A SAR image denoising method and device based on a logarithmic domain diffusion model

Abstract

The present application belongs to the technical field of remote sensing image processing, and provides a SAR image denoising method and device based on a logarithmic domain diffusion model. The denoising method comprises: preprocessing: converting an original SAR amplitude image to a logarithmic domain to obtain a starting input image for a diffusion process; forward diffusion: constructing a forward diffusion path based on a non-central Gaussian distribution, simulating a noise injection process, and generating a noisy image sequence; neural network model noise prediction: performing feature extraction and fusion on the noisy image and physical metadata through a pre-constructed neural network model, and outputting a noise residual prediction result; backward sampling reconstruction: iteratively sampling using the noise residual prediction result to reconstruct a denoised SAR image. The denoising method solves the distribution mismatch problem of traditional diffusion models for SAR multiplicative noise, improves the denoising precision, and preserves the image texture and structure information.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a SAR image denoising method and apparatus based on a logarithmic domain diffusion model. Background Technology

[0002] SAR images, due to their all-weather, all-time imaging advantages unaffected by lighting and weather conditions, are widely used in agricultural monitoring, disaster assessment, and military reconnaissance. However, due to the coherent interferometry characteristics in the imaging mechanism, SAR images commonly exhibit speckle noise, which is a type of multiplicative noise. This leads to a decrease in image quality and affects subsequent applications such as classification and change detection.

[0003] Existing denoising methods can be divided into traditional filters (such as Lee and Kuan filters), nonlocal mean (NLM) methods, variational models, and deep learning-based methods (such as GAN and U-Net). Among them, deep learning performs well in natural images, but it still faces the following problems in SAR images: (1) The speckle noise in SAR images is multiplicative gamma noise, which usually follows a Fisher-Tippett distribution in the logarithmic domain. Existing models based on the additive noise assumption are difficult to directly adapt to the multiplicative noise characteristics of SAR images; (2) Traditional diffusion probability models (such as DDPM) are built under the standard Gaussian noise assumption, and direct transfer has a distribution mismatch problem; (3) Physical metadata such as polarization and incident angle have a significant impact on the formation of SAR images and speckle noise characteristics. Traditional denoising models lack effective modeling of these metadata, resulting in unsatisfactory denoising effects and difficulty in fully suppressing speckle noise. (4) SAR images are sensitive to details, and deep networks need to remove noise while maintaining texture, which is a high modeling difficulty. Summary of the Invention

[0004] To address the multiplicative nature of speckle noise in SAR images, this invention provides a SAR image denoising method and apparatus based on a logarithmic domain diffusion model. This method solves the problems of distribution mismatch and high modeling difficulty in existing deep learning models when processing speckle noise in SAR images, achieving a balance between high-precision denoising and detail preservation.

[0005] In a first aspect, embodiments of the present invention provide a SAR image denoising method based on a logarithmic domain diffusion model, comprising the following steps:

[0006] S1: Preprocessing: Convert the original SAR amplitude image to the logarithmic domain to obtain the initial input image for the diffusion process;

[0007] S2: Forward diffusion: A forward diffusion path is constructed based on a non-central Gaussian distribution to simulate the noise injection process and generate a noisy image sequence;

[0008] S3: Neural Network Model Prediction of Noise: Through a pre-built neural network model, features are extracted and fused from noisy images and physical metadata, and the noise residual prediction results are output.

[0009] S4: Backsampling Reconstruction: Iterative sampling is performed using the noise residual prediction results to reconstruct the denoised SAR image.

[0010] The corresponding effects are as follows: Through the synergistic effect of logarithmic domain transformation, non-central Gaussian diffusion modeling, metadata fusion and iterative sampling, the distribution mismatch problem of SAR multiplicative noise in traditional diffusion models is solved, and the image texture and structural information are preserved while improving the denoising accuracy.

[0011] According to a specific implementation of an embodiment of the present invention, the formula for the logarithmic field transformation in step S1 is:

[0012] x0 = log(I);

[0013] Where I is the original SAR amplitude image, and x0 is the image input for subsequent forward diffusion.

