SAR lightweight ship tracking method and system for multi-scale feature fusion
By employing a downsampling method inspired by the concept of thermodynamic entropy and lightweight feature fusion, the limitations of information acquisition in SAR ship tracking under data scarcity and edge computing scenarios are solved, enabling efficient and stable multi-scale target tracking.
Patent Information
- Application Number
- CN202511650608.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-27
AI Technical Summary
In existing technologies, deep learning-based SAR ship tracking methods are limited in scenarios with scarce data and edge computing, making it difficult to effectively acquire large-scale contextual information. In particular, performance degrades when the target edges are blurred, occluded, or non-rigid targets are present.
Inspired by the concept of thermodynamic entropy, a downsampling method is adopted to decompose the feature map signal into different frequency band regions. The entropy value of each sub-region is calculated and weights are assigned. High-entropy regions are given priority, while low-entropy regions are compressed or ignored. The backbone network, neck structure and detection head are combined for collaborative optimization. A lightweight feature fusion module and joint convolution strategy are adopted to enhance multi-scale feature perception.
It significantly improves the model's target perception capability in complex scenarios, enhances the robustness of small target tracking and localization stability, while reducing computational costs, making the model suitable for resource-constrained platforms.
Smart Images
Figure CN121582289A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of computer vision and remote sensing image processing, and particularly relates to a SAR lightweight ship tracking method and system for multi-scale feature fusion. BACKGROUND
[0002] In recent years, target tracking methods based on deep learning have provided a new paradigm for SAR image intelligent analysis. Convolutional neural networks (CNN) can automatically learn discriminative image features due to their strong local perception ability. In the SAR ship tracking task, mainstream frameworks such as Faster R-CNN, SSD, YOLO have been widely used and have achieved remarkable results. However, the inherent limitations of CNN architecture have gradually emerged: the receptive field is limited by local convolution operations, making it difficult to obtain long-range context information, especially when the target edge is blurred, the targets are mutually occluded, or non-rigid targets exist.
[0003] To compensate for the local perception limitations of CNN, the Transformer architecture has been introduced into the field of remote sensing. Based on the self-attention mechanism, Transformer can model long-range dependencies and significantly improve the global expression ability of the model, making it particularly suitable for processing complex structures and blurred boundaries. However, its high computational cost and dependence on training data volume limit its practicality in data-scarce and edge computing scenarios. SUMMARY
[0004] The technical problem to be solved by the present application is to provide a down-sampling method inspired by the concept of thermodynamic entropy, a SAR lightweight ship tracking method and system for multi-scale feature fusion, which solves the problem of ship tracking being limited in data-scarce and edge computing scenarios in the prior art.
[0005] The technical problem to be solved by the present application is to provide a down-sampling method inspired by the concept of thermodynamic entropy, a SAR lightweight ship tracking method and system for multi-scale feature fusion, which solves the problem of ship tracking being limited in data-scarce and edge computing scenarios in the prior art.
[0006] The down-sampling method inspired by the concept of thermodynamic entropy decomposes the feature map signal into different frequency band regions. In each sub-frequency band region, the pixel value distribution entropy of each channel is calculated. According to the entropy value calculated in each sub-region, a weight is assigned to the local feature block. High-entropy regions are given higher preservation priority, and low-entropy regions are compressed or ignored. The weighted features are combined into a down-sampling output and fused, and finally the feature map is output.
[0007] The signal is decomposed into different frequency band regions using the frequency domain transformation method, and the specific formula is as follows:
[0008] , , , ,
[0009] wherein, similar to the low-energy state, , , is the high-energy state, indicating the energy distribution in different directions, capture vertical edges, capture horizontal edges, capture diagonal line information and with the inflection point.
[0010] The final output feature map is composed of the following four parts:
[0011] , , and respectively represent the batch size and the number of channels, and respectively represent the height and width of the feature map.
