A method and system for detecting weak and small targets in a SAR image

Abstract

The present application relates to the technical field of remote sensing image processing and deep learning, and in particular to a kind of SAR image weak target detection method and system, for SAR image speckle noise is strong, geometric distortion is big, weak small target contrast is low, easy to be background clutter submerged etc., the present application adopts main network + feature enhancement module + multistage detection head structure composition.Main network uses ResNet50, and it is extended to the Cascade RCNN structure for SAR rotating frame.The present application realizes high-precision, low false alarm detection of SAR image weak small target through the cooperative work of the above module.Each module function is clear, connection relationship is clear, data flow processing step is linked, and it contributes to the achievement of final technical effect, and can be widely applied to sea surface monitoring, disaster assessment, military reconnaissance and other all-weather remote sensing tasks.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and deep learning, and more specifically, to a method and system for detecting weak targets in SAR images. Background Technology

[0002] SAR imagery, short for Synthetic Aperture Radar imagery, is a remote sensing image generated by a radar system through transmitting microwave signals, receiving echoes reflected from ground objects, and processing them using synthetic aperture technology. Its core advantage is that it is unaffected by weather or lighting conditions, enabling all-day, all-weather acquisition of ground object information, making it crucial for applications in maritime surveillance and disaster assessment. However, the unique imaging mechanism of SAR images results in significant issues such as speckle noise, geometric distortion, and low signal-to-noise ratio. In particular, small targets in the images (such as vehicles and small vessels) are often extremely small and have low contrast, making them easily obscured by complex background clutter.

[0003] Currently, traditional detection algorithms based on convolutional neural networks (CNN) or factorial network (FPN) structures struggle to extract key features. SAR detection methods based on rotated bounding boxes often directly employ backbone networks from the natural image domain (such as ResNet50). While these methods possess some recognition capabilities, they still suffer from the following shortcomings: (1) Lack of feature enhancement modules for SAR directional texture structures; (2) Lack of modeling for higher-order attention distribution, weak target regions are easily ignored; (3) The lack of feature modeling training with negative samples (pure background images) makes it easy for false alarms to occur; (4) The anchor setting is simple and difficult to adapt to the multi-scale target distribution in different scenes in SAR images.

[0004] Therefore, existing technologies have failed to fundamentally solve the core contradiction of weak target features and strong background interference in SAR. There is an urgent need for a new detection method that can specifically enhance the discriminative texture of targets and actively suppress complex background interference in order to improve the detection accuracy in real complex SAR scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting weak targets in SAR images. This invention aims to solve the core problem that existing technologies have failed to fundamentally address the issue of weak target features and strong background interference in SAR. It provides a new detection method that can specifically enhance the discriminative texture of targets and actively suppress complex background interference, thereby improving the detection accuracy in real and complex SAR scenarios.

[0006] To achieve the above objectives, the embodiments of this application provide the following technical solutions: On the one hand, embodiments of this application provide a method for detecting weak targets in SAR images, the method comprising the following steps: Step S1: Obtain the SAR image data to be detected and input the SAR image data to be detected into the target detection network model. The target detection network model is a Cascade R-CNN network based on ResNet50 as the backbone network and extended to rotate boxes. Step S2: Process and extract features from the SAR image data to be detected input into the target detection network model to generate initial feature maps for multiple stages; Step S3: The initial feature maps output from the second, third, and fourth stages of the backbone network are respectively input into the SFSConv spatial-frequency separable convolution module for processing. The SFSConv spatial-frequency separable convolution module is used to enhance the directional texture features of the SAR image and suppress speckle noise, and output the corresponding multi-level enhanced feature maps. Step S4: Input the multi-level enhanced feature map obtained in step S3 into the CBAM attention module. The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module. The CBAM attention module is used to adaptively adjust the weights of the obtained multi-level enhanced feature map through the channel attention sub-module and the spatial attention sub-module in sequence to improve the feature response of the weak target region and output the attention-enhanced feature map. Step S5: Input the attention-enhanced feature map output in step S4 into the region proposal network, set a multi-scale anchor box mechanism for the region proposal network, and generate initial target candidate box data containing location information based on the multi-scale anchor box mechanism. The region proposal network is a sub-module in the Cascade R-CNN network, and the multi-scale anchor box mechanism adopts three scales (4, 6, 8). Step S6: The initial target candidate boxes generated in step S5 and the attention-enhanced feature map output in step S4 are input together into the detection head for fine classification and regression. The detection head is used to independently classify and suppress candidate boxes labeled as strong background regions, and outputs the final target rotation box data after background suppression. The final target rotation box data includes the target category confidence and position coordinates.

