A bidirectional recursive refinement fusion network structure and method for SAR target detection

Abstract

The application discloses a bidirectional recursive refinement fusion network structure and method for SAR target detection, relates to the technical field of computer vision and remote sensing image intelligent interpretation, and comprises a main feature extraction module, a bidirectional recursive refinement fusion module and a detection decoding and detection head module which are sequentially connected, wherein the bidirectional recursive refinement fusion module is internally integrated with a dynamic sampling alignment module and a plurality of reparameterization residual enhancement modules; the main feature extraction module is used for performing hierarchical convolution feature extraction on a preprocessed SAR image, and outputs multi-scale feature maps F2, F3, F4 and F5 from a high-resolution layer to a low-resolution layer and from low semantic information to high semantic information, wherein F2 is a highest-resolution feature map; the application replaces a traditional fixed interpolation up-sampling mode, effectively reduces feature misplacement and edge blur problems, and significantly improves the positioning stability of small targets and slender targets; and the application has a wide application scenario and has significant engineering application value.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent interpretation of remote sensing images, and in particular to a bidirectional recursive refinement and fusion network structure and method for SAR target detection. Background Technology

[0002] As is well known, Synthetic Aperture Radar (SAR) imaging is characterized by its all-weather, all-day operation and strong penetration capabilities through clouds, rain, and fog, and is widely used in fields such as maritime surveillance, disaster emergency response, military reconnaissance, resource surveys, and urban security. With the continuous growth in the scale of high-resolution SAR data and application demands, automatic target detection for SAR images has become one of the important foundational capabilities for intelligent remote sensing interpretation. However, the SAR imaging mechanism differs significantly from optical imaging, leading to numerous technical challenges in target detection: coherent imaging results in widespread speckle noise in SAR images, and scattering from complex terrain features generates strong clutter backgrounds, causing low contrast, weak texture, and unstable edge structures in target areas, making target echoes easily confused with background scattering; simultaneously, in practical applications, SAR images exhibit a wide range of target scales and a high proportion of small targets, and the target morphology and scattering characteristics change significantly with imaging angle and mode, further increasing the detection difficulty. Therefore, how to achieve high-precision and robust detection of multi-scale targets, especially small targets, against strong noise and clutter backgrounds is a key technical problem that urgently needs to be solved in the field of SAR target detection.

[0003] Currently, SAR target detection commonly employs deep learning-based target detection frameworks, which typically consist of three parts: 1) Backbone network: used for hierarchical convolution of SAR images to extract multi-level features; 2) Neck network: used for multi-scale feature fusion, with typical schemes such as Feature Pyramid Network (FPN), Path Aggregation Network (PAN), and Weighted Bidirectional Feature Pyramid Network (BiFPN); 3) Head: used to output the target classification results and bounding box regression parameters. Among these, multi-scale feature fusion is the key to determining the performance of SAR target detection. Small target detection relies on high-resolution shallow features, while weak target detection in complex backgrounds requires high-level semantic feature enhancement and cross-layer detail feature compensation.

[0004] In existing technologies, multi-scale feature fusion schemes mainly fall into three categories: 1) Feature Pyramid Network (FPN): It adopts a top-down fusion scheme, which transmits high-level semantic features to low-level layers and fuses them with shallow features connected laterally. This improves the problem of insufficient semantic features in shallow layers, but has limited ability to reuse fine-grained edge and speckle texture details. 2) Path Aggregation Network (PAN): It adopts a bidirectional fusion scheme, combining top-down semantic transfer and bottom-up detail feedback, which enhances the transfer effect of low-level details to high-level details. However, this fusion is a one-time aggregation process and lacks a recursive refinement mechanism. It is not adaptable to problems such as semantic-detail mismatch, speckle interference, and drastic scale changes in SAR images. 3) Weighted Bidirectional Feature Pyramid Network (BiFPN): It adopts a weighted / attention fusion scheme to improve the fusion quality through learnable weights or attention mechanisms. However, in SAR detection applications, it is still prone to feature redundancy or conflict after fusion, information loss and alignment error caused by upsampling and downsampling, and stacking multiple fusion layers will significantly increase network parameters and computational overhead, while the improvement effect is very limited.

