Lightweight target detection transformer model, method and application based on space-frequency domain joint modeling

Abstract

The application belongs to the technical field of computer vision and remote sensing image processing, and particularly relates to a lightweight Transform model target detection device, method and application based on space-frequency domain joint modeling, which solves the problems of high calculation complexity and difficulty in deployment on edge devices of an existing detection model in a complex SAR scene. The model first extracts basic features through a cascaded convolution block. The core lies in the introduction of a cross-stage spectral convolution mixer module. The module maps features to the frequency domain through fast Fourier transform, reconstructs back to the spatial domain after selective enhancement and noise suppression, realizes joint modeling of the space-frequency domain, and effectively improves the distinguishability of small targets under strong interference. Meanwhile, the model replaces the traditional down-sampling with an anchor-free down-sampling module, and through a multi-scale feature fusion mechanism, the resolution is reduced while the details of small targets are maximally retained.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing image processing technology, specifically relating to a lightweight Transformer model target detection device, method and application based on spatial-frequency domain joint modeling. Background Technology

[0002] SAR imagery, with its all-weather, all-day imaging advantages, plays an irreplaceable role in military reconnaissance, marine monitoring, and other fields. SAR image target detection technology has always been a key research focus in remote sensing. Traditional constant false alarm rate (CFAR) detection algorithms rely on manually designed statistical features, which lack generalization ability in complex backgrounds and multi-target scenarios. In recent years, the introduction of deep learning technology has significantly improved the performance of SAR target detection. Researchers have mainly explored multi-scale feature extraction and attention mechanisms. For example, Faster R-CNN achieves accurate localization through a region proposal network, the YOLO series algorithms improve detection speed through anchor box mechanisms, and Transformer-based detectors such as DETR enhance global feature modeling capabilities through self-attention mechanisms. Furthermore, the application of various attention modules has effectively improved the model's ability to focus on key features.

[0003] However, strong speckle noise and complex background interference in SAR images severely impact detection performance. High-frequency noise often obscures the subtle features of small targets, making accurate model identification difficult. Simultaneously, significant scale differences exist among targets in SAR images; small targets such as ships and vehicles are susceptible to loss of crucial details during feature extraction due to downsampling operations. Furthermore, existing detection models generally suffer from large parameter counts and high computational complexity, making real-time detection difficult on resource-constrained edge devices such as satellites and drones. These issues collectively limit the deployment effectiveness of SAR target detection algorithms in practical applications, necessitating the development of lightweight solutions that balance accuracy and efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a lightweight Transformer model target detection device, method and application based on space-frequency domain joint modeling, so as to solve the problems of low recognition accuracy, high false negative rate of small targets and large model computational complexity in existing SAR target detection methods in complex backgrounds, which makes them difficult to deploy on edge devices.

[0005] The present invention achieves the above objectives through the following technical solutions: Firstly, this invention proposes a lightweight Transformer model target detection device based on joint spatial-frequency domain modeling, comprising: The cross-stage spectral convolution mixer module is used to perform spatial-frequency joint modeling on the feature map of the input SAR image. It transforms the features to the frequency domain through fast Fourier transform for selective enhancement and noise suppression, and then reconstructs them to the spatial domain through inverse Fourier transform to generate the first enhanced feature map. An anchorless downsampling module is used to perform multi-scale feature fusion downsampling on the first enhanced feature map. After average pooling, it is split into two branches along the channel dimension, and after performing different operations on each branch, they are spliced ​​together to form the second downsampled feature map. The Transformer model decoding end is used to determine the target category and bounding box prediction result based on the second downsampled feature map; The Transformer model weights are structured and compressed using a hierarchical adaptive magnitude pruning algorithm. By calculating the global relative importance score of the weight connections and removing the connection with the lowest score, the optimal sparsity distribution of each layer is automatically determined while satisfying the preset sparsity constraints.

[0006] Furthermore, the cross-stage spectral convolution mixer module includes: The frequency domain transformation unit is used to convert the feature map of the spatial input SAR image into frequency domain features through fast Fourier transform and separate them into two components: real and imaginary parts. The frequency domain filtering unit is used to modulate the real and imaginary components of the spliced ​​frequency domain; The spatial reconstruction unit is used to reconstruct the processed frequency domain features back to the spatial domain through inverse Fourier transform, and then fuse them with the original spatial features through residual connection to obtain the first enhanced feature map.

