Lightweight multi-scale SAR image target detection method based on improved YOLO architecture

Abstract

The invention belongs to the technical field of computer vision and remote sensing image processing, and discloses a lightweight multi-scale SAR image target detection method based on an improved YOLO architecture. Firstly, an SAR image data set is constructed; then, a lightweight target detection network model improved based on the YOLO11 architecture is established; according to the lightweight target detection network model, a C3K2-ScConv module is adopted to perform space-channel reconstruction feature enhancement so as to suppress speckle noise, GhostConv lightweight convolution is introduced to generate redundant feature mapping so as to reduce the calculation amount, and multi-scale feature fusion and global context modeling are completed through an SPPF-LSKA module. A high-resolution small target detection branch is additionally arranged to improve the small-size target detection precision; then training the network model; and finally, inputting a to-be-detected picture, and outputting a target positioning and classification result based on the detection head.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, computer vision and radar image processing technology, and relates to a lightweight multi-scale SAR image target detection method based on an improved YOLO architecture. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor with the unique advantage of operating in all weather conditions and around the clock, unaffected by adverse weather conditions such as clouds, rain, and snow. It has been widely used in fields such as marine monitoring, resource exploration, and disaster assessment. However, because SAR images are based on the principle of coherent electromagnetic wave scattering, their inherent speckle noise severely interferes with target edge and texture features, resulting in low target-background contrast. Furthermore, SAR images typically contain a large amount of redundant scattering information, and targets of interest (such as ships and vehicles) are often small in scale and sparsely distributed, making them difficult to identify against complex noisy backgrounds. These characteristics cause a significant performance degradation when traditional target detection methods designed for natural optical images are directly applied to SAR images.

[0003] The development of SAR target detection technology has mainly evolved from traditional methods to deep learning-based methods.

[0004] Traditional SAR target detection methods, such as Constant False Alarm Rate (CFAR) detection, mainly rely on manually designed features such as intensity and texture. These methods perform well in simple background scenes, but in complex real-world environments, the target's scattering mechanism is difficult to accurately characterize using statistical models, resulting in limited detection performance and high false alarm and false detection rates.

[0005] In recent years, deep learning-based target detection methods have become mainstream in SAR target detection due to their powerful feature self-learning and representation capabilities. Early research attempted to directly transfer general deep learning detection frameworks, such as the single-stage detector SSD, the two-stage detector Faster R-CNN, and their variants (such as Cascade R-CNN and Mask R-CNN), to SAR images. For example, some studies have combined attention mechanisms (such as CBAM and SE) to improve the model's feature extraction capability for SAR targets (Zha, M.; Qian, W.; Yang, W.; Xu, Y. Multifeature transformation and fusion-based ship detection with small targets and complex backgrounds).IEEE Geoscience and Remote Sensing Letters 2022, 19 While these methods have made some progress, they have failed to fundamentally address the imaging mechanisms of SAR images (such as the multiplicative model of speckle noise, the sparsity and small-scale characteristics of targets) through deep adaptation at the network structure level.

[0006] With the YOLO series models gaining attention for their good balance between speed and accuracy, lightweight improvements based on the YOLO architecture have become a research hotspot. Researchers have explored this from multiple perspectives: for example, optimizing multi-scale feature fusion by designing lightweight feature extraction networks and bidirectional feature pyramids (BiFPN) (Guo, Y.; Chen, S.; Zhan, R.; Wang, W.; Zhang, J. Sar ship detection based on yolov5 using cbam and bifpn. In Proceedings of the IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 2147–2150.); or introducing frequency domain attention modules to suppress sea clutter interference (Zhang, L.; Liu, Y.; Zhao, W.; Wang, X.; Li, G.; He, Y. Frequency-adaptive learning for SAR ship detection in clutter scenes). IEEE Transactions on Geoscience and Remote Sensing 2023, 61 (1–14.). In addition, to further reduce model complexity to meet the practical deployment needs of edge computing, some works have adopted network pruning, automated architecture search, and lightweight design principles such as inverse residual modules and bottleneck structures.