[0012] The SAR lightweight ship tracking method for multi-scale feature fusion obtains the feature map of the ship target, and cooperates with the main network, neck structure and detection head to obtain real-time positioning information. Specifically, the following steps are included:
[0013] Step 1, on the main network, the down-sampling method inspired by the concept of thermodynamic entropy in claim 1 is applied to the ship image feature map for feature transformation and decomposition and reorganization;
[0014] Step 2, on the basis of the standard BiFPN architecture, through the simplification of redundant connections and fusion, while retaining the bidirectional feature flow and the learnable weight mechanism;
[0015] Step 3, adopt the joint convolution strategy, through two layers of 3x3 convolution to share weights across scales, and perform detection head calculation to enhance the robustness and stability of the regression results.
[0016] The specific process of step 2 is as follows:
[0017] Step 2.1, select three feature maps of different scales, P3(X3), P4(X4), and P5(X5), wherein P3(X3) is the top feature of higher level, and P5(X5) is the bottom feature of lower level;
[0018] Step 2.2, the higher level feature is up-sampled to the lower level, and the bottom feature is down-sampled by convolution to pass to the top layer;
[0019] Step 2.3, apply the structure constructed by weighted feature fusion to fuse P3, P4 and P5.
[0020] The specific process of step 3 is as follows:
[0021] Step 3.1, using 1x1 convolution to compress the channel of the input feature map;
[0022] Step 3.2, using Group Normalization instead of traditional Batch Normalization, and introducing SiLU activation function for training;
[0023] Step 3.3, introducing two layers of continuous 3x3 convolution layers, and sharing parameters across scales, so that the features are reused between different scales.
[0024] In the last stage of the detection head, two decoupled convolution modules and a learnable scaling factor are used, the two decoupled convolution modules are used for target box regression and class classification respectively, and the learnable scaling factor is applied to the regression output and used to adaptively adjust the size of the prediction box.
[0025] The SAR lightweight ship tracking system for multi-scale feature fusion comprises a WDNet module, a lightweight feature fusion module and an EfficientHead module.
[0026] The WDNet module constructs an entropy pool to describe the uncertainty and information confusion degree of the system state, and maps it to feature learning, while reducing feature dimension and redundancy, and maintaining the integrity and diversity of information expression;
[0027] The lightweight feature fusion module introduces channel pruning, learnable weight fusion and lightweight connection unit based on BiFPN, up-samples higher layer features to lower layer, and down-samples bottom layer features to top layer for multi-scale feature perception of feature map;
[0028] The EfficientHead module adopts a joint convolution strategy, realizes cross-scale weight sharing through two layers of 3x3 convolution, and introduces channel compression and normalization mechanism to realize feature reuse between different scales.
[0029] The lightweight feature fusion module includes a bottom-up and top-down bidirectional fusion path to realize bidirectional information flow of high semantic and detailed features.
[0030] The joint convolution strategy of the EfficientHead module includes two decoupled convolution modules, Conv_Box for target box regression, and Conv_cls for class classification.
[0031] Compared with the prior art, the present application has the following beneficial effects:
[0032] 1. Introduce a thermodynamic-inspired EntropyPool module to replace the traditional max-pooling or average-pooling operation. This module effectively preserves high-frequency information and edge details in images by simulating the conversion mechanism of high and low energy states in the process of information entropy change, solving the problem of fine-grained feature loss in the downsampling process under traditional pooling methods, and significantly improving the target perception ability of the network in complex scenes.
[0033] 2. Propose a lightweight feature fusion structure Slender-BiFPN, which, while maintaining the advantages of bidirectional feature flow, prunes redundant calculation paths and optimizes the fusion method. Compared with the standard BiFPN structure, this module effectively fuses multi-scale information with fewer parameters, improves the robustness of small target tracking, and reduces the FLOPs of the model, making this method deployable on resource-constrained platforms.
[0034] 3. The detection head part designs an EfficientHead module, which combines 1x1 convolution compression channels, GroupNormalization, and SiLU activation functions to enhance feature expression; at the same time, it introduces a Joint_Conv structure with shared parameters to realize cross-scale feature reuse, further reducing the complexity of the model. With a learnable scaling factor Size, the positioning stability and accuracy of the model under multi-scale targets are improved. BRIEF DESCRIPTION OF DRAWINGS
[0035] Figure 1 The overall structure diagram of the ship target tracking framework of the present application.