[0007] Optionally, in step S3, the processing procedure of the SFSConv spatial-frequency separable convolution module includes: The wavelet filtering branch and the convolution branch are processed in parallel. The wavelet filtering branch is used to perform a two-dimensional discrete wavelet transform on the output initial feature map to extract detailed sub-band features in the horizontal, vertical and diagonal directions to capture typical edge and texture structures in the SAR image. The convolution branch is used to output spatial features. The detailed sub-band features output by the wavelet filtering branch and the spatial features output by the convolution branch are weighted and fused to generate the enhanced feature map.

[0008] Optionally, in step S4, the CBAM attention module is placed only after the fourth stage of the backbone network.

[0009] Optionally, in step S6, the detection head is a decision module of the target detection neural network, the detection head includes a strong negative sample detection head, and the strong negative sample detection head is configured during the model training phase as follows: The system receives training data labeled with strong negative samples, which correspond to strong background regions in SAR images. In the loss function calculation, independent weights are assigned to the strong negative samples so that the model can learn the distribution characteristics of the background regions.

[0010] Optionally, the region proposal network described in step S5 is cascaded with the detection head described in step S6 to perform multi-level progressive regression on the initial target candidate box data, so as to achieve accurate positioning of the rotating box of targets with different sizes and aspect ratios.

[0011] Optionally, the weak target detected by the method refers to a target with a pixel area of ​​less than 32×32 pixels in the SAR image data.

[0012] On the other hand, embodiments of this application provide a system for detecting weak targets in SAR images, used to implement the method for detecting weak targets in SAR images described above. The system includes: A data input module, which is used to acquire and transmit SAR image data to be detected; The feature extraction and enhancement module is connected to the data input module. The feature extraction and enhancement module includes a backbone network unit and multiple SFSConv processing units, which are used to perform multi-level feature extraction and orientation texture enhancement on the SAR image data and output multi-level enhanced feature maps. An attention control module is connected to the feature extraction and enhancement module. The attention control module is used to adjust the attention weights of the received fourth-stage enhanced feature map in terms of channel and spatial dimensions, and output an attention-enhanced feature map. A candidate box generation module is connected to the attention control module. The candidate box generation module is used to generate initial target candidate box data based on the attention enhancement feature map and the multi-scale anchor box mechanism. The target identification and output module is connected to the attention control module and the candidate box generation module respectively. The target identification and output module includes a strong negative sample suppression unit, which is used to classify, regress and suppress the background of the initial target candidate box data, and output the final target rotated box data. The output of the previous module is connected to the input of the next module in sequence to form an end-to-end processing pipeline.

[0013] Optionally, the SFSConv processing unit includes a wavelet filter and a convolutional layer in parallel, as well as a feature fusion unit for weighted fusion of the outputs of the two.

[0014] Optionally, in training mode, the strong negative sample suppression unit is configured to optimize the discrimination boundary of the background category based on the input strong negative sample label data.

[0015] Optionally, the candidate box generation module has three preset anchor box sizes: 4, 6, and 8, which are adapted to small and medium-sized weak targets, respectively.