[0007] Despite the significant achievements of the aforementioned technologies, the field of SAR target detection still faces a series of interconnected and unresolved core challenges: Insufficient accuracy in feature noise suppression and enhancement: Most existing methods introduce general attention mechanisms or feature enhancement modules that are not specifically designed based on the spatial-channel coupling characteristics of speckle noise in SAR images. Noise and invalid background features are not effectively separated and suppressed, resulting in low signal-to-noise ratios in the generated feature maps, making it difficult to highlight true targets under low-contrast conditions. The contradiction between model lightweighting and accuracy preservation: Common lightweighting strategies (such as channel pruning and using basic lightweight modules) often reduce parameters and computational load, but inevitably lead to a loss of feature representation capabilities, especially unfavorable for the detection of small targets with sparse features, making "lossless" or "gain-type" lightweighting impossible. Inherent defects in multi-scale and small target detection: Standard multi-scale detection architectures (such as FPN) face severe challenges on SAR images. After multiple downsamplings, the weak feature information of small targets is almost lost; while the global context of large targets is limited by the local receptive field of traditional convolution, making it difficult to capture effectively. Existing improvements mainly focus on adjusting fusion strategies, lacking high-resolution feature preservation and dedicated enhancement mechanisms for small targets. The efficiency bottleneck of long-range dependency modeling: To improve the model's global understanding of complex scenes and large-scale targets, introducing modules such as Transformers and nonlocal networks to model long-range dependencies is an effective approach; however, its huge computational overhead conflicts with the requirements of real-time SAR processing and lightweight deployment.

[0008] In summary, existing technologies mostly focus on independent improvements in one or a few aspects, lacking a systematic solution that integrates and coordinates optimization from the feature generation source to the multi-scale prediction terminal. Therefore, designing a high-precision SAR target detection network that simultaneously achieves: 1) collaborative denoising and enhancement of target features in both spatial and channel dimensions; 2) efficient generation of rich features while reducing computational redundancy; 3) constructing efficient long-range context-aware capabilities to serve multi-scale detection; and 4) preserving and enhancing detailed information of small targets through dedicated pathways, is a critical technical problem urgently needing to be solved in this field. This invention is proposed based on this premise. Summary of the Invention

[0009] This invention aims to address the problems of severe speckle noise interference, insufficient multi-scale target feature representation, and easy loss of small targets in existing YOLO-like detection models for SAR image target detection. It proposes a multi-scale attention target detection method based on an improved YOLO architecture to achieve high-precision and low-complexity target detection in complex SAR scenes.

[0010] The technical solution of the present invention: A lightweight multi-scale SAR image target detection method based on an improved YOLO architecture includes the following steps: S1: acquire synthetic aperture radar (SAR) images, construct a sample dataset containing multiple types of targets, and perform data augmentation on the sample dataset to obtain training samples; Data augmentation processing includes at least one of random cropping, rotation, flipping, scaling, brightness adjustment, and contrast adjustment.

[0011] S2: Input the training samples into a lightweight object detection network model based on the improved YOLO11 architecture for feature extraction and fusion; the lightweight object detection network model based on the improved YOLO11 architecture includes a backbone network, a neck network, and a detection head; The input SAR image first undergoes multi-scale feature extraction through the backbone network, progressively extracting shallow, mid-level, and deep features from shallow to deep layers. Subsequently, the deep features are processed sequentially through a feature processing module consisting of GhostConv and C3K2-ScConv modules, an SPPF-LSKA module, and a C2PSA module before being fed into the neck network. Multi-scale feature fusion is performed through a bidirectional feature pyramid structure to obtain fused features. Finally, the fused features are input into four detection branches of the detection head, including three standard detection branches and a small target detection branch for tiny targets. The four detection branches calculate the spatial location and class probability in parallel, outputting the final target detection result. (1) Backbone network; The backbone network adopts a multi-scale feature extraction structure, including GhostConv module, C3K2 module, C3K2-ScConv module, SPPF-LSKA module and C2PSA module; The SAR image is input into the GhostConv module for processing, resulting in a preliminary feature map. This preliminary feature map then undergoes three consecutive feature extraction stages, progressively extracting features. Each feature extraction stage includes both the GhostConv module and the C3K2 feature processing module. The first feature extraction stage yields shallow features P3, which is then processed by the second stage to obtain mid-level features P4. The mid-level features P4 are then processed by the third stage to obtain deep features P5. Deep features P5 are processed by a feature processing module composed of the GhostConv module and the C3K2-ScConv module to obtain high-level semantic features. These high-level semantic features are then fed into the SPPF-LSKA module to generate multi-scale global context features. The SPPF-LSKA module integrates a large kernel separable attention (LSKA) mechanism on top of the SPPF architecture, thereby capturing broader regional connections. Finally, the multi-scale global context features undergo further polarization and spatial attention enhancement by the C2PSA module, forming an enhanced deep global feature map N that is more robust to interference and richer in information. Among them, the GhostConv module replaces the standard convolution in the YOLO11 architecture in the backbone network. The GhostConv module adopts a step-by-step generation strategy: first, it uses some standard convolutions to obtain basic feature maps, then it derives redundant additional feature maps through linear transformation, and finally it concatenates the basic feature maps and additional feature maps to obtain the complete output features. While significantly reducing the computational overhead, it effectively suppresses the background redundancy of SAR images and enhances the expression of target scattering features.

