SAR ship target detection method in complex environment

Abstract

The invention provides an SAR ship target detection method in a complex environment. An SAR ship detection model is adopted to identify an SAR ship target. The SAR ship detection model adopts a YOLOv11n backbone and comprises a backbone network, a neck network and a head network; the backbone network comprises five convolution modules, four lightweight feature extraction modules, an SPPF module and an anti-noise enhancement module; the neck network comprises five lightweight feature extraction modules, five connection modules, three dynamic sampling modules and two convolution modules; the head network comprises three detection modules; according to the invention, a plurality of innovative modules, including a lightweight feature extraction module (PCA), an anti-noise enhancement module (PSA-G), a dynamic sampling module (DySample) and a multi-scale small target fusion network (MSTFNet), are introduced into a backbone and neck network, so that noise resistance, light weight and high precision are taken into account.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image target detection technology, and more specifically, to a SAR ship target detection method in complex environments. Background Technology

[0002] Synthetic Aperture Radar (SAR) plays a central role in maritime surveillance and ship inspection due to its all-weather, all-day imaging capabilities. Traditional SAR ship detection mainly relies on the Constant False Alarm Rate (CFAR) algorithm, which achieves detection by statistically analyzing background noise features. However, in complex sea conditions, it often suffers from problems such as missed detection of small targets and high false alarm rates.

[0003] With the development of deep learning, single-stage detectors such as YOLO and SSD have achieved a balance between detection speed and accuracy. For example, Chinese invention patent application CN120747883A, published on October 3, 2025, entitled "A Ship Detection Method for Complex SAR Scenes," discloses a ship detection method for complex SAR scenes. This method is based on an improved YOLOv11 network. It introduces the DPSConv module into the backbone network to capture multi-scale contextual information using the receptive field while preserving subtle feature details. In the neck network, it introduces the PromptFusionMod module to perform multi-modal feature fusion through four processing stages: spatial compression and cue fusion, efficient attention mechanism, lightweight multilayer perceptron, and output refinement. The original detection head is replaced with the MscaleASFFHead detection head. This invention aims to improve the accuracy and robustness of ship detection in complex sea conditions by balancing fine-grained feature extraction, cross-scale semantic alignment, and lightweight deployment.

[0004] However, single-stage detectors such as YOLO and SSD still have significant shortcomings in their detection performance when dealing with speckle noise and multi-scale target features in SAR images. These shortcomings include weak feature extraction capabilities, large noise interference, and difficulty in balancing lightweight design with accuracy.

[0005] Therefore, how to achieve noise-resistant, lightweight, and high-precision SAR ship detection has become a research hotspot. Summary of the Invention

[0006] To address the problems of severe speckle noise interference, easy loss of target features, and difficulty in balancing model lightweighting and accuracy in complex SAR environments, this invention provides a SAR ship target detection method for complex environments, aiming to improve the robustness and accuracy of ship detection, especially enhancing the detection capability of dense small targets.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for SAR ship target detection in complex environments, which uses a SAR ship detection model to identify SAR ship targets; The SAR ship detection model adopts the YOLOv11n backbone, including the backbone network, neck network and head network; The backbone network includes five convolutional modules, four lightweight feature extraction modules, an SPPF module, and a noise reduction and enhancement module; The first convolution module, the second convolution module, the first lightweight feature extraction module, the third convolution module, the second lightweight feature extraction module, the fourth convolution module, the third lightweight feature extraction module, the fifth convolution module, the fourth lightweight feature extraction module, the SPPF module, and the noise reduction enhancement module are connected in sequence. The first lightweight feature extraction module outputs P3 layer features, the second lightweight feature extraction module outputs P4 layer features, the third lightweight feature extraction module outputs P5 layer features, and the noise reduction and enhancement module outputs P6 layer features. The neck network includes five lightweight feature extraction modules, five connection modules, three dynamic sampling modules, and two convolutional modules; The first dynamic sampling module, the first connection module, the fifth lightweight feature extraction module, the second dynamic sampling module, the second connection module, the sixth lightweight feature extraction module, the third dynamic sampling module, and the third connection module are connected in sequence; The seventh lightweight feature extraction module, the sixth convolution module, the fourth connection module, the eighth lightweight feature extraction module, the seventh convolution module, the fifth connection module, and the ninth lightweight feature extraction module are connected in sequence; The output of the P3 layer features is input to the third connection module, and the output of the third connection module is input to the seventh lightweight feature extraction module. The output P4 layer features are input to the second connection module on one hand, and to the fourth connection module on the other hand. The output P5 layer features are input to the first connection module on one hand, and to the fifth connection module on the other hand. The output P6 layer features are input into the first dynamic sampling module; The output of the fifth lightweight feature extraction module is also input to the fifth connection module, and the output of the sixth lightweight feature extraction module is also input to the fourth connection module; The seventh lightweight feature extraction module outputs small target detection features; The target detection features are output from the eighth lightweight feature extraction module. The ninth lightweight feature extraction module outputs large target detection features; The head network includes three detection modules; The small target detection features are input into the first detection module for detection, and the first detection result is output. The target detection features are input into the second detection module for detection, and the second detection result is output. Large target detection features are input into the third detection module for detection, and the third detection result is output. The lightweight feature extraction module uses partial convolution combined with ECA channel attention mechanism to focus on the effective pixel area through dynamic mask update, thereby suppressing the influence of speckle noise. The noise reduction and enhancement module extracts gradient features through the multi-scale Sobel operator and uses a lightweight threshold generator to dynamically predict the spatial adaptive threshold, thereby achieving noise gradient truncation and suppression. The dynamic sampling module uses point sampling as its core idea to achieve dynamic upsampling. It predicts the sampling offset and generates a dynamic scope through a linear layer.

