Method and system for radar image multi-source noise suppression based on variational inference

Abstract

The present application belongs to the technical field of noise suppression, and particularly relates to a radar image multi-source noise suppression method and system based on variational inference. The method comprises: preprocessing the collected original radar image; constructing a probability distribution model for each type of noise and integrating them into a unified multi-source mixed noise model; performing noise estimation based on variational inference to generate a pixel-level noise intensity map; performing multi-scale convolution feature extraction to obtain a multi-scale enhanced feature map; performing dual-domain denoising in the frequency domain and the spatial domain to obtain a dual-domain denoising feature map; performing attention enhancement on the dual-domain denoising feature map to output an attention-enhanced feature map; performing reconstruction verification and optimization using a variational autoencoder to obtain a reconstruction-optimized feature map; performing feature map to image conversion, structural similarity verification and edge preservation verification on the reconstruction-optimized feature map to output a final denoised radar image. The present application realizes efficient suppression of complex noise and accurate preservation of key target features of the radar image.

Description

Technical Field

[0001] This invention belongs to the field of noise suppression technology, specifically relating to a method and system for suppressing multi-source noise in radar images based on variational inference. Background Technology

[0002] Radar imaging plays a crucial role in military reconnaissance, maritime surveillance, aerospace remote sensing, and weather forecasting. However, in actual data acquisition, complex environments introduce various types of noise, including speckle noise, Gaussian thermal noise, and coherent interference. These noises often appear in mixed forms, leading to a significant deterioration in image quality and impacting downstream target detection and recognition tasks.

[0003] Traditional denoising methods such as mean filtering, Wiener filtering, and wavelet transform are effective in single-noise scenarios, but often perform poorly in multi-source noise environments, and can cause problems such as edge blurring and target contour loss. Deep learning-based denoising methods have been widely used in recent years, but most assume that noise follows a fixed distribution and lack the ability to dynamically model noise uncertainties. Furthermore, ordinary network models struggle to fully capture the complex coupling relationship between noise and target features.

[0004] Therefore, there is a need for a method and system that can uniformly model multi-source noise and achieve high-quality denoising through adaptive inference, so as to ensure that radar images with clear structure and complete features can still be obtained in complex environments. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for suppressing multi-source noise in radar images based on variational inference. By accurately probabilistically modeling multi-source noise and adaptively estimating by variational inference, combined with frequency-space dual-domain collaborative denoising and closed-loop optimization, the invention achieves efficient suppression of complex noise and accurate preservation of key target features in radar images.

[0006] To achieve the above objectives, this invention provides a method and system for suppressing multi-source noise in radar images based on variational inference, comprising the following steps: S1. Preprocess the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images; S2. Based on the noise characteristics of the preprocessed image, construct probability distribution models for each type of noise and integrate them into a unified multi-source mixed noise model. S3. Based on the preprocessed image and the multi-source mixed noise model, perform noise estimation based on variational inference to generate a pixel-level noise intensity map; S4. Based on the preprocessed image and pixel-level noise intensity map, multi-scale convolutional feature extraction is performed using parallel convolutional branches to obtain multi-scale enhanced feature maps. S5. Perform frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map; S6. Enhance the attention of the dual-domain denoised feature map by using channel attention branch and spatial attention branch, and output the attention-enhanced feature map; S7. The attention-enhanced feature map is reconstructed, verified, and optimized using a variational autoencoder to obtain the reconstructed and optimized feature map. S8. Obtain standard clean image samples, perform feature map to image conversion, structural similarity verification and edge preservation verification on the reconstructed optimized feature map, and output the final denoised radar image.

[0007] As a preferred embodiment of the present invention, in step S1, the original radar image is a grayscale image, and the pixel value represents the radar echo intensity. The preprocessing specifically includes: Amplitude normalization: Maps the pixel values ​​of the original radar image to the [0, 1] interval to eliminate the difference in intensity magnitude caused by target distance and scattering intensity in different regions; Phase compensation: In a normalized image, identify noise-free, uniform regions and calculate the average phase value of these regions as the reference phase. For each pixel in the normalized image, subtract the offset of the global average phase relative to the reference phase from the phase value of each pixel. This is expressed as: ; In the formula, After phase compensation, the radar image is in coordinates The phase value of the pixel at that location; After amplitude normalization, the radar image in coordinates The original phase of the pixel; This represents the average phase of the entire radar image after amplitude normalization. The average reference phase of the noise-free uniform region; Logarithmic compression: Applying a logarithmic transformation to the phase-compensated image reduces the dynamic range and prevents weak target features from being masked by strong targets. The formula is: ; In the formula, After logarithmic compression, the radar image is in coordinates The grayscale value of the pixel at that location; , To adjust the parameters; After phase compensation processing, the radar image is in coordinates The grayscale value of the pixel at that location; After logarithmic compression, a preprocessed image with uniform intensity, no obvious phase distortion, and recognizable weak target features is obtained.

[0008] As a preferred embodiment of the present invention, in S2, the various types of noise are specifically speckle noise, Gaussian thermal noise, and coherent interference. The probability distribution model is constructed as follows: for speckle noise, Rayleigh distribution is used to describe its probability distribution; for Gaussian thermal noise, normal distribution is used to describe its probability distribution; for coherent interference, Poisson-Gaussian mixed distribution is used to describe its probability distribution. By integrating Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution, a joint probability multi-source mixed noise model containing all noise features is formed. The joint probability density function of the multi-source mixed noise is the product of the single-source noise distributions.

