Efficient sar airplane detection method and system based on multi-branch cooperative calibration and frequency domain compensation

Abstract

The present application relates to synthetic aperture radar target detection technical field, especially relates to a kind of based on multi-branch collaborative calibration and frequency domain compensation efficient SAR aircraft detection method and system, construct the neural network model of multi-branch collaborative calibration and frequency domain compensation, backbone network in network model adopts hierarchical progressive structure, and each level is integrated with parallel aggregation and calibration mechanism, to parallel processing input feature and carrying out feature calibration, neck network utilizes multi-scale aggregation enhancement mechanism to fuse multi-scale feature and utilizes efficient discriminant frequency domain feedforward network to optimize multi-scale feature in frequency domain information, to in multi-scale feature fusion in spatial domain and frequency domain to feature collaborative optimization;And using training set to train neural network model, using validation set and test set to test and verify neural network model;The SAR image to be detected is input to target detection model, to obtain target category.The present application can realize efficient and accurate SAR aircraft detection under resource limited scene.

Description

Technical Field

[0001] This invention relates to the field of synthetic aperture radar target detection technology, and in particular to a highly efficient SAR aircraft detection method and system based on multi-branch collaborative calibration and frequency domain compensation. Background Technology

[0002] Current SAR aircraft detection methods mostly rely on feature extraction from a single branch or scale, making it difficult to coordinate the modeling of local details and global structural information of the target. This makes them unable to meet the needs of detecting aircraft targets with low contrast and diverse shapes in SAR images. Attention mechanisms are mostly based on first-order statistical features, which are not sensitive enough to weak signals and target boundaries, resulting in missed detections or positioning errors in low signal-to-noise ratio scenarios. Lightweight designs are mostly limited to local module optimization, lacking overall architectural coordination, and it is difficult to balance detection accuracy and computational efficiency. Moreover, they generally neglect the use of video domain information, resulting in weak detail representation capabilities in low-contrast and blurred target scenarios, and the inability to enhance the distinction between the target and the background through frequency domain compensation. Summary of the Invention

[0003] The purpose of this invention is to address the problems of incomplete feature representation, insufficient weak signal perception, difficulty in balancing lightweight design and accuracy, and low target-background discrimination in existing SAR image aircraft detection. It provides an efficient SAR aircraft detection method and system based on multi-branch collaborative calibration and frequency domain compensation. Based on multi-branch collaborative calibration, attention mechanism optimization, lightweight overall architecture, and frequency domain compensation, it achieves efficient and accurate SAR aircraft detection in resource-constrained scenarios.

[0004] According to the design scheme provided by this invention, on the one hand, an efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation is provided, comprising:

[0005] The acquired SAR image sample data is preprocessed and divided into training set, validation set and test set according to a preset ratio, and the target categories are labeled.

[0006] A neural network model for multi-branch collaborative calibration and frequency domain compensation is constructed. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion.

[0007] The neural network model is trained using the training set, tested and validated using the validation set and test set, and the final target detection model is obtained based on the test and validation results.

[0008] The SAR image to be detected is input into the target detection model, and the target detection model is used to obtain the target category in the SAR image to be detected.

[0009] As part of the efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation of this invention, the acquired SAR image sample data is further preprocessed, including:

[0010] Channel expansion replicates the original single-channel image in the sample data into a three-channel image in the channel dimension, and generates a three-channel image matrix.

[0011] Image data of different resolutions are mapped to a specified pixel specification using a bilinear interpolation algorithm.

[0012] As a highly efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation of the present invention, it further utilizes a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration, including:

[0013] The number of input feature channels is expanded by point convolution, and the input features are split into two parts, A and B, along the channel dimension;

[0014] The features of part A are extracted using three deep convolutional branches while preserving the original features. The output features of the three deep convolutional branches are concatenated and processed using pointwise convolution and the TeLU activation function. In each deep convolutional branch, element-wise multiplication is used to enhance the interaction and fusion of features within the part.

[0015] The feature data after feature processing is multiplied element-wise with the features of part B to achieve cross-branch feature calibration.