[0008] This invention achieves a balance between noise resistance, lightweight design, and high accuracy by introducing several innovative modules into the backbone and neck networks, including a lightweight feature extraction module (PCA), a noise-resistant enhancement module (PSA-G), a dynamic sampling module (DySample), and a multi-scale small target fusion network (MSTFNet).

[0009] The neck network in this invention is MSTFNet (Multi-Scale Target Fusion Network), which integrates PCA feature outputs from different levels, adds a 160×160 detection layer to focus on small ship targets, and removes the original 20×20 large target detection head. At the same time, it improves multi-scale detection capability and robustness through a cross-layer feature interaction mechanism.

[0010] Preferably, the lightweight feature extraction module includes a multi-layered alternating structure composed of an ECA channel attention module and a PConv partial convolution module, and the processing flow includes: For the input feature map, preliminary feature extraction is first performed, and then the extracted features are divided into two parts; One part of the features is first processed by halving the number of channels to effectively reduce the computational load; then the remaining spatial features after halving the number of channels are input into a multi-layer alternating structure, extracted by PConv partial convolution operation, and ECA attention mechanism is applied at the front end of PConv partial convolution to enhance feature representation. Finally, the features extracted from the multi-layer alternating structure are concatenated with another set of features along the channel dimension, and then fused and output through pointwise convolution to further reduce the amount of computation and generate the output.

[0011] The lightweight feature extraction module (Partial Convolution Attention, PCA) employs partial convolution (PConv) combined with ECA channel attention mechanism. It focuses on the effective pixel region through dynamic mask updates, suppressing the influence of speckle noise. The module structure includes multiple alternating layers of PConv and ECA, and finally outputs the feature map through pointwise convolution. While ensuring computational efficiency, it significantly improves the ship feature extraction capability in noisy or data-missing scenarios.

[0012] Preferably, the noise reduction enhancement module includes an AGT module, a first Split module, two DynamicPSA Block modules, a sixth connection module, a DWConv module, and a first CBS module; The tensors input to the noise reduction and enhancement module are first input to the AGT module; The output of the AGT module is connected to the first Split module. The output of the first Split module is processed by the first DynamicPSA module and the second DynamicPSA module connected in sequence, and then output by the second DynamicPSA module to the sixth connection module. The output of the first Split module is also directly input to the sixth connection module. The output of the sixth connection module is connected to the DWConv module and the first CBS module in sequence; The output of the first CBS module is used as the output of the noise reduction enhancement module; The AGT module includes the second Split module, the Multi-Scale Gradient Convolution module, the DynamicThreshold Generator module, the Soft Thresholding module, the Multi-Scale Fusion module, and the seventh connection module; Tensors from the AGT module are first input into the second Split module; The output of the second Split module is processed by the Multi-Scale GradientConvolution module, DynamicThreshold Generator module, Soft Thresholding module and Multi-Scale Fusion module connected in sequence, and then output by the Multi-Scale Fusion module to the seventh connection module. The output of the second Split module is also directly input to the seventh connection module. The output of the seventh connection module is used as the output of the AGT module; The DynamicPSA Block module includes a third Split module, a Multihead Attention module, two CBS modules, a scaling factor adjustment module, and an eighth connection module; Tensors from the DynamicPSA Block module are first input into the third Split module; The output of the third Split module is processed by the Multihead Attention module, the second CBS module, and the scaling factor adjustment module connected in sequence, and then output by the scaling factor adjustment module to the eighth connection module. The output of the third Split module is also directly input to the eighth connection module. The scaling factor adjustment module also feeds back the scaling factor to the Multihead Attention module to dynamically adjust the attention weight distribution. The output of the eighth connection module is input to the third CBS module, and the output of the third CBS module is used as the output of the DynamicPSABlock module.

[0013] The noise-enhancing module (Point-State Attention with Adaptive Gradient Threshold, PSA-G) extracts gradient features through a multi-scale Sobel operator and dynamically predicts spatially adaptive thresholds using a lightweight threshold generator to achieve noise gradient truncation and suppression. Subsequently, a dynamic point-state attention mechanism (DPSA) is introduced for feature weighting and dynamic calibration to strengthen the saliency representation of the target and improve the model's adaptability to complex noisy environments.