[0012] The C3K2-ScConv module replaces the standard convolutional layers of the C3K2 module in the YOLO11 architecture with ScConv units that have a spatial-channel joint reconstruction mechanism. It suppresses SAR speckle noise and enhances target saliency through spatial thinning and asymmetric channel convolution. The execution logic of the C3K2-ScConv module is as follows: First, the raw features entering the C3K2-ScConv module undergo preliminary channel refinement and splitting to generate a feature map X input to the core processing branch. The core processing branch includes a Spatial Reconstruction Unit (SRU) and a Channel Reconstruction Unit (CRU). Subsequently, the Spatial Reconstruction Unit performs group normalization on the feature map X and uses the learnable scaling parameter γ after group normalization to generate channel weights for each channel. The channel weights are then used to generate an information weight map via the Sigmoid function. Non-information weight graph The input feature map X is compared with the information weight map respectively. Non-information weight graph Perform element-wise multiplication to obtain the weighted feature map. With weighted feature map Weighted feature map With weighted feature map After splitting into two sub-feature maps, they are cross-added and reconstructed to generate an intermediate feature map. With intermediate feature map ; intermediate feature map With intermediate feature map By fusion, a spatially refined feature map is obtained. Among them, the original features are the features generated by GhostConv from the deep feature P5; Next, the Channel Reconstruction Unit (CRU) refines the input spatial feature map. Asymmetric partitioning is performed along the channel dimension at a ratio of α, where α is 0.25, to obtain the upper feature. With lower features ; respectively for the upper features With lower features Perform 1x1 convolution to compress the channels; for the upper features Abstract features are extracted using group-weighted convolution combined with pointwise convolution. ; for lower features Extracting detailed features using pointwise convolution. Then, for abstract features and detailed features Perform global average pooling separately to obtain , ; respectively using weights and right and The weighted summation is performed to obtain the final output channel enhanced feature map. ; SPPF-LSKA module: Replace the SPPF module in the YOLO11 architecture with the SPPF-LSKA module; the SPPF-LSKA module consists of a 1x1 convolutional layer, three serial max pooling layers of different sizes, a large kernel separable attention layer, and a 1x1 convolutional layer. The SPPF-LSKA module performs 1 step on the input high-level semantic features. 1. Convolutional dimensionality reduction yields an initial feature stream; then, three sequential max-pooling layers of different sizes are used to obtain the first, second, and third pooling features, respectively. The initial feature stream, the first pooling feature, the second pooling feature, and the third pooling feature are concatenated to generate a pooling feature volume that integrates multi-scale receptive field information. A large kernel separable attention (LSKA) mechanism is used to process the pooling feature volume. The processing includes: using horizontal and vertical depth decomposition of the large kernel convolutional layer to calculate the attention weight map in the spatial dimension of the pooling feature volume, and multiplying it element-wise with the pooling feature volume to establish long-range dependencies between pixels and capture global contextual information; finally, 1. 1. The convolutional layer performs channel recovery and fusion output on the enhanced features; By integrating multi-scale pooling operations into the Large Kernel Separable Attention (LSKA) mechanism, the receptive field is significantly expanded to capture global contextual information, while effectively establishing long-range dependencies between pixels, thereby enhancing the model's ability to discriminate multi-scale targets in complex backgrounds.

