SAR ship image target detection method based on C3K2-SC-Yolov11

Abstract

The invention belongs to the technical field of ship target detection, and particularly relates to a C3K2-SC-Yolov11-based SAR (Synthetic Aperture Radar) ship image target detection method. Comprising the following steps: firstly, designing a parallel convolution multi-scale feature module C3K2SC to replace a C3K2 module, thereby effectively enhancing the adaptive capability of the model, enabling the model to pay attention to channels and important regions with rich information at the same time, and capturing a significant target in a global context; secondly, an HSANPSA module is designed to replace C2PSA, and interference of coherent speckle noise is remarkably suppressed through deep fusion of hierarchical space attention and position sensitive attention; and finally, an SSSA attention mechanism is introduced to improve the CBMA, the advantages of the CBAM are maintained, the performance in the aspects of small target detection and multi-scale adaptation is remarkably improved, and the method is suitable for a target detection task in a complex scene. According to the use method of the SAR ship image target detection method based on the C3K2-SC-Yolov11, the mAP (at) 0.5 and the mAP (at) 50-95 of an improved model on an HRSID data set are improved by 5% and 9% respectively compared with that of the YOLOv11.

Description

Technical Field

[0001] This invention belongs to the field of ship target detection technology, and particularly relates to a SAR ship image target detection method based on C3K2-SC-Yolov11. Background Technology

[0002] The development of remote sensing technology has provided technical support for satellite monitoring and maritime safety. Synthetic Aperture Radar (SAR) is a crucial component of remote sensing technology. Compared to optical and infrared imaging, SAR can monitor and image various terrains around the clock, providing a more intuitive view of monitored objects. SAR echoes possess rich target information and are unaffected by weather or seasons, making them useful in various civilian fields, particularly irreplaceable in ship target detection. SAR enables automatic reconnaissance, real-time monitoring, and alerting; it can also monitor and manage ship entry and exit from ports around the clock, helping to improve transportation efficiency and monitor maritime traffic. Therefore, researching efficient SAR ship image detection methods provides technical support for specific maritime applications.

[0003] But SAR images (see) Figure 1 The imaging mechanism can cause problems such as low resolution, unclear target features, and complex background, which increases the difficulty of target detection, especially in the detection of small target hulls, where the accuracy decreases.

[0004] Therefore, those skilled in the art are dedicated to developing a SAR ship image target detection method based on C3K2-SC-Yolov11. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of low resolution, unclear target features, and complex background caused by the existing imaging mechanisms mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a SAR ship image target detection method based on C3K2-SC-Yolov11, comprising the following steps:

[0007] S1: The SimAM and CPCA attention mechanisms are combined to form a new attention mechanism, SCfusion;

[0008] S2: Combine the attention mechanism SCfusion formed in S1 with the C3K2 module to generate a new C3K2-SC module;

[0009] S3: The HSAN_PSA module is designed by introducing a multi-scale attention network;

[0010] S4: Introduce a spatially separable self-attention mechanism for CBMA.

[0011] This method improves the C3K2 and CBMA modules by combining a multi-scale attention mechanism and introducing an effective fusion of HSAN and PSA, replacing the C3K2 and C2PSA modules in Yolov11. The addition of a new attention mechanism enhances the expression and detection performance of target features. To verify the effectiveness of the improved module, data was collected using the HRSID dataset, and ablation experiments were conducted.

[0012] Preferably, the SimAM and CPCA attention mechanisms in S1 employ parallel branch convolution fusion.

[0013] SimAM is a spatial attention mechanism that enhances spatial details without considering inter-channel differences. However, in SAR detection, the contours and shapes of target features are equally important across channels. Ensuring the model learns more channel features is crucial to reducing false detection rates. Therefore, CPCA attention mechanism is introduced. CPCA captures global contextual information through dual-pooling channel attention, enhancing the model's perception of both global and local features. This allows for the capture of richer contextual information, which is beneficial for detecting small targets. CPCA employs multi-scale, deep separable convolutional modules to construct spatial attention, dynamically allocating attention weights across channel and spatial dimensions. Parallel convolutional fusion of these modules enables the model to learn more robust feature representations, enhancing features from both spatial and channel dimensions. This allows the model to adaptively focus on important features, thereby improving generalization ability on unseen data.

