A sar image complex domain self-supervised denoising method combined with physical prior

Abstract

This invention discloses a self-supervised denoising method for complex domain SAR images based on physical priors, belonging to the field of synthetic aperture radar image denoising technology. It generates independent sub-view guide maps based on physical spectrum decomposition, and extracts data from the original full-resolution image after similarity localization on the guide maps. Subsequently, a strong point mask protection mechanism based on local statistical features is introduced to construct pseudo-labels, and a multi-level wavelet network is trained using a hybrid loss function. The inference process employs amplitude domain denoising combined with original phase reconstruction to output the final single-view complex denoised image. This invention achieves full-resolution denoising in the complex domain without requiring a clean reference image, effectively suppressing speckle noise, preserving strong scattering points and physical phase information, reducing memory overhead, and is suitable for ultra-wide-swath SAR image preprocessing, possessing high engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar image denoising technology, and particularly relates to a SAR image complex domain self-supervised denoising method that combines physical priors. Background Technology

[0002] Synthetic Aperture Radar (SAR) actively transmits and receives electromagnetic waves to create images, enabling it to penetrate cloud and fog interference and achieve all-weather, day-and-night Earth observation. Therefore, SAR is a crucial tool for remote sensing data acquisition and is widely used in tasks such as target detection, scene classification, region segmentation, and change monitoring, finding broad applications in national economic development, geological exploration, environmental monitoring, and military reconnaissance.

[0003] However, because SAR imaging is based on coherent processing mechanisms, its images are inevitably affected by speckle noise. Speckle noise is generated by the coherent superposition of backscattered signals during imaging and is usually modeled as multiplicative noise. Compared to additive noise commonly found in traditional digital image processing, it has completely different statistical distribution characteristics. Speckle noise severely degrades the visualization and interpretability of SAR images, limiting the performance of downstream tasks. Therefore, speckle suppression is an indispensable step in SAR image preprocessing.

[0004] In recent years, deep learning-based image denoising methods have become mainstream. However, traditional supervised learning denoising networks require a large number of noisy-clean image pairs for training. In real SAR imaging scenarios, obtaining an absolutely clean (speckle-free) reference image is physically impossible. To address this challenge, the self-supervised learning (SSL) framework has been introduced into the field of SAR denoising. Although existing self-supervised SAR denoising methods have achieved some success, they still face the following significant challenges: Most existing denoising methods operate on the amplitude or intensity images of SAR, discarding the crucial phase information contained in single-look complex data. This approach not only destroys the complex algebraic structure of SAR signals but also ignores their physical characteristics. Furthermore, real-world SAR systems are constrained by non-ideal impulse responses, oversampling, and frequency-domain windowing, resulting in highly spatially correlated speckle noise. If self-supervised training pairs are directly constructed from the original images for network training, the network will retain spatially correlated noise as true high-frequency signals, leading to severe artifacts in the denoising results.

[0005] Therefore, overcoming the spatial correlation of speckle in SAR images, achieving high-fidelity, full-resolution denoising without a clean reference image, and accurately preserving the features and physical phase of strong scattering points while suppressing background speckle are key technical challenges that urgently need to be addressed in the field of SAR image denoising. Summary of the Invention

[0006] The purpose of this invention is to provide a self-supervised denoising method for SAR images in the complex domain that combines physical priors, and to solve the problems in the prior art of how to achieve high-fidelity, full-resolution denoising of SAR images in the complex domain under the condition of no clean reference image and spatial correlation of speckle noise, while fully preserving its complex physical properties and high-frequency structural details.

[0007] To achieve the above objectives, this invention provides a self-supervised denoising method for complex domain SAR images that incorporates physical priors, comprising the following steps: S1. Generate a sub-view guide graph that is statistically independent of the original noise based on a physical sub-aperture decomposition strategy; S2. Construct a homogeneous pixel library containing multiple independent noise observations based on the sub-view guide map generated in S1; S3. Hybrid constraint training based on multi-level wavelet network MWCNN; S4. Perform full-image denoising inference and physical phase reconstruction.

[0008] Preferably, S1 specifically includes: S11. Obtain the original high-resolution single-view complex image sequence and perform a two-dimensional fast Fourier transform; S12. Generate a low-resolution subview wizard map.

