Ship type identification method and system based on transfer learning

Abstract

The invention relates to the technical field of graph recognition, in particular to a ship type recognition method and system based on transfer learning, and the method comprises the following steps: carrying out the combined preprocessing of a visible light ship image and a synthetic aperture radar image, and separating a ship inherent feature layer and an environmental noise feature layer through an adversarial domain decoupling network; generating an interference-removed cross-domain feature matrix; enhancing space-time consistency features related to the ship motion mode, and outputting an enhanced feature matrix after enhancement; and S2, constructing a differentiable decision tree classifier based on the enhanced feature matrix in S2, selecting a category with the highest probability as a final recognition result to be output, comparing the confidence coefficient of the final recognition result with a preset threshold, and if the confidence coefficient is lower than the preset threshold, triggering incremental learning. The adaptive capacity to complex sea conditions, ship type diversity and shielding conditions is improved, the error recognition rate and the attack resisting success rate are remarkably reduced, and the reliability of the system in the open environment is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of graphic recognition technology, and in particular to a ship type recognition method and system based on transfer learning. Background Art

[0002] As the scale of global ocean transportation continues to expand, ship type identification is of great significance in application scenarios such as maritime supervision, port scheduling, maritime traffic management, and illegal behavior monitoring. Traditional ship identification methods mostly rely on single-modality images (such as visible light images or radar images) for feature extraction and classification. The recognition accuracy is low in complex sea conditions, low visibility or occluded scenes, and it is difficult to deal with the recognition bias caused by differences in data distribution between different sensors.

[0003] In recent years, with the development of deep learning technology, multimodal fusion and transfer learning methods have gradually been applied to ship recognition tasks. Some studies have tried to use multimodal images in combination with visible light images to improve recognition accuracy by taking advantage of their complementary advantages under different weather conditions. However, existing multimodal methods generally have the following technical bottlenecks: First, there is a noise interference coupling problem in cross-modal features. Images of different modalities are greatly affected by the imaging mechanism. Background clutter, light interference, sea surface reflection and other noises can easily introduce redundant features, affecting the ability to distinguish the ship's structure.

[0004] Second, the feature migration process lacks motion semantic modeling. As a dynamic target, the type of ship is characterized by differences in motion patterns in addition to structural features. Existing methods do not fully exploit the spatiotemporal features in trajectory data, making it difficult to effectively distinguish between ship types with similar structures.

[0005] Third, the classifier lacks robustness and scalability. The recognition accuracy of traditional classification models decreases when faced with unknown types of ships or interference samples. There is a lack of effective robustness enhancement mechanisms. The lack of context-constrained incremental learning mechanisms can easily lead to abnormal expansion of the model structure or blurred category boundaries, limiting its long-term adaptability in actual complex environments. Summary of the invention

[0006] The present invention provides a ship type recognition method and system based on transfer learning, which has the ship type recognition method with cross-modal noise decoupling capability, motion semantic enhancement capability and context constraint adaptive learning capability, so as to achieve a more stable, accurate and scalable ship classification effect and meet the actual needs of modern maritime intelligent management systems.

[0007] A ship type recognition method based on transfer learning comprises the following steps: S1. Cross-domain noise decoupling feature extraction: Jointly preprocess the visible light ship image and synthetic aperture radar (SAR) image, separate the ship inherent feature layer and the environmental noise feature layer through the adversarial domain decoupling network, and generate a de-interference cross-domain feature matrix; S2. Spatial-temporal perception migration enhancement: The cross-domain feature matrix output by S1 is input into the spatial-temporal gated migration network, and the feature channel weights are dynamically adjusted using the ship GPS trajectory data to enhance the spatial-temporal consistency features related to the ship's motion mode, and the enhanced enhanced feature matrix is ​​output; S3. Adversarial robust classification decision: A differentiable decision tree classifier is constructed based on the enhanced feature matrix of S2. The probability distribution of ship types is calculated through the differentiable decision tree classifier, and the category with the highest probability is selected as the final recognition result output. The confidence of the final recognition result is compared with the preset threshold. If it is lower than the preset threshold, incremental learning is triggered. At the same time, virtual adversarial samples generated by adversarial training are embedded in the classification nodes to form an anti-interference ship type discrimination boundary, thereby improving the robustness of the classifier. S4. Context-aware incremental learning: Receive the incremental learning instruction triggered by S3, generate context constraints based on the ship traffic density and navigation rules in the current sea area, and only allow new feature vectors that meet the constraints to trigger the expansion of the classifier topology structure.

