Panoramic sar image airplane target detection method, apparatus, device, and medium

Abstract

The application relates to a SAR image airplane target detection method, device and equipment and a storage medium. The method extracts corresponding prior data from a geographic prior database according to geographic position information of a SAR image, performs coarse detection on the SAR image according to a geographic polygon to obtain an airport coarse detection result, adopts an image registration method to obtain an airport fine detection result according to the airport coarse detection result and an optical image, performs target detection based on the airport fine detection result by using a trained rotating airplane target detection network to obtain a detection result, obtains airplane target category prior probability in a target detection frame according to a geographic space category probability graph, weights and modulates a classifier weight according to the airplane target category prior probability to obtain a prior-enhanced classifier weight, updates a target category probability according to the prior-enhanced classifier weight, and takes the updated target category probability and the target detection frame as a final target detection result. The method can improve the fine-grained detection performance of an airplane target in a panoramic SAR image.

Description

Technical Field

[0001] This application relates to the field of SAR image target recognition technology, and in particular to a method, apparatus, equipment and medium for detecting aircraft targets in panoramic SAR images. Background Technology

[0002] Synthetic Aperture Radar (SAR) has gained attention for its outstanding ability to capture high-resolution images under different weather conditions and has been widely used in various fields, especially in aircraft target detection and classification.

[0003] In recent years, deep learning technology has become a mainstream technology due to its ability to automatically extract image features, meeting the growing demand for automated data analysis. However, due to the unique coherent imaging mechanism, SAR images differ fundamentally from optical images, posing a series of specific challenges to aircraft detection and classification from SAR images, such as discrete appearance, attitude sensitivity, and complex background clutter interference. With the successful application of general target detection algorithms in computer vision, researchers have begun to introduce these algorithms into the task of aircraft target detection and classification in SAR images. Combining the characteristics of aircraft targets in SAR images, researchers have proposed a large number of high-performance algorithms.

[0004] When processing large-scale SAR images, airport detection is often used as a preprocessing step to identify candidate regions in order to improve the efficiency of aircraft target detection and classification. This workflow is based on a coarse-to-fine cognitive logic, first identifying potential airport areas, and then performing a fine-grained analysis of the differences between target instances. The timeliness and accuracy of airport detection are fundamental to aircraft detection and classification systems in SAR images. However, existing detection methods using deep learning suffer from several drawbacks, including the need for large amounts of labeled data, resulting in high training costs, limited robustness, and significant computational resource consumption. Summary of the Invention

[0005] Therefore, it is necessary to provide a panoramic SAR image aircraft target detection method, apparatus, equipment, and medium that can improve the fine-grained detection performance of aircraft targets in panoramic SAR images, addressing the aforementioned technical problems.

[0006] A method for aircraft target detection in panoramic SAR images, the method comprising:

[0007] Acquire the SAR image of the target to be detected;

[0008] Based on the geographic location information of the SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database.

[0009] Based on the geographic polygon, a coarse airport detection result is obtained by performing a coarse detection on the airport area in the SAR image. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0010] The trained rotating aircraft target detection network is used to perform target detection based on the fine detection results of the airport, and target detection boxes and target category probabilities are obtained. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0011] The prior probability of the aircraft category at the location of the target detection box is obtained based on the geospatial category probability map. The prior probability of the aircraft category is then used to weight and modulate the posterior probability generated by the detection head to obtain the prior-enhanced classifier weights.

[0012] The target category probability is updated based on the prior enhanced classifier weights, and the target detection result is obtained based on the updated target category probability and the target detection box.

[0013] In one embodiment, obtaining the airport fine inspection result by image registration based on the airport coarse inspection result and the optical image includes:

[0014] After downsampling the airport coarse inspection results and optical images, image registration is performed using an image feature description algorithm;

[0015] After image registration, the precise geographical location information in the coarse airport detection results is used to obtain the fine airport detection results.

[0016] In one embodiment, before performing target detection on the airport fine-grained detection results using the trained rotating aircraft target detection network:

[0017] The airport fine detection results are processed according to a preset sliding window strategy to obtain a sequence slice dataset including multiple slices;

[0018] The sequence slice dataset is input into the trained rotating aircraft target detection network for target detection.

[0019] In one embodiment, in the trained rotating aircraft target detection network:

[0020] The backbone network automatically extracts features from the sequence slice dataset to obtain feature maps of multiple sizes.

[0021] The recursive gated spatial interaction network adaptively interacts and fuses the feature maps of the three smaller sizes to obtain three corresponding fused features.

[0022] The detection head performs target detection based on the three fused features to obtain the target detection box and the target category probability.

[0023] In one embodiment, in the recursive gated spatial interaction network:

[0024] A block recursive filter is used to perform top-down feature fusion from the feature map with the smallest size to the feature map with the largest size, to obtain the bottom-level fused features and the intermediate fused features.

[0025] The block recursive filter is used to perform bottom-up feature fusion on the bottom-level fused features, intermediate fused features, and minimum-size feature maps in sequence to obtain the middle-level fused features and the high-level fused features.

[0026] The high-level fusion feature, the mid-level fusion feature, and the low-level fusion feature are used as the three fusion features input to the detection head.