[0014] Preferably, the method for generating sampling points in the dynamic sampling module is as follows: Input feature map; Generate a basic coordinate grid to generate the basic sampling position in the input space for each output feature; The input feature map is processed by a lightweight linear layer to predict and generate the original offset field G; Based on the gain parameter g of the generated original offset field G, as well as the target height sH and target width sW, calculate the static scope factor S; Add the scaled offset to the base coordinate grid; Bilinear interpolation is used to sample from the input feature map.

[0015] The Dynamic Sampling Module (DySample) implements dynamic upsampling based on point sampling. It predicts sampling offsets and generates dynamic scopes through linear layers, eliminating the need for additional convolutional kernels, thus reducing computational overhead and enhancing the recovery of details from small targets. This module is applied in the upsampling stage of the neck network, effectively mitigating the problem of feature information loss.

[0016] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the SAR ship target detection method in a complex environment as described above.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, characterized in that the computer program / instructions, when executed by a processor, implement the SAR ship target detection method in a complex environment as described above.

[0018] A fourth aspect of the present invention provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the SAR ship target detection method in a complex environment as described above.

[0019] This invention has outstanding substantive features and significant progress compared to the prior art, specifically: With the increasing demand for lightweight models and high-precision target detection, the PPDM-YOLO model proposed in this invention provides an effective solution for SAR ship detection tasks, achieving a good balance between accuracy and computational constraints. This model addresses common challenges in SAR images, such as speckle noise interference, significant variations in target size, and the problem of densely distributed small targets, by constructing a multi-module collaborative, feature-rich detection framework suitable for resource-constrained environments.

[0020] First, the PCA module analyzes the redundant relationships between feature map channels and integrates the ECA attention mechanism to compress redundant features while enhancing key information. The PCA module not only effectively reduces model parameters and computational overhead but also improves the efficiency and robustness of feature extraction. Second, to address severe noise interference in SAR images, a PSA-G module is proposed. Combining adaptive gradient thresholding and a dynamic spatial attention mechanism, it suppresses ineffective noise interference while preserving detailed features, achieving synergistic optimization between feature decoupling and dynamic calibration, further enhancing the model's adaptability to complex backgrounds.

[0021] In the feature reconstruction and upsampling stages, the DySample module is introduced. Compared with traditional upsampling methods, DySample adopts a mechanism of "dynamically distributing a single point to multiple points," paying more attention to the subtle structure of small target regions and avoiding common feature loss and inefficient learning problems, thus significantly improving the quality of feature recovery. MSTFNet employs a cross-layer feature fusion strategy, integrating features from adjacent layers while enhancing information flow between different semantic layers. This allows high-level abstract semantics to propagate to low-level detailed information, significantly improving the model's ability to detect small targets and its spatial localization accuracy.

[0022] Experimental results demonstrate that the PPDM-YOLO model achieves leading performance on two public datasets, HRSID and SSDD. On the HRSID dataset, it achieves 93.7% mAP50 and 70.3% mAP50-95, while on the SSDD dataset, it achieves 99.3% and 78.7%, respectively. Compared to YOLOv11n, this model reduces the number of parameters by 34.7% while surpassing YOLOv11n in accuracy, fully validating its comprehensive advantages in lightweight design, accuracy, and robustness.

[0023] Visualization results demonstrate that the PPDM-YOLO model exhibits strong detection performance across a variety of complex marine backgrounds, particularly excelling in scenarios with densely distributed small targets. It is worth noting that while this model significantly reduces overall computational costs, there is still room for optimization in terms of memory usage when used with extremely resource-constrained edge computing devices.

[0024] Currently, much research focuses on improving experimental accuracy through large-scale models, often neglecting the need for lightweight implementation in practical deployments. SAR ship images present complex backgrounds due to speckle noise and the presence of small, dense targets. To address these challenges, this invention's PPDM-YOLO model integrates PCA, PSA-G, MSTFNet, and DySample modules, possessing not only powerful feature extraction capabilities but also effective fusion of multi-scale features, particularly excelling in the detection of dense small targets. Compared to existing methods such as YOLOv11n, the PPDM-YOLO model achieves higher detection accuracy while significantly reducing the number of parameters. Extensive experiments on two benchmark datasets validate the effectiveness of each module (PCA, PSA-G, MSTFNet, and DySample). The PPDM-YOLO model is well-suited for deployment on various hardware platforms in diverse and complex SAR ship detection scenarios. Attached Figure Description

[0025] Figure 1 This is the overall flowchart of the PPDM-YOLO model.

[0026] Figure 2 This is a schematic diagram of the network structure of the PPDM-YOLO model.

[0027] Figure 3 This is a schematic diagram of the PCA module structure.

[0028] Figure 4 This is a schematic diagram of the PSA-G module structure.

[0029] Figure 5 This is a diagram of the DySample sampling point generation mechanism (including static and dynamic scopes).

[0030] Figure 6 This is a schematic diagram of the MSTFNet multi-scale fusion small target detection network.

[0031] Figure 7 This is a visualization comparison of the detection results from the HRSID dataset.