[0016] As a highly efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation of the present invention, it further includes parallel processing of input features and feature calibration using a parallel aggregation and calibration mechanism, and also includes:

[0017] Feature optimization is performed using a rectangular channel attention mechanism, including the following optimizations:

[0018] The first moment of each channel of the feature is extracted by global average pooling, and the skewness is obtained by combining the third central moment. The first moment is used to describe the basic strength of the feature, and the skewness is used to describe the asymmetry in the feature distribution.

[0019] The first-order moments and skewnesses are concatenated and channel weights are generated using ID convolution and the Sigmoid activation function, so as to assign a weighting coefficient to each channel.

[0020] Channel weighted optimization is performed on a pairwise basis based on weighted coefficients.

[0021] As a highly efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation of this invention, it further utilizes a multi-scale aggregation enhancement mechanism to fuse multi-scale features and utilizes an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, including:

[0022] The input features are normalized, and the number of channels of the input features is expanded by point convolution, splitting the input features into two parts A and B along the channel dimension;

[0023] The features of part A are extracted using three deep convolutional branches while preserving the original features. The output features of the three deep convolutional branches are concatenated and processed using pointwise convolution and the TeLU activation function. In each deep convolutional branch, element-wise multiplication is used to enhance the interaction and fusion of features within the part.

[0024] The processed features are residually concatenated with the input features and then normalized again.

[0025] The re-normalized features expand the number of channels through pointwise convolution and extract spatial features using depthwise convolution. The depthwise convolution uses a group convolution mechanism to ensure that each channel learns spatial information independently.

[0026] The spatial features are split into two equal-channel branches, X1 and X2. X1 is activated by TeLU to enhance nonlinearity and then multiplied element-wise with X2. The number of channels is adjusted by pointwise convolution, and the features are divided into blocks by equal size.

[0027] Perform a two-dimensional fast Fourier transform on each block feature to transform the features to the frequency domain, and weight the frequency domain components using a learnable FFT parameter matrix.

[0028] The weighted features are transformed back to the spatial domain by inverse Fourier transform and then reassembled into blocks to restore the original feature size.

[0029] Perform residual connections with the features before renormalization and output the optimized features.

[0030] As an efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation of the present invention, the detection head based on fused features to predict the target category adopts an anchorless design to directly regress the target bounding box coordinates and category probabilities from the feature map. In addition, the detection head uses CIoU as the bounding box regression loss function during the training phase. The CIoU is constrained by the intersection-union ratio, the Euclidean distance of the bounding box center point and the consistency of the aspect ratio.

[0031] As a highly efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation in this invention, the bounding box regression loss function is further expressed as: Where IoU represents the crossover-union ratio, ρ(b,b) gt ) represents the predicted bounding box b and the ground truth bounding box b. gt The Euclidean distance between the center points, c represents the diagonal length of the minimum closure region, v is a parameter used to measure the aspect ratio difference, and α is a balance coefficient used to adjust the weight of the penalty term.

[0032] Furthermore, this invention also provides a high-efficiency SAR aircraft detection system based on multi-branch collaborative calibration and frequency domain compensation, comprising: a sample acquisition module, a network construction module, a network training module, and a target detection module, wherein...

[0033] The sample acquisition module is used to preprocess the acquired SAR image sample data, divide the SAR image sample data into training set, validation set and test set according to a preset ratio, and label the target categories.

[0034] A network construction module is used to construct a neural network model for multi-branch collaborative calibration and frequency domain compensation. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion.

[0035] The network training module is used to train the neural network model using the training set, test and validate the trained neural network model using the validation set and test set, and obtain the final target detection model based on the test and validation results.

[0036] The target detection module is used to input the SAR image to be detected into the target detection model and use the target detection model to obtain the target category in the SAR image to be detected.