[0008] Optionally, the S1 specifically includes: S11, multi-source data spatiotemporal registration: align the visible light ship image with the synthetic aperture radar image, eliminate the spatial offset through feature point matching method based on the MMSI code and UTC timestamp in the ship AIS signal, and perform pixel-level alignment of the ship targets in the two modal data; S12, dual-channel adversarial decoupling: Construct a dual encoder network structure. The first encoder extracts the HSV color space texture features of the visible light image, and the second encoder extracts the backscattering intensity features of the synthetic aperture radar image. Noise separation is performed through the gradient reversal layer (GRL) and the domain classifier, where: The noise feature layer passes through the domain classifier loss function (i.e., the adversarial loss function, used to train the noise feature layer) Guidance, learning of domain-specific features associated with environmental perturbations; The inherent feature layer of the ship is reconstructed through the reconstruction loss (i.e., the reconstruction loss function is used to retain the inherent features) Constraints, retaining cross-modal common features related to the ship structure, Indicates that from the input The inherent features extracted from Indicates the image reconstruction result based on the feature; S13, cross-domain feature fusion: decouple the visible light intrinsic features Intrinsic features of synthetic aperture radar images Perform channel cascading and pass the attention-guided feature mapping weight matrix Perform dimensionality reduction fusion to generate a cross-domain feature matrix , the calculation formula is: in, represents element-wise multiplication, is the Sigmoid activation function, Indicates the channel splicing operation, is the Sigmoid activation function, express The total loss function is expressed as: ,in, , .

[0009] Optionally, the first encoder is a visible light encoder: using a ResNet-50 backbone network to extract HSV color space texture features (such as paint, rust, ship name and other information).

[0010] The second encoder is a SAR encoder: a polarization-consistent convolutional layer is designed to extract backscattering intensity features.

[0011] The domain classifier consists of a 3-layer fully connected network, which is connected to the encoder through a gradient reversal layer to identify the feature source modality and achieve feature domain adaptation.

[0012] Optionally, S2 includes sliding window sampling of the ship's trajectory data, extracting the motion state information within each time window, and constructing a motion vector sequence consisting of speed, heading angle and acceleration, inputting the motion vector sequence into a bidirectional long short-term memory network for temporal encoding, and outputting a low-dimensional spatiotemporal feature vector for characterizing the current ship motion mode.

[0013] Optionally, S2 also includes flattening the cross-domain fusion feature matrix of S1 in the spatial dimension, and performing feature-level interaction with the spatiotemporal feature vector. The interaction result is input into a perceptron composed of two layers of fully connected networks to generate gated attention weights that reflect the importance of each channel. The gated attention weights are dynamically adjusted for all channels through a normalization process so that the responsiveness between channels changes in an orderly manner with the motion state.

[0014] Optionally, the gated attention weight acts on the channel dimension of the cross-domain feature matrix of S1, enhances the feature channels relevant to the current motion mode by channel-by-channel scaling, and suppresses the expression intensity of noise interference and irrelevant features, so as to form a weighted feature matrix weighted by the gated attention weight; The weighted feature matrix is ​​input into the three-dimensional convolutional layer, local continuous structural features are extracted jointly in both time and space dimensions, motion consistency features are aggregated along the time-space dimension, and the enhanced feature matrix is ​​output. The enhanced feature matrix satisfies the constraint condition of the change amplitude between adjacent frames.

[0015] Optionally, S3 includes constructing a differentiable decision tree classifier based on an end-to-end trainable mechanism, each classification node of the decision tree classifier performs a feature splitting operation through a continuous gating function, allowing the input feature vector to be transmitted in a probabilistic form between paths, the input feature vector is the enhanced feature matrix after enhancement in S2, the gradient of the overall structure is propagated, the leaf nodes correspond to a predefined set of ship categories, and the probability distribution results on the corresponding ship category set are output as the candidate output for the final identification.