[0009] Preferably, the specific content of S11 is as follows: Obtain the original high-resolution single-view complex image sequence, denoted as ; A two-dimensional fast Fourier transform is performed on the original high-resolution single-view complex image to convert it to the frequency domain spectrum. In the formula, For azimuth frequency; For range-direction frequency.

[0010] Preferably, the specific content of S12 is as follows: Based on the spectrum segmentation strategy with guard gaps, the low-energy oversampling bands at both ends of the spectrum are removed; Based on the selected splitting dimension Adaptively calculate and remove the low-energy oversampled frequency band Trim regions at both ends of the spectrum, with a length of [missing information]. And forcibly remove the center frequency band gap region with the most concentrated energy and the most severe sidelobe leakage, with a length of ; Extracting the effective frequency band of one side The expression is as follows: ; In the formula, The function is a rectangular window; the extraction interval of the rectangular window function is... ; Only for Perform inverse Fourier transform to generate low-resolution subview wizard maps .

[0011] Preferably, the specific content of S2 is as follows: Low-resolution subview wizard map and the original high-resolution single-view complex image sequence Divide the data into tensor processing blocks of a preset size and perform tensor expansion operations; For each processing unit coordinate in the image In statistically independent low-resolution sub-view wizard maps Expanding on the amplitude feature to obtain the size is The query block; and a preset size of [size missing] around it. Extract all candidate blocks within the search window; calculate the L2 Euclidean distance between the query block and the candidate blocks. The expression is as follows: ; In the formula, For query blocks; Candidate blocks; Select the one with the smallest distance A set of relative spatial indexes of each neighbor ; Across the image domain, A high-confidence spatial index maps back to the original high-resolution single-view complex image. Extract the raw complex pixel values ​​at these locations from the corresponding search window; After traversal and recombination, the constructed dimension is A homogeneous pixel library.

[0012] Preferably, the specific content of S3 is as follows: Two independent observations are randomly selected from K homogeneous pixels in a homogeneous pixel library and used as the network input and prediction label, respectively. Constructing a multivariate loss function The expression is as follows: ; In the formula, for and The Charbonnier loss is used to drive the network to learn the statistical properties of SAR speckle noise. for The error between the pixel library's physical center and the actual pixel center provides a stable optimization anchor point; To compensate for structural similarity loss, we calculate the structural similarity within a local window to prevent a decrease in image contrast. As a gradient loss, the gradient difference between the output and the physical center is calculated to ensure the sharpness of the edges of building and road contours; For total variation loss, the sum of absolute differences between adjacent pixels is calculated, and a smoothing constraint is applied to uniform regions to eliminate residual artifacts that resemble wires or grids. , , , and To balance the weights, manual hyperparameters are used to adjust the contribution of different loss terms; Backpropagation optimizes network weights, completing self-supervised training.

[0013] Preferably, the specific content of S4 is as follows: After the model converges under multiple loss functions, the weights of the MWCNN network parameters are fixed. During the inference stage, a sliding window scanning strategy with fixed edge overlap regions is adopted to input the original amplitude image into the network in blocks. The overlapping region stitching artifacts are eliminated by a weight averaging mechanism based on the counting matrix, resulting in a spatially seamless and high-purity global denoised amplitude map. ; Extracting the intrinsic complex phase matrix of the original high-resolution single-view complex image The absolute physical recombination strategy is adopted, which uses Euler's formula to recombine the denoised pure amplitude with the original phase, as shown in the following expression: ; In the formula, For phase components; The final output is a high-fidelity denoised SAR image that retains the complete complex algebraic structure.

[0014] Therefore, the present invention employs the above-mentioned self-supervised denoising method for SAR images in the complex domain, which combines physical priors, and has the following beneficial effects: (1) This invention provides a cross-domain pixel matching mechanism, which fundamentally reduces the interference caused by the spatial correlation of SAR real speckle, and successfully breaks through the bottleneck of "noise overfitting" without pure labels. The strategy of using only low-resolution sub-views for position guidance and directly obtaining full-resolution data from the original image enables the algorithm to retain high-frequency edges such as buildings and roads while suppressing background speckle in flat areas, thus achieving high-fidelity preservation of details in the denoising results.