[0036] Figure 2 The Slender-BiFPN feature fusion module proposed in the present application is shown in the schematic diagram.
[0037] Figure 3 The EfficientHead detection head module proposed in the present application is shown in the schematic diagram. DETAILED DESCRIPTION
[0038] The structure and working process of the present application will be further described below with reference to the accompanying drawings.
[0039] In the prior art, how to design a tracking network that balances feature accuracy, computational efficiency, and multi-scale expression ability has become the core problem of current SAR ship research. On the one hand, it is necessary to effectively fuse multi-scale spatial semantic information to cope with target size changes; on the other hand, it is also necessary to enhance the response ability to edges, high frequencies, and weak targets to avoid the destruction of key features in the downsampling process. At the same time, the model structure should be kept lightweight enough to meet the deployment needs of resource-constrained platforms.
[0040] The application aims to provide a high-efficiency remote sensing ship target tracking method based on a YOLO architecture, as shown in Figure 1 The overall structure diagram shown in the figure is suitable for synthetic aperture radar (SAR) image analysis in a complex background. The method fully considers the characteristics of ship targets in spatial scale and edge details, and combines the cooperative optimization of a backbone network, a neck structure and a detection head, so as to realize high-precision and low-computing-cost target tracking.
[0041] Firstly, an entropy pool inspired by the concept of thermodynamic entropy is proposed. In thermodynamic theory, entropy describes the uncertainty and information disorder of the system state. By mapping this concept to feature learning, we expect to preserve the details of the "high information entropy region" as much as possible during downsampling, that is, to maintain the integrity and diversity of information expression while reducing feature dimension and redundancy. Specifically,
[0042] The downsampling method inspired by the concept of thermodynamic entropy decomposes the feature map signal into different frequency band regions. In each sub-frequency band region, the pixel value distribution entropy of each channel is calculated. According to the entropy value calculated in each sub-region, a weight is assigned to the local feature block. High-entropy regions are given higher priority for preservation, and low-entropy regions are compressed or ignored. The weighted features are integrated into the downsampling output, and are fused to finally output the feature map.
[0043] Further, a SAR lightweight ship tracking method and system for multi-scale feature fusion are proposed. The feature map of the ship target is obtained, and the backbone network, neck structure and detection head are cooperatively optimized to obtain real-time positioning information. Specifically, the steps include:
[0044] Step 1: In the backbone network, the downsampling method inspired by the concept of thermodynamic entropy in claim 1 is applied to perform feature transformation and decomposition and reorganization on the ship image feature map.
[0045] Step 2: On the basis of the standard BiFPN architecture, redundant connections and fusion are simplified while retaining the bidirectional feature flow and learnable weight mechanism.
[0046] Step 3: A joint convolution strategy is adopted to perform detection head calculation through two layers of 3x3 convolution to share weights across scales, thereby enhancing the robustness and stability of the regression results.
[0047] The SAR lightweight ship tracking system for multi-scale feature fusion includes a WDNet module, a lightweight feature fusion module and an EfficientHead module.
[0048] Among them, the WDNet module constructs an entropy pool to describe the uncertainty of the system state and the degree of information disorder, and maps it to feature learning, thereby reducing feature dimensionality and redundancy while maintaining the integrity and diversity of information expression.
[0049] The lightweight feature fusion module introduces channel pruning, learnable weighted fusion and lightweight connection units on the basis of BiFPN, upsampling higher-level features to lower-level features, and passing lower-level features to the top level through convolutional downsampling, which is used for multi-scale target perception of feature maps.
[0050] The EfficientHead module employs a joint convolution strategy, achieving cross-scale weight sharing through two layers of 3×3 convolutions. It also introduces channel compression and normalization mechanisms to enable feature reuse across different scales.
[0051] Specific embodiments, such as Figures 1 to 3 As shown:
[0052] like Figure 1 As shown in Figure 1, this embodiment provides an efficient ship target tracking framework. The framework, based on the design concept of "feature preservation - multi-scale fusion - efficient prediction," comprises three key modules that work together to improve overall tracking performance and computational efficiency. The three modules are:
[0053] WDNet Module: A custom module that can more effectively transform and reconstruct features, preserving fine-grained texture and structural information in images, thereby significantly improving the discriminative power of feature representations.