[0005] In summary, existing multi-scale feature fusion schemes have the following significant technical shortcomings in SAR target detection: 1) Insufficient cross-scale information transfer, making it difficult to simultaneously enhance high-level semantic features and preserve the fidelity of low-level detailed features in a single fusion; 2) Insufficient fusion refinement capability, lacking a recursive refinement correction mechanism, making it difficult to fully correct fusion errors, leading to problems such as missed detection of weak targets, clustering of dense targets, and false detection in complex speckle backgrounds; 3) Information loss and alignment errors in the upsampling / downsampling process, with traditional interpolation or stride convolution scale transformation methods easily causing loss of detailed information, which has a particularly significant impact on the detection of small and slender targets; 4) A contradiction exists between parameter overhead and detection performance, where stacking a large number of fusion layers to improve detection accuracy significantly increases network parameters and computational load, which is not conducive to engineering deployment and edge device applications.

[0006] To address the aforementioned technical issues, there is an urgent need to propose a bidirectional recursive refinement fusion network structure and method for SAR target detection. This method aims to improve the quality of multi-scale feature fusion, enhance the detection performance of weak / dense targets, and reduce the false detection rate in complex clutter backgrounds, all while controlling parameters and computational overhead. This has become a fundamental requirement for those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bidirectional recursive refinement fusion network structure and method for SAR target detection.

[0008] The technical solution adopted in this invention is: A bidirectional recursive refinement fusion network structure for SAR target detection includes a cascaded backbone feature extraction module, a bidirectional recursive refinement fusion module, and a detection decoding and detection head module. The bidirectional recursive refinement fusion module integrates a dynamic sampling alignment module and multiple reparameterized residual enhancement modules. The backbone feature extraction module is used to perform hierarchical convolution feature extraction on the preprocessed SAR image and output multi-scale feature maps F2, F3, F4, and F5 from high-resolution layer to low-resolution layer and from low semantic information to high semantic information, where F2 is the highest resolution feature map and F5 is the highest semantic feature map. The bidirectional recursive refinement fusion module BRRF is the core module of the network. It takes the multi-scale feature maps F2~F5 output by the backbone feature extraction module as input, constructs a top-down refinement propagation path and a bottom-up semantic backfeeding path, and outputs the fused multi-scale feature sets S3, S4 and S5. The dynamic sampling alignment module is set at the cross-scale upsampling point of the top-down refinement of the transmission path, and performs learnable alignment for the upsampling process from F5 to F4 and from F4 to F3. The reparameterized residual enhancement module is distributed across the fusion nodes of the bidirectional recursive refinement fusion module at each scale. It is used to perform channel alignment, residual enhancement, and noise suppression on features from different paths / scales. During the training phase, a multi-branch structure is used to enhance feature representation capabilities, while during the deployment phase, it is reparameterized into a single-branch structure to reduce inference computation costs. The detection decoding and detection head module takes the multi-scale feature set output by the bidirectional recursive refinement fusion module as input, performs target classification and location regression, and outputs the target category confidence and bounding box parameters.

[0009] Furthermore, the bidirectional recursive refinement and fusion module's bidirectional recursive refinement and fusion process includes two sub-processes: top-down refinement and bottom-up backfeedback. The top-down refinement sub-process is as follows: the high-level features are upsampled to the next scale through the dynamic sampling alignment module to make the mesh of the high-level features consistent with that of the lower-level features. The upsampled high-level features and the lateral features of the same scale are input into the reparameterized residual enhancement module for fusion enhancement to obtain the refined features of that scale. This process is passed down from F5 to F3 in this way to complete the refinement of all scales. The bottom-up refeedback process is as follows: the low-level refined features obtained from top-down refinement are downsampled and scaled before being passed to the high-level features. At each scale node, the passed low-level detailed features are fused with the existing features at that scale and input to the reparameterized residual enhancement module to obtain the refeedback high-level features. This process is repeated from F3 to F5 to form a recursive feedback loop from low to high.