[0007] Furthermore, in the frequency domain transformation unit, the fast Fourier transform is a discrete Fourier transform, used to map image features from the spatial domain to the frequency domain.

[0008] Furthermore, in the spatial reconstruction unit, after the spatial features reconstructed by inverse Fourier transform are fused with the original spatial features, they are also subjected to channel compression through pointwise convolution.

[0009] Furthermore, the anchorless downsampling module includes: The pooling and splitting unit is used to perform average pooling on the features after downsampling the first enhanced feature map and split them into left branch features and right branch features along the channel dimension. The left branch processing unit is used to perform a 3×3 convolution operation on the left branch features; The right branch processing unit is used to perform max pooling and 1×1 convolution operations on the right branch features sequentially. The output concatenation unit is used to concatenate the outputs of the left branch processing unit and the right branch processing unit along the channel dimension to obtain the second downsampled feature map.

[0010] Furthermore, in the right branch processing unit, the max pooling operation and the 1×1 convolution operation are executed in series, and no nonlinear activation function is introduced.

[0011] Furthermore, the global relative importance score is a LAMP score, which is calculated as the ratio of the squared weight magnitude of the target connection to the cumulative sum of the squared weight magnitudes of all currently alive connections in the layer to which the target connection belongs.

[0012] Furthermore, when determining the target category and bounding box prediction results, the Transformer model decoding end adopts an intersection-union ratio (IUAR)-aware query selection mechanism, dynamically selecting the decoding query based on the IUAR of the predicted box and the ground truth box.

[0013] Secondly, this invention proposes an image detection method performed by a lightweight Transformer model target detection device based on space-frequency domain joint modeling as described above, the method comprising: The feature map of the input SAR image is input into the cross-stage spectral convolution mixer module. The feature map is transformed to the frequency domain by fast Fourier transform for selective enhancement and noise suppression. Then, it is reconstructed to the spatial domain by inverse Fourier transform to generate the first enhanced feature map. The first enhanced feature map is input into the anchorless downsampling module. After average pooling, it is split into two branches along the channel dimension. Different operations are performed on each branch, and then they are spliced ​​together to form the second downsampling feature map. The second downsampled feature map is input to the decoding end of the Transformer model to determine the target category and bounding box prediction result.

[0014] Thirdly, this invention proposes an application of the lightweight Transformer model target detection device based on space-frequency domain joint modeling, as described above, for real-time SAR target detection on an edge computing platform. The edge computing platform includes a satellite or UAV platform, and the application includes: The lightweight target detection Transformer model is deployed on the edge computing device of the satellite or drone platform; The SAR image to be detected is acquired by the SAR sensor mounted on the platform and input into the system; The system processes the input SAR image and outputs in real time the category information and bounding box coordinate information of at least one target among ships, vehicles and aircraft in the image.