[0016] Optionally, the decision tree classifier introduces an adversarial sample generation mechanism during the training phase, generates a perturbation sample with the maximum classification shift risk by calculating the gradient direction of the input feature vector, and injects it into the classification structure together with the original sample for joint training. The loss function in the joint training combines the prediction loss of the original input and the stability index of the adversarial sample to enhance the stability of the classification boundary in a high noise environment; In the inference stage, the decision tree classifier calculates the probability distribution of the input feature vector on each ship type according to the path propagation result of the input feature vector in the classification tree, selects the category corresponding to the maximum probability value as the current recognition result, and judges the reliability of the output result according to the preset confidence threshold. When the confidence of the recognition result is lower than the confidence threshold, the incremental learning process is automatically triggered.

[0017] The decision tree classifier imposes continuity constraints on the feature change rate of each classification node during the inference phase, limiting the impact of feature disturbances on classification outputs.

[0018] Optionally, the S4 accesses a historical database of the ship automatic identification system (AIS), extracts the appearance of ships in the current sea area within the time window, calculates the ship traffic density in combination with the sea area, and loads the maritime rule library at the same time, and encodes the navigation restriction conditions therein into a set of logical rules; by receiving the data stream of the ship automatic identification system in real time, continuously tracking the changes in traffic density, and calculating the density change rate, when the density change rate exceeds the rate threshold, automatically switching to a high-density operation mode, and readjusting the activation conditions of incremental learning according to the current navigation environment, mapping the loaded navigation rule set into a continuous vector form to form a rule feature vector, which is used to describe the satisfaction degree of each constraint condition; When the confidence of the recognition result output by S3 is lower than the preset threshold, the corresponding enhanced feature matrix is ​​used as the new feature vector to be learned, and the new feature vector is incrementally verified. When the incremental verification passes, the new feature vector is used as a new category, a prototype node is created for the new category, and the features of the created prototype node are updated. A split path is added for the new category in the original differentiable decision tree classifier structure, and a contrast constraint mechanism is introduced to measure the feature differences between the new and old category prototypes to force a sufficient feature boundary distance to be maintained, thereby ensuring that the expanded classifier structure has clear separability and maintains the stability of the overall classification performance.

[0019] A ship type recognition system based on transfer learning is used to implement the above-mentioned ship type recognition method, including the following modules: The cross-domain feature extraction module is used to jointly pre-process the visible light image and the synthetic aperture radar image, extract the inherent features of the ship through the adversarial domain decoupling network, and generate a denoised cross-domain feature matrix; The spatiotemporal migration enhancement module is used to receive the cross-domain feature matrix and dynamically adjust the feature channel weights in combination with the ship's GPS trajectory data, and output a consistency-enhanced feature matrix that strengthens the motion pattern features; The classification decision module builds a differentiable decision tree classifier based on enhanced features, completes the probability calculation of ship types, and triggers the incremental learning mechanism when the recognition confidence is lower than the threshold; The incremental learning module generates contextual constraints based on ship traffic density and navigation rules, and performs adaptive expansion of the classifier structure when the conditions are met.

[0020] Beneficial effects of the present invention: The present invention realizes the effective separation of the inherent structural characteristics of the ship and the environmental noise by constructing a dual encoder adversarial decoupling network for visible light images and SAR images. Through guided learning of domain-specific noise and reconstruction constraints of cross-modal inherent features, the image disturbance interference caused by external factors such as illumination, sea surface reflection, and weather is suppressed, thereby improving the robustness of the model under multi-source heterogeneous input and cross-scene recognition performance.

[0021] The present invention proposes a robust classification method combining a differentiable decision tree with an adversarial embedding mechanism, introduces adversarial perturbation training and boundary smoothing constraints in the classification process, so that the classifier can still maintain the stability and clarity of the decision boundary when facing perturbed samples or samples of unknown types. Through the Lipschitz continuity restriction and prediction uncertainty estimation mechanism, the adaptability to complex sea conditions, ship type diversity and occlusion conditions is improved, and the misrecognition rate and the success rate of adversarial attacks are significantly reduced. The structure supports rapid decision-making of high-confidence samples and rejection of output of low-confidence samples, effectively improving the reliability of the system in an open environment.

