unmanned platform using the trained cross-modal migration recognition network.

Images