[0009] As a preferred embodiment of the present invention, in S3, the noise estimation based on variational inference specifically includes: A three-component Gaussian mixture distribution is introduced as an approximation of the true noise distribution; Based on the multi-source mixed noise model, Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution are respectively set as the initial prior distributions for speckle noise, Gaussian thermal noise, and coherent interference. With the goal of minimizing KL divergence and maximizing the lower bound of evidence, the Adam optimizer is used to iteratively update the variational parameters of the approximate distribution using the lower bound of evidence as the objective function. Once the optimization process converges, the noise intensity of each pixel in the preprocessed image is calculated using the optimized approximate distribution, generating a pixel-level noise intensity map that can label the noise strength pixel by pixel.

[0010] As a preferred embodiment of the present invention, the process of obtaining the multi-scale enhanced feature map in S4 is as follows: Construct a multi-branch parallel convolutional structure containing three parallel convolutional branches, each corresponding to a different scale: Small-scale branches, with 3×3 convolution kernels, capture local texture details of the image; The medium-scale branch, with a 5×5 convolution kernel, captures image edge features; Large-scale branches, with 7×7 convolution kernels, capture the global structure of the image; The preprocessed image is input into three parallel convolutional branches for feature extraction. The output feature maps of the three branches are concatenated, and channel attention is used to generate channel weights. The output features of different branches are then weighted and fused to obtain a fused feature map. After normalizing the pixel-level noise intensity map, it is multiplied pixel by pixel with the fused feature map to obtain the multi-scale enhanced feature map.

[0011] As a preferred embodiment of the present invention, the process of obtaining the dual-domain denoising feature map in S5 includes: Frequency domain denoising: Using Fourier transform, a two-dimensional fast Fourier transform is performed on each channel of the multi-scale enhanced feature map to transform the image from the spatial domain to the frequency domain. Noise frequency suppression: Calculate the frequency domain amplitude spectrum, set a threshold to distinguish between signal frequency and noise frequency, weaken high-frequency noise through a Gaussian low-pass filter, and apply amplitude attenuation to noise frequency points to preserve signal frequency; By using inverse Fourier transform, the frequency domain features after noise frequency suppression are converted back to the spatial domain to obtain the frequency domain denoising features. Spatial denoising: A residual convolutional network containing five residual blocks is used to recover the residual texture details and target structure after frequency domain denoising, thus obtaining spatial domain denoising features. Dual-domain feature fusion: After compressing the number of channels by 1×1 convolution of the frequency domain denoising features, they are added to the spatial domain denoising features pixel by pixel to obtain a dual-domain denoising feature map.

[0012] As a preferred embodiment of the present invention, the process of outputting the attention-enhanced feature map in S6 includes: In the channel attention branch, the global average value is calculated for each channel of the dual-domain denoising feature map to obtain the channel feature vector; the channel feature vector is compressed and activated through a fully connected layer to generate channel weights; the channel weights are multiplied with the dual-domain denoising feature map channel by channel to obtain the feature map after channel attention processing. In the spatial attention branch, max pooling and average pooling are performed on the dual-domain denoised feature map in the channel dimension to obtain two spatial feature maps. The two spatial feature maps are concatenated and then compressed and activated by 1×1 convolution to generate spatial weights. The spatial weights are multiplied pixel by pixel with the feature map after channel attention processing to obtain the feature map after spatial attention processing. Residual connection: Add the dual-domain denoised feature map to the feature map after spatial attention processing to output the attention-enhanced feature map.

[0013] As a preferred embodiment of the present invention, in step S7, the reconstructed and optimized feature map is obtained as follows: The variational autoencoder adopts an encoder + decoder architecture. The encoder compresses the input attention-enhanced feature map through a convolutional layer, maps it to the latent space, and outputs the mean and variance of the latent variable. The latent variable is obtained by sampling based on the mean and variance. The decoder reconstructs the latent variable into a reconstructed feature map with the same dimension as the input feature map through a transposed convolutional layer. Calculate the reconstruction error between the attention-enhanced feature map and the reconstructed feature map. If the error exceeds the preset error threshold, feed the error signal back to the multi-source mixed noise model, adjust the parameters of the multi-source mixed noise model, and repeat steps S2-S7 until the reconstruction error does not exceed the preset error threshold range.

[0014] As a preferred embodiment of the present invention, the process of outputting the final denoised radar image in S8 includes: Feature map to image conversion: The number of channels in the reconstructed and optimized feature map is compressed from 64 to 1 through 1×1 convolution to obtain a single-channel feature map. Then, the pixel values ​​are restored to the intensity range of the original radar image through linear mapping to obtain a preliminary denoised image. Structural similarity verification: Using standard clean image samples as a reference, calculate the structural similarity index SSIM between the preliminary denoised image and the standard clean image. If SSIM is greater than or equal to the set similarity threshold, the structural integrity meets the standard, and edge preservation verification is performed; if it does not meet the standard, the attention parameters of the channel attention branch and the spatial attention branch are adjusted and optimized again. Edge preservation verification: The Canny operator is used to extract edges from the preliminary denoised image and the standard clean image respectively to obtain edge maps. The intersection-union ratio (IoU) of the two edge maps is calculated. If the IoU is greater than or equal to the set IoU threshold, the edge preservation meets the standard, and the final denoised radar image is output. If it does not meet the standard, the spatial denoising parameters in the frequency-spatial dual-domain denoising are adjusted and re-optimized.

