Video sar vehicle shadow sample generation method combined with motion model

Abstract

The application discloses a video SAR vehicle shadow sample generation method combined with a motion model, relates to the field, is used for the image sample generation technical field, is used for detecting a video SAR image, and outputs a shadow target label and comprises the following steps: a vehicle instantaneous shadow projection model is established, and a physical shadow template library is generated; an image conversion network for realizing conversion from a physical template to a real SAR shadow image is constructed, a composite loss containing a region perception loss is used to train the image conversion network, an image conversion network model is constructed, and conversion enhanced samples are generated; an enhanced shadow dataset is constructed by fusing real label samples, physical template samples and conversion enhanced samples, and is used for training a vehicle shadow detection model; and video SAR vehicle shadow samples are output based on the trained vehicle shadow detection model. The application solves the problems of insufficient video SAR moving vehicle shadow label samples, lack of physical constraints in traditional data enhancement and insufficient generalization ability of a detection model.

Description

Technical Field

[0001] This invention relates to the field of image sample generation technology, and more particularly to the field of video synthetic aperture radar vehicle shadow sample generation technology, specifically a video SAR vehicle shadow sample generation method that combines motion models. Background Technology

[0002] In video synthetic aperture radar (SAR) imaging, due to the combined effects of platform motion errors, target motion, and imaging geometric changes, shadowed targets between consecutive frames often exhibit significant spatial deformation and drift, leading to a decrease in the temporal consistency of multi-frame images. Especially in vehicle motion scenarios, the shape of shadows undergoes nonlinear stretching and bending with changes in the target vehicle's speed, heading angle, and radar observation angle, making the acquisition of accurate labeled data costly and difficult to cover complex motion states.

[0003] Most existing video SAR target detection methods rely on a limited number of real labeled samples to generate training data through geometric transformation or random perturbation. These methods usually ignore the physical imaging mechanism and motion constraints in the shadow formation process, resulting in significant differences between the synthetic samples and real SAR images in terms of spatial structure and speckle statistics. This limits the generalization ability and robustness of deep learning detection models in complex scenes.

[0004] Therefore, it is necessary to construct a video SAR shadow data generation and enhancement method that can simultaneously satisfy "physical consistency" and "real texture consistency" in order to improve data quality and model performance in moving vehicle shadow detection tasks. Summary of the Invention

[0005] The purpose of this invention is to provide a video SAR vehicle shadow sample generation method that combines motion models, so as to solve the problems of difficulty in obtaining real SAR shadow sample images, low accuracy of synthesized samples, and limitation of model generalization ability in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for generating video SAR vehicle shadow samples by combining a motion model, comprising: S1. Obtain the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. Based on the geometric relationship of SAR imaging and the kinematic characteristics of the target, analyze the SAR shadow geometry of the target vehicle and construct a shadow dwell time evolution model. S2. Based on the shadow dwell time evolution model constructed in S1, a standard shadow template is generated. Shadow clutter ratio and speckle noise are introduced into the labeled shadow template to generate a physical shadow template. S3. Using the physical shadow template generated in S2 as a conditional input, construct an image transformation network and integrate a spatial adaptive normalization structure into the image transformation network. S4. Construct a composite loss function that includes region-aware loss, and use the composite loss function to train the image conversion network constructed in S3 to build the image conversion network model. S5. The image transformation network model constructed using S4 performs image transformation on the physical shadow template generated in S2 to generate transformed and enhanced shadow images, and constructs an enhanced shadow dataset. S6. Train the target detection model using the enhanced shadow dataset to generate a vehicle shadow detection model, and generate video SAR vehicle shadow samples based on the vehicle shadow detection model.

[0007] The geometric parameters in S1 include vehicle length, vehicle width, and vehicle height; the motion parameters include vehicle ground speed, vehicle heading angle, and vehicle turning angular velocity; and the radar observation parameters include radar ground sweep angle, synthetic aperture time, image resolution, and pixel scale. The geometric representation of the shadow includes the effective projection length of the target vehicle in the azimuth direction, the effective projection length of the target vehicle in the range direction, the effective projection length of the target vehicle in the range direction after radar ground sweep angle correction, and the shadow center trajectory. ; ; ; ; ; ; In the formula, Indicates the vehicle's heading angle. This represents the effective projection length of the target vehicle in the azimuth direction of the SAR image plane. Indicates the length of the vehicle. This represents the effective projection length of the target vehicle in the range direction of the SAR image plane. This indicates the effective projected length of the target vehicle in the range direction after radar sweep angle correction. Indicates the width of the vehicle. Indicates vehicle height. Indicates the radar's ground sweep angle. It is a cosine function. It is a sine function. Represents absolute value. , For time variables, representing moments in time. express The center of the shadow at that moment, , They represent Projection on the X and Y axes Indicates the initial position of the shadow center. express Vehicle ground speed at any given moment express The vehicle's heading angle at any given moment.

