Image domain aircraft target detection method based on feature enhancement and multi-scale fusion

Abstract

The application discloses an image domain airplane target detection method and device based on feature enhancement and multi-scale fusion, electronic equipment and storage medium. The method embeds a scale-aware frequency selection convolution SFSConv module, a Haar wavelet down-sampling HWD module and a PyramidMaskMamba module on the basis of a YOLOv26n network, and respectively realizes cooperative feature extraction in a frequency domain and a spatial domain, high-frequency detail reservation, long-distance dependence modeling and adaptive mask generation. Meanwhile, a small target-aware label assignment STAL strategy is used to optimize positive and negative sample matching. Experimental results show that the method effectively improves the precision and integrity of SAR image airplane target detection, and performs excellently in small target detection and complex background suppression.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and image processing technology, specifically to an image domain aircraft target detection method, device, electronic device, and storage medium based on feature enhancement and multi-scale fusion. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave imaging system, can overcome the limitations of weather and lighting conditions to achieve all-weather, day-and-night Earth observation. It has been widely used in military reconnaissance, environmental monitoring, and emergency disaster response, becoming a crucial technology for ground target monitoring. Among numerous ground targets, aircraft, as important military and civilian transportation vehicles, are of significant strategic importance for battlefield situational awareness and civil aviation management. However, the inherent speckle noise in SAR images, the discontinuous imaging characteristics of aircraft targets composed of discrete scattering centers, strong interference from man-made features against complex airport backgrounds, and intra-class diversity caused by different imaging attitudes pose serious challenges to aircraft target detection using SAR images.

[0003] Traditional aircraft target detection methods primarily rely on manually designed features and empirical thresholds, such as target segmentation methods based on the Constant False Alarm Rate (CFAR). However, CFAR, based on a single distribution, struggles to handle target detection in heavily interfered environments. While subsequent improved algorithms have enhanced detection performance to some extent, these methods are highly dependent on feature selection and construction, making them ill-suited for SAR images with diverse backgrounds and exhibiting limited generalization capabilities. In recent years, with advancements in hardware computing power and the rapid development of deep learning technology, target detection methods based on Convolutional Neural Networks (CNNs) have demonstrated significant potential in SAR image target detection tasks due to their powerful feature extraction capabilities.

[0004] Currently, CNN-based image object detection algorithms are mainly divided into two categories: one is the two-stage algorithm represented by the R-CNN series, which achieves detection through candidate region extraction and classification prediction, but candidate region generation is time-consuming and difficult to meet real-time requirements; the other is the one-stage algorithm represented by the YOLO (You Only Look Once) series, which does not require candidate box extraction and directly regresses the target category and location during the network's forward propagation process. It has a simple structure, fast inference speed, and is widely used in real-time detection scenarios. Among them, YOLO has gradually attracted researchers' attention due to its fast inference speed and high accuracy, and several improved versions of the algorithm have been proposed.

[0005] However, existing YOLO models still have the following limitations in SAR image target detection applications: First, aircraft targets have relatively complex structures, and after SAR imaging, they appear as discrete scattering centers with blurred edge features and easily lost details. The anisotropy of target scattering makes them quite sensitive to azimuth. Second, man-made features such as terminal buildings around the aircraft create strong background interference, making the model susceptible to interference in high-noise environments, leading to decreased detection accuracy. Furthermore, the scale of aircraft targets varies significantly in SAR images of different resolutions. As the number of network layers increases, the features of small-sized targets weaken or even disappear. Feature extraction networks with fixed receptive fields struggle to effectively extract small-scale target features, resulting in prominent missed detections. These problems present numerous difficulties and challenges for aircraft target detection in SAR images.

[0006] To address these issues, scholars both domestically and internationally have proposed various improvement strategies. Some studies have enhanced target features by introducing attention mechanisms, but these mostly focus on optimizing single problems and struggle to simultaneously solve the multiple challenges of scattering feature extraction, detail preservation, and overall target modeling. Other studies have improved small target detection performance by refining label assignment strategies, but these methods are not fully adapted to the imaging principles of SAR images. Existing methods lack a systematic solution for the multifaceted challenges of aircraft target detection in SAR images, making it difficult to meet the practical needs of high-precision real-time target detection. Summary of the Invention

[0007] The main purpose of this application is to provide an image domain aircraft target detection method, device, electronic device and storage medium based on feature enhancement and multi-scale fusion. It aims to solve the problems faced by existing SAR image aircraft target detection, such as discontinuous discrete scattering centers, strong speckle noise interference, large interference from complex backgrounds and easy loss of small-sized target features. Existing YOLO models lack systematic solutions, resulting in high false alarm rates, many false alarms and difficulty in balancing training efficiency and accuracy.

[0008] To achieve the above objectives, the first aspect of this application provides an image-domain aircraft target detection method based on feature enhancement and multi-scale fusion, comprising: Acquire the SAR image to be detected; The SAR image is input into the improved YOLOv26n network for feature extraction and enhancement. The backbone network of the improved YOLOv26n network is embedded with a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network contains a PyramidMaskMamba module. The frequency-selective convolution (SFSConv) module is used to perform collaborative feature extraction in the frequency and spatial domains on the intermediate feature map to enhance the discrete scattering characterization of the aircraft target and suppress speckle noise interference. At the same time, the Haar wavelet downsampling (HWD) module is used to retain high-frequency detail information during the feature map size reduction process, thereby alleviating the problem of feature loss for small-sized targets. The features processed by the backbone network are input into the PyramidMaskMamba module. The long-distance dependencies are captured through the state space model and combined with the adaptive mask generation mechanism to integrate the feature information of discrete scattering centers to enhance the overall expression of the aircraft target and suppress complex background interference. The STAL strategy of small target-aware label assignment is adopted to perform positive and negative sample matching between the predicted boxes and the ground truth boxes of the improved YOLOv26n network. The scale-aware weight factor is introduced into the improved YOLOv26n network to dynamically adjust the matching threshold of small targets. The improved YOLOv26n network outputs SAR image aircraft target detection results including bounding box coordinates, class confidence, and class labels.

[0009] Optionally, the step of using the frequency-selective convolution SFSConv module to perform collaborative feature extraction in the frequency and spatial domains of the intermediate feature map includes: inputting the same input feature map into a frequency decomposition branch and a scale-aware branch set in parallel in the frequency-selective convolution SFSConv module; in the frequency decomposition branch, converting the input feature map to the frequency domain using Fast Fourier Transform (FFT), and separating the target scattering component from speckle noise using target-learned frequency selection weights; in the scale-aware branch, extracting the spatial domain multi-scale scattering characterization features of the input feature map using parallel convolution branches with different dilation rates or convolution kernel scales, and generating adaptive fusion weights for each scale branch through global average pooling and 1×1 convolution; inputting the frequency domain scattering characterization features output by the frequency decomposition branch and the spatial domain multi-scale scattering characterization features output by the scale-aware branch into the feature fusion branch for channel importance calculation and adaptive weighted summation to obtain collaborative features in the frequency and spatial domains.