[0015] A radar image multi-source noise suppression system based on variational inference, used to implement the above method, includes: The preprocessing module preprocesses the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images. The multi-source mixing module constructs probability distribution models for various types of noise based on the noise characteristics of the preprocessed image, and integrates them into a unified multi-source mixed noise model. The noise estimation module performs variational inference-based noise estimation based on the preprocessed image and a multi-source mixed noise model, generating a pixel-level noise intensity map. The multi-scale enhancement module, based on the preprocessed image and pixel-level noise intensity map, uses parallel convolution branches to perform multi-scale convolution feature extraction to obtain a multi-scale enhanced feature map. The dual-domain denoising module performs frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map. The attention enhancement module enhances the attention of the dual-domain denoised feature map through channel attention branch and spatial attention branch, and outputs the attention-enhanced feature map; The reconstruction and optimization module uses a variational autoencoder to reconstruct, verify, and optimize the attention-enhanced feature map, resulting in a reconstructed and optimized feature map. The feature preservation module acquires standard clean image samples, performs feature map to image conversion, structural similarity verification, and edge preservation verification on the reconstructed and optimized feature maps, and outputs the final denoised radar image.

[0016] The beneficial effects of this invention are: This invention effectively overcomes the bottleneck of traditional radar image denoising methods' insufficient adaptability to complex noise. It constructs a multi-source mixed probability model for three typical noise types: speckle, Gaussian heat, and coherent interference. Combining variational inference, it approximates the true noise distribution with a three-component Gaussian mixture distribution. The Adam optimizer achieves accurate pixel-level noise intensity estimation, solving the problems of misidentification of noise type and poor suppression targeting caused by traditional single-distribution modeling. Compared to fixed-threshold filtering or single Gaussian assumption methods, it can adaptively adjust the suppression strategy according to the dominant noise type of different pixels, achieving accurate capture of noise intensity and distribution characteristics even in high-noise-density scenes, providing a reliable basis for subsequent targeted denoising.

[0017] This invention achieves synergistic optimization of noise suppression and feature preservation. It extracts multi-scale image features through parallel convolutional branches, strengthens key feature weights using a channel attention mechanism, and employs a dual-domain processing architecture—frequency-domain Gaussian low-pass filtering to suppress periodic noise and spatial-domain residual network to refine details—effectively balancing noise removal intensity and target structure fidelity. By leveraging the reconstruction error feedback of a variational autoencoder and verifying the results with SSIM and PSNR metrics, a closed-loop optimization process of modeling, estimation, suppression, and verification is formed. This avoids the edge blurring and texture loss problems common in traditional methods. The denoised image exhibits significantly improved target contour clarity and background texture integrity, directly meeting the high-fidelity radar image requirements of downstream tasks such as topographic mapping and target detection. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the principle of the method of the present invention; Figure 2 This is a schematic diagram of the modules of the system of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, the radar image multi-source noise suppression method based on variational inference includes the following steps: S1. Preprocess the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images; S2. Based on the noise characteristics of the preprocessed image, construct probability distribution models for each type of noise and integrate them into a unified multi-source mixed noise model. S3. Based on the preprocessed image and the multi-source mixed noise model, perform noise estimation based on variational inference to generate a pixel-level noise intensity map; S4. Based on the preprocessed image and pixel-level noise intensity map, multi-scale convolutional feature extraction is performed using parallel convolutional branches to obtain multi-scale enhanced feature maps. S5. Perform frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map; S6. Enhance the attention of the dual-domain denoised feature map by using channel attention branch and spatial attention branch, and output the attention-enhanced feature map; S7. The attention-enhanced feature map is reconstructed, verified, and optimized using a variational autoencoder to obtain the reconstructed and optimized feature map. S8. Obtain standard clean image samples, perform feature map to image conversion, structural similarity verification and edge preservation verification on the reconstructed optimized feature map, and output the final denoised radar image.

[0020] The acquisition of raw radar images involves matching appropriate radar hardware systems and data acquisition processes to specific application scenarios (such as military reconnaissance, disaster monitoring, and land resource exploration). First, the appropriate radar equipment is selected based on the scenario requirements. For example, airborne synthetic aperture radar (SAR) or ground-based phased array radar can be used for military reconnaissance, while satellite-borne SAR (such as the radar system carried by satellites like Sentinel-1) or airborne side-looking radar can be used for disaster monitoring and land resource exploration. Then, the selected radar equipment actively transmits electromagnetic waves towards the target area and simultaneously receives the echo signals reflected by the target. The built-in data acquisition module of the equipment converts the analog signals such as amplitude and phase of the echo signals into digital signals and records key metadata of the radar platform (such as azimuth, flight altitude, imaging mode, resolution parameters, etc.). Finally, the digital echo signals and metadata are integrated and encapsulated according to the common radar data format (such as the CEOS and HDF5 formats commonly used for SAR data) to form the raw radar grayscale image, providing input data for the subsequent multi-scale speckle noise adaptive suppression stage.

[0021] In S1, the original radar image is a grayscale image, and the pixel value represents the radar echo intensity. It suffers from large intensity differences, phase distortion, and dynamic range imbalance. The preprocessing specifically includes: Amplitude normalization: Maps the pixel values ​​of the original radar image to the [0, 1] interval to eliminate the difference in intensity magnitude caused by target distance and scattering intensity in different regions, and avoids excessive interference of strong scattering regions on the noise model; Phase compensation: In a normalized image, a noise-free, uniform region (such as a calm sea surface or open land) is identified, and the average phase value of this region is calculated as the reference phase. For the phase value of each pixel in the normalized image, the offset of the global average phase relative to the reference phase is subtracted to correct the phase distortion caused by atmospheric refraction and antenna errors, ensuring image phase consistency. This is expressed as: ; In the formula, After phase compensation, the radar image is in coordinates The phase value of the pixel at that location; After amplitude normalization, the radar image in coordinates The original phase of the pixel; This represents the average phase of the entire radar image after amplitude normalization. The average reference phase of the noise-free uniform region; Alternatively, the reference phase can be subtracted from the phase of each pixel in the original image for correction.

