Sar rotating frame target detection method and device, equipment and medium

Abstract

The application relates to a SAR rotating frame target detection method, device, equipment and medium. The method comprises the following steps: first, a diffusion model is used for forward diffusion, variance scheduling is controlled to add Gaussian noise to a SAR sample true value detection frame, and a plurality of noisy rotating boundary frames are obtained. The plurality of noisy rotating boundary frames are input into a detection network together with a SAR image, a backbone network is used for enhancing features by using a double-path gated attention unit, adjacent scale features are optimized by using Mamba, a neck network is used for fusing features, and a detection head is used for cutting ROI features to predict a category and a boundary frame. The network is trained by using KL divergence and classification loss, a detected SAR image is input into the trained network together with a random noisy frame, a target rotating frame and a category are output, and the detection precision is improved.

Description

Technical Field

[0001] This application relates to the field of SAR target detection technology, and in particular to a SAR rotating frame target detection method, apparatus, equipment and medium. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses unique imaging advantages such as all-weather, all-time operation, and resistance to severe weather interference, making it a key technology in the field of Earth observation. Relying on airborne and spaceborne SAR systems, massive amounts of high-resolution SAR data can be acquired, and target detection technology based on this data has become a core research direction in remote sensing applications. SAR target detection has broad application prospects and plays an indispensable role in scenarios such as civilian maritime monitoring and management.

[0003] In recent years, with the widespread application of deep learning in the field of target detection, researchers have successively proposed SAR target horizontal bounding box detection algorithms based on deep learning, such as Denodet, SFS-CNet, DiffDet4SAR, and MaDiNet. However, SAR targets under overhead views often exhibit arbitrary rotation characteristics. Although horizontal bounding box detection methods can quickly locate the main target, they cannot effectively characterize the target's rotation angle. This not only easily leads to redundant detection but also introduces a large amount of redundant background information when detecting targets with large aspect ratios and arbitrary orientations, such as long ships.

[0004] Therefore, research on SAR rotating target detection methods is of significant practical importance. Against this backdrop, researchers have successively proposed SAR rotating target detection schemes such as PVT-SAR, CGA-Det, SFONet, and OFF-Net. Despite significant progress in this field, SAR target detection methods still face many unresolved challenges. For example, anchor-free detection methods suffer from the lack of core visual features, insufficient bounding box adaptability, inability to match target distribution and morphological characteristics, and interference from complex background clutter in target recognition and localization. Summary of the Invention

[0005] Therefore, it is necessary to provide a SAR rotating frame target detection method, device, equipment, and medium that can improve the accuracy of target recognition and positioning in response to the above-mentioned technical problems.

[0006] A SAR rotating frame target detection method, the method comprising:

[0007] Obtain a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0008] By utilizing forward diffusion in the diffusion model and controlling variance scheduling, Gaussian noise is added to the ground truth detection box of the SAR sample image to obtain multiple noisy rotated bounding boxes.

[0009] The SAR sample images and corresponding noisy rotated bounding boxes are input into a SAR rotated bounding box target detection network for rotation detection box prediction. This network includes a backbone network, a neck network, and a detection unit. In the backbone network, a dual-path gated attention unit enhances the extracted features to obtain multi-scale feature maps. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest (ROI) features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the ROI features with category and bounding box regression for prediction, obtaining the prediction result. The dual-path gated attention unit is composed of a dual-path adaptive channel attention submodule and an enhanced spatial attention submodule. The force submodule extracts channel statistics through parallel average and extreme paths, and introduces a gating unit to learn fusion coefficients to dynamically adjust the contribution of the two paths. The enhanced spatial attention submodule extracts two types of spatial cues in the channel dimension and concatenates these two types of spatial cues. The optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window re-integration unit. The adaptive window partitioning unit is used to partition the input features into multiple sub-window features after non-overlapping partitioning. The multiple sub-window features are input to the feature processing unit for processing. The feature processing unit has a 6-layer structure, with the first 3 layers being SARSSBlock blocks and the last 3 layers being self-attention blocks. The outputs of the SARSSBlock blocks and self-attention blocks are connected to a multilayer perceptron including layer normalization. The window re-integration unit re-concatenates the processed sub-window features output by the feature processing unit into a complete feature map to obtain the enhanced feature map.

[0010] By employing KL divergence loss and classification loss, the SAR rotating box target detection network is trained based on the prediction results and the corresponding ground truth labels, resulting in a trained SAR rotating box target detection network.

[0011] Acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. Input the SAR image and multiple noisy rotated bounding boxes into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0012] In one embodiment, the backbone network uses ResNet50, and the features output by the BN layer are enhanced before the dual-path gated attention unit is embedded in the residual connections of each Bottleneck block.

[0013] In one embodiment, the optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window re-integration unit;

[0014] The adaptive window partitioning unit is used to partition the input features into multiple sub-window features after non-overlapping partitioning;

[0015] Multiple windowed features are input into the feature processing unit for processing. The feature processing unit has a 6-layer structure, with the first 3 layers being SARSSBlock blocks and the last 3 layers being self-attention blocks. The outputs of both the SARSSBlock blocks and the self-attention blocks are connected to a multilayer perceptron including layer normalization.

[0016] The window re-integration unit reassembles the processed window features output by the feature processing unit into a complete feature map, resulting in an enhanced feature map.

[0017] In one embodiment, in the feature processing unit:

[0018] The SARSSBlock block includes a first branch and a second branch in parallel. The first branch includes a linear layer, a convolutional layer, a SiLU activation function, and a SARScan2D block connected in sequence. The second branch includes a linear layer, a convolutional layer, and a SiLU activation function connected in sequence. The output data of the two branches are concatenated and then passed through a linear layer to obtain the output of the SARSSBlock block.

