Synthetic aperture radar vehicle target data generation method and device

Abstract

The application provides a synthetic aperture radar vehicle target data generation method and device, which can be applied to the fields of remote sensing image processing, computer vision and artificial intelligence technology. The method comprises the following steps: detecting and extracting a vehicle target from a source domain synthetic aperture radar image to obtain physical attribute parameters of the vehicle target; acquiring a target domain synthetic aperture radar background image and constructing a three-dimensional scattering field model of the background image; performing cross-domain adaptive fusion on the vehicle target and the background image based on the physical attribute parameters and the three-dimensional scattering field model to generate a synthetic aperture radar image; and performing physical consistency verification and recognition efficiency verification on the generated synthetic aperture radar image, and outputting generated data that passes the double verification.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing, computer vision and artificial intelligence, and more specifically to a method and apparatus for generating vehicle target data using synthetic aperture radar. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses all-weather, all-day Earth observation capabilities and is widely used in military and civilian fields such as remote sensing monitoring and target recognition. In recent years, deep learning-based target recognition methods, especially end-to-end detection frameworks represented by the YOLO series, have demonstrated excellent performance in SAR image interpretation. However, the training effect of deep learning models is highly dependent on large-scale, high-quality labeled data. For certain specific geographical areas (such as remote or sensitive areas), acquiring SAR images containing vehicle targets and their accurate annotations is often costly, time-consuming, or even impossible, resulting in an extreme scarcity of target area data. This severely restricts the application performance of deep learning-based target recognition models in these areas.

[0003] To alleviate the problem of data scarcity in specific regions, existing technologies mainly adopt the following approaches:

[0004] (1) Transfer learning method. The model is pre-trained on a source domain with sufficient data (such as the area covered by a public dataset), and then the model parameters are transferred to the target area for fine-tuning. However, SAR images are affected by many factors such as sensor parameters, imaging geometry, and ground object scattering characteristics. The data distribution between different regions is significantly different, and direct transfer can easily lead to performance loss of the model in the target domain.

[0005] (2) Data generation method based on generative adversarial networks. Generative adversarial networks are used to synthesize SAR target images to expand the training set. Most existing methods focus on image appearance transformations and fail to deeply model the physical imaging mechanisms unique to SAR images, such as the statistical characteristics of speckle noise and the coordination of scattering intensity between the target and the background. This results in insufficient physical credibility of the generated samples and very limited performance improvement for downstream recognition models.

[0006] (3) Image stitching synthesis method. The target slice detected in the source domain is directly embedded into the background image of the target domain. This method ignores the inherent consistency between the target and the background in terms of scattering intensity distribution, noise texture and three-dimensional geometry in SAR images. The synthesis results often have obvious edge artifacts and physical property mismatch, making it difficult to use for training high-performance recognition models.

[0007] Furthermore, the aforementioned methods generally lack a systematic and quantifiable evaluation mechanism for the quality of synthetic data, especially lacking verification standards oriented towards actual recognition effectiveness, making it difficult to guarantee whether the generated synthetic data truly helps improve the performance of the target recognition model. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the challenges of generating physically reliable synthetic aperture radar (SAR) image data that blends naturally with the background in the absence of real vehicle samples in the target area, and the lack of an effective quality assessment mechanism in existing technologies, this invention provides a method and apparatus for generating SAR vehicle target data. The method extracts physical attribute parameters of vehicle targets from source domain SAR images, constructs a three-dimensional scattering field model of the target domain background image, and performs cross-domain adaptive fusion of the vehicle target and background image based on the physical attribute parameters and the three-dimensional scattering field model. The generated SAR image undergoes dual screening for physical consistency verification and recognition performance verification, thereby outputting qualified generated data.

[0010] (II) Technical Solution

[0011] To address the aforementioned technical problems, embodiments of the present invention propose a method and apparatus for generating vehicle target data using synthetic aperture radar.