[0022] Logarithmic compression: To address the problem of excessive dynamic range between strong scattering targets (such as ships and aircraft) and weak scattering backgrounds (such as sea clutter), a logarithmic transformation is applied to the phase-compensated image to reduce the dynamic range and prevent weak target features from being masked by strong targets. The formula is: ; In the formula, After logarithmic compression, the radar image is in coordinates The grayscale value of the pixel at that location; , To adjust the parameters; After phase compensation processing, the radar image is in coordinates The grayscale value of the pixel at that location; After logarithmic compression, a preprocessed image with uniform intensity, no obvious phase distortion, and identifiable weak target features is obtained, eliminating the influence of acquisition error and providing a stable and uniform image input for subsequent noise modeling.

[0023] In S2, the various types of noise are specifically speckle noise, Gaussian thermal noise, and coherent interference. The probability distribution models are constructed as follows: For speckle noise (generated by the interference of scattering point echoes during radar coherent imaging, appearing as granules), Rayleigh distribution is used to describe its probability distribution; for Gaussian thermal noise (generated by the thermal motion of receiver electronic components, appearing as random fluctuations), normal distribution is used to describe its probability distribution; for coherent interference (generated by crosstalk from other radar and communication equipment signals, appearing as stripes / pulses), Poisson-Gaussian mixed distribution is used to describe its probability distribution (taking into account both pulse characteristics and randomness). Assuming the preprocessed image consists of a clean image and multi-source mixed noise, and combining the superposition characteristics of the three types of noise (speckle is multiplicative, Gaussian and coherent interference are additive), Rayleigh distribution, normal distribution and Poisson-Gaussian mixture distribution are integrated to form a joint probabilistic multi-source mixed noise model that includes all noise features. The joint probability density function of the multi-source mixed noise is the product of the single-source noise distributions, thus achieving an overall accurate description of the multi-source noise.

[0024] In S3, the noise estimation based on variational inference is specifically as follows: A three-component Gaussian mixture distribution (corresponding to three types of noise) is introduced as an approximate distribution of the true noise distribution, including variational parameters such as component weights, mean, and variance; Based on the multi-source mixed noise model, Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution are respectively set as the initial prior distributions for speckle noise, Gaussian thermal noise, and coherent interference. With the goal of minimizing KL divergence and maximizing the lower bound of evidence, the Adam optimizer is used to iteratively update the variational parameters of the approximate distribution using the lower bound of evidence as the objective function. Once the optimization process converges, the noise intensity of each pixel in the preprocessed image is calculated using the optimized approximate distribution, generating a pixel-level noise intensity map that labels the noise intensity of each pixel.

[0025] By using variational inference, the true distribution of noise is approximated, enabling adaptive estimation of the noise intensity of each pixel in the image. Since the true distribution of noise is difficult to calculate directly, a Gaussian mixture distribution is introduced as an approximation. By minimizing the KL divergence (which measures the difference between the two distributions), the approximation distribution is made closer to the true distribution. At the same time, the lower bound of evidence (ELBO) is maximized to balance the model's fit to the data with the reasonableness of the distribution.

[0026] In S4, the process of obtaining the multi-scale enhanced feature map is as follows: Construct a multi-branch parallel convolutional structure containing three parallel convolutional branches, each corresponding to a different scale: Small-scale branches, with 3×3 convolution kernels, capture local texture details of the image (such as target surface texture and small-scale clutter). The medium-scale branch, with a 5×5 convolution kernel, captures image edge features (such as target contour edges and noise stripe edges). Large-scale branches, with 7×7 convolution kernels, capture the global structure of the image (such as the overall shape of the target and the distribution of the background). The preprocessed image is input into three parallel convolutional branches for feature extraction. The output feature maps of the three branches are concatenated and channel weights are generated through channel attention (global average pooling, channel compression, and activation and restoration). The output features of different branches are then weighted and fused to obtain a fused feature map. After normalizing the pixel-level noise intensity map, it is multiplied pixel by pixel with the fused feature map to enhance the features of the noisy regions (to facilitate subsequent targeted denoising) and retain the features of the noisy regions (to avoid over-processing), resulting in a multi-scale enhanced feature map (containing local, edge, and global features, with the features of high-noise regions already enhanced).

[0027] By capturing local texture, edge details, and global structure of images through multi-scale convolution, and combining this with noise intensity-guided feature enhancement, comprehensive and targeted feature support is provided for subsequent denoising.

[0028] In S5, the process of obtaining the dual-domain denoised feature map includes: Frequency domain denoising: By using Fourier transform, a two-dimensional fast Fourier transform is performed on each channel of the multi-scale enhanced feature map to transform the image from the spatial domain to the frequency domain. Periodic noise is represented as discrete bright spots in the frequency domain, which is easy to identify. Noise frequency suppression: Calculate the frequency domain amplitude spectrum, set a threshold to distinguish between signal frequency and noise frequency, weaken high-frequency noise through a Gaussian low-pass filter, and apply amplitude attenuation to noise frequency points to preserve signal frequency; By using inverse Fourier transform, the frequency domain features after noise frequency suppression are converted back to the spatial domain to obtain the frequency domain denoising features. Spatial denoising: A ResNet residual convolutional network containing five residual blocks is used to recover the residual texture details and target structure after frequency domain denoising, thus obtaining spatial domain denoising features. Each residual block consists of convolution, batch normalization, activation, convolution, and batch normalization, and residual connections are added to avoid gradient vanishing. Dual-domain feature fusion: After compressing the number of channels by 1×1 convolution, the frequency domain denoising features are added to the spatial domain denoising features pixel by pixel to obtain a dual-domain denoising feature map, achieving complementarity between frequency domain denoising and spatial domain detail restoration.