[0037] The beneficial effects of this invention are:

[0038] 1. This invention integrates a multi-branch collaborative calibration mechanism into the SAR aircraft target detection network. Through feature interaction and calibration between multiple branches, it achieves efficient capture of local texture details and global morphological structure of aircraft targets, helping to improve the integrity of feature representation for targets with weak contrast and diverse shapes. Addressing the weak contrast and blurriness characteristics of SAR images, a frequency domain feature enhancement module is rationally introduced. The fusion logic between this module and the spatial domain feature extraction process is designed. By complementing frequency and spatial domain information, the distinction between targets and background is enhanced, assisting in solving related challenges in SAR aircraft target detection. Not limited to lightweight local modules, through overall collaborative optimization of the network architecture, a lightweight scale of 4.1M parameters and 13.7G FLOPs is achieved while ensuring a high mAP detection accuracy of 93.94%, forming a network operation foundation suitable for resource-constrained scenarios.

[0039] 2. Excellent lightweight characteristics and low resource consumption: The network model proposed in this invention has only 4.1M parameters and only 13.7GFLOPs of computation. While maintaining high detection accuracy, it significantly reduces the model's hardware resource consumption and effectively avoids the problem of excessive resource consumption caused by the large size and computational complexity of traditional high-precision detection models. This creates favorable conditions for model application in resource-limited scenarios.

[0040] 3. Leading detection performance and accurate target recognition: In the SAR-AIRcraft-1.0 dataset test, the network achieved an average accuracy (mAP) of 93.94%, which is significantly better than the current mainstream detection models such as YOLOv12s and RT-DETR-R18. This fully demonstrates that it has a stronger ability to extract and distinguish features of SAR aircraft targets and can more accurately meet the needs of detecting SAR aircraft targets with weak contrast and diverse shapes.

[0041] 4. Strong technical practicality and wide applicability: Relying on the dual core advantages of lightweight and high precision, the model of this invention can stably exert detection performance without relying on high-performance hardware, providing a practical and feasible technical path for SAR image target detection tasks, and effectively improving the adaptability and practical application value of SAR aircraft detection technology in resource-constrained scenarios. Attached image description:

[0042] Figure 1 This is a schematic diagram of an efficient SAR aircraft detection process based on multi-branch collaborative calibration and frequency domain compensation in the embodiment.

[0043] Figure 2 This is a schematic diagram of the model training process in the embodiment;

[0044] Figure 3 This is a schematic diagram of the neural network model structure in the embodiment. Detailed implementation method:

[0045] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0046] To address the problems of incomplete feature representation, insufficient weak signal detection, difficulty in balancing lightweight design and accuracy, and low target-background discrimination in existing SAR image aircraft detection, embodiments of the present invention, such as... Figure 1 As shown, an efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation is provided, which specifically includes the following:

[0047] S101. Preprocess the acquired SAR image sample data, divide the SAR image sample data into training set, validation set and test set according to the preset ratio, and label the target categories.

[0048] Specifically, the preprocessing of the acquired SAR image sample data can be designed to include:

[0049] Channel expansion replicates the original single-channel image in the sample data into a three-channel image in the channel dimension, and generates a three-channel image matrix.

[0050] Image data of different resolutions are mapped to a specified pixel specification using a bilinear interpolation algorithm.

[0051] The SAR-Aircraft-1.0 dataset can be used to construct sample data. For example... Figure 2 As shown, channel dimension expansion and image spatial size standardization are implemented for single-channel synthetic aperture radar images acquired from the SAR-Aircraft-1.0 dataset. Channel expansion involves copying the original single-channel image in the channel dimension to generate a three-channel image matrix, satisfying the structural requirements of the subsequent deep learning detection model for the number of input channels while maintaining the pixel grayscale distribution and statistical features unchanged. Size standardization involves mapping image data of different resolutions (including 600, 800, 1000, 1200, and 1500 pixel square resolutions) to a 640×640 pixel size using a bilinear interpolation algorithm. This maintains the target shape proportions while improving the smoothness of edge transitions, thereby reducing feature inconsistencies caused by different input sizes during model training.