[0012] According to a first aspect of the present invention, a method for generating vehicle target data using synthetic aperture radar (SAR) is provided, comprising: detecting and extracting a vehicle target from a source domain SAR image to obtain physical attribute parameters of the vehicle target; acquiring a target domain SAR background image and constructing a three-dimensional scattering field model of the background image; performing cross-domain adaptive fusion of the vehicle target and the background image based on the physical attribute parameters and the three-dimensional scattering field model to generate a SAR image; performing physical consistency verification and recognition performance verification on the generated SAR image, and outputting generated data that has passed both verifications.

[0013] In some exemplary embodiments, detecting and extracting vehicle targets from source domain synthetic aperture radar images includes: using a target detection network to detect vehicle targets from source domain synthetic aperture radar images, wherein the backbone network of the target detection network is equipped with a multi-directional scanning mechanism and a proxy attention mechanism based on a preset number of tokens; extracting image slices containing the detected vehicle target and its surrounding predetermined range transition zone; performing scale normalization on the image slices, and calculating the average radar cross-section of the vehicle target based on the normalized pixel intensity as a physical attribute parameter; wherein the backbone network of the target detection network is equipped with a multi-directional scanning mechanism and a proxy attention mechanism based on a preset number of tokens; the multi-directional scanning mechanism includes scanning in horizontal, vertical, and diagonal directions.

[0014] In some exemplary embodiments, constructing a three-dimensional scattering field model of the background image includes: preprocessing the synthetic aperture radar background image of the target domain by speckle noise suppression; using a neural radiation field model, representing the imaging process of the preprocessed background image as the cumulative transmittance and scattering intensity of light propagating in a voxel grid of a preset resolution; learning the attenuation coefficient and scattering intensity of each voxel point through differentiable rendering to obtain a digital model characterizing the three-dimensional distribution of the target domain terrain and scattering intensity; wherein, the neural radiation field model uses a multilayer perceptron as the network backbone, the input is the three-dimensional spatial coordinates of the voxel points, and the output is the attenuation coefficient and scattering intensity of the corresponding points; the loss function of the neural radiation field model is constructed based on the mean square error between the rendered image and the input background image.

[0015] In some exemplary embodiments, cross-domain adaptive fusion of the vehicle target and the background image includes: spatially registering an image slice of the vehicle target with the background image; wherein the image slice includes a transition zone of the vehicle target and its surrounding predetermined range; in a color space that separates luminance and color information, performing physical model-guided fusion of the luminance channel based on physical property parameters and a three-dimensional scattering field model; performing fusion of the chrominance channel using a background texture preservation strategy; and performing adversarial edge and noise optimization on the fusion result.

[0016] In some exemplary embodiments, physical model-guided fusion of the brightness channel based on physical property parameters and a three-dimensional scattering field model includes: determining the fusion weight of each pixel using a gamma distribution model based on the average radar cross section and the three-dimensional scattering field model; and performing pixel-level fusion of the brightness channel according to the fusion weight.

[0017] In some exemplary embodiments, a background texture preservation strategy is used to fuse the chroma channels, including: multiplying the chroma channel values ​​of the background image by a preset weight to obtain a background weighted component; multiplying the chroma channel values ​​of the target slice by the difference obtained by subtracting the preset weight to obtain a target weighted component; and adding the background weighted component and the target weighted component to obtain the fused chroma channel values.

[0018] In some exemplary embodiments, adversarial edge and noise optimization of the fusion result includes: inputting the fusion result into a dual-discriminator conditional generative adversarial network, wherein the first discriminator is used to evaluate edge sharpness and the second discriminator is used to evaluate speckle noise statistical characteristics; optimizing the edge and noise characteristics of the fusion result through adversarial training between the generator and the two discriminators; the loss function used for adversarial training includes a gradient penalty term.

[0019] In some exemplary embodiments, physical consistency verification of the generated synthetic aperture radar image includes: extracting the gray-level co-occurrence matrix texture features of the generated synthetic aperture radar image and the real target domain synthetic aperture radar background image; calculating the similarity between the two based on the texture features; and determining that the generated synthetic aperture radar image passes the physical consistency verification when the similarity is greater than a first preset threshold.