[0016] The beneficial effects of this invention are as follows: This invention enhances the directional texture features of SAR images by introducing the SFSConv spatial-frequency separable convolution module at multiple stages of the ResNet50 backbone network, effectively improving the separability of weak targets from the background and suppressing speckle noise. Furthermore, a CBAM attention module is embedded in the fourth stage, adaptively enhancing the target region response through a dual channel and spatial attention mechanism, avoiding noise amplification. Simultaneously, a StrongNegative BBox Head is designed to independently model pure background regions, significantly reducing the false alarm rate. In addition, a multi-scale anchor mechanism is employed to cover targets of different sizes, improving detection adaptability. During the training phase, a large number of pure background images and various data augmentation strategies are introduced to further enhance the model's generalization ability. Experiments show that in SAR weak target detection tasks, this method improves mAP by 8%–30% compared to the baseline model, reduces the false alarm rate by approximately 40%, and significantly improves recall for small targets. It can be widely applied to all-weather remote sensing detection tasks such as sea surface surveillance, disaster assessment, and military reconnaissance.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall architecture of a method for detecting weak targets in SAR images as described in an embodiment of the present invention; Figure 2 This is a schematic diagram showing the introduction location of the SFSConv enhanced feature extraction module in an embodiment of the present invention; Figure 3 A schematic diagram of the insertion position of the CBAM attention module in an embodiment of the present invention; Figure 4 A schematic diagram of the StrongNegative BBox Head location in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] It should be noted that similar reference numerals or letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] Example 1: like Figure 1-4 As shown, this embodiment provides a method for detecting weak targets in SAR images. The detection method includes the following steps: Step S1: Obtain the SAR image data to be detected and input the SAR image data to be detected into the target detection network model. The target detection network model is a Cascade R-CNN network based on ResNet50 as the backbone network and extended to rotate boxes. Step S2: Process and extract features from the SAR image data to be detected input into the target detection network model to generate initial feature maps for multiple stages; Step S3: The initial feature maps output from the second, third, and fourth stages of the backbone network are respectively input into the SFSConv spatial-frequency separable convolution module for processing. The SFSConv spatial-frequency separable convolution module is used to enhance the directional texture features of the SAR image and suppress speckle noise, and output the corresponding multi-level enhanced feature maps. Step S4: Input the multi-level enhanced feature map obtained in step S3 into the CBAM attention module. The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module. The CBAM attention module is used to adaptively adjust the weights of the obtained multi-level enhanced feature map through the channel attention sub-module and the spatial attention sub-module in sequence to improve the feature response of the weak target region and output the attention-enhanced feature map. Step S5: Input the attention-enhanced feature map output in step S4 into the region proposal network, set a multi-scale anchor box mechanism for the region proposal network, and generate initial target candidate box data containing location information based on the multi-scale anchor box mechanism. The region proposal network is a sub-module in the Cascade R-CNN network, and the multi-scale anchor box mechanism adopts three scales (4, 6, 8). Step S6: The initial target candidate boxes generated in step S5 and the attention-enhanced feature map output in step S4 are input together into the detection head for fine classification and regression. The detection head is used to independently classify and suppress candidate boxes labeled as strong background regions, and outputs the final target rotation box data after background suppression. The final target rotation box data includes the target category confidence and position coordinates.

[0023] This invention presents a method for detecting weak targets in SAR images, employing a backbone network, a feature enhancement module, and a multi-level detection head structure. The backbone network uses ResNet50 and extends it into a Cascade RCNN structure oriented towards the SAR rotating bounding box. After inputting the original SAR image, preliminary feature extraction is first performed via the ResNet50 backbone network, resulting in four stages of feature maps (C2, C3, C4, C5). Subsequently, SFSConv modules are embedded after the C2, C3, and C4 feature layers for directional texture enhancement. The enhanced C4 feature is then fed into the CBAM module for attention focusing. The processed features... Figure 1 On the one hand, the data is fed into the RPN network to generate multi-scale candidate boxes. On the other hand, these candidate boxes are fed into the cascaded detection head (including the Strong-Negative Head) for final classification, regression and background suppression, and the rotated box results are output.

