Millimeter wave radar imaging denoising collaborative optimization method based on unbiased risk estimation
By employing a generative neural network self-supervised learning method based on unbiased risk estimation, combined with the U-Net structure and Stan unbiased risk estimation theory, the noise suppression and imaging problems of millimeter-wave radar in complex scenarios are solved, achieving high-quality adaptive imaging results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 724TH RESEARCH INSTITUTE OF CHINA STATE SHIPBUILDING CORP LTD
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-17
AI Technical Summary
Existing millimeter-wave radar imaging denoising methods have limited adaptability in complex and ever-changing real-world scenarios, rely on large-scale training data, and are limited in performance under strong noise conditions, making it difficult to achieve high-quality imaging and denoising synergistic optimization.
A generative neural network with unbiased risk estimation is adopted, which is optimized on a single frame of noisy data through self-supervised learning. Features are extracted by combining the U-Net structure, and noise is modeled and suppressed by Stan unbiased risk estimation theory. A loss function is constructed and iteratively optimized.
Achieving highly robust imaging under low signal-to-noise ratio and sparse sampling conditions, reconstructing high-quality radar images with clear contours, rich details, and few artifacts, eliminating dependence on large-scale training data, and possessing excellent noise suppression capabilities.
Smart Images
Figure CN121878685A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to computational imaging and radar signal processing technology, and in particular to a collaborative optimization method for millimeter-wave radar imaging denoising based on unbiased risk estimation. Background Technology
[0002] Traditional millimeter-wave radar imaging denoising methods, such as filtering-based methods or sparse representation-based methods, are effective under certain conditions, but often rely on precise prior assumptions about the statistical properties of noise or the sparsity of the signal, and have limited adaptability in complex and ever-changing real-world scenarios.
[0003] With the development of deep learning, supervised denoising methods based on convolutional neural networks have achieved significant success. These methods learn the mapping from noisy images to noiseless images by training on a large number of noisy-noise pairs. However, their application in radar imaging is limited by two major bottlenecks: First, obtaining large-scale, high-quality paired radar training data is extremely costly, and in many scenarios, it is even impractical; second, models trained on specific datasets often have limited generalization ability, and their performance may significantly degrade when faced with new scenes or new noise types that do not match the distribution of the training data.
[0004] To avoid dependence on large-scale training data, unsupervised imaging frameworks such as Implicit Neural Representation (INR) have been proposed. These methods utilize the structural priors of deep networks to directly optimize on a single frame of noisy data. However, existing INR methods typically focus on recovering structural information from sparse sampling, with insufficient consideration for noise modeling and suppression mechanisms. This makes it difficult to achieve effective synergistic optimization between imaging and denoising, and performance remains limited in highly noisy environments.
[0005] Therefore, developing a new radar imaging method that does not require external datasets, can adaptively estimate and suppress noise from a single noisy measurement, and simultaneously achieve high-quality scene reconstruction has significant theoretical and practical value. Summary of the Invention
[0006] The purpose of this invention is to provide a millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation, so as to achieve highly robust millimeter-wave radar imaging under low signal-to-noise ratio and sparse sampling conditions.
[0007] The technical solution to achieve the objective of this invention is: a collaborative optimization method for millimeter-wave radar imaging denoising based on unbiased risk estimation, comprising the following steps:
[0008] Step 1: Acquire noisy radar echo data Initialize an untrained system with optimizable parameters. Parameterized Generative Neural Networks ;
[0009] Step 2: Extract noisy radar echo data Input Generative Neural Network The initial scene scattering rate distribution estimate is obtained. ;
[0010] Step 3: Perform a two-dimensional inverse Fourier transform on the scene scattering rate distribution estimate, transform it to the echo signal domain, and obtain the first reconstructed echo signal. ;
[0011] Step 4: Generate a random noise disturbance that conforms to preset statistical characteristics. And superimposed on the noisy echo data Above, the disturbance input is obtained. ;
[0012] Step 5: Input the disturbance Input Generative Neural Network The perturbed scene scattering rate distribution is estimated, and then transformed to the echo signal domain using a two-dimensional inverse Fourier transform to obtain the second reconstructed echo signal. ;
[0013] Step 6: Construct the loss function, which includes:
[0014] The data consistency term constrains the consistency between the first reconstructed echo signal and the noisy radar echo data;
[0015] The noise estimation term, based on Stan's unbiased risk estimation theory, estimates the noise using the first reconstructed echo signal and the second reconstructed echo signal.