[0010] Optionally, in the frequency decomposition branch, the input feature map is transformed to the frequency domain using Fast Fourier Transform (FFT), and the target scattering component and speckle noise are separated using target-learned frequency selection weights. This includes: generating a frequency domain filter kernel to be learned using the kernel function of a fractional-order Gabor filter, wherein the kernel function includes a frequency parameter controlling the filter's frequency response, a phase parameter controlling the phase shift, an attenuation parameter controlling the Gaussian window attenuation amplitude, and a direction parameter controlling the filtering direction; normalizing the generated frequency domain filter kernel to be learned, and then performing a convolution operation with the input feature map of the scale-aware frequency selection convolution (SFSConv) module to obtain the separated target scattering component and speckle noise.

[0011] Optionally, in the scale-aware branch, the spatial domain multi-scale scattering representation features of the input feature map are extracted using parallel convolutional branches with different dilation rates or convolutional kernel scales. This includes: setting four parallel convolutional branches with different dilation rates to cover the local scattering structure of aircraft targets of different sizes; compressing global context information through global average pooling operations; generating adaptive fusion weights for each scale branch by combining 1×1 convolutions; mapping the weights to between 0 and 1 to obtain the spatial domain multi-scale scattering representation features.

[0012] Optionally, the four parallel convolution branches with different dilation rates are respectively a 1×1 convolution, a 3×3 convolution with a dilation rate of 1, a 3×3 convolution with a dilation rate of 2, and a 3×3 convolution with a dilation rate of 3.

[0013] Optionally, the step of using the Haar wavelet downsampling (HWD) module to retain high-frequency detail information during the feature map size reduction process includes: performing a first-level Haar discrete wavelet transform on the input feature map, decomposing it into one low-frequency approximation component and three high-frequency detail components; concatenating the four decomposed components along the channel dimension to achieve lossless encoding of spatial information to the channel dimension, obtaining the concatenated high-channel features; compressing the concatenated high-channel features to the target dimension through 1×1 convolution and performing normalization and activation processing to obtain dimension-compressed channel features; and selecting whether to perform depthwise separable convolution feature learning based on the number of channels of the dimension-compressed channel features to further refine the high-frequency detail features.

[0014] Optionally, the step of capturing long-distance dependencies through a state-space model and combining it with an adaptive mask generation mechanism to integrate the feature information of discrete scattering centers to enhance the overall expression of the aircraft target and suppress complex background interference includes: flipping the input feature map along different directions to simulate the scanning path, and extracting four-directional scanning features by combining grouped convolution; generating four-directional adaptive directional masks through 1×1 convolution of the four-directional scanning features, and normalizing the weight of each direction to between 0 and 1; performing weighted summation of the four-directional scanning features according to the normalized adaptive directional masks to focus on the features of the target boundary and scattering center; then performing pyramid downsampling through convolution with a stride of 2 and connecting it to a mask-guided two-dimensional state-space scanning unit MGD_SS2D to model long-distance dependencies; upsampling through transposed convolution and adding it element-wise with the downsampled features of the same scale to fuse deep semantics and shallow details to obtain the features after pyramid multi-scale fusion; and concatenating and fusing the original input feature map with the features after pyramid multi-scale fusion, while adding residual connections to retain the original feature information to enhance the overall expression of the aircraft target and suppress complex background interference.

[0015] To achieve the above objectives, a second aspect of this application provides an image-domain aircraft target detection device based on feature enhancement and multi-scale fusion, comprising: The image acquisition module is used to acquire the SAR image to be detected; The feature extraction and enhancement module is used to input SAR images into the improved YOLOv26n network for feature extraction and enhancement. The backbone network of the improved YOLOv26n network is embedded with a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network contains a PyramidMaskMamba module. The dual-domain collaborative and detail-preserving module is used to perform collaborative feature extraction in the frequency domain and spatial domain on the intermediate feature map using the frequency-selective convolution SFSConv module to enhance the discrete scattering characterization of the aircraft target and suppress speckle noise interference. At the same time, the Haar wavelet downsampling HWD module is used to preserve high-frequency detail information during the feature map size reduction process. The mask generation module is used to input the features processed by the backbone network into the PyramidMaskMamba module. It captures long-distance dependencies through the state space model and integrates the feature information of discrete scattering centers with the adaptive mask generation mechanism to enhance the overall expression of the aircraft target and suppress complex background interference. The sample matching module is used to perform positive and negative sample matching between the predicted bounding boxes and the ground truth bounding boxes of the improved YOLOv26n network using the small target-aware label allocation STAL strategy. The improved YOLOv26n network introduces a scale-aware weight factor to dynamically adjust the matching threshold for small targets. The results output module is used to output SAR image aircraft target detection results containing bounding box coordinates, class confidence, and class labels on the improved YOLOv26n network.

[0016] To achieve the above objectives, a third aspect of this application provides an electronic device including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the image domain aircraft target detection method based on feature enhancement and multi-scale fusion as provided above.

[0017] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the image domain aircraft target detection method based on feature enhancement and multi-scale fusion provided above.

[0018] This application discloses an image-domain aircraft target detection method, apparatus, electronic device, and storage medium based on feature enhancement and multi-scale fusion. The method includes acquiring a SAR image to be detected; inputting the SAR image into an improved YOLOv26n network for feature extraction and enhancement processing, wherein the backbone network of the improved YOLOv26n network embeds a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network includes a PyramidMaskMamba module; using the SFSConv module to perform collaborative feature extraction in the frequency and spatial domains on the intermediate feature map to enhance the discrete scattering representation of the aircraft target and suppress speckle noise interference; and simultaneously using the HWD module to reduce the size of the feature map. The improved YOLOv26n network retains high-frequency detail information to alleviate the problem of feature loss for small targets. Features processed by the backbone network are input into the PyramidMaskMamba module, which captures long-distance dependencies through a state-space model and combines this with an adaptive mask generation mechanism to integrate feature information from discrete scattering centers, thereby enhancing the overall representation of aircraft targets and suppressing interference from complex backgrounds. A small target-aware label allocation (STAL) strategy is employed to perform positive and negative sample matching between the predicted bounding boxes and ground truth bounding boxes of the improved YOLOv26n network. A scale-aware weight factor is introduced into the improved YOLOv26n network to dynamically adjust the matching threshold for small targets. The improved YOLOv26n network outputs SAR image aircraft target detection results containing bounding box coordinates, class confidence, and class labels. The beneficial effects of this application are as follows: Strong scattering feature extraction capability: The designed SFSConv module effectively extracts the discrete scattering features of SAR aircraft targets through collaborative modeling in the frequency and spatial domains, while suppressing the interference of speckle noise and improving the discriminative power of the features; Good high-frequency detail preservation: The introduced HWD module preserves high-frequency detail information while achieving downsampling, effectively alleviating the feature loss problem of small-sized aircraft targets and improving the detection completeness of small targets; Enhanced overall target representation: The PyramidMaskMamba module effectively integrates discrete scattering through long-range dependency modeling and adaptive masking. The central feature solves the component segmentation problem and suppresses interference from complex backgrounds, improving the completeness of detection results. Small target detection performance is improved; the STAL label allocation strategy specifically optimizes sample matching for small targets, effectively reducing missed detections and false alarms in the background, and improving the detection accuracy of small-sized aircraft targets. Training efficiency and accuracy are balanced; the ProgLoss loss balancing strategy and the MuSGD hybrid optimizer effectively accelerate the model's convergence speed while improving the model's final detection accuracy. The overall improvement maintains the model's lightweight and real-time performance, meeting the deployment requirements of embedded platforms. Attached Figure Description