[0014] The corresponding effects are as follows: By using logarithmic transformation, the multiplicative Gamma noise of the SAR image is converted into additive noise that approximately follows the Fisher-Tippett distribution in the logarithmic domain, making the noise distribution easier to model. This solves the problem that traditional models cannot adapt to SAR multiplicative noise and provides a stable noise distribution basis for the subsequent diffusion process.

[0015] According to a specific implementation of the present invention, in step S2, the mean μ of the non-central Gaussian distribution increases with the diffusion time step t, and the variance σ 2 The variance is calculated using the Fisher-Tippett distribution, and the specific formula is as follows:

[0016]

[0017] Where k is the shape parameter of the original multiplicative gamma distribution spot, λ is the amplitude coefficient of the offset, γ is the set value, t is the current time step in the diffusion process, and T is the maximum time step in the diffusion process.

[0018] The corresponding effects are as follows: By dynamically adjusting the mean and variance of the non-central Gaussian distribution, it gradually approximates the true Fisher-Tippett distribution characteristics (right-skewed distribution) of SAR noise in the logarithmic domain, reducing noise distribution mismatch during the diffusion process and improving the model's simulation accuracy of SAR speckle noise.

[0019] According to a specific implementation of an embodiment of the present invention, in step S3, the pre-built neural network model adopts an improved U-Net++ network, which includes a symmetric encoder and decoder, and the encoder and decoder adopt a weighted skip connection driven by diffusion time step.

[0020] The corresponding effect is as follows: Based on the time position of the current diffusion step t, a noise weighting function γ is constructed. t By weighting the skip connection features, the diffusion neural network can identify high-noise and low-noise regions, thus prioritizing high-noise regions and weakening low-noise regions to avoid mistakenly eliminating texture details, thereby improving the network's ability to respond to noise intensities in different regions.

[0021] According to a specific implementation of an embodiment of the present invention, step S3 further includes a physical metadata injection step: the polarization mode, incident angle and frequency band information of the SAR image are encoded into a chimera vector through a multilayer perceptron (MLP), and then spliced ​​and fused with the feature map extracted by the encoder.

[0022] The corresponding effects are as follows: By encoding physical metadata (polarization, incident angle, frequency band) through MLP and injecting it into the network, the model can perceive the differences in noise characteristics under different imaging conditions (such as VV polarization being noisier than VH polarization), thereby improving the network's physical adaptability to complex imaging scenes and maintaining stable denoising performance under different incident angles (29° to 46°) and polarization modes.

[0023] According to a specific implementation of an embodiment of the present invention, the polarization method includes dual polarization VV and VH, the incident angle range is 29° to 46°, and the frequency band is C-band.

[0024] The corresponding effects are as follows: Special optimization is performed on the dual polarization (VV / VH) data and typical incident angle range (29°~46°) of C-band SAR images (the frequency band commonly used by satellites such as Sentinel-1), so that the matching degree between metadata and noise characteristics is higher.

[0025] According to a specific implementation of the present invention, the training process of the neural network model in step S3 adopts a joint loss function, wherein the joint loss function is the mean squared error (MSE) term L. MSE With the Kullback-Leibler (KL) divergence term L KL The weighted sum is calculated using the following formula:

[0026] L total =L MSE +λ KL L KL ;

[0027] Where, λ KLThis is the weighting coefficient, with a value ranging from 0.1 to 1.0.

[0028] The corresponding effects are as follows: the MSE term constrains the pixel-level residual regression accuracy, ensuring that the pixel-level error between the denoised image and the real image is minimized; the KL divergence term constrains the consistency between the predicted noise distribution and the Fisher-Tippett distribution, making the model output closer to the true statistical characteristics of SAR noise. The combination of the two reduces the denoising error rate of the model in low signal-to-noise ratio areas (such as dense urban areas) while maintaining the physical authenticity of the statistical distribution.