[0052] The preprocessed dataset is divided into training, validation, and test sets in a ratio of 8:1:1, corresponding to 80%, 10%, and 10% of the total sample data, respectively. Given that the dataset contains seven aircraft target categories (A220, A320 / 321, A330, ARJ21, Boeing 737, Boeing 787, and others) and a single image may contain multiple categories of target instances, the partitioning process must use class balance constraints to ensure that the coverage and proportion distribution of different categories in each subset are basically consistent, thereby reducing the risk of decreased detection accuracy due to sample class imbalance.

[0053] For the VOC standard XML format annotation files used in the dataset, structured parsing and file reconstruction operations are performed. In the parsing phase, the XML file is read using the DOM (Document Object Model), extracting the relative path of the image, the coordinates of the bounding rectangle of the target, and the target category information, while converting absolute paths to relative paths for storage. In the file reconstruction phase, based on the partitioning of the training and validation sets, two annotation list files, Aircraft_train.txt and Aircraft_val.txt, are generated. Each line of the file consists of the image path, the corresponding target coordinate parameters, and the category number. Multiple targets are separated by spaces to ensure that the data can be directly called in subsequent model training stages and to maintain the consistency and parsability of the data annotation format.

[0054] S102. Construct a neural network model for multi-branch collaborative calibration and frequency domain compensation. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion.

[0055] Specifically, utilizing a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration can be designed to include:

[0056] The number of input feature channels is expanded by point convolution, and the input features are split into two parts, A and B, along the channel dimension;

[0057] The features of part A are extracted using three deep convolutional branches while preserving the original features. The output features of the three deep convolutional branches are concatenated and processed using pointwise convolution and the TeLU activation function. In each deep convolutional branch, element-wise multiplication is used to enhance the interaction and fusion of features within the part.

[0058] The feature data after feature processing is multiplied element-wise with the features of part B to achieve cross-branch feature calibration.

[0059] like Figure 3 The illustrated detection architecture is based on multi-branch collaborative calibration and frequency domain compensation, with a backbone network structure based on a parallel aggregation and calibration (PAC) module for initial capture and calibration of multi-scale features. The input layer receives a preprocessed 3×640×640 SAR image, passes it through a 16-channel initial convolutional layer (stride 2), and generates a basic feature map at a scale of 1 / 2. Next, the backbone network is designed to progress in layers C2-C5, with feature scales at each layer being 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input size, respectively. Each layer integrates one PAC module.

[0060] Specifically, the PAC module first expands the number of input channels to twice the size using 1×1 point convolutions, then splits it into two parts, A and B, each accounting for 50%. Part A is further divided into three sub-branches, employing 3×3 ordinary depthwise convolutions, 5×5 ordinary depthwise convolutions and 5×5 dilated depthwise convolutions (dilation=2), and 7×7 ordinary depthwise convolutions and 7×7 dilated depthwise convolutions (dilation=3) to extract multi-scale spatial information, while retaining a copy of the original features (shortcut) in each branch. Element-wise multiplication enhances feature interaction and fusion within each branch. Subsequently, the output features of these three branches are concatenated into A', and then subjected to point-wise convolution and the TeLU activation function to further improve the complexity of feature representation and non-linear mapping capability. Finally, it is multiplied element-wise with the features of part B to achieve cross-branch feature calibration.

[0061] This includes parallel processing of input features and feature calibration using a parallel aggregation and calibration mechanism, as well as the following:

[0062] Feature optimization is performed using a rectangular channel attention mechanism, including the following optimizations:

[0063] The first moment of each channel of the feature is extracted by global average pooling, and the skewness is obtained by combining the third central moment. The first moment is used to describe the basic strength of the feature, and the skewness is used to describe the asymmetry in the feature distribution.

[0064] The first-order moments and skewnesses are concatenated and channel weights are generated using ID convolution and the Sigmoid activation function, so as to assign a weighting coefficient to each channel.

[0065] Channel weighted optimization is performed on a pairwise basis based on weighted coefficients.

[0066] The calibrated features are then embedded into the Moment Channel Attention (MCA) module.