[0008] The shadow dwell time evolution model is as follows: ; In the formula, Indicates the pixel position The intensity of shadow dwell time at that location This indicates the number of sampling frames within the sliding time window. For indicator functions, express The shadow area of ​​time, Indicates the first in the sliding window Each sampling time, by Time corresponding , , , A joint decision.

[0009] Standard shadow templates include shadow templates for stationary targets, shadow templates for targets moving in a straight line, and shadow templates for targets turning at a constant speed. When the target vehicle is stationary, the vehicle's ground speed ,in accordance with , , Calculate the shadow area and construct a shadow template for a stationary target; When the target vehicle is in a straight-line motion state, the vehicle ground speed and vehicle heading angle remain unchanged during the synthetic aperture time. Based on the vehicle ground speed and vehicle heading angle, the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state is calculated. Based on the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state, the straight-line motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the straight-line motion trajectory of the shadow to construct a shadow template for a straight-line moving target. When the target vehicle is in a constant speed turning state, the vehicle's heading angle changes with time, while the vehicle's turning angular velocity remains constant. When the vehicle's turning angular velocity approaches zero, the target vehicle's constant speed turning state degenerates into a straight-line motion state. Based on the vehicle's turning angular velocity and the sweeping displacement of the shadow in the azimuth and range directions in the straight-line motion state, the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state is calculated. Based on the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state, the uniform speed turning motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the uniform turning motion trajectory of the shadow to construct a uniform turning target shadow template.

[0010] The sweep length of the shadow in the azimuth direction during linear motion. for: ; The sweep length of the shadow in the distance direction during linear motion. for: ; The sweep length of the shadow in the azimuth direction during a constant speed turn. for: ; The sweep length of the shadow in the distance direction during a constant speed turn. for: ; ; ; In the formula, Indicates the time for synthesizing the aperture. Indicates the equivalent average sailing angle. Indicates the total angular span. Indicates the initial sailing angle. This indicates the vehicle's turning angular velocity.

[0011] The shadow image intensity in the physical shadow template is: ; ; In the formula, Indicates the pixel position The intensity of the shadow image at that location, Indicates the average clutter intensity. For shadow clutter ratio, This represents the average intensity of the background clutter region. Indicates the average intensity of the shaded area. This is speckle noise.

[0012] The image transformation network includes a generator and a multi-scale discriminator. The generator is used to generate a transformed and enhanced shadow image based on a physical shadow template, and the multi-scale discriminator is used to discriminate the transformed and enhanced shadow image generated by the generator based on the physical shadow template. The generator consists of an input layer, an encoder, a bottleneck module, a decoder, and an output layer. The encoder is used to extract multi-scale features of the physical shadow template, the bottleneck module is used to inject deep features into the spatial structure of the physical shadow template, and the decoder is used to recover the generated image. Shallow spatial information is passed between the encoder and decoder through skip connections. The output layer of the generator uses a hyperbolic tangent activation function to constrain the pixel values ​​of the transformed enhanced shadow image output by the generator. The multi-scale discriminator includes a fine-scale discriminator and a coarse-scale discriminator. The fine-scale discriminator operates at the original resolution and is used to constrain local speckle texture, shadow edges, and small-scale clutter details. The coarse-scale discriminator operates at the downsampled resolution and is used to constrain the overall shape of the shadow region and the global structural consistency.

[0013] The generator incorporates a spatial adaptive normalization module, which performs a normalization operation on the spatial structure based on the physical shadow template, and extracts the spatial features of the physical shadow template. ; In the formula, For spatial adaptive normalization, This represents the intermediate feature map in the generator. This represents the conditional feature map of the physical shadow template obtained by the lightweight convolutional extractor. Representing intermediate feature maps The channel mean, Representing intermediate feature maps The channel standard deviation, The scaling factor represents the spatial variation. Translation coefficients representing spatial changes This indicates element-wise multiplication.

[0014] In S4, the composite loss function for: ; ; ; ; ; ; ; In the formula, To counteract the loss, For pixel reconstruction loss term, For feature matching loss term, For the region perception loss term, , , They represent , , The weighting coefficients, Represents a physical shadow template. This represents the true SAR shadow image paired with the physical shadow template. This represents the transformed and enhanced shadow image output by the generator. Represents the Manhattan norm. These are intermediate parameters for calculating the feature matching loss. Indicates scale index. For the first The discriminant at each scale is at the 1st... The feature maps output by the intermediate layer The number of feature layers for normalization. Represents the loss term for the shaded area. This represents the background region loss term. , They represent , The weighting coefficients, Indicates a soft shadow mask. This represents the background mask.