[0013] (2) Neck network; The neck network adopts a bidirectional feature pyramid structure, and on the basis of the YOLO11 architecture, the last C3K2 module is replaced with a C3K2-ScConv module; Top-down path: The enhanced deep global feature map N is upsampled and spliced ​​with the deep feature P5 produced by the backbone network to obtain feature map n1; feature map n1 is refined by the C3K2 module and upsampled again and spliced ​​with the middle layer feature P4 to obtain feature map n2; feature map n2 is processed by the C3K2 module and upsampled for the third time and spliced ​​with the shallow layer feature P3 to obtain feature map n3; Bottom-up path: Feature map n3 is processed by the C3K2 module and output to the newly added small target detection branch. At the same time, it is concatenated with feature map n2 after convolution downsampling to obtain feature map m1. Feature map m1 is output by the C3K2 module to the first standard detection branch, and continues to be concatenated with feature map n1 after convolution to obtain feature map m2. Feature map m2 is output by the C3K2 module to the second standard detection branch. Finally, it is concatenated with the enhanced deep global feature map N output by the C2PSA module in the backbone network after convolution, and further suppressed by the C3K2-ScConv module before being output to the third standard detection branch.

[0014] (3) Detection head; The detection head consists of the three original standard detection branches of YOLO11 and a newly added small target detection branch. The feature map n3 from the neck network, after local feature refinement by the C3K2 module, is directly input into the small target detection branch. This branch effectively preserves the high-frequency geometric information and spatial details of small targets in the SAR image. Finally, the four detection branches (the small target detection branch and the three standard detection branches) achieve accurate classification and localization of targets of different sizes on the SAR image through parallel prediction and multi-scale feature convergence.

[0015] S3: Based on the multi-branch prediction results output by the standard detection branch and the additional small target detection branch, targets in SAR images are located, classified, and their confidence is evaluated, and the final target detection results are output.

[0016] The training process of the network model includes: inputting the data-augmented training samples into a lightweight object detection network model based on the improved YOLO11 architecture, and obtaining the prediction results through forward propagation; calculating the complete intersection-union ratio loss between the predicted bounding box and the ground truth bounding box, and the classification loss between the predicted class and the ground truth class; updating the network parameters using the backpropagation algorithm to minimize the total loss function; repeating the above process until the lightweight object detection network model based on the improved YOLO11 architecture converges; during the training process, a learning rate decay strategy and an early stopping mechanism are used to prevent overfitting, and average accuracy is used as the performance evaluation index for object detection.

[0017] The beneficial effects of this invention are as follows: By introducing the C3K2-ScConv module, this invention reconstructs features in both spatial and channel dimensions, effectively suppressing speckle noise in SAR images and improving the contrast between targets and backgrounds; by employing the Ghost convolutional structure to reduce redundant feature calculations, it significantly reduces the number of model parameters and computational complexity while maintaining detection accuracy; by using the SPPF-LSKA module to achieve multi-scale feature fusion and long-distance dependency modeling, it enhances the model's ability to detect targets at different scales; by adding a small target detection layer, it alleviates the problem of small target feature loss caused by downsampling in deep networks, improving the detection accuracy of small targets in complex scenes; and by ensuring model lightweighting and inference efficiency, it achieves high-precision detection of targets in complex SAR scenes. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the principle of the present invention.

[0019] Figure 2 This is a diagram of the backbone network structure of the improved YOLO11 of this invention.

[0020] Figure 3 This is a schematic diagram of the C3K2-ScConv module in this invention.

[0021] Figure 4 This is a schematic diagram illustrating the working principle of the GhostConv module in this invention.

[0022] Figure 5 This is a schematic diagram of the SPPF-LSKA module in this invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0024] Example The model was improved based on YOLO11n and trained and tested on the publicly available large-scale SAR target detection dataset SARDet-100k.

[0025] Step S1: Data preparation and augmentation; SAR images were acquired from the SARDet-100k dataset, containing various targets such as ships, vehicles, and aircraft. Data augmentation strategies were employed, including random horizontal flipping, random rotation (±10 degrees), and random brightness and contrast adjustments. Mosaic augmentation was also introduced to increase data diversity and improve the model's generalization ability.

[0026] Step S2: Backbone network feature extraction and C3K2-ScConv module; The preprocessed SAR image is input into a lightweight target detection network model based on an improved YOLO11 architecture. This is achieved by replacing the standard convolutional layers of the C3K2 module in the YOLO11 architecture with ScConv units that have a spatial-channel joint reconstruction mechanism.