[0022] The present invention introduces a context-driven incremental learning mechanism that combines traffic density and maritime rule constraints. In low-confidence recognition scenarios, it comprehensively evaluates the similarity between new features and historical sample distributions and the degree of satisfaction with navigation rules to decide whether to allow model structure expansion, thereby controlling the legitimacy and representativeness of newly added categories from the source. At the same time, through prototype memory updates and inter-category comparison constraints, it maintains the separability of newly added categories from existing structures, preventing category boundary blur and structural expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A schematic diagram of a recognition method flow chart of an embodiment of the present invention; Figure 2 Schematic diagram of a dual encoder network structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0026] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiments", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0027] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0028] like Figure 1-Figure 2 As shown, a ship type recognition method based on transfer learning includes the following steps: S1. Cross-domain noise decoupling feature extraction: Jointly preprocess the visible light ship image and synthetic aperture radar (SAR) image, separate the ship inherent feature layer and the environmental noise feature layer through the adversarial domain decoupling network, and generate a de-interference cross-domain feature matrix; S2. Spatial-temporal perception migration enhancement: The cross-domain feature matrix output by S1 is input into the spatial-temporal gated migration network, and the feature channel weights are dynamically adjusted using the ship GPS trajectory data to enhance the spatial-temporal consistency features related to the ship's motion mode, and the enhanced enhanced feature matrix is ​​output; S3. Adversarial robust classification decision: A differentiable decision tree classifier is constructed based on the enhanced feature matrix of S2. The probability distribution of ship types is calculated through the differentiable decision tree classifier, and the category with the highest probability is selected as the final recognition result output. The confidence of the final recognition result is compared with the preset threshold. If it is lower than the preset threshold, incremental learning is triggered. At the same time, virtual adversarial samples generated by adversarial training are embedded in the classification nodes to form an anti-interference ship type discrimination boundary, thereby improving the robustness of the classifier. S4. Context-aware incremental learning: Receive the incremental learning instruction triggered by S3, generate context constraints based on the ship traffic density and navigation rules in the current sea area, and only allow new feature vectors that meet the constraints to trigger the expansion of the classifier topology structure.

[0029] S1 specifically includes: S11, multi-source data spatiotemporal registration: align the visible light ship image with the synthetic aperture radar image, eliminate the spatial offset through feature point matching method based on the MMSI code and UTC timestamp in the ship AIS signal, and perform pixel-level alignment of the ship targets in the two modal data; S12, dual-channel adversarial decoupling: Construct a dual encoder network structure. The first encoder extracts the HSV color space texture features of the visible light image, and the second encoder extracts the backscattering intensity features of the synthetic aperture radar image. Noise separation is performed through the gradient reversal layer (GRL) and the domain classifier, where: The noise feature layer passes through the domain classifier loss function (i.e., the adversarial loss function, used to train the noise feature layer) Guidance, learning of domain-specific features associated with environmental perturbations; The inherent feature layer of the ship is reconstructed through the reconstruction loss (i.e., the reconstruction loss function is used to retain the inherent features) Constraints, retaining cross-modal common features related to the ship structure, Indicates that from the input The inherent features extracted from Indicates the image reconstruction result based on the feature; S13, cross-domain feature fusion: decouple the visible light intrinsic features Intrinsic features of synthetic aperture radar images Perform channel cascading and pass the attention-guided feature mapping weight matrix Perform dimensionality reduction fusion to generate a cross-domain feature matrix , the calculation formula is: in, represents element-wise multiplication, is the Sigmoid activation function, Indicates the channel splicing operation, is the Sigmoid activation function, express The transpose of The total loss function is expressed as: ,in, , . It is used to regulate the training weights of adversarial and reconstructive objectives, ensuring noise decoupling while retaining discriminative structural features.

[0030] The first encoder is a visible light encoder: it uses the ResNet-50 backbone network to extract HSV color space texture features (such as paint, rust, ship name, etc.); The visible light encoder (HSV texture feature extraction) is as follows: 1. Input preprocessing: Convert the visible light RGB image to HSV color space and separate the three channels of hue (H), saturation (S), and value (V); Normalize each HSV channel (H∈[0°,360°] is normalized to [0,1], and S / V∈[0,1] remains in the original range).

[0031] 2. ResNet-50 structure adjustment: Input layer transformation: replace the original ResNet-50 3-channel input layer with a 3×3 convolution kernel (adapting to HSV three-channel input); Feature extraction strategy: freeze the first three residual blocks (retain general feature extraction capabilities); Fine-tune residual blocks 4-5 to focus on HSV texture patterns unique to ships (such as rust hue changes and paint saturation differences).