[0019] The self-attention block employs a multi-head self-attention mechanism to map input features to multiple independent subspaces to calculate single-head self-attention. After concatenating the calculation results of all single-head self-attention, a projection matrix is ​​fused to obtain the output of the self-attention block.

[0020] In one embodiment, the SARScan2D block uses a four-way scanning mechanism to expand each of the sub-window features into a one-dimensional sequence, which is then processed by the corresponding S6Block block and reversed back to a two-dimensional sub-window feature.

[0021] In one embodiment, the detection unit comprises multiple detection heads connected in series;

[0022] The first detection head takes the multi-scale feature fusion map and the multiple noisy rotated bounding boxes as input to obtain the category prediction result and the rotated box prediction result.

[0023] Each subsequent detection head uses the rotation box prediction result output by the previous detection head and the multi-scale feature fusion map as input to repeat the feature interaction and prediction process;

[0024] The category prediction result and the rotation box prediction result output by the last detection head are used as the output of the detection unit.

[0025] In one embodiment, in each of the detection heads:

[0026] Based on the multi-scale feature fusion map and the prediction results of the multiple noisy rotated bounding boxes or rotated boxes, the region of interest features with different levels of matching are obtained through the rotation region of interest alignment operation.

[0027] Multi-dimensional feature interaction operations and normalization regularization processing are performed on the features of the region of interest to obtain the interactive features. The interactive features are then processed according to the dynamic interaction mechanism between instances and normalization regularization to obtain the dynamic interactive features with optimized instance information association.

[0028] Based on the dynamic interaction features, a nonlinear transformation operation is performed on the feedforward network to obtain a nonlinear transformation feature with enhanced feature representation capability. The nonlinear transformation feature is then combined with the temporal embedding information of the diffusion process to perform a temporal feature fusion and adjustment operation, resulting in an adjusted feature with fused temporal information and optimized parameters.

[0029] Based on the adjusted features, the category prediction result and the rotated bounding box prediction result are obtained through classification branch and regression branch processing, respectively.

[0030] In one embodiment, when calculating the KL divergence loss, the rotated bounding box prediction result and the corresponding ground truth detection box are mapped to Gaussian distributions respectively, and then the KL divergence value between the two Gaussian distributions is calculated.

[0031] This application also provides a SAR rotating frame target detection device based on a diffusion model, the device implementing the above-described SAR rotating frame target detection method, the device comprising:

[0032] The SAR training set acquisition module is used to acquire a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0033] The forward diffusion noise module is used to add Gaussian noise to the ground truth detection box of the SAR sample image by controlling the variance scheduling and utilizing the forward diffusion in the diffusion model to obtain multiple noisy rotated bounding boxes.

[0034] A neural network prediction module is used to input the SAR sample image and its corresponding multiple noisy rotated bounding boxes into a SAR rotated bounding box target detection network for rotation detection box prediction. The SAR rotated bounding box target detection network includes a backbone network, a neck network, and a detection unit. In the backbone network, a dual-path gated attention unit is used to enhance the extracted features and obtain a multi-scale feature map. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the region of interest features with category and bounding box regression for prediction to obtain the prediction result.

[0035] The neural network training module is used to train the SAR rotating box target detection network by using KL divergence loss and classification loss, based on the prediction results and the corresponding ground truth labels, to obtain the trained SAR rotating box target detection network.

[0036] The SAR target detection module is used to acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. The SAR image and multiple noisy rotated bounding boxes are input into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0037] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the specific steps in the above-described SAR rotating frame target detection method based on a diffusion model.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the specific steps in the above-described SAR rotating frame target detection method based on a diffusion model.

[0039] The aforementioned SAR rotating bounding box target detection method, apparatus, device, and medium utilize forward diffusion in the diffusion model. By controlling variance scheduling, Gaussian noise is added to the ground truth detection boxes of SAR sample images to obtain multiple noisy rotating bounding boxes. The SAR sample images and the corresponding multiple noisy rotating bounding boxes are input into the SAR rotating bounding box target detection network for rotating detection box prediction. The SAR rotating bounding box target detection network includes a backbone network, a neck network, and detection units. In the backbone network, dual-path gated attention units are used to enhance the extracted features and obtain multi-scale feature maps. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest features corresponding to each noisy rotating bounding box from the multi-scale feature fusion map output from the neck network, and then combines the region of interest features to perform category and bounding box regression for prediction, obtaining the prediction result. By employing KL divergence loss and classification loss, the SAR rotating bounding box target detection network is trained based on the prediction result and the corresponding ground truth labels to obtain a trained SAR rotating bounding box target detection network. The method involves acquiring a SAR image to be detected and randomly generating multiple noisy rotated bounding boxes. The SAR image and these noisy rotated bounding boxes are then input into a trained SAR rotated bounding box target detection network for target detection, yielding the target's rotated bounding box and target category. This method can obtain rotated detection boxes consistent with the target's pose, thus improving detection accuracy. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a SAR rotating frame target detection method in one embodiment;

[0041] Figure 2 This is a simplified schematic diagram of the SAR rotating box target detection network training process in one embodiment;

[0042] Figure 3 This is a schematic diagram showing the location and structure of a dual-path gating attention unit in one embodiment;

[0043] Figure 4 This is a schematic diagram of the optimized Mamba unit structure in one embodiment;

[0044] Figure 5 This is a schematic diagram of the structure of the SARSSBlock block in one embodiment;

[0045] Figure 6 This is a structural block diagram of a SAR rotating frame target detection device in one embodiment;

[0046] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] To address the problems existing in current SAR target detection methods, this application, such as Figure 1 As shown, a SAR rotating frame target detection method is proposed, which specifically includes the following steps:

[0049] Step S100: Obtain the SAR training set. The SAR training set includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0050] Step S110: Using the forward diffusion in the diffusion model, Gaussian noise is added to the ground truth detection box of the SAR sample image by controlling variance scheduling to obtain multiple noisy rotated bounding boxes.