[0024] In step S3 of this embodiment, the processing procedure of the SFSConv spatial-frequency separable convolution module includes: The wavelet filtering branch and the convolution branch are processed in parallel. The wavelet filtering branch is used to perform a two-dimensional discrete wavelet transform on the output initial feature map to extract detailed sub-band features in the horizontal, vertical and diagonal directions to capture typical edge and texture structures in the SAR image. The convolution branch is used to output spatial features. The detailed sub-band features output by the wavelet filtering branch and the spatial features output by the convolution branch are weighted and fused to generate the enhanced feature map.

[0025] In the feature enhancement stage, the SFSConv spatial-frequency separable convolution module is introduced. SFSConv is a feature enhancement layer that combines frequency domain filtering and convolution. Its core idea is to use wavelet filtering to extract directional textures in SAR images, making weak targets more separable from the background. In this invention, Daubechies wavelet (db4) is used for two-dimensional discrete wavelet transform, decomposing the input feature map into two layers to extract detail subbands in the horizontal, vertical, and diagonal directions, respectively, to capture typical edge and texture structures in SAR images. The parameters of the wavelet filtering layer and the convolutional layer share training, i.e., they are jointly optimized during backpropagation, thereby achieving adaptive fusion of frequency domain features and spatial features. Specifically, the SFSConv module consists of a wavelet filtering branch and a convolutional branch running in parallel. The output features of the two branches are weighted and fused, with the weights controlled by learnable parameters to ensure that the model can autonomously balance the contributions of spectral and spatial information during training.

[0026] In this invention, SFSConv is inserted into the convolutional blocks of the 2nd, 3rd and 4th stages of ResNet, respectively. Compared with traditional techniques, the introduction of the SFSConv enhanced feature extraction module can enhance the spatial texture details around the target; suppress high-frequency noise and improve the contrast of weak targets; reduce feature aliasing and improve the performance of the rotated bounding box regression.

[0027] In step S4 of this embodiment, the CBAM attention module is placed only after the fourth stage of the backbone network.

[0028] The Convolutional Block Attention Module (CBAM) is a lightweight module that combines channel attention and spatial attention. In this invention, CBAM is only added after the convolutional layers of ResNet Stage 4. Global average pooling and global max pooling are performed on the input feature image to obtain two 1xC (C is the number of channels) description vectors. These two vectors are then fed into a shared multilayer perceptron (MLP), and the outputs are summed and passed through a sigmoid activation function to generate a channel attention weight vector Mc. Mc is then multiplied channel-wise with the input feature map to obtain the channel-weighted feature map Fc.

[0029] Spatial Attention: Average pooling and max pooling are performed along the channel dimension of Fc to obtain two HxW (height x width) feature maps. These two feature maps are concatenated along the channel dimension and passed through a 7x7 convolutional layer and a sigmoid activation function to generate a spatial attention weight matrix Ms. Ms is then multiplied positionally by Fc to obtain the final attention-enhanced feature map F. This design allows the network to adaptively emphasize feature channels related to weak targets (such as channels with specific texture responses) and spatial regions (such as the possible locations where the target may appear).

[0030] The reason why CBAM is only added to the fourth-stage convolutional layer of ResNet is that the features in the fourth stage are more semantically informative, and introducing attention makes it easier to improve object response; it avoids the noise amplification caused by adding CBAM to shallow layers; and it keeps the overall computational cost of the model from increasing significantly. The channel attention of CBAM can adaptively improve the response of the feature channels where weak targets are located, and the spatial attention further strengthens the salience of the target region, thereby significantly improving detection accuracy.

[0031] In step S6 of this embodiment, the detection head is a decision module of the target detection neural network, and the detection head includes a strong negative sample detection head, which is configured during the model training phase as follows: The system receives training data labeled with strong negative samples, which correspond to strong background regions in SAR images. In the loss function calculation, independent weights are assigned to the strong negative samples so that the model can learn the distribution characteristics of the background regions.