[0015] The beneficial effects of this invention are as follows: The lightweight target detection Transformer model for SAR images provided by this invention has significant technical advantages and practical value. This model achieves joint feature modeling in the spatial and frequency domains through a cross-stage spectral convolution mixer module, effectively suppressing high-frequency noise and background interference in SAR images and significantly improving the discriminability of small targets in strong clutter environments. The anchor-free downsampling module, through a multi-branch feature fusion mechanism, maximizes the preservation of detailed information of small targets while reducing feature map resolution, effectively solving the problem of small target feature loss caused by traditional downsampling methods. The hierarchical adaptive pruning unit significantly reduces computational complexity while maintaining model detection accuracy by automatically determining the optimal sparsity distribution of each layer. The entire model, through its lightweight design, enables efficient real-time target detection on platforms with limited computing resources, providing a feasible technical solution for the practical application of SAR images in military reconnaissance, marine monitoring, and other fields. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the lightweight image detection Transformer model proposed in this invention; Figure 2 This is a schematic diagram of the structure of the CSSCM module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the FFCM module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the ADown module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the LiteSAR-DETR model in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the application process of the Transformer model for real-time image detection in SAR image target detection according to the present invention. Figure 7 This is a comparison chart of ship target detection results in an embodiment of the present invention; Figure 7 (a) in the diagram represents the actual area of ​​the ship target. Figure 7 (b) in the image shows the detection results of LiteSAR-DETR. Figure 7 (c) in the table represents the detection results of JRVSNet. Figure 7 (d) in the figure represents the detection result of MU-Net. Figure 7 (e) in the figure represents the detection result of CFCFNet. Figure 7 (f) in the figure represents the detection results of DMSA-Net. Figure 7 In the table, (g) represents the detection result of MSSFC-Net. Figure 7 (h) in the table represents the detection results from SPCNet. Figure 7 In the diagram, (i) represents the detection result of S2FCNet; Figure 8 This is a comparison chart of aircraft target detection results in an embodiment of the present invention; Figure 8 In the diagram, (a) represents the actual area of ​​the aircraft target. Figure 8 (b) in the image shows the detection results of LiteSAR-DETR. Figure 8 (c) in the table represents the detection results of JRVSNet. Figure 8 (d) in the figure represents the detection result of MU-Net. Figure 8 (e) in the figure represents the detection result of CFCFNet. Figure 8 (f) in the figure represents the detection results of DMSA-Net. Figure 8 In the table, (g) represents the detection result of MSSFC-Net. Figure 8 (h) in the table represents the detection results from SPCNet. Figure 8 In the diagram, (i) represents the detection result of S2FCNet; Figure 9 This is a comparison chart of vehicle target detection results in an embodiment of the present invention; Figure 9 In the diagram, (a) represents the actual area of ​​the car target. Figure 9 (b) in the image shows the detection results of LiteSAR-DETR. Figure 9 (c) in the table represents the detection results of JRVSNet. Figure 9 (d) in the figure represents the detection result of MU-Net. Figure 9 (e) in the figure represents the detection result of CFCFNet. Figure 9 (f) in the figure represents the detection results of DMSA-Net. Figure 9 In the table, (g) represents the detection result of MSSFC-Net. Figure 9 (h) in the table represents the detection results from SPCNet. Figure 9 In the diagram, (i) represents the detection result of S2FCNet. Detailed Implementation

[0017] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0018] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0019] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0020] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0021] This invention designs a lightweight target detection model for SAR images, aiming to significantly reduce computational complexity while maintaining high detection accuracy to meet the deployment requirements of edge devices. The model is named the Lightweight SAR Detection Transformer (LiteSAR-DETR). It achieves joint spatial-frequency domain modeling through a Cross-Stage Spectral Convolution Mixer (CSSCM) module, effectively suppressing background noise and enhancing the discriminability of small target features in the frequency domain. Anchor-free downsampling (ADown) module employs a multi-scale feature fusion mechanism, effectively preserving detailed information and semantic features of small targets during downsampling. Layer-adaptive Magnitude-based Pruning (LAMP) algorithm automatically determines the optimal sparsity distribution for each layer, achieving efficient structured compression of the model. Ultimately, this model significantly reduces the number of parameters while maintaining excellent detection performance in complex SAR scenarios, providing a feasible technical solution for resource-constrained platforms.

[0022] Example 1

[0023] Please see Figure 1-5 This invention proposes a lightweight Transformer model target detection device based on joint space-frequency domain modeling, comprising: The Cross-Stage Spectral Convolutional Mixer (CSSCM) module is used to perform spatial-frequency joint modeling on the feature map of the input SAR image. It transforms the features to the frequency domain through Fast Fourier Transform for selective enhancement and noise suppression, and then reconstructs them to the spatial domain through Inverse Fourier Transform to generate the first enhanced feature map.

[0024] The anchorless downsampling module (ADown module) is used to perform multi-scale feature fusion downsampling on the first enhanced feature map. After average pooling, it is split into two branches along the channel dimension, and 3×3 convolution and max pooling followed by 1×1 convolution are performed on the two branches respectively. The outputs of the two branches are concatenated to form the second downsampled feature map.

[0025] The Transformer model decoding end is used to determine the target category and bounding box prediction results based on the second downsampled feature map.

[0026] The Layered Adaptive Amplitude Pruning (LAMP) algorithm is used to perform structured compression of model weights. It calculates the LAMP score of each connection in the weight tensor and performs global pruning based on the score to automatically determine the optimal sparsity distribution of each layer.

[0027] Understandably, in the cross-stage spectral convolution mixer module, the FFCM submodule achieves the fusion of local and global features by performing multi-scale deep convolution in the spatial domain and global feature extraction in the frequency domain in parallel. This module uses learnable 1×1 convolutional layers in the frequency domain to adaptively modulate the real and imaginary components of the frequency domain features, effectively suppressing high-frequency noise in SAR images while enhancing the scattering characteristics of small targets.