[0032] Output features: Output feature maps from the 5th residual block .

[0033] The second encoder is a SAR encoder: a polarization-consistent convolutional layer is designed to extract backscatter intensity features; The SAR encoder implementation (polarization-consistent convolutional layer) is as follows: 1. Polarization data processing: The input is the scattering matrix of the fully polarized synthetic aperture radar image data , decomposed into 6-channel real input; 2. Polarization-consistent convolution design: Convolution kernel structure: Design 4 groups of 3×3 convolution kernels, each group corresponds to a polarization combination, and the relationship between polarizations is integrated through learnable weights; Feature enhancement mechanism: adding polarization coherence loss after the convolutional layer , the constraint characteristics are consistent with the scattering characteristics of the ship's metal structure ( is the theoretical value of the polarization coherence coefficient of the metal target); Output features: Output feature map after 4 layers of polarization convolution .

[0034] The domain classifier consists of a 3-layer fully connected network, which is connected to the encoder through a gradient reversal layer to identify the feature source modality and achieve feature domain adaptation.

[0035] The domain classifier implementation (noise feature separation) is as follows: 1. Network structure: Input: Noise feature layer (noise features output by the visible light / SAR encoder, dimension 256); Fully connected layer design: First layer: ,ReLU activation; Layer 2: 128 ,ReLU activation; Layer 3: 64 ,Sigmoid activation (output domain classification probability); 2. Adversarial training mechanism: Gradient Reversal Layer (GRL): GRL is inserted before the domain classifier to reverse the gradient sign during backpropagation, forcing the encoder to generate domain invariant features; Domain classifier loss function: using binary cross entropy loss ,in Indicates the source of the feature (0=visible light, 1=SAR); 3. Noise separation effect: By maximizing the domain classifier loss, the encoder is driven to separate domain-related noise (such as SAR speckle noise and visible light illumination changes) into the noise feature layer, retaining the inherent characteristics of the ship in the domain-invariant space.

[0036] S2 specifically includes: S21, trajectory spatiotemporal coding: Sliding window sampling (window length is 10 seconds, step length is 1 second) is performed on the ship GPS trajectory data to extract the speed within the time window , heading angle , acceleration Constructing motion vectors ,The time-series motion vector sequence is input into the bidirectional LSTM, and the spatiotemporal feature vector is generated through the bidirectional LSTM encoding , represents the spatiotemporal feature vector of the current ship motion mode: S22, gated weight generation: the cross-domain feature matrix output by S1 Expanded along the spatial dimension , and Perform element-by-element multiplication and generate channel-gated attention weights through two-layer perceptron MLP , expressed as: , where the two-layer perceptron MLP is a two-layer fully connected network (C , Normalize along the channel dimension so that the weight sums to 1; S23, Feature Dynamic Enhancement: Using Gated Attention Weights Perform channel weighting on the cross-domain feature matrix to strengthen the feature channels related to the motion pattern and form a dynamic feature matrix : ,in Indicates channel dimension broadcast multiplication, that is, channel-by-channel scaling; S24, spatiotemporal consistency verification: The weighted feature matrix Input 3D convolution layer (kernel size 3×3×3), aggregate motion consistency features along the time-space dimension, and output the enhanced feature matrix after enhancement , satisfying the space-time continuity constraints: ,in, is the enhanced feature of the first frame, Used to measure the amplitude of changes between frames. It is the continuity threshold, which controls the feature fluctuation range and prevents misjudgment and interference enhancement.