[0032] Figure 8 This is a visualization comparison of the detection results from the SSDD dataset.

[0033] Figure 9 This is a map showing the results of dense target detection using SAR in the open sea.

[0034] Figure 10 This is a map showing the results of dense target detection using SAR in nearshore waters. Detailed Implementation

[0035] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0036] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion.

[0037] To facilitate understanding of the technical solutions provided in this application, some technical terms involved in the embodiments of this application will be explained below.

[0038] Depthwise Separable Convolution (DSC) is not a single convolutional layer, but rather an efficient convolutional operation composed of two independent steps: depthwise convolution and pointwise convolution. It is a core component of modern lightweight neural network architectures (such as MobileNet and EfficientNet).

[0039] The module is a fundamental building block in the YOLO series of object detection models (such as YOLOv5 and YOLOv7), and its full name is the Convolution-Batch Normalization-Activation Function module. It consists of three parts connected in series: Convolutional layer (Conv): Responsible for feature extraction, scanning the input data through convolutional kernels; BatchNorm: Normalizes the convolutional output, accelerating training and improving model stability; Activation function layers (such as SiLU): introduce nonlinear factors, enabling the model to learn complex mapping relationships.

[0040] A module is a fundamental building block in the YOLO series of models, typically consisting of convolutional layers (Conv), batch normalization layers (BatchNorm), and activation functions (such as Hardswish). Its core function is to extract and transform information in the spatial and channel dimensions of the input feature map.

[0041] In YOLO, the module is mainly used to stitch feature maps of different levels together along the channel dimension, thereby fusing multi-scale information and improving the accuracy of object detection.

[0042] Example 1 like Figure 1 As shown, this embodiment provides a SAR ship target detection method in complex environments, which uses the SAR ship detection model (PPDM-YOLO) to identify SAR ship targets; The SAR ship detection model adopts the YOLOv11n backbone, which includes a backbone network, a neck network (MSTFNet), and a head network.

[0043] Backbone network The backbone network includes five convolutional modules (Conv), four lightweight feature extraction modules (PCA), an SPPF module, and a noise reduction enhancement module (PSA-G). The first Conv module, the second Conv module, the first PCA module, the third Conv module, the second PCA module, the fourth Conv module, the third PCA module, the fifth Conv module, the fourth PCA module, the SPPF module, and the PSA-G module are connected in sequence; The first PCA module outputs P3 layer features, the second PCA module outputs P4 layer features, the third PCA module outputs P5 layer features, and the PSA-G module outputs P6 layer features.

[0044] Neck network (MSTFNet, Multi-Scale Target Fusion Network) The neck network includes five lightweight feature extraction modules (PCA), five connection modules (Concat), three dynamic sampling modules (DySample), and two convolutional modules (Conv). The first DySample module, the first Concat module, the fifth PCA module, the second DySample module, the second Concat module, the sixth PCA module, the third DySample module, and the third Concat module are connected in sequence; The seventh PCA module, the sixth Conv module, the fourth Concat module, the eighth PCA module, the seventh Conv module, the fifth Concat module, and the ninth PCA module are connected in sequence; The output P3 layer features are input into the third Concat module, and the output of the third Concat module is input into the seventh PCA module. The output P4 layer features are input into the second Concat module on one hand, and into the fourth Concat module on the other. The output P5 layer features are input to the first Concat module on one hand, and to the fifth Concat module on the other. The output P6 layer features are input into the first DySample module; The output of the fifth PCA module is also input into the fifth Concat module, and the output of the sixth PCA module is also input into the fourth Concat module; The seventh PCA module outputs small target detection features; Target detection features in the output of the eighth PCA module; The ninth PCA module outputs large target detection features.

[0045] Head network The head network includes three detection modules. The small target detection features are input into the first Detect module for detection, and the first detection result is output. The target detection features are input into the second Detect module for detection, and the second detection result is output. Large target detection features are input into the third Detect module for detection, and the third detection result is output.

[0046] Module (Partial Convolution Attention) By employing partial convolution combined with ECA channel attention mechanism, the effective pixel area is focused through dynamic mask update, thereby suppressing the influence of speckle noise.

[0047] The PCA module consists of a multi-layered alternating structure composed of an ECA channel attention module and a PConv partial convolutional module. The processing flow includes: For the input feature map, preliminary feature extraction is first performed, and then the extracted features are divided into two parts; One part of the features is first processed by halving the number of channels to effectively reduce the computational load; then the remaining spatial features after halving the number of channels are input into a multi-layer alternating structure, extracted by PConv partial convolution operation, and ECA attention mechanism is applied at the front end of PConv partial convolution to enhance feature representation. Finally, the features extracted from the multi-layer alternating structure are concatenated with another set of features along the channel dimension, and then fused and output through pointwise convolution to further reduce the amount of computation and generate the output.

[0048] Module (Point-State Attention with Adaptive Gradient Threshold) Gradient features are extracted using the multi-scale Sobel operator, and a lightweight threshold generator is used to dynamically predict spatially adaptive thresholds to achieve noise gradient truncation and suppression.