[0019] Figure 1A schematic diagram of the overall structure of the improved YOLOv26n network; Figure 2 A schematic diagram of the SFSConv module structure for scale-aware frequency selection; Figure 3 This is a schematic diagram of the Haar wavelet downsampling HWD module structure; Figure 4 This is a schematic diagram of the PyramidMaskMamba module structure; Figure 5 From left to right, these are schematic diagrams of the loss curves for training set bounding box regression loss, cross-entropy loss, and distribution focus loss; Figure 6 From left to right, these are schematic diagrams of the loss curves for validation set bounding box regression loss, cross-entropy loss, and distribution focus loss. Figure 7 A graph illustrating the accuracy rate; Figure 8 This is a schematic diagram of the recall rate curve; Figure 9 This is a schematic diagram of the mAP@50 curve; Figure 10 This is a schematic diagram of the mAP@50-95 curve; Figure 11 This is a schematic diagram of the accuracy-confidence curve; Figure 12 This is a schematic diagram of the recall-confidence curve; Figure 13 This is a diagram illustrating the precision-recall curve. Figure 14 This is a schematic diagram of the F1-confidence curve; Figure 15 This is a schematic diagram of the confusion matrix for Faster R-CNN. Figure 16 This is a schematic diagram of the confusion matrix for Cascade R-CNN. Figure 17 This is a schematic diagram of the RT-DETR confusion matrix; Figure 18 This is a schematic diagram of the YOLOv12n confusion matrix; Figure 19 This is a schematic diagram of the YOLOv26n confusion matrix; Figure 20 A schematic diagram of the improved YOLOv26n confusion matrix; Figure 21 A visualization of the Faster R-CNN detection results; Figure 22 A visualization of the YOLOv26n detection results; Figure 23 A visualization of the improved YOLOv26n detection results; Figure 24 This is a flowchart of the method in this application.

[0020] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0022] Existing YOLO-based SAR image target detection methods typically employ standard convolution for feature extraction and 2-stride convolution for downsampling. However, due to the unique nature of SAR imaging, aircraft targets exhibit a discrete scattering center distribution, with blurred edges and susceptibility to speckle noise. Furthermore, man-made features in complex airport backgrounds can cause strong interference. Existing methods with fixed receptive field feature extraction networks struggle to effectively capture the discrete scattering features of small-scale targets. The downsampling process leads to the loss of high-frequency details, and the lack of ability to model long-range dependencies results in high false alarm rates, numerous false alarms, and slow training convergence for small targets.

[0023] refer to Figure 24 The first embodiment of this application provides an image-domain aircraft target detection method based on feature enhancement and multi-scale fusion to address the technical problems encountered in SAR image aircraft target detection, such as discontinuous discrete scattering centers, strong speckle noise interference, large interference from complex backgrounds, and easy loss of small-sized target features. This method can be executed by a processor, which can be located on a terminal or server. The execution process of the image-domain aircraft target detection method based on feature enhancement and multi-scale fusion is as follows: Step S101: Obtain the SAR image to be detected.

[0024] The processor can input the SAR image to be detected into the system, first perform speckle noise suppression processing, use Lee filtering to denoise the image, then perform image size normalization to adjust the image to the standard input size of 640×640, and at the same time perform pixel value normalization processing to map the pixel values ​​to between 0 and 1 to adapt to the input requirements of the model.

[0025] Step S102: Input the SAR image into the improved YOLOv26n network for feature extraction and enhancement.

[0026] Figure 1The diagram shows the improved YOLOv26n network structure, which consists of an input layer, a backbone network, a neck feature fusion network, and a head decoupled detection head. Specifically, SFSConv enhances SAR scattering features, HWD performs detail-preserving downsampling, and PyramidMaskMamba provides multi-scale holistic representation. The preprocessed image is input to the improved backbone network, sequentially passing through the backbone network's Focus layer, C2F layer, and a feature extraction layer embedding SFSConv and HWD modules for feature extraction. Each SFSConv module performs collaborative feature extraction in the frequency and spatial domains, enhancing scattering features. Each HWD module performs downsampling while preserving high-frequency details, ultimately outputting multi-scale feature maps corresponding to 8x, 16x, and 32x downsampling features, respectively.

[0027] Step S103: Use the SFSConv module to perform joint feature extraction in the frequency domain and spatial domain on the intermediate feature map.

[0028] In one embodiment of this application, the same input feature map is input to both the frequency decomposition branch and the scale-aware branch, which are set in parallel within the frequency-selective convolution (SFSConv) module. In the frequency decomposition branch, the input feature map is transformed to the frequency domain using a Fast Fourier Transform (FFT), and the target learning frequency selection weights are used to separate the target scattering component from speckle noise.

[0029] Furthermore, a frequency domain filter kernel to be learned is generated using the kernel function of a fractional-order Gabor filter. This kernel function includes frequency parameters controlling the filter's frequency response, phase parameters controlling the phase shift, attenuation parameters controlling the Gaussian window attenuation amplitude, and direction parameters controlling the filtering direction. After normalizing the generated frequency domain filter kernel, it is convolved with the input feature map of the SFSConv module to obtain the separated target scattering components and speckle noise.

[0030] In an optional embodiment of this application, in the scale-aware branch, spatial domain multi-scale scattering representation features of the input feature map are extracted using parallel convolutional branches with different dilation rates or convolutional kernel scales. Adaptive fusion weights for each scale branch are generated through global average pooling and 1×1 convolution. Further, four parallel convolutional branches with different dilation rates are set to cover the local scattering structures of aircraft targets of different sizes. Global average pooling is used to compress global contextual information, and 1×1 convolution is combined to generate adaptive fusion weights for each scale branch, mapping the weights to between 0 and 1.

[0031] For example, the four parallel convolution branches with different dilation rates are a 1×1 convolution, a 3×3 convolution with a dilation rate of 1, a 3×3 convolution with a dilation rate of 2, and a 3×3 convolution with a dilation rate of 3.