[0020] In some exemplary embodiments, the recognition performance verification of the generated synthetic aperture radar image includes: mixing the generated synthetic aperture radar image with synthetic aperture radar image samples containing vehicle targets in the real target domain according to a preset ratio to form a training dataset; training a target detection network using the training dataset, the target detection network being used to identify vehicle targets in the synthetic aperture radar image; evaluating the recognition performance index of the trained target detection network on an independent test set; calculating the improvement of the recognition performance index relative to the model trained using only real target domain samples; and determining that the generated synthetic aperture radar image passes the recognition performance verification when the improvement is greater than a second preset threshold.

[0021] According to a second aspect of the present invention, a synthetic aperture radar (SAR) vehicle target data generation apparatus is provided, comprising: a data acquisition module for detecting and extracting vehicle targets from a source domain SAR image and obtaining physical attribute parameters of the vehicle targets; a model building module for acquiring a target domain SAR background image and constructing a three-dimensional scattering field model of the background image; an image generation module for performing cross-domain adaptive fusion of the vehicle targets and the background image based on the physical attribute parameters and the three-dimensional scattering field model to generate a SAR image; and a result output module for performing physical consistency verification and recognition performance verification on the generated SAR image and outputting generated data that has passed both verifications.

[0022] (III) Beneficial Effects

[0023] As can be seen from the above technical solutions, the synthetic aperture radar vehicle target data generation method and apparatus provided by the embodiments of the present invention have at least the following beneficial effects:

[0024] (1) High fidelity of physical attributes. This invention extracts the average radar cross section of the vehicle target from the source domain synthetic aperture radar image as a physical attribute label, and combines it with the three-dimensional scattering field model of the target domain background image to guide the fusion of the brightness channel physical model. The essential scattering features of the target are embedded in the generation process, so that the generated synthetic aperture radar image is highly consistent with the real data in terms of scattering intensity distribution and speckle noise characteristics. This overcomes the shortcomings of existing methods that rely solely on appearance transformation, resulting in insufficient physical credibility.

[0025] (2) Excellent cross-domain fusion quality. This invention adopts a multi-level fusion strategy of spatial registration, physical guided fusion of the luminance channel, preservation of background texture of the chrominance channel, and adversarial edge and noise optimization. It achieves natural and coordinated fusion of the source domain vehicle target and the target domain background image in multiple dimensions such as geometric space, scattering intensity, texture features and edge noise. It effectively avoids edge artifacts and attribute mismatch problems, and significantly improves the visual quality and intrinsic consistency of the synthesized image.

[0026] (3) High Practicality of Synthetic Data. This invention establishes a dual quality screening mechanism that combines physical consistency verification and recognition performance verification. Only when the generated synthetic aperture radar image simultaneously meets the requirements of texture feature similarity and target detection network performance improvement is it output as qualified data. This mechanism ensures that the generated synthetic data not only has statistical features close to reality, but also effectively improves the recognition accuracy of the target detection network in downstream tasks, solving the problem of the lack of application performance-oriented quality assessment methods in existing technologies.

[0027] (4) The solution is complete and highly feasible. This invention provides an end-to-end modular framework from source domain feature extraction, target domain 3D modeling, cross-domain adaptive fusion to dual verification output. Each module has a clear function and works in coordination. The technical route is clear. Those skilled in the art can fully reproduce and solve the practical engineering problem of scarce synthetic aperture radar vehicle target training data in specific areas based on the contents of this invention. Attached Figure Description

[0028] The above-described features, other objects, and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0029] Figure 1 A flowchart illustrating a synthetic aperture radar vehicle target data generation method according to an embodiment of the present invention is shown schematically.

[0030] Figure 2 A flowchart illustrating the detection and extraction of vehicle targets from source-domain synthetic aperture radar images according to an embodiment of the present invention is shown.

[0031] Figure 3 A flowchart illustrating the construction of a three-dimensional scattering field model of a background image according to an embodiment of the present invention is shown.

[0032] Figure 4 This schematically illustrates a flowchart of cross-domain adaptive fusion of a vehicle target and a background image according to an embodiment of the present invention;

[0033] Figure 5 This schematically illustrates a flowchart of physical consistency verification of a generated synthetic aperture radar image according to an embodiment of the present invention;

[0034] Figure 6 This schematically illustrates a flowchart for verifying the recognition performance of a generated synthetic aperture radar image according to an embodiment of the present invention;

[0035] Figure 7 A schematic diagram of a synthetic aperture radar vehicle target data generation apparatus according to an embodiment of the present invention is shown.