[0051] Step S120: The SAR sample image and its corresponding multiple noisy rotated bounding boxes are input into the SAR rotated bounding box target detection network for rotation detection box prediction. The SAR rotated bounding box target detection network includes a backbone network, a neck network, and a detection unit. In the backbone network, the extracted features are enhanced using a dual-path gating attention unit to obtain multi-scale feature maps. The optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layer of the backbone network and inputs them into the neck network along with other multi-scale feature maps. The detection unit crops the region of interest features corresponding to each noisy rotated bounding box in the multi-scale feature fusion map output from the neck network, and then combines the region of interest features to perform category and bounding box regression for prediction to obtain the prediction result.

[0052] Step S130: By using KL divergence loss and classification loss, the SAR rotating box target detection network is trained based on the prediction results and the corresponding ground truth labels to obtain the trained SAR rotating box target detection network.

[0053] Step S140: Acquire the SAR image to be detected, and randomly generate multiple noisy rotated bounding boxes. Input the SAR image and multiple noisy rotated bounding boxes into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0054] In this method, the bounding box regression task is described as generating the bounding box's position, shape, and orientation (angle). The advantage of this approach is that it does not require the generation of prior candidate boxes, thus avoiding insufficient localization accuracy and missed detections. This paper proposes a method for detecting rotating target bounding boxes by combining a diffusion model and a SAR target detection network. Figure 2 The simplified overall workflow is demonstrated. During the training phase, Gaussian noise is added to the ground truth bounding box data by controlling variance scheduling to obtain noisy rotated bounding boxes. During inference, the learned diffusion process is reversed to adjust the noise distribution to match the learned bounding box distribution, thereby generating predicted bounding boxes.

[0055] In the diffusion model, the forward diffusion process can be defined as:

[0056] ;

[0057] By Add Gaussian noise to it. Converted into latent noise samples ,in , .

[0058] Furthermore, in this method, KL divergence is introduced to construct the loss function KLD Loss (step S140), and a joint optimization strategy is adopted to solve the target representation method. This addresses the issue of periodic angular boundaries and improves detection accuracy. KLD Loss transforms the rotated bounding box into a Gaussian distribution representation, and the properties of the Gaussian distribution can circumvent the periodic angular boundary problem. The rotated bounding box representing the target... Convert to Gaussian distribution :

[0059] ;

[0060] In the above formula, Represents the rotation matrix. This represents the characteristic diagonal matrix.

[0061] Specifically, the Gaussian distributions corresponding to the predicted and true values. and The KLD between them is represented as:

[0062] ;

[0063] Obviously, Each item in the rotating box The parameters are related to some of the parameters, and all the rotation box parameters are interconnected and influence each other during training, achieving joint optimization and automatic adjustment.

[0064] In step S100, for the entire detection task during the training process, the input is... ,in The input is a remote sensing image, i.e., a SAR image. For bounding box, This refers to the label corresponding to the bounding box. In this method, the long-side definition method is used. To indicate the first There are bounding boxes, among which Indicates the coordinates of the center point of the bounding box. This represents the long and short sides of the bounding box. The angle between the positive axis and the long side of the bounding box is denoted as . , range .

[0065] In this embodiment, the data samples are set as a set of bounding boxes. ,in , indicating that it contains A set of bounding boxes, with five dimensions corresponding to the center point coordinates, width, height, and angle of each bounding box, respectively. Train a neural network. That is, the SAR rotating bounding box target detection network makes it rotate from a noisy bounding box. Predicted target boxes And the corresponding category labels, the training process minimizes the training objective KLD loss. accomplish.

[0066] like Figure 2 As shown, the SAR rotating bounding box target detection network in step S120 includes three parts: a backbone network, a neck network, and a detection unit. The backbone network is used to extract multi-scale features of the SAR image, and the neck network performs fusion between adjacent scales to obtain multi-scale fused features. Finally, the detection unit performs target detection based on the multi-scale fused features and the corresponding multiple noisy rotating bounding boxes.

[0067] In this embodiment, the backbone network uses ResNet50, and the dual-path gated attention unit is embedded before the residual connection in each Bottleneck block, such as... Figure 3 As shown, feature enhancement is performed on the features output by the BN layer. From Figure 2 As you can see, after the SAR sample image is input into the backbone network, it goes through a convolutional layer and then through four Bottleneck blocks in sequence to obtain four feature maps of different scales.

[0068] like Figure 3As shown, the Dual-path Gated Attention Module (DGAM) unit consists of a dual-path adaptive channel attention submodule and an enhanced spatial attention submodule. The core innovation of the DGAM unit lies in constructing a dynamic dual-channel and spatially collaborative attention mechanism. By embedding gating units within the dual-path channel attention, dynamic fusion of channel weights is achieved. Simultaneously, combined with the enhanced spatial attention module, a complete feature enhancement process is formed.