[0014] Preferably, the energy function of SimAM in S1 is:

[0015]

[0016] In the formula, ω i and b i Let t ∈ i and x represent the weights and biases, respectively. i The target neuron and other neurons in the channel, M is the number of neurons in the channel, and i is the spatial index.

[0017] SimAM's attention mechanism calculates the importance of each neuron in a parameter-free manner, adaptively enhancing important features and suppressing unimportant features. Even if the target is partially occluded or small in size, the attention mechanism can highlight its key features. In the formula, the greater the difference between neuron t and its surrounding neurons, the higher its importance, indicating lower energy.

[0018] Preferably, in the CPCA attention mechanism of S1, the channel attention module weights the importance of each channel, highlighting important feature channels and suppressing irrelevant channels. After the input features are processed by average pooling and max pooling, they are mapped into the shared multilayer perceptron to obtain the channel weight vector, the formula of which is:

[0019] M c (F)=σ(Mlp(AvgPool(F))+MLP(MaxPool(F)));

[0020] In the formula, σ is the Sigmoid activation function.

[0021] Using the Sigmoid activation function as a gate aims to enhance the characteristic responses of important channels while suppressing non-significant or redundant channel information. This yields M... c Following (F), multi-scale depthwise separable convolution is used to extract spatial dependencies under different receptive fields, generating a spatial attention map. Finally, the refined output features are obtained by element-wise multiplying the channel mixing result with the channel prior. The channel-priority convolutional attention mechanism adopts a "channel first, space later" serial processing structure. This design provides a cleaner and more informative input for the subsequent spatial attention module by prioritizing the selection of key feature channels, thereby improving the efficiency and focus of feature selection in the spatial dimension.

[0022] Preferably, the C3K2-SC module in S2 is used to replace the C3K2 module in the YOLOv11 system.

[0023] In the original CPCA attention mechanism, premature channel filtering may inadvertently delete features that only become important when combined with spatial information, leading to the loss of valuable information. The SimAM attention mechanism, while focusing on enhancing spatial details, fails to adequately consider the differences between channels. To address these issues, the improved C3K2-SC module employs a dual-attention parallel fusion strategy, combining the explicit channel attention of CPCA with the adaptive spatial attention of SimAM. This effectively enhances feature extraction capabilities and overcomes the limitations of a single attention mechanism. This design systematically alleviates key bottlenecks in YOLOv11 regarding feature adaptation, multi-scale processing, and background interference, achieving an overall performance improvement. Furthermore, this module can be directly embedded without altering the overall YOLO architecture and using the same parameter configurations, exhibiting excellent plug-and-play characteristics and providing convenience for practical applications.

[0024] Preferably, the HSAN_PSA module in S3 first divides the input feature map into two branches through a 1×1 convolution. The b branch is further fed into the hybrid attention part, which consists of multiple cascaded HSAN_PSA_Blocks. Each HSAN_PSA_Block splits the input features into three paths, which are used for PSA self-attention calculation, MSCM-based multi-scale channel feature extraction, and HSAN coordinate attention to capture spatial features, respectively. Finally, this module is embedded into the residual structure to form a composite module.