[0036] Figure 8 A block diagram of an electronic device for a synthetic aperture radar vehicle target data generation method according to an embodiment of the present invention is shown schematically. Detailed Implementation

[0037] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0038] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0039] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0040] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0041] Figure 1 A flowchart illustrating a synthetic aperture radar vehicle target data generation method according to an embodiment of the present invention is shown.

[0042] like Figure 1As shown, the synthetic aperture radar vehicle target data generation method according to an embodiment of the present invention includes steps S110 to S140.

[0043] In step S110, vehicle targets are detected and extracted from the source domain synthetic aperture radar image to obtain the physical attribute parameters of the vehicle targets.

[0044] In an embodiment of the present invention, step S110 may specifically include steps S210 to S230, see below. Figure 2 .

[0045] In step S210, a target detection network is used to detect vehicle targets from the source domain synthetic aperture radar image. The backbone network of the target detection network is equipped with a multi-directional scanning mechanism and a proxy attention mechanism based on a preset number of tokens.

[0046] In an embodiment of the present invention, the backbone network of the target detection network is provided with a multi-directional scanning mechanism and a proxy attention mechanism based on a preset number of tokens; the multi-directional scanning mechanism includes scanning in horizontal, vertical and diagonal directions.

[0047] For example, the target detection network employs an improved YOLOv8 framework, with a specially designed MambaSAR module embedded in its backbone. This module enhances the network's ability to perceive the global scattering structure features of vehicle targets through a multi-directional scanning mechanism and a proxy attention mechanism. Specifically, the multi-directional scanning mechanism includes horizontal, vertical, and diagonal scanning with a fixed scanning step size, such as 3×3 pixels; the proxy attention mechanism aggregates features based on a preset number of tokens (e.g., 128). These settings effectively improve the network's robustness in detecting strong scattering points and small targets within vehicle targets.

[0048] In step S220, an image slice containing the detected vehicle target and its surrounding predetermined range transition zone is extracted.

[0049] In an embodiment of the present invention, upon detecting a vehicle target, a rectangular region centered on the target, including the target and its surrounding transition zone, is cropped as a target slice. The width of the transition zone can be set according to the actual image resolution; for example, a 50×50 pixel transition zone can be selected. This transition zone is used to smooth the boundary between the target and the background during subsequent fusion.

[0050] In step S230, the image slices are scale-normalized, and the average radar cross-section of the vehicle target is calculated based on the normalized pixel intensity as a physical attribute parameter.

[0051] Specifically, firstly, all target slices are normalized to a uniform size (e.g., 128×128 pixels) to eliminate scale differences. Then, based on the normalized backscattering intensity of each pixel within the normalized slice (usually linearly related to radar received power), its arithmetic mean is calculated as the average radar cross-section (RCS) of the vehicle target. The calculation formula is as follows:

[0052]

[0053] in, RCS represents the normalized backscattering intensity of the i-th pixel within the target slice (typically linearly related to radar received power), where N is the total number of pixels within the slice. This RCS value is stored as a key physical feature of the target and used for physical consistency constraints in subsequent fusion stages.

[0054] In step S120, a synthetic aperture radar background image of the target domain is acquired, and a three-dimensional scattering field model of the background image is constructed.

[0055] In an embodiment of the present invention, step S120 may specifically include steps S310 to S330, see below. Figure 3 .

[0056] In step S310, the background image of the target domain synthetic aperture radar is preprocessed to suppress speckle noise.

[0057] Due to the inherent speckle noise in synthetic aperture radar images, speckle noise suppression is first performed on the acquired background image of the target area. For example, Gaussian filtering (e.g., with a kernel size of 5×5) is used for initial smoothing to reduce noise interference with subsequent 3D modeling.

[0058] In step S320, the imaging process of the preprocessed background image is represented by the cumulative transmittance and scattering intensity of light propagating in a voxel grid of a preset resolution using a neural radiation field model.