[0069] Specifically, channel attention aims to filter out the feature channels that are more critical for target detection. DGAM first uses a dual-path adaptive channel attention submodule, combined with a gating unit to dynamically fuse channel weights, to achieve the discrimination and enhancement of key feature channels of SAR targets. For the input feature map... ( For batch size, For the number of channels, (For spatial dimensions), dual-path feature extraction extracts channel statistics through two parallel paths. One is the averaging path, which captures the channel mean features through global average pooling. This reflects the overall response level of the channel:

[0070] ;

[0071] The second approach is the extreme value path, which captures channel extreme value features through global max pooling. Highlighting the local response of strongly scattering targets:

[0072] ;

[0073] To avoid parameter redundancy and cross-channel interference inherent in traditional multilayer perceptrons (MLPs), the adaptive channel weights for both paths are calculated using grouped one-dimensional convolutions. The input is flattened into a one-dimensional sequence, and the convolution kernel size... By number of channels Adaptive determination:

[0074] ;

[0075] In particular, if If the value is even, add 1 to make it odd to ensure the symmetry of the convolution kernel and thus adapt to features of different channel sizes. Perform independent convolution on each channel to ensure no cross-channel information mixing, then obtain the weights through a sigmoid activation function. and .

[0076] Introducing gating units to learn fusion coefficients This allows for dynamic adjustment of the contributions from the two paths; this process is called gating dynamic fusion. and When pieced together, Convolution compression channels to Then generated via Sigmoid ,Right now:

[0077] ;

[0078] The final channel weight is a dynamic weighted sum of the two paths:

[0079] ;

[0080] in, In areas with dense noise, the extreme path weights tend to be reduced, i.e., increased, to suppress noise mean interference; in areas with stable targets, the mean path weights tend to be increased, i.e. increased, to preserve overall characteristics.

[0081] Subsequently, DGAM achieves precise positioning of key areas in the target space through an enhanced spatial attention submodule. This is based on the characteristics of channel attention enhancement. Two types of spatial cues are extracted in the channel dimension: one is obtained by averaging. To reflect the overall activation level of spatial location; secondly, to obtain the highest value. To highlight the target edges, strong scattering centers, and other high-response areas, the two are stitched together along the channel dimension to obtain... This approach fuses two types of spatial information. A CBM unit consists of convolution, batch normalization, and a Mish activation function. The CBM unit is used to process the concatenated features, which are then processed by a Sigmoid activation function to obtain spatial weights. This enhances the discriminative power of spatial features. Specifically:

[0082] ;

[0083] The final output enhancement feature is:

[0084] .

[0085] In this embodiment, an optimized Mamba unit is also set between the backbone network and the neck network. This unit integrates Mamba with a self-attention mechanism, and a MambaViSAR unit is proposed for SAR rotating target detection tasks, such as... Figure 4 As shown.

[0086] In this embodiment, the optimized Mamba unit first performs adaptive window partitioning on the input features, then sequentially passes through a SARSSBlock block and a self-attention block, and finally outputs the processed features through window recombination. The adaptive window partitioning can adapt to SAR images of different sizes, while the self-attention block further enhances the model's global perception capability.

[0087] like Figure 4 As shown, the optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window reverse unit. The adaptive window partitioning unit divides the input features into multiple non-overlapping sub-window features. These sub-window features are then fed into the feature processing unit, a 6-layer structure. The first three layers are SARSSBlock blocks, and the last three layers are self-attention blocks. The outputs of both the SARSSBlock and self-attention blocks are connected to a multilayer perceptron (MLP) with layer normalization. Finally, the window reverse unit reassembles the processed sub-window features from the feature processing unit into a complete feature map, resulting in the enhanced feature map.

[0088] Specifically, assuming input features The sequence length is Embedded latitude is Passing through the MambaViSAR unit in sequence There are several levels, among which the first The layer is the SARSSBlock module, then... If the layer is a self-attention block, then the first... The output of the layer can be represented as:

[0089] ;

[0090] In the above formula, Norm uses layer normalization without loss of generality.

[0091] In this embodiment, n is set to 6, meaning a 6-layer structure is designed in the MambaViSAR cell.

[0092] Furthermore, to adapt to the specific requirements of SAR target detection tasks, the original Mamba block structure was specifically optimized, and the SARSSBlock block was designed and proposed, with the structure as follows: Figure 5As shown. In the original Mamba model, the core causal convolution is well-suited for tasks such as language sequence modeling due to its unidirectional information transmission characteristics. However, in object detection scenarios, this unidirectional constraint is not only unnecessary but also limits the global interaction capability of features. Therefore, in this method, it is replaced with regular convolution to support bidirectional information flow.

[0093] Furthermore, considering that the intensity of the target scattering center in SAR images is highly sensitive to the observation azimuth, a four-directional scanning mechanism is integrated into the SSM architecture, proposing the SARScan2D block. This block can capture contextual information from multiple directions, enabling multi-angle and omnidirectional modeling of target features, thus better adapting to the azimuth dependence of target scattering characteristics in SAR images.

[0094] Furthermore, to compensate for potential content information loss due to constraints during SSM sequence modeling, a parallel branch symmetrical to the main branch and lacking the SARScan2D module is introduced. This branch enhances the fine-grained capture of local features through additional convolutional layers and the SiLU activation function. Finally, the output features of the two branches are concatenated, and feature projection is performed through a linear layer to integrate global context and local detail information, providing more comprehensive feature support for subsequent SAR target detection.