[0025] When detecting ship targets of different sizes using SAR, traditional methods often miss small or distant targets because C2PSA cannot effectively combine information from different scales, leading to target loss and other problems. This paper introduces a hetero-scale attention network (HSAN) combined with C2PSA. HSAN employs a grouped multi-scale convolutional structure, designed to address the characteristics of ship targets in SAR images, which exhibit significant differences in multi-scale distribution and scattering intensity. Instead of treating feature maps as a whole, this structure groups them and extracts features using convolutional kernels with different receptive fields. This allows for the simultaneous capture of weak scattering point features from small, distant ships and strong scattering structures from large, near-shore ships. After obtaining multi-scale features, HSAN further enhances the response of key targets through an attention mechanism, effectively suppressing interference from ocean clutter and speckle noise. HSAN uses a default four-group structure, effectively capturing spatial information from different receptive fields through multi-scale convolutional modules. First, depthwise separable convolutions are used to extract local contextual features within each group. Subsequently, parallel branches with long, thin convolutional kernels are used sequentially to gradually expand the receptive field, thereby fully extracting multi-scale spatial information. Finally, features from all scales are fused through convolution to generate enhanced group features. Following multi-scale convolution processing, a channel attention mechanism is further introduced to strengthen the representation of key channels while suppressing background noise interference. Simple addition or concatenation of attention mechanisms lacks adaptive weight adjustment. Therefore, HSAN_PSA introduces a progressive scale aggregation mechanism. This mechanism dynamically fuses feature maps from different groups through adaptive weight adjustment, gradually aggregating them in the feature pyramid to form a more discriminative multi-scale representation. Specifically, the outputs of the three paths are adaptively fused using learnable weights, achieving complementary enhancement of multi-dimensional features. This design, while retaining multi-scale perception capabilities, alleviates the gradient vanishing problem in deep networks through skip connections, improving training stability and feature reuse efficiency. In SAR image target detection, HSAN_PSA_Block can be used as a plug-and-play basic unit, flexibly embedded into different layers of the backbone network to support end-to-end detection of multi-scale targets. The HSAN_PSA module utilizes grouped convolutions and multi-branch structures to collaboratively capture multi-scale ship targets. Its built-in channel attention effectively filters out ocean clutter and speckle noise, while the PSA mechanism enhances the flexibility and efficiency of feature fusion through dynamic weight adjustment. This module can be embedded as a general-purpose unit into various detection networks to improve the model's generalization performance.

[0026] Preferably, the spatially separable self-attention mechanism in S4 is used to replace SMA in CBMA. It uses a local-global attention alternation mechanism to summarize local attention and generate global attention feature values ​​based on grouped computation.

[0027] The CBAM module enhances feature representation through a cascaded channel and spatial attention mechanism. Channel attention utilizes global average pooling and MLP to generate weights, highlighting key channels; spatial attention, based on global max pooling and MLP, focuses on important spatial regions. Finally, the two attention maps are multiplied to weight the original features. However, in SAR images, due to speckle noise and low contrast, the CBAM spatial attention module relies solely on simple convolution operations to capture spatial dependencies, lacking multi-scale perception capabilities and easily leading to the loss of small target features. Therefore, the SSSA module is used to replace the original spatial attention submodule.

[0028] Preferably, the spatially separable self-attention mechanism in S4 first uses local grouping attention to divide the input two-dimensional feature map into a series of non-overlapping regular sub-windows. Within each sub-window, multi-head self-attention computation is performed independently. Secondly, global sampling attention is introduced to capture long-distance and global contextual information. Then, convolution is performed on each sub-window to extract representative features, and attention interaction is performed with other sub-windows.

[0029] By replacing the SAM in CBAM with a Spatial Separable Self-Attention (SSSA) mechanism, the shortcomings of CBAM in global context modeling, attention granularity, and multi-scale adaptive capabilities are systematically addressed. SSSA, with its self-attention-based dynamic dependency modeling, multi-head fine-grained attention, and multi-scale mechanism combining local grouping and global sampling, provides the model with richer and more flexible spatial attention, making it particularly advantageous for capturing small target details that are easily overlooked in remote sensing images.

[0030] The beneficial effects of this invention are:

[0031] In this invention, the C3K2 and CBMA modules are improved by combining a multi-scale attention mechanism, and HSAN and PSA are effectively fused to replace the C3K2 and C2PSA modules in YOLOv11. The addition of a new attention mechanism enhances the expression and detection performance of target features. Experimental results show that, through multi-source data experiments and compared to other mainstream detection algorithms, the method of this invention exhibits stronger robustness and adaptability under complex sea backgrounds and diverse imaging conditions. This verifies the effectiveness of its structural improvements in enhancing the model's generalization ability, providing a reliable solution for SAR ship detection in practical applications, and thus facilitating its widespread use. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 These are SAR images representing a specific embodiment of the present invention;

[0034] Figure 2 This is an improved Yolov11 network model according to a specific embodiment of the present invention;

[0035] Figure 3 This is an improved C3K2-SC module according to a specific embodiment of the present invention;

[0036] Figure 4 This is a model of CBMA, a specific embodiment of the present invention;

[0037] Figure 5 This is a model of SSSA, a specific embodiment of the present invention;