[0027] In one embodiment, in the block recursive filter, two branches are used to extract features from the input data respectively, and the extracted features are fused and output.

[0028] In one of the branches, a recursive gate mechanism is used to gradually extract and integrate the local and global information of the input data to obtain feature representation through linear projection and iteration of local-global filtering blocks.

[0029] In one embodiment, after defining a coordinate system and resolution in the prior geographic data, a two-dimensional Gaussian distribution is used to model the probability distribution of aircraft categories in the prior geographic data to obtain the geospatial category probability map.

[0030] This application provides a panoramic SAR image aircraft target detection device, the device comprising:

[0031] The SAR image acquisition module is used to acquire SAR images of the targets to be detected.

[0032] The prior data extraction module is used to extract the corresponding geographic polygons, optical images, and geospatial category probability maps from the geographic location information of the SAR image from the geographic prior database.

[0033] The airport detection module is used to perform coarse detection of airport areas in the SAR image based on the geographic polygon to obtain coarse airport detection results, and to obtain fine airport detection results by image registration based on the coarse airport detection results and the optical image.

[0034] The target detection module is used to perform target detection based on the fine detection results of the airport using a trained rotating aircraft target detection network, and to obtain target detection boxes and target category probabilities. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0035] The classifier weight update module is used to obtain the prior probability of the aircraft category at the location of the target detection box based on the geospatial category probability map, and to perform weighted modulation on the posterior probability generated by the detection head based on the prior probability of the aircraft category to obtain the prior-enhanced classifier weights.

[0036] The target detection prediction result acquisition module is used to update the target category probability based on the prior enhanced classifier weights, and obtain the target detection result based on the updated target category probability and the target detection box.

[0037] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0038] Acquire the SAR image of the target to be detected;

[0039] Based on the geographic location information of the SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database.

[0040] Based on the geographic polygon, a coarse airport detection result is obtained by performing a coarse detection on the airport area in the SAR image. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0041] The trained rotating aircraft target detection network is used to perform target detection based on the fine detection results of the airport, and target detection boxes and target category probabilities are obtained. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0042] The prior probability of the aircraft category at the location of the target detection box is obtained based on the geospatial category probability map. The prior probability of the aircraft category is then used to weight and modulate the posterior probability generated by the detection head to obtain the prior-enhanced classifier weights.

[0043] The target category probability is updated based on the prior enhanced classifier weights, and the target detection result is obtained based on the updated target category probability and the target detection box.

[0044] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0045] Acquire the SAR image of the target to be detected;

[0046] Based on the geographic location information of the SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database.

[0047] Based on the geographic polygon, a coarse airport detection result is obtained by performing a coarse detection on the airport area in the SAR image. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0048] The trained rotating aircraft target detection network is used to perform target detection based on the fine detection results of the airport, and target detection boxes and target category probabilities are obtained. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0049] The prior probability of the aircraft category at the location of the target detection box is obtained based on the geospatial category probability map. The prior probability of the aircraft category is then used to weight and modulate the posterior probability generated by the detection head to obtain the prior-enhanced classifier weights.

[0050] The target category probability is updated based on the prior enhanced classifier weights, and the target detection result is obtained based on the updated target category probability and the target detection box.

[0051] The aforementioned panoramic SAR image aircraft target detection method, apparatus, equipment, and medium extract corresponding geographic polygons, optical images, and geospatial category probability maps from a geographic prior database based on the geographic location information of the SAR image to be detected. Coarse airport detection results are obtained by performing coarse detection on the airport area in the SAR image based on the geographic polygons. Fine airport detection results are obtained by image registration based on the coarse airport detection results and the optical image. A trained rotating aircraft target detection network is then used to perform target detection based on the fine airport detection results. Prior probability of the aircraft category within the target detection box is obtained based on the geospatial category probability. Weighted modulation of the classifier weights in the detection head is performed based on the prior aircraft category probability to obtain prior-enhanced classifier weights. The target category probability is updated based on the prior-enhanced classifier weights. Finally, the target detection result is obtained based on the updated target category probability and the target detection box. This method can improve the fine-grained detection performance of aircraft targets in panoramic SAR images. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a panoramic SAR image aircraft target detection method in one embodiment;

[0053] Figure 2 This is a schematic diagram of the overall architecture of a panoramic SAR image aircraft target detection method in one embodiment;

[0054] Figure 3 This is a visualization diagram of airport detection based on different types of bounding boxes in one embodiment, wherein, Figure 3 (a) Displays the horizontal bounding box superimposed on the optical image. Figure 3 (b) shows the geographic polygon bounding box generated by overlaying an optical image;

[0055] Figure 4 This is a flowchart illustrating a geospatial-driven airport detection method in one embodiment, wherein... Figure 4 (a) Shows the coarse airport detection phase using geographic polygon prior matching. Figure 4 (b) Demonstrates a refined airport detection phase based on geographic information matching and RIFT;

[0056] Figure 5 This is a schematic diagram of the structure of a rotating aircraft target detection network in one embodiment;

[0057] Figure 6 This is a schematic diagram of the internal structure of the SepBN detection head in one embodiment;

[0058] Figure 7 This is a schematic diagram of the block recursive filter structure in one embodiment;