[0095] In this embodiment, within the feature processing unit: the SARSSBlock block includes a parallel first branch and a second branch. The first branch includes a linear layer, a convolutional layer, a SiLU activation function, and a SARScan2D block connected in sequence. The second branch includes a linear layer, a convolutional layer, and a SiLU activation function connected in sequence. The output data of the two branches are concatenated and then passed through a linear layer to obtain the output of the SARSSBlock block. The self-attention block employs a multi-head self-attention mechanism to map the input features to multiple independent subspaces to calculate single-head self-attention. After concatenating the calculation results of all single-head self-attention, a projection matrix is ​​fused to obtain the output of the self-attention block.

[0096] Specifically, given input SARSSBlock output It can be calculated as follows:

[0097] ;

[0098] In the above formula, The input and output embedding dimensions are respectively and Linear layers. This is the activation function SiLU. Furthermore... and These represent one-dimensional regular convolution and splicing operations, respectively.

[0099] Furthermore, SAR targets generally exhibit structural heterogeneity, and their scattering characteristics undergo significant dynamic fluctuations when imaging conditions change, such as incident angle tilt and azimuth deflection, posing challenges to feature extraction and task modeling. The selective scanning operation in Mamba is essentially designed for processing one-dimensional sequence data, making it difficult to adapt to the spatial feature modeling requirements of SAR target detection, and also unable to adequately address the complex scattering characteristics of SAR targets. Therefore, to better suit the SAR target detection task, this method proposes a four-directional scanning fusion mechanism. This mechanism works in conjunction with S6 block segmentation to form the SARScan2D block. Within the SARScan2D block, the four-directional scanning mechanism unfolds the features of each sub-window into a one-dimensional sequence, which is then processed by the corresponding S6Block block and inversely restored to two-dimensional sub-window features. The specific implementation process involves cross-scanning, S6 block segmentation, and cross-fusion.

[0100] Specifically, in the SARScan2D block, cross-scanning divides the input data into multiple non-overlapping image blocks, which are then expanded along four different scan paths (from left to right, from top to bottom, from bottom to top, and from right to left). Each scan path forms a sequence, transforming the 2D image data into a one-dimensional sequence for processing. Each scan path undergoes independent S6 block feature extraction. The sequences from the four scan directions are then recombine, and a reverse operation restores the one-dimensional sequence to a two-dimensional image. This processing ensures full coverage of the input image, effectively reduces the directional sensitivity of SAR images, and provides a more robust feature base for subsequent target detection.

[0101] Furthermore, Mamba, relying on a state-space model, achieves linear complexity by compressing global context information and combining it with selective scanning. However, this compression process condenses global information into hidden states, inevitably leading to the loss of fine-grained local dependencies between feature units, which is detrimental to SAR target detection tasks. To address this, a self-attention structure is introduced in the SARSSBlock backend. This structure can focus on different regions of the image, supplementing the local features lost by Mamba due to filtering global redundant information. The synergistic effect of the self-attention mechanism and Mamba enables accurate representation of global information.

[0102] Specifically, a multi-head self-attention mechanism is used in the MambaViSAR cell, while the single-head self-attention mechanism can be represented as:

[0103] ;

[0104] In the above formula, , , These are query, key, and value matrices, respectively. The dimension of the key vector. This is used to mitigate the gradient problem caused by excessively large dot product results. Then, Attention results for each independent subspace ( For the first The projection matrices of each subspace are concatenated and output as the projection matrix. The results of the fusion are obtained as follows:

[0105] ;

[0106] This mechanism, through parallel modeling across multiple subspaces, can more comprehensively capture the global and local relationships between features.

[0107] from Figure 2 As can be seen, the MambaViSAR unit is specifically set at the output of the second layer (C3) and the third layer (C4) of the backbone network. Specifically, the features output by C3 are used as the output, processed, and then added to the feature map output by C4 before being input into the neck network for further processing.

[0108] In this embodiment, the neck network uses a Feature Pyramid Network (FPN).

[0109] In this embodiment, the detection unit comprises multiple detection heads connected in series. The first detection head takes a multi-scale feature fusion map and multiple noisy rotated bounding boxes as input to obtain a category prediction result and a rotated bounding box prediction result. Subsequent detection heads take the rotated bounding box prediction result and the multi-scale feature fusion map output by the previous detection head as input and repeat the feature interaction and prediction process. The category prediction result and the rotated bounding box prediction result output by the last detection head are used as the output of the detection unit.

[0110] In this embodiment, the target detection results output by the SAR rotating box target detection network include the target category and the target rotating detection box used for target localization.

[0111] Specifically, the detection unit consists of 6 detection heads connected in series, each detection head using multi-scale fused feature maps (i.e., Figure 2 The input consists of P5, P4, P3, P2, the predicted rotated bounding boxes, or multiple noisy rotated bounding boxes (bboxes with shape (N, nr_boxes, 5), where N is the batch size and nr_boxes is the number of candidate boxes per batch) and the time embedding vector time_emb.

[0112] Furthermore, in each detection head: based on the multi-scale feature fusion map and multiple noisy rotated bounding boxes or rotated bounding box prediction results, region of interest features with different levels of matching are obtained through rotational region of interest alignment operations. Multi-dimensional feature interaction operations and normalization regularization processing are performed on the region of interest features to obtain interactive features. Based on the interactive features, dynamic interactive features with optimized instance information association are obtained through dynamic interaction mechanisms and normalization regularization processing. Based on the dynamic interactive features, a nonlinear transformation operation is performed on the feedforward network to obtain nonlinear transformation features with enhanced feature representation capabilities. The nonlinear transformation features are then fused with the temporal embedding information of the diffusion process, and a temporal feature fusion adjustment operation is performed to obtain adjusted features with fused temporal information and optimized parameters. Based on the adjusted features, classification and regression branches are used to obtain category prediction results and rotated bounding box prediction results, respectively.