[0027] like Figure 2 As shown, the core of the C3K2-ScConv module is the ScConv unit, and its specific workflow is as follows: The input feature map X first enters the spatial reconstruction unit. The spatial reconstruction unit performs group normalization (GN) on the feature map X and calculates the channel weights for each channel using the learnable scaling parameter γ of GN. After activation by the Sigmoid function, a threshold (e.g., 0.5) is set to adjust the weights. Figure 2 Value-based extraction, separating the information weight graph Non-information weight graph Feature map X and information weight map Non-information weight graph Element-wise multiplication yields a weighted feature map. With weighted feature map Subsequently, the weighted feature map... With weighted feature map After splitting into two sub-feature maps, they are cross-added and reconstructed to generate an intermediate feature map. With intermediate feature map ; intermediate feature map With intermediate feature map By fusion, a spatially refined feature map is obtained. .

[0028] Next, the channel reconstruction unit refines the input spatial feature map. Asymmetric partitioning is performed along the channel dimension at a ratio of α, where α is 0.25, to obtain the upper feature. With lower features ; respectively for the upper features With lower features Perform 1x1 convolution to compress the channels; for the upper features Abstract features are extracted using group-weighted convolution (GWC) combined with pointwise convolution (PWC). ; for lower features Extracting detailed features using pointwise convolution. Then, for abstract features and detailed features Perform global average pooling (GAP) separately to obtain , ; respectively using weights and right and The weighted summation is performed to obtain the final output channel enhanced feature map. ; Step S3: Lightweight convolutional GhostConv module; like Figure 3 As shown, the GhostConv module is used in the backbone network to replace the standard convolutions in the YOLO11 architecture; the GhostConv module adopts a step-by-step generation strategy: first, it uses a portion of the standard convolutions to obtain the basic feature maps. Then, redundant additional feature maps f( ) are derived through linear transformation. Finally, the basic feature map and the additional feature map are concatenated to obtain the complete feature map. ; Step S4: Network feature fusion and SPPF-LSKA module; like Figure 4As shown, the SPPF-LSKA module replaces the SPPF module in the YOLO11 architecture with the SPPF-LSKA module. The SPPF-LSKA module consists of a 1x1 convolutional layer, three serial max pooling layers of different sizes, a large kernel separable attention layer, and a 1x1 convolutional layer. The SPPF-LSKA module performs 1 step on the input high-level semantic features. 1. Convolutional dimensionality reduction yields an initial feature flow; then, three sequential max-pooling layers of different sizes are used to obtain the first, second, and third pooling features, respectively. The initial feature flow, the first pooling feature, the second pooling feature, and the third pooling feature are concatenated to generate a pooling feature volume that integrates multi-scale receptive field information. A large-kernel separable attention mechanism is used to process the pooling feature volume. The processing includes: using horizontal and vertical depth decomposition of the large-kernel convolution to calculate the spatial dimension attention weight map of the pooling feature volume, and multiplying it element-wise with the pooling feature volume to establish long-range dependencies between pixels and capture global contextual information; finally, 1. 1. The convolutional layer performs channel recovery and fusion output on the enhanced features.

[0029] Step S5: Small Object Detection Branch (SODL) like Figure 5 As shown, the neck network adopts a bidirectional feature pyramid structure. Based on the YOLO11 architecture, the last C3K2 module is replaced with a C3K2-ScConv module. Top-down path: The enhanced deep global feature map N is upsampled and concatenated with the deep feature P5 produced by the backbone network to obtain feature map n1; feature map n1 is refined by the C3K2 module and upsampled again and concatenated with the middle layer feature P4 to obtain feature map n2; feature map n2 is processed by the C3K2 module and upsampled a third time and concatenated with the shallow layer feature P3 to obtain feature map n3. Bottom-up path: Feature map n3 is processed by the C3K2 module and output to the newly added small target detection branch. At the same time, it is concatenated with feature map n2 after convolution downsampling to obtain feature map m1. Feature map m1 is output by the C3K2 module to the first standard detection branch, and continues to be concatenated with feature map n1 after convolution to obtain feature map m2. Feature map m2 is output by the C3K2 module to the second standard detection branch. Finally, it is concatenated with the enhanced deep global feature map N output by the C2PSA module in the backbone network after convolution, and further suppressed by the C3K2-ScConv module before being output to the third standard detection branch.

[0030] Step S6: Target Prediction and Output Based on the multi-branch prediction results output by the four detection branches of the detection head, the target in the SAR image is located, classified and confidence evaluated, and the final target detection result is output.