[0059] Figure 8 This is a flowchart illustrating a relation-based prior reasoning method in one embodiment, wherein... Figure 8 (a) illustrates the process of constructing the geospatial category probability map. Figure 8 (b) Demonstrates the construction process of the post-processing model;

[0060] Figure 9 This is a schematic diagram illustrating the aircraft category probabilities modeled based on a Gaussian distribution in one embodiment, wherein... Figure 9 (a) represents a slice image containing three aircraft. Figure 9 (b) A three-dimensional plot showing the probability distribution of aircraft target position and category based on Gaussian distribution modeling;

[0061] Figure 10 This is a visualization of a geospatial category probability map at an airport in one embodiment, wherein... Figure 10 (a) is a SAR image of the airport. Figure 10 (b), (c), and (d) represent the geospatial category probability maps corresponding to the three different types of aircraft, respectively.

[0062] Figure 11This is a schematic diagram of the prior probability distribution of categories on a geospatial category probability map in one embodiment;

[0063] Figure 12 This is a structural block diagram of a panoramic SAR image aircraft target detection device in one embodiment;

[0064] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] Airport detection is often used as a preprocessing step to reduce the scope of image processing when detecting aircraft targets in panoramic SAR images. However, there are problems with using deep neural networks for airport detection, including: a large demand for labeled data, as high-performance deep learning methods typically require large labeled datasets for effective training. However, obtaining such large-scale and accurately labeled datasets is both challenging and costly. Limited robustness, while deep learning-based methods perform well on training and testing data under similar acquisition geometries, they may struggle to effectively detect airports when faced with images of different resolutions and bands. Significant computational resource consumption, as deep learning methods use complex deep neural networks, which, despite achieving high accuracy, are typically accompanied by enormous computational intensity and "black box" characteristics, potentially limiting their practical application.

[0067] To address the above issues, this application, such as Figure 1 As shown, a method for aircraft target detection in panoramic SAR images is provided, including the following steps:

[0068] In step S100, the SAR image to be detected is acquired.

[0069] In step S110, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database based on the geographic location information of the SAR image.

[0070] In step S120, a coarse detection of the airport area in the SAR image is performed based on the geographic polygon to obtain a coarse airport detection result. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0071] Step S130: Target detection is performed using the trained rotating aircraft target detection network based on the airport's detailed detection results, resulting in target detection boxes and target category probabilities. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0072] Step S140: Obtain the prior probability of the aircraft category at the location of the target detection box based on the geospatial category probability map, and perform weighted modulation on the posterior probability generated by the detection head based on the prior probability of the aircraft category to obtain the prior-enhanced classifier weights.

[0073] Step S150: Update the target category probability according to the prior enhanced classifier weights, and obtain the target detection result based on the updated target category probability and the target detection box.

[0074] This application first utilizes geographic location priors and geographic information matching techniques, combined with optical image and SAR image registration techniques, to realize a training-free rapid airport detection method, laying the foundation for subsequent fine-grained aircraft target detection. Second, considering the scattering characteristics of aircraft targets, a fine-grained rotating aircraft target detection network is designed, capable of accurately capturing important aircraft target features in SAR images. Finally, a relation-based prior reasoning strategy is proposed, which, through the combination of deep networks and geospatial category relationships, effectively reduces category prediction bias and improves the interpretability and credibility of decision-making.

[0075] In step S100, the image to be detected is a SAR image obtained when the target is detected from high altitude to the ground. The scenario in which the technical solution of this application is applied is to detect target aircraft in an airport, and the SAR image is a large-scale SAR image.

[0076] In step S110, based on the geographic location information inherent in the acquired SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map can be extracted from a pre-set geographic prior database.

[0077] In this embodiment, the geographic prior database is pre-set and stores multiple expectedly usable airport geographic information, geographic polygons describing the geometric outlines of the airports, optical remote sensing image information, and geospatial category probability maps generated based on multiple historical data for each airport. The airport geographic information can be used as an index for relevant data for each airport, and the corresponding data can be extracted from the geographic prior database using the geographic location information inherent in the SAR image obtained in step S100. The geographic polygons and optical image information are used to detect airport areas in the SAR image, while the geospatial category probability map is used to further improve the prediction accuracy of the target detection neural network in image post-processing. The geospatial category probability map will be specifically explained in step S140.

[0078] In the scenario where this method is applied, airport detection, as a crucial part of aircraft detection in large-scale SAR images, has become a research hotspot. Runways, as the core infrastructure of an airport, are important ground features in SAR images, making runway detection an effective method for airport location. However, traditional methods mainly rely on the specular reflection characteristics of airport runways to radar waves and their unique geometric features to extract runways. These methods typically include image segmentation-based methods, line-based methods, and saliency-based methods. Overall, traditional methods can achieve satisfactory performance in specific scenarios. However, several challenges exist: complex parameter tuning, limited robustness and generalization make them difficult to handle complex scenarios, leading to inaccurate detection.

[0079] Deep learning-based detection methods typically require large amounts of labeled data and high-performance graphics processing units (GPUs) for model training, which involves high costs for manual labeling and significant hardware resource consumption. Furthermore, the black-box nature of deep learning-based methods is also a limiting factor in their application.