[0049] The noise reduction enhancement module includes an AGT module, a first Split module, two DynamicPSA Block modules, a sixth connection module, a DWConv module, and a first CBS module; The tensors input to the noise reduction and enhancement module are first input to the AGT module; The output of the AGT module is connected to the first Split module. The output of the first Split module is processed by the first DynamicPSA module and the second DynamicPSA module connected in sequence, and then output by the second DynamicPSA module to the sixth connection module. The output of the first Split module is also directly input to the sixth connection module. The output of the sixth connection module is connected to the DWConv module and the first CBS module in sequence; The output of the first CBS module is used as the output of the noise reduction enhancement module; The AGT module includes the second Split module, the Multi-Scale Gradient Convolution module, the DynamicThreshold Generator module, the SoftThresholding module, the Multi-Scale Fusion module, and the seventh connection module; Tensors from the AGT module are first input into the second Split module; The output of the second Split module is processed by the Multi-Scale GradientConvolution module, DynamicThreshold Generator module, Soft Thresholding module and Multi-Scale Fusion module connected in sequence, and then output by the Multi-Scale Fusion module to the seventh connection module. The output of the second Split module is also directly input to the seventh connection module. The output of the seventh connection module is used as the output of the AGT module; The DynamicPSA Block module includes a third Split module, a Multihead Attention module, two CBS modules, a scaling factor adjustment module, and an eighth connection module; Tensors from the DynamicPSA Block module are first input into the third Split module; The output of the third Split module is processed by the Multihead Attention module, the second CBS module, and the scaling factor adjustment module connected in sequence, and then output by the scaling factor adjustment module to the eighth connection module. The output of the third Split module is also directly input to the eighth connection module. The scaling factor adjustment module also feeds back the scaling factor to the Multihead Attention module to dynamically adjust the attention weight distribution. The output of the eighth connection module is input to the third CBS module, and the output of the third CBS module is used as the output of the DynamicPSABlock module.

[0050] Module Dynamic upsampling is achieved by using point sampling as the core idea, and sampling offset is predicted and dynamic scope is generated through linear layers.

[0051] The method for generating sampling points in the dynamic upsampling module is as follows: Input feature map; Generate a basic coordinate grid to generate the basic sampling position in the input space for each output feature; The input feature map is processed by a lightweight linear layer to predict and generate the original offset field G; Based on the gain parameter g of the generated original offset field G, as well as the target height sH and target width sW, calculate the static scope factor S; Add the scaled offset to the base coordinate grid; Bilinear interpolation is used to sample from the input feature map.

[0052] Model Evaluation To analyze the object detection training results of the model, the model is evaluated from the following aspects: The metrics include mean average precision (mmAP), precision (P), recall (R), parameters, frames per second (FPS), and confusion matrix. The confusion matrix summarizes the true and predicted labels in matrix form, as shown in Table I. TP (TruePositive): Both predicted and true labels are positive examples; FN (False Negative): The predicted label is negative but the true label is positive; FP (False Positive): The predicted label is positive but the true label is negative; TN (TrueNegative): Both predicted and true labels are negative examples.

[0053] Table I Confusion Matrix

[0054] Precision (P): The proportion of samples that the detection model predicts to be positive, but which are actually positive. The formula is: P = TP / (TP + FP); Recall (R): The proportion of samples that are actually positive that are correctly detected as positive. The formula is: R = TP / (TP + FN); mmAP: The average mAP value across multiple categories. The average accuracy (mAP) for each category is calculated, and then the average is taken as the overall performance metric. The calculation formula is: In the formula, c represents the total number of categories.

[0055] To evaluate the performance advantages of the PPDM-YOLO model in SAR image target detection, seven representative SAR ship detection models were selected for comparative analysis under the same experimental conditions and parameter settings. These models include mainstream lightweight detection algorithms and classic deep learning detection frameworks, covering different structural designs of one-stage and two-stage approaches. Through horizontal comparison, the comprehensive capabilities of PPDM-YOLO in small target detection, complex background adaptability, and multi-scale feature extraction are fully demonstrated. These models include: SSD, Faster R-CNN, YOLOv5n, YOLOv8n, YOLOv11n, LMSD-YOLO, and MSFA-YOLO. The comparative experiments were conducted on two public datasets: HRSID and SSDD. Tables II and IV show the performance comparison results of each model on the two datasets.

[0056] Table II Comparison of object detection with other methods on the HRSID dataset

[0057] Initial experiments were conducted on the HRSID dataset. Table II shows the accuracy comparison results of each model on this dataset. It can be seen that the PPDM-YOLO model achieved the best performance with an mAP50 of 93.7%, which is 3.1% higher than YOLOv11n.