[0032] Furthermore, the frequency domain scattering characterization features output from the frequency decomposition branch and the spatial domain multi-scale scattering characterization features output from the scale perception branch are input into the feature fusion branch to calculate channel importance and perform adaptive weighted summation, thereby obtaining the synergistic features of the frequency domain and the spatial domain.

[0033] To address the issues of discontinuous scattering centers and strong speckle noise interference in SAR images, this application designs a Scale-Aware Frequency Selection Convolution (SFSConv) module to replace the standard convolution module in the original backbone network. For example... Figure 2 As shown, the SFSConv module consists of a frequency decomposition branch, a scale-aware branch, and a feature fusion branch. The frequency decomposition and scale-aware branches are computed in parallel using the same input feature map, and their outputs are then adaptively fused in the feature fusion branch. The input feature map is an intermediate feature map output from the previous network layer. Through the collaborative mechanism of frequency decomposition and scale awareness, the SFSConv module simultaneously extracts the scattering features of SAR aircraft targets in both the frequency and spatial domains, adaptively suppressing speckle noise interference.

[0034] Furthermore, the frequency decomposition branch is used to extract frequency domain scattering characteristics. This branch transforms the feature map output from the previous convolutional layer into the frequency domain using FFT, and assigns higher weights to frequency components related to the aircraft scattering center and lower weights to frequency components corresponding to speckle noise and background clutter by selecting weights based on the target learned frequencies. The aforementioned learnable weight matrix participates in backpropagation updates as network parameters, as detailed below: The kernel function of a fractional-order Gabor filter is used to generate the frequency domain filter kernel to be learned, adapting to the scattering feature extraction of SAR images, and realizing the separation of target scattering components and speckle noise. The kernel function of a fractional-order Gabor filter can be expressed as:

[0035] in, This is the kernel function output of the Gabor filter. These are learnable frequency parameters that control the frequency response of the filter. The learnable phase parameter controls the phase shift of the filter. The attenuation parameter is a learnable parameter that controls the attenuation magnitude of the Gaussian window. This is the direction parameter of the filter, which controls the direction of filtering. The spatial coordinates of the kernel function are two-dimensional local spatial coordinates with the center of the filter kernel as the origin and the pixel grid as the unit.

[0036] After normalizing the generated frequency domain filter kernel to be learned, it is compared with the input feature map of the SFSConv module. Convolution is performed to obtain the separated target scattering components and coherent speckle noise. Its expression is:

[0037] in, This is the frequency domain characteristic output after Gabor filtering. This is the input feature map for the module. It is a local constant (e.g., taking the value 1e-8) to prevent the denominator from being 0 and to ensure calculation stability. The L1 norm of the Gabor kernel. This is an abbreviation for the learnable Gabor kernel. This is represented as a convolution operation.

[0038] Furthermore, the scale-aware branch is used to extract multi-scale scattering representation features in the spatial domain. This branch sets up four parallel convolutional branches with different dilation rates or kernel scales. For example, the four parallel convolutional branches are 1×1 convolutions, 3×3 convolutions with a dilation rate of 1, 3×3 convolutions with a dilation rate of 2, and 3×3 convolutions with a dilation rate of 3. Alternatively, they can be equivalently set as 3×3, 5×5, 7×7, and dilated 3×3 convolutions to cover the local scattering structure of aircraft targets of different sizes. Subsequently, global average pooling and 1×1 convolutions are used to generate channel weights for each scale branch to enhance the branch response that better matches the current target scale, as follows: Adaptive fusion weights with four multi-scale branches are generated through global average pooling and 1×1 convolution, and their expression is as follows:

[0039] in, For the first The fusion weights of each scale branch. This is a global average pooling operation that compresses global context information. This is a 1×1 convolution operation used to generate weights. This is a normalization function that maps the weights to the range of 0 to 1. i =1,2,3,4 correspond to four scale branches with different expansion rates.

[0040] Furthermore, the feature fusion branch is used to adaptively fuse frequency domain scattering characterization features with spatial domain multi-scale scattering characterization features to obtain synergistic features in the frequency and spatial domains, providing a more discriminative target representation for the subsequent neck network and detection head, as detailed below: The features of the four branches are weighted and summed according to the dynamic weights to obtain the multi-scale fusion features in the spatial domain, which are expressed as follows:

[0041] in, This refers to the output features of the spatial perception branch. For the first i Output features of each scale branch.

[0042] Post-processing of the frequency domain features involves channel calibration and nonlinear activation of the Gabor-filtered features to enhance their expressive power. The expression is as follows:

[0043] in, The output features of the frequency-aware branch. This is a batch normalization operation to stabilize the feature distribution. It is a nonlinear activation function that enhances the nonlinear representation of features.

[0044] Channel importance calculation for spatial features. The channel importance of spatial features is measured by the mean of the absolute values ​​of the spatial dimensions.

[0045]

[0046] in, The importance score for the channel of spatial features. These represent the height and width of the feature map, respectively. Spatial features in location The pixel value.

[0047] The channel importance of frequency features is measured by the absolute mean of the spatial dimensions, and its expression is:

[0048] in, Channel importance score for frequency characteristics. Frequency characteristics at location The pixel value.

[0049] Based on the channel importance score, the fusion weights of spatial and frequency features are obtained through normalization, and their expression is as follows:

[0050] in, The fusion weights are for spatial features. For the fusion weights of frequency features, It is a local constant (e.g., taking the value 1e-8) to prevent the denominator from being 0 and to ensure calculation stability.

[0051] The spatial and frequency features are weighted and summed according to the adaptive weights to obtain the output features of the SFSConv module:

[0052] in, This is the final output feature of the SFSConv module.

[0053] Step S104: The HWD module is used to preserve high-frequency detail information during the feature map size reduction process.

[0054] In one embodiment of this application, the input feature map is subjected to a first-level Haar discrete wavelet transform, decomposing it into one low-frequency approximation component and three high-frequency detail components. The four decomposed components are concatenated along the channel dimension to achieve lossless encoding of spatial information to the channel dimension, resulting in concatenated high-channel features. The concatenated high-channel features are then compressed to the target dimension using a 1×1 convolution and normalized and activated to obtain dimension-compressed channel features. Based on the number of channels in the dimension-compressed channel features, it is selected whether to perform depthwise separable convolution feature learning to further refine the high-frequency detail features.

[0055] Specifically, to address the problem of high-frequency detail loss during traditional downsampling, this application embeds the Haar wavelet downsampling module (HWD) into the backbone network, replacing the traditional convolutional downsampling layer with a stride of 2. The HWD module downsamples features through Haar wavelet transform while preserving high-frequency detail components, thus mitigating the problem of small target feature loss during downsampling. For example... Figure 3 As shown, this module decomposes the input features into low-frequency approximate components and high-frequency detail components using Haar wavelet transform. While downsampling the low-frequency components, it retains and fuses the high-frequency detail components, and enhances the contribution of detail features through an attention mechanism. This achieves feature map size reduction while maximizing the preservation of high-frequency details such as the discrete scattering center and edge contours of the aircraft target, thus improving the feature integrity of small-sized targets. The processing procedure of the HWD module can be included as follows: (a) Haar wavelet first-order decomposition. The input feature map... It is decomposed into one low-frequency approximation component and three high-frequency detail components, preserving the global structure and local details completely.