[0010] Furthermore, the reparameterized residual enhancement module includes a fusion operator F, a first branch A, and a second branch B. The fusion operator F receives input features X1 and X2 from two different paths and fuses them to obtain an intermediate feature X. The first branch is a 1×1 convolutional branch used to perform channel alignment and linear mapping on the intermediate feature X, outputting feature A. The second branch B is a branch of a 1×1 convolutional cascaded reparameterization block. The reparameterization block is a combination of multi-branch convolutions / identity mappings during training and merged into an equivalent single convolution during deployment. The second branch B outputs feature B. Feature A and feature B are added element-wise to obtain the residual enhancement output Y of the reparameterized residual enhancement module.

[0011] Furthermore, the fusion operator is an element-wise addition operator, the reparameterization block is a RepBlock reparameterization block, the 1×1 convolution completes channel dimensionality reduction, and the RepBlock consists of a 3×3 convolution branch, a 1×1 convolution branch and an identity mapping branch during the training period. After training, the RepBlock is reparameterized into a single-branch 3×3 convolution through parameter fusion.

[0012] Furthermore, the dynamic sampling alignment module performs learnable alignment for the upsampling process from F5 to F4 and from F4 to F3. By learning the offset and weight of the sampling points, it enables high-level features to be accurately aligned to the pixel grid of low-level features, replacing the traditional bilinear interpolation upsampling.

[0013] Furthermore, the detection decoding and detection head module is an anchor-free detection head structure.

[0014] Furthermore, the resolution of the multi-scale feature maps F2, F3, F4, and F5 decreases by a factor of 2 at each step, while the number of channels increases by a factor of 2 at each step.

[0015] A bidirectional recursive refinement fusion network detection method for SAR target detection includes the following steps: Step 1: Acquire the raw SAR image and preprocess it. The preprocessing operations include amplitude normalization, and optional logarithmic transformation, size scaling / cropping, and data augmentation. The data augmentation operations include multi-scale training, random cropping, and 0~360° random rotation to adapt to the task requirements of SAR image target detection. Step 2: Input the preprocessed SAR image into the backbone feature extraction module to extract multi-scale feature maps F2, F3, F4, and F5 at different levels; the feature maps are progressively reduced in spatial resolution and progressively enhanced in semantic information, and then input into the feature encoding and decoding module to complete target detection; Step 3: Input feature maps F2~F5 into the bidirectional recursive refinement and fusion module to start the bidirectional recursive refinement and fusion process; Step 4: Execute the top-down refinement sub-process: Upsample the high-level features to the next scale through the dynamic sampling alignment module, and input them together with the horizontal features of the same scale into the reparameterized residual enhancement module for fusion enhancement to obtain the refined features of that scale. Repeat the operation step by step downward to obtain the refined features of the lower scale. Step 5: Execute the bottom-up refeeding sub-process: The low-level refined features are passed to the high-level features through downsampling / stride convolution. The downsampling adopts a 3×3 convolution with a stride of 2. At each scale node, the reparameterized residual enhancement module is used to fuse with the existing features of that scale to obtain the refeeded high-level features, forming a recursive feedback update from low to high. Step 6: Following the previous step, the bidirectional recursive refinement and fusion module outputs the fused multi-scale feature sets S3, S4, and S5; Step 7: Input the multi-scale feature sets S3, S4, and S5 into the detection decoding and detection head module, and output the target category confidence and bounding box parameters; Step 8: Post-process the output results of the detection decoding and detection head modules. The post-processing adopts ensemble prediction and one-to-one matching mechanism to obtain the final SAR target detection results.