[0037] S3 specifically includes: S31, Differentiable Decision Tree Construction: Construct a differentiable decision tree classifier based on an end-to-end trainable mechanism. Each classification node uses the Sigmoid activation function as a soft split gate, and defines the classification node split function as a differentiable Sigmoid gating function. ,in is the enhanced feature matrix output by S2, , is a learnable parameter, is the activation function; Each leaf node is associated with a probability distribution of ship categories: ,in is the category weight matrix, is the total number of known ship types, is the predicted probability of the type of ship; S32, adversarial feature embedding training: In the training phase, the fast gradient sign method is used to generate adversarial perturbations on the input feature vector to obtain adversarial samples. : , , To counter the perturbation amplitude, we will use the adversarial sample, Inject the decision tree split node and optimize the loss function together with the original samples: ,in, is the weight coefficient of the loss of the original sample and the adversarial sample, ,enforces the classification boundaries to remain stable to perturbations, is the cross entropy loss function; S33, Probability Distribution Calculation and Decision Making: For any input, the feature matrix is ​​enhanced , calculate the path probability product from the root node to each leaf node, and obtain the final ship category probability distribution vector : ,The root node is the starting node of the decision tree, located at the top of the tree structure, and is the entrance for all input features to enter the classification process. It belongs to the classification node. The leaf node is the terminal node, which no longer splits and outputs the category probability; Select the maximum probability value The corresponding category is used as the recognition result, and the confidence threshold is set ,when Trigger S4 incremental learning; S34, anti-interference boundary optimization: In order to suppress the classification deviation caused by small disturbances, Lipschitz constraints are imposed on the classification nodes in the inference stage to limit the rate of change of the gradient in the feature space: ; Generate uncertainty estimates through Monte Carlo Dropout, when the prediction variance Var , reject the output result and re-extract features.

[0038] S4 specifically includes: S41, context knowledge base construction: access to the ship AIS historical database, extract the ship traffic density in the current sea area ,in, For time window The number of ships that appear in For the sea area, load the maritime rule base and encode the navigation constraints as a logical rule set ,in, Indicates hard or soft navigation constraints such as speed limit, draft limit, and prohibited sea areas, such as the hard constraint of "prohibiting ships with a draft depth > 10m from entering the nearshore area"; S42, real-time context feature generation: based on AIS real-time data stream, calculate the traffic density change rate in the current sea area , when the traffic density change rate Activate high-density mode when mapping the logical rule set to a rule feature vector , where each dimension corresponds to the satisfaction of a specific constraint, with a continuous value range of [0,1], and the mapping method is rule vectorization based on fuzzy logic; S43, constraint-driven incremental verification: new feature vectors to be learned triggered by S3 , calculate its joint matching degree with the context feature: ,in, Representation and historical feature distribution The Mahalanobis distance similarity, Sat is the rule satisfaction, is the weight coefficient of similarity and rule constraints, ,when , allowing the triggering of classifier expansion, is the minimum score threshold that allows the classifier to expand, ; New feature vector to be learned triggered by S3 It means: in S3 probability distribution calculation and decision making, after the input feature vector is classified by the differentiable decision tree classifier, if the maximum category probability value of the recognition result is lower than the set confidence threshold (pmax <τ=0.4), it is considered that the category of the current input sample is unclear or may belong to an unknown category. At this time, the S4 incremental learning process is triggered, and the enhanced feature matrix of the low-confidence input (that is, the feature vector output by S2) is used as the new feature vector to be learned. ; S44, classifier topology expansion: for new category features that pass verification, create new category prototype nodes EMA , using exponential moving average to update the prototype: ,in, is the prototype smoothing coefficient, , For the moment Then, a new split path is added to the differentiable decision tree classifier, and a contrast loss function is introduced to constrain the original category boundary. Constrain the spacing between new and old categories ( ), is the minimum distance threshold between categories, maintaining the stability of the classifier structure, is the prototype vector of the historical category, Represents the cosine similarity between category prototypes.

[0039] A ship type recognition system based on transfer learning is used to implement the above recognition method, including the following modules: The cross-domain feature extraction module is used to jointly pre-process the visible light image and the synthetic aperture radar image, extract the inherent features of the ship through the adversarial domain decoupling network, and generate a denoised cross-domain feature matrix; The spatiotemporal migration enhancement module is used to receive the cross-domain feature matrix and dynamically adjust the feature channel weights in combination with the ship's GPS trajectory data, and output a consistency-enhanced feature matrix that strengthens the motion pattern features; The classification decision module builds a differentiable decision tree classifier based on enhanced features, completes the probability calculation of ship types, and triggers the incremental learning mechanism when the recognition confidence is lower than the threshold; The incremental learning module generates contextual constraints based on ship traffic density and navigation rules, and performs adaptive expansion of the classifier structure when the conditions are met.