[0113] Specifically, in each detection head, features and bounding boxes are first matched hierarchically and ROI features are extracted using Rotated RoI Align to obtain pro_features. Then, pro_features undergo feature interaction through a multi-head attention layer and are normalized and regularized using LayerNorm and Dropout to obtain interacted features. Next, the interacted features enter a dynamic convolutional layer (DynamicConv) to achieve dynamic feature interaction between instances, and are also processed by LayerNorm and Dropout to obtain dynamic interactive features. Afterwards, the dynamic interactive features pass through a feedforward network (Linear layer), Dropout, ReLU activation function, and another Linear layer, combined with LayerNorm to perform nonlinear feature transformation to obtain nonlinear transformed features. The nonlinear transformed features then pass through a temporal embedding fusion module (block_time_mlp), where the SiLU activation function and Linear layer fuse temporal information with features, generating scaling and offset factors to adjust the features. Simultaneously, temporal embedding also achieves the purpose of locating the time step of the diffusion process, resulting in adjusted features.

[0114] Next, the adjusted features are processed by the classification and regression branches. The classification branch is processed by a cls_module consisting of a linear layer, layer normalization, and ReLU activation function, while the regression branch is processed by a reg_module consisting of three interconnected layers. Finally, the classification branch outputs the class prediction score class_logits (shape (N, nr_boxes, num_classes)) through a linear layer (class_logits), while the regression branch outputs the rotated bounding box offset deltas through a linear layer and calculates the rotated box prediction result pred_bboxes (shape (N, nr_boxes, 5)).

[0115] Furthermore, the calculation process in each detection head is as follows:

[0116] ; ;

[0117] ;

[0118] ;

[0119] In the above formula, For the prediction results of the rotated box, The offset is deltas. As weight, This is the predicted rotating frame obtained by the detection head.

[0120] In this embodiment, the results obtained by each detection head are used to replace the bounding boxes and are passed to the next detection head. The above steps are repeated. The predicted classification result and the rotated bounding box result obtained by the last detection head are passed to the classification loss function FocalLoss and the rotated bounding box loss function KLD Loss, respectively, thereby completing the classification and detection task of the rotated target.

[0121] In step S130, the SAR rotating box target detection network is trained using KL divergence loss and classification loss, based on the prediction results and corresponding ground truth labels, to obtain the trained SAR rotating box target detection network. The KL divergence loss has already been explained above and will not be repeated here. The classification loss is a commonly used loss function in target detection training and will also not be discussed further here.

[0122] Finally, in step S140, the trained SAR rotating bounding box target detection network is used to perform target detection based on the SAR image to be detected and the randomly generated noisy rotating bounding box, so as to obtain the target category prediction result and the rotating detection box for target localization.

[0123] Inspired by the diffusion model, the above-mentioned SAR rotating bounding box target detection method, to meet the requirements of SAR rotating target detection, constructs the detection task as a denoising diffusion process for rotating bounding boxes. This method abandons complex operations such as manually setting anchor box sizes, and instead directly samples the position, shape, and angle information of the rotating box from a Gaussian distribution, thereby significantly improving positioning accuracy and better adapting to targets with different orientations and aspect ratios. Based on Mamba's core state-space model (SSM), a further optimized Mamba ViSAR unit, capable of effectively modeling global context information, is proposed and embedded together with a dual-path gated attention unit in the feature extraction network, possessing both strong long-range dependency capability and key feature extraction capability.

[0124] Specifically, this method proposes a SAR rotating bounding box target detection framework based on a diffusion model. This framework transforms the SAR rotating target detection task into a denoising diffusion process of the rotating bounding box, which can directly generate the position, shape, and angle information of the target's rotating bounding box. This design does not require pre-set a priori anchor boxes and can fully adapt to the characteristics of sparse SAR target distribution, significant aspect ratio differences, and arbitrary orientation, exhibiting extremely strong robustness.

[0125] Furthermore, this method embeds a self-developed dual-path gated attention unit and a MambaViSAR unit into the feature extraction backbone network. The dual-path gated attention unit can effectively extract key target features, while the MambaViSAR unit can fully capture long-range dependencies. The two complement each other, effectively alleviating the problem of key point isolation caused by target sparseness in SAR images, suppressing complex background interference, enhancing overall semantic representation, and significantly improving detection performance.

[0126] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0127] In one embodiment, such as Figure 6 As shown, a SAR rotating frame target detection device is provided, including: a SAR training set acquisition module 200, a forward diffusion noise addition module 210, a neural network prediction module 220, a neural network training module 230, and a SAR target detection module 240, wherein:

[0128] SAR training set acquisition module 200 is used to acquire SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0129] The forward diffusion noise module 210 is used to add Gaussian noise to the ground truth detection box of the SAR sample image by controlling variance scheduling and utilizing the forward diffusion in the diffusion model to obtain multiple noisy rotated bounding boxes.

[0130] The neural network prediction module 220 is used to input the SAR sample image and the corresponding multiple noisy rotated bounding boxes into the SAR rotated bounding box target detection network for rotation detection box prediction. The SAR rotated bounding box target detection network includes a backbone network, a neck network, and a detection unit. In the backbone network, the extracted features are enhanced using a dual-path gating attention unit to obtain a multi-scale feature map. The optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layer of the backbone network and inputs them into the neck network along with other multi-scale feature maps. The detection unit crops the region of interest features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the region of interest features to perform category and bounding box regression for prediction to obtain the prediction result.