[0016] Due to the adoption of the technical solution described above, the present invention has the following advantages: 1. The bidirectional recursive refinement fusion network structure and method for SAR target detection described in this invention sets up a dynamic sampling alignment module at the top-down cross-scale upsampling point. Through learnable sampling position and weight adjustment, it achieves adaptive alignment of high-level features to low-level grids, replacing the traditional fixed interpolation upsampling method. This effectively reduces feature misalignment and edge blurring problems, significantly improves the positioning stability of small and slender targets, and solves the problem of large sampling alignment errors in the prior art.

[0017] 2. This invention constructs a bidirectional recursive fusion mechanism of top-down refinement and bottom-up refeedback to achieve stable injection of high-level semantic features into low-level features and reverse compensation of low-level detailed features into high-level features. This solves the problem of insufficient cross-scale information transmission in existing technologies, enabling low-level features to have stronger discriminative capabilities and high-level features to have more accurate positioning information. Specifically, top-down refinement is used to improve false detections caused by "excessive detail and insufficient semantics" in low-level features, while bottom-up refeedback is used to improve the missed detection of small targets and positioning deviations caused by "excessive semantics and coarse positioning" in high-level features. It can be deduced that this invention can still improve classification and positioning accuracy simultaneously through higher-quality fusion features S3, S4, and S5 in the case of target detection without anchor boxes, thereby achieving the invention's objective of improving the robustness and accuracy of SAR target detection.

[0018] 3. This invention is designed to address the core characteristics of SAR images, such as strong noise, strong clutter, and multiple scales. Experimental verification on the OGSOD and HRSID public SAR datasets shows that the network significantly outperforms existing mainstream algorithms in overall detection accuracy, small target detection accuracy, and positioning accuracy. Moreover, it can effectively reduce false detections and missed detections in complex scenarios such as strong scattering structures in port areas / nearshore areas and dense targets in the open sea. At the same time, the network can be transferred to target detection tasks in images with strong noise and weak details, such as infrared and low-light images. It can also be adapted to the detection needs of extremely wide-swath and ultra-high-resolution SAR images through sliding window inference and overlap fusion. It has a wide range of applications and significant engineering application value. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the SAR target detection method based on BRRF of the present invention.

[0020] Figure 2 This is a schematic diagram of the overall structure of the Bidirectional Recursive Refinement Fusion Network (BRRF) of the present invention.

[0021] Figure 3 This is a schematic diagram of the reparameterized residual enhancement module (RREB) of the present invention.

[0022] Figure 4 This is a performance comparison chart of different detection methods on the OGSOD and HRSID datasets in Embodiment 1 of the present invention. Detailed Implementation

[0023] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments. However, this should not be construed as limiting the scope of protection of the present invention. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention.

[0024] The overall structure of the Bidirectional Recursive Refinement Fusion (BRRF) network for SAR target detection described in this invention is as follows: Figure 2 As shown, the system includes a cascaded backbone feature extraction module, a bidirectional recursive refinement and fusion module, and a detection decoding and detection head module. The data stream is passed sequentially from left to right. The bidirectional recursive refinement and fusion module integrates a dynamic sampling alignment module and multiple reparameterized residual enhancement modules, and finally outputs the fused multi-scale feature sets S3, S4, and S5. Since the semantic information of the multi-scale feature sets S1 and S2 is almost zero (only containing edges, corners, etc.), they are discarded. The backbone feature extraction module uses a classic convolutional neural network structure to perform hierarchical convolution on the preprocessed SAR image, outputting feature maps F2, F3, F4, and F5 at four scales, while discarding F1 (lowest semantic feature map / highest resolution). The resolution of each feature map decreases by a factor of 2, while the number of channels increases by a factor of 2. Among them, F2 has the highest resolution and contains rich details such as edge and texture information, while F5 has the strongest semantic information and can effectively distinguish the target category. The dynamic sampling alignment module is set at the cross-scale upsampling point of the top-down refinement path of the bidirectional recursive refinement fusion module. It performs learnable alignment for the upsampling process from F5 to F4 and from F4 to F3. By learning the offset and weight of the sampling points, it enables high-level features to be accurately aligned to the pixel grid of low-level features, replacing the traditional bilinear interpolation upsampling, eliminating cross-scale alignment errors and reducing the loss of detail information.