[0028] More specifically, the Cross-Stage Spectral Convolutional Mixer (CSSCM) module achieves feature enhancement through a joint spatial-frequency modeling mechanism. This module further includes: The frequency domain transformation unit is used to convert the spatial feature map into frequency domain features through Fast Fourier Transform, separating them into real and imaginary components; The frequency domain filtering unit is used to modulate the real and imaginary components of the spliced ​​frequency domain, and selective enhancement is achieved by using a 1×1 convolutional layer. The spatial reconstruction unit is used to reconstruct the processed frequency domain features back to the spatial domain through inverse Fourier transform, and then fuse them with the original spatial features through residual connection.

[0029] In the specific implementation of the scheme, the working process of the cross-stage spectral convolution mixer module is as follows: In spatial domain operations, this method first uses pointwise convolution on the input features. Dimensionality enhancement is performed, and the data is split into two groups to extract multi-scale local features, ultimately yielding the output. This process can be represented as: ; in, This indicates depthwise convolution using two different kernel sizes, 3×3 and 5×5, in the spatial domain; Pointwise convolution; express Activation function.

[0030] In frequency domain operations, spatial features The data is converted into real and imaginary components using DFT. The concatenated components are then processed by a 1×1 convolutional layer, and the real and imaginary parts are separated by modulation operations. Finally, the features are reconstructed into the spatial domain using inverse DFT. ; ; ; Finally, the output of the FFCM module is processed by the residual structure and then formed by a pointwise convolution channel compression operation: ; Therefore, the output of the CSSCM module can be represented as: ; ; ; ; ; After processing by the CSSCM module, the first feature enhancement map is obtained.

[0031] More specifically, the anchorless downsampling module (ADown module) further includes: In the ADown module, input features First, the features are processed using average pooling. The processed features are then divided into two equal parts along the channel dimension. The left branch performs only a 3×3 convolution, while the right branch first performs max pooling followed by a 1×1 convolution. After processing, the outputs of the left and right branches are concatenated to form the final output of the module. Therefore, the output of the ADown module can be represented as: ; ; ; ; ; After processing by the ADown module, the second downsampled feature map is obtained.

[0032] More specifically, the hierarchical adaptive magnitude pruning algorithm (LAMP algorithm) includes: Consider a depth of In a feedforward neural network, the weight tensors corresponding to its fully connected layers or convolutional layers are respectively... However, the weights of fully connected layers are two-dimensional matrices, while the weights of convolutional layers are four-dimensional tensors. This difference in dimensionality hinders the definition of the LAMP score. To address this issue, the weight tensors of all layers are flattened into one-dimensional vectors. After flattening, it is assumed that the weights are arranged in ascending order of the index mapping, i.e., when... When, satisfy ,in Indicates index The weights of the mapping.

[0033] At this point, the weight tensor Chinese index The corresponding LAMP score is defined as: (1) LAMP scores essentially measure the relative importance of a target connection among all surviving connections within its layer. This relativity stems from the fact that low-weight connections in that layer have been removed. When two connections have exactly the same weight values, their LAMP scores may vary depending on the index order. After calculating all LAMP scores, global pruning is performed: connections with the lowest scores are removed until a preset global sparsity constraint is met. This global pruning process is equivalent to amplitude pruning that automatically selects the layer sparsity. Verifying this conclusion only requires proving: (2) For any weight tensor and its index , This mathematical relationship always holds true. Based on equation (1), it can be deduced that equation (2) must hold true. This is because when the weight of a certain connection increases, the denominator of equation (1) decreases, while the numerator increases accordingly.

[0034] It is worth noting that the global pruning mechanism based on LAMP score and the weight magnitude-based pruning mechanism... The global pruning method differs fundamentally from traditional pruning. This difference stems from the fact that the LAMP score calculates the relative importance of surviving connections within a layer, rather than absolute weight magnitudes. Specifically, if there is exactly one connection in each layer that receives a LAMP score of 1, that score represents the highest possible score achievable by that layer.

[0035] In other words, the LAMP algorithm ensures that at least one live connection is maintained in each layer. This property is based on weight magnitude. This does not hold true in global pruning methods. Therefore, LAMP has a unique advantage in preserving critical connections within layers.