[0131] The neural network training module 230 is used to train the SAR rotating box target detection network by using KL divergence loss and classification loss, based on the prediction results and the corresponding ground truth labels, to obtain the trained SAR rotating box target detection network.

[0132] The SAR target detection module 240 is used to acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. The SAR image and multiple noisy rotated bounding boxes are input into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0133] Specific limitations regarding the SAR rotating frame target detection device based on the diffusion model can be found in the limitations of the SAR rotating frame target detection method based on the diffusion model mentioned above, and will not be repeated here. Each module in the aforementioned SAR rotating frame target detection device based on the diffusion model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0134] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a SAR rotating frame target detection method based on a diffusion model. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0135] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0136] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0137] Obtain a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0138] By utilizing forward diffusion in the diffusion model and controlling variance scheduling, Gaussian noise is added to the ground truth detection box of the SAR sample image to obtain multiple noisy rotated bounding boxes.

[0139] The SAR sample images and corresponding noisy rotated bounding boxes are input into a SAR rotated bounding box target detection network for rotation detection box prediction. This network includes a backbone network, a neck network, and a detection unit. In the backbone network, a dual-path gated attention unit enhances the extracted features to obtain multi-scale feature maps. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest (ROI) features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the ROI features with category and bounding box regression for prediction, obtaining the prediction result. The dual-path gated attention unit is composed of a dual-path adaptive channel attention submodule and an enhanced spatial attention submodule. The force submodule extracts channel statistics through parallel average and extreme paths, and introduces a gating unit to learn fusion coefficients to dynamically adjust the contribution of the two paths. The enhanced spatial attention submodule extracts two types of spatial cues in the channel dimension and concatenates these two types of spatial cues. The optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window re-integration unit. The adaptive window partitioning unit is used to partition the input features into multiple sub-window features after non-overlapping partitioning. The multiple sub-window features are input to the feature processing unit for processing. The feature processing unit has a 6-layer structure, with the first 3 layers being SARSSBlock blocks and the last 3 layers being self-attention blocks. The outputs of the SARSSBlock blocks and self-attention blocks are connected to a multilayer perceptron including layer normalization. The window re-integration unit re-concatenates the processed sub-window features output by the feature processing unit into a complete feature map to obtain the enhanced feature map.

[0140] By employing KL divergence loss and classification loss, the SAR rotating box target detection network is trained based on the prediction results and the corresponding ground truth labels, resulting in a trained SAR rotating box target detection network.

[0141] Acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. Input the SAR image and multiple noisy rotated bounding boxes into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0143] Obtain a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose.

[0144] By utilizing forward diffusion in the diffusion model and controlling variance scheduling, Gaussian noise is added to the ground truth detection box of the SAR sample image to obtain multiple noisy rotated bounding boxes.

[0145] The SAR sample images and corresponding noisy rotated bounding boxes are input into a SAR rotated bounding box target detection network for rotation detection box prediction. This network includes a backbone network, a neck network, and a detection unit. In the backbone network, a dual-path gated attention unit enhances the extracted features to obtain multi-scale feature maps. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest (ROI) features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the ROI features with category and bounding box regression for prediction, obtaining the prediction result. The dual-path gated attention unit is composed of a dual-path adaptive channel attention submodule and an enhanced spatial attention submodule. The force submodule extracts channel statistics through parallel average and extreme paths, and introduces a gating unit to learn fusion coefficients to dynamically adjust the contribution of the two paths. The enhanced spatial attention submodule extracts two types of spatial cues in the channel dimension and concatenates these two types of spatial cues. The optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window re-integration unit. The adaptive window partitioning unit is used to partition the input features into multiple sub-window features after non-overlapping partitioning. The multiple sub-window features are input to the feature processing unit for processing. The feature processing unit has a 6-layer structure, with the first 3 layers being SARSSBlock blocks and the last 3 layers being self-attention blocks. The outputs of the SARSSBlock blocks and self-attention blocks are connected to a multilayer perceptron including layer normalization. The window re-integration unit re-concatenates the processed sub-window features output by the feature processing unit into a complete feature map to obtain the enhanced feature map.

[0146] By employing KL divergence loss and classification loss, the SAR rotating box target detection network is trained based on the prediction results and the corresponding ground truth labels, resulting in a trained SAR rotating box target detection network.

[0147] Acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. Input the SAR image and multiple noisy rotated bounding boxes into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0149] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A SAR rotating frame target detection method, characterized in that, The method includes: Obtain a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose. By utilizing forward diffusion in the diffusion model and controlling variance scheduling, Gaussian noise is added to the ground truth detection box of the SAR sample image to obtain multiple noisy rotated bounding boxes. The SAR sample images and corresponding noisy rotated bounding boxes are input into a SAR rotated bounding box target detection network for rotation detection box prediction. This network includes a backbone network, a neck network, and a detection unit. In the backbone network, a dual-path gated attention unit enhances the extracted features to obtain multi-scale feature maps. An optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layers of the backbone network and inputs them along with other multi-scale feature maps into the neck network. The detection unit crops the region of interest (ROI) features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the ROI features with category and bounding box regression for prediction, obtaining the prediction result. The dual-path gated attention unit is composed of a dual-path adaptive channel attention submodule and an enhanced spatial attention submodule. The submodule extracts channel statistics through parallel average and extreme paths, and introduces a gating unit to learn fusion coefficients to dynamically adjust the contribution of the two paths. The enhanced spatial attention submodule extracts two types of spatial cues in the channel dimension and concatenates these two types of spatial cues. The optimized Mamba unit includes an adaptive window partitioning unit, a feature processing unit, and a window re-integration unit. The adaptive window partitioning unit is used to partition the input features into multiple sub-window features after non-overlapping partitioning. The multiple sub-window features are input to the feature processing unit for processing. The feature processing unit has a 6-layer structure, with the first 3 layers being SARSSBlock blocks and the last 3 layers being self-attention blocks. The outputs of the SARSSBlock blocks and self-attention blocks are connected to a multilayer perceptron including layer normalization. The window re-integration unit re-concatenates the processed sub-window features output by the feature processing unit into a complete feature map to obtain the enhanced feature map. By employing KL divergence loss and classification loss, the SAR rotating box target detection network is trained based on the prediction results and the corresponding ground truth labels, resulting in a trained SAR rotating box target detection network. Acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. Input the SAR image and multiple noisy rotated bounding boxes into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