[0032] In this embodiment, the region proposal network described in step S5 is cascaded with the detection head described in step S6 to perform multi-level progressive regression on the initial target candidate box data, so as to achieve accurate positioning of the rotating box of targets with different sizes and aspect ratios.

[0033] Large background areas, characteristic of SAR images, often contain complex clutter (such as sea surface noise and ground object scattering). Traditional detection heads struggle to fully learn the distribution of "strong background regions," leading to false detections. To address this issue, this invention introduces a StrongNegative BBox Head, which serves as the final stage in a cascaded detection head. Its structure is similar to a conventional detection head, containing fully connected layers for classification and regression, but the output dimension of its classification branch is expanded to include a "strong background" category. Data preparation is crucial during the training phase. In addition to the standard positive samples (target) and negative samples (background) annotations, a batch of "strong background regions," such as sea surface spikes, building sidelobes, and strong echo areas in mountainous regions, needs to be additionally labeled. Independent category representations are established for background regions; `strong_negative_label=-1` is explicitly added; and the number of negative samples is increased during the sampling phase to allow the model to more fully learn the background distribution.

[0034] To effectively suppress false alarms in SAR images, this invention explicitly defines "strong background regions" and models them specifically. Strong background regions mainly refer to the following two types of image regions that are prone to false detection: (1) Low-response complex background region This refers to regions in SAR images with low scattering intensity but certain texture or structural variations, such as faint sea surface ripples, cloud shadows, and low-scattering vegetation areas. Although the overall response of these regions is lower than that of the true target, their local features may still be misidentified by the detection network as target edges or weak signals.

[0035] (2) High-response targetless region This refers to regions in SAR images that exhibit strong scattering, high contrast, or obvious texture, but do not actually contain the target of interest. These mainly include: A. Sea spikes and wave scattering; B. Multiple reflections caused by building shadows, corners, and roofs; C. Areas with strong echoes from natural features such as mountains and forests; D. Sidelobe artifacts and system noise points generated during SAR imaging.

[0036] During training, strong background regions are explicitly labeled with strong negative sample labels (strong_negative_label = -1) and learned independently via the StrongNegative BBox Head, enabling the network to effectively distinguish between real targets and background interference of "similar targets". Simultaneously, by incorporating the spatial and channel attention mechanisms of the CBAM module, response suppression is applied to these regions, further enhancing the model's robustness against complex backgrounds.

[0037] In this embodiment, the weak target detected by the detection method refers to a target with a pixel area of ​​less than 32×32 pixels in the SAR image data.

[0038] Traditional RPNs use only single-scale anchors, making it difficult to cover different target scale distributions in SAR images. This invention expands the anchors to three scales [4, 6, 8], thereby achieving simultaneous coverage of small and medium-sized targets. To improve the model's generalization ability, this invention introduces an enhancement strategy: multi-angle rotation enhancement (90°, 180°, 270°). Adaptive contrast enhancement (AutoContrast); random flipping in three directions (horizontal / vertical / diagonal); introducing pure background images as negative samples to improve the model's background modeling ability.

[0039] Data preparation: The training set contains positive sample images with labeled targets and negative sample images with pure background. "Strong background" annotations are applied to easily misdetected regions in the positive sample images.

[0040] Data augmentation: Randomly apply 90°, 180°, and 270° rotations, horizontal / vertical flips, and adaptive contrast enhancement to the input image to improve the model's generalization ability.

[0041] Training process: The prepared data is input into the network. The data flow passes through the following sequence: backbone network -> SFSConv module (C2, C3, C4) -> CBAM module (after C4) -> RPN generates candidate boxes -> Cascade R-CNN detection head (including StrongNegative BBox Head). The loss signal is backpropagated from the detection head, while simultaneously optimizing all parameters in the backbone network, SFSConv, CBAM, and the detection head. The multi-scale anchor boxes ([4, 6, 8]) of the RPN are mapped to specific pixel sizes at different levels of the feature map to match targets of different sizes.