[0059] In an embodiment of the present invention, the neural radiation field model uses a multilayer perceptron as the network backbone, with the input being the three-dimensional spatial coordinates of voxel points and the output being the attenuation coefficient and scattering intensity of the corresponding points; the loss function of the neural radiation field model is constructed based on the mean square error between the rendered image and the input background image.

[0060] Specifically, a Neural Radiation Field (NeRF) model is used to model the three-dimensional scattering field of the background image. The imaging process is modeled as light emanating from the sensor and propagating in a three-dimensional voxel space, with each voxel having an attenuation coefficient (σ) and scattering intensity (c). The preset resolution of the voxel grid can be set according to actual needs, for example, 0.5 meters (horizontal) × 0.5 meters (distance) × 1 meter (height). During model training, a preset number of points (e.g., 256) are sampled for each ray, and the predicted intensity of the pixel is obtained through volume rendering integration. The NeRF model uses a multilayer perceptron (MLP) as the network backbone, with the three-dimensional spatial coordinates of the voxel points as input and the attenuation coefficient and scattering intensity of that point as output. The model's loss function is constructed based on the mean squared error (MSE) between the rendered image and the input background image, and is trained using a gradient descent optimization algorithm (e.g., Adam) in differentiable rendering, with a training epoch of up to 50 rounds.

[0061] In step S330, the attenuation coefficient and scattering intensity of each voxel are learned through differentiable rendering to obtain a digital model characterizing the three-dimensional distribution of the target domain terrain and scattering intensity.

[0062] After the above optimizations, the neural radiation field model converges, providing corresponding scattering characteristic parameters for each spatial location in the background image, forming a continuous 3D scattering field representation. This digital model not only includes the geometric information of the terrain but also encodes the backscattering intensity distribution unique to synthetic aperture radar, providing a physical reference for subsequent target placement.

[0063] In step S130, based on physical property parameters and a three-dimensional scattering field model, the vehicle target and the background image are cross-domain adaptively fused to generate a synthetic aperture radar image.

[0064] In an embodiment of the present invention, step S130 may specifically include steps S410 to S440, see below. Figure 4 .

[0065] In step S410, the image slice of the vehicle target is spatially registered with the background image; wherein the image slice includes the vehicle target and a transition zone of a predetermined range around it.

[0066] In an embodiment of the present invention, an end-to-end synthetic aperture radar image registration network is designed. This network takes target slices and background candidate regions as input, extracts features using ResNet-50, and utilizes a joint loss function for supervision, outputting the optimal affine transformation matrix. The joint loss function is a weighted sum of diversity peak loss and sparse cross-entropy loss, calculated as follows:

[0067]

[0068] The weight is set to λ. dp =0.3, λsce=0.7. Peak diversity loss L dp The aim is to encourage the distribution of feature points predicted by the network to have multiple significant peaks (i.e., diversity), which may specifically involve the negative value of the predicted heatmap entropy or the optimization of the distance between peaks; sparse cross-entropy loss L sce This network is used to supervise the accuracy of feature point classification. Through this registration network, the average registration error of target slices can be reduced by more than 65%, ensuring that the target is accurately aligned geometrically with the background before fusion.

[0069] In step S420, in the color space where brightness and color information are separated, the brightness channel is fused using a physical model guided by physical property parameters and a three-dimensional scattering field model.

[0070] In an embodiment of the present invention, step S420 may specifically include: determining the fusion weight of each pixel using a gamma distribution model based on the average radar cross section and a three-dimensional scattering field model; and performing pixel-level fusion of the brightness channel according to the fusion weight.

[0071] For example, converting the registered target slice and background image to L ab A color space is used to separate luminance (L) and chromaticity (a, b) information. For the luminance channel, an RCS-guided Gamma diffusion model is employed for fusion. Specifically, based on the average radar cross section calculated in step S230 and the three-dimensional scattering field model obtained in step S330, the theoretical scattering intensity distribution that each pixel should possess under specific geometric and imaging conditions in the target area is simulated. During the fusion process, for each pixel location, not only are they selected according to the probability weights of the Gamma distribution, but an RCS consistency constraint is also introduced to ensure that the overall and local scattering intensity of the fused target area is consistent with the theoretical distribution, thereby achieving physically reliable luminance fusion.