2. The SAR rotating frame target detection method according to claim 1, characterized in that, The backbone network uses ResNet50. Before the dual-path gated attention unit is embedded in the residual connection of each Bottleneck block, the features output by the BN layer are enhanced.

3. The SAR rotating frame target detection method according to claim 2, characterized in that, In the feature processing unit: The SARSSBlock block includes a first branch and a second branch in parallel. The first branch includes a linear layer, a convolutional layer, a SiLU activation function, and a SARScan2D block connected in sequence. The second branch includes a linear layer, a convolutional layer, and a SiLU activation function connected in sequence. The output data of the two branches are concatenated and then passed through a linear layer to obtain the output of the SARSSBlock block. The self-attention block employs a multi-head self-attention mechanism to map input features to multiple independent subspaces to calculate single-head self-attention. After concatenating the calculation results of all single-head self-attention, a projection matrix is ​​fused to obtain the output of the self-attention block.

4. The SAR rotating frame target detection method according to claim 3, characterized in that, In the SARScan2D block, the four-way scanning mechanism expands each of the sub-window features into a one-dimensional sequence, and after processing by the corresponding S6Block block, it is reversed back to a two-dimensional sub-window feature.

5. The SAR rotating frame target detection method according to any one of claims 1-4, characterized in that, The detection unit comprises multiple detection heads connected in series; The first detection head takes the multi-scale feature fusion map and the multiple noisy rotated bounding boxes as input to obtain the category prediction result and the rotated box prediction result. Each subsequent detection head uses the rotation box prediction result output by the previous detection head and the multi-scale feature fusion map as input to repeat the feature interaction and prediction process; The category prediction result and the rotation box prediction result output by the last detection head are used as the output of the detection unit.

6. The SAR rotating frame target detection method according to claim 5, characterized in that, In each of the aforementioned detection heads: Based on the multi-scale feature fusion map and the prediction results of the multiple noisy rotated bounding boxes or rotated boxes, the region of interest features with different levels of matching are obtained through the rotation region of interest alignment operation. Multi-dimensional feature interaction operations and normalization regularization processing are performed on the features of the region of interest to obtain the interactive features. The interactive features are then processed according to the dynamic interaction mechanism between instances and normalization regularization to obtain the dynamic interactive features with optimized instance information association. Based on the dynamic interaction features, a nonlinear transformation operation is performed on the feedforward network to obtain a nonlinear transformation feature with enhanced feature representation capability. The nonlinear transformation feature is then combined with the temporal embedding information of the diffusion process to perform a temporal feature fusion and adjustment operation, resulting in an adjusted feature with fused temporal information and optimized parameters. Based on the adjusted features, the category prediction result and the rotated bounding box prediction result are obtained through classification branch and regression branch processing, respectively.

7. A SAR rotating frame target detection device, characterized in that, The apparatus implements the SAR rotating frame target detection method according to any one of claims 1-6, and the apparatus comprises: The SAR training set acquisition module is used to acquire a SAR training set, which includes multiple SAR sample images and corresponding ground truth labels. The ground truth labels include ground truth detection boxes, which are rotation detection boxes corresponding to the target pose. The forward diffusion noise module is used to add Gaussian noise to the ground truth detection box of the SAR sample image by controlling the variance scheduling and utilizing the forward diffusion in the diffusion model to obtain multiple noisy rotated bounding boxes. The neural network prediction module is used to input the SAR sample image and the corresponding multiple noisy rotated bounding boxes into the SAR rotated bounding box target detection network for rotation detection box prediction. The SAR rotated bounding box target detection network includes a backbone network, a neck network, and a detection unit. In the backbone network, the extracted features are enhanced using a dual-path gating attention unit to obtain a multi-scale feature map. The optimized Mamba unit processes two adjacent scale feature maps output from the intermediate layer of the backbone network and inputs them into the neck network along with other multi-scale feature maps. The detection unit crops the region of interest features corresponding to each noisy rotated bounding box from the multi-scale feature fusion map output from the neck network, and then combines the region of interest features to perform category and bounding box regression for prediction to obtain the prediction result. The neural network training module is used to train the SAR rotating box target detection network by using KL divergence loss and classification loss, based on the prediction results and the corresponding ground truth labels, to obtain the trained SAR rotating box target detection network. The SAR target detection module is used to acquire the SAR image to be detected and randomly generate multiple noisy rotated bounding boxes. The SAR image and multiple noisy rotated bounding boxes are input into the trained SAR rotated bounding box target detection network to perform target detection and obtain the target rotated bounding box and target category.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images