[0015] In S4, after the image conversion network training is completed, the generator of the trained image conversion network model is used to convert the physical shadow template obtained in S2 to generate a converted and enhanced shadow image, and construct a converted and enhanced shadow sample. The enhanced shadow samples are fused with video SAR shadow samples and physical shadow template samples to obtain fused sample data. The fused sample data is then processed by image size normalization, shadow annotation transformation, label file generation, and training set partitioning to construct an enhanced shadow dataset. The enhanced shadow dataset is input into the object detection network for training to generate a vehicle shadow detection model; Obtain the video SAR image sequence to be detected, input the video SAR image sequence into the vehicle shadow detection model, and output the location box, category label and confidence of the vehicle shadow target to obtain video SAR vehicle shadow samples.

[0016] Compared with the prior art, the present invention has the following advantages: This invention constructs an enhanced shadow dataset by fusing real labeled samples, physical template samples, and transformed enhanced samples. Based on the enhanced shadow dataset, the target detection model is trained to generate a vehicle shadow detection model, which significantly improves the accuracy and stability of the detection model in complex motion scenes and effectively reduces the false alarm rate. Attached Figure Description

[0017] Figure 1 This is a model architecture diagram of the image conversion network model provided by the present invention; Figure 2 This is a schematic diagram of the target loss of the composite generator provided by the present invention; Figure 3 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 4 This is a system architecture diagram of Embodiment 2 provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] A method for generating video SAR vehicle shadow samples by incorporating a motion model includes: S1. Obtain the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. Based on the geometric relationship of SAR imaging and the kinematic characteristics of the target, analyze the SAR shadow geometry of the target vehicle and construct a shadow dwell time evolution model. S2. Based on the shadow dwell time evolution model constructed in S1, a standard shadow template is generated. Shadow clutter ratio and speckle noise are introduced into the labeled shadow template to generate a physical shadow template. S3. Using the physical shadow template generated in S2 as a conditional input, construct an image transformation network and integrate a spatial adaptive normalization structure into the image transformation network. S4. Construct a composite loss function that includes region-aware loss, and use the composite loss function to train the image conversion network constructed in S3 to build the image conversion network model. S5. The image transformation network model constructed using S4 performs image transformation on the physical shadow template generated in S2 to generate transformed and enhanced shadow images, and constructs an enhanced shadow dataset. S6. Train the target detection model using the enhanced shadow dataset to generate a vehicle shadow detection model, and generate video SAR vehicle shadow samples based on the vehicle shadow detection model.

[0020] The geometric parameters in S1 include vehicle length, vehicle width, and vehicle height; the motion parameters include vehicle ground speed, vehicle heading angle, and vehicle turning angular velocity; and the radar observation parameters include radar ground sweep angle, synthetic aperture time, image resolution, and pixel scale. The geometric representation of the shadow includes the effective projection length of the target vehicle in the azimuth direction, the effective projection length of the target vehicle in the range direction, the effective projection length of the target vehicle in the range direction after radar ground sweep angle correction, and the shadow center trajectory. ; ; ; ; ; ; In the formula, Indicates the vehicle's heading angle. This represents the effective projection length of the target vehicle in the azimuth direction of the SAR image plane. Indicates the length of the vehicle. This represents the effective projection length of the target vehicle in the range direction of the SAR image plane. This indicates the effective projected length of the target vehicle in the range direction after radar sweep angle correction. Indicates the width of the vehicle. Indicates vehicle height. Indicates the radar's ground sweep angle. It is a cosine function. It is a sine function. Represents absolute value. , For time variables, representing moments in time. express The center of the shadow at that moment, , They represent Projection on the X and Y axes Indicates the initial position of the shadow center. express Vehicle ground speed at any given moment express The vehicle's heading angle at any given moment.

[0021] The shadow dwell time evolution model is as follows: ; In the formula, Indicates the pixel position The intensity of shadow dwell time at that location This indicates the number of sampling frames within the sliding time window. For indicator functions, express The shadow area of ​​time, Indicates the first in the sliding window Each sampling time, by Time corresponding , , , A joint decision.

[0022] Standard shadow templates include shadow templates for stationary targets, shadow templates for targets moving in a straight line, and shadow templates for targets turning at a constant speed. When the target vehicle is stationary, the vehicle's ground speed ,in accordance with , , Calculate the shadow area and construct a shadow template for a stationary target; When the target vehicle is in a straight-line motion state, the vehicle ground speed and vehicle heading angle remain unchanged during the synthetic aperture time. Based on the vehicle ground speed and vehicle heading angle, the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state is calculated. Based on the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state, the straight-line motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the straight-line motion trajectory of the shadow to construct a shadow template for a straight-line moving target. When the target vehicle is in a constant speed turning state, the vehicle's heading angle changes with time, while the vehicle's turning angular velocity remains constant. When the vehicle's turning angular velocity approaches zero, the target vehicle's constant speed turning state degenerates into a straight-line motion state. Based on the vehicle's turning angular velocity and the sweeping displacement of the shadow in the azimuth and range directions in the straight-line motion state, the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state is calculated. Based on the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state, the uniform speed turning motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the uniform turning motion trajectory of the shadow to construct a uniform turning target shadow template.