[0080] In this method, considering the above-mentioned issues, airport-specific geographic polygons are used in step S120. This method more closely approximates the actual shape of the airport, further eliminating potential false alarms around the airport (for example, buildings outside the airport may be mistaken for aircraft in SAR images). Furthermore, to address unavoidable positioning errors during imaging, this method also introduces the RIFT algorithm, which effectively mitigates geographic information bias through precise registration of optical and SAR images. This method makes airport area extraction more accurate, providing a more reliable data foundation for subsequent airport detection and analysis.

[0081] like Figure 2As shown, geospatial-driven airport detection serves as the initial stage, receiving panoramic SAR images as input. Geographic information extracted from the SAR images is seamlessly matched with geographic polygons in the database, enabling rapid airport detection globally. Furthermore, the integration of the RIFT algorithm facilitates geographic recalibration, mitigating inherent geographic biases in the imaging system and ensuring accurate airport detection.

[0082] In this embodiment, over time, the airport possesses a relatively consistent and stable infrastructure. Based on this knowledge, geographic polygons are created using geographic big data analysis and processing software, open-source satellite imagery, and geographic location information of the airport area of ​​interest. These geographic polygons and corresponding optical images are stored as geographic priors in a geographic information database. Figure 3 As shown, geographic polygons capture the geometric contours of airports more accurately than traditional horizontal bounding boxes, further reducing the introduction of irrelevant background information. However, due to inherent positional biases during imaging, direct airport detection based on geographic information matching may lead to inaccurate location predictions and the neglect of aircraft within the airport. To address this issue, a geospatial-driven airport detection method is proposed, featuring a coarse-to-fine detection process.

[0083] In this embodiment, geographic information matching is first performed between the geographic information in the SAR image and the geographic polygons in the geographic database to obtain a coarse detection result of the airport area in the SAR image. Then, based on the coarse detection result of the airport and the corresponding optical image, image registration is used to obtain a fine detection result of the airport. Specifically, obtaining the fine detection result of the airport includes: after downsampling the coarse detection result of the airport and the optical image, image registration is performed using an image feature description algorithm. After image registration, the precise geographic location information in the coarse detection result of the airport is used to obtain the fine detection result of the airport.

[0084] Specifically, firstly, geographic polygon matching is performed based on a geographic database to quickly estimate a rough airport area and correlate it with the corresponding optical image. Then, the RIFT algorithm is used to recalibrate the geographic information in the SAR image. In RIFT, feature matching is performed between the optical and SAR images, and corresponding points are identified. Simultaneously, image downsampling is used to optimize the algorithm's efficiency. Then, the geographic information from the optical image is combined with the extracted corresponding points, and geographic transformation parameters are calculated in the SAR image to obtain refined geographic information. This process contributes to more accurate airport detection results, thus significantly improving the efficiency of fine-grained aircraft detection. The method flow is as follows: Figure 4 As shown.

[0085] like Figure 2As shown, after obtaining the detailed airport detection results, i.e. the accurate airport area, the rotating aircraft target detection network serves as the second stage to quickly obtain preliminary position and category predictions.

[0086] In step S130, before inputting the airport fine detection results into the trained rotating aircraft target detection network: the airport fine detection results are processed according to a preset sliding window strategy to obtain a sequence slice dataset including multiple slices, and the sequence slice dataset is input into the trained rotating aircraft target detection network for target detection.

[0087] In step S130, the trained rotating aircraft target detection network structure, such as Figure 5 As shown, the backbone network automatically processes the input sequence slice dataset, extracting multiple feature maps of different sizes. A recursive gated spatial interaction network adaptively interacts and fuses the three smaller feature maps to obtain three corresponding fused features. The detection head performs target detection based on the three fused features, obtaining the target detection box and the target class probability.

[0088] In this embodiment, the backbone network uses CSPNEXt, whose main task is to extract image features from the input slice image, obtaining multi-level, multi-scale feature maps P0 to P4. A recursive gated spatial interaction network (RSIM) is used to process the three smaller feature maps, P2, P3, and P4, to obtain a more comprehensive representation of the aircraft target features through adaptive spatial interaction modeling. Finally, the interactively obtained feature maps P0, P2, P3, and P4 are processed. o2 ~P o4 Input the SepBN detector head for classification and regression to obtain bounding box location and category prediction.

[0089] It should be noted that the backbone network and the detection head can also use other networks with the same capabilities, and are not limited to the two mentioned above.

[0090] In this embodiment, prioritizing the efficiency and compatibility of the backbone and neck, the rotating aircraft target detection network structure adopts the target detection network RTMDet.

[0091] Furthermore, inspired by large-kernel convolutions, RTMDet introduces 5×5 depthwise convolutions and 1×1 pointwise convolutions in the basic building blocks and neck of YOLOX to increase the effective receptive field, thereby achieving more comprehensive global information modeling. To maintain a balance between speed and accuracy, on the one hand, the model's depth and width are carefully designed, which greatly enhances the model's ability to identify and locate targets. On the other hand, a SepBN detection head is used to obtain the bounding boxes and class predictions of the targets.