[0029] In the dual-domain denoising feature map, periodic noise has been suppressed and texture details have been initially restored. Periodic noise (such as coherent interference fringes) is specifically suppressed in the frequency domain, and texture details are accurately restored in the spatial domain. Through the synergistic complementarity of the two domains, edge blurring or noise residue caused by single-domain denoising is avoided.

[0030] In S6, the process of outputting the attention-enhanced feature map includes: In the channel attention branch, the global average value is calculated for each channel of the dual-domain denoising feature map to obtain the channel feature vector; the channel feature vector is compressed and activated through a fully connected layer to generate channel weights (the larger the weight, the more significant the target feature of the channel). The channel weights are multiplied with the dual-domain denoising feature map channel by channel to obtain the feature map after channel attention processing. In the spatial attention branch, max pooling and average pooling are performed on the dual-domain denoised feature map in the channel dimension to obtain two spatial feature maps. The two spatial feature maps are concatenated and then compressed and activated by 1×1 convolution to generate spatial weights (the larger the weight, the more significant the target feature at that position). The spatial weights are multiplied pixel by pixel with the feature map after channel attention processing to obtain the feature map after spatial attention processing. Residual connection: The dual-domain denoised feature map is added to the feature map after spatial attention processing to avoid over-suppressing useful features and output an attention-enhanced feature map.

[0031] After attention enhancement, target features are significantly strengthened, while the weight of background noise is reduced. By combining channel attention and spatial attention, target features (such as strong scattering points of ships and aircraft) are strengthened, the weight of background noise regions is weakened, and the distinction between targets and noise is improved.

[0032] In S7, the reconstructed and optimized feature map is obtained as follows: The Variational Autoencoder (VAE) adopts an encoder + decoder architecture. The encoder compresses the input attention-enhanced feature map through convolutional layers and maps it to the latent space. Based on the mean and variance of the reparameterized output latent variable (latent variable, ensuring the differentiability of the sampling process), the latent variable is sampled based on the mean and variance. The decoder reconstructs the latent variable into a reconstructed feature map with the same dimension as the input feature map through transposed convolutional layers (the denoising result is stable and the error meets the threshold requirement). Calculate the reconstruction error between the attention-enhanced feature map and the reconstructed feature map. If the error exceeds a preset error threshold (e.g., 0.01), the error signal is fed back to the multi-source mixed noise model. The parameters of the multi-source mixed noise model are adjusted (e.g., fine-tuning speckle scale and Gaussian standard deviation), and steps S2-S7 are repeated until the reconstruction error does not exceed the preset error threshold range.

[0033] The denoising results are reconstructed and verified using a variational autoencoder. If the reconstruction error is too large, the noise model parameters are corrected by feedback. The process is iteratively optimized until the results converge, thereby improving the robustness of denoising.

[0034] In S8, the process of outputting the final denoised radar image includes: Feature map to image conversion: The number of channels in the reconstructed and optimized feature map is compressed from 64 to 1 through 1×1 convolution to obtain a single-channel feature map. Then, the pixel values ​​are restored to the intensity range of the original radar image through linear mapping to obtain a preliminary denoised image. Structural similarity verification: Using a standard clean image sample (containing known targets for verification) as a reference, calculate the structural similarity index SSIM (measures the degree of structural similarity) between the preliminary denoised image and the standard clean image. If SSIM is greater than or equal to the set similarity threshold (e.g., 0.9), the structural integrity meets the standard, and edge preservation verification is performed; if it does not meet the standard, the attention parameters of the channel attention branch and the spatial attention branch are adjusted and re-optimized. Edge Preservation Validation: The Canny operator is used to extract edges from the initial denoised image and the standard clean image respectively to obtain edge maps. The Intersection over Union (IoU) of the two edge maps (measures the degree of edge matching) is calculated. If the IoU is greater than or equal to the set IoU threshold (e.g., 0.85), the edge preservation meets the standard, and the final denoised radar image is output. If it does not meet the standard, the spatial denoising parameters (ResNet parameters) in the frequency-spatial dual-domain denoising are adjusted and re-optimized.

[0035] The denoising results are designed to suppress noise while preserving the edge contours and strong scattering features of key targets, thus meeting the requirements of downstream detection and recognition tasks.

[0036] Standard clean image samples can be obtained through the following methods: High-resolution spaceborne / airborne SAR (such as radar equipment suitable for military reconnaissance and disaster monitoring) is used to acquire low-noise raw radar images in low-interference scenarios (such as flat areas without complex clutter or target areas known to be free of electromagnetic interference). Precise orbital parameters (azimuth, altitude, etc.) of the radar platform are utilized to ensure the basic accuracy of the image's geometry and intensity. The acquired low-noise images are further processed by multi-source noise suppression based on variational inference to remove residual noise and correct distortion. Combined with texture type annotations from a high-resolution radar image sample library, standard clean image samples with low noise, high geometric accuracy, and realistic texture features are obtained, meeting the supervision requirements for subsequent texture reconstruction.