[0042] The SAR image to be detected is scaled to a fixed size (e.g., 1024x1024) and then input into the trained model. The network performs forward propagation, and the StrongNegative BBoxHead outputs a set of suppressed candidate boxes. These boxes are then subjected to score thresholding and non-maximum suppression to obtain the final rotated bounding box detection results for weak targets.

[0043] This invention achieves high-precision, low-false-alarm detection of weak targets in SAR images through the coordinated operation of the aforementioned modules. Each module has a clearly defined function and a clear connection between them; the data flow processing steps are interconnected, collectively contributing to the final technical achievement.

[0044] Compared with the traditional SAR detection architecture, the present invention has the following advantages: (1) SFSConv enhances the texture representation of weak targets; (2) CBAM improves the attention of key areas; (3) StrongNegative Head significantly reduces the false alarm rate; (4) Multi-scale Anchor uniformly adapts to targets of different sizes; (5) The introduction of a large number of background samples makes the model more stable and reliable.

[0045] This invention enhances the directional texture features of SAR images by introducing the SFSConv spatial-frequency separable convolution module at multiple stages of the ResNet50 backbone network, effectively improving the separability of weak targets from the background and suppressing speckle noise. Furthermore, a CBAM attention module is embedded in the fourth stage, adaptively enhancing the target region response through a dual channel and spatial attention mechanism, avoiding noise amplification. Simultaneously, a StrongNegative BBox Head is designed to independently model pure background regions, significantly reducing the false alarm rate. In addition, a multi-scale anchor mechanism is employed to cover targets of different sizes, improving detection adaptability. During the training phase, a large number of pure background images and various data augmentation strategies are introduced to further enhance the model's generalization ability. Experiments show that in SAR weak target detection tasks, this method improves mAP by 8%–30% compared to the baseline model, reduces the false alarm rate by approximately 40%, and significantly improves recall for small targets. It can be widely applied to all-weather remote sensing detection tasks such as sea surface surveillance, disaster assessment, and military reconnaissance.

[0046] Example 2 This embodiment provides a SAR image weak target detection system based on Embodiment 1, used to implement the detection method described above. The system includes: A data input module, which is used to acquire and transmit SAR image data to be detected; The feature extraction and enhancement module is connected to the data input module. The feature extraction and enhancement module includes a backbone network unit and multiple SFSConv processing units, which are used to perform multi-level feature extraction and orientation texture enhancement on the SAR image data and output multi-level enhanced feature maps. An attention control module is connected to the feature extraction and enhancement module. The attention control module is used to adjust the attention weights of the received fourth-stage enhanced feature map in terms of channel and spatial dimensions, and output an attention-enhanced feature map. A candidate box generation module is connected to the attention control module. The candidate box generation module is used to generate initial target candidate box data based on the attention enhancement feature map and the multi-scale anchor box mechanism. The target identification and output module is connected to the attention control module and the candidate box generation module respectively. The target identification and output module includes a strong negative sample suppression unit, which is used to classify, regress and suppress the background of the initial target candidate box data, and output the final target rotated box data. The output of the previous module is connected to the input of the next module in sequence to form an end-to-end processing pipeline.

[0047] The SFSConv processing unit described in this embodiment includes a wavelet filter and a convolutional layer connected in parallel, as well as a feature fusion unit for weighted fusion of the outputs of the two.

[0048] In this embodiment, the strong negative sample suppression unit is configured in training mode to optimize the discrimination boundary of the background category based on the input strong negative sample label data.

[0049] In this embodiment, the candidate box generation module has three preset anchor box sizes: 4, 6, and 8, which are adapted to small and medium-sized weak targets, respectively.

[0050] The following section, by combining the differences between the two models, demonstrates the optimization of this invention from the baseline model to the improved model: (1) Baseline model: roi-trans-le90_r50_fpn_1x_rsar.py (2) Improved Model 1: roi-trans-le90_r50_fpn_1x_rsar_SFSConv.py (3) Improved Model 2: roi-trans-le90_r50_fpn_30e_rsar_cbam_anchor_stage4_with_negatives.py The two models improved the backbone and other parts of the baseline model from different directions, and both achieved improvements to varying degrees.