[0031] Experimental verification Experimental results on the SARDet-100k test set show that, compared with the baseline model YOLO11n, the proposed method (LSG-YOLO) improves mAP@0.5 from 88.6% to 91.12% and mAP@[0.5:0.95] from 60.0% to 61.4%, while maintaining the number of parameters at 2.58M. The computational cost has increased slightly but remains within a controllable range. Furthermore, it demonstrates comprehensive improvements in detection accuracy for small targets (AP_S), medium targets (AP_M), and large targets (AP_L), proving the effectiveness of the proposed method.

[0032] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A lightweight multi-scale SAR image target detection method based on an improved YOLO architecture, characterized in that, Includes the following steps: S1: Acquire SAR images, construct a sample dataset containing multiple target categories, and perform data augmentation on the sample dataset to obtain training samples; S2: Input the training samples into a lightweight object detection network model based on the improved YOLO11 architecture for feature extraction and fusion; the lightweight object detection network model based on the improved YOLO11 architecture includes a backbone network, a neck network, and a detection head; The input SAR image first undergoes multi-scale feature extraction through the backbone network, progressively extracting shallow, mid-level, and deep features from shallow to deep layers. Subsequently, the deep features are processed sequentially through a feature processing module consisting of GhostConv and C3K2-ScConv modules, an SPPF-LSKA module, and a C2PSA module before being fed into the neck network. Multi-scale feature fusion is performed through a bidirectional feature pyramid structure to obtain fused features. Finally, the fused features are input into four detection branches of the detection head, including three standard detection branches and a small target detection branch for tiny targets. The four detection branches calculate the spatial location and class probability in parallel, outputting the final target detection result. S3: Based on the multi-branch prediction results output by the four detection branches of the detection head, the target in the SAR image is located, classified and its confidence is evaluated, and the final target detection result is output.

2. The lightweight multi-scale SAR image target detection method based on the improved YOLO architecture according to claim 1, characterized in that, The backbone network adopts a multi-scale feature extraction structure, including GhostConv module, C3K2 module, C3K2-ScConv module, SPPF-LSKA module and C2PSA module; The SAR image is input into the GhostConv module for processing, resulting in a preliminary feature map. This preliminary feature map then undergoes three consecutive feature extraction stages, progressively extracting features. Each feature extraction stage includes both the GhostConv module and the C3K2 feature processing module. The first feature extraction stage yields shallow features P3, which is then processed by the second stage to obtain mid-level features P4. The mid-level features P4 are then processed by the third stage to obtain deep features P5. Deep features P5 are processed by a feature processing module composed of the GhostConv module and the C3K2-ScConv module to obtain high-level semantic features. These high-level semantic features are then fed into the SPPF-LSKA module to generate multi-scale global context features. The SPPF-LSKA module integrates a large-kernel separable attention mechanism based on the SPPF architecture. Finally, the multi-scale global context features undergo further polarization and spatial attention enhancement by the C2PSA module, forming an enhanced deep global feature map N with stronger anti-interference capabilities and richer information. GhostConv module: The GhostConv module is used in the backbone network to replace the standard convolution in the YOLO11 architecture. The GhostConv module adopts a step-by-step generation strategy: first, it uses some standard convolutions to obtain basic feature maps, then it derives redundant additional feature maps through linear transformation, and finally the basic feature maps and additional feature maps are concatenated to obtain the complete output features. The C3K2-ScConv module replaces the standard convolutional layers of the C3K2 module in the YOLO11 architecture with ScConv units that have a spatial-channel joint reconstruction mechanism. The execution logic of the C3K2-ScConv module is as follows: First, the raw features entering the C3K2-ScConv module undergo preliminary channel refinement and splitting to generate a feature map X input to the core processing branch. The core processing branch includes spatial reconstruction units and channel reconstruction units. Subsequently, the spatial reconstruction unit performs group normalization on the feature map X and uses the learnable scaling parameter γ after group normalization to generate channel weights for each channel. The channel weights are then used to generate an information weight map using the Sigmoid function. Non-information weight graph The input feature map X is compared with the information weight map respectively. Non-information weight graph Perform element-wise multiplication to obtain the weighted feature map. With weighted feature map Weighted feature map With weighted feature map After splitting into two sub-feature maps, they are cross-added and reconstructed to generate an intermediate feature map. With intermediate feature map ; intermediate feature map With intermediate feature map By fusion, a spatially refined feature map is obtained. The original features are generated by GhostConv from the deep features P5; then, the channel reconstruction unit refines the input spatial feature map. Asymmetric partitioning is performed along the channel dimension at a ratio of α, where α is 0.25, to obtain the upper feature. With lower features ; respectively for the upper features With lower features Perform 1x1 convolution to compress the channels; for the upper features Abstract features are extracted using group-weighted convolution combined with pointwise convolution. ; for lower features Extracting detailed features using pointwise convolution. Then, for abstract features and detailed features Perform global average pooling separately to obtain , ; respectively using weights and right and The weighted summation is performed to obtain the final output channel enhanced feature map. ; SPPF-LSKA module: Replace the SPPF module in the YOLO11 architecture with the SPPF-LSKA module; the SPPF-LSKA module consists of a 1x1 convolutional layer, three serial max pooling layers of different sizes, a large kernel separable attention layer, and a 1x1 convolutional layer. The SPPF-LSKA module performs 1 step on the input high-level semantic features.