[0056]

[0057] in, A It is a low-frequency approximation component that preserves the global contour and structural information of the target. H For high-frequency detail components in the horizontal direction, edge and scattering details in the horizontal direction are preserved. V For high-frequency detail components in the vertical direction, preserve the edge and scattering details in the vertical direction. D For high-frequency detail components in the diagonal direction, edge and scattering details in the diagonal direction are preserved. This is a first-order Haar discrete wavelet transform operation. X 1 represents the input feature map of the HWD module.

[0058] (b) Channel splicing of high and low frequency components. The four components obtained from the decomposition are spliced ​​along the channel dimension to achieve lossless encoding of spatial information to the channel dimension, resulting in the spliced ​​high-channel features.

[0059]

[0060] in, The concatenated feature map has a spatial size that is half that of the input and a channel count that is four times that of the input. This is a concatenation operation along the channel dimension, where dim represents the concatenation dimension. (c) Channel Dimension Compression. The concatenated high-channel features are compressed to the target dimension using 1×1 convolutions, while normalization and activation are performed to avoid dimensionality explosion. The expression is as follows:

[0061] in, This is the feature map after channel compression.

[0062] Based on the number of channels in the channel features after dimensionality compression, we select whether to perform depthwise separable convolution feature learning, and further refine the detailed features as follows:

[0063] in, This is the final output feature of the HWD module. This refers to the number of output channels of the module. The 64-channel count was manually designed. This is a depthwise separable convolution operation used to extract detailed features.

[0064] Step S105: Input the features processed by the backbone network into the PyramidMaskMamba module.

[0065] In one embodiment of this application, a state-space model is used to capture long-range dependencies, and an adaptive mask generation mechanism is combined to integrate the feature information of discrete scattering centers to enhance the overall representation of the aircraft target and suppress interference from complex backgrounds. Further, the input feature map is flipped along different directions to simulate a scanning path, and four-directional scanning features are extracted using grouped convolution. The four-directional scanning features are then processed by 1×1 convolution to generate adaptive directional masks for four directions, with the weight of each direction normalized to between 0 and 1. The four-directional scanning features are weighted and summed according to the directional masks to focus on the features of the target boundary and scattering centers. Then, pyramid downsampling is performed using a convolution with a stride of 2, and a mask-guided two-dimensional state-space scanning unit MGD_SS2D is used to model long-range dependencies. Upsampling is performed by transposed convolution, and deep semantics and shallow details are fused element-wise with the downsampled features at the same scale to obtain the pyramid multi-scale fused features. The original input feature map and the pyramid multi-scale fused features are then concatenated and fused, with residual connections added to retain the original feature information, thereby enhancing the overall representation of the aircraft target and suppressing interference from complex backgrounds.

[0066] To address the issues of component segmentation and strong background interference caused by discrete scattering of aircraft targets, this application designs a PyramidMaskMamba module in the feature fusion network. The PyramidMaskMamba module captures long-range dependencies based on a state-space model and combines multi-scale mask modeling to enhance the overall representation of the aircraft target and suppress complex background interference. Figure 4 As shown, this module is built on a state-space model. Through multi-scale feature input, SSM feature enhancement, adaptive mask generation, and feature fusion, it captures the long-range dependencies of the target, integrates the feature information of discrete scattering centers, and strengthens the overall representation of the aircraft target. Simultaneously, it suppresses interference features from complex backgrounds through adaptive masking, solving the problems of incomplete detection results and target fragmentation. The processing procedure of the PyramidMaskMamba module can be included as follows: (a) Four-directional scanning feature generation. Features are generated by simulating scanning paths in different directions through feature flipping, combined with grouped convolution to extract features in each direction. It can be represented as:

[0067] in, For the first i Scanning features in each direction. To the feature map along The direction is flipped to simulate scanning in different directions. For the first i There are four scanning directions: down, up, left, and diagonal. This indicates a convolution operation with a 3×3 kernel. The feature map for the input PyramidMaskMamba module is generated after the neck network is fused by FPN / PAN.

[0068] (b) Adaptive orientation mask generation. Four adaptive masks are generated using 1×1 convolutions for weighted fusion of directional features. m It can be represented as:

[0069] in, m For the direction mask, The activation function normalizes the weights in each direction to between 0 and 1.

[0070] (c) Mask-guided directional feature fusion: Based on the directional mask, the features in the four directions are weighted and summed to focus on the features of the target boundary and the scattering center, and the output features of the mask-guided scanning unit are then fused. It can be represented as:

[0071] in, The output characteristics of the mask-guided scanning unit. Let be the mask weight for the i-th direction.

[0072] (d) Pyramid downsampling feature extraction. Downsampling is performed using convolutions with a stride of 2, and simultaneously connected to MGD_SS2D. The MGD_SS2D unit is a mask-guided 2D state-space scanning unit within the PyramidMaskMamba module, used to model long-range dependencies at the corresponding scale. Its expression is:

[0073] in, For the first pyramid l Input features of the layer. For the first pyramid l+ Output features of layer 1 This indicates a convolutional downsampling operation with a stride of 2. This indicates that the mask guides the two-dimensional state space scanning unit. (For details on MGD_SS2D operations, please refer to Liu et al.'s paper VMamba: Visual StateSpace Model).

[0074] Pyramid upsampling feature fusion involves upsampling via transposed convolution and then adding it element-wise with downsampling features of the same scale, thus fusing deep semantics with shallow details. Its expression is:

[0075] in, The first after upsampling and fusion l Layer features. This is a transposed convolution operation used for upsampling of the feature map.

[0076] (e) Residual fusion output of the PMM module. The original input to the PyramidMaskMamba module (i.e., the fusion feature map to be enhanced obtained after the neck network is fused by FPN / PAN) is input to the PyramidMaskMamba module. The features are then combined with those from the pyramid and spliced ​​together, with residual connections added to preserve the original feature information. The expression is as follows:

[0077] in, This is the final output feature of the PMM module. These are features resulting from the fusion of multiple scales within the pyramid. For the residual connection branch, channel projection is performed on the input features.

[0078] Step S106: The small target perception label allocation STAL strategy is used to perform positive and negative sample matching between the predicted boxes and the ground truth boxes of the improved YOLOv26n network.