[0040] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0041] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A ship type identification method based on transfer learning, characterized in that: The following steps are involved: S1. Cross-domain noise decoupling feature extraction: Jointly preprocess the visible light ship image and synthetic aperture radar image, separate the ship inherent feature layer and the environmental noise feature layer through the adversarial domain decoupling network, and generate a de-interference cross-domain feature matrix, which includes: S11, multi-source data spatiotemporal registration, aligning the visible light ship image with the synthetic aperture radar image; S12, dual-channel adversarial decoupling, constructing a dual encoder network structure. The first encoder extracts the HSV color space texture features of the visible light image, and the second encoder extracts the backscattering intensity features of the synthetic aperture radar image. Noise separation is performed through the gradient reversal layer and domain classifier; S13, performing cross-domain feature fusion to generate a cross-domain feature matrix; S2. Spatial-temporal perception migration enhancement: The cross-domain feature matrix output by S1 is input into the spatial-temporal gated migration network, and the feature channel weights are dynamically adjusted using the ship GPS trajectory data to enhance the spatial-temporal consistency features related to the ship's motion mode, and the enhanced enhanced feature matrix is ​​output; S3. Adversarial robust classification decision: A differentiable decision tree classifier is constructed based on the enhanced feature matrix of S2. The probability distribution of ship types is calculated through the differentiable decision tree classifier. The category with the highest probability is selected as the final recognition result output. The confidence of the final recognition result is compared with the preset threshold. If it is lower than the preset threshold, incremental learning is triggered. At the same time, virtual adversarial samples generated by adversarial training are embedded in the classification node to form an anti-interference ship type discrimination boundary. S4. Context-aware incremental learning: Receive the incremental learning instruction triggered by S3, generate context constraints based on the ship traffic density and navigation rules in the current sea area, and only allow new feature vectors that meet the constraints to trigger the expansion of the classifier topology structure.

2. A ship type identification method based on transfer learning according to claim 1, characterized in that: The alignment operation of the visible light ship image and the synthetic aperture radar image includes eliminating the spatial offset by a feature point matching method based on the MMSI code and the UTC timestamp in the ship AIS signal, and performing pixel-level alignment on the ship targets in the two modal data; The noise feature layer is passed through the domain classifier loss function Guidance, learning of domain-specific features associated with environmental perturbations; The inherent characteristic layer of the ship is reconstructed through loss Constraints, retaining cross-modal common features related to the ship structure, Indicates that from the input The inherent features extracted from Indicates the image reconstruction result based on the feature; In the cross-domain feature fusion, the decoupled visible light intrinsic features Intrinsic characteristics of synthetic aperture radar images Perform channel cascading and pass the attention-guided feature mapping weight matrix Perform dimensionality reduction fusion to generate a cross-domain feature matrix , the calculation formula is: in, represents element-wise multiplication, is the Sigmoid activation function, Indicates the channel splicing operation, is the Sigmoid activation function, express The transpose of The total loss function is expressed as: ,in, , .

3. A ship type identification method based on transfer learning according to claim 2, characterized in that: The first encoder is a visible light encoder: a ResNet-50 backbone network is used to extract HSV color space texture features; The second encoder is a SAR encoder: a polarization-consistent convolution layer is designed to extract backscatter intensity features; The domain classifier consists of a 3-layer fully connected network, which is connected to the encoder through a gradient reversal layer to identify the feature source modality and achieve feature domain adaptation.

4. A ship type identification method based on transfer learning according to claim 1, characterized in that: The S2 includes sliding window sampling of the ship's trajectory data, extracting the motion state information in each time window, and constructing a motion vector sequence consisting of speed, heading angle and acceleration, inputting the motion vector sequence into a bidirectional long short-term memory network for temporal encoding, and outputting a low-dimensional spatiotemporal feature vector for characterizing the current ship motion mode.

5. A ship type identification method based on transfer learning according to claim 4, characterized in that: The S2 also includes flattening the cross-domain fusion feature matrix of S1 in the spatial dimension and performing feature-level interaction with the spatiotemporal feature vector. The interaction result is input into a perceptron composed of two layers of fully connected networks to generate gated attention weights that reflect the importance of each channel. The gated attention weights are dynamically adjusted for all channels through a normalization process so that the responsiveness between channels changes in an orderly manner with the motion state.