1. Convolutional dimensionality reduction yields an initial feature flow; then, three sequential max-pooling layers of different sizes are used to obtain the first, second, and third pooling features, respectively. The initial feature flow, the first pooling feature, the second pooling feature, and the third pooling feature are concatenated to generate a pooling feature volume that integrates multi-scale receptive field information. A large-kernel separable attention mechanism is used to process the pooling feature volume. The processing includes: using horizontal and vertical depth decomposition of the large-kernel convolution to calculate the spatial dimension attention weight map of the pooling feature volume, and multiplying it element-wise with the pooling feature volume to establish long-range dependencies between pixels and capture global contextual information; finally, 1.

1. The convolutional layer performs channel recovery and fusion output on the enhanced features.

3. The lightweight multi-scale SAR image target detection method based on the improved YOLO architecture according to claim 2, characterized in that, The neck network adopts a bidirectional feature pyramid structure, and on the basis of the YOLO11 architecture, the last C3K2 module is replaced with a C3K2-ScConv module; Top-down path: The enhanced deep global feature map N is upsampled and concatenated with the deep feature P5 produced by the backbone network to obtain feature map n1; feature map n1 is refined by the C3K2 module and upsampled again and concatenated with the middle layer feature P4 to obtain feature map n2; feature map n2 is processed by the C3K2 module and upsampled a third time and concatenated with the shallow layer feature P3 to obtain feature map n3. Bottom-up path: Feature map n3 is processed by the C3K2 module and output to the newly added small target detection branch. At the same time, it is concatenated with feature map n2 after convolution downsampling to obtain feature map m1. Feature map m1 is output by the C3K2 module to the first standard detection branch, and continues to be concatenated with feature map n1 after convolution to obtain feature map m2. Feature map m2 is output by the C3K2 module to the second standard detection branch. Finally, it is concatenated with the enhanced deep global feature map N output by the C2PSA module in the backbone network after convolution, and further suppressed by the C3K2-ScConv module before being output to the third standard detection branch.

4. The lightweight multi-scale SAR image target detection method based on the improved YOLO architecture according to claim 3, characterized in that, The detection head consists of the three original standard detection branches of YOLO11 and the newly added small target detection branch. The feature map n3 in the neck network is directly input into the small target detection branch after local feature refinement by the C3K2 module. Finally, the four detection branches achieve accurate classification and localization of targets of different sizes on SAR images through parallel prediction and multi-scale feature convergence.

5. The lightweight multi-scale SAR image target detection method based on the improved YOLO architecture according to claim 4, characterized in that, The training process of the lightweight object detection network model based on the improved YOLO11 architecture includes: inputting the data-augmented training samples into the lightweight object detection network model based on the improved YOLO11 architecture, and obtaining the prediction results through forward propagation; calculating the perfect intersection-union ratio loss between the predicted bounding box and the ground truth bounding box, and the classification loss between the predicted class and the ground truth class; updating the parameters of the lightweight object detection network model based on the improved YOLO11 architecture using the backpropagation algorithm to minimize the total loss function; repeating the above process until the lightweight object detection network model based on the improved YOLO11 architecture converges; during the training process, a learning rate decay strategy and an early stopping mechanism are used to prevent overfitting, and average accuracy is used as the object detection performance evaluation index.

6. The lightweight multi-scale SAR image target detection method based on the improved YOLO architecture according to claim 5, characterized in that, Data augmentation processing includes at least one of random cropping, rotation, flipping, scaling, brightness adjustment, and contrast adjustment.

Images