[0058] To further verify the model's robustness in complex environments, a visual analysis was performed on targets near the shore in the HRSID dataset. For example... Figure 7 As shown, three images with dense targets and complex backgrounds were selected. These scenes pose a significant challenge to target detection. The green boxes represent ground scene annotations, the red boxes represent actually detected ship targets (TP), the blue boxes represent missed ship targets (FN), and the yellow boxes represent falsely detected ship targets (FP). Despite the dense occlusion and coastline interference, the PPDM-YOLO model still demonstrates excellent detection capabilities, stably and accurately identifying various targets. The specific detection performance of different methods in this scene is listed in Table III, further confirming the superior performance of the PPDM-YOLO model in complex scenes.

[0059] Table III Performance Evaluation of Various Ship Inspection Methods on HRSID

[0060] In the visualization results of the HRSID dataset, significant differences in detection performance were observed among the three baseline models in complex nearshore backgrounds. In labeled image 1 (containing 13 ship targets), experimental results show that YOLOv5n produced 4 false positives (FP), while YOLOv8n and YOLOv11n reduced the number of false positives to 3. In labeled image 2 (containing 9 targets), YOLOv5n produced 1 false positive and exhibited target splitting, i.e., one real target was misclassified as two independent targets. YOLOv8n produced 2 false positives, while YOLOv11n optimized to only 1 false positive. In labeled image 3 (dense small target scene, 11 targets), YOLOv5n and YOLOv11n both produced 1 false negative (FN) and 1 false positive, while YOLOv8n performed relatively better with only 1 false positive. The present invention showed no false negatives or false positives in all labeled images.

[0061] Next, a comparative experiment was conducted on the SSDD dataset. Compared to the HRSID dataset, the SSDD dataset has a relatively simpler background.

[0062] Table IV compares object detection with other methods on the SSDD dataset.

[0063] As shown in Table IV, all models exhibited high accuracy. The model of this invention slightly outperformed the others on this dataset, achieving an mAP50 of 99.4%. Furthermore, the model has only 1.71M parameters and a size of only 4.2MB. Although the accuracy differences between models were small, the model of this invention maintained high accuracy while possessing low computational overhead and moderate FPS. Similarly, as... Figure 8 As shown, nearshore targets in the SSDD dataset are visualized. Table V presents detailed results for vessel target detection using different methods on this dataset.

[0064] Table V Performance evaluation of various ship inspection methods on SSDD

[0065] In the visualization analysis of the SSDD dataset, significant differences in target detection performance were observed among the seven models in complex nearshore backgrounds. YOLOv5n had the highest number of missed detections, missing a total of 4 targets (FN). YOLOv8n performed slightly better, missing 2 targets; YOLOv11n further improved, reducing the number of missed detections to 1, but producing 2 false positives (FP). In contrast, the model of this invention performed best, missing only 1 target and detecting almost all targets. A similar trend was observed in labeled image 2 containing 11 ship targets. YOLOv5n performed the worst, missing 6 targets; YOLOv8n reduced the number of missed detections by 1; YOLOv11n performed relatively well, producing only 1 false positive. In labeled image 3 containing 3 ship targets, the three benchmark models YOLOv5n, YOLOv8n, and YOLOv11n all failed to achieve perfect detection, each producing 1 false positive.

[0066] In the model proposed in this invention, the PCA and PSA-G modules effectively address the issues of target feature loss and insufficient feature extraction caused by speckle noise in complex backgrounds of SAR ship images. Furthermore, the introduction of MSTFNet and DySample significantly improves the detection accuracy of small targets. This invention's model significantly outperforms other models while reducing the number of parameters.

[0067] To verify the performance of the method of the present invention in dense target detection, a comparative experiment was conducted using dense SAR ship images in two typical maritime scenarios: nearshore environment and high seas environment. Figure 9 and Figure 10 Visualized detection results are presented in nearshore and high-sea SAR scenarios. Tables VI and VII systematically compare the detection metrics of the model of this invention with three typical CNN models.

[0068] Table VI Comparison of different models in dense object detection against an open ocean background on the HRSID dataset

[0069] Table VII Comparison of Dense Target Detection Results of Different Models in HRSID High Sea Background

[0070] In large-scale multi-target detection in the high seas (see Table VI), the model of this invention successfully detected 49 vessels with 0 false positives (FP) and only 2 false negatives (FN), achieving 100% precision and 96.1% recall. This result significantly outperforms other methods, demonstrating excellent anti-interference capabilities in complex sea clutter environments. As shown in Table VII, in the more challenging high-resolution nearshore SAR scene, the method of this invention still maintains its leading position, achieving 94.4% precision and 91.1% recall, thanks to the optimized feature extraction module and effective noise suppression strategy. However, there are still 3 false negatives (FN) and 5 false positives (FP) in the nearshore scene.

[0071] It is important to emphasize that nearshore detection faces greater challenges compared to high seas conditions. Quantitative analysis shows that the performance of all four models degrades in nearshore scenarios, primarily due to two factors: 1) strong scattering interference from land structures and port facilities reduces target contrast; 2) occlusion effects from dock facilities or large vessels result in incomplete target morphology, increasing the difficulty of feature recognition. The model proposed in this invention effectively addresses these challenges through multi-scale feature fusion and a context-aware module, demonstrating stronger environmental adaptability than traditional CNN methods.