[0036] Hierarchical amplitude pruning is essentially a hierarchical optimal solution to the Frobenius norm distortion minimization problem. This optimal solution can be viewed as a relaxed form of minimizing the L2 norm distortion of the output for the worst-case input. This finding leads to a core hypothesis: reducing the L2 norm distortion induced by pruning the model output in the worst case may improve the performance of retrained models. This hypothesis also explains why amplitude pruning works well in practice.

[0037] Based on the above assumptions, the LAMP algorithm is designed as a natural extension of hierarchical amplitude pruning. Its core lies in automatically determining the hierarchical sparsity to achieve a global pruning scheme. To rigorously demonstrate the properties of LAMP, consider a depth of... A fully connected neural network, which is sensitive to input The output can be represented as: (3) in, Represents the ReLU activation function. Indicates the first The weight matrix of the layer, This represents the set of all weight matrices.

[0038] First, we focus on a single fully connected layer, investigating how to minimize the L2 norm distortion of the layer's output caused by pruning under a given worst-case input signal. Theoretical analysis shows that this problem can be relaxed into minimizing the Frobenius norm distortion of the weight tensor, and its solution is precisely equivalent to layer amplitude pruning. Let... The input vector of the fully connected layer. Its weight tensor. The pruning goal is to reduce the original weight tensor to its weight tensor. Transform into sparse tensors ,in It satisfies the global sparsity constraint. A binary mask matrix.

[0039] The current optimization objective is to find a pruning mask that minimizes the worst-case L2 norm bounded input that causes L2 norm distortion in the output. ,Right now: (4) Equation (4) provides an upper bound for the minimum expected L2 norm distortion corresponding to any unit sphere input distribution. Therefore, this maximum-minimum distortion can be regarded as the design basis for a data-independent loss minimization pruning scheme, where the loss function takes a squared form. According to the definition of the spectral norm, Equation (4) is equivalent to: (5) in Let represent the spectral norm. For any matrix ... All satisfy ( (representing the Frobenius norm), equation (5) can be relaxed to the problem of minimizing Frobenius norm distortion: (6) in , They represent the weight tensors respectively. With mask matrix exist The element of the position. From the expression on the right side of equation (6), we can see that: layered amplitude pruning, i.e. Time Retention forward big The location is precisely the optimal solution for minimizing the Frobenius norm distortion induced by pruning. This finding establishes hierarchical amplitude pruning as an approximate optimal solution for minimizing output L2 norm distortion, and also reveals a potential correlation between smaller output L2 norm distortion and improved subnetwork performance after pruning and retraining.

[0040] Therefore, a core question arises: how to choose the sparsity of the pruning amplitude at each layer to minimize model-level output distortion? This problem can be formally defined as the following minimization problem: (7) in This represents the model-level sparsity constraints specified by operational requirements. Indicates the first The weight matrix after layer pruning.

[0041] The nonlinear nature of the activation function makes equation (7) difficult to solve precisely. To address this problem, the following greedy algorithm is used for an approximate solution: (1) Estimate the distortion of the model output caused by removing each individual connection in the network; (2) Remove the connection with the lowest current rating (i.e., distortion); (3) Based on the pruned model, repeat steps (1) and (2) to calculate the score and remove the connection until the preset pruning termination condition is met.

[0042] Assuming that only one connection is pruned in each iteration, the above problem can be relaxed by outputting an upper bound on the distortion using the following model. Let Indicates the first The weight matrix after layer pruning is as follows: (8) The right side of equation (8) has two key advantages: (1) The expression is completely independent of the activation function and is mathematically equivalent to hierarchical amplitude pruning; (2) The score can be calculated in advance, without having to be calculated repeatedly after each pruning.

[0043] In particular, the multiplicative terms It does not affect the pruning decision. For the weight tensor... Each index Its denominator can use cumulative sum. Pre-calculation. It is this calculation technique that ultimately yields the LAMP score in equation (1).

[0044] Example 2

[0045] Please see Figure 6 This embodiment proposes an image detection method performed by a lightweight Transformer model target detection device based on spatial-frequency domain joint modeling, as described in Embodiment 1. The method includes: The feature map of the input SAR image is input into the cross-stage spectral convolution mixer module. The feature map is transformed to the frequency domain by fast Fourier transform for selective enhancement and noise suppression. Then, it is reconstructed to the spatial domain by inverse Fourier transform to generate the first enhanced feature map. The first enhanced feature map is input into the anchorless downsampling module. After average pooling, it is split into two branches along the channel dimension. Different operations are performed on each branch, and then they are spliced ​​together to form the second downsampling feature map. The second downsampled feature map is input to the decoding end of the Transformer model to determine the target category and bounding box prediction result.