[0038] Figure 6 This is an improved CBSA module according to a specific embodiment of the present invention;

[0039] Figure 7 This is the HSAN module in a specific embodiment of the present invention;

[0040] Figure 8 This is the HSAN_PSA module in a specific embodiment of the present invention;

[0041] Figure 9 This is the HSAN_PSA_Block module in a specific embodiment of the present invention;

[0042] Figure 10 This refers to SAR ship tag information in a specific embodiment of the present invention;

[0043] Figure 11 These are the detection results of each model in a specific embodiment of the present invention. Detailed Implementation

[0044] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0045] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0046] Please see Figures 1 to 11 A method for using a SAR ship image target detection method based on C3K2-SC-Yolov11 includes the following steps:

[0047] S1: The SimAM and CPCA attention mechanisms are combined to form a new attention mechanism, SCfusion;

[0048] S2: Combine the attention mechanism SCfusion formed in S1 with the C3K2 module to generate a new C3K2-SC module;

[0049] S3: The HSAN_PSA module is designed by introducing a multi-scale attention network;

[0050] S4: Introduce a spatially separable self-attention mechanism for CBMA.

[0051] In S1, the SimAM and CPCA attention mechanisms employ parallel branched convolution fusion.

[0052] The energy function of SimAM in S1 is:

[0053]

[0054] In the formula, ω i and b i Let t ∈ i and x represent the weights and biases, respectively. i The target neuron and other neurons in the channel, M is the number of neurons in the channel, and i is the spatial index.

[0055] In the CPCA attention mechanism of S1, the channel attention module weights the importance of each channel, highlighting important feature channels and suppressing irrelevant channels. The input features are then processed through average pooling and max pooling before being mapped into a shared multilayer perceptron to obtain the channel weight vector, as shown in the formula:

[0056] M c(F)=σ(Mlp(AvgPool(F))+MLP(MaxPool(F)));

[0057] In the formula, σ is the Sigmoid activation function.

[0058] The C3K2-SC module in S2 is used to replace the C3K2 module in the YOLOv11 system.

[0059] The HSAN_PSA module in S3 first divides the input feature map into two branches using a 1×1 convolution. The b branch is further fed into the hybrid attention part, which consists of multiple cascaded HSAN_PSA_Blocks. Each HSAN_PSA_Block splits the input features into three paths, which are used for PSA self-attention calculation, MSCM-based multi-scale channel feature extraction, and HSAN coordinate attention to capture spatial features. Finally, this module is embedded into the residual structure to form a composite module.

[0060] The spatially separable self-attention mechanism in S4 is used to replace SMA in CBMA. It uses a local-global attention alternation mechanism to summarize local attention and generate global attention feature values ​​based on grouped computation.

[0061] The spatially separable self-attention mechanism in S4 first uses local grouping attention to divide the input two-dimensional feature map into a series of non-overlapping regular sub-windows. Within each sub-window, multi-head self-attention computation is performed independently. Secondly, global sampling attention is introduced to capture long-distance and global contextual information. Then, convolution is performed on each sub-window to extract representative features, and attention interaction is performed with other sub-windows.

[0062] The following is in conjunction with the appendix Figure 1-11 Specific embodiments are described in detail below. The content of this invention is not limited to any specific embodiment, nor does it represent the best embodiment; common alternatives well known to those skilled in the art are also covered within the scope of protection of this invention.

[0063] Example 1

[0064] S1: The SimAM and CPCA attention mechanisms are combined to form a new attention mechanism, SCfusion;

[0065] S2: Combine the attention mechanism SCfusion formed in S1 with the C3K2 module to generate a new C3K2-SC module;

[0066] S3: The HSAN_PSA module is designed by introducing a multi-scale attention network;

[0067] S4: Introduce a spatially separable self-attention mechanism for CBMA.

[0068] To verify the effectiveness of the improved modules, simulation experiments were conducted on each group of added modules. The experiments were all conducted under the same environment and parameters. The results are shown in Table 1 below, which effectively demonstrates the role of each module in improving detection accuracy.