[0067] The MCA module extracts the first moment (mean M1) of each channel in the input feature map using global average pooling, and calculates the skewness (M2) by combining it with the third central moment. The formula is as follows:

[0068]

[0069] Where, σ c This is the standard deviation of channel c. These two statistics represent the shape of the channel's characteristic distribution—the mean represents the basic strength, while the skewness represents the asymmetry in the distribution.

[0070] Then, channel weights F are generated using 1D convolution (kernel size 7) and a sigmoid activation function:

[0071] F=σ(Conv1D(Concat(M1,M2))) (3)

[0072] This assigns different weighting coefficients to each channel. Finally, by optimizing the input feature X through channel-by-channel weighting, the output feature is obtained:

[0073] Y = X·F (4)

[0074] Specifically, by using a multi-scale aggregation enhancement mechanism to fuse multi-scale features and by utilizing an efficient discriminative frequency domain feedforward network to optimize the frequency domain information of the multi-scale features, it can be designed to include:

[0075] The input features are normalized, and the number of channels of the input features is expanded by point convolution, splitting the input features into two parts A and B along the channel dimension;

[0076] The features of part A are extracted using three deep convolutional branches while preserving the original features. The output features of the three deep convolutional branches are concatenated and processed using pointwise convolution and the TeLU activation function. In each deep convolutional branch, element-wise multiplication is used to enhance the interaction and fusion of features within the part.

[0077] The processed features are residually concatenated with the input features and then normalized again.

[0078] The re-normalized features expand the number of channels through pointwise convolution and extract spatial features using depthwise convolution. The depthwise convolution uses a group convolution mechanism to ensure that each channel learns spatial information independently.

[0079] The spatial features are split into two equal-channel branches, X1 and X2. X1 is activated by TeLU to enhance nonlinearity and then multiplied element-wise with X2. The number of channels is adjusted by pointwise convolution, and the features are divided into blocks by equal size.

[0080] Perform a two-dimensional fast Fourier transform on each block feature to transform the features to the frequency domain, and weight the frequency domain components using a learnable FFT parameter matrix.

[0081] The weighted features are transformed back to the spatial domain by inverse Fourier transform and then reassembled into blocks to restore the original feature size.

[0082] Perform residual connections with the features before renormalization and output the optimized features.

[0083] Upsampling is performed on the feature map C5, the deepest output of the backbone network, to improve spatial resolution. The core of this operation is to match the spatial size of the feature map with the mid-layer output C4 of the backbone network, laying the scale foundation for subsequent feature alignment and fusion. Next, the upsampled deep features are concatenated with the mid-layer feature maps along the channel dimension, fusing deep semantic and mid-layer structural information to form the initial fused features.

[0084] After splicing, the fused features enter the Multi-Scale Aggregation Enhancement (MSAE) module. MSAE is based on a residual structure: the input features are first normalized by Norm, and then flow into the PAC module—this PAC module removes the calibration connections of the three branches of part A from the backbone network version; the processed features are residually connected with the input features of MSAE, and then normalized by Norm before entering the next stage. The embedded High-Efficiency Discriminative Frequency-Domain Feedforward Network (EDFFN) operates sequentially: first, it expands the number of channels to twice the original number through pointwise convolution, and then extracts spatial features through 3×3 depthwise convolution (using a grouped convolution mechanism to ensure that each channel learns spatial information independently); then it splits into two equal-channel branches, X1 and X2. X1 is activated by TeLU to enhance nonlinearity and then multiplied element-wise with X2 to strengthen feature association; then, it adjusts the number of channels through pointwise convolution and enters the frequency domain for processing—dividing into blocks of 5 equal parts (achieved by dimensional rearrangement), performing a two-dimensional fast Fourier transform (FFT) on each block to transform to the frequency domain, using a learnable FFT parameter matrix to weight the frequency domain components (highlighting target-related high-frequency components and suppressing noise-related low-frequency components), and then transforming back to the spatial domain through an inverse Fourier transform (IFFT); finally, it reassembles the blocks to the original feature map size, performs residual connections (element-wise addition) with the features before the second Norm, and outputs the optimized features.