[0037] Example 2: In S2, the joint probability density function of multi-source mixed noise is improved (it is no longer the product of single-source noise distributions, but the joint probability density of coupled multi-source mixed noise), by introducing two types of coupling coefficients: The first coupling coefficient a represents the linear coupling strength between speckle noise and Gaussian thermal noise. It is obtained by calculating the pixel value covariance of the two types of noise in the region without coherent interference and combining it with maximum likelihood estimation. The value range of a is [0, 0.5]. When a=0, it degenerates into an uncoupled state. The second coupling coefficient b represents the nonlinear coupling strength of coherent interference to the speckle noise-Gaussian thermal noise combination. It is obtained by statistically analyzing the rate of change of the Rayleigh distribution scale parameter of speckle noise and the variance of the normal distribution of Gaussian thermal noise when the coherent interference pulse occurs, and then fitting it using the least squares method. The value range of b is [0, 0.3]. ; In the formula, This represents the joint probability density of coupled multi-source mixed noise; , , These represent the noise intensities of speckle noise, Gaussian thermal noise, and coherent interference, respectively. , , Let represent the probability density functions of the Rayleigh distribution, the normal distribution, and the Poisson-Gaussian mixture distribution, respectively. , , , respectively, represent the Rayleigh distribution scale parameter of speckle noise, the standard deviation of the normal distribution of Gaussian thermal noise, and the characteristic parameters of the Poisson-Gaussian mixture distribution of coherent interference; This is the normalized coupling term for speckle-Gaussian noise; This is the pulse coupling attenuation term for coherent interference; In the variational inference noise estimation process in step S3, the coupling coefficients a and b are included in the variational parameter optimization range and are updated synchronously with the component weights, mean, and variance of the three-component Gaussian mixture distribution. During each iteration, a and b are corrected based on the statistical characteristics of the current pixel-level noise intensity map until the KL divergence converges and the fluctuation of the coupling coefficient is less than a preset threshold (e.g., 0.001).

[0038] The improvements in this embodiment can increase the noise fitting accuracy of the multi-source hybrid noise model for complex scenarios (such as maritime surveillance with high electromagnetic interference and land reconnaissance with strong clutter) by 15%-25%, thereby improving the accuracy of subsequent pixel-level noise estimation and reducing the problem of excessive smoothing of target edges caused by modeling bias.

[0039] Example 3: As Figure 2 As shown, the radar image multi-source noise suppression system based on variational inference, used to implement the method in Embodiment 1 or Embodiment 2, includes: The preprocessing module preprocesses the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images. The multi-source mixing module constructs probability distribution models for various types of noise based on the noise characteristics of the preprocessed image, and integrates them into a unified multi-source mixed noise model. The noise estimation module performs variational inference-based noise estimation based on the preprocessed image and a multi-source mixed noise model, generating a pixel-level noise intensity map. The multi-scale enhancement module, based on the preprocessed image and pixel-level noise intensity map, uses parallel convolution branches to perform multi-scale convolution feature extraction to obtain a multi-scale enhanced feature map. The dual-domain denoising module performs frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map. The attention enhancement module enhances the attention of the dual-domain denoised feature map through channel attention branch and spatial attention branch, and outputs the attention-enhanced feature map; The reconstruction and optimization module uses a variational autoencoder to reconstruct, verify, and optimize the attention-enhanced feature map, resulting in a reconstructed and optimized feature map. The feature preservation module acquires standard clean image samples, performs feature map to image conversion, structural similarity verification, and edge preservation verification on the reconstructed and optimized feature maps, and outputs the final denoised radar image.

[0040] Example 4: A radar image multi-source noise suppression device based on variational inference, comprising: One or more processors; Memory, used to store one or more computer programs; When one or more programs are executed by one or more processors, the one or more processors perform the method in Embodiment 1 or Embodiment 2.

[0041] Example 5: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method in Example 1 or Example 2.