6. A ship type identification method based on transfer learning according to claim 5, characterized in that: The gated attention weight acts on the channel dimension of the cross-domain feature matrix of S1, enhances the feature channels relevant to the current motion mode by channel-by-channel scaling, and suppresses the expression intensity of noise interference and irrelevant features, thereby forming a weighted feature matrix weighted by the gated attention weight; The weighted feature matrix is ​​input into the three-dimensional convolutional layer, local continuous structural features are extracted jointly in both time and space dimensions, motion consistency features are aggregated along the time-space dimension, and the enhanced feature matrix is ​​output. The enhanced feature matrix satisfies the constraint condition of the change amplitude between adjacent frames.

7. A ship type identification method based on transfer learning according to claim 1, characterized in that: The S3 includes constructing a differentiable decision tree classifier based on an end-to-end trainable mechanism. Each classification node of the decision tree classifier performs a feature splitting operation through a continuous gating function, allowing the input feature vector to be transmitted in a probabilistic form between paths. The input feature vector is the enhanced feature matrix after enhancement in S2. The gradient of the overall structure is propagated, and the leaf nodes correspond to a predefined set of ship categories, and the probability distribution results on the corresponding set of ship categories are output as the candidate output for final recognition.

8. A ship type identification method based on transfer learning according to claim 7, characterized in that: During the training phase, the decision tree classifier introduces an adversarial sample generation mechanism, generates perturbation samples with the maximum classification shift risk by calculating the gradient direction of the input feature vector, and injects the perturbation samples together with the original samples into the classification structure for joint training. The loss function in the joint training combines the prediction loss of the original input and the stability index of the adversarial sample to enhance the stability of the classification boundary in a high noise environment. In the inference stage, the decision tree classifier calculates the probability distribution of the input feature vector on each ship type according to the path propagation result of the input feature vector in the classification tree, selects the category corresponding to the maximum probability value as the current recognition result, and determines the reliability of the output result according to the preset confidence threshold. When the confidence of the recognition result is lower than the confidence threshold, the incremental learning process is automatically triggered; The decision tree classifier imposes continuity constraints on the feature change rate of each classification node during the inference phase, limiting the impact of feature disturbances on classification outputs.

9. The ship type identification method based on transfer learning according to claim 1 is characterized in that: The S4 accesses the historical database of the automatic identification system of ships, extracts the appearance of ships in the time window in the current sea area, calculates the ship traffic density in combination with the sea area, and loads the maritime rule library at the same time, and encodes the navigation restriction conditions therein into a set of logical rules; by receiving the data stream of the automatic identification system of ships in real time, continuously tracks the changes in traffic density, and calculates the density change rate. When the density change rate exceeds the rate threshold, it automatically switches to the high-density operation mode, and readjusts the activation conditions of incremental learning according to the current navigation environment, maps the loaded navigation rule set into a continuous vector form, and constitutes a rule feature vector, which is used to describe the satisfaction degree of various constraint conditions; When the confidence of the recognition result output by S3 is lower than the preset threshold, the corresponding enhanced feature matrix is ​​used as the new feature vector to be learned, and the new feature vector is incrementally verified. When the incremental verification passes, the new feature vector is used as a new category, a prototype node is created for the new category, and the features of the created prototype node are updated. A split path is added for the new category in the original differentiable decision tree classifier structure, and a contrast constraint mechanism is introduced to measure the feature differences between the new and old category prototypes to force a sufficient feature boundary distance.

10. A ship type identification system based on transfer learning, used to implement a ship type identification method based on transfer learning as claimed in any one of claims 1 to 9, characterized in that: Includes the following modules: The cross-domain feature extraction module is used to jointly pre-process the visible light image and the synthetic aperture radar image, extract the inherent features of the ship through the adversarial domain decoupling network, and generate a denoised cross-domain feature matrix; The spatiotemporal migration enhancement module is used to receive the cross-domain feature matrix and dynamically adjust the feature channel weights in combination with the ship's GPS trajectory data, and output a consistency-enhanced feature matrix that strengthens the motion pattern features; The classification decision module builds a differentiable decision tree classifier based on enhanced features, completes the probability calculation of ship types, and triggers the incremental learning mechanism when the recognition confidence is lower than the threshold; The incremental learning module generates contextual constraints based on ship traffic density and navigation rules, and performs adaptive expansion of the classifier structure when the conditions are met.

Images