[0072] In step S430, the chroma channels are fused using a background texture preservation strategy.

[0073] Since synthetic aperture radar (SAR) images are typically grayscale, chromaticity information primarily originates from background texture when converting to pseudo-color or undergoing multi-channel processing. To maintain the natural consistency of the background, a weighted retention strategy is employed for chromaticity channel fusion: the chromaticity channel values ​​of the background image are multiplied by a preset weight to obtain the background weighted component; the chromaticity channel values ​​of the target slice are multiplied by the difference between one and the preset weight to obtain the target weighted component; the background weighted component and the target weighted component are then added together to obtain the fused chromaticity channel values. The calculation formula is as follows:

[0074]

[0075]

[0076] in, , and , Here, (u, v) represents the a and b channel values ​​of the target slice and the background image, respectively, and (u, v) represents the discrete pixel coordinates. The background weight coefficient is fixed at w = 0.8. This strategy ensures that the background texture dominates the fusion result, thereby avoiding the disruption of overall texture harmony caused by the introduction of the target.

[0077] In step S440, adversarial edge and noise optimization is performed on the fusion result.

[0078] To further enhance the realism of the synthesized images, a Dual Discriminator Conditional Generative Adversarial Network (DDM-CGAN) is employed for post-processing of the initial fusion results. This network comprises one generator and two discriminators: the first discriminator focuses on evaluating edge sharpness, while the second discriminator evaluates the statistical characteristics of global speckle noise. During training, a Wasserstein GAN loss function with a gradient penalty term (WGAN-GP) is used, with the generator's learning rate set to 0.0002 and optimized iteratively 30 times. Through adversarial training between the generator and the two discriminators, the final output synthesized image is indistinguishable from real synthetic aperture radar images in terms of edge naturalness and noise characteristics.

[0079] In step S140, the generated synthetic aperture radar image is subjected to physical consistency verification and recognition performance verification, and the generated data that has passed the dual verification is output.

[0080] In embodiments of the present invention, physical consistency verification of the generated synthetic aperture radar image may specifically include steps S510 to S530, see below. Figure 5 .

[0081] In step S510, the gray-level co-occurrence matrix texture features of the generated synthetic aperture radar image and the synthetic aperture radar background image of the real target domain are extracted.

[0082] For example, it can be used to calculate second-order statistics such as contrast, correlation, energy, and homogeneity. The gray-level co-occurrence matrix (GLCM) describes the roughness, regularity, and other properties of texture by counting the occurrences of pixel pairs with specific spatial relationships and gray levels in an image.

[0083] In step S520, the similarity between the two is calculated based on texture features.

[0084] For example, GLCM features can be vectorized, and similarity scores can be calculated using metrics such as Euclidean distance, cosine similarity, or Mahalanobis distance.

[0085] In step S530, when the similarity is greater than the first preset threshold, the generated synthetic aperture radar image is determined to have passed the physical consistency verification.

[0086] In an embodiment of the present invention, the threshold can be set to 85%.

[0087] In embodiments of the present invention, verifying the recognition effectiveness of the generated synthetic aperture radar image may specifically include steps S610 to S650, see below. Figure 6 .

[0088] In step S610, the generated synthetic aperture radar image is mixed with synthetic aperture radar image samples containing vehicle targets in the real target domain according to a preset ratio to form a training dataset.

[0089] For example, the ratio of real to synthetic samples is 1:5, with 12 real samples per class.

[0090] In step S620, a target detection network is trained using a training dataset. The target detection network is used to identify vehicle targets in synthetic aperture radar images.

[0091] In step S630, the recognition performance metrics of the trained object detection network are evaluated on an independent test set.

[0092] In step S640, the improvement in recognition performance metrics relative to the model trained using only real target domain samples is calculated.

[0093] In step S650, when the increase is greater than the second preset threshold, it is determined that the generated synthetic aperture radar image passes the recognition performance verification.

[0094] Only when the synthesized synthetic aperture radar image passes both the physical consistency verification and the recognition performance verification mentioned above is it deemed qualified data and included in the final output library for downstream target detection model training; otherwise, the synthesized data is discarded or regenerated.