[0092] Furthermore, the internal structure of the SepBN detection head is as follows: Figure 6 As shown, the 3×3 convolutional layer achieves weight parameter sharing and calculates batch normalization (BN) separately on different feature maps, reducing the number of parameters and computational complexity. Next, three parallel 1×1 convolutional layers predict the target position, angle, and class confidence, respectively.

[0093] It should be noted that RTMDet is preferred as the main framework of the rotating aircraft target detection network in this method because of its high accuracy and speed. However, other target detection network structures can be used as the main framework for aircraft target detection networks in different application scenarios, and this method is not limited to RTMDet.

[0094] However, further consideration of the inherent complexity of SAR images, such as the discrete appearance, attitude sensitivity, and scale heterogeneity of aircraft targets, poses a significant challenge to efficient feature extraction. These challenges result in substantial false alarms and missed detections when directly using mainstream general-purpose deep learning algorithms for fine-grained aircraft detection tasks in SAR images, urgently requiring further efforts to improve their accuracy.

[0095] To address the aforementioned issues, considering the discreteness and scale heterogeneity of aircraft targets, multi-scale spatial information interaction is crucial for detecting fine-grained aircraft from SAR images. While self-attention mechanisms can effectively simulate spatial interactions in visual data, their enormous computational resources hinder their widespread application. To overcome this obstacle, this method proposes a recursively gated spatial interaction network, RSIM.

[0096] like Figure 5 As shown, RSIM demonstrates a bidirectional feature fusion strategy that effectively establishes a complementary relationship between low-level details and high-level semantic information. The Block Recursive Filter (BRF), as a core component of RSIM, effectively encodes the relationships between discrete scattering features of aircraft targets through global information modeling and efficient adaptive spatial interaction, thereby providing robust knowledge for the detection head.

[0097] In this embodiment, in the recursive gated spatial interaction network: firstly, a block recursive filter is used to perform top-down feature fusion from the smallest feature map (high-level features) to the largest feature map (low-level features), obtaining intermediate fused features and low-level fused features. Then, a block recursive filter is used to sequentially perform bottom-up feature fusion of the low-level fused features, intermediate fused features, and high-level features, obtaining mid-level fused features and high-level fused features. Finally, the high-level fused features, mid-level fused features, and low-level fused features are input into the detection head for multi-scale prediction, outputting the target detection result and category prediction.

[0098] refer to Figure 5In the recursive gated spatial interaction network structure, during top-down feature fusion: the smallest feature map P4 among the three smaller features is upsampled after passing through a 1×1 convolutional layer and concatenated with the middle-sized feature map P3 via channel concatenation. This is then fused through a block recursive filter, followed by another 1×1 convolutional layer to obtain the intermediate fused feature. This intermediate fused feature is then upsampled and concatenated with the largest feature map P2 via channel concatenation, fused through a block recursive filter, and finally passed through a 3×3 convolutional layer to obtain the bottom-level fused feature P. o2 .

[0099] Next, during bottom-up feature fusion: the bottom layer fuses feature P o2 After downsampling, the features are concatenated with the intermediate fusion features and then fused through a block recursive filter. Finally, the fusion is performed through a 3×3 convolutional layer to obtain the intermediate fusion feature P. o3 .

[0100] Next, the middle layer fusion feature P o3 After downsampling, the feature map P4 after passing through a 1×1 convolutional layer is concatenated by channel, then fused using a block recursive filter, and finally passed through a 3×3 convolutional layer to obtain the high-level feature fusion P. o4 .

[0101] In this embodiment, in the block recursive filter, two branches are used to extract features from the input data respectively, and the extracted features are fused and output. In one of the branches, a recursive gate mechanism is used to gradually extract and integrate the local and global information of the input data to obtain feature representation through linear projection and iteration of local-global filter blocks.

[0102] like Figure 7 As shown, in the block recursive filter, the input feature map is divided into two distinct branches along the channel dimension and operated independently. Then, the features extracted from the different branches are effectively fused to enhance feature information representation while reducing computational cost. In one branch, a recursive gate mechanism is employed, iterating through linear projection and local-global filtering (LGF) blocks to progressively extract and integrate local and global information, obtaining a finer-grained feature representation.

[0103] Furthermore, let the input features X of the block recursive filter be... in and output feature X out The operations in a block recursive filter can be represented as:

[0104]

[0105] In the above formula, φ proj (·) is the projection function, S k(k = 0, 1, 2) represents a set of projected features derived from the input features S. L(·) and g(·) represent the LGF block and feature dimension matching function, respectively. ⊙ is the Hadamard product. 1×1 (·) and concat(·) are 1×1 convolution and channel concatenation operations, respectively.

[0106] In this embodiment, the Local-Global Filter (LGF) block integrates local and global perspectives to enhance the network's ability to capture fine-grained features of aircraft targets in SAR images. For example... Figure 7 As shown in (a), in the global branch, a two-dimensional Fast Fourier Transform (FFT) is used to transform spatial features to the frequency domain. Then, a trainable global filter is used to learn features in the frequency domain, which is equivalent to performing a global convolution in the spatial domain, establishing a global receptive field, and comprehensively capturing feature distributions and patterns. Furthermore, the processed features are returned to the spatial domain through a two-dimensional Inverse Fourier Transform (IFFT) to maintain spatial consistency and feature integrity. In the local branch, a 3×3 depthwise convolution is used to focus on local features. Finally, 1×1 convolutions and channel concatenation are performed on the local and global information to improve the network's ability to capture aircraft targets at different scales.