[0085] After obtaining this fusion result, the channel structure is adjusted through pointwise convolution to form a feature map with a uniform number of channels. This feature map is then upsampled and concatenated with the mid-layer extracted feature C3 from the backbone stage along the channel dimension to form a new fused feature. This new feature is then input into the MSAE module for further processing. This process deepens cross-level information integration and strengthens the coherence of the feature structure, outputting feature C3'. C3' undergoes convolution to generate a new mid-layer representation, which is then concatenated with the output feature channels of the first MSAE to construct a new fused feature map. This map is then fed into the MSAE module for optimization, generating a fused feature C4' that covers full-scale information. After convolution, C4' is concatenated with the deep output channel C5 from the backbone stage to form the final fused feature map. The MSAE module performs a final optimization to obtain the fused feature C5' containing full-scale information.

[0086] The neck structure outputs fused features in three sets of sizes: large, medium, and small. These three sets of features are respectively connected to the subsequent detection head, providing multi-scale and multi-semantic level support for the detection module, helping to improve the detection accuracy and robustness of aircraft targets in complex airport backgrounds.

[0087] Among them, the detection head that integrates features to predict the target category adopts an anchorless design to directly regress the target bounding box coordinates and category probabilities from the feature map. In the training phase, the detection head uses CIoU as the bounding box regression loss function, which is constrained by the intersection-union ratio, the Euclidean distance of the bounding box center point, and the aspect ratio consistency.

[0088] The three-level fused feature maps are fed into the detection head for processing to complete the localization and classification prediction of aircraft targets. The detection head adopts an anchor-free design, which, compared with traditional anchor-based methods, can get rid of the constraints of hyperparameters such as preset anchor size and matching strategies, alleviate the problem of positive and negative sample imbalance, simplify the network structure and training process, and reduce computational overhead and memory usage. At the same time, the anchor-free mechanism needs to learn target localization at the pixel level, which has higher requirements for the detailed expression and semantic consistency of the feature maps. The aforementioned multi-scale fused features play a key supporting role here, providing stable input for accurate localization by means of hierarchical expression from shallow details to deep semantics.

[0089] During the training phase, the detection head uses Complete IoU (CIoU) as the bounding box regression loss function to quantify the matching degree between the predicted and ground truth boxes. Unlike traditional IoU or GIoU losses, which only focus on overlapping regions, CIoU, in addition to calculating the intersection-union ratio (IoU), further incorporates the Euclidean distance between the bounding box centers and aspect ratio consistency as supplementary constraints, thus more comprehensively reflecting the geometric similarity between the two. Its mathematical expression is:

[0090]

[0091] Wherein, ρ(b,b) gt The distance between the predicted bounding box and the center point of the ground truth bounding box is represented by , c represents the diagonal length of the minimum closure region, v measures the aspect ratio difference, and α is a balancing coefficient to adjust the weight of the penalty term. This loss function can effectively improve the accuracy of bounding box regression, and is particularly suitable for optimizing the localization of weak contrast targets in SAR images.

[0092] S103. Train the neural network model using the training set, test and validate the trained neural network model using the validation set and test set, and obtain the final target detection model based on the test and validation results.

[0093] Specifically, the weight parameters and bias terms of the network model can be initialized first. A random initialization strategy is used to assign initial values ​​to the parameters of each layer, providing the model with basic learning capabilities. For parameter optimization, the Adaptive Moment Estimator (Adam) optimizer is selected to handle the iterative update of the network parameters. This optimizer dynamically adjusts the learning rate of each parameter by simultaneously tracking the first moment (mean) and second moment (variance) of the parameter gradient. This effectively avoids oscillations during convergence and ensures iterative efficiency. It is well-suited to the lightweight deep network structure in this design and can drive the loss function to converge efficiently to its minimum value.