[0025] The structure of the reparameterized residual enhancement module is as follows: Figure 3 As shown, fusion nodes are distributed across scales F3, F4, and F5. Each fusion node's reparameterized residual enhancement module receives input features X1 and X2 from two different paths. These are first fused by the fusion operator F to obtain an intermediate feature X. The intermediate feature X is divided into two parallel branches. The first branch A is a Conv1×1 convolutional branch, completing channel alignment and linear feature mapping, outputting feature A. The second branch B is a Conv1×1 convolutional cascaded reparameterized block, with 1×1 convolutions performing channel dimensionality reduction. During training, the RepBlock consists of a Conv3×3 convolutional branch and a 1×1 convolutional branch. The RepBlock consists of a convolutional branch and an identity mapping branch, which enhances the feature representation capability. After training, RepBlock is reparameterized into a single-branch 3×3 convolution through parameter fusion. The second branch finally outputs feature B. Feature A and feature B are added element-wise to achieve residual enhancement and output the fused feature Y. This module ensures feature representation capability during training and is reparameterized into a single-branch convolution during deployment, which greatly reduces the amount of inference computation. The fusion operator F is an element-wise addition operator. The reparameterized block is the RepBlock reparameterized block. The 1×1 convolution completes channel dimensionality reduction. During training, RepBlock consists of a 3×3 convolutional branch, a 1×1 convolutional branch, and an identity mapping branch. After training, RepBlock is reparameterized into a single-branch 3×3 convolution through parameter fusion.

[0026] The core of the bidirectional recursive refinement fusion module is the bidirectional recursive refinement fusion mechanism. The bidirectional recursive refinement fusion module BRRF is the core module of the network. It takes the multi-scale feature maps F2~F5 output by the backbone feature extraction module as input, constructs a top-down refinement propagation path and a bottom-up semantic feedback path, and outputs the fused multi-scale feature sets S3, S4, and S5. Among them, the top-down refinement path is responsible for passing high-level semantic information to low-level scales step by step, and the bottom-up semantic feedback path is responsible for feeding back low-level detailed information to high-level scales step by step, so as to complete the recursive refinement of cross-scale features with fewer new parameters.

[0027] In this embodiment, the bidirectional recursive refinement and fusion process described above is as follows: The top-down refinement sub-process is as follows: taking the F5 output from the backbone as the initial high-level feature, upsampling it to the F4 scale through the dynamic sampling alignment module, and inputting it together with the lateral feature F4 into the reparameterized residual enhancement module at the F4 scale to obtain the refined feature at the F4 scale; upsampling the F4 scale refined feature to the F3 scale through the dynamic sampling alignment module, and inputting it together with the lateral feature F3 into the reparameterized residual enhancement module at the F3 scale to obtain the refined feature at the F3 scale; The bottom-up refeedback process is as follows: the refined features at the F3 scale are downsampled to the F4 scale through a 3×3 convolution with a stride of 2, and input together with the refined features at the F4 scale into the reparameterized residual enhancement module at the F4 scale to update the F4 scale features; the updated F4 scale features are downsampled to the F5 scale through a 3×3 convolution with a stride of 2, and input together with the original F5 features into the reparameterized residual enhancement module at the F5 scale to update the F5 scale features; the above top-down refinement and bottom-up refeedback process outputs the fused multi-scale feature sets S3, S4, and S5.

[0028] The detection decoding and detection head module adopts an anchor-free detection head structure. The multi-scale feature sets S3, S4, and S5 are input into the detection head to perform target category prediction and bounding box position regression, and output the target category confidence and bounding box parameters at each scale.