[0107] In this embodiment, the network is optimized by minimizing the loss during training. During the testing and actual detection phases, a non-maximum suppression algorithm for rotation is used to remove overlapping bounding boxes to obtain prediction results.

[0108] like Figure 2 As shown, after obtaining the preliminary prediction results of the location and category of aircraft targets, in the third stage, a geospatial category probability map is constructed using the ground truth labels in historical data, which organizes the relationship between the parking locations of aircraft in the airport and the corresponding aircraft target categories.

[0109] Because aircraft targets in SAR images have discrete appearances and are sensitive to attitude, potential aircraft targets may exhibit low confidence levels and be very similar to background clutter or other similar aircraft types, leading to classification ambiguity. To overcome this obstacle, a relation-based prior reasoning strategy is proposed. This strategy organizes prior knowledge about geospatial category relationships and integrates it into the classification process to improve the performance of CNNs, such as... Figure 8 As shown, relation-based prior reasoning comprises two main components: a geospatial category probability graph and a post-processing model.

[0110] In this embodiment, considering that aircraft are typically parked in specific locations and orientations at airports, different types of aircraft are assigned to designated areas to optimize operational efficiency. Therefore, the probability of encountering a specific type of aircraft at a specific location can be reasonably inferred, enabling effective target category prediction. In this method, a probability map is used to represent the geospatial category relationships of different aircraft types.

[0111] Furthermore, when creating the geospatial category probability map, after defining the coordinate system and resolution of the geospatial category probability map, a two-dimensional Gaussian distribution is used to model the category probability distribution of aircraft targets in the prior geographic data to obtain the geospatial category probability map.

[0112] Specifically, when constructing the geospatial category probability map for each airport of interest, the probability relationship between aircraft category and geospatial coordinates is obtained from historical images associated with each airport. Determining the coordinate system and map resolution is crucial in constructing the geospatial category probability map. First, this paper adopts a latitude-longitude coordinate system, which better reflects the expression of geographic information. Then, latitude and longitude, using a distance of 5 meters from the ground surface as an example, are used as the axis unit for detailed representation. Therefore, the longitude resolution R of the geospatial category probability map is... lon and latitude resolution R lat The calculation is as follows:

[0113]

[0114]

[0115] In the above formula, R is the radius of the Earth, (lon1, lat1, lon2, lat2) are the latitude and longitude range of the airport, and La is the latitude of the airport's regional center. S lon and S lat These are the surface distances along the longitude and latitude directions of the airport, respectively. λ is the surface resolution of the map, which is set to 5m by default.

[0116] After defining the coordinate system and resolution of the geospatial category probability map, a two-dimensional Gaussian distribution (GD) is used to model the aircraft category probability distribution. While there are other alternative methods for modeling geospatial probability maps, the Gaussian distribution has unique advantages in resolving complex spatial relationships. On the one hand, the well-understood nature of GD makes it well-suited for probabilistic modeling. On the other hand, despite inherent positioning biases in SAR images, the continuous probability distribution of GD helps to effectively characterize aircraft distribution information. Specifically, the center longitude and latitude of the ground truth boxes in the training set are used as reference coordinates. Then, by calculating the deviation between the provided longitude and latitude coordinates and the reference coordinates, the aircraft category probability distribution is determined, such as... Figure 9 As shown. The calculation formula is as follows:

[0117]

[0118] In the above formula, P(C i |X,G) is a category C i The probability density function (PDF) at position G, (G x G y These are the geographic coordinates (i.e., longitude and latitude) corresponding to the location (x, y) on the input image. (Box) ctrlon Box ctrlon ) represents the center longitude and latitude of the truth box, and δ represents the standard deviation.

[0119] Building upon this foundation, as historical images are progressively traversed, Gaussian distributions are overlaid based on the aircraft's type and location. This results in a geospatial category probability map with multiple Gaussian distributions, where aircraft parking locations have higher probability values, while irrelevant areas such as the background have lower probability values. This allows for a comprehensive summary of the geospatial context relationships between different types of aircraft in historical images.

[0120] Figure 10 This is the geospatial category probability map for Airport I. Figure 10 In the geospatial category probability map, each grid represents the result of the superposition of multiple Gaussian distributions associated with its neighboring targets in multiple historical images.

[0121] Through the above construction process, a geospatial category probability map is constructed for airports of interest based on relevant historical data, and the map is pre-stored in a geographic prior database.

[0122] In this embodiment, after extracting the geospatial category probability map related to the SAR image to be detected from the geographic prior database, in step S140, the posterior probability of the classifier weights in the detection head and the geographic category prior probability are modulated and enhanced according to the geospatial category probability map, thereby improving the accuracy and stability of the algorithm's classification prediction. The calculation formula is as follows:

[0123]

[0124] In the above formula, X is the input image, and C i This represents the i-th class. β is the weighting factor. On the geospatial category probability map, it is located at geographical location G, category C. i The accumulation of prior probabilities, the calculation principle diagram is as follows: Figure 11 As shown.