[0094] After completing parameter initialization and optimizer configuration, the formal training process of the model is initiated. During the training phase, the batch size is set to 16, and mini-batch stochastic gradient descent is used for parameter updates to balance training efficiency and gradient estimation stability. The total number of training epochs is set to 500, allowing the model to fully learn the scattering characteristics of different aircraft targets (A220, A320 / 321, A330, ARJ21, Boeing 737, Boeing 787, and others) in the SAR-Aircraft-1.0 dataset, the target presentation patterns under different airport backgrounds (such as runways, aprons, and vegetated areas), and the feature distribution of multi-scale and multi-angle targets, gradually improving the model's detection accuracy and generalization ability for aircraft targets in complex scenes.

[0095] During training, after each iteration, the model performance is evaluated in real time on an independent validation set. The evaluation metric is the mean accuracy (mAP) of the validation set, supplemented by the trend of the loss function value. When the validation set mAP reaches its optimum and the loss function tends to stabilize, the corresponding model parameter weights are saved and determined as the baseline parameters for the final detection model. This optimal weight file will be used for subsequent performance validation on the test set to systematically evaluate the target detection accuracy, weak signal target recognition ability, and robustness of the proposed method under complex electromagnetic environments and diverse background interference on the SAR-Aircraft-1.0 dataset.

[0096] S104. Input the SAR image to be detected into the target detection model, and use the target detection model to obtain the target category in the SAR image to be detected.

[0097] By integrating a multi-branch collaborative calibration mechanism into the SAR aircraft target detection network, and through feature interaction and calibration between multiple branches, the network can efficiently capture local texture details and global morphological structure of aircraft targets, thereby helping to improve the integrity of feature representation of targets with weak contrast and diverse shapes.

[0098] Furthermore, based on the above method, this embodiment of the invention also provides a high-efficiency SAR aircraft detection system based on multi-branch collaborative calibration and frequency domain compensation, comprising: a sample acquisition module, a network construction module, a network training module, and a target detection module, wherein,

[0099] The sample acquisition module is used to preprocess the acquired SAR image sample data, divide the SAR image sample data into training set, validation set and test set according to a preset ratio, and label the target categories.

[0100] A network construction module is used to construct a neural network model for multi-branch collaborative calibration and frequency domain compensation. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion.

[0101] The network training module is used to train the neural network model using the training set, test and validate the trained neural network model using the validation set and test set, and obtain the final target detection model based on the test and validation results.

[0102] The target detection module is used to input the SAR image to be detected into the target detection model and use the target detection model to obtain the target category in the SAR image to be detected.

[0103] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0104] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0105] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0106] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0107] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A highly efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation, characterized in that, Include: The acquired SAR image sample data is preprocessed and divided into training set, validation set and test set according to a preset ratio, and the target categories are labeled. A neural network model for multi-branch collaborative calibration and frequency domain compensation is constructed. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion. The neural network model is trained using the training set, tested and validated using the validation set and test set, and the final target detection model is obtained based on the test and validation results. The SAR image to be detected is input into the target detection model, and the target detection model is used to obtain the target category in the SAR image to be detected. The parallel aggregation and calibration mechanism is used to process input features in parallel and perform feature calibration, including: The input feature channels are expanded using point convolution, and the input features are split into two parts, A and B, along the channel dimension. The features of part A are extracted using three deep convolution branches while preserving the original features. The output features of the three deep convolution branches are concatenated and processed using pointwise convolution and the TeLU activation function. Element-wise multiplication is used in each deep convolution branch to enhance feature interaction and fusion within the branch. The processed feature data is then multiplied element-wise with the features of part B to achieve cross-branch feature calibration. Multi-scale features are fused using a multi-scale aggregation enhancement mechanism, and the mid-frequency domain information of the multi-scale features is optimized using an efficient discriminative frequency domain feedforward network, including: The input features are normalized, and the number of input feature channels is expanded using pointwise convolution, splitting the input features into two parts, A and B, along the channel dimension. The features in part A are then processed using three deep convolution branches to extract multi-scale spatial information while preserving the original features. The output features of the three deep convolution branches are concatenated and processed using pointwise convolution and the TeLU activation function. Element-wise multiplication is used in each deep convolution branch to enhance feature interaction and fusion within the component. The processed features are then residually connected to the input features and normalized again. The re-normalized features are further expanded using pointwise convolution and deep convolution is then applied. Spatial features are extracted, and the depthwise convolution employs a grouped convolution mechanism to ensure that each channel learns spatial information independently. The spatial features are split into two equal-channel branches, X1 and X2. X1 is activated by TeLU to enhance nonlinearity and then multiplied element-wise with X2. The number of channels is adjusted by pointwise convolution, and the features are divided into blocks with equal dimensions. A two-dimensional fast Fourier transform is performed on each block of features to transform the features to the frequency domain, and the frequency domain components are weighted using a learnable FFT parameter matrix. The weighted features are transformed back to the spatial domain by inverse Fourier transform, and the blocks are reassembled to restore the original feature size. Residual connections are performed with the features before re-normalization, and the optimized features are output.