Claims

1. A method for suppressing multi-source noise in radar images based on variational inference, characterized in that... Includes the following steps: S1. Preprocess the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images; S2. Based on the noise characteristics of the preprocessed image, probability distribution models are constructed for each type of noise and integrated into a unified multi-source mixed noise model. The various types of noise are speckle noise, Gaussian thermal noise and coherent interference. S3. Based on the preprocessed image and the multi-source mixed noise model, perform noise estimation based on variational inference to generate a pixel-level noise intensity map, specifically: A three-component Gaussian mixture distribution is introduced as an approximation of the true noise distribution; Based on the multi-source mixed noise model, Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution are respectively set as the initial prior distributions for speckle noise, Gaussian thermal noise, and coherent interference. With the goal of minimizing KL divergence and maximizing the lower bound of evidence, the Adam optimizer is used to iteratively update the variational parameters of the approximate distribution using the lower bound of evidence as the objective function. Once the optimization process converges, the noise intensity of each pixel in the preprocessed image is calculated using the optimized approximate distribution, generating a pixel-level noise intensity map that can label the noise strength pixel by pixel. S4. Based on the preprocessed image and pixel-level noise intensity map, multi-scale convolutional feature extraction is performed using parallel convolution branches to obtain multi-scale enhanced feature maps. The process is as follows: Construct a multi-branch parallel convolutional structure containing three parallel convolutional branches, each corresponding to a different scale: Small-scale branches, with 3×3 convolution kernels, capture local texture details of the image; The medium-scale branch, with a 5×5 convolution kernel, captures image edge features; Large-scale branches, with 7×7 convolution kernels, capture the global structure of the image; The preprocessed image is input into three parallel convolutional branches for feature extraction. The output feature maps of the three branches are concatenated, and channel attention is used to generate channel weights. The output features of different branches are then weighted and fused to obtain a fused feature map. After normalizing the pixel-level noise intensity map, it is multiplied pixel by pixel with the fused feature map to obtain the multi-scale enhanced feature map; S5. Perform frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map. The process includes: Frequency domain denoising: Using Fourier transform, a two-dimensional fast Fourier transform is performed on each channel of the multi-scale enhanced feature map to transform the image from the spatial domain to the frequency domain. Noise frequency suppression: Calculate the frequency domain amplitude spectrum, set a threshold to distinguish between signal frequency and noise frequency, weaken high-frequency noise through a Gaussian low-pass filter, and apply amplitude attenuation to noise frequency points to preserve signal frequency; By using inverse Fourier transform, the frequency domain features after noise frequency suppression are converted back to the spatial domain to obtain the frequency domain denoising features. Spatial denoising: A residual convolutional network containing five residual blocks is used to recover the residual texture details and target structure after frequency domain denoising, thus obtaining spatial domain denoising features. Dual-domain feature fusion: After compressing the number of channels by 1×1 convolution of the frequency domain denoising features, they are added to the spatial domain denoising features pixel by pixel to obtain a dual-domain denoising feature map. S6. Enhance the attention of the dual-domain denoised feature map by using channel attention branch and spatial attention branch, and output the attention-enhanced feature map; S7. The attention-enhanced feature map is reconstructed, verified, and optimized using a variational autoencoder to obtain the reconstructed and optimized feature map. S8. Obtain standard clean image samples, perform feature map to image conversion, structural similarity verification, and edge preservation verification on the reconstructed and optimized feature maps, and output the final denoised radar image. The process includes: Feature map to image conversion: The number of channels in the reconstructed and optimized feature map is compressed from 64 to 1 through 1×1 convolution to obtain a single-channel feature map. Then, the pixel values ​​are restored to the intensity range of the original radar image through linear mapping to obtain a preliminary denoised image. Structural similarity verification: Using standard clean image samples as a reference, calculate the structural similarity index SSIM between the preliminary denoised image and the standard clean image. If SSIM is greater than or equal to the set similarity threshold, the structural integrity meets the standard, and edge preservation verification is performed; if it does not meet the standard, the attention parameters of the channel attention branch and the spatial attention branch are adjusted and optimized again. Edge preservation verification: The Canny operator is used to extract edges from the preliminary denoised image and the standard clean image respectively to obtain edge maps. The intersection-union ratio (IoU) of the two edge maps is calculated. If the IoU is greater than or equal to the set IoU threshold, the edge preservation meets the standard, and the final denoised radar image is output. If it does not meet the standard, the spatial denoising parameters in the frequency-spatial dual-domain denoising are adjusted and re-optimized.

2. The radar image multi-source noise suppression method based on variational inference according to claim 1, characterized in that, In S1, the original radar image is a grayscale image, and the pixel value represents the radar echo intensity. The preprocessing specifically includes: Amplitude normalization: Maps the pixel values ​​of the original radar image to the [0, 1] interval to eliminate the difference in intensity magnitude caused by target distance and scattering intensity in different regions; Phase compensation: In a normalized image, identify noise-free, uniform regions and calculate the average phase value of these regions as the reference phase. For each pixel in the normalized image, subtract the offset of the global average phase relative to the reference phase from the phase value of each pixel. This is expressed as: ； In the formula, After phase compensation, the radar image is in coordinates The phase value of the pixel at that location; After amplitude normalization, the radar image in coordinates The original phase of the pixel; This represents the average phase of the entire radar image after amplitude normalization. The average reference phase of the noise-free uniform region; Logarithmic compression: Applying a logarithmic transformation to the phase-compensated image reduces the dynamic range and prevents weak target features from being masked by strong targets. The formula is: ； In the formula, After logarithmic compression, the radar image is in coordinates The grayscale value of the pixel at that location; , To adjust the parameters; After phase compensation processing, the radar image is in coordinates The grayscale value of the pixel at that location; After logarithmic compression, a preprocessed image with uniform intensity, no obvious phase distortion, and recognizable weak target features is obtained.

3. The radar image multi-source noise suppression method based on variational inference according to claim 1, characterized in that, In S2, the probability distribution model is constructed as follows: for speckle noise, Rayleigh distribution is used to describe its probability distribution; for Gaussian thermal noise, normal distribution is used to describe its probability distribution; for coherent interference, Poisson-Gaussian mixture distribution is used to describe its probability distribution. By integrating Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution, a joint probability multi-source mixed noise model containing all noise features is formed. The joint probability density function of the multi-source mixed noise is the product of the single-source noise distributions.

4. The radar image multi-source noise suppression method based on variational inference according to claim 1, characterized in that, In S6, the process of outputting the attention-enhanced feature map includes: In the channel attention branch, the global average value is calculated for each channel of the dual-domain denoising feature map to obtain the channel feature vector; the channel feature vector is compressed and activated through a fully connected layer to generate channel weights; the channel weights are multiplied with the dual-domain denoising feature map channel by channel to obtain the feature map after channel attention processing. In the spatial attention branch, max pooling and average pooling are performed on the dual-domain denoised feature map in the channel dimension to obtain two spatial feature maps. The two spatial feature maps are concatenated and then compressed and activated by 1×1 convolution to generate spatial weights. The spatial weights are multiplied pixel by pixel with the feature map after channel attention processing to obtain the feature map after spatial attention processing. Residual connection: Add the dual-domain denoised feature map to the feature map after spatial attention processing to output the attention-enhanced feature map.