[0023] The sweep length of the shadow in the azimuth direction during linear motion. for: ; The sweep length of the shadow in the distance direction during linear motion. for: ; The sweep length of the shadow in the azimuth direction during a constant speed turn. for: ; The sweep length of the shadow in the distance direction during a constant speed turn. for: ; ; ; In the formula, Indicates the time for synthesizing the aperture. Indicates the equivalent average sailing angle. Indicates the total angular span. Indicates the initial sailing angle. This indicates the vehicle's turning angular velocity.

[0024] The shadow image intensity in the physical shadow template is: ; ; In the formula, Indicates the pixel position The intensity of the shadow image at that location, Indicates the average clutter intensity. For shadow clutter ratio, This represents the average intensity of the background clutter region. Indicates the average intensity of the shaded area. This is speckle noise.

[0025] The image transformation network includes a generator and a multi-scale discriminator. The generator is used to generate a transformed and enhanced shadow image based on a physical shadow template, and the multi-scale discriminator is used to discriminate the transformed and enhanced shadow image generated by the generator based on the physical shadow template. The generator consists of an input layer, an encoder, a bottleneck module, a decoder, and an output layer. The encoder is used to extract multi-scale features of the physical shadow template, the bottleneck module is used to inject deep features into the spatial structure of the physical shadow template, and the decoder is used to recover the generated image. Shallow spatial information is passed between the encoder and decoder through skip connections. The output layer of the generator uses a hyperbolic tangent activation function to constrain the pixel values ​​of the transformed enhanced shadow image output by the generator. The multi-scale discriminator includes a fine-scale discriminator and a coarse-scale discriminator. The fine-scale discriminator operates at the original resolution and is used to constrain local speckle texture, shadow edges, and small-scale clutter details. The coarse-scale discriminator operates at the downsampled resolution and is used to constrain the overall shape of the shadow region and the global structural consistency.

[0026] The generator incorporates a spatial adaptive normalization module, which performs a normalization operation on the spatial structure based on the physical shadow template, and extracts the spatial features of the physical shadow template. ; In the formula, For spatial adaptive normalization, This represents the intermediate feature map in the generator. This represents the conditional feature map of the physical shadow template obtained by the lightweight convolutional extractor. Representing intermediate feature maps The channel mean, Representing intermediate feature maps The channel standard deviation, The scaling factor represents the spatial variation. Translation coefficients representing spatial changes This indicates element-wise multiplication.

[0027] In S4, the composite loss function for: ; ; ; ; ; ; ; In the formula, To counteract the loss, For pixel reconstruction loss term, For feature matching loss term, For the region perception loss term, , , They represent , , The weighting coefficients, Represents a physical shadow template. This represents the true SAR shadow image paired with the physical shadow template. This represents the transformed and enhanced shadow image output by the generator. Represents the Manhattan norm. These are intermediate parameters for calculating the feature matching loss. Indicates scale index. For the first The discriminant at each scale is at the 1st... The feature maps output by the intermediate layer The number of feature layers for normalization. Represents the loss term for the shaded area. This represents the background region loss term. , They represent , The weighting coefficients, Indicates a soft shadow mask. This represents the background mask.

[0028] In S4, after the image conversion network training is completed, the generator of the trained image conversion network model is used to convert the physical shadow template obtained in S2 to generate a converted and enhanced shadow image, and construct a converted and enhanced shadow sample. The enhanced shadow samples are fused with video SAR shadow samples and physical shadow template samples to obtain fused sample data. The fused sample data is then processed by image size normalization, shadow annotation transformation, label file generation, and training set partitioning to construct an enhanced shadow dataset. The enhanced shadow dataset is input into the object detection network for training to generate a vehicle shadow detection model; Obtain the video SAR image sequence to be detected, input the video SAR image sequence into the vehicle shadow detection model, and output the location box, category label and confidence of the vehicle shadow target to obtain video SAR vehicle shadow samples.

[0029] Example 1 The azimuth direction is the direction of satellite flight, and the range direction is perpendicular to the direction of satellite flight, i.e., the direction in which the radar line of sight is projected onto the ground. The vehicle heading angle is the angle between the vehicle's longitudinal axis (direction of the vehicle's front) and the SAR azimuth direction (i.e., the direction of radar satellite flight). The radar ground-sweep angle is the angle between the radar line of sight and the ground, which is equivalent to the angle between the direction of the radar beam and the ground plane when the radar beam illuminates the ground.