[0079] In one embodiment of this application, a scale-aware weight factor is introduced into the improved YOLOv26n network to dynamically adjust the matching threshold for small targets, optimizing sample matching between small and difficult targets and reducing false alarms in the background. Furthermore, a progressive loss balancing strategy (ProgLoss) is employed during training to dynamically adjust the weights of classification, regression, and fine-grained feature losses, adapting to the optimization needs of different training stages. Simultaneously, a hybrid optimizer (MuSGD) is used for iterative parameter optimization, accelerating model convergence and improving the final detection accuracy.

[0080] Next, this application designs a small target-aware label assignment STAL strategy and a progressive loss balancing ProgLoss strategy, and uses a hybrid optimizer MuSGD to construct a training optimization scheme.

[0081] To address the issue of misclassification of small target samples in traditional labeling strategies, a Small Target Aware Labeling Strategy (STAL) is designed. This strategy introduces a scale-aware weight factor to enhance the matching weight of ground truth bounding boxes for small-sized aircraft targets. Simultaneously, it employs a multi-dimensional matching criterion, dynamically adjusting the positive and negative sample matching thresholds based on target scale and scattering feature similarity to optimize sample matching between small and difficult targets, thereby reducing missed detections of small targets and false alarms in the background.

[0082] Furthermore, a scale-aware weighting factor is introduced to enhance the matching weight of small targets, addressing the issue of misclassification of small target samples. Its expression is:

[0083] in, For scale-aware weights, the smaller the target size, the greater the weight. s The scale value of the target. This is a scaling factor used to control the magnitude of weight adjustment.

[0084] A progressive loss balancing strategy, ProgLoss, is designed. During training, the weights of classification loss, regression loss, and fine-grained feature loss are dynamically adjusted to adapt to the optimization needs of different training stages, accelerating model convergence. Simultaneously, a hybrid optimizer, MuSGD, is employed, combining the optimization advantages of SGD and Adam to improve training efficiency and the model's final detection accuracy.

[0085] The weights of classification, regression, and hard sample loss are dynamically adjusted to achieve balanced convergence of the loss during training. The expression for this is:

[0086] in, This represents the total loss value. For classification loss, cross-entropy loss is used to optimize target classification. For regression loss, CIoU+DFL loss is ultimately used to optimize bounding box localization. , , All are weighted coefficients.

[0087] Step S107: Output the SAR image aircraft target detection results, including bounding box coordinates, class confidence, and class label, into the improved YOLOv26n network.

[0088] The enhanced fusion feature input is decoupled from the detection head, and the classification branch and regression branch are processed in parallel to output the target's class probability, confidence score and bounding box coordinates, respectively. Redundant detection boxes are removed by non-maximum suppression (NMS), and finally the position and class information of all aircraft targets in the image are output.

[0089] A SAR aircraft target detection dataset was constructed, containing SAR images with different imaging angles, resolutions, and background scenes, covering various types of aircraft targets. The dataset was preprocessed before being input into the network for training. The preprocessing methods are well-known to those skilled in the art and will not be elaborated further.

[0090] In another embodiment of this application, the image domain aircraft target detection method based on feature enhancement and multi-scale fusion further includes initializing model parameters using a transfer learning strategy and fine-tuning by loading pre-trained YOLOv26n weights. During training, the STAL label allocation strategy and ProgLoss loss function are used, and the MuSGD optimizer is used to train the model end-to-end to optimize network parameters, enabling the model to simultaneously adapt to tasks such as scattering feature extraction, small target detection, and complex background anti-interference.

[0091] The training loss and validation loss curves of the model show a good convergence trend, with no overfitting or underfitting.

[0092] from Figure 5 It can be seen that the training loss includes bounding box loss (box_loss), classification loss (cls_loss), and distribution focus loss (dfi_loss). The three types of sub-losses continuously decrease and gradually stabilize with the increase of the number of iterations, indicating that the model continuously learns the target features on the training set and the parameter optimization is effective.

[0093] from Figure 6 As can be seen, the validation loss includes bounding box loss (box_loss), classification loss (cls_loss), and distribution focus loss (dfi_loss). All three types of sub-losses decrease synchronously and stabilize at low levels. The trend of the validation set loss is consistent with that of the training set loss, indicating that the model has good generalization ability and there is no overfitting problem on the training set.

[0094] The loss curve is smooth overall without drastic fluctuations, proving that the training process is stable and the model has good consistency and reliability in learning data features.

[0095] like Figures 7-10 As shown, the model training performance is analyzed from the precision, recall, mAP@50, and mAP@50-95 curves: Precision and recall are the core metrics for measuring the false positives and false negatives of a model. Figure 7 and Figure 8 These are schematic diagrams of precision and recall curves, respectively.

[0096] Both precision and recall curves rise rapidly with training iterations and then plateau. The core conclusions are as follows: In the early stages of model training, precision and recall improved rapidly, indicating that the model quickly captured key features of the target and its detection capability was initially formed.

[0097] The fact that both metrics remained stable at high levels during the later stages of training indicates that after the model converged, the ability to accurately identify targets and recall targets completely was balanced, which can effectively reduce false positives and false negatives.

[0098] mAP is a core evaluation metric for object detection. This experiment statistically analyzed mAP@50 (mean precision at IOU=0.5) and mAP50-95 (mean precision from IOU 0.5 to 0.95). The results are as follows: Figure 9 and Figure 10 As shown: from Figure 9 As can be seen, the mAP@50 curve rises rapidly and stabilizes at a high level, indicating that the model achieves extremely high accuracy in target localization and classification under a relaxed IOU threshold, demonstrating excellent recognition performance in typical detection scenarios. Figure 10 As can be seen, the mAP50-95 curve also continues to rise and converge, indicating that the model still has excellent performance under strict IOU thresholds and multi-scale detection scenarios, and has strong adaptability to tasks with high target localization accuracy requirements. The simultaneous achievement of dual mAP indicators verifies that the model has both classification accuracy and localization precision, and has good robustness.

[0099] The training performance of the model will be analyzed below using the precision-confidence curve, recall-confidence curve, precision-recall curve, and F1-confidence curve.

[0100] Figure 11 The curve shows the change in model precision versus confidence threshold. Precision steadily increases with increasing confidence. The precision for all categories reaches 1.00 at a confidence level of 0.961. Within the high confidence range, the model detects no errors, indicating extremely high reliability of the output results. For each model category, the precision is close to 1 at high confidence, while the precision remains at a reasonable level in the low confidence range. This suggests that the model can obtain relatively reliable detection results by increasing the confidence threshold. Figure 12 The curve shows the model recall rate versus the confidence threshold. The recall rate decreases gradually as the confidence level increases. The overall recall rate reaches 0.93 at a confidence level of 0.000, indicating that the model can cover the vast majority of targets at extremely low thresholds with a very low false negative rate. The recall rate for various aircraft types remains high overall, with mainstream models such as the A320 / 321 and A330 showing better recall performance. Even with increased confidence, the recall rate does not drop significantly, demonstrating the model's stable ability to capture various targets.