[0054] Slender-BiFPN Module: A custom module designed in a lightweight manner based on the standard BiFPN architecture. It simplifies redundant connections and fusion strategies while preserving bidirectional feature streams and learnable weighting mechanisms.
[0055] EfficientHead module: A custom module that uses a joint convolution strategy. It shares weights across two 3×3 convolution layers at different scales, effectively reducing redundant computations in the detection head.
[0056] Specifically, the above three modules are used in a lightweight SAR ship tracking method based on multi-scale feature fusion, including:
[0057] S1, such as Figure 1 As shown, the WDNet module proposed in this invention enables the model to pay more attention to edge feature information, especially enhancing the localization capability of ship targets, particularly in complex backgrounds or noisy environments.
[0058] S2, such as Figure 2As shown, the Slender-BiFPN feature fusion module provided by the application aims to significantly reduce the calculation complexity and improve the feature fusion efficiency on the basis of maintaining the BiFPN bidirectional feature flow advantage.
[0059] S3、As Figure 3 As shown, the EfficientHead detection head module provided by the application adopts a highly structured lightweight design, aiming to further compress the model calculation amount and improve the inference efficiency while maintaining the accuracy. To cope with the challenge of increasing model complexity, the efficient detection head method of the present solution ingeniously reduces the network parameter amount, effectively reduces the memory consumption without sacrificing accuracy.
[0060] The WDNet module in S1 specifically comprises the following steps:
[0061] S11, the EntropyPool mechanism is an innovative structure designed in the application to solve the problem of high-frequency information loss caused by traditional down-sampling methods. Its inspiration comes from the concept of 'entropy' in thermodynamics, which measures the amount of information that measures the degree of chaos of a system. This module aims to retain as much texture, edge and high-frequency detail information in the image as possible while reducing the resolution of the feature map, thereby enhancing the feature expression capability.
[0062] The S11 specifically comprises the following steps:
[0063] S111, assuming that the input feature map is: , where B is the batch size, C is the number of channels, H and W are the height and width of the feature map, respectively. By frequency domain transformation, the different frequency components of the feature map are decomposed, and the formula used is: , , , , similar to adjusting the distribution of high-energy and low-energy states in a thermodynamic system, so that information is transmitted in a more balanced way during down-sampling.
[0064] S112, for the cth input channel , respectively, four kernels are used to do convolution with a step size of 2 to obtain four down-sampled feature maps. The formula is: , , , .
[0065] S113, therefore, the final output is composed of these four parts: , EntropyPool evaluates the information entropy in each local region and guides the sampling process to preserve the most discriminative regions. The signal is decomposed into different frequency bands, which can effectively separate high-frequency edge information and low-frequency structural information, facilitating subsequent entropy calculation. In each sub-region, the pixel value distribution entropy of each channel is calculated, and according to the entropy value calculated in each sub-region, a weight is assigned to the local feature block. High-entropy regions are given higher preservation priority, and low-entropy regions are compressed or ignored. The weighted features are combined into a downsampled output, and a learnable weighted convolution or attention mechanism is used for fusion, and finally the feature map is output. In Table 1, it can be concluded that the WDNet module proposed by us has higher precision and smaller parameters.
[0066] In the embodiment, the Slender-BiFPN module specifically comprises the following steps:
[0067] In the S2, the Slender-BiFPN module specifically comprises the following steps:
[0068] S21, the Slender-BiFPN module introduces channel pruning, learnable weighted fusion and lightweight connection strategy based on BiFPN, aiming to reduce model complexity while enhancing cross-scale information interaction capability, realizing more effective multi-scale target perception.
[0069] S22, in the feature fusion process, each path passes through a learnable weighted parameter to automatically determine the importance of each input feature map. The formula is: wherein is a learnable fusion weight, and ε is a very small constant, which can be selected as any value in 1-10 to prevent division by zero error. This module can realize information interaction and weighted fusion between multi-scale features, effectively enhancing the model's perception ability for small targets and multi-scale targets in complex scenes.