[0030] In S1, the shadow dwell intensity of the shadow dwell time evolution model is equal to the proportion of the number of sampling frames in which the pixel is covered by the instantaneous shadow region within the sliding time window, satisfying the following condition: . In S2, when the target vehicle is turning at a constant speed, the vehicle's heading angle changes over time: ; In the formula, express The vehicle heading angle at any given time is the angle between the vehicle's direction of travel and its azimuth.

[0031] Three types of standard shadow templates are generated based on the target's motion state, constructing a physical shadow template library covering three motion states: stationary, linear motion, and uniform turning. Shadow clutter ratio and speckle noise are introduced into the labeled shadow templates to give the generated samples clear motion physical constraints and basic SAR intensity characteristics.

[0032] The image transformation network is a mapping network from physical shadow templates to real SAR shadow images. It is used to convert physical shadow templates into transformed enhanced shadow images with real SAR speckle texture and background clutter statistical characteristics. Introducing a spatial adaptive normalization module into the generator of the image transformation network avoids the weakening of the spatial structure of the physical shadow template by traditional normalization operations, allowing the generated image to retain the texture and clutter features of real SAR images while maintaining the geometric structure. In the spatial adaptive normalization module, intermediate feature maps... Channel mean and intermediate feature maps Channel standard deviation The generator targets intermediate feature maps. For each channel, the global mean and standard deviation are calculated at all spatial locations. The scaling factor for spatial variation is also included. Translation coefficients of spatial variation Both are generated through a lightweight convolutional layer. The scaling factor of spatial variation determines the contrast strength of the transformed and enhanced shadow image at the corresponding pixel, and the translation factor of spatial variation determines the base brightness level of the transformed and enhanced shadow image at the corresponding pixel. The scaling factor and translation factor of spatial variation have the same spatial dimension as the intermediate feature map.

[0033] In the region-aware loss function, the pixel reconstruction loss term constrains the low-frequency consistency between the generated image and the real SAR shadow samples, while the feature matching loss term constrains the consistency between real and generated sample pairs on the intermediate feature layer of the discriminator. The region-aware loss term constrains the shadow region with a physical template, maintaining the geometric accuracy of the shadow; simultaneously, it constrains the background region with real SAR samples to preserve the true clutter statistical characteristics. Soft shadow mask. and background mask Generated from a physical shadow template.

[0034] Soft shadow mask for: ; Background mask for: ; In the formula, This represents the Sigmoid activation function. Indicates the brightness threshold. This indicates the mask sharpness control parameter.

[0035] During the training of the image transformation network using a composite loss function, sampled data from a parameter range covering three motion modes: stationary, linear motion, and uniform turning, to construct paired training data. The Adam optimizer is used to jointly optimize the generator and discriminator. The learning rate remains constant for a preset number of epochs, then decays linearly. Gradient clipping is used to stabilize the training process. Based on the composite loss function, region-differential supervision ensures that the transformed and enhanced shadow image simultaneously satisfies the geometric consistency of the shadow region and the realistic texture consistency of the background region.

[0036] Example 2 A video SAR vehicle shadow sample generation system incorporating a motion model is disclosed. This system can be deployed in computer equipment, servers, graphics processing unit workstations, or other electronic devices with data processing capabilities. It is used to execute the video SAR vehicle shadow sample generation method incorporating a motion model as described in Example 1. Figure 4 As shown, the system includes a parameter input module, a physical shadow template generation module, a real data access module, a paired data preparation module, an image conversion and augmentation module, an enhanced sample library construction module, a detection model training module, and a shadow detection output module.

[0037] The parameter input module receives vehicle geometric parameters, vehicle motion parameters, and radar observation parameters. Vehicle geometric parameters include vehicle length. Vehicle width and vehicle height Vehicle motion parameters include the target vehicle's ground speed. Vehicle heading angle and vehicle turning angular velocity Radar observation parameters include radar ground-sweep angle. The module includes parameters such as synthesis aperture time, image size (image resolution), and pixel scale. It connects to the physical shadow template generation module to provide fundamental parameters for shadow template generation.

[0038] The physical shadow template generation module is connected to the parameter input module and is used to generate static shadow templates, linear motion target (LMT) shadow templates, and constant turning rate (CTR) shadow templates based on the imaging geometry of Video Synthetic Aperture Radar (VSAR) and the target motion state. The physical shadow templates output by this module are used to characterize the spatial range, trailing shape, and geometric constraints of vehicle shadows under different motion states.

[0039] The real data access module is used to import real video SAR images and their corresponding vehicle shadow annotation data. This module connects to the paired data preparation module to provide real SAR background texture, speckle statistical characteristics, and supervision annotation information.

[0040] The paired data preparation module is connected to both the physical shadow template generation module and the real data access module. It is used to pair and organize the physical shadow template with real SAR shadow samples to form the training input for the image conversion network. The function of this module is to establish the correspondence between the "physical template and real sample," providing training data for subsequent image conversion augmentation.