[0029] According to a specific implementation of an embodiment of the present invention, the formula for the backsampling reconstruction in step S4 is:

[0030]

[0031] Where, x t For the noisy image at step t, α t The signal retention rate coefficient. Let t be the noise residual predicted by the neural network at step t.

[0032] The corresponding effect is as follows: By combining noise residual iterative correction with random sampling in the reconstruction method, the excessive smoothing of the image caused by deterministic sampling is avoided, and the naturalness and realism of the image are improved while preserving texture details.

[0033] Secondly, embodiments of the present invention provide a SAR image denoising device based on a logarithmic domain diffusion model, comprising:

[0034] The preprocessing module is used to convert the original SAR amplitude image to the logarithmic domain to obtain the initial input image for the diffusion process;

[0035] The forward diffusion module is used to establish a forward diffusion path based on a non-central Gaussian distribution, simulate noise injection, and generate a noisy image;

[0036] The neural network module is used to extract and fuse features from noisy images and physical metadata, and output noise residual prediction results.

[0037] The backsampling module is used to perform iterative sampling using the noise residual prediction results to reconstruct the denoised SAR image. Attached Figure Description

[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0039] Figure 1 The flowchart illustrates the steps of a SAR image denoising method based on a logarithmic domain diffusion model provided in an embodiment of the present invention.

[0040] Figure 2 This diagram illustrates a structural block diagram of a SAR image denoising device based on a logarithmic domain diffusion model provided in an embodiment of the present invention.

[0041] Figure 3 A partial schematic diagram of the S2FS remote sensing image training set in an embodiment of the present invention is shown;

[0042] Figure 4 A schematic diagram showing the effect comparison of SAR image denoising methods in embodiments of the present invention is illustrated. Detailed Implementation

[0043] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0044] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0045] Figure 1 A flowchart illustrating the steps of a SAR image denoising method based on a logarithmic domain diffusion model provided in this embodiment of the invention is shown below. Figure 1 The method includes the following steps:

[0046] S1: Preprocessing: Convert the original SAR amplitude image to the logarithmic domain to obtain the initial input image for the diffusion process.

[0047] The original amplitude image is transformed to the logarithmic domain using a logarithmic transformation, as follows:

[0048] x0 = log(I);

[0049] Let I be the original SAR amplitude image, and x0 be the image input for subsequent forward diffusion. Essentially, x0 is injected with noise through a Markov chain over multiple steps, gradually degenerating it into sampleable noise x. T .

[0050] S2: Forward diffusion: A forward diffusion path is established based on a non-central Gaussian distribution to simulate noise injection and generate a noisy image.

[0051] The coherent noise after logarithmic transformation is no longer distributed according to a standard normal distribution, but rather according to a Fisher-Tippett distribution. A non-central Gaussian distribution is used to approximate the Fisher-Tippett distribution. The non-central Gaussian distribution is known as follows:

[0052]

[0053] in:

[0054] x t It is the noisy image after the t-th diffusion step, x t-1 For the image state at step t-1, α t This is a coefficient used to control the signal retention rate. Control signal retention rate, β s To determine the noise proportion for each step, following cosine modulation, we decide how much noise to add in each step, α. t Decide how much of the original content to retain. ε t The random noise variable injected in step t has its distribution controlled by expectation and variance.

[0055] Let T be the maximum time step of diffusion (total number of diffusion steps), and t in the master equation represent the current time step, taking the range t = 0, 1, 2, ..., T. s is the summation index used to solve for the cumulative sum from step 1 to step t, ranging from 1 to t. To generate the noise ratio β for each time step... s Cosine tuning is employed.

[0056]

[0057]

[0058] β s =1-α s (4);

[0059] Where u is greater than or equal to 0, it is an offset constant, commonly u = 0.008, which controls the noise addition at both ends to be smoother and the middle section to be faster. This yields the cumulative retention rate at step t.