[0072] Furthermore, this invention investigates the lightweight design of the PCA module by comparing the backbone-based PPDM-YOLO model with YOLOv11n. Table VIII shows that in the first five layers, the advantage is not significant due to the limited number of channels. However, as the number of channels increases, the PPDM-YOLO model exhibits a significant improvement in efficiency, particularly in layers 7 and 9, reducing parameters by 40,708 and 160,512, respectively. Overall, the PPDM-YOLO model contains 1.71M parameters, achieving a 34.7% parameter reduction compared to YOLOv11n (2.62M).

[0073] Table VIII compares the parameters of the two models on the backbone network.

[0074] ablation experiment To further evaluate the effectiveness of the PCA module, a comparative analysis was conducted with three alternative attention mechanism models: SE, coordinate attention (CA), and CBAM. Detailed experimental results are shown in Table IX.

[0075] Data shows that compared with the baseline model, mAP50 is improved to 92.5%, and mAP50-95 is improved to 68.4%, proving the effectiveness of the designed lightweight network structure of the PCA module.

[0076] Table IX Comparison Results of Different Attention Mechanism Models

[0077] In Experiment 3, the function of the PSA-G module was investigated. Data showed that mAP50-95 increased by 0.9%, while mAP50 increased by 1%.

[0078] Table X presents the experimental results of different upsampling modules based on YOLOv11n. Compared with three mainstream upsampling methods—Content-Aware Feature Reconstruction (CARAFE), Efficient Upconvolutional Block (EUCB), and Nearest Neighbor Interpolation (NNI)—the proposed method shows significant improvements on the HRSID ship dataset. Specifically, compared to the baseline model, accuracy is improved by 1%, recall by 0.5%, and mAP50 by 0.8%. The DySample module enables the model to focus attention on key regions through dynamic feature adaptation, thereby improving overall detection performance. Comparative analysis shows that the DySample module not only outperforms existing upsampling techniques but also demonstrates powerful feature extraction capabilities and excellent adaptability, especially in complex marine environments with scale variations and occlusion, effectively handling diverse ship identification scenarios.

[0079] Table X: Performance Comparison of Different Upsampling Modules

[0080] In Experiment 4, this invention investigated the effect of MSTFNet. Data shows that this module improved mAP50-95 by 3%, mAP50 by 1.9%, and reduced the number of parameters by 0.66M. This module employs a lightweight feature fusion technique to improve efficiency and reduce computational load.

[0081] Finally, ablation experiments were conducted on the HRSID dataset, combining all modules. As shown in Table XI, the effectiveness and impact of PCA, PSA-G, DySample, and MSTFNet were evaluated using YOLOv11n as the baseline model. The contribution of each module is marked with "√". Compared to YOLOv11n, our model reduced the number of parameters by 34.7%, improved mAP50 by 3%, and improved mAP50-95 by 3.6%, while maintaining a moderate FPS. These ablation experiments highlight the importance of each module in the PPDM-YOLO framework, emphasizing their complementarity and effectiveness in improving the performance of YOLOv11n.

[0082] Table XI Model component ablation experiments on HRSID data

[0083] Example 2 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the SAR ship target detection method in a complex environment as described in Embodiment 1.

[0084] Example 3 This embodiment provides a computer-readable storage medium storing a computer program / instruction thereon, characterized in that the computer program / instruction, when executed by a processor, implements the SAR ship target detection method in a complex environment as described in Embodiment 1.