[0046] After completing the above module design, a lightweight detection Transformer model for SAR images can be obtained, and target image detection can be performed.

[0047] In this embodiment, a neural network model is built using the PyTorch framework. Figure 5 A schematic diagram of the LiteSAR-DETR model network architecture in an embodiment of the present invention is shown. Figure 5 As shown, the LiteSAR-DETR model takes a SAR image as input. First, it uses cascaded 3×3 convolutional blocks to extract basic features. Then, a cross-stage spectral convolutional mixer (CSSCM) module enhances the spatial and frequency domain features. Next, an anchorless downsampling (ADown) module reduces the feature map resolution while preserving details of small targets. Then, a hierarchical adaptive amplitude pruning (LAMP) algorithm is applied to structurally compress the model, significantly reducing the number of parameters and computational complexity. Finally, an intersection-union-ratio (IU)-aware query selection mechanism is introduced at the decoding end to further optimize model efficiency while ensuring detection accuracy. Combining these modules, the model generates target prediction results through the decoding head.

[0048] Example 3

[0049] An application of a lightweight Transformer model target detection device based on space-frequency domain joint modeling, as described in Example 1, for real-time SAR target detection on an edge computing platform, wherein the edge computing platform includes a satellite or UAV platform, and the application includes: The lightweight target detection Transformer model is deployed on the edge computing device of the satellite or UAV platform; the SAR image to be detected is acquired by the SAR sensor on the platform and input into the system; the system processes the input SAR image and outputs in real time the category information and bounding box coordinate information of at least one target in the image, including ships, vehicles and aircraft.

[0050] The above scheme will be explained and analyzed in conjunction with a specific simulation experiment.

[0051] To compare the performance of the LiteSAR-DETR model, seven other models were built: JRVSNet, MU-Net, CFCFNet, DMSA-Net, MSSFC-Net, SPCNet, and S2FCNet.

[0052] Simulation Experiment 1 The model provided by this invention is used to identify three types of targets: ships, aircraft, and vehicles. The detection performance of LiteSAR-DETR is verified by comparing the detection results of different models.

[0053] Figure 7 The diagram shows the results of a comparative test of ship targets in an embodiment of the present invention. In a complex near-shore scenario involving multiple ship targets, Figure 7The LiteSAR-DETR algorithm shown in (b) demonstrates superior detection performance, accurately detecting all ship targets, while other algorithms suffer from missed detections and false detections. LiteSAR-DETR's accurate detection is attributed to the multi-scale feature fusion mechanism of the ADown module, which successfully preserves the contour details of small-sized ships, avoiding the feature loss problem caused by traditional downsampling methods. Compared with other models, LiteSAR-DETR not only has fewer parameters and lower computational cost, but also superior detection accuracy, with no missed detections or false detections, fully validating the effectiveness of its lightweight design.

[0054] Figure 8 The diagram shows the results of a comparative experiment on aircraft targets in an embodiment of the present invention. Only LiteSAR-DETR accurately identified all targets in the aircraft target detection task, while all other models failed to detect any aircraft targets. This is mainly due to the frequency-domain selective enhancement mechanism of the CSSCM module, which strengthens the representation ability of key scattering features such as aircraft wingtips and overcomes the problem of difficult object recognition in low-contrast scenes. These improvements fully demonstrate LiteSAR-DETR's accurate detection capability for small targets in complex backgrounds.

[0055] Figure 9 The accompanying diagram shows the comparative test results of vehicle targets in an embodiment of the present invention. In low-contrast scenes, LiteSAR-DETR successfully detected all four vehicle targets, achieving an average detection probability of 0.87, significantly outperforming other comparative models. Despite having fewer parameters than JRVS-Net, this model maintains leading detection accuracy, primarily due to the intelligent preservation strategy of key connections during model compression in the LAMP algorithm. The comparative results demonstrate that through the synergistic optimization of noise suppression in the CSSCM module, detail preservation in the ADown module, and efficient compression in the LAMP algorithm, LiteSAR-DETR achieves stable detection of low-contrast targets while maintaining a lightweight design, providing reliable technical support for the practical application of SAR images on edge devices.