[0051] 1. Comparison with baseline model The baseline model uses a ResNet50 + FPN + Cascade RCNN architecture and does not include any enhancement mechanisms specifically for SAR images.

[0052] The present invention has made the following improvements: (1) Insert the SFSConv direction frequency domain enhancement module; (2) Add CBAM attention to Stage 4; (3) Reconstruct the Anchor scale; (4) Introduce strong-negative background modeling; (5) Add a large number of pure background training samples; (6) Adjust the sampler, loss function weights and training strategy.

[0053] 2. Experimental Results Compared to the baseline model, the method of this invention significantly improves the overall performance of SAR weak target detection tasks, especially in terms of small target recall and background suppression capability (FAR).

[0054] 3. Dataset Description (1) Data source: The experimental data were a fusion of the publicly available SSDD (SAR Ship Detection Dataset) and the SARWeak-12 dataset developed by our research group. SSDD provides typical surface ship targets, while SARWeak-12 supplements this with land vehicles and more complex background types.

[0055] (2) Data partitioning and scale: Training set: A total of 18,000 SAR images. Among them, 13,200 images contain targets and are labeled, and 4,800 images are plain background images.

[0056] Test set: A total of 5000 SAR images, all with fine annotations. The test set and training set have no overlap in scene, target type, and imaging conditions.

[0057] (3) Target category and size distribution: Target categories: Includes two types of small targets: ships and vehicles.

[0058] Size distribution: Over 90% of the target bounding box pixels are smaller than 32×32. Specifically, approximately 60% of the target area is concentrated between 10×10 and 25×25 pixels, and approximately 30% of the target area is smaller than 10×10 pixels, which meets the definition of a small target. The aspect ratio of the targets ranges from 1:1 to 1:5, covering different shapes of ships (slender) and vehicles (approximately rectangular).

[0059] (4) Background complexity: The dataset covers a variety of challenging scenarios to comprehensively evaluate model performance: High clutter sea surface: includes sea surface textures and sea spikes under different wind conditions.

[0060] Complex landforms: urban building complexes (producing strong reflections and overlaps), mountain shadows, and forest areas.

[0061] Uniform background: Low-scattering areas such as calm sea surface and flat farmland are used to test false alarm suppression capability.

[0062] Special challenge set: The test set specifically includes 2000 images with extremely high clutter, whose background response intensity is similar to that of the target, constituting strong interference.

[0063] (5) Image specifications and preprocessing: All images were uniformly processed to 1024×1024 pixels. During the training phase, in addition to the aforementioned rotation, flipping, and adaptive contrast enhancement, speckle noise simulation based on Gamma distribution and local block random brightness transformation were added to enhance the model's adaptability to SAR image degradation effects.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting weak targets in SAR images, characterized in that, The detection method includes the following steps: Step S1: Obtain the SAR image data to be detected and input the SAR image data to be detected into the target detection network model. The target detection network model is a Cascade R-CNN network based on ResNet50 as the backbone network and extended to rotate boxes. Step S2: Process and extract features from the SAR image data to be detected input into the target detection network model to generate initial feature maps for multiple stages; Step S3: The initial feature maps output from the second, third, and fourth stages of the backbone network are respectively input into the SFSConv spatial-frequency separable convolution module for processing. The SFSConv spatial-frequency separable convolution module is used to enhance the directional texture features of the SAR image and suppress speckle noise, and output the corresponding multi-level enhanced feature maps. Step S4: Input the multi-level enhanced feature map obtained in step S3 into the CBAM attention module. The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module. The CBAM attention module is used to adaptively adjust the weights of the obtained multi-level enhanced feature map through the channel attention sub-module and the spatial attention sub-module in sequence to improve the feature response of the weak target region and output the attention-enhanced feature map. Step S5: Input the attention-enhanced feature map output in step S4 into the region proposal network, set a multi-scale anchor box mechanism for the region proposal network, and generate initial target candidate box data containing location information based on the multi-scale anchor box mechanism. The region proposal network is a sub-module in the Cascade R-CNN network, and the multi-scale anchor box mechanism adopts three scales (4, 6, 8). Step S6: The initial target candidate boxes generated in step S5 and the attention-enhanced feature map output in step S4 are input together into the detection head for fine classification and regression. The detection head is used to independently classify and suppress candidate boxes labeled as strong background regions, and outputs the final target rotation box data after background suppression. The final target rotation box data includes the target category confidence and position coordinates.