[0125] Finally, in step S150, the target category probability is updated according to the prior enhanced classifier weights, and the accurate aircraft target detection result is obtained based on the updated target category probability and the target detection box.

[0126] In this paper, the effectiveness of the proposed method (the proposed method) is also verified through experiments. The experimental results are demonstrated by comparing the performance of the proposed method with mainstream target detection algorithms. For fairness, the principle of controlled variables is adopted, and the proposed geospatial-driven airport detection algorithm is used for preprocessing. Accuracy and speed are quantified in mAP and FPS, respectively, as detailed in Table 1.

[0127] Table 1 Comparison of detection performance with state-of-the-art methods

[0128]

[0129] According to Table 1, RoI Transformer and Oriented R-CNN demonstrate strong performance in fine-grained aircraft detection of SAR images, achieving mAPs of 68.5% and 69.2%, respectively. In contrast, single-stage algorithms such as RetinaNet-OBB and OrientedRepPoints offer better efficiency than two-stage detection algorithms but exhibit weaker detection performance. RTMDet, a high-precision, low-latency single-stage detector, was developed by integrating insights from the YOLO series and other advanced detectors. Therefore, RTMDet performs well among mainstream algorithms, achieving an mAP of 74.7%. Compared to mainstream algorithms, the proposed method is competitive, with an mAP of 81.2%. Furthermore, the proposed method also demonstrates good timeliness in terms of FPS.

[0130] In the aforementioned SAR image aircraft target detection methods, a novel geospatial context prior knowledge reasoning framework is proposed for fine-grained aircraft target detection in SAR images. This framework leverages geographic location priors, the characteristics of aircraft targets in SAR images, and the geospatial category relationship between airports and aircraft. First, a geospatial-driven airport detection model utilizes geographic location priors for fast airport detection. Then, the rotating aircraft target detection network RTMDet-RSIM generates high-level representations of aircraft with different orientations and sizes, thereby improving the performance of fine-grained aircraft detection. Finally, a reasoning strategy based on relational priors is proposed to improve classification performance by modeling the geospatial category relationships of historical scenes. Experiments demonstrate the superiority of the proposed method and its potential for practical applications.

[0131] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0132] In one embodiment, such as Figure 12 As shown, a panoramic SAR image aircraft target detection device is provided, comprising: a SAR image acquisition module 200, a prior data extraction module 210, an airport detection module 220, a target detection module 230, a classifier weight update module 240, and a target detection prediction result acquisition module 250, wherein:

[0133] SAR image acquisition module 200 is used to acquire SAR images of targets to be detected;

[0134] The prior data extraction module 210 is used to extract the corresponding geographic polygon, optical image and geospatial category probability map from the geographic prior database based on the geographic location information of the SAR image.

[0135] Airport detection module 220 is used to perform coarse detection of airport areas in the SAR image based on the geographic polygon to obtain coarse airport detection results, and to obtain fine airport detection results by image registration based on the coarse airport detection results and the optical image.

[0136] The target detection module 230 is used to perform target detection based on the fine detection results of the airport using a trained rotating aircraft target detection network, and to obtain target detection boxes and target category probabilities. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0137] The classifier weight update module 240 is used to obtain the prior probability of the aircraft category at the location of the target detection box based on the geospatial category probability map, and to perform weighted modulation on the classifier weights in the detection head based on the prior probability of the aircraft category to obtain the prior-enhanced classifier weights.

[0138] The target detection prediction result acquisition module 250 is used to update the target category probability according to the prior enhanced classifier weights, and obtain the target detection result based on the updated target category probability and the target detection box.

[0139] Specific limitations regarding the panoramic SAR image aircraft target detection device can be found in the limitations of the panoramic SAR image aircraft target detection method described above, and will not be repeated here. Each module in the aforementioned panoramic SAR image aircraft target detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0140] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores geographic prior knowledge database data. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a SAR image aircraft target detection method.

[0141] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0142] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0143] Acquire the SAR image of the target to be detected;

[0144] Based on the geographic location information of the SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database.

[0145] Based on the geographic polygon, a coarse airport detection result is obtained by performing a coarse detection on the airport area in the SAR image. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0146] The trained rotating aircraft target detection network is used to perform target detection based on the fine detection results of the airport, and target detection boxes and target category probabilities are obtained. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0147] The prior probability of the aircraft category at the location of the target detection box is obtained based on the geospatial category probability map. The prior probability of the aircraft category is then used to weight and modulate the posterior probability generated by the detection head to obtain the prior-enhanced classifier weights.

[0148] The target category probability is updated based on the prior enhanced classifier weights, and the target detection result is obtained based on the updated target category probability and the target detection box.

[0149] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0150] Acquire the SAR image of the target to be detected;

[0151] Based on the geographic location information of the SAR image, the corresponding geographic polygon, optical image, and geospatial category probability map are extracted from the geographic prior database.

[0152] Based on the geographic polygon, a coarse airport detection result is obtained by performing a coarse detection on the airport area in the SAR image. Based on the coarse airport detection result and the optical image, an image registration method is used to obtain a fine airport detection result.