[0095] In combination with the above Figures 1 to 6 Method and process Figure 7 From the perspective of functional modular design, the logical architecture of the system of the present invention during actual deployment is further demonstrated.

[0096] Figure 7 A schematic block diagram of a synthetic aperture radar vehicle target data generation apparatus according to an embodiment of the present invention is shown.

[0097] like Figure 7As shown, a synthetic aperture radar vehicle target data generation device 700 of this embodiment includes a data acquisition module 710, a model building module 720, an image generation module 730, and a result output module 740.

[0098] The data acquisition module 710 is used to detect and extract vehicle targets from source domain synthetic aperture radar images to obtain the physical attribute parameters of the vehicle targets. In one embodiment, the data acquisition module 710 can be used to perform the operation S110 described above, which will not be repeated here.

[0099] The model building module 720 is used to acquire a synthetic aperture radar background image of the target domain and construct a three-dimensional scattering field model of the background image. In one embodiment, the model building module 720 can be used to perform the operation S120 described above, which will not be repeated here.

[0100] The image generation module 730 is used to perform cross-domain adaptive fusion of the vehicle target and the background image based on physical property parameters and a three-dimensional scattering field model to generate a synthetic aperture radar image. In one embodiment, the image generation module 730 can be used to perform the operation S130 described above, which will not be repeated here.

[0101] The result output module 740 is used to perform physical consistency verification and recognition performance verification on the generated synthetic aperture radar image, and output the generated data that has passed the dual verification. In one embodiment, the result output module 740 can be used to perform the operation S140 described above, which will not be repeated here.

[0102] According to embodiments of the present invention, any plurality of modules among the data acquisition module 710, model building module 720, image generation module 730, and result output module 740 may be combined into one module, or any one of these modules may be split into multiple modules. Alternatively, at least a portion of the functionality of one or more of these modules may be combined with at least a portion of the functionality of other modules and implemented in one module. According to embodiments of the present invention, at least one of the data acquisition module 710, model building module 720, image generation module 730, and result output module 740 may be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in hardware or firmware, or in any one of software, hardware, and firmware implementations, or in a suitable combination of any of these. Alternatively, at least one of the data acquisition module 710, model building module 720, image generation module 730, and result output module 740 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0103] Figure 8 A block diagram of an electronic device for a synthetic aperture radar vehicle target data generation method according to an embodiment of the present invention is shown schematically.

[0104] like Figure 8 As shown, an electronic device 800 according to an embodiment of the present invention includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0105] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 802 and / or RAM 803. It should be noted that programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in one or more memories.

[0106] According to an embodiment of the present invention, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0107] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0108] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include ROM 802 and / or RAM 803 and / or one or more memories other than ROM 802 and RAM 803 described above.

[0109] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A method for generating vehicle target data using synthetic aperture radar, characterized in that, include: Vehicle targets are detected and extracted from source domain synthetic aperture radar images to obtain the physical attribute parameters of the vehicle targets; Acquire a synthetic aperture radar background image of the target domain and construct a three-dimensional scattering field model of the background image; Based on the physical property parameters and the three-dimensional scattering field model, the vehicle target and the background image are cross-domain adaptively fused to generate a synthetic aperture radar image. The generated synthetic aperture radar image is subjected to physical consistency verification and recognition performance verification, and the generated data that has passed the dual verification is output.

2. The method according to claim 1, characterized in that, The detection and extraction of vehicle targets from source domain synthetic aperture radar images includes: A target detection network is used to detect vehicle targets from source domain synthetic aperture radar images. The backbone network of the target detection network is equipped with a multi-directional scanning mechanism and a proxy attention mechanism based on a preset number of tokens. Extract an image slice containing the detected vehicle target and its surrounding predetermined transition zone; The image slices are scale-normalized, and the average radar cross section of the vehicle target is calculated based on the normalized pixel intensity, which is used as the physical attribute parameter. The backbone network of the target detection network is equipped with a multi-directional scanning mechanism and an agent attention mechanism based on a preset number of tokens; the multi-directional scanning mechanism includes scanning in horizontal, vertical and diagonal directions.