[0029] The flowchart of the SAR target detection method based on the above network is as follows: Figure 1 As shown, the specific steps are as follows: Step 1: Acquire the original SAR image and preprocess it. The preprocessing operations include amplitude normalization to map pixel values ​​to the [0,1] interval, logarithmic transformation to suppress speckle noise, scaling the image to a fixed size of 640×640, and data augmentation. The data augmentation operations include multi-scale training, random cropping, and 0~360° random rotation to adapt to the task requirements of SAR image target detection. Step 2: Input the preprocessed SAR image into the backbone feature extraction module, and extract multi-scale features through hierarchical convolution to obtain feature maps F2, F3, F4, and F5 with the resolution decreasing by a factor of 2 and the number of channels increasing by a factor of 2. Step 3: Input F2~F5 into the bidirectional recursive refinement and fusion module to start the bidirectional refinement and fusion process; Step 4: Perform top-down refinement. Upsample the high-level features to the next scale through the dynamic sampling alignment module, and input them together with the horizontal features of the same scale into the reparameterized residual enhancement module for fusion enhancement to obtain the refined features of that scale. Repeat the operation step by step downward to obtain the refined features of the lower scale. Step 5: Perform bottom-up refeedback, passing the low-level refined features to the high-level features through downsampling / stride convolution. The downsampling uses a 3×3 convolution with a stride of 2. At each scale node, the reparameterized residual enhancement module fuses the features with the existing features at that scale to obtain the refeeded high-level features, forming a recursive feedback update from low to high. Step 6: After completing steps 4 and 5, output the multi-scale feature sets S3, S4, and S5; Step 7: Input the multi-scale feature set S3~S5 into the detection decoding and detection head module, and output the target category confidence and bounding box parameters; Step 8: Post-process the output results of the detection decoding and detection head modules to obtain the final SAR target detection result; the post-processing algorithm is based on the ensemble prediction mechanism to complete target detection and generates the final detection result through a one-to-one matching method. Example

[0030] To verify the effectiveness and generalization ability of the technical solution proposed in this invention under complex scattering scenarios, this embodiment selects two representative public remote sensing datasets, OGSOD and HRSID, for comparative verification. OGSOD provides paired samples of spatially registered optical and SAR images, including targets such as bridges, oil tanks, and ports, and is significantly affected by scattering mechanisms, speckle noise, and sidelobes in the SAR mode. HRSID is a high-resolution SAR ship dataset, covering diverse scenarios such as ports / nearshore and offshore, and additionally provides pure background images for robustness testing. Both datasets are divided into training, validation, and test sets in an 8:1:1 ratio to verify the robustness and transferability of the proposed solution from two complementary perspectives: cross-modal differences and sea clutter / structural background interference. This embodiment performs end-to-end training in a unified experimental environment using PyTorch 2.1, CUDA 12.1, and NVIDIA A800. The results are comprehensively evaluated using standard accuracy metrics AP, AP_50, and AP_75, and efficiency metrics Size, GFLOPs, and FPS. Figure 4 As shown: On the OGSOD dataset, the embodiments of this invention achieve an overall accuracy (AP) of 51.8%, significantly outperforming two-stage methods such as Faster R-CNN and Cascade R-CNN, as well as typical single-stage methods such as ATSS and RetinaNet. Furthermore, it maintains its advantage in more stringent localization metrics, achieving 90.2% AP_50 and 50.7% AP_75, and achieving 45.9% AP_s in small target detection. This demonstrates a targeted improvement on the key challenge of SAR detection: "small targets are easily overwhelmed by noise." The experimental results are shown in the table below:

[0031] On the HRSID dataset, the AP of the present invention reaches 71.2%, AP_50 is 94.3%, AP_75 is 81.1%, the AP_s of small targets reaches 59.9%, and the competitive indicators of medium / large targets remain competitive with AP_m at 82.9% and AP_l at 83.7%, indicating that the present invention has good balance and stability on targets of different scales.