2. The efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation according to claim 1, characterized in that, The acquired SAR image sample data is preprocessed, including: Channel expansion replicates the original single-channel image in the sample data into a three-channel image in the channel dimension, and generates a three-channel image matrix. Image data of different resolutions are mapped to a specified pixel specification using a bilinear interpolation algorithm.

3. The efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation according to claim 1, characterized in that, The parallel aggregation and calibration mechanism is used to process input features in parallel and perform feature calibration, and also includes: Feature optimization is performed using a rectangular channel attention mechanism, including the following optimizations: The first moment of each channel of the feature is extracted by global average pooling, and the skewness is obtained by combining the third central moment. The first moment is used to describe the basic strength of the feature, and the skewness is used to describe the asymmetry in the feature distribution. The first-order moments and skewnesses are concatenated and channel weights are generated using ID convolution and the Sigmoid activation function, so as to assign a weighting coefficient to each channel. Channel weighted optimization is performed on a pairwise basis based on weighted coefficients.

4. The efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation according to claim 1, characterized in that, The detection head based on fused features to predict target categories adopts an anchorless design to directly regress the target bounding box coordinates and category probabilities from the feature map. During the training phase, the detection head uses CIoU as the bounding box regression loss function, which is constrained by the intersection-union ratio, the Euclidean distance of the bounding box center point, and the consistency of aspect ratio.

5. The efficient SAR aircraft detection method based on multi-branch collaborative calibration and frequency domain compensation according to claim 4, characterized in that, The bounding box regression loss function is expressed as: Where IoU represents the intersection-union ratio, Represents the prediction box With real frame Euclidean distance between the centers This represents the diagonal length of the smallest closure region. This is a parameter used to measure aspect ratio differences. This is a balancing coefficient used to adjust the weight of the penalty term.

6. A high-efficiency SAR aircraft detection system based on multi-branch collaborative calibration and frequency domain compensation, characterized in that, The method described in claim 1 includes: a sample acquisition module, a network construction module, a network training module, and an object detection module, wherein... The sample acquisition module is used to preprocess the acquired SAR image sample data, divide the SAR image sample data into training set, validation set and test set according to a preset ratio, and label the target categories. A network construction module is used to construct a neural network model for multi-branch collaborative calibration and frequency domain compensation. The neural network model includes: an input layer for performing initial convolution operations on the input image; a backbone network for initially capturing multi-scale features and performing feature calibration output; a neck network for fusing multi-scale features; and a detection head for predicting target categories based on fused features. The backbone network adopts a hierarchical structure, and each layer integrates a parallel aggregation and calibration mechanism to process input features in parallel and perform feature calibration. The neck network uses a multi-scale aggregation enhancement mechanism to fuse multi-scale features and uses an efficient discriminative frequency domain feedforward network to optimize the frequency domain information in the multi-scale features, so as to perform collaborative optimization of features in the spatial and frequency domains during multi-scale feature fusion. The network training module is used to train the neural network model using the training set, test and validate the trained neural network model using the validation set and test set, and obtain the final target detection model based on the test and validation results. The target detection module is used to input the SAR image to be detected into the target detection model and use the target detection model to obtain the target category in the SAR image to be detected.

7. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 5.

Images