[0069] Table 1 Ablation Experiment

[0070]

[0071] To verify the effectiveness of each improved module, a systematic ablation experiment was conducted in a unified experimental environment, and the results are shown in Table 1. The experimental results show that all improved modules contribute positively to the model performance. Specifically, introducing the C3K2-SC module alone (Model A) significantly improves precision (P), recall (R), and both mAP metrics. Further adding the HSAN_PSA module (Model B) or the CBSA attention mechanism (Model C) further optimizes the model across multiple metrics, particularly in precision and mAP50. Finally, the complete model integrating all three modules (Model D) achieves the best overall performance, with an average improvement of approximately 4 percentage points in its main detection metrics compared to the original YOLOv11 baseline model. This effectively verifies the effectiveness and synergistic effect of the C3K2-SC structure, the HSAN_PSA module, and the CBSA attention mechanism in enhancing the model's detection accuracy and generalization ability.

[0072] To verify the detection performance of the C3K2-SC-Yolov11 model, it was compared with the Yolov11, Yolov10s, ACyolov11, and RTDETR detection models, respectively. All experiments were conducted under the same environment. The results are shown in Table 2 below.

[0073] Table 2 Comparison of experimental results

[0074]

[0075] As shown in Table 2, the C3K2-SC-YOLOv11 model proposed in this application achieved the best performance in four core metrics—precision (P), recall (R), mAP50, and mAP50-95—in a horizontal comparison with various mainstream detection algorithms.

[0076] The YOLOv10s and RT-DETR models performed relatively poorly in this experimental environment, with all metrics falling below the baseline model YOLOv11. The baseline YOLOv11 model significantly outperformed the former, especially in accuracy, where it led by approximately 3 percentage points. ACYOLOv11 demonstrated strong competitiveness, but its overall accuracy, particularly the mAP50-95 metric, still lagged behind that of this application.

[0077] C3K2-SC-YOLOv11 achieved comprehensive superiority, with its mAP50 and mAP50-95 improving by approximately 6 and 9.6 percentage points respectively compared to the benchmark YOLOv11. This result fully demonstrates that by introducing modules such as C3K2-SC, HSAN_PSA, and CBSA, the model's capabilities in feature extraction, complex background suppression, and multi-scale target fusion have been substantially enhanced. This significantly improves detection accuracy and overall generalization performance while maintaining high recall, validating the effectiveness of its structural improvements and demonstrating its potential for application in real-world scenarios.

[0078] To verify the performance of the C3K2-SC-Yolov11 model in ship target detection, a comparative analysis of the visualization results revealed significant differences among the models. The detection results are as follows: Figure 11 As shown.

[0079] Figure 11 The image shows the detection results of various models. (a) shows the true labels of ships in the image, and (b), (c), (d), (e), and (f) show the detection results of the models, respectively. As shown in the figure, false detections occurred in the detection results based on Yolov11. In the red box, there were 8 false positives when there were only 4 labels. False positives occurred in the detection results based on Yolov10 and ACYOLOv11, indicating missed detections. RTDETR had the most false positives, suggesting that the Transformer-based detector has poor generalization performance in the current task, possibly due to the distribution of training data. The C3K2-SC-Yolov11 model proposed in this application has a false positive confidence of only 0.3, far below the effective detection threshold. In practical applications, these false positives can be easily filtered out through post-processing, and all true targets were correctly detected. The C3K2-SC-Yolov11 model shows significant advantages in detection accuracy and stability. Its improved C3K2 module effectively enhances feature extraction capabilities, and the SC attention mechanism strengthens the focus on key regions, thus achieving more accurate ship localization in complex marine environments. Compared with the baseline model, this application significantly reduces the false positive rate while maintaining a high recall rate, demonstrating the effectiveness and practicality of the proposed improvement strategy.

[0080] To verify the generalization ability of the proposed algorithm, it was tested on the SSDD dataset. The SSDD dataset contains 1160 images and 2456 ship targets, with an average of 2.12 ships per image. These images were mainly acquired by RadarSat-2, TerraSAR-X, and Sentinel-1 sensors, covering four polarization modes: HH, HV, VV, and VH, with resolutions ranging from 1m to 15m. The images in the dataset cover a large area of ​​sea and nearshore regions, demonstrating the diversity of ship targets against different backgrounds. The results of the comparison using the aforementioned detection algorithm are shown in the table below.