[0016] Regularization term, used to impose sparsity on the scattering rate distribution estimate;
[0017] Step 7: Calculate the loss function and update the parameters of the generative neural network using the backpropagation algorithm. ;
[0018] Step 8: Iterate through steps 2 to 7 until the preset number of iterations or convergence condition is reached;
[0019] Step 9: Output the final optimized generative neural network's estimate of the scene scattering rate distribution generated by the noisy echo data, as the denoised millimeter-wave radar imaging result.
[0020] Furthermore, the generative neural network An encoder-decoder structure is adopted, and the features extracted by the encoder at different scales are fused with the output of the decoder at the corresponding scales through skip connections.
[0021] Furthermore, the encoder-decoder structure is a U-Net network.
[0022] Furthermore, the random noise disturbance It is Gaussian white noise.
[0023] Furthermore, the objective function The specific form is:
[0024]
[0025] Where i represents the radar echo signal number. For noise variance, is the regularization coefficient, and m is the number of echo points.
[0026] Furthermore, the Monte Carlo method is used to approximate the divergence term in the Stan unbiased risk estimate.
[0027] A hybrid model predictive control method for addressing control delay in large vehicles is provided. The hybrid model predictive control method for addressing control delay in large vehicles is implemented to achieve the hybrid model predictive control method for addressing control delay in large vehicles.
[0028] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the hybrid model predictive control method for large vehicle control delay, thereby realizing the hybrid model predictive control method for large vehicle control delay.
[0029] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the hybrid model predictive control method for large vehicle control delay is implemented, thereby realizing the hybrid model predictive control method for large vehicle control delay.
[0030] Compared with existing technologies, the significant advantages of this invention are: 1) Self-supervised and zero-shot learning: This method requires no "noisy-noise-free" training data pairs and optimizes directly on a single frame of noisy echo to be processed, eliminating the dependence on large-scale labeled datasets. 2) Excellent noise suppression capability: By introducing SURE theory and using a noise perturbation estimation mechanism, this method can adaptively model and suppress noise. 3) High-quality and robust imaging: By co-optimizing the imaging and denoising processes and combining U-Net to fully utilize echo structure information, this method can reconstruct high-quality radar images with clear contours, rich details, and few artifacts, even under the dual challenges of sparse sampling and strong noise. Attached Figure Description
[0031] Figure 1 This is a flowchart illustrating the overall framework of the imaging method in this embodiment of the invention.
[0032] Figure 2 This is a schematic diagram of the U-Net network structure used in this embodiment of the invention.
[0033] Figure 3 This is the imaging scenario of the method of this invention in the actual measured data experiment.
[0034] Figure 4 The image shows the imaging results of the method of the present invention under actual measured data. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0036] A millimeter-wave radar imaging denoising collaborative optimization network based on unbiased risk estimation includes:
[0037] Step 1: Acquire noisy radar echo data Let the number of iterations be n = 1; and n = 1, ..., num; where num is the total number of iterations. Construct an untrained dataset with optimizable parameters. Parameterized Generative Neural Networks It is configured to output the scene scattering coefficient after the input echo.
[0038] The generative neural network mentioned in step 1 Its structure is based on the classic encoder-decoder U-Net architecture, and its overall design is as follows: Figure 2 As shown, the entire network mainly consists of two paths: an encoding path for extracting semantic features and a decoding path for progressively restoring spatial resolution to generate the output image.
[0039] The generative neural network uses the UNet structure as follows:
[0040] During the encoding phase, the input image first undergoes a series of convolutional layers for feature extraction. As the network depth increases, the number of channels in the feature map expands layer by layer, gradually increasing from an initial 1 channel to 64, 128, 256, 512, and finally reaching 1024 channels. Each encoding unit typically includes a 3×3 convolutional layer for extracting local spatial features. This is followed by batch normalization layers to stabilize the training process and accelerate convergence; subsequently, a non-linear mapping is introduced through the ReLU activation function to improve the model's fitting ability and expressive power. Simultaneously, the network uses max-pooling layers for downsampling, gradually reducing the spatial size of the feature map, thereby achieving stronger information compression and a larger receptive field.