[0041] The image transformation and augmentation module is connected to the paired data preparation module, preferably employing an improved Pix2PixHD image transformation network. This module includes a generator, a Spatially Adaptive Denormalization (SPADE) unit, a multi-scale discriminator, and a region-aware loss calculation unit. The generator transforms the physical shadow template into a shadow image with realistic SAR speckle texture; the SPADE unit preserves the spatial structure within the physical template; the multi-scale discriminator constrains local texture and overall structure; and the region-aware loss calculation unit separately constrains shadow and background regions.

[0042] The enhanced sample library construction module is connected to the image transformation and augmentation module. It receives transformed and enhanced shadow images and organizes the ground truth labeled samples, physical shadow samples, and transformed and enhanced samples into a unified format to generate an enhanced shadow dataset. This module can perform image size normalization, label format conversion, training set partitioning, and sample index management.

[0043] The detection model training module is connected to the augmented sample library construction module, and is used to train the vehicle shadow detection model using the augmented shadow dataset. The detection model can employ the YOLO11n detector or other existing object detection networks. This module enables the model to learn real texture features, physical structure priors, and transformed augmented sample features simultaneously through joint training with multi-source samples.

[0044] The shadow detection output module is connected to the detection model training module. It is used to perform inference on the SAR image sequence of the video to be detected and output the bounding box, category information, and confidence score of the vehicle shadow target. Through the above structural connection, this device can realize a complete closed loop from physical shadow modeling and realistic image conversion to detection model training and result output.

[0045] First, based on the geometric relationships of SAR imaging and the kinematic characteristics of the target, a unified shadow dwell time evolution model is established. Closed-form analytical shadow geometric expressions are derived for three typical motion states: stationary, linear motion, and uniform turning, thereby generating physically constrained physical shadow templates. Second, the generated physical shadow templates are used as conditional inputs to an improved Pix2PixHD generation network. By fusing a Spatial Adaptive Normalization (SPADE) structure, a multi-scale discriminator, and a region-aware loss function, a high-fidelity mapping from the geometric template to a shadow image with true SAR speckle statistical characteristics is achieved.

[0046] Finally, by jointly training physically generated samples, network-generated samples, and a small amount of real labeled data, an augmented dataset for video SAR vehicle shadow detection is constructed, which significantly improves the accuracy and stability of the detection model in complex motion scenarios and effectively reduces the false alarm rate.

[0047] Figure 1 The model architecture diagram of the image conversion network model provided by this invention is as follows: Figure 1 As shown, the image transformation network model includes an image transformation network and a composite loss function (equivalent to the loss of the generator's overall objective). The image transformation network includes a generator and a multi-scale discriminator. The generator generates transformed and enhanced shadow images based on physical shadow templates, and the multi-scale discriminator discriminates the transformed and enhanced shadow images generated by the generator based on the physical shadow templates. The generator includes an input layer, an encoder, a conditional extractor, a bottleneck module, a decoder, and an output layer. The multi-scale discriminator includes an input layer, a pooling layer, and at least two sub-discriminators. The encoder of the generator adopts a U-Net structure for the input template. Downsampling compression is performed to extract global and local deep semantic features of the image. A conditional extractor is used to extract the input template. The global conditional information is obtained and the conditional condition (cond) is fed in parallel into the bottleneck module and decoder. The bottleneck module and decoder incorporate the Spatial Adaptive Normalization (SPADE) structure. The bottleneck module uses the external conditional condition to adjust the mean and variance of the feature map pixel by pixel, ensuring that the features strictly conform to the spatial layout of the input template; it is the core module for conditional image generation. A skip connection between the encoder and decoder directly transmits the shallow detail features of each layer of the encoder to the subsequent decoder, ensuring that edge and texture details of the generated image are not lost. The output layer outputs... The resulting image uses a physical shadow template as a conditional input. The image transformation network model incorporates an improved Pix2PixHD generative network model, integrating a Spatial Adaptive Normalization (SPADE) structure, a multi-scale discriminator, and a region-aware loss function to achieve a high-fidelity mapping from a geometric template to a shadow image with realistic SAR speckle statistics. The Pix2PixHD generative network model is an image generation model, essentially a conditional generative adversarial network consisting of a generator, discriminator, and encoder. The generator includes a global generator network and a local enhancement network, enabling the model to grasp global semantics and optimize local details with low computational cost.

[0048] Figure 2 A schematic diagram of the composite generator target loss, including region-aware loss, is shown below. Figure 2 As shown, during the training process, samples are taken from the parameter range covering three motion modes: stationary, linear motion, and uniform turning, to construct pairs of training data; the Adam optimizer is used to jointly optimize the generator and discriminator; the learning rate remains constant within a preset number of rounds, and then decays linearly; and the training process is stabilized by gradient clipping.