5. The radar image multi-source noise suppression method based on variational inference according to claim 1, characterized in that, In S7, the reconstructed and optimized feature map is obtained as follows: The variational autoencoder adopts an encoder + decoder architecture. The encoder compresses the input attention-enhanced feature map through a convolutional layer, maps it to the latent space, and outputs the mean and variance of the latent variable. The latent variable is obtained by sampling based on the mean and variance. The decoder reconstructs the latent variable into a reconstructed feature map with the same dimension as the input feature map through a transposed convolutional layer. Calculate the reconstruction error between the attention-enhanced feature map and the reconstructed feature map. If the error exceeds the preset error threshold, feed the error signal back to the multi-source mixed noise model, adjust the parameters of the multi-source mixed noise model, and repeat steps S2-S7 until the reconstruction error does not exceed the preset error threshold range.

6. A radar image multi-source noise suppression system based on variational inference, used to implement the method according to any one of claims 1-5, characterized in that, include: The preprocessing module preprocesses the acquired raw radar images to eliminate the influence of acquisition errors and obtain the preprocessed images. The multi-source mixing module constructs probability distribution models for various types of noise based on the noise characteristics of the preprocessed image, and integrates them into a unified multi-source mixed noise model. The various types of noise are speckle noise, Gaussian thermal noise, and coherent interference. The noise estimation module, based on the preprocessed image and a multi-source mixed noise model, performs variational inference-based noise estimation to generate a pixel-level noise intensity map. Specifically: A three-component Gaussian mixture distribution is introduced as an approximation of the true noise distribution; Based on the multi-source mixed noise model, Rayleigh distribution, normal distribution, and Poisson-Gaussian mixture distribution are respectively set as the initial prior distributions for speckle noise, Gaussian thermal noise, and coherent interference. With the goal of minimizing KL divergence and maximizing the lower bound of evidence, the Adam optimizer is used to iteratively update the variational parameters of the approximate distribution using the lower bound of evidence as the objective function. Once the optimization process converges, the noise intensity of each pixel in the preprocessed image is calculated using the optimized approximate distribution, generating a pixel-level noise intensity map that can label the noise strength pixel by pixel. The multi-scale enhancement module, based on the preprocessed image and pixel-level noise intensity map, uses parallel convolution branches to perform multi-scale convolutional feature extraction to obtain a multi-scale enhanced feature map. The process is as follows: Construct a multi-branch parallel convolutional structure containing three parallel convolutional branches, each corresponding to a different scale: Small-scale branches, with 3×3 convolution kernels, capture local texture details of the image; The medium-scale branch, with a 5×5 convolution kernel, captures image edge features; Large-scale branches, with 7×7 convolution kernels, capture the global structure of the image; The preprocessed image is input into three parallel convolutional branches for feature extraction. The output feature maps of the three branches are concatenated, and channel attention is used to generate channel weights. The output features of different branches are then weighted and fused to obtain a fused feature map. After normalizing the pixel-level noise intensity map, it is multiplied pixel by pixel with the fused feature map to obtain the multi-scale enhanced feature map; The dual-domain denoising module performs frequency-spatial dual-domain denoising on the multi-scale enhanced feature map to obtain a dual-domain denoised feature map. The process includes: Frequency domain denoising: Using Fourier transform, a two-dimensional fast Fourier transform is performed on each channel of the multi-scale enhanced feature map to transform the image from the spatial domain to the frequency domain. Noise frequency suppression: Calculate the frequency domain amplitude spectrum, set a threshold to distinguish between signal frequency and noise frequency, weaken high-frequency noise through a Gaussian low-pass filter, and apply amplitude attenuation to noise frequency points to preserve signal frequency; By using inverse Fourier transform, the frequency domain features after noise frequency suppression are converted back to the spatial domain to obtain the frequency domain denoising features. Spatial denoising: A residual convolutional network containing five residual blocks is used to recover the residual texture details and target structure after frequency domain denoising, thus obtaining spatial domain denoising features. Dual-domain feature fusion: After compressing the number of channels by 1×1 convolution of the frequency domain denoising features, they are added to the spatial domain denoising features pixel by pixel to obtain a dual-domain denoising feature map. The attention enhancement module enhances the attention of the dual-domain denoised feature map through channel attention branch and spatial attention branch, and outputs the attention-enhanced feature map; The reconstruction and optimization module uses a variational autoencoder to reconstruct, verify, and optimize the attention-enhanced feature map, resulting in a reconstructed and optimized feature map. The feature preservation module acquires standard clean image samples, performs feature map-to-image conversion, structural similarity verification, and edge preservation verification on the reconstructed and optimized feature maps, and outputs the final denoised radar image. The process includes: Feature map to image conversion: The number of channels in the reconstructed and optimized feature map is compressed from 64 to 1 through 1×1 convolution to obtain a single-channel feature map. Then, the pixel values ​​are restored to the intensity range of the original radar image through linear mapping to obtain a preliminary denoised image. Structural similarity verification: Using standard clean image samples as a reference, calculate the structural similarity index SSIM between the preliminary denoised image and the standard clean image. If SSIM is greater than or equal to the set similarity threshold, the structural integrity meets the standard, and edge preservation verification is performed; if it does not meet the standard, the attention parameters of the channel attention branch and the spatial attention branch are adjusted and optimized again. Edge preservation verification: The Canny operator is used to extract edges from the preliminary denoised image and the standard clean image respectively to obtain edge maps. The intersection-union ratio (IoU) of the two edge maps is calculated. If the IoU is greater than or equal to the set IoU threshold, the edge preservation meets the standard, and the final denoised radar image is output. If it does not meet the standard, the spatial denoising parameters in the frequency-spatial dual-domain denoising are adjusted and re-optimized.

Images