2. The detection method according to claim 1, characterized in that, In step S3, the processing procedure of the SFSConv spatial-frequency separable convolution module includes: The wavelet filtering branch and the convolution branch are processed in parallel. The wavelet filtering branch is used to perform a two-dimensional discrete wavelet transform on the output initial feature map to extract detailed sub-band features in the horizontal, vertical and diagonal directions to capture typical edge and texture structures in the SAR image. The convolution branch is used to output spatial features. The detailed sub-band features output by the wavelet filtering branch and the spatial features output by the convolution branch are weighted and fused to generate the enhanced feature map.

3. The detection method according to claim 1, characterized in that, In step S4, the CBAM attention module is placed only after the fourth stage of the backbone network.

4. The detection method according to claim 1, characterized in that, In step S6, the detection head is a decision module of the target detection neural network, and the detection head includes a strong negative sample detection head, which is configured during the model training phase as follows: The system receives training data labeled with strong negative samples, which correspond to strong background regions in SAR images. In the loss function calculation, independent weights are assigned to the strong negative samples so that the model can learn the distribution characteristics of the background regions.

5. The detection method according to claim 1, characterized in that, The region proposal network described in step S5 is cascaded with the detection head described in step S6 to perform multi-level progressive regression on the initial target candidate box data, so as to achieve accurate positioning of the rotating box of targets with different sizes and aspect ratios.

6. The detection method according to any one of claims 1 to 5, characterized in that, The weak targets detected by the method refer to targets with a pixel area of ​​less than 32×32 pixels in the SAR image data.

7. A system for detecting weak targets in SAR images, used to implement the detection method according to any one of claims 1 to 6, characterized in that, The system includes: A data input module, which is used to acquire and transmit SAR image data to be detected; The feature extraction and enhancement module is connected to the data input module. The feature extraction and enhancement module includes a backbone network unit and multiple SFSConv processing units, which are used to perform multi-level feature extraction and orientation texture enhancement on the SAR image data and output multi-level enhanced feature maps. An attention control module is connected to the feature extraction and enhancement module. The attention control module is used to adjust the attention weights of the received fourth-stage enhanced feature map in terms of channel and spatial dimensions, and output an attention-enhanced feature map. A candidate box generation module is connected to the attention control module. The candidate box generation module is used to generate initial target candidate box data based on the attention enhancement feature map and the multi-scale anchor box mechanism. The target identification and output module is connected to the attention control module and the candidate box generation module respectively. The target identification and output module includes a strong negative sample suppression unit, which is used to classify, regress and suppress the background of the initial target candidate box data, and output the final target rotated box data. The output of the previous module is connected to the input of the next module in sequence to form an end-to-end processing pipeline.

8. The system according to claim 7, characterized in that, The SFSConv processing unit includes a wavelet filter and a convolutional layer in parallel, as well as a feature fusion unit for weighted fusion of the outputs of the two.

9. The system according to claim 7, characterized in that, In training mode, the strong negative sample suppression unit is configured to optimize the discrimination boundary of the background category based on the input strong negative sample label data.

10. The system according to claim 7, characterized in that, The candidate box generation module has three preset anchor box sizes: 4, 6, and 8, which are adapted to small and medium-sized weak targets, respectively.

Images