[0060] To make the predicted noise as close as possible to the speckle noise, μ is required. t (mean) and (Variance) approximates the Fisher-Tippett distribution. Simulating the right-skewed characteristic of the Fisher-Tippett distribution, a mean shift term that increases with the time step t is designed:

[0061]

[0062] The variance is calculated directly from the variance of the Fisher-Tippett distribution:

[0063]

[0064] Where k is the shape parameter of the original multiplicative gamma distribution spot. λ is the magnitude coefficient of the offset, γ is a set value, here γ = 0.5772, t is the current time step in the diffusion process, T is the maximum time step in the diffusion process, and β t The noise proportion at step t can also be understood as the noise intensity. (To simulate the right-skewed characteristic of the Fisher-Tippett distribution, the expectation of the standard Gaussian distribution (expectation of 0) is shifted to the right by a value greater than 0).

[0065] From the initial non-central Gaussian distribution x t and x t-1 Relationship between them:

[0066]

[0067] ε t The random noise variable injected in step t follows a non-central Gaussian distribution, the distribution of which is controlled by the expectation and variance. To satisfy a non-central Gaussian distribution.

[0068] It can be known that x t The relationship between x and x0 is that each x... t-1 The superposition of; therefore:

[0069]

[0070] α t：s Let be the cumulative retention rate from s to t, representing the proportion of noise injected in step s that is retained by step t. β k The noise intensity at step k can also be understood as the proportion of noise injected at step k. Meanwhile, the expectation and variance satisfy:

[0071]

[0072]

[0073] In the formula, E[x t [x] is the image at step t during the diffusion process. t Expectation; Var(x) t ) represents the variance of the image at step t of the diffusion process.

[0074] To make the non-central Gaussian distribution approximate the Fisher-Tippett distribution, a constraint is imposed on E[x]. T ]≈μ total and Finally, we get x tThe probability distribution of μ. total Let μ be the theoretical mean of the target Fisher-Tippett distribution. total =m+γλ, γ=0.5772, where m is the position parameter, λ is the amplitude coefficient of the offset, and γ is the set value.

[0075] S3: Neural Network Model Prediction of Noise: Through a pre-built neural network model, features are extracted and fused from noisy images and physical metadata, and the noise residual prediction results are output.

[0076] The denoising neural network employs an improved U-Net++ architecture, consisting of a symmetrical encoder and decoder, and incorporating intermediate skip connections to enhance multi-scale feature fusion capabilities. The encoder comprises multiple convolutional layers (ConvBlock), each extracting spatial semantic features at different scales. The decoder progressively restores the image size through upsampling operations and fuses the high-resolution features from the skip connections. The network's final output is a noise residual prediction, used to guide image reconstruction during the backdiffusion process.

[0077] 3.1: Since the imaging characteristics of SAR images are affected by various physical parameters (such as incident angle and polarization), this method introduces a metadata guidance mechanism into the network to enhance the physical consistency of the model. Given a noisy image x t Given the current time step t and SAR metadata, predict the noise residual ε in the current image. θ This allows us to gradually obtain the original, denoised image x0.

[0078] x t Input the multi-layer encoder and perform feature extraction:

[0079]

[0080] The output features of the encoder's layer l are represented by ConvBlock, which is a convolutional block operation and the basic unit for feature extraction in the encoder. It includes operations such as convolution and activation functions, and is used to perform convolution calculations and nonlinear transformations on the input features.

[0081] The structured physical metadata (such as incident angle and polarization mode encoding) is input into a multilayer perceptron (MLP) to obtain the following splicing vector:

[0082] f meta =MLP(metadata) (11);

[0083] f metaThis is a chimera vector obtained by encoding metadata using a multilayer perceptron (MLP), used to integrate structured metadata information into network features. Metadata refers to the structured physical metadata of SAR images, including physical parameter information such as incident angle (e.g., 29°-46°), polarization mode (e.g., dual polarization VV and VH), and frequency band (e.g., C-band). This information is closely related to the formation and noise characteristics of SAR images.

[0084] The feature extraction data is combined with the metadata of the SAR image for information fusion, and then the combined data is injected into the encoder.