[0032] In addition to improved accuracy, this invention also offers clear benefits in engineering deployment: it achieves an inference speed of 77.6 FPS on OGSOD while maintaining a deployable range in terms of parameter and computational load (73.0M parameters, 68.6 GFLOPs), meeting the speed and resource requirements of practical applications; compared to the baseline RT-DETR, this invention maintains a higher operating speed while improving accuracy, demonstrating better deployment availability.

[0033] The comparative results show that, under conditions of strong scattering structures and dense target distribution in port areas / nearshore areas, the present invention can reduce false detections and missed detections caused by background structural clutter, and maintain more stable detection of weakly scattering, small-scale targets. The detection box is more consistent with the target position / scale, and there are fewer redundant boxes, thereby improving the reliability and usability of the output results in complex scenarios.

[0034] The parts of this invention not described in detail are prior art.

[0035] The embodiments selected herein for the purpose of disclosing the inventive objectives are currently considered suitable; however, it should be understood that the invention is intended to include all variations and modifications of the embodiments that fall within the scope of this concept and invention.

Claims

1. A bidirectional recursive refinement and fusion network structure for SAR target detection; characterized in that: It includes a cascaded backbone feature extraction module, a bidirectional recursive refinement and fusion module, and a detection decoding and detection head module. The bidirectional recursive refinement and fusion module integrates a dynamic sampling alignment module and multiple reparameterized residual enhancement modules. The backbone feature extraction module is used to perform hierarchical convolution feature extraction on the preprocessed SAR image and output multi-scale feature maps F2, F3, F4, and F5 from high-resolution layer to low-resolution layer and from low semantic information to high semantic information, where F2 is the highest resolution feature map and F5 is the highest semantic feature map. The bidirectional recursive refinement fusion module BRRF is the core module of the network. It takes the multi-scale feature maps F2~F5 output by the backbone feature extraction module as input, constructs a top-down refinement propagation path and a bottom-up semantic backfeeding path, and outputs the fused multi-scale feature sets S3, S4 and S5. The dynamic sampling alignment module is set at the cross-scale upsampling point of the top-down refinement of the transmission path, and performs learnable alignment for the upsampling process from F5 to F4 and from F4 to F3. The reparameterized residual enhancement module is distributed across the fusion nodes of the bidirectional recursive refinement fusion module at each scale. It is used to perform channel alignment, residual enhancement, and noise suppression on features from different paths / scales. During the training phase, a multi-branch structure is used to enhance feature representation capabilities, while during the deployment phase, it is reparameterized into a single-branch structure to reduce inference computation costs. The detection decoding and detection head module takes the multi-scale feature set output by the bidirectional recursive refinement fusion module as input, performs target classification and location regression, and outputs the target category confidence and bounding box parameters.

2. The bidirectional recursive refinement and fusion network structure for SAR target detection according to claim 1, characterized in that: The bidirectional recursive refinement and fusion module's bidirectional recursive refinement and fusion process includes two sub-processes: top-down refinement and bottom-up backfeedback. The top-down refinement sub-process is as follows: the high-level features are upsampled to the next scale through the dynamic sampling alignment module to make the mesh of the high-level features consistent with that of the lower-level features. The upsampled high-level features and the lateral features of the same scale are input into the reparameterized residual enhancement module for fusion enhancement to obtain the refined features of that scale. This process is passed down from F5 to F3 in this way to complete the refinement of all scales. The bottom-up refeedback process is as follows: the low-level refined features obtained from top-down refinement are downsampled and scaled before being passed to the high-level features. At each scale node, the passed low-level detailed features are fused with the existing features at that scale and input to the reparameterized residual enhancement module to obtain the refeedback high-level features. This process is repeated from F3 to F5 to form a recursive feedback loop from low to high.