[0101] Figure 13The precision-recall curves show that the average precision (mAP) at 0.5 for all categories reaches 0.872, indicating excellent and stable overall detection performance. Comparing the AP values ​​of individual aircraft models, the A320 / 321 ranks first with an AP value of 0.980, demonstrating the best recognition and localization performance for this aircraft type. The AP values ​​for mainstream large passenger aircraft such as the A330 (0.901) and Boeing 787 (0.917) both exceed 0.9, demonstrating strong feature capture capabilities for conventional passenger aircraft with sufficient samples and distinct size characteristics. For aircraft with relatively small sample sizes (such as the ARJ21 and A220), the AP values ​​reach 0.852 and 0.858 respectively, without a significant decline in precision, proving that the model training has no significant class bias and does not overfit to aircraft with large sample sizes, exhibiting good balance in multi-class detection. Overall, the precision-recall curves are close to the upper right corner, proving that the model achieves an optimal trade-off between precision and recall, and that the detection performance is reliable.

[0102] Figure 14 The graph shows the F1 score versus confidence threshold, visually illustrating the overall balance between precision and recall at different confidence levels. As can be seen, the F1 score for all categories peaks at 0.83 at a confidence level of 0.438, representing the optimal confidence threshold for the model. Among the various aircraft types, mainstream models such as the A320 / 321, A330, and Boeing 787 exhibit significantly higher peak F1 scores than other categories. The ARJ21, A220, and other categories also maintain relatively high F1 levels.

[0103] Figures 15-20 The diagrams show the confusion matrices of Faster R-CNN, Cascade R-CNN, RT-DETR, YOLOv12n, YOLOv26n, and the improved YOLOv26n, respectively. It is easy to see from the diagonal values ​​that the improved YOLOv26n outperforms the other algorithms in detection performance on all models.

[0104] The SAR image to be detected is input into the trained improved model. First, feature extraction is performed through the backbone network, with the SFSConv module enhancing scattering features and the HWD module preserving high-frequency details. Then, multi-scale feature fusion is performed in the neck region, followed by feature enhancement through the PyramidMaskMamba module. Finally, the detection result, including the target's bounding box coordinates, class confidence score, and class label, is output through the decoupled detection head, completing the detection and identification of SAR aircraft targets.

[0105] like Figures 21-23The figures show the visualization detection of SAR images under the same imaging conditions using Faster R-CNN, YOLOv26n, and improved YOLOv26n algorithms, respectively. As can be seen from the figures, the Faster R-CNN algorithm has missed detections and low detection confidence, YOLOv26n has no missed detections but has low confidence in the detection of some targets, and the improved YOLOv26n has no missed detections and high detection confidence. This indicates that the improved YOLOv26n has good recognition performance for various models.

[0106] To address the aforementioned technical problems, a second embodiment of this application provides an image domain aircraft target detection device based on feature enhancement and multi-scale fusion to solve the same technical problems as the method embodiment. This device may include the following modules: The image acquisition module is used to acquire the SAR image to be detected.

[0107] The feature extraction and enhancement module is used to input SAR images into the improved YOLOv26n network for feature extraction and enhancement. The backbone network of the improved YOLOv26n network is embedded with a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network contains a PyramidMaskMamba module.

[0108] The dual-domain collaborative and detail-preserving module is used to perform collaborative feature extraction in the frequency and spatial domains of the intermediate feature map using the SFSConv module to enhance the discrete scattering characterization of the aircraft target and suppress speckle noise interference. At the same time, the HWD module is used to preserve high-frequency detail information during the feature map size reduction process.

[0109] The mask generation module is used to input the features processed by the backbone network into the PyramidMaskMamba module. It captures long-distance dependencies through the state space model and integrates the feature information of discrete scattering centers with the adaptive mask generation mechanism to enhance the overall representation of the aircraft target and suppress complex background interference.

[0110] The sample matching module is used to perform positive and negative sample matching between the predicted bounding boxes and the ground truth bounding boxes of the improved YOLOv26n network using the small target-aware label allocation STAL strategy. The improved YOLOv26n network introduces a scale-aware weight factor to dynamically adjust the matching threshold for small targets.

[0111] The results output module is used to output SAR image aircraft target detection results containing bounding box coordinates, class confidence, and class labels on the improved YOLOv26n network.

[0112] It is understood that the above-described device embodiments correspond to the method embodiments. Therefore, the device embodiments have all the beneficial effects of the method embodiments, and will not be repeated here.

[0113] To address the aforementioned technical problems, a third embodiment of this application provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the image domain aircraft target detection method based on feature enhancement and multi-scale fusion as provided above.

[0114] To address the aforementioned technical problems, the fourth embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the image domain aircraft target detection method based on feature enhancement and multi-scale fusion provided above.

[0115] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An image-domain aircraft target detection method based on feature enhancement and multi-scale fusion, characterized in that, include: Acquire the SAR image to be detected; The SAR image is input into the improved YOLOv26n network for feature extraction and enhancement. The backbone network of the improved YOLOv26n network is embedded with a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network contains a PyramidMaskMamba module. The frequency-selective convolution (SFSConv) module is used to perform collaborative feature extraction in the frequency and spatial domains on the intermediate feature map to enhance the discrete scattering characterization of the aircraft target and suppress speckle noise interference. At the same time, the Haar wavelet downsampling (HWD) module is used to retain high-frequency detail information during the feature map size reduction process, thereby alleviating the problem of feature loss for small-sized targets. The features processed by the backbone network are input into the PyramidMaskMamba module. The long-distance dependencies are captured through the state space model and combined with the adaptive mask generation mechanism to integrate the feature information of discrete scattering centers to enhance the overall expression of the aircraft target and suppress complex background interference. The method of capturing long-range dependencies through a state-space model and combining it with an adaptive mask generation mechanism to integrate the feature information of discrete scattering centers to enhance the overall representation of the aircraft target and suppress complex background interference includes: The input feature map is flipped along different directions to simulate the scanning path, and four-way scanning features are extracted by grouped convolution. The four-way scanning features are generated into adaptive directional masks in four directions through 1×1 convolution, and the weights of each direction are normalized to between 0 and 1. The four-directional scanning features are weighted and summed according to the normalized adaptive directional mask to focus on the features of the target boundary and the scattering center. Then, pyramid downsampling is performed through convolution with a stride of 2, and the long-distance dependency is modeled by a mask-guided two-dimensional state space scanning unit MGD_SS2D. The long-distance dependency is then modeled by transposed convolution upsampling and element-wise addition of the downsampled features at the same scale to fuse deep semantics and shallow details, thus obtaining the features after pyramid multi-scale fusion. The original input feature map is spliced ​​and fused with the features obtained by multi-scale fusion of the pyramid, and residual connections are added to retain the original feature information, so as to enhance the overall expression of the aircraft target and suppress interference from complex backgrounds. The STAL strategy of small target-aware label assignment is adopted to perform positive and negative sample matching between the predicted boxes and the ground truth boxes of the improved YOLOv26n network. In the improved YOLOv26n network, a scale-aware weight factor is introduced to dynamically adjust the matching threshold of small targets. The improved YOLOv26n network outputs SAR image aircraft target detection results including bounding box coordinates, class confidence, and class labels.