[0085]

[0086] The concatenation operation joins two or more feature vector feature maps along a specified dimension, fusing information from different sources. Broadcast(f meat The first step is to encode the metadata using an MLP to obtain the vector f. meta Then f meta Spread to the feature map of the encoder's l-th layer Same space size (H) l W l Finally, splicing is performed on the channel dimension to obtain the fusion features, which are then spliced ​​and fused.

[0087] 3.2: To improve the network's response to noise intensities in different regions, a time-step-driven skip connection weighting mechanism is proposed. Traditional skip connections directly pass the encoder feature map to the decoder, which can easily cause high-noise regions to be "prematurely exposed," leading to misleading decoding. To address this issue, this method constructs a noise weight function γ based on the time position of the current diffusion step t. t By weighting the skip connection features, the diffusion neural network can identify high-noise and low-noise regions. Thus, high-noise regions are given priority emphasis, while low-noise regions are weakened to avoid mistakenly eliminating texture details.

[0088]

[0089] in,

[0090] 3.3: Each step t in the diffusion process represents a different noise intensity stage. To enable the network to "perceive the current diffusion stage," this method introduces a sinusoidal positional encoding (PE) mechanism. Specifically, time step t is mapped to a fixed-dimensional sine and cosine embedding vector:

[0091] PE(t)=[sin(w1t),cos(w1t),...,sin(w kt),cos(w k t)] (14);

[0092] The vector is then concatenated to the input of the intermediate layer of the decoder, working together with the skip connection features in the reverse reconstruction process. This design enables the network to adaptively adjust its "aggressive or conservative" denoising strategy, allowing early steps to focus on coarse modeling and later steps to enhance detail restoration.

[0093] The final result is:

[0094]

[0095] The neural network model constructed above has the following characteristics: 1. The encoder part adopts a multi-layer convolutional structure to extract multi-scale features from the input image at time t; 2. The metadata information is encoded by MLP and broadcast to each layer of feature map; 3. Skip connections and weight adjustment modules are used to weightedly fuse features at different time steps; 4. The decoder finally outputs the prediction noise residual.

[0096] S4: Backsampling Reconstruction: Iterative sampling is performed using the noise residual prediction results to reconstruct the denoised SAR image.

[0097] Backdiffusion is the inverse of forward diffusion, aiming to progressively recover a clean image from a Gaussian noise vector. In backdiffusion, the noise residual value predicted by the network is used as input, and the next image x is iteratively reconstructed from the current diffusion state using a fixed sampling formula. t-1 Continue until x0 is restored. The specific formula is as follows:

[0098]

[0099] x t For the noisy image at step t, α t The signal retention rate coefficient. The noise residual at step t, predicted by the neural network, is used for random sampling during the backdiffusion process, increasing the randomness and generalization ability of the model.

[0100] The overall algorithm for the SAR image denoising method described above is shown in the table below:

[0101]

[0102] Furthermore, the neural network model in S3 mentioned above is trained using the MSE+KL joint loss function to optimize the network parameters.

[0103] The designed loss function consists of two parts: a mean squared error (MSE) term for pixel-by-pixel denoising accuracy and a Kullback-Leibler (KL) divergence term for statistical distribution alignment.

[0104] Total loss is defined as:

[0105] L total =L MSE +λ KL L KL (17);

[0106] Where λ KL These are weighting coefficients determined through experimental verification, with values ​​ranging from 0 to 1. Here, λ... KL =0.1.

[0107] L MSE This item is used to constrain the difference between the noise residuals predicted by the neural network and the actual noise added during the forward diffusion process, ensuring that the network can accurately learn the residual regression task.

[0108] For N training samples in a batch, the expanded form is:

[0109]

[0110] in Let t be the noise residual predicted by the network at step t. This is the true residual obtained from forward diffusion.

[0111] L KL The term is used to constrain the statistical distribution of the network output residual to be consistent with the target Fisher-Tippett noise distribution in terms of mean and scale parameters, ensuring that the network not only fits single-sample noise, but also statistically reproduces the right-biased noise characteristics of the SAR logarithmic domain.