[0085] Example 4 This embodiment provides a computer program product, including a computer program / instruction, characterized in that, when the computer program / instruction is executed by a processor, it implements the SAR ship target detection method in a complex environment as described in Embodiment 1.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A SAR ship target detection method in complex environments, characterized in that, SAR ship detection model is used to identify SAR ship targets; The SAR ship detection model adopts the YOLOv11n backbone, including the backbone network, neck network and head network; The backbone network includes five convolutional modules, four lightweight feature extraction modules, an SPPF module, and a noise reduction and enhancement module; The first convolution module, the second convolution module, the first lightweight feature extraction module, the third convolution module, the second lightweight feature extraction module, the fourth convolution module, the third lightweight feature extraction module, the fifth convolution module, the fourth lightweight feature extraction module, the SPPF module, and the noise reduction enhancement module are connected in sequence. The first lightweight feature extraction module outputs P3 layer features, the second lightweight feature extraction module outputs P4 layer features, the third lightweight feature extraction module outputs P5 layer features, and the noise reduction and enhancement module outputs P6 layer features. The neck network includes five lightweight feature extraction modules, five connection modules, three dynamic sampling modules, and two convolutional modules; The first dynamic sampling module, the first connection module, the fifth lightweight feature extraction module, the second dynamic sampling module, the second connection module, the sixth lightweight feature extraction module, the third dynamic sampling module, and the third connection module are connected in sequence; The seventh lightweight feature extraction module, the sixth convolution module, the fourth connection module, the eighth lightweight feature extraction module, the seventh convolution module, the fifth connection module, and the ninth lightweight feature extraction module are connected in sequence; The output of the P3 layer features is input to the third connection module, and the output of the third connection module is input to the seventh lightweight feature extraction module. The output P4 layer features are input to the second connection module on one hand, and to the fourth connection module on the other hand. The output P5 layer features are input to the first connection module on one hand, and to the fifth connection module on the other hand. The output P6 layer features are input into the first dynamic sampling module; The output of the fifth lightweight feature extraction module is also input to the fifth connection module, and the output of the sixth lightweight feature extraction module is also input to the fourth connection module; The seventh lightweight feature extraction module outputs small target detection features; The target detection features are output from the eighth lightweight feature extraction module. The ninth lightweight feature extraction module outputs large target detection features; The head network includes three detection modules; The small target detection features are input into the first detection module for detection, and the first detection result is output. The target detection features are input into the second detection module for detection, and the second detection result is output. Large target detection features are input into the third detection module for detection, and the third detection result is output. The lightweight feature extraction module uses partial convolution combined with ECA channel attention mechanism to focus on the effective pixel area through dynamic mask update, thereby suppressing the influence of speckle noise. The noise reduction and enhancement module extracts gradient features through the multi-scale Sobel operator and uses a lightweight threshold generator to dynamically predict the spatial adaptive threshold, thereby achieving noise gradient truncation and suppression. The dynamic sampling module uses point sampling as its core idea to achieve dynamic upsampling. It predicts the sampling offset and generates a dynamic scope through a linear layer.

2. The SAR ship target detection method in complex environments according to claim 1, characterized in that: The lightweight feature extraction module includes a multi-layered alternating structure consisting of an ECA channel attention module and a PConv partial convolution module. The processing flow includes: For the input feature map, preliminary feature extraction is first performed, and then the extracted features are divided into two parts; One part of the features is first processed by halving the number of channels, and then the remaining spatial features after halving the number of channels are input into a multi-layer alternating structure, extracted by PConv partial convolution operation, and ECA attention mechanism is applied at the front end of PConv partial convolution. Finally, the features extracted from the multi-layer alternating structure are concatenated with another set of features along the channel dimension, and then fused and output through pointwise convolution.

3. The SAR ship target detection method in complex environments according to claim 1, characterized in that: The noise reduction enhancement module includes an AGT module, a first Split module, two DynamicPSA Block modules, a sixth connection module, a DWConv module, and a first CBS module; The tensors input to the noise reduction and enhancement module are first input to the AGT module; The output of the AGT module is connected to the first Split module. The output of the first Split module is processed by the first DynamicPSA module and the second DynamicPSA module connected in sequence, and then output by the second DynamicPSA module to the sixth connection module. The output of the first Split module is also directly input to the sixth connection module. The output of the sixth connection module is connected to the DWConv module and the first CBS module in sequence; The output of the first CBS module is used as the output of the noise reduction enhancement module; The AGT module includes the second Split module, the Multi-Scale Gradient Convolution module, the DynamicThreshold Generator module, the Soft Thresholding module, the Multi-Scale Fusion module, and the seventh connection module; Tensors from the AGT module are first input into the second Split module; The output of the second Split module is processed by the Multi-Scale Gradient Convolution module, Dynamic Threshold Generator module, Soft Thresholding module and Multi-Scale Fusion module connected in sequence, and then output by the Multi-Scale Fusion module to the seventh connection module. The output of the second Split module is also directly input to the seventh connection module. The output of the seventh connection module is used as the output of the AGT module; The DynamicPSA Block module includes a third Split module, a Multihead Attention module, two CBS modules, a scaling factor adjustment module, and an eighth connection module; Tensors from the DynamicPSA Block module are first input into the third Split module; The output of the third Split module is processed by the Multihead Attention module, the second CBS module, and the scaling factor adjustment module connected in sequence, and then output by the scaling factor adjustment module to the eighth connection module. The output of the third Split module is also directly input to the eighth connection module. The scaling factor adjustment module also feeds back the scaling factor to the Multihead Attention module to dynamically adjust the attention weight distribution. The output of the eighth connection module is input to the third CBS module, and the output of the third CBS module is used as the output of the DynamicPSA Block module.

4. The SAR ship target detection method in complex environments according to claim 1, characterized in that: The sampling point generation method of the dynamic sampling module is as follows: Input feature map; Generate a basic coordinate grid to generate the basic sampling position in the input space for each output feature; The input feature map is processed by a lightweight linear layer to predict and generate the original offset field G; Based on the gain parameter g of the generated original offset field G, as well as the target height sH and target width sW, calculate the static scope factor S; Add the scaled offset to the base coordinate grid; Bilinear interpolation is used to sample from the input feature map.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the SAR ship target detection method in complex environments as described in any one of claims 1 to 7.

6. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the SAR ship target detection method in a complex environment as described in any one of claims 1 to 7.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the SAR ship target detection method in a complex environment as described in any one of claims 1 to 7.

Images