2. The image domain aircraft target detection method based on feature enhancement and multi-scale fusion according to claim 1, characterized in that, The frequency-selective convolution SFSConv module is used to perform joint feature extraction in the frequency and spatial domains on the intermediate feature map, including: The same input feature map is fed into the frequency decomposition branch and the scale-aware branch, which are set in parallel in the frequency-selective convolution SFSConv module. In the frequency decomposition branch, the input feature map is transformed to the frequency domain by Fast Fourier Transform (FFT), and the target learning frequency selection weights are used to separate the target scattering component from the speckle noise. In the scale-aware branch, the spatial domain multi-scale scattering representation features of the input feature map are extracted by parallel convolution branches with different dilation rates or convolution kernel scales, and adaptive fusion weights for each scale branch are generated by global average pooling and 1×1 convolution. The frequency domain scattering characterization features output by the frequency decomposition branch and the spatial domain multi-scale scattering characterization features output by the scale perception branch are input into the feature fusion branch for channel importance calculation and adaptive weighted summation to obtain the synergistic features of the frequency domain and spatial domain.

3. The image domain aircraft target detection method based on feature enhancement and multi-scale fusion according to claim 2, characterized in that, In the frequency decomposition branch, the input feature map is transformed to the frequency domain using Fast Fourier Transform (FFT), and the target scattering component and speckle noise are separated using target-learned frequency selection weights, including: A frequency domain filter kernel to be learned is generated using the kernel function of a fractional Gabor filter. The kernel function includes a frequency parameter that controls the frequency response of the filter, a phase parameter that controls the phase shift, an attenuation parameter that controls the attenuation amplitude of the Gaussian window, and a direction parameter that controls the filtering direction. After normalizing the generated frequency domain filter kernel to be learned, it is convolved with the input feature map of the scale-aware frequency selection convolution (SFSConv) module to obtain the separated target scattering component and speckle noise.

4. The image domain aircraft target detection method based on feature enhancement and multi-scale fusion according to claim 2, characterized in that, In the scale-aware branch, the spatial domain multi-scale scattering representation features of the input feature map are extracted using parallel convolutional branches with different dilation rates or convolutional kernel scales, including: Four parallel convolutional branches with different dilation rates are set up to cover the local scattering structure of aircraft targets of different sizes. Global average pooling is used to compress global context information, and 1×1 convolution is combined to generate adaptive fusion weights for each scale branch. The weights are mapped to the range of 0 to 1 to obtain the spatial domain multi-scale scattering representation features.

5. The image domain aircraft target detection method based on feature enhancement and multi-scale fusion according to claim 4, characterized in that, The four parallel convolution branches with different dilation rates are a 1×1 convolution, a 3×3 convolution with a dilation rate of 1, a 3×3 convolution with a dilation rate of 2, and a 3×3 convolution with a dilation rate of 3.

6. The image domain aircraft target detection method based on feature enhancement and multi-scale fusion according to claim 1, characterized in that, The method of using the Haar wavelet downsampling (HWD) module to retain high-frequency detail information during feature map size reduction includes: The input feature map is subjected to a first-order Haar discrete wavelet transform, which decomposes it into one low-frequency approximation component and three high-frequency detail components. The four components obtained from the decomposition are concatenated along the channel dimension to achieve lossless encoding of spatial information to the channel dimension, resulting in the concatenated high-channel features. The concatenated high-channel features are compressed to the target dimension by 1×1 convolution and then normalized and activated to obtain dimension-compressed channel features. Based on the number of channels in the channel features of dimensionality compression, we can choose whether to perform feature learning using depthwise separable convolution to further refine high-frequency detail features.

7. An image-domain aircraft target detection device based on feature enhancement and multi-scale fusion, characterized in that, include: The image acquisition module is used to acquire the SAR image to be detected; The feature extraction and enhancement module is used to input SAR images into the improved YOLOv26n network for feature extraction and enhancement. The backbone network of the improved YOLOv26n network is embedded with a scale-aware frequency-selective convolution (SFSConv) module and a Haar wavelet downsampling (HWD) module, and the neck network contains a PyramidMaskMamba module. The dual-domain collaborative and detail-preserving module is used to perform collaborative feature extraction in the frequency domain and spatial domain on the intermediate feature map using the frequency-selective convolution SFSConv module to enhance the discrete scattering characterization of the aircraft target and suppress speckle noise interference. At the same time, the Haar wavelet downsampling HWD module is used to preserve high-frequency detail information during the feature map size reduction process. The mask generation module is used to input the features processed by the backbone network into the PyramidMaskMamba module. It captures long-range dependencies through a state-space model and integrates the feature information of discrete scattering centers using an adaptive mask generation mechanism to enhance the overall representation of the aircraft target and suppress interference from complex backgrounds. Specifically, the step of capturing long-range dependencies through a state-space model and integrating the feature information of discrete scattering centers using an adaptive mask generation mechanism to enhance the overall representation of the aircraft target and suppress interference from complex backgrounds includes: The input feature map is flipped along different directions to simulate the scanning path, and four-way scanning features are extracted by grouped convolution. The four-way scanning features are generated into adaptive directional masks in four directions through 1×1 convolution, and the weights of each direction are normalized to between 0 and 1. The four-directional scanning features are weighted and summed according to the normalized adaptive directional mask to focus on the features of the target boundary and the scattering center. Then, pyramid downsampling is performed through convolution with a stride of 2, and the long-distance dependency is modeled by a mask-guided two-dimensional state space scanning unit MGD_SS2D. The long-distance dependency is then modeled by transposed convolution upsampling and element-wise addition of the downsampled features at the same scale to fuse deep semantics and shallow details, thus obtaining the features after pyramid multi-scale fusion. The original input feature map is spliced ​​and fused with the features obtained by multi-scale fusion of the pyramid, and residual connections are added to retain the original feature information, so as to enhance the overall expression of the aircraft target and suppress interference from complex backgrounds. The sample matching module is used to perform positive and negative sample matching between the predicted bounding boxes and the ground truth bounding boxes of the improved YOLOv26n network using the small target-aware label allocation STAL strategy. The improved YOLOv26n network introduces a scale-aware weight factor to dynamically adjust the matching threshold for small targets. The results output module is used to output SAR image aircraft target detection results containing bounding box coordinates, class confidence, and class labels on the improved YOLOv26n network.

8. An electronic device, characterized in that, The method includes a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the image domain aircraft target detection method based on feature enhancement and multi-scale fusion as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the image domain aircraft target detection method based on feature enhancement and multi-scale fusion as described in any one of claims 1 to 6.

Images