[0112] By minimizing the predicted noise distribution p θ With the target Fisher-Tippett distribution q FT The KL divergence makes the model output closer to the true statistical distribution of SAR noise, and can be approximated as:

[0113]

[0114] The final total loss function balances pixel-level denoising accuracy with statistical distribution consistency, effectively improving the model's restoration capability and physical adaptability in complex SAR noise backgrounds.

[0115] Embodiments of the present invention provide a SAR image denoising method based on a logarithmic domain diffusion model, which integrates noise statistical distribution modeling, deep neural networks, physical metadata, and diffusion sampling mechanisms, and has at least the following technical effects:

[0116] 1. Using the Fisher-Tippett distribution to approximate and describe the noise characteristics of SAR images in the logarithmic domain;

[0117] 2. Construct a forward diffusion process based on a non-central Gaussian distribution to simulate noise injection into SAR images;

[0118] 3. Construct a residual neural network to predict the noise term, and design a joint MSE+KL divergence loss to improve the distribution approximation ability;

[0119] 4. Introduce SAR metadata such as polarization mode, incident angle, and frequency band, and inject it into the network after encoding through a multilayer perceptron (MLP) module;

[0120] 5. A skip connection weighting mechanism is proposed to guide feature recovery at different diffusion time steps.

[0121] Figure 2 This is a structural block diagram of a SAR image denoising device based on a logarithmic domain diffusion model provided in an embodiment of the present invention. The device includes:

[0122] The preprocessing module is used to convert the original SAR amplitude image to the logarithmic domain to obtain the initial input image for the diffusion process;

[0123] The forward diffusion module is used to establish a forward diffusion path based on a non-central Gaussian distribution, simulate noise injection, and generate a noisy image;

[0124] The neural network module is used to extract and fuse features from noisy images and physical metadata, and output noise residual prediction results.

[0125] The backsampling module is used to perform iterative sampling using the noise residual prediction results to reconstruct the denoised SAR image.

[0126] Figure 2 The functions of each module in the embodiments correspond to the contents of their respective method embodiments, and will not be repeated here.

[0127] It should be noted that the arrangement of the modules in a flow layout is merely one embodiment of the present invention, and other arrangements may also be used, which are not limited in the present invention.

[0128] Verification of the effectiveness of the above SAR image denoising method:

[0129] Because the underlying principle of this SAR-DDPM model is to approximate a non-central Gaussian distribution.

[0130] The Fisher-Tippett distribution clearly indicates that simulated SAR images are insufficient to meet the model's training requirements for various parameters. Therefore, the training of the SAR-DDPM model requires a remote sensing image dataset containing clean reference images. This invention uses S2FS, the Sentinel-2 optical dataset, solely for pre-training / prior learning and simulation experiments, and not for evaluating real SAR results. Real SAR images of Sentinel-1GRD (IW mode, 10m, VV / VH) are used as validation data; the S2FS remote sensing image dataset consists of 2400 non-overlapping images, each with a noise-free control, and a size of [missing information].

[0131] These images, each 256×256 pixels with a ground resolution of 10 meters, were all taken by the Sentinel-2 satellite and cropped at 16 locations. They comprise four main categories: cities, farms, forests, and ports, with 600 images in each category. Training used 800 satellite images from the S2FS remote sensing image set, with the remaining images used as the test set. The training batch size was 128 images, with a batch size of 410, and the training duration was approximately 22 hours. Some of the images used in the training are shown below. Figure 3 As shown.

[0132] Randomly selected images from the test set were denoised using the aforementioned SAR image denoising method, with good results. (Next) Figure 4 This is a comparison image showing the denoising effect of SAR-DDPM on images. For example... Figure 4 As shown, the aforementioned SAR image denoising method demonstrates excellent performance in both noise suppression and image fidelity. This method significantly improves noise suppression in uniform regions (such as water bodies and plains), and while suppressing noise, it also significantly enhances the preservation of texture details, dynamically sensing differences in the scattering characteristics of different terrain features. Combined with... Figure 4 It can be observed that the above-mentioned SAR image denoising method can avoid introducing artifacts in edge and weak texture areas (such as vegetation cover areas) and preserve the original ground features well.