[0056] As shown in Table 1, LiteSAR-DETR demonstrated comprehensive advantages in global performance comparisons. It achieved optimal levels in all key evaluation metrics: AP (55.8%), AP... 50 (87.4%) and AP 75 (60.2%) is the highest value among all comparative models.

[0057] In terms of model efficiency, LiteSAR-DETR also demonstrates outstanding performance. Its parameter count (25.8M) and computational cost (75.3G) are the lowest among all models, while its inference speed (175FPS) is the fastest, fully reflecting the advantages of its lightweight design. This result shows that through the synergistic effect of frequency domain feature enhancement from the CSSCM module, detail-preserving downsampling from the ADown module, and structured compression from the LAMP algorithm, the model achieves extreme efficiency optimization while maintaining high accuracy.

[0058] While some models performed similarly on individual metrics, LiteSAR-DETR achieved the best results across all six core metrics, demonstrating a comprehensive performance advantage with no weaknesses. This overall leading performance proves that LiteSAR-DETR not only excels in detection accuracy but also achieves breakthroughs in model complexity and inference efficiency, providing an ideal technical solution for real-time detection of SAR images on edge devices.

[0059] Table 1

[0060] By comparing with other state-of-the-art models, the LiteSAR-DETR model demonstrated the best overall performance across various SAR target detection tasks. This indicates that LiteSAR-DETR exhibits superior prediction accuracy in detecting different types of targets, further validating the model's reliability and stability.

[0061] In summary, the LiteSAR-DETR model proposed in this invention significantly improves the model's multi-scale recognition capability for SAR images by employing a CSSCM module for joint spatial-frequency domain modeling to enhance target features and suppress background noise, a multi-scale feature fusion mechanism in the ADown module to preserve small target details during downsampling, and an LAMP algorithm for efficient model compression through adaptive sparsity allocation. This is achieved through a comprehensive comparison with seven advanced models—JRVSNet, MU-Net, CFCFNet, DMSA-Net, MSSFC-Net, SPCNet, and S2FCNet8—demonstrating its superior performance in detection accuracy and inference efficiency. Compared to other deep learning models, LiteSAR-DETR significantly reduces computational resource requirements while maintaining extremely high detection accuracy, effectively overcoming the technical challenges of high false negative and false positive rates for multiple targets in SAR images. It provides an innovative lightweight technical solution for all-weather, all-time remote sensing target detection in resource-constrained environments.

[0062] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0063] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0064] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A lightweight target detection device based on spatial-frequency domain joint modeling using a Transformer model, characterized in that, include: A backbone network for extracting multi-scale features, the backbone network comprising: a feature extraction architecture consisting of a cross-stage spectral convolutional mixer module and an anchorless downsampling module alternately set up; The cascaded configuration of the feature extraction architecture is as follows: a first cross-stage spectral convolution mixer module, a first anchorless downsampling module, a second cross-stage spectral convolution mixer module, a second anchorless downsampling module, and a third cross-stage spectral convolution mixer module are connected in sequence. The cross-stage spectral convolution mixer module is used to perform spatial-frequency joint modeling on the feature map of the input SAR image. It uses pointwise convolution to increase the dimensionality of the input features and split them into two groups. It uses depth convolution with two different kernel sizes, 3×3 and 5×5, to extract multi-scale local features. After activation function and pointwise convolution processing, the features are concatenated to obtain spatial features. The spatial features are transformed to the frequency domain by fast Fourier transform for selective enhancement and noise suppression. Then, they are reconstructed to the spatial domain by inverse Fourier transform to generate the first enhanced feature map corresponding to the first cross-stage spectral convolution mixer module, the second cross-stage spectral convolution mixer module, and the third cross-stage spectral convolution mixer module, respectively. The anchorless downsampling module is used to perform multi-scale feature fusion downsampling on the first enhanced feature map. After average pooling, it is split into two branches along the channel dimension, and after performing different operations on each branch, they are spliced ​​together to form the second downsampled feature map. The Transformer model decoding end is used to determine the target category and bounding box prediction result based on the second downsampled feature map; The Transformer model weights are structured and compressed using a hierarchical adaptive magnitude pruning algorithm. By calculating the global relative importance score of the weight connections and removing the connection with the lowest score, the optimal sparsity distribution of each layer is automatically determined while satisfying the preset sparsity constraints.

2. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 1, characterized in that, The cross-stage spectral convolution mixer module includes: The frequency domain transformation unit is used to convert the feature map of the spatial input SAR image into frequency domain features through fast Fourier transform and separate them into two components: real and imaginary parts. The frequency domain filtering unit is used to modulate the real and imaginary components of the spliced ​​frequency domain; The spatial reconstruction unit is used to reconstruct the processed frequency domain features back to the spatial domain through inverse Fourier transform, and then fuse them with the original spatial features through residual connection to obtain the first enhanced feature map.

3. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 2, characterized in that, In the frequency domain transformation unit, the Fast Fourier Transform is a Discrete Fourier Transform, used to map image features from the spatial domain to the frequency domain.

4. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 2, characterized in that, In the spatial reconstruction unit, the spatial features reconstructed by inverse Fourier transform are fused with the original spatial features, and then channel compression is performed through pointwise convolution.

5. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 1, characterized in that, The anchorless downsampling module includes: The pooling and splitting unit is used to perform average pooling on the features after downsampling the first enhanced feature map and split them into left branch features and right branch features along the channel dimension. The left branch processing unit is used to perform a 3×3 convolution operation on the left branch features; The right branch processing unit is used to perform max pooling and 1×1 convolution operations on the right branch features sequentially. The output concatenation unit is used to concatenate the outputs of the left branch processing unit and the right branch processing unit along the channel dimension to obtain the second downsampled feature map.

6. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 5, characterized in that, In the right branch processing unit, the max pooling operation and the 1×1 convolution operation are executed in series, and no non-linear activation function is introduced.

7. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 1, characterized in that, The global relative importance score is the LAMP score, which is calculated as the ratio of the squared weight magnitude of the target connection to the cumulative sum of the squared weight magnitudes of all currently alive connections in the layer to which the target connection belongs.

8. The lightweight Transformer model target detection device based on spatial-frequency domain joint modeling according to claim 1, characterized in that, When determining the target category and bounding box prediction results, the Transformer model decoder uses an intersection-union-ratio (IUU)-aware query selection mechanism to dynamically select the decoding query based on the IUU of the predicted bounding box and the ground truth bounding box.

9. An image detection method performed by a lightweight Transformer model target detection device based on spatial-frequency domain joint modeling as described in any one of claims 1-8, characterized in that, The method includes: Multi-scale features are extracted through a backbone network, which includes a feature extraction architecture consisting of a cross-stage spectral convolution mixer module and an anchorless downsampling module that are alternately set up. The cascaded configuration of the feature extraction architecture is as follows: a first cross-stage spectral convolution mixer module, a first anchorless downsampling module, a second cross-stage spectral convolution mixer module, a second anchorless downsampling module, and a third cross-stage spectral convolution mixer module are connected in sequence. The feature map of the input SAR image is input to the cross-stage spectral convolution mixer module. Pointwise convolution is used to increase the dimensionality of the input features and split them into two groups. Multi-scale local features are extracted by depth convolution with two different kernel sizes, 3×3 and 5×5. After activation function and pointwise convolution processing, the features are concatenated to obtain spatial features. The spatial features are transformed to the frequency domain by fast Fourier transform for selective enhancement and noise suppression. Then, they are reconstructed to the spatial domain by inverse Fourier transform, generating the first enhanced feature map corresponding to the first cross-stage spectral convolution mixer module, the second cross-stage spectral convolution mixer module, and the third cross-stage spectral convolution mixer module, respectively. The first enhanced feature map is input into the anchorless downsampling module. After average pooling, it is split into two branches along the channel dimension. Different operations are performed on each branch, and then they are spliced ​​together to form the second downsampling feature map. The second downsampled feature map is input to the decoding end of the Transformer model to determine the target category and bounding box prediction result.

10. An application of a lightweight Transformer model target detection device based on spatial-frequency domain joint modeling as described in any one of claims 1-8 for real-time SAR target detection on an edge computing platform, characterized in that, The edge computing platform includes a satellite or drone platform, and the applications include: The lightweight target detection Transformer model is deployed on the edge computing device of the satellite or drone platform; The SAR image to be detected is acquired by the SAR sensor mounted on the platform and input to the device; The device is used to process the input SAR image and output in real time the category information and bounding box coordinate information of at least one target among ships, vehicles and aircraft in the image.

Images