3. The bidirectional recursive refinement and fusion network structure for SAR target detection according to claim 1, characterized in that: The reparameterized residual enhancement module includes a fusion operator F, a first branch A, and a second branch B. The fusion operator F is used to receive input features X1 and X2 from two different paths and fuse them to obtain an intermediate feature X. The first branch is a 1×1 convolutional branch, used to perform channel alignment and linear mapping on the intermediate feature X, and output feature A; The second branch B is a branch of a 1×1 convolutional cascaded reparameterization block. The reparameterization block is a combination of multi-branch convolutions / identity mappings during training and is merged into an equivalent single convolution during deployment. The second branch B outputs feature B. Feature A and feature B are added element by element to obtain the residual enhancement output Y of the reparameterized residual enhancement module.

4. The bidirectional recursive refinement fusion network structure for SAR target detection according to claim 1, characterized in that: The fusion operator is an element-wise addition operator, the reparameterization block is a RepBlock reparameterization block, the 1×1 convolution completes channel dimensionality reduction, and the RepBlock consists of a 3×3 convolution branch, a 1×1 convolution branch and an identity mapping branch during the training period. After training, the RepBlock is reparameterized into a single-branch 3×3 convolution through parameter fusion.

5. The bidirectional recursive refinement and fusion network structure for SAR target detection according to claim 1, characterized in that: The dynamic sampling alignment module performs learnable alignment for the upsampling process from F5 to F4 and from F4 to F3. By learning the offset and weight of the sampling points, it enables high-level features to be accurately aligned to the pixel grid of low-level features, replacing the traditional bilinear interpolation upsampling.

6. The bidirectional recursive refinement fusion network structure for SAR target detection according to claim 1, characterized in that: The detection decoding and detection head module is an anchor-free detection head structure.

7. The bidirectional recursive refinement fusion network structure for SAR target detection according to claim 1, characterized in that: The resolution of the multi-scale feature maps F2, F3, F4, and F5 decreases by a factor of 2, while the number of channels increases by a factor of 2.

8. A bidirectional recursive refinement fusion network detection method for SAR target detection according to any one of claims 1-7, comprising the following steps: Step 1: Acquire the raw SAR image and preprocess it. The preprocessing operations include amplitude normalization, and optional logarithmic transformation, size scaling / cropping, and data augmentation. The data augmentation operations include multi-scale training, random cropping, and 0~360° random rotation to adapt to the task requirements of SAR image target detection. Step 2: Input the preprocessed SAR image into the backbone feature extraction module to extract multi-scale feature maps F2, F3, F4, and F5 at different levels; the feature maps are progressively reduced in spatial resolution and progressively enhanced in semantic information, and then input into the feature encoding and decoding module to complete target detection; Step 3: Input feature maps F2~F5 into the bidirectional recursive refinement and fusion module to start the bidirectional recursive refinement and fusion process; Step 4: Execute the top-down refinement sub-process: Upsample the high-level features to the next scale through the dynamic sampling alignment module, and input them together with the horizontal features of the same scale into the reparameterized residual enhancement module for fusion enhancement to obtain the refined features of that scale. Repeat the operation step by step downward to obtain the refined features of the lower scale. Step 5: Execute the bottom-up refeeding sub-process: The low-level refined features are passed to the high-level features through downsampling / stride convolution. The downsampling adopts a 3×3 convolution with a stride of 2. At each scale node, the reparameterized residual enhancement module is used to fuse with the existing features of that scale to obtain the refeeded high-level features, forming a recursive feedback update from low to high. Step 6: Following the previous step, the bidirectional recursive refinement and fusion module outputs the fused multi-scale feature sets S3, S4, and S5; Step 7: Input the multi-scale feature sets S3, S4, and S5 into the detection decoding and detection head module, and output the target category confidence and bounding box parameters; Step 8: Post-process the output results of the detection decoding and detection head modules. The post-processing adopts ensemble prediction and one-to-one matching mechanism to obtain the final SAR target detection results.

Images