[0015] (2) This invention, through the strong point mask protection mechanism and weighted physical center guidance, combined with multi-level wavelet network and gradient hybrid loss function, effectively overcomes the defect of traditional denoising network that easily darkens bright targets, maintains the radiation mean ratio before and after denoising, and can effectively protect the strong scattering points of the target.

[0016] (3) In the data construction and inference process, the present invention designs a memory-safe block segmentation mechanism based on tensor rearrangement and a block overlapping inference architecture, which greatly reduces the memory overhead during model operation, making it possible for this solution to process ultra-wide SAR remote sensing images with billions of pixels on consumer-grade graphics cards, and has high engineering feasibility as a general SAR preprocessing module.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a self-supervised denoising method for complex domain SAR images that incorporates physical priors, according to the present invention. Figure 2 This is a schematic diagram of the overall process of the present invention; Figure 3 A schematic diagram of the mechanism for building a cross-domain homogeneous pixel library. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0020] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0021] The following is combined with Figures 1-3 The embodiments of the present invention will be described in detail below.

[0022] Example 1 like Figure 1The diagram shows a flowchart of a self-supervised denoising method for complex domain SAR images based on physical priors, according to the present invention. The steps include: A two-dimensional Fourier transform is performed on the original single-look complex SAR image, and a physical sub-aperture decomposition strategy is adopted in the frequency domain to generate a sub-lookahead guide map that is statistically independent of the original noise. Spatial feature similarity is calculated using the generated sub-lookahead guide map to obtain a homogeneous pixel index with similar structure. Then, based on this index, numerical extraction is performed directly on the original high-resolution SLC image to construct a homogeneous pixel library containing multiple independent noise observations.

[0023] During the inference phase, the trained denoising network is used to perform block-overlapping inference on the original image to obtain the denoised amplitude, and combined with the original phase to perform complex domain physical reconstruction, outputting the final denoised SLC image.

[0024] The physical sub-aperture decomposition strategy does not directly divide the spectrum at the center, but rather removes the oversampled invalid frequency bands at both ends of the spectrum and forcibly discards the center frequency band where the spectral energy is most concentrated and the coupling is most severe. This physically blocks the frequency domain coupling caused by the non-ideal system response, ensuring that the speckle noise between the generated sub-view and the original image strictly satisfies statistical independence.

[0025] The homogeneous pixel library is constructed using a block-based streaming process. On the wizard graph, the L2 distance of local feature blocks is calculated through an unfolding operation to find the nearest neighbor index. This index is then mapped back to the original full-resolution image to extract the center pixel. This "cross-domain matching" mechanism ensures that the captured data retains high-frequency information at full resolution while avoiding mismatches caused by noise correlation in the original image.

[0026] The denoising network uses a multi-level complex wavelet neural network (MWCNN) and uses discrete wavelet transform (DWT) to replace the traditional pooling layer for downsampling, so as to retain the high-frequency subbands (LH, HL, HH) containing edges and noise without loss, thereby improving the network's ability to extract features from complex SAR textures.

[0027] Its overall process structure diagram is as follows: Figure 2 As shown, the specific steps are as follows: S1, Physical sub-aperture decomposition: Obtain the original high-resolution single-view complex image sequence (Original SLC), denoted as Perform a two-dimensional Fast Fourier Transform (FFT) on it to convert it to the frequency domain spectrum. .

[0028] In an ideal, noise-free scenario, Goodman's statistical optics theory assumes that the complex amplitudes of different frequency components are uncorrelated. However, due to the non-ideal roll-off effect of antenna patterns in real SAR imaging systems, and the oversampling operation used to avoid aliasing, simple frequency domain division leads to strong frequency coupling at the segmentation edges, resulting in significant noise correlation in the generated subviews.

[0029] To address this, the present invention designs a spectrum segmentation strategy with a guard gap: in the selected segmentation dimension (such as azimuth length)... On the spectrum, adaptive calculation and trimming of low-energy oversampled bands (Trim regions, length) at both ends of the spectrum are performed. ), and forcibly remove the center frequency band (Gap region, length) where the energy is most concentrated and the sidelobe leakage is most severe. The specific operation involves extracting the effective frequency band on one side: ; in To extract the interval as The rectangular window function. Only for... Perform Inverse Fourier Transform (IFFT) to generate a low-resolution subview wizard map. This strategy physically blocks frequency domain coupling, ensuring that the speckle pattern of the guide map and the speckle pattern of the remaining original signal are strictly statistically independent in the spatial domain, with the correlation coefficient approaching the ideal value of 0.