[0070] In the S21, the Slender-BiFPN module specifically comprises the following steps:
[0071] S211, input different layer feature maps from Backbone, usually including three scales (e.g. P3, P4, P5), representing high-resolution detail features, middle-level semantic information and low-resolution high-level semantic features respectively. Higher-level features need to be upsampled to lower levels to enhance the semantic information of lower-level features.
[0072] In the S211, the Slender-BiFPN module specifically comprises the following steps:
[0073] S2111, High-level features need to be upsampled to low-level features to enhance the semantic information of low-level features. To enhance the semantic information of low-level features. In our architecture, we assume The I-layer feature map represents the backbone network. Slender-BiFPN mainly processes three different scale feature maps, namely P3 (X3), P4 (X4), which can be represented as: , , .
[0074] S2112, The method of upsampling P5 to P4 using the nearest neighbor interpolation method can be represented as: The bottom layer feature X is downsampled by convolution, which introduces more detailed information. A 3x3 convolution operation with a step size of 2 is used for feature down-sampling. A 3x3 convolution operation with a step size of 2 is used for feature down-sampling: .
[0075] S2113, The fused features are further passed to the C2f structure, which uses a cross-stage partial channel crosslinking technique to increase feature diversity. Finally, it is delivered to the detection head for final target prediction.
[0076] S212, A bottom-up and top-down bidirectional fusion path is constructed to realize the bidirectional flow of high semantic and detailed features. This structure can enhance the context modeling ability between different layers, especially for small targets or distant targets. From Table 1, we can see that our Slender-BiFPN module reduces the parameter amount of the model without losing accuracy.
[0077] The S3 mentioned in the application, the EfficientHead module, specifically includes the following steps:
[0078] S31, The application proposes a joint convolution design concept based on the reconstruction of feature paths, shared feature extractors, and the introduction of structure reparameterization methods to improve performance and reduce computational burden. As shown in Figure 3 For the original input feature map, a set of joint convolutions is used to preliminarily refine it to form a high semantic unified representation.
[0079] The S31 mentioned in the application, specifically includes the following steps:
[0080] S311, In order to control the parameter size and computational amount, we use 1x1 convolution to compress the input feature map, which effectively reduces redundant information and improves feature processing efficiency.
[0081] S312、On this basis, group normalization (GN) is used to replace the traditional batch normalization (BN) to alleviate the performance degradation caused by the instability of BN statistics. At the same time, the SiLU activation function is introduced, and its adaptive characteristics make the gradient smoother in the propagation process, which helps to improve the nonlinear expression ability and training convergence speed.
[0082] S313、After channel compression, we introduce a joint convolution module to further improve the expression efficiency of multi-scale features. This module consists of two consecutive 3x3 convolutions, sharing parameters across scales, allowing cross-scale feature reuse. This design reduces the number of parameters and memory overhead, and maintains the consistency of key features at different resolutions, thereby improving overall expression ability and accuracy, especially suitable for resource-constrained scenarios.
[0083] S314、In the final stage of the detection head, EfficientHead adopts two decoupled convolution modules, Conv_Box for target box regression, and Conv_cls for class classification. In addition, we also design a learnable scaling factor to improve the positioning accuracy of the model at different target scales. This factor is applied to the regression output and is used to adaptively adjust the size of the predicted box. This mechanism can effectively offset the prediction bias caused by scale changes, improving the robustness and stability of the regression results. From Table 1, we can see that our method has higher accuracy and smaller model parameters.
[0084] To benchmark this method, it was evaluated on three synthetic aperture radar datasets, SSDD, HRSID, and MIDSHIP, which cover complex coastal and port scenarios with multi-scale ship targets.
[0085] The SSDD dataset consists of 1160 images with a size of 500x500 pixels and 2456 annotated ship targets.
[0086] The HRSID dataset contains 5,604 images (800x800px) and 16,951 annotated ships, with spatial resolutions of 0.5 meters, 1 meter, and 3 meters.