[0041] The decoding stage is symmetrical to the encoding stage, with the main goal of restoring the spatial resolution of the image layer by layer while fusing feature information extracted in the encoding stage to preserve key structural details. In terms of channel settings, the decoding path is symmetrical to the encoding path, with the number of channels decreasing sequentially from 1024 to 512, 256, 128, and 64. Finally, a 1×1 convolutional layer maps the number of channels to the required output dimension, ensuring that the output feature map matches the category or channel requirements of the imaging task. Each decoding module performs upsampling through bilinear interpolation, doubling the feature map size compared to the previous layer. Simultaneously, feature maps from the corresponding layers in the encoding path are extracted and fused into the current decoding layer through channel concatenation to enrich the feature information. A 3×3 convolutional layer further refines the local structure, and batch normalization and the ReLU activation function further enhance expressive power.
[0042] Step 2: Extract noisy radar echo data The input to the generative neural network is mapped to a scene scattering rate distribution to obtain the initial scene scattering rate distribution. ;
[0043] Step 3: Perform a two-dimensional inverse Fourier transform on the initial scene scattering rate distribution estimate to transform it into the echo signal domain, obtaining the first reconstructed echo signal. ;
[0044] Step 4: Generate a random noise disturbance that conforms to preset statistical characteristics. This is then superimposed onto the noisy radar echo data to obtain a disturbance input. ;
[0045] Step 5: Input the disturbance The input is fed into the network, and the output scene scattering rate distribution is subjected to inverse Fourier transform: ;
[0046] Step 6: Define the imaging problem as an unsupervised risk minimization problem and construct the loss function. The loss function comprises three parts: the first part is a data consistency term used to ensure that the reconstructed echo matches the actual acquired radar echo; the second part is a generative neural network-based loss function for noisy echo data. and disturbance input The noise estimation term for the response difference estimation; the third part is the regularization term for the sparsity estimation of the constrained scene scattering rate distribution;
[0047] Let the observed noisy echo be Design an unsupervised imaging module To make its output As close as possible to noise-free echo .
[0048] The entire unsupervised imaging module It consists of two cascaded parts: .in For a parameter Controlled deep neural network, input is noisy echo The output is the scene scattering distribution. . This is a physical backward transform operator, specifically the two-dimensional inverse Fourier transform (IFFT) in this invention, which transforms the scattering rate distribution... Transform back into the echo domain to obtain the denoised echo estimate. .
[0049] According to the SURE theory, in the unknown In the case of mean square error An unbiased estimate It can be represented as:
[0050] in yes about The divergence. Directly calculating the divergence is very difficult; this invention uses the Monte Carlo method to approximate it:
[0051] in τ is a random vector sampled from a standard normal distribution N(0,1), where τ is a small positive number.
[0052] By applying a sparsity prior L1 norm to the scattering coefficients of the network output, the final optimization objective function L(θ) is defined as:
[0053]
[0054] Right now:
[0055] The first term of the loss function constrains the correlation between the scattering distribution of the network output and the input signal, thus ensuring that the network input contains noisy echoes. Then, the output scene scattering distribution is transformed from the sampled signal. The signal should be consistent with the input echo. The second term of the loss function, the Monte Carlo estimation term, constrains the network's denoising ability by injecting random noise samples into the input data to generate multiple perturbation versions of the data, and utilizes the differences in the network's response to these perturbation data to reduce noise. The third term of the loss function applies a sparse prior L1 norm to the scattering coefficients of the network output, constraining the sparsity of the scattering coefficients.
[0056] Step 7: Calculate the loss function and iteratively optimize the parameters of the generative neural network using the backpropagation algorithm. ;
[0057] Step 8: Increment the value of n by 1, and repeat steps 2 through 7 until n = num;
[0058] Step 9: Output the final optimized scattering rate distribution estimate generated by the generative neural network. This is the imaging result of millimeter-wave radar.