[0030] S2, Physically Guided Homogeneous Pixel Library Construction: Combination Figure 3 As shown. This aims to solve the "noise overfitting" deadlock caused by traditional self-supervised algorithms searching within a single noisy domain. To prevent memory overflow caused by ultra-large remote sensing images, a guide graph is used. And the original image Divide into preset sizes (e.g.) It processes tensor blocks and uses tensor unfolding to achieve streaming processing.

[0031] For each processing unit coordinate in the image First, in the statistical independent wizard graph Expanding on the amplitude feature to obtain the size is The query patch, with a preset size around it. Extract all candidate blocks within the search window. Calculate the L2 Euclidean distance between the query block and the candidate blocks: ; Select the one with the smallest distance A set of relative spatial indexes of each neighbor Since the noise in the wizard map generated by S1 is independent and unbiased, the similarity of the tiles in the wizard map highly represents the similarity of the real topological structure of the ground features, rather than the superposition of artifacts from spatially correlated noise.

[0032] Subsequently, across the image domain, this A high-confidence spatial index maps back to the original full-resolution SLC image. In the corresponding search window, extract the original complex pixel values ​​at these locations. After traversal and recombination, the resulting structure has dimensions of... The multidimensional tensors constitute a homogeneous pixel bank. This mechanism not only utilizes the physical independence of the guide graph to achieve unbiased structural localization, but also preserves the original image through "cross-domain grasping". Full-resolution broadband textures.

[0033] S3. Hybrid constraint training based on Multi-level Wavelet Network (MWCNN): To fully extract the complex textures of SAR images, the denoising network employs a MWCNN architecture based on Discrete Wavelet Transform (DWT). Unlike the pooling layers of traditional CNNs, which lose information, DWT possesses orthogonal invertibility, enabling lossless decomposition of feature maps into low-frequency subbands and multiple high-frequency directional subbands. This allows the network to accurately separate edge structures and high-frequency speckle in the deep feature space.

[0034] During self-supervised training, two independent observations are randomly selected from K homogeneous pixels in the pixel library each time, serving as the network's input and prediction label, respectively. Since both contain the same real-world features but have independent and uncorrelated random speckle patterns, the network learns to output clean ground feature signals to minimize prediction error.

[0035] To prevent excessive smoothing of the network from causing the image to lose sharpness, this scheme constructs a multi-variable loss function: ; in: for and The Charbonnier loss is used to drive the network to learn the statistical properties of SAR speckle noise. for The error between the pixel library's physical center and the actual pixel center provides a stable optimization anchor point; Calculate structural similarity within a local window to prevent image contrast degradation; Calculate the gradient difference between the output and the physical center to ensure the sharpness of the edges of building and road outlines; The sum of absolute differences between adjacent pixels is calculated, and a smoothing constraint is applied to uniform regions to eliminate residual artifacts that resemble streaks or grids.

[0036] S4. Full-image denoising inference and physical phase reconstruction: After the model converges under multiple loss function-driven conditions, the weights of the MWCNN network parameters are fixed. To adapt to the processing of ultra-large remote sensing images without memory limitations, a sliding window scanning strategy with fixed edge overlap regions is adopted during the inference stage. The original amplitude image is input into the network in blocks, and the overlapping region stitching artifacts are eliminated through a weight averaging mechanism based on the counting matrix, thereby obtaining a spatially seamless, high-purity global denoised amplitude map. .

[0037] Finally, in order to preserve the complex properties of the data, the intrinsic complex phase matrix of the original SLC image is extracted. Because the phase of SAR images contains valuable physical information such as elevation, microwave polarization, and surface deformation, and exhibits characteristics that conform to... Given uniformly distributed, strongly random variables, this invention abandons direct network prediction of phase and instead employs an absolute physical recombination strategy. The denoised, purified amplitude is recombined with the original phase using Euler's formula: ; The final output is a high-fidelity denoised SAR image that retains the complete complex algebraic structure. This strategy enhances the visual appeal of the result while preserving the physical properties of the SAR image.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A self-supervised denoising method for complex domain SAR images that incorporates physical priors, characterized in that, Includes the following steps: S1. Generate a sub-view guide graph that is statistically independent of the original noise based on a physical sub-aperture decomposition strategy; S2. Construct a homogeneous pixel library containing multiple independent noise observations based on the sub-view guide map generated in S1; S3. Hybrid constraint training based on multi-level wavelet network MWCNN; S4. Perform full-image denoising inference and physical phase reconstruction.

2. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 1, characterized in that, S1 specifically includes: S11. Obtain the original high-resolution single-view complex image sequence and perform a two-dimensional fast Fourier transform; S12. Generate a low-resolution subview wizard map.

3. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 2, characterized in that, The specific content of S11 is as follows: Obtain the original high-resolution single-view complex image sequence, denoted as ; A two-dimensional fast Fourier transform is performed on the original high-resolution single-view complex image to convert it to the frequency domain spectrum. In the formula, For azimuth frequency; For range-direction frequency.

4. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 3, characterized in that, The specific content of S12 is as follows: Based on the spectrum segmentation strategy with guard gaps, the low-energy oversampling bands at both ends of the spectrum are removed; Based on the selected splitting dimension Adaptively calculate and remove the low-energy oversampled frequency band Trim regions at both ends of the spectrum, with a length of [missing information]. And forcibly remove the center frequency band gap region with the most concentrated energy and the most severe sidelobe leakage, with a length of ; Extracting the effective frequency band of one side The expression is as follows: ； In the formula, The function is a rectangular window; the extraction interval of the rectangular window function is... ; Only for Perform inverse Fourier transform to generate low-resolution subview wizard maps .

5. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 4, characterized in that, The specific details of S2 are as follows: Low-resolution subview wizard map and the original high-resolution single-view complex image sequence Divide the data into tensor processing blocks of a preset size and perform tensor expansion operations; For each processing unit coordinate in the image In statistically independent low-resolution subview wizard maps Expanding on the amplitude feature to obtain the size is The query block; and a preset size of [size missing] around it. Extract all candidate blocks within the search window; calculate the L2 Euclidean distance between the query block and the candidate blocks. The expression is as follows: ； In the formula, For query blocks; Candidate blocks; Select the one with the smallest distance A set of relative spatial indexes of each neighbor ; Across the image domain, A high-confidence spatial index maps back to the original high-resolution single-view complex image. Extract the raw complex pixel values ​​at these locations from the corresponding search window; After traversal and recombination, the constructed dimension is A homogeneous pixel library.

6. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 5, characterized in that, The specific details of S3 are as follows: Two independent observations are randomly selected from K homogeneous pixels in a homogeneous pixel library and used as the network input and prediction label, respectively. Constructing a multivariate loss function The expression is as follows: ； In the formula, for and The Charbonnier loss is used to drive the network to learn the statistical properties of SAR speckle noise. for The error between the pixel library's physical center and the actual pixel center provides a stable optimization anchor point; To mitigate structural similarity loss, structural similarity within a local window is calculated to prevent image contrast degradation. As a gradient loss, the gradient difference between the output and the physical center is calculated to ensure the sharpness of the edges of building and road outlines; For total variation loss, the sum of absolute differences between adjacent pixels is calculated, and a smoothing constraint is applied to uniform regions to eliminate residual artifacts that resemble wires or grids. , , , and To balance the weights, manual hyperparameters are used to adjust the contribution of different loss terms; Backpropagation optimizes network weights, completing self-supervised training.

7. The SAR image complex domain self-supervised denoising method combining physical priors as described in claim 6, characterized in that, The specific details of S4 are as follows: After the model converges under multiple loss functions, the weights of the MWCNN network parameters are fixed. During the inference stage, a sliding window scanning strategy with fixed edge overlap regions is adopted to input the original amplitude image into the network in blocks. The overlapping region stitching artifacts are eliminated by a weight averaging mechanism based on the counting matrix, resulting in a spatially seamless and high-purity global denoised amplitude map. ; Extracting the intrinsic complex phase matrix of the original high-resolution single-view complex image The absolute physical recombination strategy is adopted, which uses Euler's formula to recombine the denoised pure amplitude with the original phase, as shown in the following expression: ； In the formula, For phase components; The final output is a high-fidelity denoised SAR image that retains the complete complex algebraic structure.

Images