[0087] SSDD contains 1,160 images (500 × 500 px) and 2,456 ship annotations.
[0088] MIDShip contains 5,673 images with 135,884 fine-labeled ship object instances. The data is collected from 43 different channel regions, covering a variety of maritime traffic scenarios, including ship encounter, berthing operation, small target aggregation, and partial occlusion. For the three datasets, we use a 7:2:1 ratio for training, validation, and testing. The method specifically includes the following steps:
[0089] Step 1, the present application adopts multi-dimensional indicators to comprehensively evaluate the performance of the model. In terms of accuracy, mAP.50 and mAP.50:.95 are selected as evaluation criteria, respectively reflecting the ability of the model under fixed threshold and dynamic threshold; in terms of calculation efficiency, the complexity of the model is measured by the number of parameters to evaluate its lightweight degree. This evaluation system can comprehensively investigate the performance of the model under different positioning accuracy requirements, and quantify its calculation resource occupation, ensuring the optimal balance between accuracy and efficiency. Among them, mAP.50 reflects the basic performance, mAP.50:.95 evaluates the adaptability of the model to positioning accuracy, and the number of parameters directly reflects the storage and calculation requirements of the model.
[0090] Step 2, we conducted ablation experiments on the public dataset SSDD to evaluate the performance improvement of WDNet, Slender-BiFPN and EfficientHead in network detection. mAP.50, mAP.50:.95 and Params are used as quantitative measures. The influence of each component is listed in Table 1.
[0091] Table 1
[0092]
[0093] Experiments show that our method achieves excellent detection accuracy in multiple datasets. In addition, our unique lightweight design enables the model to have fewer parameters and faster inference.
[0094] Step 3, in order to verify the effectiveness and superiority of the proposed method, we conducted comprehensive comparison experiments on multiple datasets under the same conditions using the most advanced method. The results for the dataset SSDD are shown in Table 2.
[0095] Table 2
[0096]
[0097] In summary, the core innovations of the present application are summarized as follows:
[0098] 1.The entropy pooling module can dynamically calculate the information entropy of the local image region, and then weighted sampling is performed according to the information entropy, replacing the conventional maximum pooling or average pooling operation.The method can adaptively retain the high-frequency features such as key texture and edge in the downsampling process, and significantly improve the trackability of weak targets or small-scale targets in SAR images.EntropyPool can be seamlessly integrated into each level of feature extraction layer in the YOLO main structure, and improve the expression ability of the overall network to the details of the ship contour.
[0099] 2.The Slender-BiFPN structure is proposed to solve the problem of large variation of ship target size in SAR images, which effectively improves the efficiency and quality of feature fusion of different scales through channel compression, fusion path simplification and feature selection mechanism; at the same time, the EfficientHead detection head adopts the joint convolution structure and the lightweight task decoupling strategy, which significantly reduces the calculation cost in the inference stage through structure reparameterization, and balances the real-time performance while maintaining the accuracy. The two are optimized in the efficiency of feature utilization, which is the key component to realize light weight and high efficiency.
[0100] Those skilled in the art should understand that those skilled in the art can realize the variations in combination with the prior art and the above embodiments, and such variations do not affect the essential content of the scheme, which will not be described here.
[0101] It should be understood that the present scheme is not limited to the above specific embodiments, and the devices and structures not fully described should be understood as being implemented in the ordinary way in the art; any person skilled in the art can make many possible changes and modifications to the present scheme, or modify it into equivalent embodiments without departing from the scope of the present scheme, which does not affect the essential content of the present scheme. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present scheme, without departing from the content of the present scheme, still belongs to the protection scope of the present scheme.
Claims
1. A downsampling method inspired by the concept of thermodynamic entropy, characterized by: The feature map signal is decomposed into different frequency band regions. Within each sub-frequency band region, the pixel value distribution entropy of each channel is calculated. Based on the entropy value calculated for each sub-region, weights are assigned to local feature blocks. High-entropy regions are given higher retention priority, while low-entropy regions are compressed or ignored. The weighted features are synthesized into a downsampled output and fused to finally output the feature map.