[0153] The trained rotating aircraft target detection network is used to perform target detection based on the fine detection results of the airport, and target detection boxes and target category probabilities are obtained. The trained rotating aircraft target detection network includes a backbone network, a recursive gating spatial interaction network, and a detection head.

[0154] The prior probability of the aircraft category at the location of the target detection box is obtained based on the geospatial category probability map. The prior probability of the aircraft category is then used to weight and modulate the posterior probability generated by the detection head to obtain the prior-enhanced classifier weights.

[0155] The target category probability is updated based on the prior enhanced classifier weights, and the target detection result is obtained based on the updated target category probability and the target detection box.

[0156] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting aircraft targets in panoramic SAR images, characterized in that, The method comprises: acquiring a SAR image to be subjected to target detection; extracting a corresponding geographic polygon, optical image and geographic spatial category probability map from a geographic priori database according to geographic position information of the SAR image; coarsely detecting an airport region in the SAR image according to the geographic polygon to obtain an airport coarse detection result, and using image registration to obtain an airport fine detection result according to the airport coarse detection result and the optical image; using a trained rotating airplane target detection network to perform target detection based on the airport fine detection result to obtain a target detection frame and a target category probability, wherein the trained rotating airplane target detection network comprises a backbone network, a recursive gate spatial interaction network and a detection head; obtaining an airplane category priori probability of the target detection frame position according to the geographic spatial category probability map, and weighting and modulating a posteriori probability generated by the detection head according to the airplane category priori probability to obtain a priori enhanced classifier weight; updating the target category probability according to the priori enhanced classifier weight, and obtaining a target detection result according to the updated target category probability and the target detection frame.

2. The panoramic SAR image aircraft target detection method according to claim 1, characterized in that, The method of using image registration to obtain the airport fine detection result according to the airport coarse detection result and the optical image comprises: after performing down-sampling processing on the airport coarse detection result and the optical image, using an image feature description algorithm to perform image registration; after the image registration, using accurate geographic position information in the airport coarse detection result to obtain the airport fine detection result.

3. The panoramic SAR image aircraft target detection method according to claim 1, characterized in that, Before using the trained rotating airplane target detection network to perform target detection on the airport fine detection result: processing the airport fine detection result according to a preset sliding window strategy to obtain a sequence slice data set comprising multiple slices; inputting the sequence slice data set into the trained rotating airplane target detection network to perform target detection.

4. The panoramic SAR image aircraft target detection method according to claim 3, characterized in that, In the trained rotating airplane target detection network: the backbone network performs automatic feature extraction on the sequence slice data set to obtain multiple size feature maps; the recursive gate spatial interaction network performs adaptive interaction fusion on three smaller size feature maps to obtain corresponding three fusion features; the detection head performs target detection according to the three fusion features to obtain the target detection frame and the target category probability.

5. The panoramic SAR image aircraft target detection method according to claim 4, characterized in that, In the recursive gate spatial interaction network: a block recursive filter is used to perform top-down feature fusion from the smallest size feature map to the largest size feature map to obtain bottom layer fusion features and intermediate layer fusion features; the block recursive filter is used to sequentially perform bottom-up feature fusion on the bottom layer fusion features, the intermediate layer fusion features and the smallest size feature map to obtain middle layer fusion features and high layer fusion features; the high layer fusion features, the middle layer fusion features and the bottom layer fusion features are used as input three fusion features of the detection head.

6. The panoramic SAR image aircraft target detection method according to claim 5, characterized in that, In the block recursive filter, two branches are used to extract features of input data respectively, and the extracted features are fused and output. In one of the branches, a recursive gating mechanism is employed to progressively extract and integrate local and global information of the input data through linear projection and iteration of local-global filtering blocks to obtain the feature representation.

7. The panoramic SAR image aircraft target detection method according to any one of claims 1-6, characterized in that, By defining a coordinate system and resolution in the prior geographical data, a two-dimensional Gaussian distribution is used to model the category probability distribution of the aircraft in the prior geographical data to obtain the geographical category probability map.

8. An apparatus for detecting an airplane target in a panoramic SAR image, characterized by comprising: The device comprises: An SAR image acquisition module, configured to acquire an SAR image to be subjected to target detection; An prior data extraction module, configured to extract corresponding geographical polygons, optical images and geographical category probability maps from a geographical prior database according to geographical position information of the SAR image; An airport detection module, configured to perform coarse detection on an airport region in the SAR image according to the geographical polygons to obtain an airport coarse detection result, and to obtain an airport fine detection result by using image registration based on the airport coarse detection result and the optical images; A target detection module, configured to perform target detection based on the airport fine detection result by using a trained rotating aircraft target detection network to obtain a target detection frame and a target category probability, wherein the trained rotating aircraft target detection network comprises a backbone network, a recursive gating spatial interaction network and a detection head; A classifier weight updating module, configured to obtain an aircraft category prior probability of the target detection frame position according to the geographical category probability map, and to weight and modulate a posterior probability generated by the detection head according to the aircraft category prior probability to obtain a prior-enhanced classifier weight; A target detection prediction result obtaining module, configured to update the target category probability according to the prior-enhanced classifier weight, and to obtain a target detection result according to the updated target category probability and the target detection frame. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1 to 7.

Images