[0081] Table 3 Comparison of Model Generalization Ability Results

[0082]

[0083] As shown in Table 3, in the cross-dataset SAR image ship detection task, YOLOv10 outperforms YOLOv11 and RTDETR, indicating that YOLOv10 has stronger generalization ability than both, achieving a 7 percentage point improvement in precision and mAP50. While YOLOv10s outperform YOLOv11 in precision, its recall is lower; RTDETR's overall performance is poor; ACYOLOv11 demonstrates excellent generalization performance, with precision and mAP50 second only to the method in this application, but its recall remains a weakness. The C3K2-SC-Yolov11 method proposed in this application exhibits excellent generalization performance, significantly outperforming other comparative methods in four metrics: precision, recall, mAP50, and mAP50-95. Particularly in precision and mAP50, C3K2-SC-Yolov11 shows improvements of approximately 16% and 10% respectively compared to the baseline model Yolov11, demonstrating its ability to maintain high detection accuracy and robustness even under complex sea surface backgrounds and multi-source SAR imaging conditions. In summary, C3K2-SC-Yolov11 significantly improves detection precision and average accuracy while maintaining good recall, validating its good adaptability and generalization ability across sensor, multi-polarization, and multi-resolution SAR images, and its potential for application in practical ship detection tasks.

[0084] Experimental results show that, through multi-source data experiments and compared with other mainstream detection algorithms, this application demonstrates stronger robustness and adaptability under complex sea surface backgrounds and diverse imaging conditions, verifying the effectiveness of its structural improvements in enhancing the model's generalization ability and providing a reliable solution for SAR ship detection in practical applications. However, the model has high computational complexity, and a lightweight model needs to be designed to reduce computational load and facilitate integration into system equipment.

[0085] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A SAR ship image target detection method based on C3K2-SC-Yolov11, characterized in that, Includes the following steps: S1: The SimAM and CPCA attention mechanisms are combined to form a new attention mechanism, SCfusion; S2: Combine the attention mechanism SCfusion formed in S1 with the C3K2 module to generate a new C3K2-SC module; S3: The HSAN_PSA module is designed by introducing a multi-scale attention network; S4: Introduce a spatially separable self-attention mechanism for CBMA.

2. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, In S1, the SimAM and CPCA attention mechanisms employ parallel branched convolution fusion.

3. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, The energy function of SimAM in S1 is: In the formula, ω i and b i Let t ∈ i and x represent the weights and biases, respectively. i The target neuron and other neurons in the channel, M is the number of neurons in the channel, and i is the spatial index.

4. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, In the CPCA attention mechanism of S1, the channel attention module weights the importance of each channel, highlighting important feature channels and suppressing irrelevant channels. The input features are then processed through average pooling and max pooling before being mapped into a shared multilayer perceptron to obtain the channel weight vector, as shown in the formula: M c (F)<σ(Mlp(AvgPool(F))+MLP(MaxPool(F))); In the formula, σ is the Sigmoid activation function.

5. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, The C3K2-SC module in S2 is used to replace the C3K2 module in the YOLOv11 system.

6. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, The HSAN_PSA module in S3 first divides the input feature map into two branches through a 1×1 convolution. The b branch is further fed into the hybrid attention part, which consists of multiple cascaded HSAN_PSA_Blocks. Each HSAN_PSA_Block splits the input features into three paths, which are used for PSA self-attention calculation, MSCM-based multi-scale channel feature extraction, and HSAN coordinate attention to capture spatial features. Ultimately, this module is embedded into the residual structure to form a composite module.

7. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, The spatially separable self-attention mechanism in S4 is used to replace SMA in CBMA. It uses a local-global attention alternation mechanism to summarize local attention and generate global attention feature values ​​based on grouped computation.

8. The SAR ship image target detection method based on C3K2-SC-Yolov11 as described in claim 1, characterized in that, The spatially separable self-attention mechanism in S4 first uses local grouping attention to divide the input two-dimensional feature map into a series of non-overlapping regular sub-windows. Within each sub-window, multi-head self-attention computation is performed independently. Secondly, global sampling attention is introduced to capture long-distance and global contextual information. Then, convolution is performed on each sub-window to extract representative features, and attention interaction is performed with other sub-windows.

Images