3. The method according to claim 1, characterized in that, The construction of the three-dimensional scattering field model of the background image includes: The target domain synthetic aperture radar background image is preprocessed with speckle noise suppression. Using the neural radiation field model, the imaging process of the preprocessed background image is represented as the cumulative transmittance and scattering intensity of light propagating in a voxel grid of a preset resolution. By learning the attenuation coefficient and scattering intensity of each voxel through differentiable rendering, a digital model characterizing the three-dimensional distribution of the terrain and scattering intensity of the target domain is obtained. The neural radiation field model uses a multilayer perceptron as the network backbone, with the input being the three-dimensional spatial coordinates of voxel points and the output being the attenuation coefficient and scattering intensity of the corresponding points; the loss function of the neural radiation field model is constructed based on the mean square error between the rendered image and the input background image.

4. The method according to claim 2, characterized in that, The step of performing cross-domain adaptive fusion of the vehicle target and the background image includes: Spatially register the image slice of the vehicle target with the background image; wherein the image slice includes a transition zone of the vehicle target and its surrounding predetermined range; In a color space that separates luminance and color information, the luminance channel is fused using a physical model-guided method based on the physical property parameters and the three-dimensional scattering field model. A background texture preservation strategy is used to blend the chroma channels; Adversarial edge and noise optimization is performed on the fusion results.

5. The method according to claim 4, characterized in that, The physical model-guided fusion of the brightness channel based on the physical property parameters and the three-dimensional scattering field model includes: Based on the average radar cross section and the three-dimensional scattering field model, the fusion weight of each pixel is determined by the gamma distribution model. The luminance channel is fused at the pixel level according to the fusion weight.

6. The method according to claim 4, characterized in that, The fusion of chroma channels using a background texture preservation strategy includes: The background image's chroma channel values ​​are multiplied by a preset weight to obtain the background weighted component; Multiply the chroma channel value of the target slice by the difference obtained by subtracting the preset weight to obtain the target weighted component; The background weighted component and the target weighted component are added together to obtain the fused chroma channel values.

7. The method according to claim 4, characterized in that, The adversarial edge and noise optimization of the fusion result includes: The fusion result is input into a dual-discriminator conditional generative adversarial network, where the first discriminator is used to evaluate edge sharpness and the second discriminator is used to evaluate speckle noise statistical characteristics. The edge and noise characteristics of the fusion result are optimized by adversarial training between the generator and two discriminators. The loss function used in the adversarial training includes a gradient penalty term.

8. The method according to claim 1, characterized in that, The physical consistency verification of the generated synthetic aperture radar image includes: Extract the gray-level co-occurrence matrix texture features of the generated synthetic aperture radar image and the real target domain synthetic aperture radar background image; The similarity between the two is calculated based on the texture features; When the similarity is greater than a first preset threshold, the generated synthetic aperture radar image is determined to have passed the physical consistency verification.

9. The method according to claim 1, characterized in that, The recognition performance of the generated synthetic aperture radar images is verified, including: The generated synthetic aperture radar images are mixed with synthetic aperture radar image samples containing vehicle targets in the real target domain according to a preset ratio to form a training dataset; A target detection network is trained using the training dataset, and the target detection network is used to identify vehicle targets in synthetic aperture radar images; The recognition performance metrics of the trained object detection network were evaluated on an independent test set. Calculate the improvement of the recognition performance index relative to the model trained using only real target domain samples; When the improvement is greater than the second preset threshold, the generated synthetic aperture radar image is determined to have passed the recognition performance verification.

10. A synthetic aperture radar vehicle target data generation device, characterized in that, include: The data acquisition module is used to detect and extract vehicle targets from source domain synthetic aperture radar images and obtain the physical attribute parameters of the vehicle targets. The model building module is used to acquire the background image of the synthetic aperture radar in the target domain and construct a three-dimensional scattering field model of the background image; The image generation module is used to perform cross-domain adaptive fusion of the vehicle target and the background image based on the physical property parameters and the three-dimensional scattering field model to generate a synthetic aperture radar image. The results output module is used to verify the physical consistency and recognition performance of the generated synthetic aperture radar image, and output the generated data that has passed the dual verification.

Images