[0133] This method effectively addresses the distribution mismatch problem in existing deep learning models when processing speckle noise in SAR images. By constructing a non-central Gaussian diffusion process in the logarithmic domain and introducing a Fisher-Tippett distribution for noise modeling, it improves denoising accuracy while preserving image texture and structural information. Experimental results show that this method significantly improves upon existing models (such as SAR-DnCNN) in terms of equivalent number of views (ENL) and structural similarity (SSIM), making it suitable for SAR image enhancement tasks in complex scattering scenarios.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A SAR image denoising method based on a logarithmic domain diffusion model, characterized in that, Includes the following steps: S1： Preprocessing: The original SAR amplitude image is converted to the logarithmic domain to obtain the initial input image for the diffusion process; S2: Forward diffusion: A forward diffusion path is constructed based on a non-central Gaussian distribution to simulate the noise injection process and generate a noisy image sequence; S3: Neural network model for noise prediction: Through a pre-built neural network model, features are extracted and fused from the noisy image and physical metadata, and the noise residual prediction result is output. The physical metadata includes incident angle, polarization mode, and frequency band. S4: Backsampling reconstruction: Iterative sampling is performed using the noise residual prediction results to reconstruct the denoised SAR image.

2. The SAR image denoising method according to claim 1, characterized in that: In step S1, the formula for the logarithmic field transformation is: ； Where I is the original SAR amplitude image, The image is used as input for subsequent forward diffusion.

3. The SAR image denoising method according to claim 2, characterized in that: In step S2, the mean μ of the non-central Gaussian distribution increases with the diffusion time step t, and the variance σ 2 The variance is calculated using the Fisher-Tippett distribution, and the specific formula is as follows: ； ； ； in, γ is the amplitude coefficient of the expected offset of the Gaussian distribution, and γ is a set value. Here, represents the shape parameter of the original multiplicative gamma distribution spot, t represents the current time step in the diffusion process, and T represents the maximum time step in the diffusion process. This represents the noise ratio at each step. Let be the noise ratio at step t, and s be the summation index used to calculate the cumulative sum from step 1 to step t.

4. The SAR image denoising method according to claim 1, characterized in that: In step S3, the pre-built neural network model uses an improved U-Net++ network, which includes a symmetric encoder and decoder, and the encoder and decoder use a weighted skip connection driven by a diffusion time step.

5. The SAR image denoising method according to claim 4, characterized in that: Step S3 also includes a physical metadata injection step: the polarization mode, incident angle and frequency band information of the SAR image are encoded into a chimeric vector by a multilayer perceptron (MLP) and then spliced ​​and fused with the feature map extracted by the encoder.

6. The SAR image denoising method according to claim 5, characterized in that: The polarization mode includes dual polarization VV and VH, with an incident angle range of 29°~46° and a frequency band of C-band.

7. The SAR image denoising method according to claim 4, characterized in that: The training process of the neural network model in step S3 employs a joint loss function, which is the mean squared error term. With KL divergence term The weighted sum is calculated using the following formula: ； in, These are the weighting coefficients.

8. The SAR image denoising method according to claim 1, characterized in that: The formula for backsampling reconstruction in step S4 is: ； Where, x t For the noisy image at step t, The signal retention rate coefficient. Let t be the noise residual predicted by the neural network at step t.

9. A SAR image denoising device based on a logarithmic domain diffusion model, characterized in that, The SAR image denoising method according to any one of claims 1-8 includes: The preprocessing module is used to convert the original SAR amplitude image to the logarithmic domain to obtain the initial input image for the diffusion process; The forward diffusion module is used to establish a forward diffusion path based on a non-central Gaussian distribution, simulate noise injection, and generate a noisy image; The neural network module is used to extract and fuse features from noisy images and physical metadata, and output noise residual prediction results. The physical metadata includes incident angle, polarization mode, and frequency band. The backsampling module is used to perform iterative sampling using the noise residual prediction results to reconstruct the denoised SAR image.

Images