[0059] Example
[0060] To verify the effectiveness of the present invention, the following experimental design was conducted.
[0061] Real-world data acquisition was conducted using a TI AWR2243 millimeter-wave radar, mounted on the side roof of a vehicle. The radar scanned a parking lot scene to obtain data. The radar system's carrier frequency was 77 GHz, and the modulation frequency was 70.295 THz / s. During the experiment, the target angle and the vehicle were located within 6-8 meters of the radar.
[0062] See the experimental scenario Figure 3 As shown, the experimental results are as follows: Figure 4 As shown, from Figure 4 Visually, the method proposed in this invention performs well in target imaging. The target area within the red-marked box shows strong signal and clear edges, with a complete vehicle target outline and almost no obvious artifact interference. Simultaneously, the suppression of background noise is also very significant, resulting in a clean image background and good contrast between the target and background, making the target stand out more. Experimental results strongly demonstrate that the noise perturbation estimation-based imaging algorithm proposed in this invention can adaptively cope with the challenges posed by strong noise and data sparsity without any prior training.
[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0064] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A collaborative optimization method for millimeter-wave radar imaging denoising based on unbiased risk estimation, characterized in that, Includes the following steps: Step 1: Acquire noisy radar echo data Initialize an untrained system with optimizable parameters. Parameterized Generative Neural Networks ; Step 2: Extract noisy radar echo data Input Generative Neural Network The initial scene scattering rate distribution estimate is obtained. ; Step 3: Perform a two-dimensional inverse Fourier transform on the scene scattering rate distribution estimate, transform it to the echo signal domain, and obtain the first reconstructed echo signal. ; Step 4: Generate a random noise disturbance that conforms to preset statistical characteristics. And superimposed on the noisy echo data Above, the disturbance input is obtained. ; Step 5: Input the disturbance Input Generative Neural Network The perturbed scene scattering rate distribution is estimated, and then transformed to the echo signal domain using a two-dimensional inverse Fourier transform to obtain the second reconstructed echo signal. ; Step 6: Construct the loss function, which includes: The data consistency term constrains the consistency between the first reconstructed echo signal and the noisy radar echo data; The noise estimation term, based on Stan's unbiased risk estimation theory, estimates the noise using the first reconstructed echo signal and the second reconstructed echo signal. Regularization term, used to impose sparsity on the scattering rate distribution estimate; Step 7: Calculate the loss function and update the parameters of the generative neural network using the backpropagation algorithm. ; Step 8: Iterate through steps 2 to 7 until the preset number of iterations or convergence condition is reached; Step 9: Output the final optimized generative neural network's estimate of the scene scattering rate distribution generated by the noisy echo data, as the denoised millimeter-wave radar imaging result.
2. The millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation according to claim 1, characterized in that, The generative neural network An encoder-decoder structure is adopted, and the features extracted by the encoder at different scales are fused with the output of the decoder at the corresponding scales through skip connections.
3. The millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation according to claim 1, characterized in that, The encoder-decoder structure is a U-Net network.
4. The millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation according to claim 1, characterized in that, The random noise disturbance It is Gaussian white noise.
5. The millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation according to claim 1, characterized in that, The objective function The specific form is as follows: ； Where i represents the radar echo signal number. For noise variance, is the regularization coefficient, and m is the number of echo points.
6. The millimeter-wave radar imaging denoising collaborative optimization method based on unbiased risk estimation according to claim 1, characterized in that, An approximation of the divergence term in the Stan unbiased risk estimate was made using the Monte Carlo method.
7. A hybrid model predictive control method for addressing control delay in large vehicles, characterized in that, Implement the hybrid model predictive control method for large vehicle control delay as described in any one of claims 1-6 to realize the hybrid model predictive control method for large vehicle control delay.
8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the hybrid model predictive control method for large vehicle control delay as described in any one of claims 1-6, thereby realizing the hybrid model predictive control method for large vehicle control delay.
9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the hybrid model predictive control method for large vehicle control delay as described in any one of claims 1-6, thereby realizing the hybrid model predictive control method for large vehicle control delay.