2. The downsampling method inspired by the concept of thermodynamic entropy according to claim 1, characterized in that: The signal is decomposed into different frequency bands using frequency domain transformation, as shown in the following formula: ， ， , in, Similar to a low-energy state, , , , representing a high-energy state, indicating energy distribution in different directions. Capture vertical edges, Snap to horizontal edges Capture diagonal information and inflection points.
3. The downsampling method inspired by the concept of thermodynamic entropy according to claim 2, characterized in that: The final output feature map consists of the following four parts. composition: , , and These represent the batch size and the number of channels, respectively. and These represent the height and width of the feature map, respectively.
4. A lightweight SAR ship tracking method for multi-scale feature fusion, characterized in that: The process involves acquiring a feature map of the ship target, combining it with the backbone network, neck structure, and detection head for collaborative optimization, and obtaining real-time positioning information. Specifically, this includes the following steps: Step 1: On the backbone network, the downsampling method inspired by the concept of thermodynamic entropy as described in claim 1 is applied to perform feature transformation and decomposition and reconstruction on the ship image feature map; Step 2: Based on the standard BiFPN architecture, simplify redundant connections and fusion while retaining bidirectional feature flow and learnable weight mechanism; Step 3: Adopt a joint convolution strategy, using two 3×3 convolution layers to share weights across scales for detector head calculation, thereby enhancing the robustness and stability of the regression results.
5. The SAR lightweight ship tracking method for multi-scale feature fusion according to claim 4, characterized in that: The specific process of step 2 is as follows: Step 2.1: Select three feature maps of different scales, namely P3 (X3), P4 (X4), and P5 (X5), where P3 (X3) is the higher-level top-level feature and P5 (X5) is the lower-level bottom-level feature; Step 2.2: Upsample higher-level features to lower levels, and pass lower-level features to the top level through convolutional downsampling; Step 2.3: Apply the structure constructed by weighted feature fusion to fuse P3, P4 and P5.
6. The SAR lightweight ship tracking method for multi-scale feature fusion according to claim 5, characterized in that: The specific process of step 3 is as follows: Step 3.1: Compress the input feature map using a 1×1 convolution; Step 3.2: Replace traditional Batch Normalization with Group Normalization, and introduce the SiLU activation function for training; Step 3.3: Introduce two consecutive 3×3 convolutional layers and share parameters across scales to enable feature reuse across different scales.
7. The SAR lightweight ship tracking method for multi-scale feature fusion according to claim 4, characterized in that: In the final stage of the detection head, two decoupled convolutional modules and a learnable scaling factor are used. The two decoupled convolutional modules are used for object box regression and class classification, respectively, and the learnable scaling factor is applied to the regression output to adaptively adjust the size of the prediction box.
8. A lightweight SAR ship tracking system for multi-scale feature fusion, characterized in that: It includes the WDNet module, the lightweight feature fusion module, and the EfficientHead module. Among them, the WDNet module constructs an entropy pool to describe the uncertainty of the system state and the degree of information disorder, and maps it to feature learning, thereby reducing feature dimensionality and redundancy while maintaining the integrity and diversity of information expression. The lightweight feature fusion module introduces channel pruning, learnable weighted fusion and lightweight connection units on the basis of BiFPN, upsampling higher-level features to lower-level features, and passing lower-level features to the top level through convolutional downsampling, which is used for multi-scale target perception of feature maps. The EfficientHead module employs a joint convolution strategy, achieving cross-scale weight sharing through two layers of 3×3 convolutions. It also introduces channel compression and normalization mechanisms to enable feature reuse across different scales.
9. The SAR lightweight ship tracking system for multi-scale feature fusion according to claim 8, characterized in that: The lightweight feature fusion module includes a bottom-up and top-down bidirectional fusion path to achieve bidirectional information flow between high semantic and detailed features.
10. The SAR lightweight ship tracking system for multi-scale feature fusion according to claim 8, characterized in that: The joint convolution strategy of the EfficientHead module includes two decoupled convolution modules: Conv_Box, used for target box regression; and Conv_cls, used for class classification.