[0049] Figure 3 This is an overall flowchart of a video SAR vehicle shadow sample generation method combining a motion model, as described in Example 1. Figure 3 As shown, when obtaining video SAR vehicle shadow samples using a video SAR vehicle shadow sample generation method that combines motion models, the following steps are taken: First, an instantaneous shadow projection model of the vehicle is established; second, a physical shadow template library covering stationary, linear motion, and uniform turning states is generated using a dwell time map; then, an improved Pix2PixHD network is constructed to convert the physical templates into real SAR shadow images; next, an image conversion network is trained using a region-aware composite loss; finally, real labeled samples, physical template samples, and converted and enhanced samples are fused to construct an enhanced shadow dataset, which is then used to train the vehicle shadow detection model. This scheme, through the process of "physical modeling—image conversion—regional constraint—multi-source training," ensures that the generated samples not only conform to the video SAR shadow formation mechanism but also possess the speckle texture and background clutter statistical characteristics of real SAR images, thereby solving the problems of insufficient labeled samples for moving vehicle shadows in video SAR, lack of physical constraints in traditional data augmentation, and insufficient generalization ability of the detection model.

[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating video SAR vehicle shadow samples by combining motion models, characterized in that, include: S1. Obtain the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. Based on the geometric relationship of SAR imaging and the kinematic characteristics of the target, analyze the SAR shadow geometry of the target vehicle and construct a shadow dwell time evolution model. S2. Based on the shadow dwell time evolution model constructed in S1, a standard shadow template is generated. Shadow clutter ratio and speckle noise are introduced into the labeled shadow template to generate a physical shadow template. S3. Using the physical shadow template generated in S2 as a conditional input, construct an image transformation network and integrate a spatial adaptive normalization structure into the image transformation network. S4. Construct a composite loss function that includes region-aware loss, and use the composite loss function to train the image conversion network constructed in S3 to build the image conversion network model. S5. The image transformation network model constructed using S4 performs image transformation on the physical shadow template generated in S2 to generate transformed and enhanced shadow images, and constructs an enhanced shadow dataset. S6. Train the target detection model using the enhanced shadow dataset to generate a vehicle shadow detection model, and generate video SAR vehicle shadow samples based on the vehicle shadow detection model.

2. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 1, characterized in that, The geometric parameters in S1 include vehicle length, vehicle width, and vehicle height; the motion parameters include vehicle ground speed, vehicle heading angle, and vehicle turning angular velocity; and the radar observation parameters include radar ground sweep angle, synthetic aperture time, image resolution, and pixel scale. The geometric representation of the shadow includes the effective projection length of the target vehicle in the azimuth direction, the effective projection length of the target vehicle in the range direction, the effective projection length of the target vehicle in the range direction after radar ground sweep angle correction, and the shadow center trajectory. ； ； ； ； ； ； In the formula, Indicates the vehicle's heading angle. This represents the effective projection length of the target vehicle in the azimuth direction of the SAR image plane. Indicates the length of the vehicle. This represents the effective projection length of the target vehicle in the range direction of the SAR image plane. This indicates the effective projected length of the target vehicle in the range direction after radar sweep angle correction. Indicates the width of the vehicle. Indicates vehicle height. Indicates the radar's ground sweep angle. It is a cosine function. It is a sine function. Represents absolute value. , For time variables, representing moments in time. express The center of the shadow at that moment, , They represent Projection on the X and Y axes Indicates the initial position of the shadow center. express Vehicle ground speed at any given moment express The vehicle's heading angle at any given moment.

3. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 2, characterized in that, The shadow dwell time evolution model is as follows: ； In the formula, Indicates the pixel position The intensity of shadow dwell time at that location This indicates the number of sampling frames within the sliding time window. For indicator functions, express The shadow area of ​​time, Indicates the first in the sliding window Each sampling time, by Time corresponding , , , A joint decision.

4. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 3, characterized in that, Standard shadow templates include shadow templates for stationary targets, shadow templates for targets moving in a straight line, and shadow templates for targets turning at a constant speed. When the target vehicle is stationary, the vehicle's ground speed ,in accordance with , , Calculate the shadow area and construct a shadow template for a stationary target; When the target vehicle is in a straight-line motion state, the vehicle ground speed and vehicle heading angle remain unchanged during the synthetic aperture time. Based on the vehicle ground speed and vehicle heading angle, the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state is calculated. Based on the sweep displacement of the shadow in the azimuth and range directions under the straight-line motion state, the straight-line motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the straight-line motion trajectory of the shadow to construct a shadow template for a straight-line moving target. When the target vehicle is in a constant speed turning state, the vehicle's heading angle changes with time, while the vehicle's turning angular velocity remains constant. When the vehicle's turning angular velocity approaches zero, the target vehicle's constant speed turning state degenerates into a straight-line motion state. Based on the vehicle's turning angular velocity and the sweeping displacement of the shadow in the azimuth and range directions in the straight-line motion state, the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state is calculated. Based on the sweeping displacement of the shadow in the azimuth and range directions in the constant speed turning state, the uniform speed turning motion trajectory of the shadow is generated. The instantaneous shadow region is calculated based on the geometric parameters, motion parameters, and radar observation parameters of the target vehicle. The instantaneous shadow region is then accumulated along the uniform turning motion trajectory of the shadow to construct a uniform turning target shadow template.

5. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 4, characterized in that, The sweep length of the shadow in the azimuth direction during linear motion. for: ； The sweep length of the shadow in the distance direction during linear motion. for: ； The sweep length of the shadow in the azimuth direction during a constant speed turn. for: ； The sweep length of the shadow in the distance direction during a constant speed turn. for: ； ； ； In the formula, Indicates the time for synthesizing the aperture. Indicates the equivalent average sailing angle. Indicates the total angular span. Indicates the initial sailing angle. This indicates the vehicle's turning angular velocity.

6. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 5, characterized in that, The shadow image intensity in the physical shadow template is: ； ； In the formula, Indicates the pixel position The intensity of the shadow image at that location, Indicates the average clutter intensity. For shadow clutter ratio, This represents the average intensity of the background clutter region. Indicates the average intensity of the shaded area. This is speckle noise.

7. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 6, characterized in that, The image transformation network includes a generator and a multi-scale discriminator. The generator is used to generate a transformed and enhanced shadow image based on a physical shadow template, and the multi-scale discriminator is used to discriminate the transformed and enhanced shadow image generated by the generator based on the physical shadow template. The generator consists of an input layer, an encoder, a bottleneck module, a decoder, and an output layer. The encoder is used to extract multi-scale features of the physical shadow template, the bottleneck module is used to inject deep features into the spatial structure of the physical shadow template, and the decoder is used to recover the generated image. Shallow spatial information is passed between the encoder and decoder through skip connections. The output layer of the generator uses a hyperbolic tangent activation function to constrain the pixel values ​​of the transformed enhanced shadow image output by the generator. The multi-scale discriminator includes a fine-scale discriminator and a coarse-scale discriminator. The fine-scale discriminator operates at the original resolution and is used to constrain local speckle texture, shadow edges, and small-scale clutter details. The coarse-scale discriminator operates at the downsampled resolution and is used to constrain the overall shape of the shadow region and the global structural consistency.

8. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 7, characterized in that, The generator incorporates a spatial adaptive normalization module, which performs a normalization operation on the spatial structure based on the physical shadow template, and extracts the spatial features of the physical shadow template. ； In the formula, For spatial adaptive normalization, This represents the intermediate feature map in the generator. This represents the conditional feature map of the physical shadow template obtained by the lightweight convolutional extractor. Representing intermediate feature maps The channel mean, Representing intermediate feature maps The channel standard deviation, The scaling factor represents the spatial variation. Translation coefficients representing spatial changes This indicates element-wise multiplication.

9. The method for generating video SAR vehicle shadow samples by combining a motion model according to claim 8, characterized in that, In S4, the composite loss function for: ； ； ； ； ； ； ； In the formula, To counteract the loss, For pixel reconstruction loss term, For feature matching loss term, For the region perception loss term, , , They represent , , The weighting coefficients, Represents a physical shadow template. This represents the true SAR shadow image paired with the physical shadow template. This represents the transformed and enhanced shadow image output by the generator. Represents the Manhattan norm. These are intermediate parameters for calculating the feature matching loss. Indicates scale index. For the first The discriminant at each scale is at the 1st... The feature maps output by the intermediate layer The number of feature layers for normalization. Represents the loss term for the shaded area. This represents the background region loss term. , They represent , The weighting coefficients, Indicates a soft shadow mask. This represents the background mask.

10. A method for generating video SAR vehicle shadow samples by combining a motion model according to claim 9, characterized in that, In S4, after the image conversion network training is completed, the generator of the trained image conversion network model is used to convert the physical shadow template obtained in S2 to generate a converted and enhanced shadow image, and construct a converted and enhanced shadow sample. The enhanced shadow samples are fused with video SAR shadow samples and physical shadow template samples to obtain fused sample data. The fused sample data is then processed by image size normalization, shadow annotation transformation, label file generation, and training set partitioning to construct an enhanced shadow dataset. The enhanced shadow dataset is input into the object detection network for training to generate a vehicle shadow detection model; Obtain the video SAR image sequence to be detected, input the video SAR image sequence into the vehicle shadow detection model, and output the location box, category label and confidence of the vehicle shadow target to obtain video SAR vehicle shadow samples.

Images